A Guide To Modern Econometrics

User Manual:

Open the PDF directly: View PDF .
Page Count: 448

Download
Open PDF In Browser	View PDF

A Guide to
Modern
Econometrics
2nd edition

Marno Verbeek
Erasmus University Rotterdam

A Guide to
Modern
Econometrics

A Guide to
Modern
Econometrics
2nd edition

Marno Verbeek
Erasmus University Rotterdam

Copyright  2004

John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester,
West Sussex PO19 8SQ, England
Telephone (+44) 1243 779777

Email (for orders and customer service enquiries): cs-books@wiley.co.uk
Visit our Home Page on www.wileyeurope.com or www.wiley.com
All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system or
transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning
or otherwise, except under the terms of the Copyright, Designs and Patents Act 1988 or under the
terms of a licence issued by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road,
London W1T 4LP, UK, without the permission in writing of the Publisher. Requests to the
Publisher should be addressed to the Permissions Department, John Wiley & Sons Ltd, The Atrium,
Southern Gate, Chichester, West Sussex PO19 8SQ, England, or emailed to permreq@wiley.co.uk,
or faxed to (+44) 1243 770620.
This publication is designed to provide accurate and authoritative information in regard to the
subject matter covered. It is sold on the understanding that the Publisher is not engaged in rendering
professional services. If professional advice or other expert assistance is required, the services of a
competent professional should be sought.
Other Wiley Editorial Ofﬁces
John Wiley & Sons Inc., 111 River Street, Hoboken, NJ 07030, USA
Jossey-Bass, 989 Market Street, San Francisco, CA 94103-1741, USA
Wiley-VCH Verlag GmbH, Boschstr. 12, D-69469 Weinheim, Germany
John Wiley & Sons Australia Ltd, 33 Park Road, Milton, Queensland 4064, Australia
John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02-01, Jin Xing Distripark, Singapore 129809
John Wiley & Sons Canada Ltd, 22 Worcester Road, Etobicoke, Ontario, Canada M9W 1L1
Wiley also publishes its books in a variety of electronic formats. Some content that appears
in print may not be available in electronic books.
Library of Congress Cataloging-in-Publication Data
Verbeek, Marno.
A guide to modern econometrics / Marno Verbeek. – 2nd ed.
p. cm.
Includes bibliographical references and index.
ISBN 0-470-85773-0 (pbk. : alk. paper)
1. Econometrics. 2. Regression analysis. I. Title.
HB139.V465 2004
330 .01 5195 – dc22
2004004222
British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library
ISBN 0-470-85773-0
Typeset in 10/12pt Times by Laserwords Private Limited, Chennai, India
Printed and bound in Great Britain by TJ International, Padstow, Cornwall
This book is printed on acid-free paper responsibly manufactured from sustainable forestry
in which at least two trees are planted for each one used for paper production.

Contents
Preface
1 Introduction
1.1 About Econometrics
1.2 The Structure of this Book
1.3 Illustrations and Exercises
2 An Introduction to Linear Regression
2.1 Ordinary Least Squares as an Algebraic Tool
2.1.1 Ordinary Least Squares
2.1.2 Simple Linear Regression
2.1.3 Example: Individual Wages
2.1.4 Matrix Notation
2.2 The Linear Regression Model
2.3 Small Sample Properties of the OLS Estimator
2.3.1 The Gauss–Markov Assumptions
2.3.2 Properties of the OLS Estimator
2.3.3 Example: Individual Wages (Continued)
2.4 Goodness-of-ﬁt
2.5 Hypothesis Testing
2.5.1 A Simple t-test
2.5.2 Example: Individual Wages (Continued)
2.5.3 Testing One Linear Restriction
2.5.4 A Joint Test of Signiﬁcance of Regression Coefﬁcients
2.5.5 Example: Individual Wages (Continued)
2.5.6 The General Case
2.5.7 Size, Power and p-Values

xiii
1
1
3
4
7
8
8
10
12
12
14
16
16
17
20
20
23
23
25
25
27
28
30
31

CONTENTS

vi

2.6 Asymptotic Properties of the OLS Estimator
2.6.1 Consistency
2.6.2 Asymptotic Normality
2.6.3 Small Samples and Asymptotic Theory
2.7 Illustration: The Capital Asset Pricing Model
2.7.1 The CAPM as a Regression Model
2.7.2 Estimating and Testing the CAPM
2.8 Multicollinearity
2.8.1 Example: Individual Wages (Continued)
2.9 Prediction
Exercises

32
32
34
36
38
38
39
42
44
44
46

3 Interpreting and Comparing Regression Models
3.1 Interpreting the Linear Model
3.2 Selecting the Set of Regressors
3.2.1 Misspecifying the Set of Regressors
3.2.2 Selecting Regressors
3.2.3 Comparing Non-nested Models
3.3 Misspecifying the Functional Form
3.3.1 Nonlinear Models
3.3.2 Testing the Functional Form
3.3.3 Testing for a Structural Break
3.4 Illustration: Explaining House Prices
3.5 Illustration: Explaining Individual Wages
3.5.1 Linear Models
3.5.2 Loglinear Models
3.5.3 The Effects of Gender
3.5.4 Some Words of Warning
Exercises

51
51
55
55
56
59
62
62
63
63
65
68
68
71
74
76
77

4 Heteroskedasticity and Autocorrelation
4.1 Consequences for the OLS Estimator
4.2 Deriving an Alternative Estimator
4.3 Heteroskedasticity
4.3.1 Introduction
4.3.2 Estimator Properties and Hypothesis Testing
4.3.3 When the Variances are Unknown
4.3.4 Heteroskedasticity-consistent Standard Errors for OLS
4.3.5 A Model with Two Unknown Variances
4.3.6 Multiplicative Heteroskedasticity
4.4 Testing for Heteroskedasticity
4.4.1 Testing Equality of Two Unknown Variances
4.4.2 Testing for Multiplicative Heteroskedasticity
4.4.3 The Breusch–Pagan Test
4.4.4 The White Test
4.4.5 Which Test?

79
79
81
82
82
84
85
87
88
89
90
90
91
91
92
92

CONTENTS

4.5 Illustration: Explaining Labour Demand
4.6 Autocorrelation
4.6.1 First Order Autocorrelation
4.6.2 Unknown ρ
4.7 Testing for First Order Autocorrelation
4.7.1 Asymptotic Tests
4.7.2 The Durbin–Watson Test
4.8 Illustration: The Demand for Ice Cream
4.9 Alternative Autocorrelation Patterns
4.9.1 Higher Order Autocorrelation
4.9.2 Moving Average Errors
4.10 What to do When you Find Autocorrelation?
4.10.1 Misspeciﬁcation
4.10.2 Heteroskedasticity-and-autocorrelation-consistent
Standard Errors for OLS
4.11 Illustration: Risk Premia in Foreign Exchange Markets
4.11.1 Notation
4.11.2 Tests for Risk Premia in the One-month Market
4.11.3 Tests for Risk Premia Using Overlapping Samples
Exercises
5 Endogeneity, Instrumental Variables and GMM
5.1 A Review of the Properties of the OLS Estimator
5.2 Cases Where the OLS Estimator Cannot be Saved
5.2.1 Autocorrelation with a Lagged Dependent Variable
5.2.2 An Example with Measurement Error
5.2.3 Simultaneity: the Keynesian Model
5.3 The Instrumental Variables Estimator
5.3.1 Estimation with a Single Endogenous Regressor
and a Single Instrument
5.3.2 Back to the Keynesian Model
5.3.3 Back to the Measurement Error Problem
5.3.4 Multiple Endogenous Regressors
5.4 Illustration: Estimating the Returns to Schooling
5.5 The Generalized Instrumental Variables Estimator
5.5.1 Multiple Endogenous Regressors with an Arbitrary
Number of Instruments
5.5.2 Two-stage Least Squares and the Keynesian Model
Again
5.5.3 Speciﬁcation Tests
5.5.4 Weak Instruments
5.6 The Generalized Method of Moments
5.6.1 Example
5.6.2 The Generalized Method of Moments
5.6.3 Some Simple Examples
5.7 Illustration: Estimating Intertemporal Asset
Pricing Models

vii

92
97
98
100
101
101
102
103
106
106
107
108
108
110
112
112
113
116
119
121
122
125
126
127
129
131
131
135
136
136
137
142
142
145
146
147
148
149
150
153
154

CONTENTS

viii

5.8 Concluding Remarks
Exercises

157
158

6 Maximum Likelihood Estimation and Speciﬁcation Tests
6.1 An Introduction to Maximum Likelihood
6.1.1 Some Examples
6.1.2 General Properties
6.1.3 An Example (Continued)
6.1.4 The Normal Linear Regression Model
6.2 Speciﬁcation Tests
6.2.1 Three Test Principles
6.2.2 Lagrange Multiplier Tests
6.2.3 An Example (Continued)
6.3 Tests in the Normal Linear Regression Model
6.3.1 Testing for Omitted Variables
6.3.2 Testing for Heteroskedasticity
6.3.3 Testing for Autocorrelation
6.4 Quasi-maximum Likelihood and Moment Conditions Tests
6.4.1 Quasi-maximum Likelihood
6.4.2 Conditional Moment Tests
6.4.3 Testing for Normality
Exercises

161
162
162
166
169
170
171
171
173
177
178
178
179
181
182
182
184
185
186

7 Models with Limited Dependent Variables
7.1 Binary Choice Models
7.1.1 Using Linear Regression?
7.1.2 Introducing Binary Choice Models
7.1.3 An Underlying Latent Model
7.1.4 Estimation
7.1.5 Goodness-of-ﬁt
7.1.6 Illustration: the Impact of Unemployment Beneﬁts on
Recipiency
7.1.7 Speciﬁcation Tests in Binary Choice Models
7.1.8 Relaxing Some Assumptions in Binary Choice Models
7.2 Multi-response Models
7.2.1 Ordered Response Models
7.2.2 About Normalization
7.2.3 Illustration: Willingness to Pay for Natural Areas
7.2.4 Multinomial Models
7.3 Models for Count Data
7.3.1 The Poisson and Negative Binomial Models
7.3.2 Illustration: Patents and R&D Expenditures
7.4 Tobit Models
7.4.1 The Standard Tobit Model
7.4.2 Estimation

189
190
190
190
192
193
194
197
199
201
202
203
204
205
208
211
211
215
218
218
220

CONTENTS

7.5

7.6

7.7
7.8

ix

7.4.3 Illustration: Expenditures on Alcohol and Tobacco
(Part 1)
7.4.4 Speciﬁcation Tests in the Tobit Model
Extensions of Tobit Models
7.5.1 The Tobit II Model
7.5.2 Estimation
7.5.3 Further Extensions
7.5.4 Illustration: Expenditures on Alcohol and Tobacco
(Part 2)
Sample Selection Bias
7.6.1 The Nature of the Selection Problem
7.6.2 Semi-parametric Estimation of the Sample Selection
Model
Estimating Treatment Effects
Duration Models
7.8.1 Hazard Rates and Survival Functions
7.8.2 Samples and Model Estimation
7.8.3 Illustration: Duration of Bank Relationships
Exercises

8 Univariate Time Series Models
8.1 Introduction
8.1.1 Some Examples
8.1.2 Stationarity and the Autocorrelation Function
8.2 General ARMA Processes
8.2.1 Formulating ARMA Processes
8.2.2 Invertibility of Lag Polynomials
8.2.3 Common Roots
8.3 Stationarity and Unit Roots
8.4 Testing for Unit Roots
8.4.1 Testing for Unit Roots in a First Order Autoregressive
Model
8.4.2 Testing for Unit Roots in Higher Order Autoregressive
Models
8.4.3 Extensions
8.4.4 Illustration: Annual Price/Earnings Ratio
8.5 Illustration: Long-run Purchasing Power Parity (Part 1)
8.6 Estimation of ARMA Models
8.6.1 Least Squares
8.6.2 Maximum Likelihood
8.7 Choosing a Model
8.7.1 The Autocorrelation Function
8.7.2 The Partial Autocorrelation Function
8.7.3 Diagnostic Checking
8.7.4 Criteria for Model Selection
8.7.5 Illustration: Modelling the Price/Earnings Ratio

222
225
227
228
230
232
233
237
237
239
240
244
245
247
249
251
255
256
256
258
261
261
264
265
266
268
269
271
273
274
276
279
279
280
281
281
283
284
285
286

CONTENTS

x

8.8 Predicting with ARMA Models
8.8.1 The Optimal Predictor
8.8.2 Prediction Accuracy
8.9 Illustration: The Expectations Theory of the Term Structure
8.10 Autoregressive Conditional Heteroskedasticity
8.10.1 ARCH and GARCH Models
8.10.2 Estimation and Prediction
8.10.3 Illustration: Volatility in Daily Exchange Rates
8.11 What about Multivariate Models?
Exercises

288
288
291
293
297
298
301
303
305
306

9 Multivariate Time Series Models
9.1 Dynamic Models with Stationary Variables
9.2 Models with Nonstationary Variables
9.2.1 Spurious Regressions
9.2.2 Cointegration
9.2.3 Cointegration and Error-correction Mechanisms
9.3 Illustration: Long-run Purchasing Power Parity (Part 2)
9.4 Vector Autoregressive Models
9.5 Cointegration: the Multivariate Case
9.5.1 Cointegration in a VAR
9.5.2 Example: Cointegration in a Bivariate VAR
9.5.3 Testing for Cointegration
9.5.4 Illustration: Long-run Purchasing Power Parity (Part 3)
9.6 Illustration: Money Demand and Inﬂation
9.7 Concluding Remarks
Exercises

309
310
313
313
314
318
319
321
324
325
327
328
331
333
339
339

10 Models Based on Panel Data
10.1 Advantages of Panel Data
10.1.1 Efﬁciency of Parameter Estimators
10.1.2 Identiﬁcation of Parameters
10.2 The Static Linear Model
10.2.1 The Fixed Effects Model
10.2.2 The Random Effects Model
10.2.3 Fixed Effects or Random Effects?
10.2.4 Goodness-of-ﬁt
10.2.5 Alternative Instrumental Variables Estimators
10.2.6 Robust Inference
10.2.7 Testing for Heteroskedasticity and Autocorrelation
10.3 Illustration: Explaining Individual Wages
10.4 Dynamic Linear Models
10.4.1 An Autoregressive Panel Data Model
10.4.2 Dynamic Models with Exogenous Variables
10.5 Illustration: Wage Elasticities of Labour Demand

341
342
343
344
345
345
347
351
352
353
355
357
358
360
360
365
366

CONTENTS

10.6 Nonstationarity, Unit Roots and Cointegration
10.6.1 Panel Data Unit Root Tests
10.6.2 Panel Data Cointegration Tests
10.7 Models with Limited Dependent Variables
10.7.1 Binary Choice Models
10.7.2 The Fixed Effects Logit Model
10.7.3 The Random Effects Probit Model
10.7.4 Tobit Models
10.7.5 Dynamics and the Problem of Initial Conditions
10.7.6 Semi-parametric Alternatives
10.8 Incomplete Panels and Selection Bias
10.8.1 Estimation with Randomly Missing Data
10.8.2 Selection Bias and Some Simple Tests
10.8.3 Estimation with Nonrandomly Missing Data
Exercises

xi

368
369
372
373
373
375
376
377
378
380
380
381
383
385
385

A Vectors and Matrices
A.1 Terminology
A.2 Matrix Manipulations
A.3 Properties of Matrices and Vectors
A.4 Inverse Matrices
A.5 Idempotent Matrices
A.6 Eigenvalues and Eigenvectors
A.7 Differentiation
A.8 Some Least Squares Manipulations

389
389
390
391
392
393
394
394
395

B

397
397
398
399
400
401
403
405

Statistical and Distribution Theory
B.1 Discrete Random Variables
B.2 Continuous Random Variables
B.3 Expectations and Moments
B.4 Multivariate Distributions
B.5 Conditional Distributions
B.6 The Normal Distribution
B.7 Related Distributions

Bibliography

409

Index

421

Preface
Emperor Joseph II: “Your work is ingenious. It’s quality work. And there are simply
too many notes, that’s all. Just cut a few and it will be perfect.”
Wolfgang Amadeus Mozart: “Which few did you have in mind, Majesty?”
from the movie Amadeus, 1984 (directed by Milos Forman)

The ﬁeld of econometrics has developed rapidly in the last two decades, while the use
of up-to-date econometric techniques has become more and more standard practice in
empirical work in many ﬁelds of economics. Typical topics include unit root tests,
cointegration, estimation by the generalized method of moments, heteroskedasticity
and autocorrelation consistent standard errors, modelling conditional heteroskedasticity,
models based on panel data, and models with limited dependent variables, endogenous regressors and sample selection. At the same time econometrics software has
become more and more user friendly and up-to-date. As a consequence, users are
able to implement fairly advanced techniques even without a basic understanding of
the underlying theory and without realizing potential drawbacks or dangers. In contrast, many introductory econometrics textbooks pay a disproportionate amount of
attention to the standard linear regression model under the strongest set of assumptions. Needless to say that these assumptions are hardly satisﬁed in practice (but
not really needed either). On the other hand, the more advanced econometrics textbooks are often too technical or too detailed for the average economist to grasp the
essential ideas and to extract the information that is needed. This book tries to ﬁll
this gap.
The goal of this book is to familiarize the reader with a wide range of topics
in modern econometrics, focusing on what is important for doing and understanding
empirical work. This means that the text is a guide to (rather than an overview of)
alternative techniques. Consequently, it does not concentrate on the formulae behind
each technique (although the necessary ones are given) nor on formal proofs, but on
the intuition behind the approaches and their practical relevance. The book covers a
wide range of topics that is usually not found in textbooks at this level. In particular, attention is paid to cointegration, the generalized method of moments, models

xiv

PREFACE

with limited dependent variables and panel data models. As a result, the book discusses developments in time series analysis, cross-sectional methods as well as panel
data modelling. Throughout, a few dozen full-scale empirical examples and illustrations are provided, taken from ﬁelds like labour economics, ﬁnance, international
economics, consumer behaviour, environmental economics and macro-economics. In
addition, a number of exercises are of an empirical nature and require the use of
actual data.
For the second edition, I have tried to ﬁne-tune and update the text, adding additional
discussion, material and more recent references, whenever necessary or desirable. The
material is organized and presented in a similar way as in the ﬁrst edition. Some topics
that were not or only limitedly included in the ﬁrst edition now receive much more
attention. Most notably, new sections covering count data models, duration models and
the estimation of treatment effects in Chapter 7, and panel data unit root and cointegration tests in Chapter 10 are added. Moreover, Chapter 2 now includes a subsection
on Monte Carlo simulation. At several places, I pay more attention to the possibility
that small sample distributions of estimators and test statistics may differ from their
asymptotic approximations. Several new tests have been added to Chapters 3 and 5,
and the presentation in Chapters 6 and 8 has been improved. At a number of places,
empirical illustrations have been updated or added. As before, (almost) all data sets
are available through the book’s website.
This text originates from lecture notes used for courses in Applied Econometrics in
the M.Sc. programs in Economics at K. U. Leuven and Tilburg University. It is written for an intended audience of economists and economics students that would like to
become familiar with up-to-date econometric approaches and techniques, important for
doing, understanding and evaluating empirical work. It is very well suited for courses
in applied econometrics at the masters or graduate level. At some schools this book
will be suited for one or more courses at the undergraduate level, provided students
have a sufﬁcient background in statistics. Some of the later chapters can be used in
more advanced courses covering particular topics, for example, panel data, limited
dependent variable models or time series analysis. In addition, this book can serve as
a guide for managers, research economists and practitioners who want to update their
insufﬁcient or outdated knowledge of econometrics. Throughout, the use of matrix
algebra is limited.
I am very much indebted to Arie Kapteyn, Bertrand Melenberg, Theo Nijman, and
Arthur van Soest, who all have contributed to my understanding of econometrics and
have shaped my way of thinking about many issues. The fact that some of their ideas
have materialized in this text is a tribute to their efforts. I also owe many thanks to
several generations of students who helped me to shape this text into its current form. I
am very grateful to a large number of people who read through parts of the manuscript
and provided me with comments and suggestions on the basis of the ﬁrst edition. In
particular, I wish to thank Peter Boswijk, Bart Capéau, Geert Dhaene, Tom Doan,
Peter de Goeij, Joop Huij, Ben Jacobsen, Jan Kiviet, Wim Koevoets, Erik Kole, Marco
Lyrio, Konstantijn Maes, Wessel Marquering, Bertrand Melenberg, Paulo Nunes, Anatoly Peresetsky, Max van de Sande Bakhuyzen, Erik Schokkaert, Arthur van Soest,
Frederic Vermeulen, Guglielmo Weber, Olivier Wolthoorn, Kuo-chun Yeh and a number of anonymous reviewers. Of course I retain sole responsibility for any remaining

PREFACE

xv

errors. Special thanks go to Jef Flechet for his help with many empirical illustrations
and his constructive comments on many previous versions. Finally, I want to thank
my wife Marcella and our three children, Timo, Thalia and Tamara, for their patience
and understanding for all the times that my mind was with this book, while it should
have been with them.

1
1.1

Introduction

About Econometrics

Economists are frequently interested in relationships between different quantities, for
example between individual wages and the level of schooling. The most important job
of econometrics is to quantify these relationships on the basis of available data and
using statistical techniques, and to interpret, use or exploit the resulting outcomes appropriately. Consequently, econometrics is the interaction of economic theory, observed
data and statistical methods. It is the interaction of these three that makes econometrics interesting, challenging and, perhaps, difﬁcult. In the words of a seminar speaker,
several years ago: ‘Econometrics is much easier without data’.
Traditionally econometrics has focused upon aggregate economic relationships.
Macro-economic models consisting of several up to many hundreds equations were
speciﬁed, estimated and used for policy evaluation and forecasting. The recent
theoretical developments in this area, most importantly the concept of cointegration,
have generated increased attention to the modelling of macro-economic relationships
and their dynamics, although typically focusing on particular aspects of the economy.
Since the 1970s econometric methods are increasingly employed in micro-economic
models describing individual, household or ﬁrm behaviour, stimulated by the
development of appropriate econometric models and estimators which take into account
problems like discrete dependent variables and sample selection, by the availability of
large survey data sets, and by the increasing computational possibilities. More recently,
the empirical analysis of ﬁnancial markets has required and stimulated many theoretical
developments in econometrics. Currently econometrics plays a major role in empirical
work in all ﬁelds of economics, almost without exception, and in most cases it is no
longer sufﬁcient to be able to run a few regressions and interpret the results. As a
result, introductory econometrics textbooks usually provide insufﬁcient coverage for
applied researchers. On the other hand, the more advanced econometrics textbooks are
often too technical or too detailed for the average economist to grasp the essential ideas
and to extract the information that is needed. Thus there is a need for an accessible
textbook that discusses the recent and relatively more advanced developments.

2

INTRODUCTION

The relationships that economists are interested in are formally speciﬁed in mathematical terms, which lead to econometric or statistical models. In such models there is
room for deviations from the strict theoretical relationships due to, for example, measurement errors, unpredictable behaviour, optimization errors or unexpected events.
Broadly, econometric models can be classiﬁed in a number of categories.
A ﬁrst class of models describes relationships between present and past. For example,
how does the short-term interest rate depend on its own history? This type of model,
typically referred to as a time series model, usually lacks any economic theory and
is mainly built to get forecasts for future values and the corresponding uncertainty
or volatility.
A second type of model considers relationships between economic quantities over a
certain time period. These relationships give us information on how (aggregate) economic quantities ﬂuctuate over time in relation to other quantities. For example, what
happens to the long-term interest rate if the monetary authority adjusts the short-term
one? These models often give insight into the economic processes that are operating.
Third, there are models that describe relationships between different variables measured at a given point in time for different units (for example households or ﬁrms).
Most of the time, this type of relationship is meant to explain why these units are different or behave differently. For example, one can analyse to what extent differences in
household savings can be attributed to differences in household income. Under particular conditions, these cross-sectional relationships can be used to analyse ‘what if’
questions. For example, how much more would a given household, or the average
household, save if income would increase by 1%?
Finally, one can consider relationships between different variables measured for
different units over a longer time span (at least two periods). These relationships
simultaneously describe differences between different individuals (why does person 1
save much more than person 2?), and differences in behaviour of a given individual over
time (why does person 1 save more in 1992 than in 1990?). This type of model usually
requires panel data, repeated observations over the same units. They are ideally suited
for analysing policy changes on an individual level, provided that it can be assumed
that the structure of the model is constant into the (near) future.
The job of econometrics is to specify and quantify these relationships. That is, econometricians formulate a statistical model, usually based on economic theory, confront it
with the data, and try to come up with a speciﬁcation that meets the required goals. The
unknown elements in the speciﬁcation, the parameters, are estimated from a sample
of available data. Another job of the econometrician is to judge whether the resulting
model is ‘appropriate’. That is, check whether the assumptions made to motivate the
estimators (and their properties) are correct, and check whether the model can be used
for what it is made for. For example, can it be used for prediction or analysing policy
changes? Often, economic theory implies that certain restrictions apply to the model
that is estimated. For example, (one version of) the efﬁcient market hypothesis implies
that stock market returns are not predictable from their own past. An important goal of
econometrics is to formulate such hypotheses in terms of the parameters in the model
and to test their validity.
The number of econometric techniques that can be used is numerous and their validity often depends crucially upon the validity of the underlying assumptions. This book
attempts to guide the reader through this forest of estimation and testing procedures,

THE STRUCTURE OF THIS BOOK

3

not by describing the beauty of all possible trees, but by walking through this forest
in a structured way, skipping unnecessary side-paths, stressing the similarity of the
different species that are encountered, and by pointing out dangerous pitfalls. The
resulting walk is hopefully enjoyable and prevents the reader from getting lost in the
econometric forest.

1.2

The Structure of this Book

The ﬁrst part of this book consists of Chapters 2, 3 and 4. Like most textbooks, it starts
with discussing the linear regression model and the OLS estimation method. Chapter 2
presents the basics of this important estimation method, with some emphasis on its
validity under fairly weak conditions, while Chapter 3 focuses on the interpretation of
the models and the comparison of alternative speciﬁcations. Chapter 4 considers two
particular deviations from the standard assumptions of the linear model: autocorrelation and heteroskedasticity of the error terms. It is discussed how one can test for
these phenomena, how they affect the validity of the OLS estimator and how this can
be corrected. This includes a critical inspection of the model speciﬁcation, the use
of adjusted standard errors for the OLS estimator and the use of alternative (GLS)
estimators. These three chapters are essential for the remaining part of this book and
should be the starting point in any course.
In Chapter 5 another deviation from the standard assumptions of the linear model is
discussed which is, however, fatal for the OLS estimator. As soon as the error term in
the model is correlated with one or more of the explanatory variables all good properties
of the OLS estimator disappear and we necessarily have to use alternative estimators.
The chapter discusses instrumental variables (IV) estimators and, more generally, the
generalized method of moments (GMM). This chapter, at least its earlier sections, is
also recommended as an essential part of any econometrics course.
Chapter 6 is mainly theoretical and discusses maximum likelihood (ML) estimation.
Because in empirical work maximum likelihood is often criticized for its dependence
upon distributional assumptions, it is not discussed in the earlier chapters where alternatives are readily available that are either more robust than maximum likelihood or
(asymptotically) equivalent to it. Particular emphasis in Chapter 6 is on misspeciﬁcation tests based upon the Lagrange multiplier principle. While many empirical studies
tend to take the distributional assumptions for granted, their validity is crucial for consistency of the estimators that are employed and should therefore be tested. Often these
tests are relatively easy to perform, although most software does not routinely provide
them (yet). Chapter 6 is crucial for understanding Chapter 7 on limited dependent
variable models and for a small number of sections in Chapters 8 to 10.
The last part of this book contains four chapters. Chapter 7 presents models that
are typically (though not exclusively) used in micro-economics, where the dependent
variable is discrete (e.g. zero or one), partly discrete (e.g. zero or positive) or a duration.
It also includes discussions of the sample selection problem and the estimation of
treatment effects that go further than their typical textbook treatment.
Chapters 8 and 9 discuss time series modelling including unit roots, cointegration
and error-correction models. These chapters can be read immediately after Chapter 4 or
5, with the exception of a few parts that relate to maximum likelihood estimation. The

INTRODUCTION

4

theoretical developments in this area over the last 20 years have been substantial and
many recent textbooks seem to focus upon it almost exclusively. Univariate time series
models are covered in Chapter 8. In this case models are developed that explain an
economic variable from its own past. This includes ARIMA models, as well as GARCH
models for the conditional variance of a series. Multivariate time series models that
consider several variables simultaneously are discussed in Chapter 9. This includes
vector autoregressive models, cointegration and error-correction models.
Finally, Chapter 10 covers models based on panel data. Panel data are available if
we have repeated observations of the same units (for example households, ﬁrms or
countries). The last decade the use of panel data has become important in many areas
of economics. Micro-economic panels of households and ﬁrms are readily available
and, given the increase in computing resources, more manageable than in the past. In
addition, it is more and more common to pool time series of several countries. One of
the reasons for this may be that researchers believe that a cross-sectional comparison
of countries provides interesting information, in addition to a historical comparison of
a country with its own past. This chapter also discusses the recent developments on
unit roots and cointegration in a panel data setting.
At the end of the book the reader will ﬁnd two short appendices discussing mathematical and statistical results that are used at several places in the book. This includes
a discussion of some relevant matrix algebra and distribution theory. In particular, a
discussion of properties of the (bivariate) normal distribution, including conditional
expectations, variances and truncation is provided.
In my experience the material in this book is too much to be covered in a single course. Different courses can be scheduled on the basis of the chapters that
follow. For example, a typical graduate course in applied econometrics would cover
Chapters 2, 3, 4, parts of Chapter 5, and then continue with selected parts of Chapters 8
and 9 if the focus is on time series analysis, or continue with Section 6.1 and Chapter 7
if the focus is on cross-sectional models. A more advanced undergraduate or graduate
course may focus attention to the time series chapters (Chapters 8 and 9), the microeconometric chapters (Chapters 6 and 7) or panel data (Chapter 10 with some selected
parts from Chapters 6 and 7).
Given the focus and length of this book, I had to make many choices of which
material to present or not. As a general rule I did not want to bother the reader with
details that I considered not essential or do not have empirical relevance. The main
goal was to give a general and comprehensive overview of the different methodologies
and approaches, focusing on what is relevant for doing and understanding empirical
work. Some topics are only very brieﬂy mentioned and no attempt is made to discuss
them at any length. To compensate for this I have tried to give references at appropriate
places to other, often more advanced, textbooks that do cover these issues.

1.3

Illustrations and Exercises

In most chapters a variety of empirical illustrations is provided in separate sections or
subsections. While it is possible to skip these illustrations essentially without losing
continuity, these sections do provide important aspects concerning the implementation
of the methodology discussed in the preceding text. In addition, I have attempted to

ILLUSTRATIONS AND EXERCISES

5

provide illustrations that are of economic interest in themselves, using data that are
typical for current empirical work and covering a wide range of different areas. This
means that most data sets are used in recently published empirical work and are fairly
large, both in terms of number of observations and number of variables. Given the
current state of computing facilities, it is usually not a problem to handle such large
data sets empirically.
Learning econometrics is not just a matter of studying a textbook. Hands-on experience is crucial in the process of understanding the different methods and how and when
to implement them. Therefore, readers are strongly encouraged to get their hands dirty
and to estimate a number of models using appropriate or inappropriate methods, and
to perform a number of alternative speciﬁcation tests. With modern software becoming
more and more user-friendly, the actual computation of even the more complicated
estimators and test statistics is often surprisingly simple, sometimes dangerously simple. That is, even with the wrong data, the wrong model and the wrong methodology,
programs may come up with results that are seemingly all right. At least some expertise is required to prevent the practitioner from such situations and this book plays an
important role in this.
To stimulate the reader to use actual data and estimate some models, almost all data
sets used in this text are available through the web site http://www.wileyeurope.com/
go/verbeek2ed. Readers are encouraged to re-estimate the models reported in this text
and check whether their results are the same, as well as to experiment with alternative
speciﬁcations or methods. Some of the exercises make use of the same or additional
data sets and provide a number of speciﬁc issues to consider. It should be stressed
that for estimation methods that require numerical optimization, alternative programs,
algorithms or settings may give slightly different outcomes. However, you should get
results that are close to the ones reported.
I do not advocate the use of any particular software package. For the linear regression model any package will do, while for the more advanced techniques each package
has its particular advantages and disadvantages. There is typically a trade-off between
user-friendliness and ﬂexibility. Menu driven packages often do not allow you to compute anything else than what’s on the menu, but if the menu is sufﬁciently rich that
may not be a problem. Command driven packages require somewhat more input from
the user, but are typically quite ﬂexible. For the illustrations in the text, I made use of
EViews 3.0, LIMDEP 7.0, MicroFit 4.0, RATS 5.1 and Stata 7.0. Several alternative
econometrics programs are available, including ET, PcGive, TSP and SHAZAM. Journals like the Journal of Applied Econometrics and the Journal of Economic Surveys
regularly publish software reviews.
The exercises included at the end of each chapter consist of a number of questions
that are primarily intended to check whether the reader has grasped the most important
concepts. Therefore, they typically do not go into technical details nor ask for derivations or proofs. In addition, several exercises are of an empirical nature and require
the reader to use actual data.

2

An Introduction to
Linear Regression

One of the cornerstones of econometrics is the so-called linear regression model
and the ordinary least squares (OLS) estimation method. In the ﬁrst part of this
book we shall review the linear regression model with its assumptions, how it can
be estimated, how it can be used for generating predictions and for testing economic
hypotheses.
Unlike many textbooks, I do not start with the statistical regression model with
the standard, Gauss–Markov, assumptions. In my view the role of the assumptions
underlying the linear regression model is best appreciated by ﬁrst treating the most
important technique in econometrics, ordinary least squares, as an algebraic tool rather
than a statistical one. This is the topic of Section 2.1. The linear regression model is
then introduced in Section 2.2, while Section 2.3 discusses the properties of the OLS
estimator in this model under the so-called Gauss–Markov assumptions. Section 2.4
discusses goodness-of-ﬁt measures for the linear model, and hypothesis testing is
treated in Section 2.5. In Section 2.6, we move to cases where the Gauss–Markov
conditions are not necessarily satisﬁed and the small sample properties of the OLS
estimator are unknown. In such cases, the limiting behaviour of the OLS estimator when – hypothetically – the sample size becomes inﬁnitely large, is commonly
used to approximate its small sample properties. An empirical example concerning
the capital asset pricing model (CAPM) is provided in Section 2.7. Sections 2.8 and
2.9 discuss multicollinearity and prediction, respectively. Throughout, an empirical
example concerning individual wages is used to illustrate the main issues. Additional
discussion on how to interpret the coefﬁcients in the linear model, how to test some
of the model’s assumptions and how to compare alternative models, is provided in
Chapter 3.

AN INTRODUCTION TO LINEAR REGRESSION

8

2.1

Ordinary Least Squares as an Algebraic Tool

2.1.1 Ordinary Least Squares

Suppose we have a sample with N observations on individual wages and some background characteristics. Our main interest lies in the question how in this sample wages
are related to the other observables. Let us denote wages by y and the other K − 1 characteristics by x2 , . . . , xK . It will become clear below why this numbering of variables
is convenient. Now we may ask the question: which linear combination of x2 , . . . , xK
and a constant gives a good approximation of y? To answer this question, ﬁrst consider
an arbitrary linear combination, including a constant, which can be written as
β̃1 + β̃2 x2 + · · · + β̃K xK ,

(2.1)

where β̃1 , . . . , β̃K are constants to be chosen. Let us index the observations by i such
that i = 1, . . . , N . Now, the difference between an observed value yi and its linear
approximation is
yi − [β̃1 + β̃2 xi2 + · · · + β̃K xiK ].
(2.2)
To simplify the derivations we shall introduce some short-hand notation. Appendix A
provides additional details for readers unfamiliar with the use of vector notation. First,
we collect the x-values for individual i in a vector xi , which includes the constant.
That is,
xi = (1 xi2 xi3 . . . xiK ) .
Collecting the β̃ coefﬁcients in a K-dimensional vector β̃ = (β̃1 , . . . , β̃K ) , we can
brieﬂy write (2.2) as
(2.3)
yi − xi β̃.
Clearly, we would like to choose values for β̃1 , . . . , β̃K such that these differences are
small. Although different measures can be used to deﬁne what we mean by ‘small’, the
most common approach is to choose β̃ such that the sum of squared differences is as
small as possible. That is, we determine β̃ to minimize the following objective function
S(β̃) ≡

N


(yi − xi β̃)2 .

(2.4)

i=1

This approach is referred to as the ordinary least squares or OLS approach. Taking
squares makes sure that positive and negative deviations do not cancel out when taking
the summation.
To solve the minimization problem, we can look at the ﬁrst order conditions, obtained
by differentiating S(β̃) with respect to the vector β̃. (Appendix A discusses some rules
on how to differentiate a scalar expression, like (2.4), with respect to a vector.) This
gives the following system of K conditions:
−2

N

i=1

xi (yi − xi β̃) = 0

(2.5)

ORDINARY LEAST SQUARES AS AN ALGEBRAIC TOOL

or



N



xi xi

β̃ =

i=1

N


9

xi yi .

(2.6)

i=1

These equations are sometimes referred to as normal equations. As this system has K
unknowns, one can obtain a unique solution for β̃ provided that the symmetric matrix

N

i=1 xi xi , which contains sums of squares and cross products of the regressors xi , can
be inverted. For the moment, we shall assume that this is the case. The solution to the
minimization problem, which we shall denote by b, is then given by

b=

N


−1
xi xi

i=1

N


xi yi .

(2.7)

i=1

By checking the second order conditions, it is easily veriﬁed that b indeed corresponds
to a minimum.
The resulting linear combination of xi is thus given by
ŷi = xi b,
which is the best linear approximation of y from x2 , . . . , xK and a constant. The
phrase ‘best’ refers to the fact that the sum of squared differences (approximation
errors) is minimal for the least squares solution b.
In deriving the linear approximation we have not used any economic or statistical
theory. It is simply an algebraic tool and it holds irrespective of the way the data are
generated. That is, given a set of variables we can always determine the best linear
approximation of one variable using the other variables. The only assumption that
we had to make (which is directly checked from the data) is that the K × K matrix

N

i=1 xi xi is invertible. This says that none of the xik s is an exact linear combination
of the other ones and thus redundant. This is usually referred to as the no-multicollinearity assumption. It should be stressed that the linear approximation is an
in-sample result (that is, in principle it does not give information about observations
(individuals) that are not included in the sample) and, in general, there is no direct
interpretation of the coefﬁcients.
Despite these limitations, the algebraic results on the least squares method are very
useful. Deﬁning a residual ei as the difference between the observed and the approximated value, ei = yi − ŷi = yi − xi b, we can decompose the observed yi as
yi = ŷi + ei = xi b + ei .

(2.8)

This allows us to write the minimum value for the objective function as
S(b) =

N

i=1

ei2 ,

(2.9)

AN INTRODUCTION TO LINEAR REGRESSION

10

which is referred to as the residual sum of squares. It can be shown that the approximated value xi b and the residual ei satisfy certain properties by construction. For
example, if we rewrite (2.5), substituting the OLS solution b, we obtain
N


xi (yi − xi b) =

i=1

N


xi ei = 0.

(2.10)

i=1

This means that the vector e = (e1 , . . . , eN ) is orthogonal1 to each vector ofobservations on an x-variable. For example, if xi contains a constant, it implies that N
i=1 ei =
0. That is, the average residual is zero. This is an intuitively appealing result. If the
average residual were nonzero, this would mean that we could improve upon the
approximation by adding or subtracting the same constant for each observation, i.e. by
changing b1 . Consequently, for the average observation it follows that
ȳ = x̄  b,
(2.11)
N

where ȳ = (1/N ) N
i=1 yi and x̄ = (1/N )
i=1 xi , a K-dimensional vector of sample
means. This shows that for the average observation there is no approximation error.
Similar interpretations hold for the other x-variables: if the derivative
of the sum of
squared approximation errors with respect to β̃k is positive, that is if N
i=1 xik ei > 0
it means that we can improve the objective function by decreasing β̃k .
2.1.2 Simple Linear Regression
In the case where K = 2 we only have one regressor and a constant. In this case,
the observations2 (yi , xi ) can be drawn in a two-dimensional graph with x-values on
the horizontal axis and y-values on the vertical one. This is done in Figure 2.1 for
a hypothetical data set. The best linear approximation of y from x and a constant is
obtained by minimizing the sum of squared residuals, which – in this two-dimensional
case – equal the vertical distances between an observation and the ﬁtted value. All
ﬁtted values are on a straight line, the regression line.
Because a 2 × 2 matrix can be inverted analytically we can derive solutions for b1
and b2 in this special case from the general expression for b above. Equivalently, we
can minimize the residual sum of squares with respect to the unknowns directly. Thus
we have,
N

(yi − β̃1 − β̃2 xi )2 .
(2.12)
S(β̃1 , β̃2 ) =
i=1

The basic elements in the derivation of the OLS solutions are the ﬁrst order conditions:
∂S(β̃1 , β̃2 )
∂ β̃1
∂S(β̃1 , β̃2 )
∂ β̃2

= −2

N


(yi − β̃1 − β̃2 xi ) = 0,

(2.13)

xi (yi − β̃1 − β̃2 xi ) = 0.

(2.14)

i=1

= −2

N

i=1


Two vectors x and y are said to be orthogonal if x  y = 0, that is if i xi yi = 0 (see Appendix A).
2
In this subsection xi will be used to denote the single regressor, so that it does not include the constant.
1

ORDINARY LEAST SQUARES AS AN ALGEBRAIC TOOL

11

y

.2

0

−.2
−.1

0

.1
x

.2

.3

Figure 2.1 Simple linear regression: ﬁtted line and observation points

From (2.13) we can write
N
N
1 
1 
b1 =
y − b2
x = ȳ − b2 x̄,
N i=1 i
N i=1 i

(2.15)

where b2 is solved from combining (2.14) and (2.15). First, from (2.14) write
N


xi yi − b1

i=1

N


xi −

i=1

 N



xi2 b2 = 0

i=1

and then substitute (2.15) to obtain
N



xi yi − N x̄ ȳ −

i=1

N



xi2

− N x̄

b2 = 0

2

i=1

such that we can solve for the slope coefﬁcient b2 as
b2 =

N

i=1 (xi − x̄)(yi −
N
2
i=1 (xi − x̄)

ȳ)

.

(2.16)

Through adding a factor 1/(N − 1) to numerator and denominator it appears that the
OLS solution b2 is the ratio of the sample covariance between x and y and the sample
variance of x. From (2.15), the intercept is determined so as to make the average
approximation error (residual) equal to zero.

AN INTRODUCTION TO LINEAR REGRESSION

12

2.1.3 Example: Individual Wages

An example that will appear frequently in this chapter is based on a sample of individual
wages with background characteristics, like gender, race and years of schooling. We
use a subsample of the US National Longitudinal Survey (NLS) that relates to 1987
and we have a sample of 3294 young working individuals, of which 1569 are female.3
The average hourly wage rate in this sample equals $6.31 for males and $5.15 for
females. Now suppose we try to approximate wages by a linear combination of a
constant and a 0–1 variable denoting whether the individual is male or not. That is,
xi = 1 if individual i is male and zero otherwise. Such a variable, which can only take
on the values of zero and one, is called a dummy variable. Using the OLS approach
the result is
ŷi = 5.15 + 1.17xi .
This means that for females our best approximation is $5.15 and for males it is $5.15 +
$1.17 = $6.31. It is not a coincidence that these numbers are exactly equal to the
sample means in the two subsamples. It is easily veriﬁed from the results above that
b1 = ȳf
b2 = ȳm − ȳf ,


where
ȳm = 
i xi yi /
i xi is the sample average of the wage for males, and ȳf =

(1
−
x
)y
/
(1
−
x
i i
i ) is the average for females.
i
i
2.1.4 Matrix Notation

Because econometricians make frequent use of matrix expressions as shorthand notation, some familiarity with this matrix ‘language’ is a prerequisite to reading the
econometrics literature. In this text, we shall regularly rephrase results using matrix
notation, and occasionally, when the alternative is extremely cumbersome, restrict
attention to matrix expressions only. Using matrices, deriving the least squares solution
is faster, but it requires some knowledge of matrix differential calculus. We introduce
the following notation:


1
.
X =  ..
1

x12
..
.
xN2

 
   
x1
y1
x1K
..  =  ..  , y =  ..  .
 . 
.   . 
xN
yN
. . . xNK
...

So, in the N × K matrix X the i-th row refers to observation i, and the k-th column
refers to the k-th explanatory variable (regressor). The criterion to be minimized, as
given in (2.4), can be rewritten in matrix notation using that the inner product of a
given vector a with itself (a  a) is the sum of its squared elements (see Appendix A).
That is,
S(β̃) = (y − Xβ̃) (y − Xβ̃) = y  y − 2y  Xβ̃ + β̃  X Xβ̃,
(2.17)
3

The data for this example are available as WAGES1.

ORDINARY LEAST SQUARES AS AN ALGEBRAIC TOOL

13

from which the least squares solution follows from differentiating4 with respect to β̃
and setting the result to zero:
∂S(β̃)
∂ β̃

= −2(X y − X Xβ̃) = 0.

(2.18)

Solving (2.18) gives the OLS solution
b = (X X)−1 X y,

(2.19)

which is exactly the same as the one derived in (2.7)
written in matrix notation.
 but now

Note that we again have to assume that X X = N
x
x
is
invertible, i.e. that there
i=1 i i
is no exact (or perfect) multicollinearity.
As before, we can decompose y as
y = Xb + e,

(2.20)

where e is an N-dimensional vector of residuals. The ﬁrst order conditions imply that
X (y − Xb) = 0 or
X e = 0,
(2.21)
which means that each column of the matrix X is orthogonal to the vector of residuals.
With (2.19) we can also write (2.20) as
y = Xb + e = X(X X)−1 X y + e = ŷ + e

(2.22)

so that the predicted value for y is given by
ŷ = Xb = X(X X)−1 X y = PX y.

(2.23)

In linear algebra, the matrix PX ≡ X(X X)−1 X is known as a projection matrix (see
Appendix A). It projects the vector y upon the columns of X (the column space
of X ). This is just the geometric translation of ﬁnding the best linear approximation of y from the columns (regressors) in X. The residual vector of the projection
e = y − Xb = (I − PX )y = MX y is the orthogonal complement. It is a projection of
y upon the space orthogonal to the one spanned by the columns of X. This interpretation is sometimes useful. For example, projecting twice on the same space should
leave the result unaffected, so that it holds that PX PX = PX and MX MX = MX . More
importantly, it holds that MX PX = 0 as the column space of X and its orthogonal
complement do not have anything in common (except the null vector). This is an
alternative way to interpret the result that ŷ and e and also X and e are orthogonal.
The interested reader is referred to Davidson and MacKinnon (1993, Chapter 1) for an
excellent discussion on the geometry of least squares.
4

See Appendix A for some rules for differentiating matrix expressions with respect to vectors.

14

2.2

AN INTRODUCTION TO LINEAR REGRESSION

The Linear Regression Model

Usually, economists want more than just ﬁnding the best linear approximation of
one variable given a set of others. They want economic relationships that are more
generally valid than the sample they happen to have. They want to draw conclusions
about what happens if one of the variables actually changes. That is: they want to
say something about things that are not observed (yet). In this case, we want the
relationship that is found to be more than just a historical coincidence; it should reﬂect
a fundamental relationship. To do this it is assumed that there is a general relationship
that is valid for all possible observations from a well-deﬁned population (for example
all US households, or all ﬁrms in a certain industry). Restricting attention to linear
relationships, we specify a statistical model as

or

yi = β1 + β2 xi2 + · · · + βK xiK + εi

(2.24)

yi = xi β + εi ,

(2.25)

where yi and xi are observable variables and εi is unobserved and referred to as an
error term or disturbance term. The elements in β are unknown population parameters. The equality in (2.25) is supposed to hold for any possible observation, while
we only observe a sample of N observations. We shall consider this sample as one
realization of all potential samples of size N that could have been drawn from the same
population. In this way we can view yi and εi (and often xi ) as random variables.
Each observation corresponds to a realization of these random variables. Again we can
use matrix notation and stack all observations to write
y = Xβ + ε,

(2.26)

where y and ε are N-dimensional vectors and X, as before, is of dimension N × K.
Notice the difference between this equation and (2.20).
In contrast to (2.20), equations (2.25) and (2.26) are population relationships, where
β is a vector of unknown parameters characterizing the population. The sampling process describes how the sample is taken from the population and, as a result, determines
the randomness of the sample. In a ﬁrst view, the xi variables are considered as ﬁxed
and non-stochastic, which means that every new sample will have the same X matrix.
In this case one refers to xi as being deterministic. A new sample only implies new
values for εi , or – equivalently – for yi . The only relevant case where the xi s are truly
deterministic is in a laboratory setting, where a researcher can set the conditions of a
given experiment (e.g. temperature, air pressure). In economics we will typically have
to work with non-experimental data. Despite this, it is convenient and in particular
cases appropriate in an economic context to treat the xi variables as deterministic. In
this case, we will have to make some assumptions about the sampling distribution
of εi . A convenient one corresponds to random sampling where each error εi is a
random drawing from the population distribution, independent of the other error terms.
We shall return to this issue below.

THE LINEAR REGRESSION MODEL

15

In a second view, a new sample implies new values for both xi and εi , so that each
time a new set of N observations for (yi , xi ) is drawn. In this case random sampling
means that each set (xi , εi ) or (yi , xi ) is a random drawing from the population distribution. In this context, it will turn out to be important to make assumptions about
the joint distribution of xi and εi , in particular regarding the extent to which the distribution of εi is allowed to depend upon X. The idea of a (random) sample is most
easily understood in a cross-sectional context, where interest lies in a large and ﬁxed
population, for example all UK households in January 1999, or all stocks listed at the
New York Stock Exchange on a given date. In a time series context, different observations refer to different time periods and it does not make sense to assume that we
have a random sample of time periods. Instead, we shall take the view that the sample
we have is just one realization of what could have happened in a given time span and
the randomness refers to alternative states of the world. In such a case we will need
to make some assumptions about the way the data are generated (rather than the way
the data are sampled).
It is important to realize that without additional restrictions the statistical model in
(2.25) is a tautology: for any value for β one can always deﬁne a set of εi s such that
(2.25) holds exactly for each observation. We thus need to impose some assumptions
to give the model a meaning. A common assumption is that the expected value of εi
given all the explanatory variables in xi is zero, that is E{εi |xi } = 0. Usually, people
refer to this as the assumption saying that the x-variables are exogenous. Under this
assumption it holds that
(2.27)
E{yi |xi } = xi β,
so that the regression line xi β describes the conditional expectation of yi given the
values for xi . The coefﬁcients βk measure how the expected value of yi is changed if the
value of xik is changed, keeping the other elements in xi constant (the ceteris paribus
condition). Economic theory, however, often suggests that the model in (2.25) describes
a causal relationship, in which the β coefﬁcients measure the changes in yi caused
by a ceteris paribus change in xik . In such cases, εi has an economic interpretation
(not just a statistical one) and imposing that it is uncorrelated with xi , as we do by
imposing E{εi |xi } = 0, may not be justiﬁed. As in many cases it can be argued that
unobservables in the error term are related to observables in xi , we should be cautious
interpreting our regression coefﬁcients as measuring causal effects. We shall come back
to these issues in Chapter 5.
Now that our β coefﬁcients have a meaning, we can try to use the sample (yi , xi ),
i = 1, . . . , N to say something about it. The rule which says how a given sample is
translated into an approximate value for β is referred to as an estimator. The result for
a given sample is called an estimate. The estimator is a vector of random variables,
because the sample may change. The estimate is a vector of numbers. The most widely
used estimator in econometrics is the ordinary least squares (OLS) estimator. This
is just the ordinary least squares rule described in Section 2.1 applied to the available
sample. The OLS estimator for β is thus given by

b=

N

i=1

−1
xi xi

N

i=1

xi yi .

(2.28)

AN INTRODUCTION TO LINEAR REGRESSION

16

Because we have assumed an underlying ‘true’ model (2.25), combined with a sampling
scheme, b is now a vector of random variables. Our interest lies in the true unknown
parameter vector β, and b is considered an approximation to it. While a given sample
only produces a single estimate, we evaluate the quality of it through the properties
of the underlying estimator. The estimator b has a sampling distribution because its
value depends upon the sample that is taken (randomly) from the population.

2.3

Small Sample Properties of the OLS Estimator

2.3.1 The Gauss–Markov Assumptions

In this section we shall discuss several important properties of the OLS estimator b.
To do so, we need to make some assumptions about the error term and the explanatory
variables xi . The ﬁrst set of assumptions we consider are the so-called Gauss–Markov
assumptions. These assumptions are usually standard in the ﬁrst chapters of econometrics textbooks, although – as we shall see below – they are not all strictly needed to
justify the use of the ordinary least squares estimator. They just constitute a simple
case in which the small sample properties of b are easily derived.
The standard set of Gauss–Markov assumptions is given by
E{εi } = 0,

i = 1, . . . , N

{ε1 , . . . , εN } and {x1 , . . . , xN } are independent
V {εi } = σ ,
2

cov{εi , εj } = 0,

i = 1, . . . , N

i, j = 1, . . . , N, i = j.

(A1)
(A2)
(A3)
(A4)

Assumption (A1) says that the expected value of the error term is zero, which means
that, on average, the regression line should be correct. Assumption (A3) states that all
error terms have the same variance, which is referred to as homoskedasticity, while
assumption (A4) imposes zero correlation between different error terms. This excludes
any form of autocorrelation. Taken together, (A1), (A3) and (A4) imply that the error
terms are uncorrelated drawings from a distribution with expectation zero and constant
variance σ 2 . Using the matrix notation from above, it is possible to rewrite these three
conditions as
(2.29)
E{ε} = 0 and V {ε} = σ 2 IN ,
where IN is the N × N identity matrix. This says that the covariance matrix of the
vector of error terms ε is a diagonal matrix with σ 2 on the diagonal. Assumption (A2)
implies that X and ε are independent. This is a fairly strong assumption, which can be
relaxed somewhat (see below). It implies that
E{ε|X} = E{ε} = 0

(2.30)

V {ε|X} = V {ε} = σ 2 IN .

(2.31)

and

SMALL SAMPLE PROPERTIES OF THE OLS ESTIMATOR

17

That is, the matrix of regressor values X does not provide any information about the
expected values of the error terms or their (co)variances. The two conditions (2.30) and
(2.31) combine the necessary elements from the Gauss–Markov assumptions needed
for the results below to hold. Often, assumption (A2) is stated as: the regressor matrix
X is a deterministic nonstochastic matrix. The reason for this is that the outcomes in
the matrix X can be taken as given without affecting the properties of ε, that is, one
can derive all properties conditional upon X. For simplicity, we shall take this approach
in this section and Section 2.5. Under the Gauss–Markov assumptions (A1) and (A2),
the linear model can be interpreted as the conditional expectation of yi given xi , i.e.
E{yi |xi } = xi β. This is a direct implication of (2.30).
2.3.2 Properties of the OLS Estimator

Under assumptions (A1)–(A4), the OLS estimator b for β has several desirable properties. First of all, it is unbiased. This means that, in repeated sampling, we can
expect that our estimator is on average equal to the true value β. We formulate this as
E{b} = β. It is instructive to see the proof:
E{b} = E{(X X)−1 X y} = E{β + (X X)−1 X ε}
= β + E{(X X)−1 X ε} = β.
The latter step here is essential and it follows from
E{(X X)−1 X ε} = E{(X X)−1 X }E{ε} = 0,
because X and ε are independent and E{ε} = 0. Note that we did not use assumptions
(A3) and (A4) in the proof. This shows that the OLS estimator is unbiased as long
as the error terms are mean zero and independent of all explanatory variables, even if
heteroskedasticity or autocorrelation are present. We shall come back to this issue in
Chapter 4.
In addition to knowing that we are, on average, correct, we would also like to make
statements about how (un)likely it is to be far off in a given sample. This means we
would like to know the distribution of b. First of all, the variance of b (conditional
upon X ) is given by



−1

V {b|X} = σ (X X)
2

=σ

2

N


−1
xi xi

,

(2.32)

i=1

which, for simplicity, we shall denote by V {b}. Implicitly, this means that we treat X
as deterministic. The proof is fairly easy and goes as follows:
V {b} = E{(b − β)(b − β) } = E{(X X)−1 X εε X(X X)−1 }
= (X X)−1 X (σ 2 IN )X(X X)−1 = σ 2 (X X)−1 .

AN INTRODUCTION TO LINEAR REGRESSION

18

Without using matrix notation the proof goes as follows:

 
−1
−1 

−1

 




V {b} = V
xi xi
xi εi =
xi xi
V
xi εi
xi xi



=

i


i

xi xi

−1


σ2

i




xi xi

i

i



−1

xi xi


= σ2

i

i



−1
xi xi

i

.

(2.33)

i

The last result is collected in the Gauss–Markov Theorem, which says that under
assumptions (A1)–(A4) the OLS estimator b is the best linear unbiased estimator for
β. In short we say that b is BLUE for β. To appreciate this result, consider the class of
linear unbiased estimators. A linear estimator is a linear function of the elements in y
and can be written as b̃ = Ay, where A is a K × N matrix. The estimator is unbiased
if E{Ay} = β. (Note that the OLS estimator is obtained for A = (X X)−1 X .) Then
the theorem states that the difference between the covariance matrices of b̃ = Ay and
the OLS estimator b is always positive semi-deﬁnite. What does this mean? Suppose
we are interested in some linear combination of β coefﬁcients, given by d  β where d
is a K-dimensional vector. Then the Gauss–Markov result implies that the variance of
the OLS estimator d  b for d  β is at least as large as the variance of any other linear
unbiased estimator d  b̃, that is
V {d  b̃} ≥ V {d  b} for any vector d.
As a special case this holds for the k-th element and we have that
V {b̃k } ≥ V {bk }.
Thus, under the Gauss–Markov assumptions, the OLS estimator is the most accurate
(linear) unbiased estimator for β. More details on the Gauss–Markov result can be
found in Stewart and Gill (1998, Section 2.4), Hayashi (2000, Section 1.3) or Greene
(2003, Section 4.4).
To estimate the variance of b we need to replace the unknown error variance σ 2 by an
estimate. An obvious candidate is the sample variance of the residuals ei = yi − xi b,
that is
N
1  2
s̃ 2 =
e
(2.34)
N − 1 i=1 i
(recalling that the average residual is zero). However, because ei is different from εi ,
it can be shown that this estimator is biased for σ 2 . An unbiased estimator is given by
s2 =

N

1
e2 .
N − K i=1 i

(2.35)

This estimator has a degrees of freedom correction as it divides by the number of
observations minus the number of regressors (including the intercept). An intuitive

SMALL SAMPLE PROPERTIES OF THE OLS ESTIMATOR

19

argument for this is that K parameters were chosen so as to minimize the residual sum
of squares and thus to minimize the sample variance of the residuals. Consequently,
s̃ 2 is expected to underestimate the variance of the error term σ 2 . The estimator s 2 ,
with a degrees of freedom correction, is unbiased under assumptions (A1)–(A4); see
Hayashi (2000, Section 1.3) or Greene (2003, Section 4.6) for a proof. The variance
of b can thus be estimated by


−1

V̂ {b} = s (X X)
2

=s

2

 N


−1
xi xi

.

(2.36)

i=1

The estimated variance of an element bk is given by s 2 ckk where ckk is the (k, k)
element in (i xi xi )−1 . The square root of this estimated variance is usually referred to
as the standard error of bk . We shall denote it as se(bk ). It is the estimated standard
deviation of bk and is a measure for the accuracy of the estimator. Under assumptions
√
(A1)–(A4), it holds that se(bk ) = s ckk . When the error terms are not homoskedastic
or exhibit autocorrelation, the standard error of the OLS estimator bk will have to be
computed in a different way (see Chapter 4).
So far, we made no assumption about the shape of the distribution of the error terms
εi , except that they were mutually uncorrelated, independent of X, had zero mean and a
constant variance. For exact statistical inference from a given sample of N observations,
explicit distributional assumptions have to be made.5 The most common assumption
is that the errors are jointly normally distributed.6 In this case the uncorrelatedness
of (A4) is equivalent to independence of all error terms. The precise assumption is
as follows
ε ∼ N(0, σ 2 IN ),
(A5)
saying that ε has an N-variate normal distribution with mean vector 0 and covariance
matrix σ 2 IN . Assumption (A5) thus replaces (A1), (A3) and (A4). An alternative way
of formulating (A5) is
εi ∼ NID(0, σ 2 ),
(A5 )
which is a shorthand way of saying that the error terms εi are independent drawings
from a normal distribution (n.i.d.) with mean zero and variance σ 2 . Even though error
terms are unobserved, this does not mean that we are free to make any assumption
we like. For example, if error terms are assumed to follow a normal distribution this
means that yi (for a given value of xi ) also follows a normal distribution. Clearly,
we can think of many variables whose distribution (conditional upon a given set of
xi variables) is not normal, in which case the assumption of normal error terms is
inappropriate. Fortunately, not all assumptions are equally crucial for the validity of
the results that follow and, moreover, the majority of the assumptions can be tested
empirically; see Chapters 3, 4 and 6 below.
To make things simpler let us consider the X matrix as ﬁxed and deterministic or,
alternatively, let us work conditional upon the outcomes X. Then the following result
5
6

Later we shall see that for approximate inference in large samples this is not necessary.
The distributions used in this text are explained in Appendix B.

AN INTRODUCTION TO LINEAR REGRESSION

20

Table 2.1

OLS results wage equation

Dependent variable: wage
Variable

Estimate

Standard error

constant
male

5.1469
1.1661

0.0812
0.1122

s = 3.2174 R 2 = 0.0317 F = 107.93

holds. Under assumptions (A2) and (A5) the OLS estimator b is normally distributed
with mean vector β and covariance matrix σ 2 (X X)−1 , i.e.
b ∼ N(β, σ 2 (X X)−1 ).

(2.37)

The proof of this follows directly from the result that b is a linear combination of all
εi and is omitted here. From this it also follows that each element in b is normally
distributed, for example
(2.38)
bk ∼ N(βk , σ 2 ckk ),
where ckk is the (k, k ) element in (X X)−1 . These results provide the basis for statistical
tests based upon the OLS estimator b.
2.3.3 Example: Individual Wages (Continued)
Let us now turn back to our wage example. We can formulate a statistical model as

wage i = β1 + β2 male i + εi ,

(2.39)

where wage i denotes the hourly wage rate of individual i and male i = 1 if i is male and
0 otherwise. Imposing that E{εi } = 0 and E{εi |male i } = 0 gives β1 the interpretation
of the expected wage rate for females, while E{wage i |male i = 1} = β1 + β2 is the
expected wage rate for males. These are unknown population quantities and we may
wish to estimate them. Assume that we have a random sample, implying that different
observations are independent. Also assume that εi is independent of the regressors,
in particular, that the variance of εi does not depend upon gender (male i ). Then the
OLS estimator for β is unbiased and its covariance matrix is given by (2.32). The
estimation results are given in Table 2.1. In addition to the OLS estimates, identical
to those presented before, we now also know something about the accuracy of the
estimates, as reﬂected in the reported standard errors. We can now say that our estimate
of the expected hourly wage differential β2 between males and females is $1.17 with a
standard error of $0.11. Combined with the normal distribution, this allows us to make
statements about β2 . For example, we can test the hypothesis that β2 = 0. If this is the
case, the wage differential between males and females in our sample is nonzero only
by chance. Section 2.5 discusses how to test hypotheses regarding β.

2.4

Goodness-of-ﬁt

Having estimated a particular linear model, a natural question that comes up is: how
well does the estimated regression line ﬁt the observations? A popular measure for the
goodness-of-ﬁt is the proportion of the (sample) variance of y that is explained by the

GOODNESS-OF-FIT

21

model. This variable is called the R 2 (R squared) and is deﬁned as
R =

V̂ {ŷi }

1/(N − 1)

N

i=1 (ŷi

− ȳ)2

,
(2.40)
1/(N − 1) i=1 (yi − ȳ)2
V̂ {yi }

where ŷi = xi b and ȳ = (1/N ) i yi denotes the sample mean of yi . Note that ȳ also
corresponds with the sample mean of ŷi , because of (2.11).
From the ﬁrst order conditions (compare (2.10)) it follows directly that
2

N


=

ei xik = 0,

N

k = 1, . . . , K.

i=1


Consequently, we can write yi = ŷi + ei where i ei ŷi = 0. In the most relevant case
where the model contains an intercept term, it holds that
V̂ {yi } = V̂ {ŷi } + V̂ {ei },
where V̂ {ei } = s̃ 2 . Using this, the R 2 can be rewritten as

2
1/(N − 1) N
V̂ {ei }
i=1 ei
=1−
R2 = 1 −
.
N
1/(N − 1) i=1 (yi − ȳ)2
V̂ {yi }

(2.41)

(2.42)

Equation (2.41) shows how the sample variance of yi can be decomposed into the
sum of the sample variances of two orthogonal components: the predictor ŷi and the
residual ei . The R 2 thus indicates which proportion of the sample variation in yi is
explained by the model.
If the model of interest contains an intercept term, the two expressions for R 2 in
(2.40) and (2.42) are equivalent. Moreover, in this case it can be shown that 0 ≤ R 2 ≤
1. Only if all ei = 0 it holds that R 2 = 1, while the R 2 is zero if the model does not
explain anything in addition to the sample mean of yi . That is, the R 2 of a model with
just an intercept term is zero by construction. In this sense, the R 2 indicates how much
better the model performs than a trivial model with only a constant term.
From the results in Table 2.1, we see that the R 2 of the very simple wage equation
is only 0.0317. This means that only approximately 3.2% of the variation in individual
wages can be attributed to gender differences. Apparently, many other observable and
unobservable factors affect a person’s wage besides gender. This does not automatically
imply that the model that was estimated in Table 2.1 is incorrect or useless: it just
indicates the relative (un)importance of gender in explaining individual wage variation.
In the exceptional cases that the model does not contain an intercept term, the two
expressions
for R 2 are not equivalent. The reason is that (2.41) is violated because
N
2
i=1 ei is no longer equal to zero. In this situation it is possible that the R computed
from (2.42) becomes negative. An alternative measure, which is routinely computed by
some software packages if there is no intercept, is the uncentred R 2 , which is deﬁned as
N 2
N 2
e
i=1 ŷi
2
(2.43)
uncentred R = N 2 = 1 − Ni=1 i2 .
i=1 yi
i=1 yi
Generally, the uncentred R 2 is higher than the standard R 2 .

AN INTRODUCTION TO LINEAR REGRESSION

22

Because the R 2 measures the explained variation in yi it is also sensitive to the
deﬁnition of this variable. For example, explaining wages is something different than
explaining log wages, and the R 2 s will be different. Similarly, models explaining
consumption, changes in consumption or consumption growth will not be directly
comparable in terms of their R 2 s. It is clear that some sources of variation are
much harder to explain than others. For example, variation in aggregate consumption for a given country is usually easier to explain than the cross-sectional variation in consumption over individual households. Consequently, there is no absolute
benchmark to say that an R 2 is ‘high’ or ‘low’. A value of 0.2 may be high in
certain applications but low in others, and even a value of 0.95 may be low in certain contexts.
Sometimes the R 2 is interpreted as a measure of quality of the statistical model,
while in fact it measures nothing more than the quality of the linear approximation. As
the OLS approach is developed to give the best linear approximation, irrespective of
the ‘true’ model and the validity of its assumptions, estimating a linear model by OLS
will always give the best R 2 possible. Any other estimation method, and we will see
several below, will lead to lower R 2 values even though the corresponding estimator
may have much better statistical properties under the assumptions of the model. Even
worse, when the model is not estimated by OLS the two deﬁnitions (2.40) and (2.42)
are not equivalent and it is not obvious how an R 2 should be deﬁned. For later use, we
shall present an alternative deﬁnition of the R 2 , which for OLS is equivalent to (2.40)
and (2.42), and for any other estimator is guaranteed to be between zero and one. It is
given by

2
N
i=1 (yi − ȳ)(ŷi − ȳ)
 
,
R 2 = corr2 {yi , ŷi } = 
(2.44)
N
N
2
2
(y
−
ȳ)
(
ŷ
−
ȳ)
i
i
i=1
i=1
which denotes the squared (sample) correlation coefﬁcient between the actual and
ﬁtted values. Using (2.41) it is easily veriﬁed that for the OLS estimator (2.44) is
equivalent to (2.40). Written in this way, the R 2 can be interpreted to measure how
well the variation in ŷi relates to variation in yi . Despite this alternative deﬁnition,
the R 2 reﬂects the quality of the linear approximation and not necessarily that of the
statistical model we are interested in. As a result, the R 2 is typically not the most
important aspect of our estimation results.
Another drawback of the R 2 is that it will never decrease if the number of regressors
is increased, even if the additional variables have no real explanatory power. A common
way to solve this is to correct the variance estimates in (2.42) for the degrees of
freedom. This gives the so-called adjusted R 2 , or R̄ 2 , deﬁned as

2
1/(N − K) N
i=1 ei
R̄ = 1 −
.
N
1/(N − 1) i=1 (yi − ȳ)2
2

(2.45)

This goodness-of-ﬁt measure has some punishment for the inclusion of additional
explanatory variables in the model and, therefore, does not automatically increase
when regressors are added to the model (see Chapter 3). In fact, it may decline when
a variable is added to the set of regressors. Note that, in extreme cases, the R̄ 2 may

HYPOTHESIS TESTING

23

become negative. Also note that the adjusted R 2 is strictly smaller than R 2 unless the
model only includes a constant term and both measures are zero.

2.5

Hypothesis Testing

Under the Gauss–Markov assumptions (A1)–(A4) and normality of the error terms
(A5), we saw that the OLS estimator b has a normal distribution with mean β and
covariance matrix σ 2 (X X)−1 . We can use this result to develop tests for hypotheses
regarding the unknown population parameters β. Starting from (2.38), it follows that
the variable
b −β
(2.46)
z = k√ k
σ ckk
has a standard normal distribution (i.e. a normal distribution with mean 0 and variance 1).
If we replace the unknown σ by its estimate s, this is no longer exactly true. It can be
shown7 that the unbiased estimator s 2 deﬁned in (2.35) is independent of b and has a
Chi-squared distribution with N − K degrees of freedom. In particular,8
2
(N − K)s 2 /σ 2 ∼ χN−K
.

(2.47)

Consequently, the following random variable
tk =

bk − βk
√
s ckk

(2.48)

is the ratio of a standard normal variable and the square root of an independent Chisquared variable and therefore follows Student’s t distribution with N − K degrees of
freedom. The t distribution is close to the standard normal distribution except that it
has fatter tails, particularly when the number of degrees of freedom N − K is small.
The larger N − K the closer the t distribution resembles the standard normal, and for
sufﬁciently large N − K, the two distributions are identical.
2.5.1 A Simple t-test

The result above can be used to construct test statistics and conﬁdence intervals. The
general idea of hypothesis testing is as follows. Starting from a given hypothesis, the
null hypothesis, a test statistic is computed that has a known distribution under the
assumption that the null hypothesis is valid. Next, it is decided whether the computed
value of the test statistic is unlikely to come from this distribution, which indicates that
the null hypothesis is unlikely to hold. Let us illustrate this with an example. Suppose
we have a null hypothesis that speciﬁes the value of βk , say H0 : βk = βk0 , where βk0
7

The proof of this is beyond the scope of this text. The basic idea is that a sum of squared normals is
Chi-squared distributed (see Appendix B).
8
See Appendix B for details about the distributions in this section.

AN INTRODUCTION TO LINEAR REGRESSION

24

is a speciﬁc value chosen by the researcher. If this hypothesis is true we know that
the statistic
b − βk0
tk = k
(2.49)
se(bk )
has a t distribution with N − K degrees of freedom. If the null hypothesis is not true,
the alternative hypothesis H1 : βk = βk0 holds. As there are no unknown values in tk ,
it becomes a test statistic that can be computed from the estimate bk and its standard
error se(bk ). The usual testing strategy is to reject the null hypothesis if tk realizes
a value that is very unlikely if the null hypothesis is true. In this case this means
very large absolute values for tk . To be precise, one rejects the null hypothesis if the
probability of observing a value of |tk | or larger is smaller than a given signiﬁcance
level α, often 5%. From this, one can deﬁne the critical values tN−K;α/2 using
P {|tk | > tN−K;α/2 } = α.
For N − K not too small, these critical values are only slightly larger than those of
the standard normal distribution, for which the two-tailed critical value for α = 0.05
is 1.96. Consequently, at the 5% level the null hypothesis will be rejected if
|tk | > 1.96.
The above test is referred to as a two-sided test because the alternative hypothesis
allows for values of βk on both sides of βk0 . Occasionally, the alternative hypothesis is
one-sided, for example: the expected wage for a man is larger than that for a woman.
Formally, we deﬁne the null hypothesis as H0 : βk ≤ βk0 with alternative H1 : βk > βk0 .
Next we consider the distribution of the test statistic tk at the boundary of the null
hypothesis (i.e. under βk = βk0 , as before) and we reject the null hypothesis if tk is too
large (note that large values for bk lead to large values for tk ). Large negative values
for tk are compatible with the null hypothesis and do not lead to its rejection. Thus for
this one-sided test, the critical value is determined from
P {tk > tN−K;α } = α.
Using the standard normal approximation again, we reject the null hypothesis at the
5% level if
tk > 1.64.
Regression packages typically report the following t-value,
tk =

bk
,
se(bk )

sometimes referred to as the t-ratio, which is the point estimate divided by its standard
error. The t-ratio is the t-statistic one would compute to test the null hypothesis that
βk = 0, which may be a hypothesis that is of economic interest as well. If it is rejected,
it is said that ‘bk differs signiﬁcantly from zero’, or that the corresponding variable

HYPOTHESIS TESTING

25

‘xik has a signiﬁcant impact on yi ’. Often we simply say that (the effect of) ‘xik is
signiﬁcant’.
A conﬁdence interval can be deﬁned as the interval of all values for βk0 for which
the null hypothesis that βk = βk0 is not rejected by the t-tests. Loosely speaking, a
conﬁdence interval gives a range of values for the true βk that are not unlikely given
the data, i.e. given the estimate bk and the associated standard error. This implies the
following inequalities that hold with probability 1 − α
−tN−K;α/2 <
or

bk − βk
< tN−K;α/2 ,
se(bk )

bk − tN−K;α/2 se(bk ) < βk < bk + tN−K;α/2 se(bk ).

(2.50)
(2.51)

Consequently, using the standard normal approximation, a 95% conﬁdence interval for
βk is given by the interval
[bk − 1.96se(bk ), bk + 1.96se(bk )].

(2.52)

In repeated sampling, 95% of these intervals will contain the true value βk which is a
ﬁxed but unknown number (and thus not stochastic).
2.5.2 Example: Individual Wages (Continued)

From the results in Table 2.1 we can compute t-ratios and perform simple tests. For
example, if we want to test whether β2 = 0, we construct the t-statistic as the estimate
divided by its standard error to get t = 10.38. Given the large number of observations,
the appropriate t distribution is virtually identical to the standard normal one, so the
5% two-tailed critical value is 1.96. This means that we clearly have to reject the
null hypothesis that β2 = 0. We thus have to reject that in the population the expected
wage differential between males and females is zero. We can also compute a conﬁdence
interval, which has bounds 1.17 ± 1.96 × 0.11. This means that with 95% conﬁdence
we can say that over the entire population the expected wage differential between
males and females is between $0.95 and $1.39 per hour.
2.5.3 Testing One Linear Restriction

The test discussed above involves a restriction on a single coefﬁcient. Often, a hypothesis of economic interest implies a linear restriction on more than one coefﬁcient, such
as9 β2 + β3 + · · · + βK = 1. In general, we can formulate such a linear hypothesis as
H0 : r1 β1 + · · · + rK βK = r  β = q,

(2.53)

for some scalar value q and a K-dimensional vector r. We can test the hypothesis in
(2.53) using the result that r  b is the BLUE for r  β with variance V {r  b} = r  V {b}r.
9

For example, in a Cobb–Douglas production function, written as a linear regression model in logs, constant
returns to scale corresponds to the sum of all slope parameters (the coefﬁcients for all log inputs) being
equal to one.

AN INTRODUCTION TO LINEAR REGRESSION

26

Replacing σ 2 in the covariance matrix V {b} by its estimate s 2 produces the estimated
covariance matrix, denotedV̂ {b}. Consequently, the standard error of the linear com-

bination r  b is se(r  b) = s r  V̂ {b}r. As b is K-variate normal, r  b is normal as well
(see Appendix B), so that we have
r b − r β
∼ tN−K ,
se(r  b)

(2.54)

which is a straightforward generalization of (2.48).10 The test statistic for H0 follows as
t=

r b − q
,
se(r  b)

(2.55)

which has a tN−K distribution under the null hypothesis. At the 5% level, absolute
values of t in excess of 1.96 (the normal approximation) lead to rejection of the null.
This represents the most general version of the t-test.
Computation of the standard error se(r  b) requires the estimated covariance matrix
of the vector b. Unfortunately, some regression packages do not provide easy ways
to obtain this standard error or the t-test statistic directly. In such cases, a convenient
way to obtain the same test statistic is by a reparametrization of the original model,
such that the linear restriction in H0 corresponds to a restriction of the usual form, say
βk∗ = 0. For example, consider
yi = β1 + β2 xi2 + β3 xi3 + εi
and suppose the restriction of interest is β2 = β3 . Then, we can rewrite the model as11
yi = β1 + (β2 − β3 )xi2 + β3 (xi3 + xi2 ) + εi
or

yi = β1 + β2∗ xi2 + β3 (xi3 + xi2 ) + εi .

From the deﬁnition of OLS as minimizing the residual sum of squares, it follows that
it is invariant to linear reparametrizations. Consequently, the OLS estimator for β3 in
both formulations of the model will be identical, and the estimator for β2∗ is given
by b2 − b3 . The advantage of the reparametrization is that the null hypothesis can be
written as a zero restriction on one of the regression coefﬁcients, i.e. H0 : β2∗ = 0.
Consequently, it can be tested using the standard t-ratio for β2∗ in the reparametrized
model, given by
b2 − b3
b2∗
t=
=
,
se(b2∗ )
se(b2 − b3 )
where b2∗ is the OLS estimator for β2∗ .
10
11

The statistic is the same if r is a K-dimensional vector of zeroes with a 1 on the k-th position.
This reparametrization is not unique.

HYPOTHESIS TESTING

27

2.5.4 A Joint Test of Signiﬁcance of Regression Coefﬁcients

A standard test that is often automatically supplied by a regression package as well is
a test for the joint hypothesis that all coefﬁcients, except the intercept β1 , are equal to
zero. We shall discuss this procedure slightly more generally by testing the null that
J of the K coefﬁcients are equal to zero (J < K). Without loss of generality, assume
that these are the last J coefﬁcients in the model,
H0 : βK−J +1 = · · · = βK = 0.

(2.56)

The alternative hypothesis in this case is that H0 is not true, i.e. that at least one of
these J coefﬁcients is not equal to zero.
The easiest test procedure in this case is to compare the sum of squared residuals
of the full model with the sum of squared residuals of the restricted model (which is
the model with the last J regressors omitted). Denote the residual sum of squares of
the full model by S1 and that of the restricted model by S0 . If the null hypothesis is
correct one would expect that the sum of squares with the restriction imposed is only
slightly larger than that in the unrestricted case. A test statistic can be obtained by
using the following result, which we present without proof. Under the null hypothesis
and assumptions (A1)–(A5) it holds that
S0 − S1
∼ χJ2 .
σ2

(2.57)

2
From earlier results we know that (N − K)s 2 /σ 2 = S1 /σ 2 ∼ χN−K
. Moreover, under
2
the null hypothesis it can be shown that S0 − S1 and s are independent. Consequently,
we can deﬁne the following test statistic,

f =

(S0 − S1 )/J
.
S1 /(N − K)

(2.58)

Under the null hypothesis f has an F distribution with J and N − K degrees of
J
. If we use the deﬁnition of the R 2 from (2.42), we can also
freedom, denoted FN−K
write this f statistic as
(R12 − R02 )/J
f =
,
(2.59)
(1 − R12 )/(N − K)
were R12 and R02 are the usual goodness-of-ﬁt measures for the unrestricted and the
restricted model, respectively.
It is clear that in this case only very large values for the test statistic imply rejection
of the null hypothesis. Despite the two-sided alternative hypothesis, the critical values
J
FN−K;α
for this so-called F-test are one-sided, and deﬁned by the following equality
J
P {f > FN−K;α
} = α,

where α is the signiﬁcance level of the test. For example, if N − K = 60 and J = 3
the critical value at the 5% level is 2.76.

AN INTRODUCTION TO LINEAR REGRESSION

28

In most applications the estimators for different elements in the parameter vector will
be correlated, which means that the explanatory powers of the explanatory variables
overlap. Consequently, the marginal contribution of each explanatory variable, when
added last, may be quite small. Hence, it is perfectly possible for the t-tests on each
variable’s coefﬁcient not to be signiﬁcant, while the F-test for a number of these
coefﬁcients is highly signiﬁcant. That is, it is possible that the null hypothesis β1 = 0
is as such not unlikely, that the null β2 = 0 is not unlikely, but that the joint null
β1 = β2 = 0 is quite unlikely to be true. As a consequence, in general, t-tests on each
restriction separately may not reject, while a joint F-test does. The converse is also
true: it is possible that individual t-tests do reject the null, while the joint test does not.
The section on multicollinearity below illustrates this point.
A special case of this F-test is sometimes misleadingly referred to as the model test,12
where one tests the signiﬁcance of all regressors, i.e. one tests H0 : β2 = β2 = · · · =
βK = 0, meaning that all partial slope coefﬁcients are equal to zero. The appropriate
test statistic in this case is
f =

(S0 − S1 )/(K − 1)
,
S1 /(N − K)

(2.60)


where S1 is the residual sum of squares of the model, that is S1 = i ei2 , and S0 is the
residual 
sum of squares of the restricted model containing only an intercept term, that
is S0 = i (yi − ȳ)2 .13 Because the restricted model has an R 2 of zero by construction
the test statistic can also be written as
F =

R 2 /(K − 1)
,
(1 − R 2 )/(N − K)

(2.61)

where we used the convention to denote this statistic as F. Note that it is a simple
function of the R 2 of the model. If the test based on F does not reject the null
hypothesis, one can conclude that the model performs rather poorly: a ‘model’ with just
an intercept term would not do statistically worse. However, the converse is certainly
not true: if the test does reject the null, one cannot conclude that the model is good,
perfect, valid or the best. An alternative model may perform much better. Chapter 3
pays more attention to this issue.
2.5.5 Example: Individual Wages (Continued)

The fact that we concluded above that there was a signiﬁcant difference between
expected wage rates for males and females does not necessarily point to discrimination.
It is possible that working males and females differ in terms of their characteristics,
for example, their years of schooling. To analyse this, we can extend the regression
model with additional explanatory variables, for example school i , which denotes the
12

This terminology is misleading as it does not in any way test whether the restrictions imposed by the
model are correct. The only thing tested is whether all coefﬁcients, excluding the intercept, are equal to
zero, in which case one would have a trivial model with an R 2 of zero. As shown in (2.61), the test
statistic associated with the model test is simply a function of the R 2 .
13
Using the deﬁnition of the OLS estimator, it is easily veriﬁed that the intercept term in a model without
regressors is estimated as the sample average ȳ. Any other choice would result in a larger S value.

HYPOTHESIS TESTING

29

years of schooling, and exper i , which denotes experience in years. The model is now
interpreted to describe the conditional expected wage of an individual given his gender,
years of schooling and experience, and can be written as
wage i = β1 + β2 male i + β3 school i + β4 exper i + εi .
The coefﬁcient β2 for male i now measures the difference in expected wage between a
male and a female with the same schooling and experience. Similarly, the coefﬁcient β3
for school i gives the expected wage difference between two individuals with the same
experience and gender where one has one additional year of schooling. In general,
the coefﬁcients in a multiple regression model can only be interpreted under a ceteris
paribus condition, which says that the other variables that are included in the model
are constant.
Estimation by OLS produces the results given in Table 2.2. The coefﬁcient for male i
now suggests that if we compare an arbitrary male and female with the same years
of schooling and experience, the expected wage differential is $1.34 compared to
$1.17 before. With a standard error of $0.11, this difference is still statistically highly
signiﬁcant. The null hypothesis that schooling has no effect on a person’s wage, given
gender and experience, can be tested using the t-test described above, with a test
statistic of 19.48. Clearly the null hypothesis has to be rejected. The estimated wage
increase from one additional year of schooling, keeping years of experience ﬁxed, is
$0.64. It should not be surprising, given these results, that the joint hypothesis that
all three partial slope coefﬁcients are zero, that is wages are not affected by gender,
schooling or experience, has to be rejected as well. The F-statistic takes the value of
167.6, the appropriate 5% critical value being 2.60.
Finally, we can use the above results to compare this model with the simpler one in
Table 2.1. The R 2 has increased from 0.0317 to 0.1326, which means that the current
model is able to explain 13.3% of the within sample variation in wages. We can
perform a joint test on the hypothesis that the two additional variables, schooling and
experience, both have zero coefﬁcients, by performing the F-test described above. The
test statistic in (2.59) can be computed from the R 2 s reported in Tables 2.1 and 2.2 as
f =

(0.1326 − 0.0317)/2
= 191.35.
(1 − 0.1326)/(3294 − 4)

With a 5% critical value of 3.00, the null hypothesis is obviously rejected. We can
thus conclude that the model that includes gender, schooling and experience performs
signiﬁcantly better than the model which only includes gender.
Table 2.2

OLS results wage equation

Dependent variable: wage
Variable

Estimate

Standard error

t-ratio

constant
male
school
exper

−3.3800
1.3444
0.6388
0.1248

0.4650
0.1077
0.0328
0.0238

−7.2692
12.4853
19.4780
5.2530

s = 3.0462 R 2 = 0.1326 R̄ 2 = 0.1318 F = 167.63

AN INTRODUCTION TO LINEAR REGRESSION

30

2.5.6 The General Case

The most general linear null hypothesis is a combination of the previous two cases
and comprises a set of J linear restrictions on the coefﬁcients. We can formulate these
restrictions as
Rβ = q,
where R is a J × K matrix, assumed to be of full row rank,14 and q is a J-dimensional
vector. An example of this is the set of restrictions β2 + β3 + · · · + βK = 1 and β2 =
β3 , in which case J = 2 and

R=

0 1
0 1

1 ···
−1 0

···
···

1
0



 
1
.
, q=
0

In principle it is possible to estimate the model imposing the above restrictions, such
that the test procedure of Subsection 2.5.4 can be employed. However, in many cases
these restrictions are such that it is hard to estimate under the null hypothesis (i.e.
imposing Rβ = q). In such a case, one can use the result that
Rb ∼ N(Rβ, σ 2 R(X X)−1 R  ),

(2.62)

such that a quadratic form can be constructed that has a Chi-squared distribution under
the null hypothesis, i.e.
ξ=

(Rb − q) (R(X X)−1 R  )−1 (Rb − q)
∼ χJ2 .
σ2

(2.63)

As σ 2 is unknown we have to replace it by its estimate s 2 . There are two ways to
continue. The ﬁrst one simply replaces σ 2 in (2.63) by s 2 and uses that the resulting
statistic is approximately χ 2 distributed (under the null, of course).15 Often, one refers
to this as a Wald test. The second way of going on uses (2.47) again such that a test
statistic can be deﬁned as the ratio of two independent χ 2 variables, i.e.
f =
=

(Rb − q) (σ 2 R(X X)−1 R  )−1 (Rb − q)/J
[(N − K)s 2 /σ 2 ]/(N − K)
(Rb − q) (R(X X)−1 R  )−1 (Rb − q)
,
J s2

(2.64)

which, under H0 , follows an F distribution with J and N − K degrees of freedom. As
before, large values of f lead to rejection of the null. It can be shown that the f statistic
in (2.64) is algebraically identical to the ones in (2.58) and (2.59) given above. It is
simply a matter of computational ease which one to use. Note that in large samples
f ≈ ξ /J .
14
15

Full row rank implies that the restrictions do not exhibit any linear dependencies.
The approximate result is obtained from the asymptotic distribution, and also holds if normality of the
error terms is not imposed (see below). The approximation is more accurate if the sample size is large.

HYPOTHESIS TESTING

31

2.5.7 Size, Power and p-Values

When an hypothesis is statistically tested two types of errors can be made. The ﬁrst
one is that we reject the null hypothesis while it is actually true, and is referred to as a
type I error. The second one, a type II error, is that the null hypothesis is not rejected
while the alternative is true. The probability of a type I error is directly controlled by
the researcher through his choice of the signiﬁcance level α. When a test is performed
at the 5% level, the probability of rejecting the null hypothesis while it is true is 5%.
This probability (signiﬁcance level) is often referred to as the size of the test. The
probability of a type II error depends upon the true parameter values. Intuitively, if the
truth deviates much from the stated null hypothesis, the probability of such an error
will be relatively small, while it will be quite large if the null hypothesis is close to the
truth. The reverse probability, that is, the probability of rejecting the null hypothesis
when it is false, is known as the power of the test. It indicates how ‘powerful’ a test
is in ﬁnding deviations from the null hypothesis (depending upon the true parameter
value). In general, reducing the size of a test will decrease its power, so that there is
a trade-off between type I and type II errors.
Suppose that we are testing the hypothesis that β2 = 0, while its true value is in fact
0.1. It is clear that the probability that we reject the null hypothesis depends upon the
standard error of our OLS estimator b2 and thus, among other things, upon the sample
size. The larger the sample the smaller the standard error and the more likely we are
to reject. This implies that type II errors become increasingly unlikely if we have large
samples. To compensate for this, researchers typically reduce the probability of type
I errors (that is of incorrectly rejecting the null hypothesis) by lowering the size α of
their tests. This explains why in large samples it is more appropriate to choose a size
of 1% or less rather than the ‘traditional’ 5%. Similarly, in very small samples we may
prefer to work with a signiﬁcance level of 10%.
Commonly, the null hypothesis that is chosen is assumed to be true unless there
is convincing evidence of the contrary. This suggests that if a test does not reject,
for whatever reason, we stick to the null hypothesis. This view is not completely
appropriate. A range of alternative hypotheses could be tested (for example β2 = 0,
β2 = 0.1 and β2 = 0.5), with the result that none of them is rejected. Obviously,
concluding that these three null hypotheses are simultaneously true would be ridiculous.
The only appropriate conclusion is that we cannot reject that β2 is 0, nor that it is
0.1 or 0.5. Sometimes, econometric tests are simply not very powerful and very large
sample sizes are needed to reject a given hypothesis.
A ﬁnal probability that plays a role in statistical tests is usually referred to as the
p-value. This p or probability value denotes the minimum size for which the null
hypothesis would still be rejected. It is deﬁned as the probability, under the null, to
ﬁnd a test statistic that (in absolute value) exceeds the value of the statistic that is
computed from the sample. If the p-value is smaller than the signiﬁcance level α, the
null hypothesis is rejected. Many modern software packages supply such p-values and
in this way allow researchers to draw their conclusions without consulting or computing
the appropriate critical values. It also shows the sensitivity of the decision to reject
the null hypothesis, with respect to the choice of signiﬁcance level. For example, a
p-value of 0.08 indicates that the null hypothesis is rejected at the 10% signiﬁcance
level, but not at the 5% level.

AN INTRODUCTION TO LINEAR REGRESSION

32

2.6

Asymptotic Properties of the OLS Estimator

In many cases, the small sample properties of the OLS estimator may deviate from
those discussed above. For example, if the error terms in the linear model εi do not
follow a normal distribution, it is no longer the case that the sampling distribution of
the OLS estimator b is normal. If assumption (A2) of the Gauss–Markov conditions is
violated, it can no longer be shown that b has an expected value of β. In fact, the linear
regression model under the Gauss–Markov assumptions and with normal error terms
is one of the very few cases in econometrics where the exact sampling distribution of
the parameter estimators is known. As soon as we relax some of these assumptions
or move to alternative models, the small sample properties of our estimators are typically unknown. In such cases we use an alternative approach to evaluate the quality
of our estimators, which is based on asymptotic theory. Asymptotic theory refers to
the question what happens if, hypothetically, the sample size grows inﬁnitely large.
Asymptotically, econometric estimators usually have nice properties, like normality,
and we use the asymptotic properties to approximate the properties in the ﬁnite sample that we happen to have. This section presents a ﬁrst discussion of the asymptotic
properties of the OLS estimator.
2.6.1 Consistency

Let us start with the linear model under the Gauss–Markov assumptions. In this case
we know that the OLS estimator b has the following ﬁrst two moments
E{b} = β
V {b} = σ 2



N


(2.65)

−1
xi xi

= σ 2 (X X)−1 .

(2.66)

i=1

Unless we assume that the error terms are normal, the shape of the distribution of
b is unknown. It is, however, possible to say something about the distribution of b,
at least approximately. A ﬁrst starting point is the so-called Chebycheff inequality,
which says that the probability that a random variable z deviates more than a positive
number δ from its mean, is bounded by its variance divided by δ 2 , that is
P {|z − E{z}| > δ} <

V {z}
, for all δ > 0.
δ2

(2.67)

For the OLS estimator this implies that its k-th element satisﬁes
P {|bk − βk | > δ} <

V {bk }
σ 2 ckk
=
for all δ > 0,
δ2
δ2

(2.68)


 −1
where ckk , as before, is the (k, k) element in (X X)−1 = ( N
i=1 xi xi ) . In most applications, the above inequality is not very useful as the upper bound on the probability
is larger than one. Let us, however, look at this inequality keeping δ ﬁxed and letting,
in our mind, the sample size N grow to inﬁnity. Then what happens? It is clear that

ASYMPTOTIC PROPERTIES OF THE OLS ESTIMATOR

33

N

xi xi increases as the number of terms grows, so that the variance of b decreases
as the sample size increases. If we assume that16
i=1

N
1 
x x  converges to a ﬁnite nonsingular matrix xx
N i=1 i i

(A6)

if the sample size N becomes inﬁnitely large, it follows directly from the above inequality that
(2.69)
lim P {|bk − βk | > δ} = 0 for all δ > 0.
N→∞

This says that, asymptotically, the probability that the OLS estimator deviates more
that δ from the true parameter value is zero. We usually refer to this property as ‘the
probability limit of b is β’, or ‘b converges in probability to β’, or just17
plim b = β.

(2.70)

Note that b is a vector of random variables, whose distribution depends on N, and β is
a vector of ﬁxed (unknown) numbers. When an estimator for β converges to the true
value, we say that it is a consistent estimator. Any estimator that satisﬁes (2.69) is a
consistent estimator for β, even if it is biased.
Consistency is a so-called large sample property and, loosely speaking, says that
if we obtain more and more observations, the probability that our estimator is some
positive number away from the true value β becomes smaller and smaller. Values that
b may take that are not close to β become increasingly unlikely. In many cases, one
cannot prove that an estimator is unbiased, and it is possible that no unbiased estimator exists (for example in nonlinear or dynamic models). In these cases, a minimum
requirement for an estimator to be useful appears to be that it is consistent. In the
sequel we shall therefore mainly be concerned with consistency of our estimators, not
with their (un)biasedness.
A useful property of probability limits (plims) is the following. If plim b = β and
g(.) is a continuous function (at least in the true value β), it also holds that
plim g(b) = g(β).

(2.71)

This guarantees that the parametrization employed is irrelevant for consistency. For
example, if s 2 is a consistent estimator for σ 2 , then s is a consistent estimator for σ .
Note that this result does not hold for unbiasedness, as E{s}2 = E{s 2 } (see Appendix B).
The non-singularity of xx requires that, asymptotically, there is no multicollinearity. The requirement
that the limit is ﬁnite is a ‘regularity’ condition, which will be satisﬁed in most empirical applications.
A sufﬁcient condition is that the x-variables are independent drawings from the same distribution with
a ﬁnite variance. Violations typically occur in time series contexts where one or more of the x-variables
may be trended. We shall return to this issue in Chapters 8 and 9.
17
Unless indicated otherwise, lim and plim refer to the (probability) limit for the sample size N going to
inﬁnity (N → ∞).
16

AN INTRODUCTION TO LINEAR REGRESSION

34

The OLS estimator is consistent under substantially weaker conditions than the ones
employed above. To see this, let us write the OLS estimator as

b=

N
1 
x x
N i=1 i i

−1

N
1 
xy =β+
N i=1 i i



N
1 
x x
N i=1 i i

−1

N
1 
xε.
N i=1 i i

(2.72)

In this expression, the sample averages of xi xi and xi εi play a role. If the sample size
increases the sample averages are taken over increasingly more observations. It seems
reasonable to assume, and it can be shown to be true under very weak conditions,18 that
in the limit these sample averages converge to the corresponding population means.
Now, under assumption (A6), we have that
−1
plim(b − β) = xx
E{xi εi },

(2.73)

which shows that the OLS estimator is consistent if it holds that
E{xi εi } = 0.

(A7)

This simply says that the error term is mean zero and uncorrelated with any of the
explanatory variables. Note that E{εi |xi } = 0 implies (A7), while the converse is not
necessarily true.19 Thus we can conclude that the OLS estimator b is consistent for β
under conditions (A6) and (A7). Typically, these conditions are much weaker than the
Gauss–Markov conditions (A1)–(A4) required for unbiasedness. We shall discuss the
relevance of this below.
Similarly, the least squares estimator s 2 for the error variance σ 2 is consistent under
conditions (A6), (A7) and (A3) (and some weak regularity conditions). The intuition
is that with b converging to β the residuals ei become asymptotically equivalent to
the error terms εi , so that the sample variance of ei will converge to the error variance
σ 2 , as deﬁned in (A3).
2.6.2 Asymptotic Normality

If the small sample distribution of an estimator is unknown, the best we can do is try
to ﬁnd some approximation. In most cases, one uses an asymptotic approximation (for
N going to inﬁnity) based on the asymptotic distribution. Most estimators in econometrics can be shown to be asymptotically normally distributed (under weak regularity
conditions). By√the asymptotic distribution of a consistent estimator β̂ we√mean the
distribution of N (β̂ − β) as N goes to inﬁnity. The reason for the factor N is that
asymptotically β̂ is equal to β with probability one for all consistent estimators. That
is, β̂ − β has a degenerate distribution for N → ∞ with all probability mass at zero. If
18

The result that sample averages converge to population means is provided in several versions of the
law of large numbers (see Davidson and MacKinnon, 1993, Section 4.5; Hayashi, 2000, Section 2.1;
or Greene, 2003, Appendix D).
19
To be precise, E{εi |xi } = 0 implies that E{εi g(xi )} = 0 for any function g (see Appendix B).

ASYMPTOTIC PROPERTIES OF THE OLS ESTIMATOR

35

√
√
we multiply by N and consider the asymptotic distribution √
of N (β̂ − β), this will
usually be a non-degenerate normal distribution. In that case N is referred to as the
rate of convergence and it is sometimes said that the corresponding estimator is rootN-consistent. In later chapters we shall see a few cases where the rate of convergence
differs from root N.
For the OLS estimator it can be shown that under the Gauss–Markov conditions
(A1)–(A4) combined with (A6) we have that
√



−1
N (b − β) → N 0, σ 2 xx
,

(2.74)

where → means ‘is asymptotically distributed as’. Thus, the OLS estimator b is asymp−1
. In practice,
totically normally distributed with variance–covariance matrix σ 2 xx
where we necessarily have a ﬁnite sample, we can use this result to approximate the
distribution of b as


a
−1
b ∼ N β, σ 2 xx
/N ,
(2.75)
a

where ∼ means ‘is approximately distributed as’.
Because
matrix xx will be consistently estimated by the sample mean
 the unknown

x
x
,
this
approximate
distribution is estimated as
(1/N ) N
i=1 i i

a

b ∼ N β, s 2

 N


−1 
xi xi

.

(2.76)

i=1

In (2.76) we have a distributional result for the OLS estimator b based upon the
asymptotic results, which is approximately valid in small samples. The quality of
the approximation increases as the sample size grows and typically it is hoped that
the sample size is sufﬁciently large for the approximation to be reasonably accurate.
Because the result in (2.76) corresponds exactly to what is used in the case of the
Gauss–Markov assumptions combined with the assumption of normal error terms, it
follows that all the distributional results for the OLS estimator reported above, including those for t- and F-statistics, are approximately valid, even if the errors are not
normally distributed.
Because asymptotically, a tN−K distributed variable converges to a standard normal
one, it is not uncommon to use the critical values from a standard normal distribution
(like the 1.96 at the 5% level) for all inferences, while not imposing normality of the
J
errors. Similarly, if f has an FN−K
distribution then asymptotically ξ = Jf has a χ 2
distribution with J degrees of freedom. To test a set of J linear restrictions on β, we
can thus use J times the f statistics and use the critical values from the asymptotic
Chi-squared distribution (compare (2.63) and (2.64)).
It is possible to further relax the assumptions without affecting the validity of the
results in (2.74) and (2.76). In particular, we can relax assumption (A2) to
xi and εi are independent.

(A8)

36

AN INTRODUCTION TO LINEAR REGRESSION

This condition does not rule out dependence between xi and εj for i = j , which is
of interest for models with lagged dependent variables. Note that (A8) implies (A7).
Further discussion on the asymptotic distribution of the OLS estimator and how it can
be estimated is provided in Chapters 4 and 5.
2.6.3 Small Samples and Asymptotic Theory

The linear regression model under the Gauss–Markov conditions is one of the very
few cases in econometrics in which the ﬁnite sample properties of the estimator and
test statistics are known. In many other circumstances and models, it is not possible or
extremely harsh to derive small sample properties of an econometric estimator. In such
cases, most econometricians are (necessarily) satisﬁed with knowing ‘approximate’
properties. As discussed above, such approximate properties are typically derived from
asymptotic theory in which one considers what happens to an estimator or test statistic
if the size of the sample is (hypothetically) growing to inﬁnity. As a result, one expects
that approximate properties based on asymptotic theory work reasonably well if the
sample size is sufﬁciently large.
Unfortunately, there is no unambiguous deﬁnition of what is ‘sufﬁciently large’. In
simple circumstances a sample size of 30 may be sufﬁcient, whereas in more complicated or extreme cases a sample of 1000 may still be insufﬁcient for the asymptotic
approximation to be reasonably accurate. To obtain some idea about the small sample
properties, Monte Carlo simulation studies are often performed. In a Monte Carlo
study, a large number (e.g. 1000) of simulated samples are drawn from a data generating process, speciﬁed by the researcher. Each (pseudo) random sample is used to
compute an estimator and/or a test statistic, and the distributional characteristics over
the different replications are analysed.
As an illustration, consider the following data generating process
yi = β1 + β2 xi + εi ,
corresponding to the simple linear regression model. To conduct a simulation, we need
to choose the distribution of xi , or ﬁx a set of values for xi , we need to specify the
values for β1 and β2 , and we need to specify the distribution of εi . Suppose we consider
samples of size N, with ﬁxed values xi = 1 for i = 1, . . . , N/2 (males, say) and xi = 0
otherwise (females).20 If εi ∼ NID(0, 1), independent of xi , the endogenous variable
yi is also normally distributed with mean β1 + β2 xi and unit variance. Given these
assumptions, a computer can easily generate a sample of N values for yi . Next, we
use this sample to compute the OLS estimator. Replicating this R times, with R newly
drawn samples, produces R estimators for β, b(1) , . . . , b(R) , say. Assuming β1 = 0 and
β2 = 1, Figure 2.2 presents a histogram of R = 1000 OLS estimates for β2 based on
1000 simulated samples of size N = 100. Because we know that the OLS estimator
is unbiased under these assumptions, we expect that b(r) is, on average, close to the
true value of 1. Moreover, from the results in Subsection 2.3.2 and because the R
replications are generated independently, we know that the slope coefﬁcient in b(r) is
distributed as
b2(r) ∼ NID(β2 , c22 ),
20

N is taken to be an even number.

ASYMPTOTIC PROPERTIES OF THE OLS ESTIMATOR

where β2 = 1 and


c22 = σ 2

N

(xi − x̄)2

37

−1
= 4/N.

i=1

The larger the number of replications, the more the histogram in Figure 2.2 will resemble the normal distribution. For ease of comparison, the normal density is also drawn.
A Monte Carlo study allows us to investigate the statistical properties of the OLS
estimator, or any other estimator we may be interested in, as a function of the way in
which the data are generated. Further, we can analyse the properties of test statistics,
e.g. for the t or F-test. Most interestingly this is done in cases where one of the
model assumptions (A2), (A3), (A4) or (A5) is violated. For example, what happens
if εi has a skewed or fat-tailed distribution (like a t-distribution with 10 degrees of
freedom)? Asymptotic theory teaches us that the distribution of the OLS estimator
will be approximately normal, but what is ‘approximately’, and how does it depend
upon the sample size N ? Further, we can use the Monte Carlo study to analyse the
distribution of a test statistic when the null hypothesis is false. For example, we may
analyse the probability that the null hypothesis that β2 = 0.5 is rejected, as a function
of the true value of β2 (and the sample size N ). If the true value is 0.5, this gives
us the (actual) size of the test, whereas for β2 = 0.5, we obtain the power of the test.
Finally, we can use a simulation study to analyse the properties of an estimator on the
basis of a model that deviates from the data generating process, for example a model
that omits a relevant explanatory variable.
While Monte Carlo studies are useful, their results usually strongly depend upon
the choices for xi , β, σ 2 and the sample size N, and therefore cannot necessarily be
extrapolated to different settings. Nevertheless, they provide interesting information
about the statistical properties of an estimator or test statistic under controlled circumstances. Fortunately, for the linear regression model the asymptotic approximation
0.12

0.10

0.08

0.06

0.04

0.02

0.00
0

Figure 2.2

0.5

1

1.5

2

Histogram of 1000 OLS estimates with normal density (Monte Carlo results)

AN INTRODUCTION TO LINEAR REGRESSION

38

usually works quite well. As a result, for most applications it is reasonably safe to state
that the OLS estimator is approximately normally distributed. More information about
Monte Carlo experiments is provided in Davidson and MacKinnon (1993, Chapter 21),
while a simple illustration is provided in Patterson (2000, Section 8.2).

2.7

Illustration: The Capital Asset Pricing Model

One of the most important models in ﬁnance is the Capital Asset Pricing Model
(CAPM). The CAPM is an equilibrium model which assumes that all investors compose their asset portfolio on the basis of a trade-off between the expected return and
the variance of the return on their portfolio. This implies that each investor holds a
so-called mean variance efﬁcient portfolio, a portfolio that gives maximum expected
return for a given variance (level of risk). If all investors hold the same beliefs about
expected returns and (co)variances of individual assets, and in the absence of transaction costs, taxes and trading restrictions of any kind, it is also the case that the
aggregate of all individual portfolios, the market portfolio, is mean variance efﬁcient.
In this case, it can be shown that expected returns on individual assets are linearly
related to the expected return on the market portfolio. In particular, in holds that21
E{rj t − rf } = βj E{rmt − rf },

(2.77)

where rj t is the risky return on asset j in period t, rmt the risky return on the market
portfolio, and rf denotes the riskless return, which we assume to be time-invariant for
simplicity. The proportionality factor βj is given by
βj =

cov{rj t , rmt }
V {rmt }

(2.78)

and indicates how strong ﬂuctuations in the returns on asset j are related to movements
of the market as a whole. As such, it is a measure of systematic risk (or market risk).
Because it is impossible to eliminate systematic risk through a diversiﬁcation of one’s
portfolio without affecting the expected return, investors are compensated for bearing
this source of risk through a risk premium E{rmt − rf } > 0.
In this section, we consider the CAPM and see how it can be rewritten as a linear
regression model, which allows us to estimate and test it. A more extensive discussion
of empirical issues related to the CAPM can be found in Berndt (1991) or, more technically, in Campbell, Lo and MacKinlay (1997, Chapter 5) and Gourieroux and Jasiak
(2001, Section 4.2). More details on the CAPM can be found in ﬁnance textbooks, for
example Elton and Gruber (2003).
2.7.1 The CAPM as a Regression Model

The relationship in (2.77) is an ex ante equality in terms of unobserved expectations.
Ex post, we only observe realized returns on the different assets over a number of
21

Because the data correspond to different time periods, we index the observations by t, t = 1, 2, . . . , T ,
rather than i.

ILLUSTRATION: THE CAPITAL ASSET PRICING MODEL

39

periods. If, however, we make the usual assumption that expectations are rational, so
that expectations of economic agents correspond to mathematical expectations, we can
derive a relationship from (2.77) that involves actual returns. To see this, let us deﬁne
the unexpected returns on asset j as
uj t = rj t − E{rj t },
and the unexpected returns on the market portfolio as
umt = rmt − E{rmt }.
Then, it is possible to rewrite (2.77) as
rj t − rf = βj (rmt − rf ) + εj t ,

(2.79)

where
εj t = uj t − βj umt .
Equation (2.79) is a regression model, without an intercept, where εj t is treated as an
error term. This error term is not something that is just added to the model, but it has
a meaning, being a function of unexpected returns. It is easy to show, however, that it
satisﬁes some minimal requirements for a regression error term, as given in (A7). For
example, it follows directly from the deﬁnitions of umt and uj t that it is mean zero, i.e.
E{εj t } = E{uj t } − βj E{umt } = 0.

(2.80)

Furthermore, it is uncorrelated with the regressor rmt − rf . This follows from the
deﬁnition of βj , which can be written as
βj =

E{uj t umt }
V {umt }

,

(note that rf is not stochastic) and the result that
E{εj t (rmt − rf )} = E{(uj t − βj umt )umt } = E{uj t umt } − βj E{u2mt }.
From the previous section, it then follows that the OLS estimator provides a consistent
estimator for βj . If, in addition, we impose assumption (A8) that εj t is independent of
rmt − rf and assumptions (A3) and (A4) stating that εj t does not exhibit autocorrelation
or heteroskedasticity, we can use the asymptotic result in (2.74) and the approximate
distributional result in (2.76). This implies that routinely computed OLS estimates,
standard errors and tests are appropriate, by virtue of the asymptotic approximation.
2.7.2 Estimating and Testing the CAPM

The CAPM describes the expected returns on any asset or portfolio of assets as a
function of the (expected) return on the market portfolio. In this subsection, we consider
the returns on three different industry portfolios, while approximating the return on the

AN INTRODUCTION TO LINEAR REGRESSION

40

market portfolio by the return on a value-weighted stock market index. Returns for
the period January 1960 to December 2002 (516 months) for the food, durables and
construction industries were obtained from the Center for Research in Security Prices
(CRSP).22 The industry portfolios are value-weighted and are rebalanced once every
year. While theoretically the market portfolio should include all tradeable assets, we
shall assume that the CRSP value-weighted index is a good approximation. The riskless
rate is approximated by the return on 1-month treasury bills. Although this return is
time-varying, it is known to investors while making their decisions. All returns are
expressed in percentage per month.
First, we estimate the CAPM relationship (2.79) for these three industry portfolios. That is, we regress excess returns on the industry portfolios (returns in excess
of the riskless rate) upon excess returns on the market index proxy, not including an
intercept. This produces the results presented in Table 2.3. The estimated beta coefﬁcients indicate how sensitive the value of the industry portfolios are to general market
movements. This sensitivity is relatively low for the food industry, but fairly high for
durables and construction: an excess return on the market of, say, 10% corresponds to
an expected excess return on the food, durables and construction portfolios of 7.9, 11.1
and 11.6% respectively. It is not surprising to see that the durables and construction
industries are more sensitive to overall market movements than is the food industry.
Assuming that the conditions required for the distributional results of the OLS estimator are satisﬁed, we can directly test the hypothesis (which has some economic interest)
that βj = 1 for each of the three industry portfolios. This results in t-values of −7.38,
3.89 and 6.21, respectively, so that we reject the null hypothesis for each of the three
industry portfolios.
As the CAPM implies that the only relevant variable in the regression is the excess
return on the market portfolio, any other variable (known to the investor when making
his decisions) should have a zero coefﬁcient. This also holds for a constant term. To
check whether this is the case, we can re-estimate the above models while including an
intercept term. This produces the results in Table 2.4. From these results, we can test
the validity of the CAPM by testing whether the intercept term is zero. For food, the
appropriate t-statistic is 2.66, which implies that we reject the validity of the CAPM
at the 5% level. The point estimate of 0.339 implies that the food industry portfolio
is expected to have a return that is 0.34% per month higher than the CAPM predicts.
Note that the estimated beta coefﬁcients are very similar to those in Table 2.3 and that
the R 2 s are close to the uncentred R 2 s.
Table 2.3 CAPM regressions (without intercept)
Dependent variable: excess industry portfolio returns
Industry
excess market return
uncentred R 2
s

Food

Durables

Construction

0.790
(0.028)

1.113
(0.029)

1.156
(0.025)

0.601
2.902

0.741
2.959

0.804
2.570

Note: Standard errors in parentheses.
22

The data for this illustration are available as CAPM2.

ILLUSTRATION: THE CAPITAL ASSET PRICING MODEL

41

Table 2.4 CAPM regressions (with intercept)
Dependent variable: excess industry portfolio returns
Industry

Food

Durables

Construction

constant

0.339
(0.128)
0.783
(0.028)

0.064
(0.131)
1.111
(0.029)

−0.053
(0.114)
1.157
(0.025)

0.598
2.885

0.739
2.961

0.803
2.572

excess market return
R2
s

Note: Standard errors in parentheses.

The R 2 s in these regressions have an interesting economic interpretation.
Equation (2.79) allows us to write that
V {rj t } = βj2 V {rj t } + V {εj t },
which shows that the variance of the return on a stock (portfolio) consists of two parts:
a part related to the variance of the market index and an idiosyncratic part. In economic
terms, this says that total risk equals market risk plus idiosyncratic risk. Market risk
is determined by βj and is rewarded: stocks with a higher βj provide higher expected
returns because of (2.77). Idiosyncratic risk is not rewarded because it can be eliminated
by diversiﬁcation: if we construct a portfolio that is well diversiﬁed, it will consist of a
large number of assets, with different characteristics, so that most of the idiosyncratic
risk will cancel out and mainly market risk matters. The R 2 , being the proportion of
explained variation in total variation, is an estimate of the relative importance of market
risk for each of the industry portfolios. For example, it is estimated that 59.8% of the
risk (variance) of the food industry portfolio is due to the market as a whole, while
40.2% is idiosyncratic (industry-speciﬁc) risk. The durables and construction industries
appear to be better diversiﬁed.
Finally, we consider one deviation from the CAPM that is often found in empirical
work: the existence of a January effect. There is some evidence that, ceteris paribus,
returns in January are higher than in any of the other months. We can test this within
the CAPM framework by including a dummy in the model for January and testing
Table 2.5

CAPM regressions (with intercept and January dummy)

Dependent variable: excess industry portfolio returns
Industry

Food

Durables

Construction

constant

0.417
(0.133)
−0.956
(0.456)
0.788
(0.028)

0.069
(0.137)
−0.063
(0.473)
1.112
(0.029)

−0.094
(0.118)
0.498
(0.411)
1.155
(0.025)

0.601
2.876

0.739
2.964

0.804
2.571

January dummy
excess market return
R2
s

Note: Standard errors in parentheses.

AN INTRODUCTION TO LINEAR REGRESSION

42

whether it is signiﬁcant. By doing this, we obtain the results in Table 2.5. Computing
the t-statistics corresponding to the January dummy shows that for two of the three
industry portfolios we do not reject the absence of a January effect at the 5% level. For
the food industry, however, the January effect appears to be negative and statistically
signiﬁcant. Consequently, the results do not provide support for the existence of a
positive January effect.

2.8

Multicollinearity

In general, there is nothing wrong with including variables in your model that are
correlated. In an individual wage equation, for example, we may want to include both
age and experience, although it can be expected that older persons, on average, have
more experience. However, if the correlation between two variables is too high, this
may lead to problems. Technically, the problem is that the matrix X X is close to
being not invertible. This may lead to unreliable estimates with high standard errors
and of unexpected sign or magnitude. Intuitively, the problem is also clear. If age and
experience are highly correlated it may be hard for the model to identify the individual
impact of these two variables, which is exactly what we are trying to do. In such a case,
a large number of observations with sufﬁcient variation in both age and experience
may help us to get sensible answers. If this is not the case and we do get poor estimates
(for example: t-tests show that neither age nor experience are individually signiﬁcant),
we can only conclude that there is insufﬁcient information in the sample to identify
the effects we would like to identify. In the wage equation, we are trying to identify
the effect of age, keeping experience and the other included variables constant, as
well as the effect of experience, keeping age and the other variables constant (the
ceteris paribus condition). It is clear that in the extreme case that people with the same
age would have the same level of experience we would not be able to identify these
effects. In the case where age and experience are highly but not perfectly correlated,
the estimated effects are likely to be highly inaccurate.
In general, the term multicollinearity is used to describe the problem when an
approximate linear relationship among the explanatory variables leads to unreliable
regression estimates. This approximate relationship is not restricted to two variables
but can involve more or even all regressors. In the wage equation, for example, the
problems may be aggravated if we include years of schooling in addition to age and
years of experience. In the extreme case, one explanatory variable is an exact linear
combination of one or more other explanatory variables (including the intercept). This
is usually referred to as exact multicollinearity, in which case the OLS estimator is
not uniquely deﬁned from the ﬁrst order conditions of the least squares problem (the
matrix X X is not invertible).
The use of too many dummy variables (which are either zero or one) is a typical
cause for exact multicollinearity. Consider the case where we would like to include
a dummy for males (male i ), a dummy for females (female i ) as well as a constant.
Because male i + female i = 1 for each observation (and 1 is included as the constant),
the X X matrix becomes singular. Exact multicollinearity is easily solved by excluding
one of the variables from the model and estimating the model including either male i and
a constant, female i and a constant, or both male i and female i but no constant. The latter

MULTICOLLINEARITY

43

approach is not recommended because standard software tends to compute statistics
like the R 2 and the F-statistic in a different way if the constant is suppressed; see the
illustration in the next subsection. Another useful example of exact multicollinearity
in this context is the inclusion of the variables age, years of schooling and potential
experience, deﬁned as age minus years of schooling minus six. Clearly, this leads to
a singular X X matrix if a constant is included in the model (see Section 5.4 for an
illustration).
To illustrate the effect of multicollinearity on the OLS estimator in more detail,
consider the following example. Let the following regression model be estimated,
yi = β1 xi1 + β2 xi2 + εi ,
where it is assumed that the sample means ȳ = x̄1 = x̄2 = 0.23 Moreover, assume
that the sample variances of xi1 and xi2 are equal to 1, while the sample covariance
(correlation coefﬁcient) is r12 . Then the variance of the OLS estimator can be written as
1
V {b} = σ
N
2



1
r12

r12
1

−1

σ 2 /N
=
2
1 − r12



1
−r12

−r12
1


.

It is clear that the variances of both b1 and b2 increase if the absolute value of the
correlation coefﬁcient between x1 and x2 increases.24 Due to the increased variance of
the OLS estimator, t-statistics will be smaller. If xi1 and xi2 show a strong positive
correlation (r12 > 0), the estimators b1 and b2 will be negatively correlated.
Another consequence of multicollinearity is that some linear combinations of the
parameters are pretty accurately estimated, while other linear combinations are highly
inaccurate. Usually, when regressors are positively correlated, the sum of the regression
coefﬁcients can be rather precisely determined, while the difference cannot. In the above
example we have for the variance of b1 + b2 that
V {b1 + b2 } =

σ 2 /N
σ 2 /N
(2 − 2r12 ) = 2
,
2
1 + r12
1 − r12

while for the variance of the difference we have
V {b1 − b2 } =

σ 2 /N
σ 2 /N
(2
+
2r
)
=
2
.
12
2
1 − r12
1 − r12

So if r12 is close to 1, the variance of b1 − b2 is many times higher than the variance of
b1 + b2 . For example, if r12 = 0.95 the ratio of the two variances is 39. An important
consequence of this result is that for prediction purposes, in particular the accuracy of
prediction, multicollinearity typically has little impact. This is a reﬂection of the fact
that the ‘total impact’ of all explanatory variables is accurately identiﬁed.
23

This can be achieved by deducting the sample mean from all variables. In this case, there is no need for
a constant term because the OLS estimator of the intercept will be equal to zero.
24
Note that this also holds if the true value of one of the regression coefﬁcients is zero. Thus, including
unnecessary regressors in a model decreases the precision of the OLS estimator for the other coefﬁcients
(see Chapter 3).

AN INTRODUCTION TO LINEAR REGRESSION

44

Table 2.6 Alternative speciﬁcations with dummy variables
Dependent variable: wage
Speciﬁcation
constant
male
female
R 2 , uncentred R 2

A

B

C

5.147
(0.081)
1.166
(0.112)
–

6.313
(0.078)
–

–

0.0317

−1.166
(0.112)

6.313
(0.078)
5.147
(0.081)

0.0317

0.7640

Note: Standard errors in parentheses.

In summary, high correlations between (linear combinations of) explanatory variables
may result in multicollinearity problems. If this happens, one or more parameters we
are interested in are estimated highly inaccurately. Essentially, this means that our
sample does not provide sufﬁcient information about these parameters. To alleviate the
problem, we are therefore forced to use more information, for example by imposing
some a priori restrictions on the vector of parameters. Commonly, this means that one
or more variables are omitted from the model. Another solution, which is typically
not practical, is to extend the sample size. As illustrated by the above example all
variances decrease as the sample size increases. An extensive and critical survey of the
multicollinearity problem and the (in)appropriateness of some mechanical procedures
to solve it, is provided in Maddala (2001, Chapter 7).
2.8.1 Example: Individual Wages (Continued)
Let us go back to the simple wage equation of Subsection 2.3.3. As explained above,
the addition of a female dummy to the model would cause exact multicollinearity.
Intuitively, it is also obvious that with only two groups of people one dummy variable
and a constant are sufﬁcient to capture them. The choice of whether to include the male
or the female dummy is arbitrary. The fact that the two dummy variables add up to one
for each observation does not imply multicollinearity if the model does not contain an
intercept term. Consequently, it is possible to include both dummies while excluding
the intercept term. To illustrate the consequences of these alternative choices, consider
the estimation results in Table 2.6.
As speciﬁcation C does not contain an intercept term, the uncentred R 2 is provided
rather than the R 2 , which explains its high value. As before, the coefﬁcient for the
male dummy in speciﬁcation A denotes the expected wage differential between men
and women. Similarly, the coefﬁcient for the female dummy in the second speciﬁcation
denotes the expected wage differential between women and men. For speciﬁcation C,
however, the coefﬁcients for male and female reﬂect the expected wage for men and
women, respectively. It is quite clear that all three speciﬁcations are equivalent, while
their parametrization is somewhat different.

2.9

Prediction

An econometrician’s work does not end after having produced the coefﬁcient estimates
and corresponding standard errors. A next step is to interpret the results and to use the

PREDICTION

45

model for the goals it was intended. One of these goals, particularly with time series
data, is prediction. In this section we consider prediction using the regression model,
that is, we want to predict the value for the dependent variable at a given value for
the explanatory variables, x0 . Given that the model is assumed to hold for all potential
observations, it will also hold that
y0 = x0 β + ε0 ,
where ε0 satisﬁes the same properties as all other error terms. The obvious predictor
for y0 is ŷ0 = x0 b. As E{b} = β it is easily veriﬁed that this is an unbiased predictor,
i.e.25 E{ŷ0 − y0 } = 0. Under assumptions (A1)–(A4), the variance of the predictor is
given by
V {ŷ0 } = V {x0 b} = x0 V {b}x0 = σ 2 x0 (X X)−1 x0 .
(2.81)
This variance, however, is only an indication of the variation in the predictor if different
samples would be drawn, that is the variation in the predictor due to variation in b. To
analyse how accurate the predictor is, we need the variance of the prediction error,
which is deﬁned as
y0 − ŷ0 = x0 β + ε0 − x0 b = ε0 − x0 (b − β).

(2.82)

The prediction error has variance
V {y0 − ŷ0 } = σ 2 + σ 2 x0 (X X)−1 x0

(2.83)

provided that it can be assumed that b and ε0 are uncorrelated. This is usually not
a problem because ε0 is not used in the estimation of β. In the simple regression
model (with one explanatory variable xi ), one can rewrite the above expression as
(see Maddala, 2001, Section 3.7)


(x0 − x̄)2
1
2
2
+
V {y0 − ŷ0 } = σ + σ
.
2
N
i (xi − x̄)
Consequently, the further the value of x0 is from the sample mean x̄, the larger the
variance of the prediction error. This is a sensible result: if we want to predict y for
extreme values of x we cannot expect it to be very accurate.
Finally, we can compute a so-called prediction interval. A 95% prediction interval
for y0 is given by




x0 b − 1.96s 1 + x0 (X X)−1 x0 , x0 b + 1.96s 1 + x0 (X X)−1 x0 ,
(2.84)
where, as before, 1.96 is the critical value of a standard normal distribution. With a
probability of 95%, this interval contains the true unobserved value y0 .
As one of the important goals of dynamic models is forecasting, we shall return to
the prediction issue in Chapter 8.
25

In this expectation both ŷ0 and y0 are treated as random variables.

AN INTRODUCTION TO LINEAR REGRESSION

46

Exercises
Exercise 2.1 (Regression)

Consider the following linear regression model:
yi = β1 + β2 xi2 + β3 xi3 + εi = xi β + εi .
a.
b.
c.
d.
e.
f.
g.
h.
i.

j.

Explain how the ordinary least squares estimator for β is determined and derive
an expression for b.
Which assumptions are needed to make b an unbiased estimator for β?
Explain how a conﬁdence interval for β2 can be constructed. Which additional
assumptions are needed?
Explain how one can test the hypothesis that β3 = 1.
Explain how one can test the hypothesis that β2 + β3 = 0.
Explain how one can test the hypothesis that β2 = β3 = 0.
Which assumptions are needed to make b a consistent estimator for β?
Suppose that xi2 = 2 + 3xi3 . What will happen if you try to estimate the
above model?
∗
Suppose that the model is estimated with xi2
= 2xi2 − 2 included rather than xi2 .
How are the coefﬁcients in this model related to those in the original model? And
the R 2 s?
Suppose that xi2 = xi3 + ui , where ui and xi3 are uncorrelated. Suppose that the
model is estimated with ui included rather than xi2 . How are the coefﬁcients in
this model related to those in the original model? And the R 2 s?

Exercise 2.2 (Individual Wages)

Using a sample of 545 full-time workers in the USA, a researcher is interested in
the question whether women are systematically underpaid compared to men. First, she
estimates the average hourly wages in the sample for men and women, which are $5.91
and $5.09, respectively.
a.

Do these numbers give an answer to the question of interest? Why not? How could
one (at least partially) correct for this?

The researcher also runs a simple regression of an individual’s wage on a male dummy,
equal to 1 for males and 0 for females. This gives the results reported in Table 2.7.
Table 2.7

Hourly wages explained from gender: OLS results

Variable

Estimate

Standard error

t-ratio

constant
male

5.09
0.82

0.58
0.15

8.78
5.47

N = 545 s = 2.17 R 2 = 0.26

EXERCISES

47

b. How can you interpret the coefﬁcient estimate of 0.82? How do you interpret the
estimated intercept of 5.09?
c. How do you interpret the R 2 of 0.26?
d. Explain the relationship between the coefﬁcient estimates in the table and the
average wage rates of males and females.
e. A student is unhappy with this model as ‘a female dummy is omitted from the
model’. Comment upon this criticism.
f. Test, using the above results, the hypothesis that men and women have, on average,
the same wage rate, against the one-sided alternative that women earn less. State
the assumptions required for this test to be valid.
g. Construct a 95% conﬁdence interval for the average wage differential between
males and females in the population.
Subsequently, the above ‘model’ is extended to include differences in age and education, by including the variables age (age in years) and educ (education level, from 1
to 5). Simultaneously, the endogenous variable is adjusted to be the natural logarithm
of the hourly wage rate. The results are reported in Table 2.8.
h. How do you interpret the coefﬁcients of 0.13 for the male dummy, and 0.09
for age?
i. Test the joint hypothesis that gender, age and education do not affect a person’s wage.
j. A student is unhappy with this model as ‘the effect of education is rather restrictive’. Can you explain this criticism? How could the model be extended or
changed to meet the above criticism? How can you test whether the extension
has been useful?
The researcher re-estimates the above model including age2 as an additional regressor.
The t-value on this new variable becomes −1.14, while R 2 = 0.699 and R̄ 2 increases
to 0.683.
k. Could you give a reason why the inclusion of age2 might be appropriate?
l. Would you retain this new variable given the R 2 and the R̄ 2 measures? Would
you retain age2 given its t-value? Explain this apparent conﬂict in conclusions.
Table 2.8 Log hourly wages explained from gender, age
and education level: OLS results
Variable

Coefﬁcient

Standard error

t-ratio

constant
male
age
educ

−1.09
0.13
0.09
0.18

0.38
0.03
0.02
0.05

2.88
4.47
4.38
3.66

N = 545 s = 0.24 R 2 = 0.691 R̄ 2 = 0.682

AN INTRODUCTION TO LINEAR REGRESSION

48

Exercise 2.3 (Asset Pricing – Empirical)

In the recent ﬁnance literature it is suggested that asset prices are fairly well described
by a so-called factor model, where excess returns are linearly explained from excess
returns on a number of ‘factor portfolios’. As in the CAPM, the intercept term should
be zero, just like the coefﬁcient for any other variable included in the model the value
of which is known in advance (e.g. a January dummy). The data set ASSETS2 contains
excess returns on four factor portfolios for January 1960 to December 2002:26
rmrf:
smb:
hml:
umd:

excess return on a value-weighted market proxy
return on a small-stock portfolio minus the return
on a large-stock portfolio (Small minus Big)
return on a value-stock portfolio minus the return
on a growth-stock portfolio (High minus Low)
return on a high prior return portfolio minus the return
on a low prior return portfolio (Up minus Down)

All data are for the USA. Each of the last three variables denotes the difference in
returns on two hypothetical portfolios of stocks. These portfolios are re-formed each
month on the basis of the most recent available information on ﬁrm size, book-tomarket value of equity and historical returns, respectively. The hml factor is based
on the ratio of book value to market value of equity, and reﬂects the difference in
returns between a portfolio of stocks with a high book-to-market ratio (value stocks)
and a portfolio of stocks with a low book-to-market ratio (growth stocks). The factors
are motivated by empirically found anomalies of the CAPM (for example, small ﬁrms
appear to have higher returns than large ones, even after the CAPM risk correction).
In addition to the excess returns on these four factors, we have observations on
the returns on ten different ‘assets’ which are ten portfolios of stocks, maintained by
the Center for Research in Security Prices (CRSP). These portfolios are size-based,
which means that portfolio 1 contains the 10% smallest ﬁrms listed at the New York
Stock Exchange and portfolio 10 contains the 10% largest ﬁrms that are listed. Excess
returns (in excess of the riskfree rate) on these portfolios are denoted by r1 to r10,
respectively.
In answering the following questions use r1, r10 and the returns on two additional
portfolios that you select.
a.

b.
c.
d.
e.

26

Regress the excess returns on your four portfolios upon the excess return on the
market portfolio (proxy), noting that this corresponds to the CAPM. Include a
constant in these regressions.
Give an economic interpretation of the estimated β coefﬁcients.
Give an economic and a statistical interpretation of the R 2 s.
Test the hypothesis that βj = 1 for each of the four portfolios. State the assumptions you need to make for the tests to be (asymptotically) valid.
Test the validity of the CAPM by testing whether the constant terms in the four
regressions are zero.

All data for this exercise are taken from the website of Kenneth French; see http://mba.tuck/dartmouth.edu/
pages/faculty/ken.french.

EXERCISES

49

f. Test for a January effect in each of the four regressions.
g. Next, estimate the four-factor model
rj t = αj + βj 1 rmrft + βj 2 smbt + βj 3 hmlt + βj 4 umdt + εj t
by OLS. Compare the estimation results with those obtained from the one-factor
(CAPM) model. Pay attention to the estimated partial slope coefﬁcients and
the R 2 s.
h. Perform F-tests for the hypothesis that the coefﬁcients for the three new factors
are jointly equal to zero.
i. Test the validity of the four-factor model by testing whether the constant terms in
the four regressions are zero. Compare your conclusions with those obtained from
the CAPM.
Exercise 2.4 (Regression – True or False?)

Carefully read the following statements. Are they true or false? Explain.
a. Under the Gauss–Markov conditions, OLS can be shown to be BLUE. The phrase
‘linear’ in this acronym refers to the fact that we are estimating a linear model.
b. In order to apply a t-test, the Gauss–Markov conditions are strictly required.
c. A regression of the OLS residual upon the regressors included in the model by
construction yields an R 2 of zero.
d. The hypothesis that the OLS estimator is equal to zero can be tested by means of
a t-test.
e. From asymptotic theory, we learn that – under appropriate conditions – the error
terms in a regression model will be approximately normally distributed if the
sample size is sufﬁciently large.
f. If the absolute t-value of a coefﬁcient is smaller than 1.96, we accept the null
hypothesis that the coefﬁcient is zero, with 95% conﬁdence.
g. Because OLS provides the best linear approximation of a variable y from a set of
regressors, OLS also gives best linear unbiased estimators for the coefﬁcients of
these regressors.
h. If a variable in a model is signiﬁcant at the 10% level, it is also signiﬁcant at the
5% level.

3

Interpreting and
Comparing Regression
Models

In the previous chapter attention was paid to the estimation of linear regression models.
In particular, the ordinary least squares approach was discussed, including its properties
under several sets of assumptions. This allowed us to estimate the vector of unknown
parameters β and to test parametric restrictions, like βk = 0. In the ﬁrst section of
this chapter we pay additional attention to the interpretation of regression models and
their coefﬁcients. In Section 3.2, we discuss how we can select the set of regressors
to be used in our model and what the consequences are if we misspecify this set.
This also involves comparing alternative models. Section 3.3 discusses the assumption
of linearity and how it can be tested. To illustrate the main issues, this chapter is
concluded with two empirical examples. Section 3.4 describes a model to explain house
prices, while Section 3.5 discusses the estimation and speciﬁcation of an individual
wage equation.

3.1

Interpreting the Linear Model

As already stressed in the previous chapter, the linear model
yi = xi β + εi

(3.1)

has little meaning unless we complement it with additional assumptions on εi . It is
common to state that εi has expectation zero and that the xi s are taken as given. A
formal way of stating this is that it is assumed that the expected value of εi given X,
or the expected value of εi given xi , is zero, that is
E{εi |X} = 0 or

E{εi |xi } = 0,

(3.2)

52

INTERPRETING AND COMPARING REGRESSION MODELS

respectively, where the latter condition is implied by the ﬁrst. Under E{εi |xi } = 0,
we can interpret the regression model as describing the conditional expected value of
yi given values for the explanatory variables xi . For example, what is the expected
wage for an arbitrary woman of age 40, with a university education and 14 years of
experience? Or, what is the expected unemployment rate given wage rates, inﬂation
and total output in the economy? The ﬁrst consequence of (3.2) is the interpretation
of the individual β coefﬁcients. For example, βk measures the expected change in yi
if xik changes with one unit but all the other variables in xi do not change. That is
∂E{yi |xi }
= βk .
∂xik

(3.3)

It is important to realize that we had to state explicitly that the other variables in xi did
not change. This is the so-called ceteris paribus condition. In a multiple regression
model single coefﬁcients can only be interpreted under ceteris paribus conditions. For
example, βk could measure the effect of age on the expected wage of a woman, if the
education level and years of experience are kept constant. An important consequence
of the ceteris paribus condition is that it is not possible to interpret a single coefﬁcient
in a regression model without knowing what the other variables in the model are. If
interest is focussed on the relationship between yi and xik , the other variables in xi act
as control variables. For example, we may be interested in the relationship between
house prices and the number of bedrooms, controlling for differences in lot size and
location. Depending upon the question of interest, we may decide to control for some
factors but not for all (see Wooldridge, 2003, Section 6.3, for more discussion).
Sometimes these ceteris paribus conditions are hard to maintain. For example, in
the wage equation case, it may be very common that a changing age almost always
corresponds to changing years of experience. Although the βk coefﬁcient in this case
still measures the effect of age keeping years of experience (and the other variables)
ﬁxed, it may not be very well identiﬁed from a given sample, due to the collinearity
between the two variables. In some cases, it is just impossible to maintain the ceteris
paribus condition, for example if xi includes both age and age-squared. Clearly, it is
ridiculous to say that a coefﬁcient βk measures the effect of age given that age-squared
is constant. In this case, one should go back to the derivative (3.3). If xi β includes,
say, age i β2 + age 2i β3 , we can derive
∂E{yi |xi }
= β2 + 2 age i β3 ,
∂age i

(3.4)

which can be interpreted as the marginal effect of a changing age if the other variables in xi (excluding age 2i ) are kept constant. This shows how the marginal effects
of explanatory variables can be allowed to vary over the observations by including
additional terms involving these variables (in this case age 2i ). For example, we can
allow the effect of age to be different for men and women by including an interaction
term age i male i in the regression, where male i is a dummy for males. Thus, if the
model includes age i β2 + age i male i β3 the effect of a changing age is
∂E{yi |xi }
= β2 + male i β3 ,
∂age i

(3.5)

INTERPRETING THE LINEAR MODEL

53

which is β2 for females and β2 + β3 for males. Sections 3.4 and 3.5 will illustrate the
use of such interaction terms.
Frequently, economists are interested in elasticities rather than marginal effects.
An elasticity measures the relative change in the dependent variable due to a relative
change in one of the xi variables. Often, elasticities are estimated directly from a linear
regression model involving the logarithms of most explanatory variables (excluding
dummy variables), that is
log yi = (log xi ) γ + vi ,
(3.6)
where log xi is shorthand notation for a vector with elements (1, log xi2 , . . . , log xiK )
and it is assumed that E{vi | log xi } = 0. We shall call this a loglinear model. In
this case,
xik
∂E{yi |xi }
∂E{log yi | log xi }
·
= γk ,
(3.7)
≈
∂xik
E{yi |xi }
∂ log xik
where the ≈ is due to the fact that E{log yi | log xi } = E{log yi |xi } = log E{yi |xi }. Note
that (3.3) implies that in the linear model
∂E{yi |xi }
xik
x
·
= ik βk ,
∂xik
E{yi |xi }
xi β

(3.8)

which shows that the linear model implies that elasticities are nonconstant and vary
with xi , while the loglinear model imposes constant elasticities. While in many cases
the choice of functional form is dictated by convenience in economic interpretation,
other considerations may play a role. For example, explaining log yi rather than yi
may help reducing heteroskedasticity problems, as illustrated in Section 3.5 below.
In Section 3.3 we shall brieﬂy consider statistical tests for a linear versus a loglinear
speciﬁcation.
If xik is a dummy variable (or another variable that may take nonpositive values)
we cannot take its logarithm and we include the original variable in the model. Thus
we estimate
log yi = xi β + εi .
(3.9)
Of course, it is possible to include some explanatory variables in logs and some in
levels. In (3.9) the interpretation of a coefﬁcient βk is the relative change in yi due
to an absolute change of one unit in xik . So if xik is a dummy for males, βk is the
(ceteris paribus) relative wage differential between men and women. Again this holds
only approximately, see Subsection 3.5.2.
The inequality of E{log yi |xi } and log E{yi |xi } also has some consequences for prediction purposes. Suppose we start from the loglinear model (3.6) with E{vi | log xi } =
0. Then, we can determine the predicted value of log yi as (log xi ) γ . However, if we
are interested in predicting yi rather than log yi , it is not the case that exp{(log xi ) γ }
is a good predictor for yi in the sense that it corresponds to the expected value of yi ,
given xi . That is, E{yi |xi } = exp{E{log yi |xi }} = exp{(log xi ) γ }. The reason is that
taking logarithms is a nonlinear transformation, while the expected value of a nonlinear function is not this nonlinear function of the expected value. The only way to get
around this problem is to make distributional assumptions. If, for example, it can be

54

INTERPRETING AND COMPARING REGRESSION MODELS

assumed that vi in (3.6) is normally distributed with mean zero and variance σv2 , it
implies that the conditional distribution of yi is lognormal (see Appendix B) with mean




E{yi |xi } = exp E{log yi |xi } + 12 σv2 = exp (log xi ) γ + 12 σv2 .

(3.10)

Sometimes, the additional half-variance term is also added when the error terms are
not assumed to be normal. Often, it is simply omitted. Additional discussion on predicting yi when the dependent variable is log(yi ) is provided in Wooldridge (2003,
Section 6.4).
It should be noted that the assumption that E{εi |xi } = 0 is also important, as it says
that changing xi should not lead to changes in the expected error term. There are many
cases in economics where this is hard to maintain and the models we are interested
in do not correspond to conditional expectations. We shall come back to this issue in
Chapter 5.
Another consequence of (3.2) is often overlooked. If we change the set of explanatory
variables xi to zi , say, and estimate another regression model,
yi = zi γ + vi

(3.11)

with the interpretation that E{yi |zi } = zi γ , there is no conﬂict with the previous model
that said that E{yi |xi } = xi β. Because the conditioning variables are different, both
conditional expectations could be correct in the sense that both are linear in the conditioning variables. Consequently, if we interpret the regression models as describing
the conditional expectation given the variables that are included there can never be
any conﬂict between them. It is just two different things we might be interested in.
For example, we may be interested in the expected wage as a function of gender only,
but also in the expected wage as a function of gender, education and experience. Note
that, because of a different ceteris paribus condition, the coefﬁcients for gender in
these two models do not have the same interpretation. Often, researchers implicitly
or explicitly make the assumption that the set of conditioning variables is larger than
those that are included. Sometimes, it is suggested that the model contains all relevant
observable variables (implying that observables that are not included in the model are
in the conditioning set but irrelevant). If it would be argued, for example, that the two
linear models above should be interpreted as
E{yi |xi , zi } = zi γ
and

E{yi |xi , zi } = xi β,

respectively, then the two models are typically in conﬂict and at most one of them can
be correct.1 Only in such cases, it makes sense to compare the two models statistically
and to test, for example, which model is correct and which one is not. We come back
to this issue in Subsection 3.2.3.
1

We abstract from trivial exceptions, like xi = −zi and β = −γ .

SELECTING THE SET OF REGRESSORS

3.2

55

Selecting the Set of Regressors

3.2.1 Misspecifying the Set of Regressors

If one is (implicitly) assuming that the conditioning set of the model contains more
variables than the ones that are included, it is possible that the set of explanatory
variables is ‘misspeciﬁed’. This means that one or more of the omitted variables are
relevant, i.e. have nonzero coefﬁcients. This raises two questions: what happens when
a relevant variable is excluded from the model and what happens when an irrelevant
variable is included in the model? To illustrate this, consider the following two models
yi = xi β + zi γ + εi ,

(3.12)

yi = xi β + vi ,

(3.13)

and

both interpreted as describing the conditional expectation of yi given xi , zi (and may
be some additional variables). The model in (3.13) is nested in (3.12) and implicitly
assumes that zi is irrelevant (γ = 0). What happens if we estimate model (3.13) while
in fact model (3.12) is the correct model? That is, what happens when we omit zi from
the set of regressors?
The OLS estimator for β based on (3.13), denoted b2 , is given by

b2 =

N


−1
xi xi

i=1

N


xi yi .

(3.14)

i=1

The properties of this estimator under model (3.12) can be determined by substituting
(3.12) into (3.14) to obtain

b2 = β +

N

i=1

−1
xi xi

N

i=1


xi zi γ

+

N


−1
xi xi

i=1

N


xi εi .

(3.15)

i=1

Depending upon the assumptions made for model (3.12), the last term in this expression
will have an expectation or probability limit of zero.2 The second term on the right
hand side, however, corresponds to a bias (or asymptotic bias) in the OLS estimator
due to estimating the incorrect model (3.13). This is referred to as an omitted variable
bias. As expected, there will be no bias if γ = 0 (implying that the two models are
identical), but there
 is one more case in which the estimator for β will not be biased
and that is when N
i=1 xi zi = 0, or, asymptotically, when E{xi zi } = 0. If this happens
we say that xi and zi are orthogonal. This does not happen very often in economic
applications. Note, for example, that the presence of an intercept in xi implies that
E{zi } should be zero.
The converse is less of a problem. If we estimate model (3.12) while in fact model
(3.13) is appropriate, that is, we needlessly include the irrelevant variables zi , we would
simply be estimating the γ coefﬁcients, which are zero. In this case, however, it would
2

Compare the derivations of the properties of the OLS estimator in Chapter 2.

56

INTERPRETING AND COMPARING REGRESSION MODELS

be preferable to estimate β from the restricted model (3.13) rather than from (3.12)
because the latter estimator for β will usually have a higher variance and thus be less
reliable. While the derivation of this result requires some tedious matrix manipulations,
it is intuitively obvious: model (3.13) imposes more information, so that we can expect
that the estimator that exploits this information is, on average, more accurate than one
which does not. Thus, including irrelevant variables in your model, even though they
have a zero coefﬁcient, will typically increase the variance of the estimators for the
other model parameters. Including as many variables as possible in a model is thus not
a good strategy, while including too few variables has the danger of biased estimates.
This means we need some guidance on how to select the set of regressors.
3.2.2 Selecting Regressors

Again, it should be stressed that if we interpret the regression model as describing the
conditional expectation of yi given the included variables xi , there is no issue of a
misspeciﬁed set of regressors, although there might be a problem of functional form
(see the next section). This implies that statistically there is nothing to test here. The
set of xi variables will be chosen on the basis of what we ﬁnd interesting and often
economic theory or common sense guides us in our choice. Interpreting the model in a
broader sense implies that there may be relevant regressors that are excluded or irrelevant ones that are included. To ﬁnd potentially relevant variables we can use economic
theory again. For example, when specifying an individual wage equation we may use
the human capital theory which essentially says that everything that affects a person’s
productivity will affect his or her wage. In addition, we may use job characteristics
(blue or white collar, shift work, public or private sector, etc.) and general labour
market conditions (e.g. sectorial unemployment).
It is good practice to select the set of potentially relevant variables on the basis of
economic arguments rather than statistical ones. Although it is sometimes suggested
otherwise, statistical arguments are never certainty arguments. That is, there is always
a small (but not ignorable) probability of drawing the wrong conclusion. For example,
there is always a probability (corresponding to the size of the test) of rejecting the null
hypothesis that a coefﬁcient is zero, while the null is actually true. Such type I errors
are rather likely to happen if we use a sequence of many tests to select the regressors
to include in the model. This process is referred to as data snooping or data mining
(see Leamer, 1978; Lovell, 1983; or Charemza and Deadman, 1999, Chapter 2), and in
economics it is not a compliment if someone accuses you of doing it. In general, data
snooping refers to the fact that a given set of data is used more than once to choose a
model speciﬁcation and to test hypotheses. You can imagine, for example, that if you
have a set of 20 potential regressors and you try each one of them, that it is quite likely
to conclude that one of them is signiﬁcant, even though there is no true relationship
between any of these regressors and the variable you are explaining. Although statistical
software packages sometimes provide mechanical routines to select regressors, these
are not recommended in economic work. The probability of making incorrect choices
is high and it is not unlikely that your ‘model’ captures some peculiarities in the
data that have no real meaning outside the sample. In practice, however, it is hard
to avoid that some amount of data snooping enters your work. Even if you do not
perform your own speciﬁcation search and happen to ‘know’ which model to estimate,

SELECTING THE SET OF REGRESSORS

57

this ‘knowledge’ may be based upon the successes and failures of past investigations.
Nevertheless, it is important to be aware of the problem. In recent years, the possibility
of data snooping biases plays an important role in empirical studies that model stock
returns. Lo and MacKinlay (1990), for example, analyse such biases in tests of ﬁnancial
asset pricing models, while Sullivan, Timmermann and White (2001) analyse to what
extent the presence of calendar effects in stock returns, like the January effect discussed
in Section 2.7, can be attributed to data snooping.
The danger of data mining is particularly high if the speciﬁcation search is from
simple to general. In this approach, you start with a simple model and you include
additional variables or lags of variables until the speciﬁcation appears adequate. That
is, until the restrictions imposed by the model are no longer rejected and you are
happy with the signs of the coefﬁcient estimates and their signiﬁcance. Clearly, such
a procedure may involve a very large number of tests. An alternative is the generalto-speciﬁc modelling approach, advocated by Professor David Hendry and others,
typically referred to as the LSE methodology.3 This approach starts by estimating a
general unrestricted model (GUM), which is subsequently reduced in size and complexity by testing restrictions that can be imposed; see Charemza and Deadman (1999)
for an extensive treatment. The idea behind this approach is appealing. Assuming that
a sufﬁciently general and complicated model can describe reality, any more parsimonious model is an improvement if it conveys all of the same information in a simpler,
more compact form. The art of model speciﬁcation in the LSE approach is to ﬁnd
models that are valid restrictions of the GUM, and that cannot be reduced to even
more parsimonious models that are also valid restrictions. While the LSE methodology involves a large number of (mis)speciﬁcation tests, it can be argued to be relatively
insensitive to data-mining problems. The basic argument, formalized by White (1990),
is that as the sample size grows to inﬁnity only the true speciﬁcation will survive all
speciﬁcation tests. This assumes that the ‘true speciﬁcation’ is a special case of the
GUM that a researcher starts with. Rather than ending up with a speciﬁcation that is
most likely incorrect, due to an accumulation of type I and type II errors, the generalto-speciﬁc approach in the long run would result in the correct speciﬁcation. While
this asymptotic result is insufﬁcient to assure that the LSE approach works well with
sample sizes typical for empirical work, Hoover and Perez (1999) show that it may
work pretty well in practice in the sense that the methodology recovers the correct
speciﬁcation (or a closely related speciﬁcation) most of the time. An automated version of the general-to-speciﬁc approach is developed by Krolzig and Hendry (2001)
and available in PcGets (see Bårdsen, 2001, or Owen, 2003, for a review).
In practice, most applied researchers will start somewhere ‘in the middle’ with a
speciﬁcation that could be appropriate and, ideally, then test (1) whether restrictions
imposed by the model are correct and test (2) whether restrictions not imposed by the
model could be imposed. In the ﬁrst category are misspeciﬁcation tests for omitted
variables, but also for autocorrelation and heteroskedasticity (see Chapter 4). In the
second category are tests of parametric restrictions, for example that one or more
explanatory variables have zero coefﬁcients.
3

The adjective LSE derives from the fact that there is a strong tradition of time-series econometrics at the
London School of Economics (LSE), starting in the 1960s. Currently, the practitioners of LSE econometrics
are widely dispersed among institutions throughout the world.

58

INTERPRETING AND COMPARING REGRESSION MODELS

In presenting your estimation results, it is not a ‘sin’ to have insigniﬁcant variables
included in your speciﬁcation. The fact that your results do not show a signiﬁcant
effect on yi of some variable xik is informative to the reader and there is no reason
to hide it by re-estimating the model while excluding xik . Of course, you should be
careful including many variables in your model that are multicollinear so that, in the
end, almost none of the variables appears individually signiﬁcant.
Besides formal statistical tests there are other criteria that are sometimes used to
select a set of regressors. First of all, the R 2 , discussed in Section 2.4, measures the
proportion of the sample variation in yi that is explained by variation in xi . It is
clear that if we were to extend the model by including zi in the set of regressors, the
explained variation would never decrease, so that also the R 2 will never decrease if
we include additional variables in the model. Using the R 2 as criterion would thus
favour models with as many explanatory variables as possible. This is certainly not
optimal, because with too many variables we will not be able to say very much about
the model’s coefﬁcients, as they may be estimated rather inaccurately. Because the R 2
does not ‘punish’ the inclusion of many variables, one would better use a measure
which incorporates a trade-off between goodness-of-ﬁt and the number of regressors
employed in the model. One way to do this is to use the adjusted R 2 (or R̄ 2 ), as
discussed in the previous chapter. Writing it as
R̄ 2 = 1 −


2
1/(N − K) N
i=1 ei
N
1/(N − 1) i=1 (yi − ȳ)2

(3.16)

and noting that the denominator in this expression is unaffected by the model under
consideration, shows
the adjusted R 2 provides a trade-off between goodness-of-ﬁt,
N that
2
as measured by i=1 ei , and the simplicity or parsimony of the model, as measured by
the number of parameters K. There exist a number of alternative criteria that provide
such a trade-off, the most common ones being Akaike’s Information Criterion (AIC),
proposed by Akaike (1973), given by
N
1  2 2K
AIC = log
e +
N i=1 i
N

(3.17)

and the Schwarz Bayesian Information Criterion (BIC), proposed by Schwarz
(1978), which is given by
BIC = log

N
1  2 K
e + log N.
N i=1 i
N

(3.18)

Models with a lower AIC or BIC are typically preferred. Note that both criteria add a
penalty that increases with the number of regressors. Because the penalty is larger for
BIC, the latter criterion tends to favour more parsimonious models than AIC. The use
of either of these criteria is usually restricted to cases where alternative models are not
nested (see Subsection 3.2.3) and economic theory provides no guidance on selecting
the appropriate model. A typical situation is the search for a parsimonious model that
describes the dynamic process of a particular variable (see Chapter 8).

SELECTING THE SET OF REGRESSORS

59

Alternatively, it is possible to test whether the increase in R 2 is statistically signiﬁcant. Testing this is exactly the same as testing whether the coefﬁcients for the newly
added variables zi are all equal to zero, and we have seen a test for that in the previous
chapter. Recall from (2.59) that the appropriate F-statistic can be written as
f =

(R12 − R02 )/J
,
(1 − R12 )/(N − K)

(3.19)

where R12 and R02 denote the R 2 in the model with and without zi , respectively, and J
is the number of variables in zi . Under the null hypothesis that zi has zero coefﬁcients,
the f statistic has an F distribution with J and N − K degrees of freedom, provided we
can impose conditions (A1)–(A5) from Chapter 2. The F-test thus provides a statistical
answer to the question whether the increase in R 2 due to including zi in the model
was signiﬁcant or not. It is also possible to rewrite f in terms of adjusted R 2 s. This
would show that R̄12 > R̄02 if and only if f exceeds a certain threshold. In general, these
thresholds do not correspond to 5% or 10% critical values of the F distribution, but are
substantially smaller. In particular, it can be shown that R̄12 > R̄02 if and only if the f
statistic is larger than one. For a single variable (J = 1) this implies that the adjusted
R 2 will increase if the additional variable has a t-ratio with an absolute value larger
than unity. (Recall that for a single restriction t 2 = f .) This reveals that the adjusted
R 2 would lead to the inclusion of more variables than standard t or F-tests.
Direct tests of the hypothesis that the coefﬁcients γ for zi are zero can be obtained
from the t and F-tests discussed in the previous chapter. Compared to f above, a test
statistic can be derived which is more generally appropriate. Let γ̂ denote the OLS
estimator for γ and let V̂ {γ̂ } denote an estimated covariance matrix for γ̂ . Then, it
can be shown that under the null hypothesis that γ = 0 the test statistic
ξ = γ̂  V̂ {γ̂ }−1 γ̂

(3.20)

has an asymptotic χ 2 distribution with J degrees of freedom. This is similar to the
Wald test described in Chapter 2 (compare (2.63)). The form of the covariance matrix
of γ̂ depends upon the assumptions we are willing to make. Under the Gauss–Markov
assumptions we would obtain a statistic that satisﬁes ξ = Jf .
It is important to recall that two single tests are not equivalent to one joint test. For
example, if we are considering the exclusion of two single variables with coefﬁcients
γ1 and γ2 , the individual t-tests may reject neither γ1 = 0 nor γ2 = 0, whereas the
joint F-test (or Wald test) rejects the joint restriction γ1 = γ2 = 0. The message here
is that if we want to drop two variables from the model at the same time, we should
be looking at a joint test rather than at two separate tests. Once the ﬁrst variable is
omitted from the model, the second one may appear signiﬁcant. This is particularly of
importance if collinearity exists between the two variables.
3.2.3 Comparing Non-nested Models

Sometimes econometricians want to compare two different models that are not nested.
In this case neither of the two models is obtained as a special case of the other. Such

60

INTERPRETING AND COMPARING REGRESSION MODELS

a situation may arise if two alternative economic theories lead to different models for
the same phenomenon. Let us consider the following two alternative speciﬁcations:

and

Model A: yi = xi β + εi

(3.21)

Model B: yi = zi γ + vi ,

(3.22)

where both are interpreted as describing the conditional expectation of yi given xi
and zi . The two models are non-nested if zi includes a variable that is not in xi and
vice versa. Because both models are explaining the same endogenous variable, it is
possible to use the R̄ 2 , AIC or BIC criteria, discussed in the previous subsection. An
alternative and more formal idea that can be used to compare the two models is that of
encompassing (see Mizon, 1984; Mizon and Richard, 1986): if model A is believed
to be the correct model it must be able to encompass model B, that is, it must be
able to explain model B’s results. If model A is unable to do so, it has to be rejected.
Vice versa, if model B is unable to encompass model A, it should be rejected as well.
Consequently, it is possible that both models are rejected, because neither of them
is correct. If model A is not rejected, we can test it against another rival model and
maintain it as long as it is not rejected.
The encompassing principle is very general and it is legitimate to require a model
to encompass its rivals. If these rival models are nested within the current model, they
are automatically encompassed by it, because a more general model is always able to
explain results of simpler models (compare (3.15) above). If the models are not nested
encompassing is nontrivial. Unfortunately, encompassing tests for general models are
fairly complicated, but for the regression models above things are relatively simple.
We shall consider two alternative tests. The ﬁrst is the non-nested F -test or encom 
passing F-test. Writing xi = (x1i
x2i ) where x1i is included in zi (and x2i is not), model
B can be tested by constructing a so-called artiﬁcial nesting model as

yi = zi γ + x2i
δA + vi .

(3.23)

This model typically has no economic rationale, but reduces to model B if δA = 0.
Thus, the validity of model B (model B encompasses model A) can be tested using
an F-test for the restrictions δA = 0. In a similar fashion, we can test the validity of
model A by testing δB = 0 in

yi = xi β + z2i
δB + εi ,

(3.24)

where z2i contains the variables from zi that are not included in xi . The null hypotheses
that are tested here state that one model encompasses the other. The outcome of the
two tests may be that both models have to be rejected. On the other hand, it is also
possible that neither of the two models is rejected. Thus the fact that model A is
rejected should not be interpreted as evidence in favour of model B. It just indicates
that something is captured by model B which is not adequately taken into account in
model A.

SELECTING THE SET OF REGRESSORS

61

A more parsimonious non-nested test is the J-test. Let us start again from an artiﬁcial
nesting model that nests both model A and model B, given by
yi = (1 − δ)xi β + δzi γ + ui ,

(3.25)

where δ is a scalar parameter and ui denotes the error term. If δ = 0, equation (3.25)
corresponds to model A and if δ = 1 it reduces to model B. Unfortunately, the nesting
model (3.25) cannot be estimated because in general β, γ and δ cannot be separately
identiﬁed. One solution to this problem (suggested by Davidson and MacKinnon, 1981)
is to replace the unknown parameters γ by γ̂ , the OLS estimates from model B, and
to test the hypothesis that δ = 0 in
yi = xi β ∗ + δzi γ̂ + ui = xi β ∗ + δ ŷiB + ui ,

(3.26)

where ŷiB is the predicted value from model B and β ∗ = (1 − δ)β. The J-test for the
validity of model A uses the t-statistic for δ = 0 in this last regression. Computationally,
it simply means that the ﬁtted value from the rival model is added to the model that
we are testing and that we test whether its coefﬁcient is zero using a standard t-test.
Compared to the non-nested F-test, the J-test involves only one restriction. This means
that the J-test may be more attractive (have more power) if the number of additional
regressors in the non-nested F-test is large. If the non-nested F-test involves only one
additional regressor, it is equivalent to the J-test. More details on non-nested testing can
be found in Davidson and MacKinnon (1993, Section 11.3) and the references therein.
Another relevant case with two alternative models that are non-nested is the choice
between a linear and loglinear functional form. Because the dependent variable is
different (yi and log yi , respectively) a comparison on the basis of goodness-of-ﬁt
measures, including AIC and BIC, is inappropriate. One way to test the appropriateness of the linear and loglinear models involves nesting them in a more general
model using the so-called Box–Cox transformation (see Davidson and MacKinnon,
1993, Section 14.6), and comparing them against this more general alternative. Alternatively, an approach similar to the encompassing approach above can be chosen by
making use of an artiﬁcial nesting model. A very simple procedure is the PE test, suggested by MacKinnon, White and Davidson (1983). First, estimate both the linear and
loglinear models by OLS. Denote the predicted values by ŷi and log ỹi , respectively.
Then the linear model can be tested against its loglinear alternative by testing the null
hypothesis that δLIN = 0 in the test regression
yi = xi β + δLIN (log ŷi − log ỹi ) + ui .
Similarly, the loglinear model corresponds to the null hypothesis δLOG = 0 in
log yi = (log xi ) γ + δLOG (ŷi − exp{log ỹi }) + ui .
Both tests can simply be based on the standard t-statistics, which under the null hypothesis have an approximate standard normal distribution. If δLIN = 0 is not rejected, the
linear model may be preferred. If δLOG = 0 is not rejected, the loglinear model is
preferred. If both hypotheses are rejected, neither of the two models appears to be

INTERPRETING AND COMPARING REGRESSION MODELS

62

appropriate and a more general model should be considered, for example by generalizing the functional form of the xi variables in either the linear or the loglinear model.4
An empirical illustration using the PE test is provided in Section 3.4.

3.3

Misspecifying the Functional Form

Although the assumptions made in interpreting the models are fairly weak, there is one
important way in which the models may be misspeciﬁed and that is in their linearity.
The interpretation that E{yi |xi } = xi β implies that no other functions of xi are relevant
in explaining the expected value of yi . This is restrictive and the main motivation for
linear speciﬁcations is their convenience.
3.3.1 Nonlinear Models

Nonlinearities can arise in two different ways. In a ﬁrst case, the model is still linear in
the parameters but nonlinear in its explanatory variables. This means that we include
nonlinear functions of xi as additional explanatory variables, for example the variables
age 2i and age i male i could be included in an individual wage equation. The resulting
model is still linear in the parameters and can still be estimated by ordinary least
squares. In a second case, the model is nonlinear in its parameters and estimation is
less easy. In general, this means that E{yi |xi } = g(xi , β), where g(.) is a regression
function nonlinear in β. For example, for a single xi we could have that
β

(3.27)

g(xi , β) = β1 xi12 xi23 ,

(3.28)

g(xi , β) = β1 + β2 xi 3
or for a two-dimensional xi

β

β

which corresponds to a Cobb–Douglas production function with two inputs. As the
second function is linear in parameters after taking logarithms (assuming β1 > 0), it is
a common strategy in this case to model log yi rather than yi . This does not work for
the ﬁrst example.
Nonlinear models can also be estimated by a nonlinear version of the least squares
method, by minimizing the objective function
S(β̃) =

N


(yi − g(xi , β̃))2

(3.29)

i=1

with respect to β̃. This is called nonlinear least squares estimation. Unlike in the
linear case, it is generally not possible to analytically solve for the value of β̃ that
minimizes S(β̃) and we need to use numerical procedures to obtain the nonlinear least
4

It may be noted that with sufﬁciently general functional forms it is possible to obtain models for yi and
log yi that are both correct in the sense that they represent E{yi |xi } and E{log yi |xi }, respectively. It
is not possible, however, that both speciﬁcations have a homoskedastic error term (see the example in
Section 3.5).

MISSPECIFYING THE FUNCTIONAL FORM

63

squares estimator. A necessary condition for consistency is that there exists a unique
global minimum for S(β̃), which means that the model is identiﬁed. An excellent
treatment of such nonlinear models is given in Davidson and MacKinnon (1993) and
we will not pursue it here.
It is possible to rule out functional form misspeciﬁcations completely, by saying that
one is interested in the linear function of xi that approximates yi as well as possible.
This goes back to the initial interpretation of ordinary least squares as determining
the linear combination of x-variables that approximates a variable y as well as possible. We can do the same thing in a statistical setting by relaxing the assumption
that E{εi |xi } = 0 to E{εi xi } = 0. Recall that E{εi |xi } = 0 implies that E{εi g(xi )} = 0
for any function g (provided expectations exist), so that it is indeed a relaxation of
assumptions. In this case, we can interpret the linear regression model as describing
the best linear approximation of yi from xi . In many cases, we would interpret the
linear approximation as an estimate for its population equivalent rather than just an
in-sample result. Note that the condition E{εi xi } = 0 corresponds to condition (A7)
from Chapter 2 and is necessary for consistency of the OLS estimator.
3.3.2 Testing the Functional Form

A simple way to test the functional form of
E{yi |xi } = xi β

(3.30)

would be to test whether additional nonlinear terms in xi are signiﬁcant. This can be
done using standard t-tests, F-tests, or, more generally, Wald tests. This only works if
one can be speciﬁc about the alternative. If the number of variables in xi is large the
number of possible tests is also large.
Ramsey (1969) has suggested a test based upon the idea that under the null hypothesis
nonlinear functions of ŷi = xi b should not help explaining yi . In particular, he tests
whether powers of ŷi have nonzero coefﬁcients in the auxiliary regression
yi = xi β + α2 ŷi2 + α3 ŷi3 + · · · + αQ ŷiQ + vi .

(3.31)

An auxiliary regression, and we shall see several below, is typically used to compute
a test statistic only, and is not meant to represent a meaningful model. In this case
we can use a standard F-test for the Q − 1 restrictions in H0 : α2 = · · · = αQ = 0,
or a more general Wald test (with an asymptotic χ 2 distribution with Q − 1 degrees
of freedom). These tests are usually referred to as RESET tests (regression equation
speciﬁcation error tests). Often, a test is performed for Q = 2 only. It is not unlikely
that a RESET test rejects because of the omission of relevant variables from the model
(in the sense deﬁned earlier) rather than just a functional form misspeciﬁcation. That
is, the inclusion of an additional variable may capture the nonlinearities indicated by
the test.
3.3.3 Testing for a Structural Break

So far, we assumed that the functional form of the model was the same for all observations in the sample. As shown in Section 3.1, interacting dummy variables with other

64

INTERPRETING AND COMPARING REGRESSION MODELS

explanatory variables provides a useful tool to allow the marginal effects in the model
to be different across subsamples. Sometimes, it is interesting to consider an alternative
speciﬁcation in which all the coefﬁcients are different across two or more subsamples.
In a cross-sectional context, we can think of subsamples containing males and females
or married and unmarried workers. In a time-series application, the subsamples are
typically deﬁned by time. For example, the coefﬁcients in the model may be different
before and after a major change in macro-economic policy. In such cases, the change
in regression coefﬁcients is referred to as a structural break.
Let us consider an alternative speciﬁcation consisting of two groups, indicated by
gi = 0 and gi = 1, respectively. A convenient way to express the general speciﬁcation
is given by
(3.32)
yi = xi β + gi xi γ + εi ,
where the K-dimensional vector gi xi contains all explanatory variables (including the
intercept), interacted with the indicator variable gi . This equation says that the coefﬁcient vector for group 0 is β, whereas for group 1 it is β + γ . The null hypothesis is
γ = 0, in which case the model reduces to the restricted model.
A ﬁrst way to test γ = 0 is obtained by using the F-test from Subsection 2.5.4. Its
test statistic is given by
(S − SUR )/K
f = R
,
SUR /(N − 2K)
where K is the number of regressors in the restricted model (including the intercept)
and SUR and SR denote the residual sums of squares of the unrestricted and the restricted
model, respectively. Alternatively, the general unrestricted model can be estimated by
running a separate regression for each subsample. This leads to identical coefﬁcient
estimates as in (3.32), and consequently, the unrestricted residual sum of squares can be
obtained as SUR = S0 + S1 , where Sg denotes the residual sum of squares in subsample
g; see Section 3.5 for an illustration. The above F-test is typically referred to as the
Chow test for structural change (Chow, 1960).5 When using (3.32), it can easily be
adjusted to check for a break in a subset of the coefﬁcients by including only those
interactions in the general model. Note that the degrees of freedom of the test should
be adjusted accordingly.
Application of the Chow test is useful if one has some a priori idea that the regression
coefﬁcients may be different across two well-deﬁned subsamples. In a time-series
application, this requires a known break date, that is, a time period that indicates when
the structural change occurred. Sometimes there are good economic reasons to identify
the break dates, for example, the German uniﬁcation in 1990, or the end of the Bretton
Woods system of ﬁxed exchange rates in 1973. If the date of a possible break is not
known exactly, it is possible to adjust the Chow test by testing for all possible breaks
in a given time interval. While the test statistic is easily obtained as the maximum of all
F-statistics, its distribution is nonstandard; see Stock and Watson (2003, Section 12.7)
for additional discussion.
5

The above version of the Chow test assumes homoskedastic error terms under the null hypothesis. That
is, it assumes that the variance of εi is constant and does not vary across subsamples or with xi . A version
that allows for heteroskedasticity can be obtained by applying the Wald test to (3.32), combined with a
heteroskedasticity-robust covariance matrix; see Subsections 4.3.2 and 4.3.4 below.

ILLUSTRATION: EXPLAINING HOUSE PRICES

3.4

65

Illustration: Explaining House Prices

In this section we consider an empirical illustration concerning the relationship between
sale prices of houses and their characteristics. The resulting price function can be
referred to as a hedonic price function, because it allows the estimation of hedonic
prices (see Rosen, 1974). A hedonic price refers to the implicit price of a certain
attribute (e.g. the number of bedrooms) as revealed by the sale price of a house. In
this context, a house is considered as a bundle of such attributes. Typical products
for which hedonic price functions are estimated are computers, cars and houses. For
our purpose, the important conclusion is that a hedonic price function describes the
expected price (or log price) as a function of a number of characteristics. Berndt (1990,
Chapter 4) discusses additional economic and econometric issues relating to the use,
interpretation and estimation of such price functions.
The data we use6 are taken from a study by Anglin and Gençay (1996) and contain
sale prices of 546 houses, sold during July, August and September of 1987, in the city
of Windsor, Canada, along with their important features. The following characteristics
are available: the lot size of the property in square feet, the numbers of bedrooms, full
bathrooms and garage places, and the number of stories. In addition there are dummy
variables for the presence of a driveway, recreational room, full basement and central
air conditioning, for being located in a preferred area and for using gas for hot water
heating. To start our analysis, we shall ﬁrst estimate a model that explains the log of the
sale price from the log of the lot size, the numbers of bedrooms and bathrooms and the
presence of air conditioning. OLS estimation produces the results in Table 3.1. These
results indicate a reasonably high R 2 of 0.57 and fairly high t-ratios for all coefﬁcients.
The coefﬁcient for the air conditioning dummy indicates that a house that has central
air conditioning is expected to sell at a 21% higher price than a house without it, both
houses having the same number of bedrooms and bathrooms and the same lot size. A
10% larger lot, ceteris paribus, increases the expected sale price by about 4%, while
an additional bedroom is estimated to raise the price by almost 8%. The expected log
sale price of a house with four bedrooms, one full bathroom, a lot size of 5000 sq. ft
and no air conditioning can be computed as
7.094 + 0.400 log(5000) + 0.078 × 4 + 0.216 = 11.028,
which corresponds to an expected price of exp{11.028 + 0.5 × 0.24562 } = 63 460
Canadian dollars. The latter term in this expression corresponds to one half of the
estimated error variance (s 2 ) and is based upon the assumption that the error term is
normally distributed (see (3.10)). Omitting this term produces an expected price of only
61 575 dollars. To appreciate the half-variance term, consider the ﬁtted values of our
model. Taking the exponential of these ﬁtted values produces predicted prices for the
houses in our sample. The average predicted price is 66 679 dollars, while the sample
average of actual prices is 68 122. This indicates that without any corrections we would
systematically underpredict prices. When the half-variance term is added, the average
predicted price based on the model explaining log prices increases to 68 190, which is
fairly close to the actual average.
6

The data are available as HOUSING.

INTERPRETING AND COMPARING REGRESSION MODELS

66

Table 3.1 OLS results hedonic price function
Dependent variable: log(price)
Variable
constant
log(lot size)
bedrooms
bathrooms
air conditioning

Estimate

Standard error

t-ratio

7.094
0.400
0.078
0.216
0.212

0.232
0.028
0.015
0.023
0.024

30.636
14.397
5.017
9.386
8.923

s = 0.2456 R = 0.5674 R̄ 2 = 0.5642 F = 177.41

Table 3.2

OLS results hedonic price function, extended model

Dependent variable: log(price)
Variable
constant
log(lot size)
bedrooms
bathrooms
air conditioning
driveway
recreational room
full basement
gas for hot water
garage places
preferred area
stories

Estimate

Standard error

t-ratio

7.745
0.303
0.034
0.166
0.166
0.110
0.058
0.104
0.179
0.048
0.132
0.092

0.216
0.027
0.014
0.020
0.021
0.028
0.026
0.022
0.044
0.011
0.023
0.013

35.801
11.356
2.410
8.154
7.799
3.904
2.225
4.817
4.079
4.178
5.816
7.268

s = 0.2104 R 2 = 0.6865 R̄ 2 = 0.6801 F = 106.33

To test the functional form of this simple speciﬁcation, we can use the RESET test.
This means that we generate predicted values from our model, take powers of them,
include them in the original equation and test their signiﬁcance. Note that these latter
regressions are run for testing purposes only and are not meant to produce a meaningful
model. Including the squared ﬁtted value produces a t-statistic of 0.514 (p = 0.61), and
including the squared and cubed ﬁtted values gives an F-statistic of 0.56 (p = 0.57).
Both tests do not indicate particular misspeciﬁcations of our model. Nevertheless, we
may be interested in including additional variables in our model because prices may
also be affected by characteristics like the number of garage places or the location
of the house. To this end, we include all other variables in our model to obtain the
speciﬁcation that is reported in Table 3.2. Given that the R 2 increases to 0.68 and
that all the individual t-statistics are larger than 2 this extended speciﬁcation appears
to perform signiﬁcantly better in explaining house prices than the previous one. A
joint test on the hypothesis that all seven additional variables have a zero coefﬁcient is
provided by the F-test, where the test statistic is computed on the basis of the respective
R 2 s as
(0.6865 − 0.5674)/7
= 28.99,
f =
(1 − 0.6865)/(546 − 12)

ILLUSTRATION: EXPLAINING HOUSE PRICES

67

which is highly signiﬁcant for an F distribution with 7 and 532 degrees of freedom
(p = 0.000). Looking at the point estimates, the ceteris paribus effect of a 10% larger
lot size is now estimated to be only 3%. This is almost certainly due to the change in
ceteris paribus condition, for example because houses with larger lot sizes tend to have
a driveway relatively more often.7 Similarly, the estimated impact of the other variables
is reduced compared to the estimates in Table 3.1. As expected, all coefﬁcient estimates
are positive and relatively straightforward to interpret. Ceteris paribus, a house in a
preferred neighbourhood of the city is expected to sell at a 13% higher price than a
house located elsewhere.
As before we can test the functional form of the speciﬁcation by performing one
or more RESET tests. With a t-value of 0.06 for the squared ﬁtted values and an
F-statistic of 0.04 for the squared and cubed terms, there is again no evidence of misspeciﬁcation of the functional form. It is possible though to consider more speciﬁc
alternatives when testing the functional form. For example, one could hypothesize
that an additional bedroom implies a larger price increase when the house is in a
preferred neighbourhood. If this is the case, the model should include an interaction term between the location dummy and the number of bedrooms. If the model is
extended to include this interaction term, the t-test on the new variable produces a
highly insigniﬁcant value of −0.131. Overall, the current model appears surprisingly
well speciﬁed.
The model allows us to compute the expected log sale price of an arbitrary house
in Windsor. If you would own a 2-story house on a lot of 10 000 square feet, located
in a preferred neighbourhood of the city, with 4 bedrooms, 1 bathroom, 2 garage
places, a driveway, recreational room, air conditioning and a full and ﬁnished basement,
using gas for water heating, the expected log price is 11.87. This indicates that the
hypothetical price of your house, if sold in the summer of 1987, is estimated to be
slightly more than 146 000 Canadian dollars.
Instead of modelling log prices, we could also consider explaining prices. Table 3.3
reports the results of a regression model where prices are explained as a linear function
of lot size and all other variables. Compared to the previous model the coefﬁcients
now reﬂect absolute differences in prices rather than relative differences. For example,
the presence of a driveway (ceteris paribus) is expected to increase the house price
by 6688 dollars, while Table 3.2 implies an estimated increase of 11%. It is not
directly clear from a comparison of the results in Tables 3.2 and 3.3 which of the
two speciﬁcations is preferable. Recall that the R 2 does not provide an appropriate
means of comparison. As discussed in Subsection 3.2.3, it is possible to test these
two non-nested models against each other. Using the PE test we can test the two
hypotheses that the linear model is appropriate and that the loglinear model is appropriate. When testing the linear model we obtain a test statistic of −6.196. Given the
critical values of a standard normal distribution, this implies that the speciﬁcation in
Table 3.3 has to be rejected. This does not automatically imply that the speciﬁcation
in Table 3.2 is appropriate. Nevertheless, when testing the loglinear model (where
only price and lot size are in logs) we ﬁnd a test statistic of −0.569, so that it is
not rejected.

7

The sample correlation coefﬁcient between log lot size and the driveway dummy is 0.33.

INTERPRETING AND COMPARING REGRESSION MODELS

68

Table 3.3

OLS results hedonic price function, linear model

Dependent variable: price
Variable
constant
lot size
bedrooms
bathrooms
air conditioning
driveway
recreational room
full basement
gas for hot water
garage places
preferred area
stories

Estimate

Standard error

t-ratio

−4038.35
3.546
1832.00
14335.56
12632.89
6687.78
4511.28
5452.39
12831.41
4244.83
9369.51
6556.95

3409.47
0.350
1047.00
1489.92
1555.02
2045.25
1899.96
1588.02
3217.60
840.54
1669.09
925.29

−1.184
10.124
1.750
9.622
8.124
3.270
2.374
3.433
3.988
5.050
5.614
7.086

s = 15423 R 2 = 0.6731 R̄ 2 = 0.6664 F = 99.97

3.5

Illustration: Explaining Individual Wages

It is a well-known fact that the average hourly wage rates of males are higher than
those of females for almost all industrialized countries. In this section, we analyse this
phenomenon for Belgium. In particular, we want to ﬁnd out whether factors such as
education level and experience can explain the wage differential. For this purpose we
use a data set consisting of 1472 individuals, randomly sampled from the working
population in Belgium for the year 1994. The data set, taken from the Belgian part
of the European Community Household Panel, contains 893 males and 579 females.8
The analysis is based on the following four variables:
wage
male
educ

exper

before tax hourly wage rate, in euro per hour
1 if male, 0 if female
education level, 1 = primary school,
2 = lower vocational training, 3 = intermediate level,
4 = higher vocational training, 5 = university level
experience in years

Some summary statistics of these variables are given in Table 3.4. We see, for
example, that the average wage rate for men is ¤11.56 per hour, while for women it
is only ¤10.26 per hour, which corresponds to a difference of ¤1.30 or almost 13%.
Because the average years of experience in the sample is lower for women than for
men, this does not necessarily imply wage discrimination against women.
3.5.1 Linear Models

A ﬁrst model to estimate the effect of gender on the hourly wage rate, correcting for
differences in experience and education level, is obtained by regressing wage upon
8

The data for this illustration are available as BWAGES.

ILLUSTRATION: EXPLAINING INDIVIDUAL WAGES

69

Table 3.4 Summary statistics, 1472 individuals
Mean
wage
educ
exper

Males
Standard dev.

11.56
3.24
18.52

4.75
1.26
10.25

Table 3.5

Mean

Females
Standard dev.

10.26
3.59
15.20

3.81
1.09
9.70

OLS results speciﬁcation 1

Dependent variable: wage
Variable

Estimate

Standard error

t-ratio

constant
male
educ
exper

0.214
1.346
1.986
0.192

0.387
0.193
0.081
0.010

0.552
6.984
24.629
20.064

s = 3.55 R 2 = 0.3656 R̄ 2 = 0.3643 F = 281.98

male, exper and educ, the results of which are given in Table 3.5. If we interpret this
model as describing the expected wage given gender, experience and education level,
the ceteris paribus effect of gender is virtually identical to the average wage differential. Apparently, adjusting for differences in education and experience does not change
the expected wage differential between males and females. Note that the difference is
statistically highly signiﬁcant, with a t-ratio of 6.984. As expected, the effect of experience, keeping the education level ﬁxed, is positive: an additional year of experience
increases the expected wage by somewhat more than ¤0.19 per hour. Similarly, higher
education levels substantially increase the expected wage. If we compare two people
with two adjacent education levels but of the same gender and having the same experience, the expected wage differential is approximately ¤1.99 per hour. Given the high
t-ratios, both the effects of exper and educ are statistically highly signiﬁcant. The R 2
of the estimated model is 0.3656, which implies that more than 36% of the variation
in individual wages can be attributed (linearly) to differences in gender, experience
and education.
It could be argued that experience affects a person’s wage nonlinearly: after many
years of experience, the effect of an additional year on one’s wage may become
increasingly smaller. To model this, we can include the square of experience in the
model, which we expect to have a negative coefﬁcient. The results of this are given in
Table 3.6. The additional variable exper 2 has a coefﬁcient that is estimated to be negative, as expected. With a t-ratio of −5.487 we can safely reject the null hypothesis that
squared experience has a zero coefﬁcient and we can conclude that including exper 2
signiﬁcantly improves the model. Note that the adjusted R 2 has increased from 0.3643
to 0.3766. Given the presence of both experience and its square in the speciﬁcation, we
cannot interpret their coefﬁcients in isolation. One way to describe the effect of experience is to say that the expected wage difference through a marginal increase of experience is, ceteris paribus, given by (differentiate with respect to experience as in (3.4)):
0.358 − 0.0044 × 2 × exper i

INTERPRETING AND COMPARING REGRESSION MODELS

70

Table 3.6

OLS results speciﬁcation 2

Dependent variable: wage
Variable
constant
male
educ
exper
exper 2

Estimate
−0.892
1.334
1.988
0.358
−0.0044

Standard error
0.433
0.191
0.080
0.032
0.0008

t-ratio
−2.062
6.988
24.897
11.309
−5.487

s = 3.51 R 2 = 0.3783 R̄ 2 = 0.3766 F = 223.20

which shows that the effect of experience differs with its level. Initially, it is as big
as ¤0.36 per hour, but it reduces to less than ¤0.10 for a person with 30 years of
experience. Alternatively, we can simply compare predicted wages for a person with,
say, 30 years of experience and one with 31 years. The estimated wage difference is
then given by
0.358 − 0.0044(312 − 302 ) = 0.091
which produces a slightly lower estimate. The difference is caused by the fact that
the ﬁrst number is based on the effect of a ‘marginal’ change in experience (it is a
derivative), while an increase of 1 year is not really marginal.
Before continuing our statistical analysis, it is important to analyse to what extent
the assumptions regarding the error terms are satisﬁed in this example. Recall that for
the standard errors and statistical tests to be valid, we need to exclude both autocorrelation and heteroskedasticity. Given that there is no natural ordering in the data and
individuals are randomly sampled, autocorrelation is not an issue, but heteroskedasticity could be problematic. While we shall see some formal tests for heteroskedasticity
in Chapter 4, a quick way to get some insight into the likelihood of the failure of the
homoskedasticity assumption is to make a graph of the residuals of the model against
the predicted values. If there is no heteroskedasticity, we can expect that the dispersion
of residuals does not vary with different levels of the ﬁtted values. For the model in
Table 3.6, we present such a graph in Figure 3.1.
Figure 3.1 clearly shows an increased variation in the residuals for higher ﬁtted
values and thus casts serious doubt on the assumption of homoskedasticity. This implies
that the routinely computed standard errors and corresponding t-tests are not appropriate.
One way to eliminate or reduce the heteroskedasticity problem is provided by changing the functional form and use log wages rather than wages as the explanatory variable.
Why this may help solving the problem can be seen as follows. Let us denote the current
model as
wi = g(xi ) + εi ,
(3.33)
where g(xi ) is a function of xi that predicts the wage wi (e.g. xi β) and εi is an error
term that has mean zero (conditional upon xi ). This is an additive model in the sense
that the error term is added to the predicted value. It is also possible to consider a

ILLUSTRATION: EXPLAINING INDIVIDUAL WAGES

71

40
30

Residual

20
10
0
−10
−20
0

5

10

15

20

Fitted value

Figure 3.1

Residuals versus ﬁtted values, linear model

multiplicative model of the form
wi = g(xi ) exp{ηi },

(3.34)

where ηi is an error term that has mean zero (conditional upon xi ). It is easily veriﬁed
that the two models are equivalent if g(xi )[exp{ηi } − 1] = εi . If ηi is homoskedastic,
it is clear that εi is heteroskedastic with a variance that depends upon g(xi ). If we thus
ﬁnd heteroskedasticity in the additive model, it could be the case that a multiplicative
model is appropriate with a homoskedastic error term. The multiplicative model can
easily be written as an additive model, with an additive error term, by taking logarithms.
This gives
log wi = log g(xi ) + ηi = f (xi ) + ηi .
(3.35)
In our case g(xi ) = xi β. Estimation of (3.35) becomes simple if we assume that the
function f is such that log g(xi ) is a linear function of the parameters. Typically, this
involves the inclusion of logs of the x-variables (excluding dummy variables), so that
we obtain a loglinear model (compare (3.6)).
3.5.2 Loglinear Models

For our next speciﬁcation, we estimate a loglinear model that explains the log of the
hourly wage rate from gender, the log of experience, the squared log of experience
and the log of education. (Note that the log of experience squared is perfectly collinear
with the log of experience.) This gives the results in Table 3.7. Because the endogenous
variable is different, the R 2 is not really comparable to those for the models that explain
the hourly wage rate, but it happens to be almost the same. The interpretation of the
coefﬁcient estimates is also different from before. The coefﬁcient of male now measures
the relative difference in expected wages for males and females. In particular, the
ceteris paribus difference of the expected log wage between men and women is 0.118.

INTERPRETING AND COMPARING REGRESSION MODELS

72

Table 3.7

OLS results speciﬁcation 3

Dependent variable: log(wage)
Variable
constant
male
log(educ)
log(exper)
log2 (exper)

Estimate

Standard error

t-ratio

1.263
0.118
0.442
0.110
0.026

0.066
0.016
0.018
0.054
0.011

19.033
7.574
24.306
2.019
2.266

s = 0.286 R 2 = 0.3783 R̄ 2 = 0.3766 F = 223.13 S = 120.20

If a woman is expected to earn an amount w ∗ , a comparable man is expected to earn
exp{log w ∗ + 0.118} = w ∗ exp{0.118} = w ∗ 1.125, which corresponds to a difference
of approximately 12%. Because exp(a) ≈ 1 + a if a is close to zero, it is common in
loglinear models to make the direct transformation from the estimated coefﬁcients to
percentage changes. Thus a coefﬁcient of 0.118 for males is interpreted as an expected
wage differential of approximately 11.8%.
Before continuing, let us consider the issue of heteroskedasticity again. A plot of
the residuals of the loglinear model against the predicted log wages is provided in
Figure 3.2. While there appear to be some traces of heteroskedasticity still, the graph
is much less pronounced than for the additive model. Therefore, we shall continue to
work with speciﬁcations that explain log wages rather than wages and, where needed,
assume that the errors are homoskedastic. In particular, we shall assume that standard
errors and routinely computed t and F-tests are appropriate. Chapter 4 provides some
additional discussion on tests for heteroskedasticity and how it can be handled.
The coefﬁcients for log experience and its square are somewhat hard to interpret.
If log2 (exper) were excluded, the estimated coefﬁcient for log(exper) would simply

1

Residual

0

−1

−2
5

5.5

6

6.5

Fitted value

Figure 3.2

Residuals against ﬁtted values, loglinear model

ILLUSTRATION: EXPLAINING INDIVIDUAL WAGES

73

imply an expected wage increase of approximately 0.11% for an experience increase
of 1%. In the current case, we can estimate the elasticity as
0.110 + 2 × 0.026 log(exper).
It is surprising to see that this elasticity is increasing with experience. This, however,
is not in conﬂict with our earlier ﬁnding that suggested that the effect of experience is
positive but decreasing with its level. The effects of log(exper) and log2 (exper) are,
individually, marginally signiﬁcant at the 5% level but insigniﬁcant at the 1% level.
(Note that given the large number of observations a size of 1% may be considered
more appropriate.) This does not necessarily mean that experience has no signiﬁcant
effect upon wages. To that end, we need to consider a joint test for the two restrictions.
The test statistic can be computed from the R 2 s of the above model and a restricted
model that excludes both log(exper) and log2 (exper). This restricted model has an R 2
of only 0.1798, such that an F-statistic can be computed as
f =

(0.3783 − 0.1798)/2
= 234.2,
(1 − 0.3783)/(1472 − 5)

(3.36)

which indicates a remarkably strong rejection. Because the two variables that involve
experience are individually insigniﬁcant at the 1% level, we could consider dropping
one of them. If we drop log2 (exper), we obtain the results in Table 3.8, which show
that the resulting model only has a slightly worse ﬁt.
Let us consider this reduced speciﬁcation in more detail. Because the effect of
education is restricted to be linear in the log of the education level, the ceteris paribus
difference in expected log wages between two persons with education levels educ1 and
educ2, respectively, is 0.437(log(educ1) − log(educ2)). So compared to the lowest
education level 1, the effects of levels 2–5 are estimated as 0.30, 0.48, 0.61 and 0.70,
respectively. It is also possible to unrestrictedly estimate these four effects by including
four dummy variables corresponding to the four higher education levels. The results
of this are provided in Table 3.9. Note that with ﬁve educational levels, the inclusion
of four dummies is sufﬁcient to capture all effects. By including ﬁve dummies, we
would fall into the so-called dummy variable trap, and exact multicollinearity would
arise. Which of the ﬁve dummy variables is excluded is immaterial; it only matters for
the economic interpretation of the other dummies’ coefﬁcients. The omitted category
acts as reference group and all effects are relative to this group. In this example, the
reference category has education level one.
Table 3.8

OLS results speciﬁcation 4

Dependent variable: log(wage)
Variable
constant
male
log(educ)
log(exper)

Estimate

Standard error

t-ratio

1.145
0.120
0.437
0.231

0.041
0.016
0.018
0.011

27.798
7.715
24.188
21.488

s = 0.287 R 2 = 0.3761 R̄ 2 = 0.3748 F = 294.96 S = 120.63

INTERPRETING AND COMPARING REGRESSION MODELS

74

Table 3.9

OLS results speciﬁcation 5

Dependent variable: log(wage)
Variable
constant
male
educ = 2
educ = 3
educ = 4
educ = 5
log(exper)

Estimate

Standard error

t-ratio

1.272
0.118
0.114
0.305
0.474
0.639
0.230

0.045
0.015
0.033
0.032
0.033
0.033
0.011

28.369
7.610
4.306
9.521
14.366
19.237
21.804

s = 0.282 R 2 = 0.3976 R̄ 2 = 0.3951 F = 161.14 S = 116.47

Looking at the results in Table 3.9, we see that each of the four dummy variables
is individually highly signiﬁcant, with coefﬁcients that deviate somewhat from the
effects estimated on the basis of the restricted model. In fact, the previous model is
nested within the current model and imposes three restrictions. While it is somewhat
complicated to determine analytical expressions for these three restrictions, we can
easily test them using the R 2 version of the F-test. This gives
f =

(0.3976 − 0.3761)/3
= 17.358.
(1 − 0.3976)/(1472 − 7)

(3.37)

As the 1% critical value for an F distribution with 3 and 1465 degrees of freedom
is given by 3.78, the null hypothesis has to be rejected. That is, speciﬁcation 5 with
educational dummies is a signiﬁcant improvement over speciﬁcation 4 with the log
education level.
3.5.3 The Effects of Gender

Until now the effect of gender was assumed to be constant, irrespective of a person’s
experience or education level. As it is possible, for example, that men are differently
rewarded than women for having more education, this may be restrictive. It is possible
to allow for such differences by interacting each of the explanatory variables with the
gender dummy. One way to do so is to include the original regressor variables as well
as the regressors multiplied by male. This way the coefﬁcients for the latter set of
variables measure to what extent the effect is different for males.
Including interactions for all ﬁve variables produces the results in Table 3.10. This
is the unrestricted speciﬁcation used in the Chow test, discussed in Subsection 3.3.3.
An exactly equivalent set of results would have been obtained if we estimate the model
separately for the two subsamples of males and females. The only advantage of estimating over the subsamples is the fact that in computing the standard errors it is assumed
that the error terms are homoskedastic within each subsample, while the pooled model
in Table 3.10 imposes homoskedasticity over the entire sample. This explains why
estimated standard errors will be different, a large difference corresponding to strong
heteroskedasticity. The coefﬁcient estimates are exactly identical. This follows directly
from the deﬁnition of the OLS estimator: minimizing the sum of squared residuals

ILLUSTRATION: EXPLAINING INDIVIDUAL WAGES

75

Table 3.10 OLS results speciﬁcation 6
Dependent variable: log(wage)
Variable

Estimate

Standard error

t-ratio

constant
male
educ = 2
educ = 3
educ = 4
educ = 5
log(exper)
educ = 2 × male
educ = 3 × male
educ = 4 × male
educ = 5 × male
log(exper) × male

1.216
0.154
0.224
0.433
0.602
0.755
0.207
−0.097
−0.167
−0.172
−0.146
0.041

0.078
0.095
0.068
0.063
0.063
0.065
0.017
0.078
0.073
0.074
0.076
0.021

15.653
1.615
3.316
6.851
9.585
11.673
12.535
−1.242
−2.272
−2.317
−1.935
1.891

s = 0.281 R 2 = 0.4032 R̄ 2 = 0.3988 F = 89.69 S = 115.37

with different coefﬁcients for two subsamples is exactly equivalent to minimizing for
each subsample separately.
The results in Table 3.10 do not indicate important signiﬁcant differences between
men and women in the effect of experience. There are some indications, however,
that the effect of education is lower for men than for women, as two of the four
education dummies interacted with male are signiﬁcant at the 5% level, though not at
the 1% level. Note that the coefﬁcient for male no longer reﬂects the gender effect, as
the other variables are a function of gender as well. The estimated wage differential
between a male and female of, say, 20 years of experience and education level 2 can
be computed as
0.154 + 0.041 log(20) − 0.097 = 0.180,
corresponding to somewhat more than 18%. To statistically test the joint hypothesis
that each of the ﬁve coefﬁcients of the variables interacted with male are zero, we can
easily compute an F-test from the R 2 s in Tables 3.10 and 3.9. This is equivalent to
the Chow test for a structural break (between the two subsamples deﬁned by gender).
This results in
(0.4032 − 0.3976)/5
f =
= 2.7399,
(1 − 0.4032)/(1472 − 12)
which does not exceed the 1% critical value of 3.01, but does reject at the 5% level.
As a more general speciﬁcation test, we can perform Ramsey’s RESET test. Including
the square of the ﬁtted value to the speciﬁcation in Table 3.10, produces a t-statistic
of 3.989, which implies rejection at both the 5% and 1% level.
A ﬁnal speciﬁcation that we explore involves interaction terms between experience
and education, which allows the effect of education to be different across education
levels and at the same time allows the effects of different education levels to vary with
experience. To do so, we interacted log(exper) with each of the four education dummies.
The results are reported in Table 3.11. The coefﬁcient for log(exper) interacted with
educ = 2 measures to what extent the effect of experience is different for education
level 2 in comparison with the reference category, being education level 1. The results

INTERPRETING AND COMPARING REGRESSION MODELS

76

Table 3.11 OLS results speciﬁcation 7
Dependent variable: log(wage)
Variable
constant
male
educ = 2
educ = 3
educ = 4
educ = 5
log(exper)
log(exper) × educ
log(exper) × educ
log(exper) × educ
log(exper) × educ

=2
=3
=4
=5

Estimate

Standard error

t-ratio

1.489
0.116
0.067
0.135
0.205
0.341
0.163
0.019
0.050
0.088
0.100

0.212
0.015
0.226
0.219
0.219
0.218
0.065
0.070
0.068
0.069
0.068

7.022
7.493
0.297
0.618
0.934
1.565
2.494
0.274
0.731
1.277
1.465

s = 0.281 R 2 = 0.4012 R̄ 2 = 0.3971 F = 97.90 S = 115.76

do not indicate any important interaction effects between experience and education.
Individually, each of the four coefﬁcients does not differ signiﬁcantly from zero, and
jointly, the F-test produces the insigniﬁcant value of 2.196.
Apparently, this last speciﬁcation suffers from multicollinearity. Almost none of the
individual coefﬁcients is signiﬁcant, while the R 2 is reasonably large. Note that a joint
test on all coefﬁcients, except the intercept, being zero produces the highly signiﬁcant value of 97.90. Finally, we perform a RESET test (with Q = 2) on this model,
which produces a t-value of 2.13, which is insigniﬁcant at the 1% level. Nevertheless,
speciﬁcation 6 in Table 3.10 seems more appropriate than the current one.
3.5.4 Some Words of Warning

Despite our relatively careful statistical analysis, we still have to be cautious in interpreting the resulting estimates economically. The educational level, for example, will
to a large extent capture differences in the type of jobs people are employed in. That is,
the effect of education, as measured by the models’ coefﬁcients, will typically operate
through a person’s job characteristics. Thus the educational effect cannot be interpreted
to hold for people that have the same job, besides having the same experience and gender. Of course, this is a direct consequence of not including ‘job type’ in the model,
such that it is not captured by our ceteris paribus condition.
Another issue is that the model is only estimated for the subpopulation of working
males and females. There is no reason why it would be valid to extend the estimation
results to also explain wages of non-workers that consider entering the labour market. It
may well be the case that selection into the labour market is nonrandom and depends
upon potential wages, which would lead to a so-called selection bias in the OLS
estimator. To take this into account, it is possible to model wages jointly with the
decision to join the labour market and we shall discuss a class of models for such
problems in Chapter 7.
We should also be careful of interpreting the coefﬁcient for education as measuring
the causal effect. That is, if we would increase the education level of an arbitrary
person in the sample, the expected effect upon his or her wage may not correspond

EXERCISES

77

to the estimated coefﬁcient. The reason is that education is typically correlated with
unobserved characteristics (intelligence, ability) that also determine a person’s wage.
In this sense, the effect of education as estimated by OLS is partly due to differences in
unobserved characteristics of people attaining the different education levels. Chapter 5
comes back to this problem.

Exercises
Exercise 3.1 (Speciﬁcation Issues)

a. Explain what is meant by ‘data mining’.
b. Explain why it is inappropriate to drop two variables from the model at the same
time on the basis of their t-ratios only.
c. Explain the usefulness of the R̄ 2 , AIC and BIC criteria to compare two models
that are nested.
d. Consider two non-nested regression models explaining the same variable yi . How
can you test one against the other?
e. Explain why a functional form test (like Ramsey’s RESET test) may indicate an
omitted variable problem.
Exercise 3.2 (Regression – Empirical)

In the data set CLOTHING information is contained on sales, size and other characteristics of 400 Dutch men’s fashion stores. The goal is to explain sales per square metre
(sales) from the characteristics of the shop (number of owners, full-time and part-time
workers, number of hours worked, shop size, etc.).
a. Estimate a linear model (model A) that explains sales from total number of hours
worked (hoursw), shop size in square metres (ssize) and a constant. Interpret
the results.
b. Perform Ramsey’s RESET test with Q = 2.
c. Test whether the number of owners (nown) affects shop sales, conditional upon
hoursw and ssize.
d. Also test whether the inclusion of the number of part-time workers (npart) improves
the model.
e. Estimate a linear model (model B) that explains sales from the number of owners,
full-time workers (nfull), part-time workers and shop size. Interpret the results.
f. Compare model A and model B on the basis of R̄ 2 , AIC and BIC.
g. Perform a non-nested F-test of model A against model B. Perform a non-nested
F-test of model B against model A. What do you conclude?
h. Repeat the above test using the J-test. Does your conclusion change?
i. Include the numbers of full-time and of part-time workers in model A to obtain
model C. Estimate this model. Interpret the results and perform a RESET test. Are
you satisﬁed with this speciﬁcation?

78

INTERPRETING AND COMPARING REGRESSION MODELS

Exercise 3.3 (Regression – Empirical)

The data set HOUSING contains the data of the models estimated in Section 3.4.
a.

b.
c.
d.
e.

Create four dummy variables relating to the number of bedrooms, corresponding
to 2 or less, 3, 4, and 5 or more. Estimate a model for log prices that includes log
lot size, the number of bathrooms, the air conditioning dummy and three of these
dummies. Interpret the results.
Why is the model under a not nested in the speciﬁcation that is reported in
Table 3.1?
Perform two non-nested F-tests to test these two speciﬁcations against each other.
What do you conclude?
Include all four dummies in the model and re-estimate it. What happens? Why?
Suppose that lot size would be measured in square metres rather than square feet.
How would this affect the estimation results in Table 3.2? Pay attention to the
coefﬁcient estimates, the standard errors and the R 2 . How would the results in
Table 3.3 be affected by this? Note: the conversion is 1 m2 = 10.76 ft2 .

4

Heteroskedasticity and
Autocorrelation

In many empirical cases, the Gauss–Markov conditions (A1)–(A4) from Chapter 2 will
not all be satisﬁed. As we have seen in Subsection 2.6.1, this is not necessarily fatal
for the OLS estimator in the sense that it is consistent under fairly weak conditions. In
this chapter we will discuss the consequences of heteroskedasticity and autocorrelation,
which imply that the error terms in the model are no longer independently and identically distributed. In such cases, the OLS estimator may still be unbiased or consistent,
but its covariance matrix is different from that derived in Chapter 2. Moreover, the
OLS estimator may be relatively inefﬁcient and no longer have the BLUE property.
In Section 4.1, we discuss the general consequences for the OLS estimator of an error
covariance matrix that is not a constant times the identity matrix, while Section 4.2
presents, in a general matrix notation, an alternative estimator that is best linear
unbiased in this more general case. Heteroskedasticity is treated in Sections 4.3–4.5,
while the remaining sections of this chapter are devoted to autocorrelation. Examples of
heteroskedasticity and its consequences are discussed in Section 4.3, while Section 4.4
describes a range of alternative tests. An empirical illustration involving heteroskedastic
error terms is presented in Section 4.5.
The basics of autocorrelation are treated in Sections 4.6 and 4.7, while a fairly simple
illustration is given in Section 4.8. In Sections 4.9 and 4.10 attention is paid to some
additional issues concerning autocorrelation, which includes a discussion of moving
average error terms and so-called Newey–West standard errors. Finally, Section 4.11
has an extensive illustration on uncovered interest rate parity, which involves autocorrelation due to a so-called overlapping samples problem.

4.1

Consequences for the OLS Estimator

The model of interest is unchanged and given by
yi = xi β + εi ,

(4.1)

HETEROSKEDASTICITY AND AUTOCORRELATION

80

which can be written as
y = Xβ + ε.

(4.2)

The essential Gauss–Markov assumptions from (A1)–(A4) can be summarized as
E{ε|X} = E{ε} = 0

(4.3)

V {ε|X} = V {ε} = σ I,
2

(4.4)

which say that the conditional distribution of the errors given the matrix of explanatory
variables has zero means, constant variances and zero covariances. In particular this
means that each error has the same variance and that two different error terms are
uncorrelated. These assumptions imply that E{εi |xi } = 0, so that the model corresponds
to the conditional expectation of yi given xi . Moreover, it was shown that the OLS
estimator was the best linear unbiased estimator for β.
Both heteroskedasticity and autocorrelation imply that (4.4) no longer holds. Heteroskedasticity arises if different error terms do not have identical variances, so that the
diagonal elements of the covariance matrix are not identical. For example, it is possible
that different groups in the sample have different variances. It can also be expected
that the variation of unexplained household savings increases with income, just as the
level of savings will increase with income. Autocorrelation almost exclusively arises
in cases where the data have a time dimension. It implies that the covariance matrix is
nondiagonal such that different error terms are correlated. The reason could be persistence in the unexplained part of the model. Both of these problems will be discussed
in more detail below, but for the moment it is important to note that they both violate
(4.4). Let us assume that the error covariance matrix can more generally be written as
V {ε|X} = σ 2 ,

(4.5)

where  is a positive deﬁnite matrix, which, for the sake of argument, we will sometimes assume to be known. It is clear from the above that it may depend upon X.
If we reconsider the proof of unbiasedness of the OLS estimator, it is immediately
clear that only assumption (4.3) was used. As this assumption is still imposed, assuming
(4.5) instead of (4.4) will not change the result that the OLS estimator b is an unbiased
estimator for β. However, the simple expression for the covariance matrix of b is
no longer valid. Recall that the OLS estimator can be written as b = (X X)−1 X y =
β + (X X)−1 X ε. Conditional upon X, the covariance matrix of b thus depends upon
the conditional covariance matrix of ε, given in (4.5). In particular, we obtain (for a
given matrix X )
V {b|X} = V {(X X)−1 X ε|X} = (X X)−1 X V {ε|X}X(X X)−1
= σ 2 (X X)−1 X X(X X)−1 ,

(4.6)

which only reduces to the simpler expression σ 2 (X X)−1 if  is the identity matrix.
Consequently, although the OLS estimator is still unbiased, its routinely computed
variance and standard errors will be based on the wrong expression. Thus, standard
t- and F-tests will no longer be valid and inferences will be misleading. In addition,

DERIVING AN ALTERNATIVE ESTIMATOR

81

the proof of the Gauss–Markov result that the OLS estimator is BLUE also breaks
down, so that the OLS estimator is unbiased, but no longer best.
These consequences indicate two ways of handling the problems of heteroskedasticity and autocorrelation. The ﬁrst implies the derivation of an alternative estimator that
is best linear unbiased. The second implies sticking to the OLS estimator but somehow adjust the standard errors to allow for heteroskedasticity and/or autocorrelation.
In fact, there is also a third way of eliminating the problems. The reason is that in
many cases you may ﬁnd heteroskedasticity and (particularly) autocorrelation because
the model you are estimating is misspeciﬁed in one way or the other. If this is the
case, detecting heteroskedasticity or autocorrelation should lead you to reconsider the
model to evaluate to what extent you are conﬁdent in its speciﬁcation. Examples of
this will be discussed below.
For pedagogical purposes we shall ﬁrst, in the next section, consider the derivation
of an alternative estimator. It should be stressed, however, that this is in many cases
not the most natural thing to do.

4.2

Deriving an Alternative Estimator

In this section we shall derive the best linear unbiased estimator for β under assumption
(4.5) assuming that  is completely known. The idea behind the derivation is that
we know the best linear unbiased estimator under the Gauss–Markov assumptions
(A1)–(A4), so that we transform the model such that it satisﬁes the Gauss–Markov
conditions again, i.e. such that we obtain error terms that are homoskedastic and exhibit
no autocorrelation. We start this by writing
 −1 = P  P ,

(4.7)

for some square, nonsingular matrix P, not necessarily unique. For the moment, it is
not important how to ﬁnd such a matrix P. It sufﬁces to note that because  is positive
deﬁnite there will always exist a matrix P that satisﬁes (4.7). Using (4.7) it is possible
to write
 = (P P )−1 = P −1 (P  )−1
P P  = PP −1 (P  )−1 P  = I.
Consequently, it holds for the error term vector ε premultiplied by the transformation
matrix P that
E{P ε|X} = PE {ε|X} = 0
V {P ε|X} = PV {ε|X}P  = σ 2 P P  = σ 2 I.
In other words, P ε satisﬁes the Gauss–Markov conditions. Consequently, we can
transform the entire model by this P matrix to obtain
P y = P Xβ + P ε

or

y ∗ = X∗ β + ε∗ ,

(4.8)

HETEROSKEDASTICITY AND AUTOCORRELATION

82

where the error term vector ε∗ satisﬁes the Gauss–Markov conditions. We know that
applying ordinary least squares in this transformed model produces the best linear
unbiased estimator for β.1 This, therefore, is automatically the best linear unbiased
estimator for β in the original model with assumptions (4.3) and (4.5). The resulting
estimator is given by
β̂ = (X∗ X∗ )−1 X∗ y ∗ = (X  −1 X)−1 X  −1 y.

(4.9)

This estimator is referred to as the generalized least squares (GLS) estimator. It is
easily seen that it reduces to the OLS estimator if  = I . Moreover, the choice of P is
irrelevant for the estimator; only  −1 matters. We shall see several examples of GLS
estimators below which are easier to interpret than this general formula. The point to
remember from this expression is that all the GLS estimators that we will see below
are special cases of (4.9).
Clearly, we can only compute the GLS estimator if the matrix  is known. In
practice this will typically not be the case and  will have to be estimated ﬁrst. Using
an estimated version of  in (4.9) results in a feasible generalized least squares
estimator for β, typically referred to as FGLS or EGLS (with the ‘E’ for estimated).
This raises some additional issues that we will consider below.
The fact that the GLS estimator can be obtained as the OLS estimator in some
transformed model not only has theoretical interest. On the contrary, it is fairly common
to transform the observable variables yourself and apply standard OLS routines. The
advantage of deriving the GLS estimator in this way is also that we do not have to
derive a new covariance matrix or a new estimator for σ 2 : we can simply use all
the standard OLS results after replacing the original variables by their transformed
counterparts. For example, the covariance matrix of β̂ (for a given X ) is given by
V {β̂} = σ 2 (X∗ X∗ )−1 = σ 2 (X  −1 X)−1 ,

(4.10)

where σ 2 can be estimated by dividing the residual sum of squares by the number of
observations minus the number of regressors, i.e.
σ̂ 2 =

1
1
(y ∗ − X∗ β̂) (y ∗ − X∗ β̂) =
(y − Xβ̂)  −1 (y − Xβ̂).
N −K
N −K

(4.11)

The fact that β̂ is BLUE implies that it has a smaller variance than the OLS estimator
b. Indeed, it can be shown that the OLS covariance matrix (4.6) is larger than the GLS
covariance matrix (4.10), in the sense that the matrix difference is positive semi-deﬁnite.

4.3

Heteroskedasticity

4.3.1 Introduction
The case where V {ε|X} is diagonal, but not equal to σ 2 times the identity matrix, is
referred to as heteroskedasticity. It means that the error terms are mutually uncorrelated, while the variance of εi may vary over the observations. This problem is
1

Alternative transformation matrices P can be found such that the vector P ε does not exhibit autocorrelation
or heteroskedasticity. The requirement that P is nonsingular guarantees that no information is lost in the
transformation.

HETEROSKEDASTICITY

83

frequently encountered in cross-sectional models. For example, consider the case where
yi denotes expenditure on food and xi consists of a constant and disposable income
DPI i . An Engel curve for food is expected to be upward sloping (with decreasing
slope). Thus, on average higher income corresponds with higher expenditure on food.
In addition, one can expect that the variation of food expenditures among high-income
households is much larger than the variation among low-income households. If this
is the case, the variance of εi increases with income. This kind of heteroskedasticity
could be modelled as
V {εi |DPI i } = σi2 = σ 2 exp{α2 DPI i } = exp{α1 + α2 DPI i }

(4.12)

for some α2 and α1 = log σ 2 . For the moment, we will not make additional assumptions
about the form of heteroskedasticity. We just assume that
V {εi |X} = V {εi |xi } = σ 2 h2i ,

(4.13)

where all h2i are known and positive. Combining this with the assumed absence of
autocorrelation, we can formulate the new assumption as
V {ε|X} = σ 2 Diag{h2i } = σ 2 ,

(A9)

where Diag{h2i } is a diagonal matrix with elements h21 , . . . , h2N . Assumption (A9)
replaces assumptions (A3) and (A4) from Chapter 2. Clearly, if the variances of our
error terms depend upon the explanatory variables, we can no longer assume independence, as in (A2). Therefore, we replace assumptions (A1) and (A2) by
E{ε|X} = 0,

(A10)

which is weaker. Note that (A10) is still substantially stronger than (A7) which says
that E{εi xi } = 0.
We are interested in the best linear unbiased estimator for β in the model
yi = xi β + εi ,

i = 1, . . . , N

(4.14)

under assumptions (A9) and (A10). To this end, we can use the general matrix expressions from above. From the structure of  it is easily seen that an appropriate
transformation matrix P is given by
P = Diag{h−1
i },

(4.15)

−1
which is a diagonal matrix with elements h−1
1 , . . . , hN . Typical elements in the trans∗
formed data vector Py are thus yi = yi / hi (and similar for the elements in xi and
εi ). The GLS estimator for β is thus obtained by running OLS on the following
transformed model
yi∗ = xi∗  β + εi∗
(4.16)

HETEROSKEDASTICITY AND AUTOCORRELATION

84

or

yi
=
hi



xi
hi



β+

εi
.
hi

(4.17)

It is easily seen that the transformed error term is homoskedastic. The resulting least
squares estimator is given by

β̂ =

N


−1

h−2
i xi xi

i=1

N


h−2
i xi yi .

(4.18)

i=1

(Note that this is a special case of (4.9).) This GLS estimator is sometimes referred
to as a weighted least squares estimator, because it is a least squares estimator in
which each observation is weighted by (a factor proportional to) the inverse of the
error variance. It can be derived directly from minimizing the residual sum of squares
in (2.4) after dividing each element in the sum by h2i . Under assumptions (A9) and
(A10), the GLS estimator is the best linear unbiased estimator for β. The use of weights
implies that observations with a higher variance get a smaller weight in estimation.
Loosely speaking, the greatest weights are given to observations of the highest quality
and the smallest weights to those of the lowest quality. It is important to note that
in the transformed model all variables are transformed, including the intercept term.
This implies that the new model does not contain an intercept term. It should also
be stressed that the transformed regression is only employed to easily determine the
GLS estimator and not necessarily has an interpretation of itself. That is, the parameter
estimates are to be interpreted in the context of the original untransformed model.
4.3.2 Estimator Properties and Hypothesis Testing

Because the GLS estimator is simply an OLS estimator in a transformed model that
satisﬁes the Gauss–Markov properties, we can immediately determine the properties
of β̂ from the standard properties of the OLS estimator, after replacing all variables by
their transformed counterparts. For example, the covariance matrix of β̂ is given by

V {β̂} = σ

2

N


−1

h−2
i xi xi

,

(4.19)

i=1

where the unknown error variance σ 2 can be estimated unbiasedly by
σ̂ 2 =

N

1
h−2 (yi − xi β̂)2 .
N − K i=1 i

(4.20)

If, in addition to assumptions (A9) and (A10), we assume normality of the error terms
as in (A5), it also follows that β̂ has a normal distribution with mean zero and variance
(4.19). This can be used to derive tests for linear restrictions on the β coefﬁcients. For
example, to test the hypothesis H0 : β2 = 1 against H1 : β2 = 1, we can use the t-statistic
given by
β̂ − 1
t2 = 2
.
(4.21)
se(β̂2 )

HETEROSKEDASTICITY

85

Because we assumed that all h2i s are known, estimating the error variance by σ̂ 2
has the usual consequence of changing the standard normal distribution into a tN−K
distribution. If normality of the errors is not assumed, the normal distribution is only
asymptotically valid. The null hypothesis would be rejected at the 5% level if |t2 | is
larger than the critical value of the standard normal distribution, which is 1.96.
As before, the F-test can be used to test a number of linear restrictions on β,
summarized as H0 : Rβ = q, where R is of dimension J × K. For example, we could
test β2 + β3 + β4 = 1 and β5 = 0 simultaneously (J = 2). The alternative is H1 : Rβ =
q (which means that the equality sign does not hold for at least one element). The test
statistic is based upon the GLS estimator β̂ and requires the (estimated) variance of
R β̂, which is given by V {R β̂} = RV {β̂}R  . It is given by
ξ = (R β̂ − q) (R V̂ {β̂}R  )−1 (R β̂ − q).

(4.22)

Under H0 this statistic has an asymptotic χ 2 distribution with J degrees of freedom.
This test is usually referred to as a Wald test (compare Chapters 2 and 3). Because
V̂ {β̂} is obtained from V {β̂} by replacing σ 2 by its estimate σ̂ 2 , we can also construct
a version of this test that has an exact F-distribution (imposing normality of the error
terms), as in the standard case (compare Subsection 2.5.6). The test statistic is given
by f = ξ /J , which under the null hypothesis has an F distribution with J and N − K
degrees of freedom.
4.3.3 When the Variances are Unknown

Obviously, it is hard to think of any economic example in which the variances of
the error terms would be known up to a proportionality factor. Probably the only
relevant case arises when the heteroskedasticity is related to one observed variable
only, for example
2
V {εi |xi } = σ 2 xi2
,
(4.23)
where xi2 is an observed exogenous variable (satisfying xi2 > 0). In this case hi = xi2
and the transformed regression is given by
yi
=
xi2



xi
xi2



β+

εi
,
xi2

while the variance of the new disturbance term is
 

σ2
εi 
V
xi = 2i = σ 2 .

xi2
xi2

(4.24)

(4.25)

If the h2i s are unknown, it is no longer possible to compute the GLS estimator. In this
case β̂ is only of theoretical interest. The obvious solution seems to be to replace the
unknown h2i s by unbiased or consistent estimates and hope that this does not affect
the properties of the (pseudo) GLS estimator. This is not as simple as it seems. The
main problem is that there are N unknown h2i s and only N observations to estimate
them. In particular, for any observation i there is only one residual ei to estimate the

HETEROSKEDASTICITY AND AUTOCORRELATION

86

variance of εi . As a consequence, we cannot expect to ﬁnd consistent estimators for
the h2i s unless additional assumptions are made. These assumptions relate to the form
of heteroskedasticity and will usually specify the N unknown variances as a function
of observed (exogenous) variables and a small number of unknown parameters.
Often the variance of the error term may be related to more than just a single exoge2
nous variable. In addition, the relationship between σi2 and xik
may not be proportional.
Therefore, more general forms than (4.23) are often employed. For example,

or

α
V {εi } = σ 2 xik

(4.26)

α1
α2
V {εi } = σ 2 (xik
+ xi
),

(4.27)

where (xik , xi ) are two observed exogenous variables. The speciﬁcations in (4.26) and
(4.27) contain additional unknown parameters that have to be estimated ﬁrst to apply
the GLS procedure with estimated values of h2i . Suppose for the moment that we have
consistent estimates for α1 and α2 . Then, we can compute ĥ2i , which is a consistent
estimator for h2i , and subsequently compute the estimator

∗

β̂ =

N


−1

ĥ−2
i xi xi

i=1

N


ĥ−2
i xi yi .

(4.28)

i=1

This estimator is a feasible (or estimated) generalized least squares estimator (FGLS,
EGLS), because it is based on estimated values for h2i . Provided the unknown parameters in h2i are consistently estimated it holds (under some weak regularity conditions)
that the EGLS estimator β̂ ∗ and the GLS estimator β̂ are asymptotically equivalent.
This just means that asymptotically we can ignore the fact that the unknown weights
are replaced by consistent estimates. Unfortunately, the EGLS estimator does not share
the small sample properties of the GLS estimator, so that we cannot say that β̂ ∗ is
BLUE. In fact, β̂ ∗ will usually be a nonlinear estimator as ĥ2i is a nonlinear function
of yi s. Thus, although we can expect that in reasonably large samples the behaviour
of the EGLS and the GLS estimator are fairly similar, there is no guarantee that the
EGLS estimator outperforms the OLS estimator in small samples (although usually
it does).
What we can conclude is that under assumptions (A9) and (A10), together with
an assumption about the form of heteroskedasticity, the feasible GLS estimator is
consistent for β and asymptotically best (asymptotically efﬁcient). Its covariance matrix
can be estimated by
 N
−1

−2
∗
2

V̂ {β̂ } = σ̂
ĥi xi xi
,
(4.29)
i=1

where σ̂ 2 is the standard estimator for the error variance from the transformed regression (based on (4.20) but replacing β̂ by β̂ ∗ ).
In the remaining part of our discussion on heteroskedasticity, we shall pay attention
to three issues. First, we shall see that we can apply ordinary least squares and adjust
its standard errors for heteroskedasticity, without making any assumptions about its

HETEROSKEDASTICITY

87

form. Second, we shall see how assumptions on the form of heteroskedasticity can be
exploited to consistently estimate the unknown parameters in h2i in order to determine
the EGLS estimator. Third, in Section 4.4, we discuss a range of alternative tests for
the detection of heteroskedasticity.
4.3.4 Heteroskedasticity-consistent Standard Errors for OLS

Reconsider the model with heteroskedastic errors,
yi = xi β + εi ,

(4.30)

with E{εi |X} = 0 and V {εi |X} = σi2 . In matrix notation this can be written as
y = Xβ + ε,
with V {ε|X} = σ 2  = Diag{σi2 }. If we apply ordinary least squares in this model, we
know from the general results above that this estimator is unbiased and consistent for
β. From (4.6), the appropriate covariance matrix is given by
V {b|X} = (X X)−1 X Diag{σi2 }X(X X)−1 .

(4.31)

It seems that to estimate this covariance matrix we also need to estimate all σi2 s, which
is impossible without additional assumptions. However, in an important paper, White
(1980) argues that only a consistent estimator of the K × K matrix
≡

N
1 
1  2 
σ xx
X Diag{σi2 }X =
N
N i=1 i i i

(4.32)

is required. Under very general conditions, it can be shown that
N
1  2 
e xx,
S≡
N i=1 i i i

(4.33)

where ei is the OLS residual, is a consistent2 estimator for
V̂ {b} = (X X)−1

=

N


ei2 xi xi (X X)−1

i=1
N


−1

xi xi

i=1

. Therefore,

N

i=1


ei2 xi xi

N


−1
xi xi

(4.34)

i=1

can be used as an estimate of the true variance of the OLS estimator. This result shows
that we can still make appropriate inferences based upon b without actually specifying
the type of heteroskedasticity. All we have to do is replace the standard formula
2

To be precise, the probability limit of S −

equals a null matrix.

HETEROSKEDASTICITY AND AUTOCORRELATION

88

for computing the OLS covariance matrix with the one in (4.34), which is a simple
option in most modern software packages. Standard errors computed as the square
root of the diagonal elements in (4.34) are usually referred to as heteroskedasticityconsistent standard errors or simply White standard errors.3 The use of such robust
standard errors has become a standard practice in many areas of application. Because
the resulting tests statistics are (asymptotically) appropriate, whether or not the errors
have a constant variance, this is referred to as ‘heteroskedasticity-robust inference’. If
you have some idea about the form of heteroskedasticity (i.e. how hi depends upon
observables and unknown parameters), feasible generalized least squares may provide
a more efﬁcient estimator. The following two subsections provide examples of this.
4.3.5 A Model with Two Unknown Variances

In this subsection we consider a simple case where the sample consists of two separate
groups which can have a different error variance. As examples, consider samples
of developed and developing countries, single-person and multi-person households,
male and female workers, etc. A linear wage equation for this latter sample could be
speciﬁed as
yi = xi β + εi ,
where E{εi |xi } = 0 and V {εi |xi } = σA2 if i belongs to group A (males) and V {εi |xi } =
σB2 if i belongs to group B (females). If we would know4 σA2 and σB2 , GLS estimation
would be straightforward. When σA2 and σB2 are unknown, they can be estimated fairly
simply. Just split the sample into the two groups (males/females) and run separate
regressions. Using the residuals from these regressions, the error variance can be estimated in the usual way, because within each subsample the error term is homoskedastic.
Assume that there are NA observations in the ﬁrst group and NB in the second group.
The OLS estimator for β based on group A observations is given by

bA =



−1
xi xi

i∈A



xi yi

i∈A

where the summations are over all observations in group A. Similarly, we obtain bB .
The error variance is estimated in the standard way, viz.
sA2 =


1
(y − xi bA )2
NA − K i∈A i

(4.35)

and similarly for sB2 . These are unbiased and consistent estimators for σA2 and σB2 . The
EGLS estimator for β is given by

β̂ ∗ =


i∈A

3

sA−2 xi xi +


i∈B

−1 
sB−2 xi xi


i∈A

sA−2 xi yi +




sB−2 xi yi .

(4.36)

i∈B

This covariance matrix estimate is also attributed to Eicker (1967), so that some authors refer to the
corresponding standard errors as the Eicker–White standard errors.
4
To compute the GLS estimator, it is actually sufﬁcient to know the ratio σA2 /σB2 .

HETEROSKEDASTICITY

89

It is easily seen that (4.36) is a special case of (4.28). Moreover, it can be shown that
(4.36) is a matrix-weighted average of the two least squares estimators bA and bB . In
particular, β̂ ∗ = W bA + (I − W )bB , where I is the K-dimensional identity matrix and
W =



i∈A

sA−2 xi xi +



−1
sB−2 xi xi

i∈B



sA−2 xi xi .

(4.37)

i∈A

The weighting matrices W and I − W are inversely related to the (estimated) variance
matrices of the respective estimators. Thus the more accurate estimate gets a higher
weight than the less accurate (higher variance) one.
4.3.6 Multiplicative Heteroskedasticity

A common form of heteroskedasticity employed in practice is that of multiplicative
heteroskedasticity. Here it is assumed that the error variance is related to a number of
exogenous variables, gathered in a J-dimensional vector zi (not including a constant).
To guarantee positivity of the error variance for all parameter values, an exponential
function is used. In particular, it is assumed that
V {εi |xi } = σi2 = σ 2 exp{α1 zi1 + · · · + αJ ziJ } = σ 2 exp{zi α}

(4.38)

where zi is a vector of observed variables that is a function of xi (usually a subset of
xi variables or a transformation thereof). In this model the error variance is related to
one or more exogenous variables, as in the Engel curve example above. Note that in
the special case when J = 1 and zi1 is a dummy variable (e.g. a dummy for males),
we obtain the model with two unknown variances.
To be able to compute the EGLS estimator, we need consistent estimators for the
unknown parameters in h2i = exp{zi α}, that is for α, which can be based upon the
OLS residuals. To see how, ﬁrst note that log σi2 = log σ 2 + zi α. One can expect that
the OLS residuals ei = yi − xi b have something to tell about σi2 . Indeed it can be
shown that
log ei2 = log σ 2 + zi α + vi ,
(4.39)
where vi = log(ei2 /σi2 ) is an error term which is (asymptotically) homoskedastic and
uncorrelated with zi . One problem is that vi does not have zero expectation (not even
asymptotically). However, this will only affect the estimation of the constant log σ 2 ,
which is irrelevant. Consequently, the EGLS estimator for β can be obtained along the
following steps.
1. Estimate the model with OLS. This gives the least squares estimator b.
2. Compute log ei2 = log(yi − xi b)2 from the least square residuals.
3. Estimate (4.39) with least squares, i.e. regress log ei2 upon zi and a constant. This
gives consistent estimators α̂ for α.
4. Compute ĥ2i = exp{zi α̂} and transform all observations to obtain
yi /ĥi = (xi /ĥi ) β + (εi /ĥi ),

HETEROSKEDASTICITY AND AUTOCORRELATION

90

and run OLS on the transformed model. Do not forget to transform the constant.
This yields the EGLS estimator β̂ ∗ for β.
5. The scalar σ 2 can be estimated consistently by
N

(yi − xi β̂ ∗ )2
1
.
N − K i=1
ĥ2i

σ̂ 2 =

6. Finally, a consistent estimator for the covariance matrix of β̂ ∗ is given by

∗

V̂ {β̂ } = σ̂

2

N

xi xi
i=1

−1

ĥ2i

.

This corresponds to the least squares covariance matrix in the transformed regression
that is automatically computed in regression packages.

4.4

Testing for Heteroskedasticity

In order to judge whether in a given model the OLS results are misleading because of
inappropriate standard errors due to heteroskedasticity, a number of alternative tests
are available. If these tests do not reject the null, there is no need to suspect our
least squares results. If rejections are found, we may consider the use of an EGLS
estimator, heteroskedasticity-consistent standard errors for the OLS estimator, or we
may revise the speciﬁcation of our model. In this section, we discuss several tests
that are designed to test the null hypothesis of homoskedasticity against a variety of
alternative hypotheses of heteroskedasticity.
4.4.1 Testing Equality of Two Unknown Variances

The ﬁrst test we consider concerns the situation of two unknown variances as discussed
above, i.e. the variance of εi equals σA2 if observation i belongs to group A and equals
σB2 if observation i belongs to group B. The null hypothesis is the hypothesis that the
variance is constant, i.e. H0 : σA2 = σB2 . A test for H0 can be derived by using the result
that (approximately or exactly if we assume normality of the errors)
(Nj − K)

sj2
σj2

∼ χN2 j −K ,

j = A, B.

(4.40)

Moreover, sA2 and sB2 are independent, so we have that (see Appendix B)
sA2 /σA2
−K
∼ FNNBA−K
.
sB2 /σB2

(4.41)

Under the null hypothesis this reduces to
−K
.
λ = sA2 /sB2 ∼ FNNBA−K

(4.42)

TESTING FOR HETEROSKEDASTICITY

91

In the case of a two-sided alternative, H1 : σA2 = σB2 , the null hypothesis of homoskedasticity is thus rejected if the ratio of the two estimated variances is either too small or
too large. For a one-sided alternative, H1 : σA2 > σB2 , we reject if λ is too large. If
the alternative hypothesis speciﬁes that σA2 < σB2 we can simply interchange the role
of groups A and B in computing the test statistic. This test is a special case of the
Goldfeld–Quandt test (Goldfeld and Quandt, 1965; see Greene, 2003, Section 11.4).
4.4.2 Testing for Multiplicative Heteroskedasticity

For this test, the alternative hypothesis is well-speciﬁed and is given by (4.38), i.e.
σi2 = σ 2 exp{zi α},

(4.43)

where zi is a J-dimensional vector as before. The null hypothesis of homoskedasticity
corresponds with α = 0, so the problem under test is
H0 : α = 0 versus H1 : α = 0.
This hypothesis can be tested using the results of the least squares regression in
(4.39). There are several (asymptotically equivalent) ways to perform this test, but
the simplest one is based on the standard F-test in (4.39) for the hypothesis that all
coefﬁcients, except the constant, are equal to zero. This statistic is usually automatically
provided in a regression package. Because the error term in (4.39) does not satisfy the
Gauss–Markov conditions exactly, the F-distribution (with J and N − J − 1 degrees of
freedom) holds only by approximation. Another approximation is based on the asymptotic χ 2 -distribution (with J degrees of freedom) of the test statistic after multiplication
by J (compare Subsection 2.5.6).
4.4.3 The Breusch–Pagan Test

In this test, proposed by Breusch and Pagan (1980), the alternative hypothesis is less
speciﬁc and generalizes (4.38). It is given by
σi2 = σ 2 h(zi α),

(4.44)

where h is an unknown, continuously differentiable function (that does not depend
on i ), such that h(.) > 0 and h(0) = 1. As a special case (if h(t) = exp{t}) we obtain
(4.38). A test for H0 : α = 0 versus H1 : α = 0 can be derived independently of the
function h. The simplest variant of the Breusch–Pagan test can be computed as the
number of observations multiplied by the R 2 of an auxiliary regression, in particular the
R 2 of a regression of ei2 (the squared OLS residuals) on zi and a constant. The resulting
test statistic, given by ξ = N R 2 is asymptotically χ 2 distributed with J degrees of
freedom. The Breusch–Pagan test is a Lagrange multiplier test for heteroskedasticity.
The main characteristics of Lagrange multiplier tests are that they do not require that
the model is estimated under the alternative and that they are often simply computed
from the R 2 of some auxiliary regression (see Chapter 6).

HETEROSKEDASTICITY AND AUTOCORRELATION

92

4.4.4 The White Test

All tests for heteroskedasticity above test for deviations from the null of homoskedasticity in particular directions. That is, it is necessary to specify the nature of heteroskedasticity one is testing for. The White test (White, 1980) does not require
additional structure on the alternative hypothesis and exploits further the idea of a
heteroskedasticity-consistent covariance matrix for the OLS estimator. As we have
seen, the correct covariance matrix of the least squares estimator is given by (4.31),
which can be estimated by (4.34). The conventional estimator is

V̂ {b} = s

2

N


−1
xi xi

.

(4.45)

i=1

If there is no heteroskedasticity, (4.45) will give a consistent estimator of V {b}, while
if there is, it will not. White has devised a statistical test based on this observation. A
simple operational version of this test is carried out by obtaining N R 2 in the regression
of ei2 on a constant and all (unique) ﬁrst moments, second moments and cross products
of the original regressors. The test statistic is asymptotically distributed as Chi-squared
with P degrees of freedom, where P is the number of regressors in the auxiliary
regression, excluding the intercept.
The White test is a generalization of the Breusch–Pagan test, which also involves
an auxiliary regression of squared residuals, but excludes any higher order terms.
Consequently, the White test may detect more general forms of heteroskedasticity than
the Breusch–Pagan test. In fact, the White test is extremely general. Although this
is a virtue, it is, at the same time, a potentially serious shortcoming. The test may
reveal heteroskedasticity, but it may instead simply identify some other speciﬁcation
error (such as an incorrect functional form). On the other hand, the power of the
White test may be rather low against certain alternatives, particularly if the number of
observations is small.
4.4.5 Which Test?

In practice, the choice of an appropriate test for heteroskedasticity is determined by how
explicit we want to be about the form of heteroskedasticity. In general, the more explicit
2
we are, e.g. σi2 = σ 2 xik
, the more powerful the test will be, i.e. the more likely it is that
the test will correctly reject the null hypothesis. However, if the true heteroskedasticity
is of a different form, the chosen test may not indicate the presence of heteroskedasticity
at all. The most general test, the White test, has limited power against a large number of
alternatives, while a speciﬁc test, like the one for multiplicative heteroskedasticity, has
more power but only against a limited number of alternatives. In some cases, a visual
inspection of the residuals (e.g. a plot of OLS residuals against one or more exogenous
variables) or economic theory can help us in choosing the appropriate alternative. You
may also refer to the graphs presented in Section 3.5.

4.5

Illustration: Explaining Labour Demand

In this section we consider a simple model to explain labour demand of Belgian ﬁrms.
To this end, we have a cross-sectional data set of 569 ﬁrms that includes information

ILLUSTRATION: EXPLAINING LABOUR DEMAND

93

for 1996 on the total number of employees, their average wage, the amount of capital
and a measure of output. The following four variables play a role:5
labour:
capital :
wage:
output:

total employment (number of workers);
total ﬁxed assets (in million euro);
total wage costs divided by number of workers (in 1000 euro);
value added (in million euro).

To set ideas, let us start from a simple production function6
Q = f (K, L),
where Q denotes output and K and L denote the capital and labour input, respectively.
The total production costs are rK + w L, where r denotes the costs of capital and w
denotes the wage rate. Taking r and w and the output level Q as given, minimizing
total costs (with respect to K and L) subject to the production function, results in
demand functions for capital and labour. In general form, the labour demand function
can be written as
L = g(Q, r, w )
for some function g. Because observations on the costs of capital are not easily available
and typically do not exhibit much cross-sectional variation, we will, in estimation,
approximate r by the capital stock K. The inclusion of capital stock in a labour demand
equation may also be motivated by more advanced theoretical models (see Layard and
Nickell, 1986).
First, we shall assume that the function g is linear in its arguments and add an
additive error term. Estimating the resulting linear regression model using the sample
of 569 ﬁrms yields the results reported in Table 4.1. The coefﬁcient estimates all have
the expected sign: higher wages ceteris paribus lead to a reduction of labour input,
while more output requires more labour.
Before interpreting the associated standard errors and other statistics, it is useful to check for the possibility of heteroskedasticity. We do this by performing a
Breusch–Pagan test using the alternative hypothesis that the error variance depends
upon the three explanatory variables. Running an auxiliary regression of the squared
Table 4.1 OLS results linear model
Dependent variable: labour
Variable

Estimate

Standard error

t-ratio

constant
wage
output
capital

287.72
−6.742
15.40
−4.590

19.64
0.501
0.356
0.269

14.648
−13.446
43.304
−17.067

s = 156.26 R 2 = 0.9352 R̄ 2 = 0.9348 F = 2716.02
5
6

The data are available in LABOUR2.
An excellent overview of production functions with cost minimization, in an applied econometrics context,
is given in Wallis (1979).

HETEROSKEDASTICITY AND AUTOCORRELATION

94

Table 4.2

Auxiliary regression Breusch–Pagan test

Dependent variable: ei2
Variable

Estimate

constant
wage
output
capital

−22719.51
228.86
5362.21
−3543.51

Standard error

t-ratio

11838.88
302.22
214.35
162.12

−1.919
0.757
25.015
−21.858

s = 94182 R 2 = 0.5818 R̄ 2 = 0.5796 F = 262.05

OLS residuals upon wage, output and capital, including a constant, leads to the results
in Table 4.2. The high t-ratios as well as the relatively high R 2 are striking and indicate
that the error variance is unlikely to be constant. We can compute the Breusch–Pagan
test statistic by computing N = 569 times the R 2 of this auxiliary regression, which
gives 331.0. As the asymptotic distribution under the null hypothesis is a Chi-squared
with three degrees of freedom, this implies a very sound rejection of homoskedasticity.
It is actually quite common to ﬁnd heteroskedasticity in situations like this, in which
the size of the observational units differs substantially. For example, our sample contains ﬁrms with one employee and ﬁrms with over 1000 employees. We can expect
that large ﬁrms have larger absolute values of all variables in the model, including the
unobservables collected in the error term. A common approach to alleviate this problem is to use logarithms of all variables rather than their levels (compare Section 3.5).
Consequently, our ﬁrst step in handling the heteroskedasticity problem is to consider a
loglinear model. It can be shown that the loglinear model is obtained if the production
function is of the Cobb–Douglas type, i.e. Q = AK α Lβ .
The OLS estimation results for the loglinear model are given in Table 4.3. Recall that
in the loglinear model the coefﬁcients have the interpretation of elasticities. The wage
elasticity of labour demand is estimated to be −0.93, which is fairly high. It implies
that a 1% increase in wages, ceteris paribus, results in almost 1% decrease in labour
demand. The elasticity of the demand for labour with respect to output has an estimate
of approximately unity, so that 1% more output requires 1% more labour input.
If the error term in the loglinear model is heteroskedastic, the standard errors and
t-ratios in Table 4.3 are not appropriate. We can perform a Breusch–Pagan test in a
similar way as before: the auxiliary regression of squared OLS residuals upon the three
explanatory variables (in logs) leads to an R 2 of 0.0136. The resulting test statistic is
7.74, which is on the margin of being signiﬁcant at the 5% level. A more general test is
the White test. To compute the test statistic we run an auxiliary regression of squared
Table 4.3 OLS results loglinear model
Dependent variable: log(labour)
Variable

Estimate

Standard error

t-ratio

constant
log(wage)
log(output)
log(capital )

6.177
−0.928
0.990
−0.004

0.246
0.071
0.026
0.019

25.089
−12.993
37.487
−0.197

s = 0.465 R 2 = 0.8430 R̄ 2 = 0.8421 F = 1011.02

ILLUSTRATION: EXPLAINING LABOUR DEMAND

95

OLS residuals upon all original regressors, their squares and all their interactions. The
results are presented in Table 4.4. With an R 2 of 0.1029, the test statistic takes the
value of 58.5, which is highly signiﬁcant for a Chi-squared variable with 9 degrees
of freedom. Looking at the t-ratios in this regression, the variance of the error term
appears to be signiﬁcantly related to output and capital.
As the White test strongly indicates the presence of heteroskedasticity, it seems
appropriate to compute heteroskedasticity-consistent standard errors for the OLS estimator. This is a standard option in most modern software packages and the results
are presented in Table 4.5. Clearly, the adjusted standard errors are larger than the
incorrect ones, reported in Table 4.3. Note that the F-statistic is also adjusted and
uses the heteroskedasticity-consistent covariance matrix. (Some software packages
simply reproduce the F-statistic from Table 4.3.) Qualitatively, the conclusions are
not changed: wages and output are signiﬁcant in explaining labour demand, capital
is not.
If we are willing to make assumptions about the form of heteroskedasticity, the use of
the more efﬁcient EGLS estimator is an option. Let us consider the multiplicative form
in (4.38), where we choose zi = xi . That is, the variance of εi depends upon log(wage),
log(output) and log(capital ). We can estimate the parameters of the multiplicative
heteroskedasticity by computing the log of the squared OLS residuals and then running
a regression of log ei2 upon zi and a constant. This gives the results in Table 4.6.
The variables log(capital ) and log(output) appear to be important in explaining the
Table 4.4 Auxiliary regression White test
Dependent variable: ei2
Variable

Estimate

Standard error

t-ratio

constant
log(wage)
log(output)
log(capital )
log2 (wage)
log2 (output)
log2 (capital )
log(wage)log(output)
log(wage)log(capital )
log(output)log(capital )

2.545
−1.299
−0.904
1.142
0.193
0.138
0.090
0.138
−0.252
−0.192

3.003
1.753
0.560
0.376
0.259
0.036
0.014
0.163
0.105
0.037

0.847
−0.741
−1.614
3.039
0.744
3.877
6.401
0.849
−2.399
−5.197

s = 0.851 R 2 = 0.1029 R̄ 2 = 0.0884 F = 7.12
Table 4.5 OLS results loglinear model with White standard errors
Dependent variable: log(labour)
Variable

Estimate

constant
log(wage)
log(output)
log(capital )

6.177
−0.928
0.990
−0.004

Heteroskedasticity-consistent
Standard error
t-ratio
0.294
0.087
0.047
0.038

s = 0.465 R 2 = 0.8430 R̄ 2 = 0.8421 F = 544.73

21.019
−10.706
21.159
−0.098

HETEROSKEDASTICITY AND AUTOCORRELATION

96

Table 4.6 Auxiliary regression multiplicative heteroskedasticity
Dependent variable: log ei2
Variable

Estimate

Standard error

t-ratio

constant
log(wage)
log(output)
log(capital )

−3.254
−0.061
0.267
−0.331

1.185
0.344
0.127
0.090

−2.745
−0.178
2.099
−3.659

s = 2.241 R 2 = 0.0245 R̄ 2 = 0.0193 F = 4.73

variance of the error term. Also note that the F-value of this auxiliary regression
leads to rejection of the null hypothesis of homoskedasticity. To check whether this
speciﬁcation for the form of heteroskedasticity is not too restrictive, we estimated
a version where the three squared terms are also included. An F-test on the three
restrictions implied by the model presented in Table 4.6 produced an f-statistic of 1.85
(p = 0.137), so that the null hypothesis cannot be rejected.
Recall that the previous regression produces consistent estimates for the parameters
describing the multiplicative heteroskedasticity, excluding the constant. The exponential of the predicted values of the regression can be used to transform the original
data. As the inconsistency of the constant affects all variables equiproportionally, it
does not affect the estimation results based on the transformed data. Transforming all
variables and using an OLS procedure on the transformed equation yields the EGLS
estimates presented in Table 4.7. If we compare the results in Table 4.7 with the OLS
results with heteroskedasticity-consistent standard errors in Table 4.5 we see that the
efﬁciency gain is substantial. The standard errors for the EGLS approach are much
smaller. Note that a comparison with the results in Table 4.3 is not appropriate, as the
standard errors in the latter table are only valid in the absence of heteroskedasticity.
The EGLS coefﬁcient estimates are fairly close to the OLS ones. A remarkable difference is that the effect of capital is now signiﬁcant at the 5% level, while we did not
ﬁnd statistical evidence for this effect before. We can test the hypothesis that the wage
elasticity equals minus one by computing the t-statistic (−0.856 + 1)/0.072 = 2.01,
which implies a (marginal) rejection at the 5% level.
The fact that the R 2 in Table 4.7 is larger than in the OLS case is misleading for
two reasons. First, the transformed model does not contain an intercept term so that
the uncentred R 2 is computed. Second, the R 2 is computed for the transformed model
with a transformed endogenous variable. If one would compute the implied R 2 for the
original model, it would be smaller than the one obtained by running OLS. It is known
Table 4.7 EGLS results loglinear model
Dependent variable: log(labour)
Variable

Estimate

Standard error

t-ratio

constant
log(wage)
log(output)
log(capital )

5.895
−0.856
1.035
−0.057

0.248
0.072
0.027
0.022

23.806
−11.903
37.890
−2.636

s = 2.509 R 2 = 0.9903 R̄ 2 = 0.9902 F = 14401.3

AUTOCORRELATION

97

from Chapter 2 that the alternative deﬁnitions of the R 2 do not give the same outcome
if the model is not estimated by OLS. Using the deﬁnition that
R 2 = corr2 {yi , ŷi },

(4.46)

where ŷi = xi β̂ ∗ , the above example produces an R 2 of 0.8403, which is only slightly
lower than the OLS value. Because OLS is deﬁned to minimize the residual sum of
squares, it automatically maximizes the R 2 . Consequently, the use of any other estimator will never increase the R 2 , and the R 2 is not a good criterion to compare alternative
estimators. (Of course, there are more important things in an econometrician’s life than
a high R 2 .)

4.6

Autocorrelation

We will now look at another case where V {ε} = σ 2 I is violated, viz. when the covariances between different error terms are not all equal to zero. The most relevant example
of this occurs when two or more consecutive error terms are correlated, and we say that
the error term is subject to autocorrelation or serial correlation. Given our general
discussion above, as long as it can be assumed that E{ε|X} = 0 (assumption (A9)),
the consequences of autocorrelation are similar to those of heteroskedasticity: OLS
remains unbiased, but it becomes inefﬁcient and its standard errors are estimated in
the wrong way.
Autocorrelation normally occurs only when using time series data. To stress this,
we shall follow the literature and index the observations from t = 1, 2, . . . , T rather
than from i = 1, 2, . . . , N . The most important difference is that now the order of the
observations does matter and the index reﬂects a natural ordering. In general, the error
term εt picks up the inﬂuence of those variables affecting the dependent variables that
have not been included in the model. Persistence of the effects of excluded variables
is therefore a frequent cause of positive autocorrelation. If such excluded variables are
observed and could have been included in the model, we can also interpret the resulting
autocorrelation as an indication of a misspeciﬁed model. This explains why tests for
autocorrelation are very often interpreted as misspeciﬁcation tests. Incorrect functional
forms, omitted variables and an inadequate dynamic speciﬁcation of the model may
all lead to ﬁndings of autocorrelation.
Suppose you are using monthly data to estimate a model that explains the demand for
ice cream. Typically, the state of the weather will be an important factor hidden in the
error term εt . In this case, you are likely to ﬁnd a pattern of observations that is like the
one in Figure 4.1. In this ﬁgure we plot ice cream consumption against time, while the
connected points describe the ﬁtted values of a regression model that explains ice cream
consumption from aggregate income and a price index.7 Clearly, positive and negative
residuals group together. In macro-economic analyses, business cycle movements may
have very similar effects. In most economic applications, autocorrelation is positive,
but sometimes it will be negative: a positive error for one observation is likely to be
followed by a negative error for the next, and vice versa.
7

The data used in this ﬁgure are taken from Hildreth and Lu (1960) and are available in ICECREAM; see
also Section 4.8.

HETEROSKEDASTICITY AND AUTOCORRELATION

98

Consumption

.548

.256
1

30
Time

Figure 4.1 Actual and ﬁtted consumption of ice cream, March 1951–July 1953

4.6.1 First Order Autocorrelation

There are many forms of autocorrelation and each one leads to a different structure for
the error covariance matrix V {ε}. The most popular form is known as the ﬁrst-order
autoregressive process. In this case the error term in
yt = xt β + εt

(4.47)

is assumed to depend upon its predecessor as follows
εt = ρεt−1 + vt ,

(4.48)

where vt is an error term with mean zero and constant variance σv2 , that exhibits no
serial correlation. This assumes that the value of the error term in any observation
is equal to ρ times its value in the previous observation plus a fresh component vt .
This fresh component is assumed to have zero mean and constant variance, and to
be independent over time. Furthermore, assumption (A2) from Chapter 2 is imposed
which implies that all explanatory variables are independent of all error terms. The
parameters ρ and σv2 are typically unknown, and, along with β we may wish to estimate
them. Note that the statistical properties of vt are the same as those assumed for εt in
the standard case: thus if ρ = 0, εt = vt and the standard Gauss–Markov conditions
(A1)–(A4) from Chapter 2 are satisﬁed.
To derive the covariance matrix of the error term vector ε, we need to make an
assumption about the distribution of the initial period error, ε1 . Most commonly, it is
assumed that ε1 is mean zero with the same variance as all other εt s. This is consistent
with the idea that the process has been operating for a long period in the past and that
|ρ| < 1. When the condition |ρ| < 1 is satisﬁed we say that the ﬁrst-order autoregressive process is stationary. A stationary process is such that the mean, variances and

AUTOCORRELATION

99

covariances of εt do not change over time (see Chapter 8 below). Imposing stationarity
it easily follows from
E{εt } = ρE{εt−1 } + E{vt }
that E{εt } = 0. Further, from
V {εt } = V {ρεt−1 + vt } = ρ 2 V {εt−1 } + σv2 ,
we obtain that the variance of εt , denoted as σε2 , is given by
σε2 = V {εt } =

σv2
.
1 − ρ2

(4.49)

The nondiagonal elements in the variance–covariance matrix of ε follow from
2
} + E{εt−1 vt } = ρ
cov{εt , εt−1 } = E{εt εt−1 } = ρE{εt−1

σv2
.
1 − ρ2

(4.50)

The covariance between error terms two periods apart is
E{εt εt−2 } = ρE{εt−1 εt−2 } + E{εt−2 vt } = ρ 2

σv2
,
1 − ρ2

(4.51)

and in general we have, for non-negative values of s,
E{εt εt−s } = ρ s

σv2
.
1 − ρ2

(4.52)

This shows that for 0 < |ρ| < 1 all elements in ε are mutually correlated with a decreasing covariance if the distance in time gets large (i.e. if s gets large). The covariance
matrix of ε is thus a full matrix (a matrix without zero elements). From this matrix an
appropriate transformation matrix can be derived, as discussed in Section 4.2. However,
looking at (4.47) and (4.48) directly, it is immediately apparent which transformation
is appropriate. Because εt = ρεt−1 + vt , where vt satisﬁes the Gauss–Markov conditions, it is obvious that a transformation like εt − ρεt−1 will generate homoskedastic
non-autocorrelated errors. That is, all observations should be transformed as yt − ρyt−1
and xt − ρxt−1 . Consequently, the transformed model is given by
yt − ρyt−1 = (xt − ρxt−1 ) β + vt ,

t = 2, 3, . . . , T .

(4.53)

Because the model in (4.53) satisﬁes the Gauss–Markov conditions, estimation with
OLS yields the GLS estimator (assuming ρ is known). However, this statement is not
entirely correct, since the transformation in (4.53) cannot be applied to the ﬁrst observation (because y0 and x0 are not observed). The information in this ﬁrst observation
is lost and OLS in (4.53) produces only an approximate GLS estimator.8 Of course,
8

Technically, the implicit transformation matrix P that is used here is not a square matrix and thus
not invertible.

HETEROSKEDASTICITY AND AUTOCORRELATION

100

when the number of observations is large, the loss of a single observation will typically
not have a large impact on the results.
The ﬁrst observation can be rescued by noting that the error term for the ﬁrst observation, ε1 , is uncorrelated with all vt s, t = 2, . . . , T . However, the variance of ε1 (given
in (4.49)) is much larger than the variance of the transformed errors (v2 , . . . , vT ), particularly when ρ is close to unity. To obtain homoskedastic and non-autocorrelated
errors in a transformed model (which includes the ﬁrst observation), this ﬁrst observation should be transformed by multiplying it by 1 − ρ 2 . The complete transformed
model is thus given by
1 − ρ 2 y1 =

1 − ρ 2 x1 β +

1 − ρ 2 ε1 ,

(4.54)

and by (4.53) for observations 2 to T. It is easily veriﬁed that the transformed error
in (4.54) has the same variance as vt . OLS applied on (4.53) and (4.54) produces the
GLS estimator β̂, which is the best linear unbiased estimator (BLUE) for β.
In early work (Cochrane and Orcutt, 1949) it was common to drop the ﬁrst (transformed) observation and to estimate β from the remaining T − 1 transformed observations. As said, this yields only an approximate GLS estimator and it will not be as
efﬁcient as the estimator using all T observations. However, if T is large the difference between the two estimators is negligible. Estimators not using the ﬁrst transformed
observations are often referred to as Cochrane–Orcutt estimators. Similarly, the transformation not including the ﬁrst observation is referred to as the Cochrane–Orcutt
transformation. The estimator that uses all transformed observations is sometimes called
the Prais–Winsten (1954) estimator.
4.6.2 Unknown ρ

In practice it is of course highly uncommon that the value of ρ is known. In that case
we will have to estimate it. Starting from
εt = ρεt−1 + vt ,

(4.55)

where vt satisﬁes the usual assumptions, it seems natural to estimate ρ from a regression
of the OLS residual et on et−1 . The resulting OLS estimator for ρ is given by

ρ̂ =

T

t=2

−1 
2
et−1

T



et et−1 .

(4.56)

t=2

While this estimator for ρ is typically biased, it is a consistent estimator for ρ under
weak regularity conditions. If we use ρ̂ instead of ρ to compute the feasible GLS
(EGLS) estimator β̂ ∗ , the BLUE property is no longer retained. Under the same conditions as before, it holds that the EGLS estimator β̂ ∗ is asymptotically equivalent to
the GLS estimator β̂. That is, for large sample sizes we can ignore the fact that ρ
is estimated.
A related estimation procedure is the so-called iterative Cochrane–Orcutt procedure,
which is applied in many software packages. In this procedure ρ and β are recursively
estimated until convergence, i.e. having estimated β with EGLS (by β̂ ∗ ), the residuals

TESTING FOR FIRST ORDER AUTOCORRELATION

101

are recomputed and ρ is estimated again using the residuals from the EGLS step. With
this new estimate of ρ, EGLS is applied again and one obtains a new estimate of β.
This procedure goes on until convergence, i.e. until both the estimate for ρ and the
estimate for β do not change anymore. One can expect that this procedure increases
the efﬁciency (i.e. decreases the variance) of the estimator for ρ. However, there is no
guarantee that it will increase the efﬁciency of the estimator for β as well. We know that
asymptotically it does not matter that we estimate ρ, and – consequently – it does not
matter (asymptotically) how we estimate it either, as long as it is estimated consistently.
In small samples, however, iterated EGLS typically performs somewhat better than its
two-step variant.

4.7

Testing for First Order Autocorrelation

When ρ = 0 no autocorrelation is present and OLS is BLUE. If ρ = 0 inferences based
on the OLS estimator will be misleading because standard errors will be based on the
wrong formula. Therefore, it is common practice with time series data to test for autocorrelation in the error term. Suppose we want to test for ﬁrst order autocorrelation
indicated by ρ = 0 in (4.48). We will present several alternative tests for autocorrelation below. The ﬁrst set of tests are relatively simple and based on asymptotic
approximations, while the last test has a known small sample distribution.
4.7.1 Asymptotic Tests

The OLS residuals from (4.47) provide useful information about the possible presence
of serial correlation in the equation’s error term. An intuitively appealing starting point
is to consider the regression of the OLS residual et upon its lag et−1 . This regression
may be done with or without an intercept term (leading to marginally different results).
This auxiliary regression not only produces an estimate for the ﬁrst order autocorrelation coefﬁcient, ρ̂, but also routinely provides a standard error to this estimate. In
the absence of lagged dependent variables in (4.47), the corresponding t-test is asymptotically valid. In fact, the resulting test statistic can be shown to be approximately
equal to
√
t ≈ T ρ̂,
(4.57)
which provides an alternative way of computing the test statistic. Consequently, at the
5% signiﬁcance level we reject the null hypothesis of no autocorrelation against a twosided alternative if |t| > 1.96. If the alternative hypothesis is positive autocorrelation
(ρ > 0), which is often expected a priori, the null hypothesis is rejected at the 5%
level if t > 1.64 (compare Subsection 2.5.1).
Another alternative is based upon the R 2 of the auxiliary regression (including an
intercept term). If we take the R 2 of this regression and multiply it by the effective
number of observations T − 1 we obtain a test statistic that, under the null hypothesis,
has a χ 2 distribution with one degree of freedom. Clearly an R 2 close to zero in
this regression implies that lagged residuals are not explaining current residuals and
a simple way to test ρ = 0 is by computing (T − 1)R 2 . This test is a special case of
the Breusch (1978)–Godfrey (1978) Lagrange multiplier test (see Chapter 6) and is

HETEROSKEDASTICITY AND AUTOCORRELATION

102

easily extended to higher orders of autocorrelation (by including additional lags of the
residual and adjusting the degrees of freedom accordingly).
If the model of interest includes a lagged dependent variable (or other explanatory
variables that are correlated with lagged error terms), the above tests are still appropriate provided that the regressors xt are included in the auxiliary regression. This
takes account of the possibility that xt and ut−1 are correlated, and makes sure that
the test statistics have the appropriate approximate distribution. When it is suspected
that the error term in the equation of interest is heteroskedastic, such that the variance of εt depends upon xt , the t-versions of the autocorrelation tests can be made
heteroskedasticity-consistent by using White standard errors (see Subsection 4.3.4) in
the auxiliary regression to construct the test statistics.

4.7.2 The Durbin–Watson Test

A popular test for ﬁrst order autocorrelation is the Durbin–Watson test (Durbin and
Watson, 1950), which has a known small sample distribution under a restrictive set
of conditions. Two important assumptions underlying this test are that we can treat
the xt s as deterministic and that xt contains an intercept term. The ﬁrst assumption is
important because it requires that all error terms are independent of all explanatory
variables (assumption (A2)). Most importantly, this excludes the inclusion of lagged
dependent variables in the model.
The Durbin–Watson test statistic is given by
dw =

T
2
t=2 (et − et−1 )
,
T
2
t=1 et

(4.58)

where et is the OLS residual (notice the different indices for the summations). Straightforward algebra shows that
dw ≈ 2 − 2ρ̂,
(4.59)
where the approximation sign is due to small differences in the observations over
which summations are taken. Consequently, a value of dw close to 2 indicates that
the ﬁrst order autocorrelation coefﬁcient ρ is close to zero. If dw is ‘much smaller’
than 2, this is an indication for positive autocorrelation (ρ > 0); if dw is much larger
than 2 then ρ < 0. Even under H0 : ρ = 0, the distribution of dw depends not only
upon the sample size T and the number of variables K in xt , but also upon the actual
values of the xt s. Consequently, critical values cannot be tabulated for general use.
Fortunately, it is possible to compute upper and lower limits for the critical values
of dw that depend only upon sample size T and number of variables K in xt . These
values, dL and dU , were tabulated by Durbin and Watson (1950) and Savin and White
(1977), and are partly reproduced in Table 4.8. The true critical value dcrit is between
the bounds that are tabulated, that is dL < dcrit < dU . Under H0 we thus have that (at
the 5% level)
P {dw < dL } ≤ P {dw < dcrit } = 0.05 ≤ P {dw < dU }.

ILLUSTRATION: THE DEMAND FOR ICE CREAM

103

Table 4.8 Lower and upper bounds for 5% critical values of the Durbin–Watson test (Savin
and White, 1977)
Number of
observations
T
T
T
T
T

= 25
= 50
= 75
= 100
= 200

dL

K =3

1.206
1.462
1.571
1.634
1.748

dU

1.550
1.628
1.680
1.715
1.789

Number of regressors (incl. intercept)
K =5
K=7
dL
dU
dL
dU
1.038
1.378
1.515
1.592
1.728

1.767
1.721
1.739
1.758
1.810

0.868
1.291
1.458
1.550
1.707

2.012
1.822
1.801
1.803
1.831

dL

K=9

0.702
1.201
1.399
1.506
1.686

dU

2.280
1.930
1.867
1.850
1.852

For a one-sided test against positive autocorrelation (ρ > 0), there are three possibilities:
a. dw is less than dL . In this case, it is certainly lower than the true critical value dcrit ,
so you would reject H0 ;
b. dw is larger than dU . In this case, it is certainly larger than dcrit and you would not
reject H0 ;
c. dw lies between dL and dU . In this case it might be larger or smaller than the critical
value. Because you cannot tell which, you are unable to accept or reject H0 . This
is the so-called ‘inconclusive region’.
The larger the sample size, the smaller the inconclusive region. For K = 5 and T =
25 we have dL;5% = 1.038 and dU ;5% = 1.767; for T = 100 these numbers are 1.592
and 1.758.
The existence of an inclusive region and the requirement that the Gauss–Markov
conditions, including normality of the error terms, are satisﬁed are important drawbacks of the Durbin–Watson test. Nevertheless, because it is routinely supplied by
most regression packages it typically provides a quick indication of the potential presence of autocorrelation. Values substantially less than 2 are an indication of positive
autocorrelation (as they correspond to ρ̂ > 0). Note that the asymptotic tests are approximately valid, even without normal error terms, and can be extended to allow for lagged
dependent variables in xt .
In the less common case where the alternative hypothesis is the presence of negative
autocorrelation (ρ < 0), the true critical value is between 4 − dU and 4 − dL , so that
no additional tables are required.

4.8

Illustration: The Demand for Ice Cream

This empirical illustration is based on one of the founding articles on autocorrelation,
viz. Hildreth and Lu (1960). The data used in this study are time series data with
30 four-weekly observations from 18 March 1951 to 11 July 1953 on the following
variables:9
9

Data available in ICECREAM.

HETEROSKEDASTICITY AND AUTOCORRELATION

104

cons:
income:
price:
temp:

consumption of ice cream per head (in pints);
average family income per week (in US Dollars);
price of ice cream (per pint);
average temperature (in Fahrenheit).

A graphical illustration of the data is given in Figure 4.2, where we see the time series
patterns of consumption, price and temperature (divided by 100). The graph clearly
suggests that the temperature is an important determinant for the consumption of ice
cream, which supports our expectations.
The model used to explain consumption of ice cream is a linear regression model
with income, price and temp as explanatory variables. The results of a ﬁrst OLS regression are given in Table 4.9. While the coefﬁcient estimates have the expected signs, the
Durbin–Watson statistic is computed as 1.0212. For a one-sided Durbin–Watson test
for H0 : ρ = 0, against the alternative of positive autocorrelation, we have at the 5%
level (α = 0.05) that dL = 1.21 (T = 30, K = 4) and dU = 1.65. The value of 1.02
clearly implies that the null hypothesis should be rejected against the alternative of
positive autocorrelation. When we plot the observed values of cons and the predicted
Consumption
Temp /100
Price

.8

.6

.4

.2
0

10

20

30

Time

Figure 4.2 Ice cream consumption, price and temperature/100
Table 4.9

OLS results

Dependent variable: cons
Variable

Estimate

Standard error

t-ratio

constant
price
income
temp

0.197
−1.044
0.00331
0.00345

0.270
0.834
0.00117
0.00045

0.730
−1.252
2.824
7.762

s = 0.0368 R 2 = 0.7190 R̄ 2 = 0.6866 F = 22.175
dw = 1.0212

ILLUSTRATION: THE DEMAND FOR ICE CREAM

105

.6

Consumption

.5

.4

.3

.2
0

10

20

30

Time

Figure 4.3

Actual and ﬁtted values (connected) ice cream consumption

values according to the model, as in Figure 4.3, we see that positive (negative) values for the error term are more likely to be followed by positive (negative) values.
Apparently, the inclusion of temp in the model is insufﬁcient to capture the seasonal
ﬂuctuation in ice cream consumption.
The ﬁrst order autocorrelation coefﬁcient in
εt = ρεt−1 + vt
is easily estimated by saving the residuals from the previous regression and running
a least squares regression of et on et−1 (without a constant).10 This gives an estimate
ρ̂ = 0.401 with an R 2 of 0.149.
The asymptotic test for H0 : ρ = 0 against ﬁrst order
√
autocorrelation is based on T ρ̂ = 2.19. This is larger than the 5% critical value from
the standard normal distribution given by 1.96, so again we have to reject the null
hypothesis of no serial correlation. The Breusch–Godfrey test produces a test statistic
of (T − 1)R 2 = 4.32, which exceeds the 5% critical value of 3.84 of a Chi-squared
distribution with one degree of freedom.
These rejections imply that OLS is no longer the best linear unbiased estimator for
β and, most importantly, that the routinely computed standard errors are not correct. It
is possible to make correct and more accurate statements about the price elasticity of
ice cream if we choose a more efﬁcient estimation method, like (estimated) GLS. The
iterative Cochrane–Orcutt method yields the results presented in Table 4.10. Note that
the EGLS results conﬁrm our earlier results which indicate that income and temperature
are important determinants in the consumption function. It should be stressed that the
statistics in Table 4.10 that are indicated by an asterisk correspond to the transformed
model and are not directly comparable to their equivalents in Table 4.9 that reﬂect the
untransformed model. This also holds for the Durbin–Watson statistic, which is no
longer appropriate in Table 4.10.
10

There is no need to include a constant because the average OLS residual is zero.

HETEROSKEDASTICITY AND AUTOCORRELATION

106

Table 4.10 EGLS (iterative Cochrane–Orcutt) results
Dependent variable: cons
Variable

Estimate

Standard error

t-ratio

constant
price
income
temp

0.157
−0.892
0.00320
0.00356

0.300
0.830
0.00159
0.00061

0.524
−1.076
2.005
5.800

0.401

0.2079

ρ̂

s = 0.0326∗ R 2 = 0.7961∗
dw = 1.5486∗

R̄ 2 = 0.7621∗

1.927
F = 23.419

Table 4.11 OLS results extended speciﬁcation
Dependent variable: cons
Variable

Estimate

Standard error

t-ratio

constant
price
income
temp
temp t−1

0.189
−0.838
0.00287
0.00533
−0.00220

0.232
0.688
0.00105
0.00067
0.00073

0.816
−1.218
2.722
7.953
−3.016

s = 0.0299 R 2 = 0.8285 R̄ 2 = 0.7999 F = 28.979
dw = 1.5822

As mentioned before, the ﬁnding of autocorrelation may be an indication that there is
something wrong with the model, like the functional form or the dynamic speciﬁcation.
A possible way to eliminate the problem of autocorrelation is to change the speciﬁcation
of the model. It seems natural to consider including one or more lagged variables in
the model. In particular, we will include the lagged temperature temp t−1 in the model.
OLS in this extended model produces the results in Table 4.11.
Compared to Table 4.9, the Durbin–Watson test statistic has increased to 1.58, which
is in the inconclusive region (α = 0.05) given by (1.14, 1.74). As the value is fairly
close to the upper bound, we may choose not to reject the null of no autocorrelation. Apparently lagged temperature has a signiﬁcant negative effect on ice cream
consumption, while the current temperature has a positive effect. This may indicate
an increase of demand when the temperature rises, which is not fully consumed and
reduces expenditures one period later.11

4.9

Alternative Autocorrelation Patterns

4.9.1 Higher Order Autocorrelation

First order autoregressive errors are not uncommon in macro-economic time series
models and in most cases allowing for ﬁrst order autocorrelation will eliminate the
problem. However, when we have quarterly or monthly data for example, it is possible
11

What is measured by cons is expenditures on ice cream, not actual consumption.

ALTERNATIVE AUTOCORRELATION PATTERNS

107

that there is a periodic (quarterly or monthly) effect that is causing the errors across
the same periods but in different years to be correlated. For example, we could have
(in the case of quarterly data) that
εt = γ εt−4 + vt ,

(4.60)

εt = γ1 εt−1 + γ2 εt−2 + γ3 εt−3 + γ4 εt−4 + vt ,

(4.61)

or, more generally,

which is known as fourth order autocorrelation. Essentially, this is a straightforward
generalization of the ﬁrst order process and estimation by EGLS follows along the
same lines. As long as the explanatory variables are uncorrelated with all error terms,
EGLS is based on a ﬁrst step OLS estimation of (4.60) or (4.61) where the errors
are replaced by least squares residuals et . The appropriate transformation to derive
the EGLS estimator for β will be clear from (4.60) or (4.61). Note that the ﬁrst four
observations will be lost in the transformation.
4.9.2 Moving Average Errors

As discussed, an autoregressive speciﬁcation of the errors, as in (4.48), (4.60) or (4.61),
implies that all error terms are mutually correlated, although the correlation between
terms that are many periods apart will be negligibly small. In some cases, (economic)
theory suggests a different form of autocorrelation, in which only particular error terms
are correlated, while all others have a zero correlation. This can be modelled by a socalled moving average error process. Moving average structures often arise when the
sampling interval (e.g. one month) is smaller than the interval for which the variables
are deﬁned. Consider the problem of estimating an equation to explain the value of
some ﬁnancial instrument such as 90-day treasury bills or 3-month forward contracts
on foreign exchange. If one uses monthly data, then any innovation occurring in month
t would affect the value of instruments maturing in months t, t + 1 and t + 2 but would
not affect the value of instruments maturing later, because the latter would not yet have
been issued. This suggests correlation between the error terms one and two months
apart, but zero correlation between terms further apart.
Another example is the explanation of the yearly change in prices (inﬂation),
observed every 6 months. Suppose we have observations on the change in consumer
prices compared to the level one year ago, at 1 January and 1 July. Also suppose
that background variables (for example money supply) included in xt are observed
half-yearly. If the ‘true’ model is given by
yt = xt β + vt ,

t = 1, 2, . . . , T (half-yearly),

(4.62)

where yt is the half-yearly change in prices and the error term vt satisﬁes the
Gauss–Markov conditions, it holds for the change on a yearly level, yt∗ = yt + yt−1 that
yt∗ = (xt + xt−1 ) β + vt + vt−1 ,
or



yt∗ = xt∗ β + εt ,

t = 1, 2, . . . , T ,

t = 1, 2, . . . , T ,

(4.63)
(4.64)

108

HETEROSKEDASTICITY AND AUTOCORRELATION

where εt = vt + vt−1 and xt∗ = xt + xt−1 . If we assume that vt has a variance σv2 , the
properties of the error term in (4.64) are the following
E{εt } = E{vt } + E{vt−1 } = 0
V {εt } = V {vt + vt−1 } = 2σv2
cov{εt , εt−1 } = cov{vt + vt−1 , vt−1 + vt−2 } = σv2
cov{εt , εt−s } = cov{vt + vt−1 , vt−s + vt−1−s } = 0,

s = 2, 3, . . .

Consequently, the covariance matrix of the error term vector contains a large number
of zeros. On the diagonal we have 2σv2 (the variance) and just below and above the
diagonal we have σv2 (the ﬁrst order autocovariance), while all other covariances are
equal to zero. We call this a ﬁrst order moving average process (for εt ). In fact, this is
a restricted version because the correlation coefﬁcient between εt and εt−1 is a priori
ﬁxed at 0.5. A general ﬁrst order moving average process would be speciﬁed as
εt = vt + αvt−1
for some α, |α| < 1; see the discussion in Chapter 8 on time series models.
It is generally somewhat harder to estimate regression models with moving average
errors than with autoregressive errors. This is because the transformation generating ‘Gauss–Markov errors’ is complicated. Some software packages have specialized
procedures available, but if appropriate software is lacking, estimation can be quite difﬁcult. A possible solution is to apply ordinary least squares while correcting standard
errors for the presence of autocorrelation (of whatever nature) in εt . This will be discussed in the next section. An empirical example involving moving average errors is
provided in Section 4.11.

4.10

What to do When you Find Autocorrelation?

In many cases the ﬁnding of autocorrelation is an indication that the model is misspeciﬁed. If this is the case, the most natural route is not to change your estimator
(from OLS to EGLS) but to change your model. Typically, three (interrelated) types
of misspeciﬁcation may lead to a ﬁnding of autocorrelation in your OLS residuals:
dynamic misspeciﬁcation, omitted variables and functional form misspeciﬁcation.
If we leave the case where the error term is independent of all explanatory variables,
there is another reason why GLS or EGLS may be inappropriate. In particular, it is
possible that the GLS estimator is inconsistent because the transformed model does not
satisfy the minimal requirements for the OLS estimator to be consistent. This situation
can arise even if OLS applied to the original equation is consistent. Section 4.11
provides an empirical example of this issue.
4.10.1 Misspeciﬁcation
Let us start with functional form misspeciﬁcation. Suppose that the true linear relationship is between yt and log xt as

yt = β1 + β2 log xt + εt

WHAT TO DO WHEN YOU FIND AUTOCORRELATION?

109

y

2.01065

0.097666
1

40
Time

Figure 4.4

Actual and ﬁtted values when true model is yt = 0.5 log t + εt

and suppose, for illustrative purposes, that xt increases with t. If we nevertheless
estimate a linear model that explains yt from xt we could ﬁnd a situation as depicted in
Figure 4.4. In this ﬁgure, based upon simulated data with xt = t and yt = 0.5 log xt plus
a small error, the ﬁtted values of a linear model are connected while the actual values
are not. Very clearly, residuals of the same sign group together. The Durbin–Watson
statistic corresponding to this example is as small as 0.193. The solution in this case
is not to re-estimate the linear model using feasible generalized least squares but to
change the functional form and include log xt rather than xt .
As discussed above, the omission of a relevant explanatory variable may also lead
to a ﬁnding of autocorrelation. For example, in Section 4.8 we saw that excluding
sufﬁcient variables that reﬂect the seasonal variation of ice cream consumption resulted
in such a case. In a similar fashion, an incorrect dynamic speciﬁcation may result in
autocorrelation. In such cases, we have to decide whether the model of interest is
supposed to be static or dynamic. To illustrate this, start from the (static) model
yt = xt β + εt

(4.65)

with ﬁrst order autocorrelation εt = ρεt−1 + vt . We can interpret the above model as
describing E{yt |xt } = xt β. However, we may also be interested in forecasting on the
basis of current xt values as well as lagged observations on xt−1 and yt−1 , that is
E{yt |xt , xt−1 , yt−1 }. For the above model, we obtain

β)
E{yt |xt , xt−1 , yt−1 } = xt β + ρ(yt−1 − xt−1

(4.66)

and we can write a dynamic model as

yt = xt β + ρyt−1 − ρxt−1
β + vt ,

(4.67)

the error term which does not exhibit any autocorrelation. The model in (4.67) shows
that the inclusion of a lagged dependent variable and lagged exogenous variables results

HETEROSKEDASTICITY AND AUTOCORRELATION

110

in a speciﬁcation that does not suffer from autocorrelation. Conversely, we may ﬁnd
autocorrelation in (4.65) if the dynamic speciﬁcation is similar to (4.67) but includes,
for example, only yt−1 or some elements of xt−1 . In such cases, the inclusion of these
‘omitted’ variables will resolve the autocorrelation problem.
The static model (4.65) with ﬁrst order autocorrelation provides us with E{yt |xt }
as well as the dynamic forecast E{yt |xt , xt−1 , yt−1 } and may be more parsimonious
compared to a full dynamic model with several lagged variables included (with unrestricted coefﬁcients). It is a matter of choice whether we are interested in E{yt |xt } or
E{yt |xt , xt−1 , yt−1 } or both. For example, explaining a person’s wage from his wage
in the previous year may be fairly easy, but may not provide answers to the questions
we are interested in. In many applications, though, the inclusion of a lagged dependent
variable in the model will eliminate the autocorrelation problem. It should be emphasized, though, that the Durbin–Watson test is inappropriate in a model where a lagged
dependent variable is present. In Subsection 5.2.1 particular attention is paid to models
with both autocorrelation and a lagged dependent variable.
4.10.2 Heteroskedasticity-and-autocorrelation-consistent Standard
Errors for OLS

Let us reconsider our basic model
yt = xt β + εt ,

(4.68)

where εt is subject to autocorrelation. If this is the model we are interested in, for
example because we want to know the conditional expectation of yt given a wellspeciﬁed xt , we can choose to apply the GLS approach or apply ordinary least squares
while adjusting its standard errors. This last approach is particularly useful when the
correlation between εt and εt−s can be argued to be (virtually) zero after some lag
length H and/or when the conditions for consistency of the GLS estimator happen to
be violated.
If E{xt εt } = 0 and E{εt εt−s } = 0 for s = H, H + 1, . . . , the OLS estimator is consistent and its covariance matrix can be estimated by

V̂ ∗ {b} =

T

t=1

−1
xt xt


T S∗

T


−1
xt xt

,

(4.69)

t=1

where
S∗ =

T
H −1
T

1 2 
1 

et xt xt +
wj
es es−j (xs xs−j
+ xs−j xs ).
T t=1
T j =1
s=j +1

(4.70)

Note that we obtain the White covariance matrix, as discussed in Subsection 4.3.4, if
wj = 0, so that (4.69) is a generalization. In the standard case wj = 1, but this may
lead to an estimated covariance matrix in ﬁnite samples that is not positive deﬁnite.
To prevent this, it is common to use Bartlett weights, as suggested by Newey and
West (1987). These weights decrease linearly with j as wj = 1 − j/H . The use of

WHAT TO DO WHEN YOU FIND AUTOCORRELATION?

111

such a set of weights is compatible with the idea that the impact of the autocorrelation
of order j diminishes with |j |. Standard errors computed from (4.69) are referred
to as heteroskedasticity-and-autocorrelation-consistent (HAC) standard errors or
simply Newey–West standard errors. Sometimes HAC standard errors are also used
when the autocorrelation is strictly speaking not restricted to H lags, for example
with an autoregressive structure. Theoretically, this can be justiﬁed by applying an
asymptotic argument that H increases with T as T goes to inﬁnity (but not as fast as
T ). Empirically, this may not work very well in small samples.
To obtain some intuition for the expression in (4.70), it is instructive to note that S ∗ is
an estimator for the asymptotic covariance matrix of the sample mean (1/T ) Tt=1 xt εt
(compare (2.33) in Chapter 2). Suppose that εt would be observable, then one could
think of estimating this covariance matrix as
1
ε ε x x ,
T s,t t s t s
where the summation is over all relevant elements (symmetric in s and t). This estimator is actually inconsistent, because, for example, the covariance between x1 ε1 and
xT εT is estimated from one data point. This explains why we need to restrict the autocorrelation structure. With zero autocorrelation at lag length H or more, the summation
is over |s − t| ≤ H − 1 only and the above estimator becomes consistent.
The Bartlett weights guarantee that the estimator S ∗ is positive deﬁnite in every
sample. This can be understood by looking at the covariance matrix of a short-run sum
H −1
j =0 xt−j εt−j , which is given by

V


−1
H


j =0

xt−j εt−j





= HE {εt2 xt xt }

+ (H − 1)[E{εt εt−1 xt xt−1
} + E{εt−1 εt xt−1 xt }] + · · ·


+ [E{εt εt−H +1 xt xt−H
+1 } + E{εt−H +1 εt xt−H +1 xt }]

=H

H
−1



(1 − j/H )[E{εt εt−j xt xt−j
} + E{εt−j εt xt−j xt }],

j =0

which is positive deﬁnite by construction. Dividing by H, and replacing the expectation operators by sample averages and εt by et , produces the S ∗ matrix. Because
with one sample there is only one sample mean (1/T ) Tt=1 xt εt to look at, we estimate its variance by looking at the variance, within the sample, of short-run averages,
−1
S ∗ is estimating
(1/H ) H
j =0 xt−j εt−j and divide this by the sample size T. Because
√
an asymptotic covariance matrix, i.e. the covariance matrix of T times the sample
average, this (1/T ) factor disappears again in (4.70). Note that in the absence of autocorrelation we would estimate the variance of the sample mean as the sample variance
of xt εt , divided by T. Finally, the fact that the estimator uses the OLS residuals et
rather than the unobservable εt has no asymptotic consequences.

112

4.11

HETEROSKEDASTICITY AND AUTOCORRELATION

Illustration: Risk Premia in Foreign Exchange
Markets

A trader who orders goods abroad that have to be paid at some later date can settle his
required payments in different ways. As an example, consider a European trader who at
the end of the current month buys an amount of coffee at the price of US$100 000, to be
paid by the end of next month. A ﬁrst strategy to settle his account is to buy dollars now
and hold these in deposit until the end of next month. This has the obvious consequence
that the trader does not get the European (one month) interest rate during this month,
but the US one (assuming he holds the dollar amount in a US deposit). A second
strategy is to buy dollars at the so-called forward market. There a price (exchange
rate) is determined, which has to be paid for dollars when delivered at the end of next
month. This forward rate is agreed upon in the current period and has to be paid at
delivery (one month from now). Assuming that the forward contract is riskless (ignoring
default risk, which is usually very small), the trader will be indifferent between the
two strategies. Both possibilities are without risk and therefore it is expected that both
yield the same return at the end of next month. If not, arbitrage possibilities would
generate riskless proﬁts. The implied equality of the interest rate differential (European
and US rates) and the difference between the forward rate and the spot rate is known
as the covered interest rate parity (CIP) condition.
A third possibility for the trader to pay his bill in dollars is simply to wait until
the end of next month and then buy US dollars at a yet unknown exchange rate. If
the usual assumption is made that the trader is risk-averse, it will only be attractive
to take the additional exchange rate risk if it can be expected that the future spot rate
(expressed in euro per dollar) is lower than the forward rate. If this is the case, we say
that the market is willing to pay a risk premium. In the absence of a risk premium
(the forward rate equals the expected spot rate), the covered interest rate parity implies
the uncovered interest rate parity (UIP), which says that the interest rate differential
between two countries equals the expected relative change in the exchange rate. In
this section we consider tests for the existence of risk premia in the forward exchange
market, based upon regression models.
4.11.1 Notation
For a European investor it is possible to hedge against currency risk by buying at time t
the necessary amount of US dollars for delivery at time t + 1 against a known rate Ft ,
the forward exchange rate. Thus, Ft is the rate at time t against which dollars can be
bought and sold (through a forward contract) at time t + 1. The riskless interest rates
US
E
and Rf,t+1
, respectively. For the European
for Europe and the US are given by Rf,t+1
investor, the investment in US deposits can be made riskless through hedging on the
forward exchange market. That is, a riskless investment for the German investor would
give return
US
+ log Ft − log St ,
(4.71)
Rf,t+1

where St is the current spot (exchange) rate. To avoid riskless arbitrage opportunities
(and unlimited proﬁts for investors), this return should equal the riskless return on
European deposits, i.e. it should hold that
E
US
− Rf,t+1
= log Ft − log St .
Rf,t+1

(4.72)

ILLUSTRATION: RISK PREMIA IN FOREIGN EXCHANGE MARKETS

113

The right hand side of (4.72) is known as the (negative of the) forward discount, while
the left hand side is referred to as the interest differential. Condition (4.72) is known
as covered interest rate parity and is a pure no-arbitrage condition which is therefore
almost surely satisﬁed in practice (if transaction costs are negligible).
An alternative investment corresponds to an investment in US deposits without
hedging the currency risk. The return on this risky investment is
US
Rf,t+1
+ log St+1 − log St ,

(4.73)

the expected value of which equals (4.71) if
Et {log St+1 } = log Ft

or

Et {st+1 } = ft ,

where small letters denote the log of capital letters, and Et {.} denotes the conditional expectation given all available information at time t. The equality Et {st+1 } = ft
together with covered interest rate parity implies the uncovered interest rate parity
condition, which says that the interest differential between two countries equals the
expected exchange rate change, i.e.
E
US
Rf,t+1
− Rf,t+1
= Et {log St+1 } − log St .

(4.74)

Many macro-economic models employ this UIP condition. One of its consequences is
that a small country cannot both control its domestic interest rate level and its exchange
rates. In the sequel attention will be paid to the question whether uncovered interest
rate parity holds, i.e. whether risk premia on the forward exchange markets exist.
The reason why the expected future spot rate Et {st+1 } may differ from the forward
rate ft is the existence of a risk premium. It is possible that the market is willing to
pay a risk premium for taking the exchange rate risk in (4.73). In the absence of a
risk premium, hedging against currency risk is free and any investor can eliminate his
exchange rate risk completely without costs. Because the existence of a positive risk
premium for a European investor implies that a US investor can hedge exchange rate
risk against the euro while receiving a discount, it is not uncommon to assume that
neither investor pays a risk premium. In this case, the foreign exchange market is often
referred to as being (risk-neutral) ‘efﬁcient’ (see Taylor, 1995).
Note that the risk premium is deﬁned as the difference between the expected log
of the future spot rate and the log of the forward rate. Dropping the logarithm has
the important objection that expressing exchange rates in one or the other currency is
−1
no longer irrelevant. In the logarithmic case this is irrelevant because Et {log St+1
}−
−1
log Ft = −Et {log St+1 } + log Ft .
4.11.2 Tests for Risk Premia in the One-month Market

One approach to test for the presence of risk premia is based on a simple regression
framework. In this subsection we shall discuss tests for the presence of a risk premium
in the 1-month forward market using monthly data. That is, the sampling frequency
corresponds exactly with the length of the term contract. Empirical results will be
presented for 1-month forwards on the US$/¤ and US$/£Sterling exchange rates, using

114

HETEROSKEDASTICITY AND AUTOCORRELATION

monthly data from January 1979 to December 2001. The pre-euro exchange rates are
based on the German Mark. The use of monthly data to test for risk premia on the
3-month forward market is discussed in the next subsection.
The hypothesis that there is no risk premium can be written as
H0 : Et−1 {st } = ft−1 .

(4.75)

A simple way to test this hypothesis exploits the well known result that the difference
between a random variable and its conditional expectation given a certain information
set is uncorrelated with any variable from this information set, i.e.
E{(st − Et−1 {st })xt−1 } = 0

(4.76)

for any xt−1 that is known at time t − 1. From this we can write the following regression model

st − ft−1 = xt−1
β + εt ,
(4.77)
where εt = st − Et−1 {st }. If H0 is correct and if xt−1 is known at time t − 1, it should
hold that β = 0. Consequently, H0 is easily tested by testing whether β = 0 for a given
choice of xt−1 variables. Below we shall choose as elements in xt−1 a constant and the
forward discount st−1 − ft−1 .
Because st−1 − ft−2 is observed in period t − 1, εt−1 is also an element of the
information set at time t − 1. Therefore, (4.76) also implies that under H0 the error
terms in (4.77) exhibit no autocorrelation. Autocorrelation in εt is thus an indication for the existence of a risk premium. Note that the hypothesis does not imply
anything about the variance of εt , which suggests that imposing homoskedasticity
may not be appropriate and heteroskedasticity-consistent standard errors could be
employed.
The data employed12 are taken from Datastream and cover the period January
1979–December 2001. We use the US$/¤ rate and the US$/£ rate, which are visualized in Figure 4.5. From this ﬁgure, we can infer the weakness of the dollar in the
beginning of the 1980s and in 2000/2001. In Figure 4.6 the monthly forward discount
st − ft is plotted for both exchange rates. Typically, the forward discount is smaller
than 1% in absolute value. For the euro, the dollar spot rate is in almost all months
below the forward rate, which implies, given the covered interest rate parity argument,
that the US nominal interest rate exceeds the European one. Only during 1993–1994
and at the end of 2001, the converse appears to be the case.
Next, equation (4.77) is estimated by OLS taking xt−1 = (1, st−1 − ft−1 ) . The
results for the US$/£ rate are given in Table 4.12. Because the forward discount
has the properties of a lagged dependent variable (st−1 − ft−1 is correlated with
εt−1 ), the Durbin–Watson test is not appropriate. The simplest alternative is to use
the Breusch–Godfrey test, which is based upon an auxiliary regression of et upon
et−1 , st−1 − ft−1 and a constant (see above) and then taking13 T R 2 . We can test for
higher order autocorrelations by including additional lags, like et−2 and et−3 . This
12
13

The data for this illustration are available in FORWARD2.
Below we use the effective number of observations in the auxiliary regressions to determine T in T R 2 .

ILLUSTRATION: RISK PREMIA IN FOREIGN EXCHANGE MARKETS

115

2.5

2.0
US$/GBP
1.5

1.0
US$/EUR
0.5
80 82 84 86 88 90 92 94 96 98 00

Figure 4.5 US$/EUR and US$/GBP exchange rates, January 1979–December 2001

0.010
US$/GBP

0.005
0.000
−0.005
−0.010
−0.015

US$/EUR
−0.020
80

Figure 4.6

82

84 86

88 90

92

94 96

98

00

Forward discount, US$/EUR and US$/GBP, January 1979–December 2001
Table 4.12

OLS results US$/£Sterling

Dependent variable: st − ft−1
Variable

Estimate

Standard error

t-ratio

constant
st−1 − ft−1

−0.0051
3.2122

0.0024
0.8175

−2.162
3.929

s = 0.0315 R 2 = 0.0535 R̄ 2 = 0.0501 F = 15.440

way, the null hypothesis of no autocorrelation can be tested against the alternatives
of 1st and (up to) 12th order autocorrelation, with test statistics of 0.22 and 10.26.
2
With 5% critical values of 3.84 and 21.0 (for a χ12 and χ12
, respectively), this does not
imply rejection of the null hypotheses. The t-statistics in the regression indicate that

116

HETEROSKEDASTICITY AND AUTOCORRELATION

the intercept term is signiﬁcantly different from zero, while the forward discount has
a signiﬁcantly positive coefﬁcient. A joint test on the two restrictions β = 0 results
in an f-statistic of 7.74 (p = 0.0005), so that the null hypothesis of no risk premium
is rejected. The numbers imply that if the nominal UK interest rate exceeds the US
interest rate such that the forward discount st−1 − ft−1 exceeds 0.16% (e.g. in the early
1990s), it is found that Et−1 {st } − ft−1 is positive. Thus, UK investors can sell their
pounds on the forward market at a rate of, say, $1.75, while the expected spot rate is,
say, $1.77. UK importers that want to hedge against exchange rate risk for their orders
in the US have to pay a risk premium. On the other hand, US traders proﬁt from this;
they can hedge against currency risk and cash (!) a risk premium at the same time.14
The t-tests employed above are only asymptotically valid if εt exhibits no autocorrelation, which is guaranteed by (4.76), and if εt is homoskedastic. The Breusch–Pagan
test statistic for heteroskedasticity can be computed as TR 2 of an auxiliary regression
of et2 upon a constant and st−1 − ft−1 , which yields a value of 7.26, which implies a
clear rejection of the null hypothesis. The use of more appropriate heteroskedasticityconsistent standard errors does not result in qualitatively different conclusions.
In a similar way we can test for a risk premium in the US$/¤ forward rate. The
results of this regression are as follows:
st − ft−1 = − 0.0023 + 0.485 (st−1 − ft−1 ) + et ,
(0.0031) (0.766)
BG(1) = 0.12,

R 2 = 0.0015

BG(12) = 14.12.

Here BG(h) denotes the Breusch–Godfrey test statistic for up to h-th order autocorrelation. For the US$/¤ rate no risk premium is found: both the regression coefﬁcients
are not signiﬁcantly different from zero and the hypothesis of no autocorrelation is
not rejected.

4.11.3 Tests for Risk Premia Using Overlapping Samples

The previous subsection was limited to an analysis of the one-month forward market for foreign exchange. Of course, forward markets exist with other maturities, for
example 3 months or 6 months. In this subsection we shall pay attention to the question to what extent the techniques discussed in the previous section can be used to test
for the presence of a risk premium in the 3-month forward market. The frequency of
observation is, still, one month.
Let us denote the log price of a 3-month forward contract by ft3 . The null hypothesis
of no risk premium can then be formulated as
3
.
H0 : Et−3 {st } = ft−3
14

(4.78)

There is no fundamental problem with the risk premium being negative. While this means that the
expected return is lower than that of a riskless investment, the actual return may still exceed the riskless
rate in situations that are particularly interesting to the investor. For example, a ﬁre insurance on your
house typically has a negative expected return, but a large positive return in the particular case that your
house burns down.

ILLUSTRATION: RISK PREMIA IN FOREIGN EXCHANGE MARKETS

117

Using similar arguments as before, a regression model similar to (4.77) can be written as
3

st − ft−3
= xt−3
β + εt ,
(4.79)
where εt = st − Et−3 {st }. If xt−3 is observed at time t − 3, the vector β in (4.79)
should equal zero under H0 . Simply using OLS to estimate the parameters in (4.79)
with xt−3 = (1, st−3 − ft−3 ) gives the following results for the US$/£ rate:
3
3
st − ft−3
= − 0.014 + 3.135 (st−3 − ft−3
) + et ,
(0.004) (0.529)

BG(1) = 119.69,

R 2 = 0.1146

BG(12) = 173.67.

and for the US$/¤ rate:
3
3
= − 0.011 + 0.006 (st−3 − ft−3
) + et ,
st − ft−3
(0.006) (0.535)

BG(1) = 130.16,

R 2 = 0.0000

BG(12) = 177.76.

These results seem to suggest the clear presence of a risk premium in both markets:
the Breusch–Godfrey tests for autocorrelation indicate strong autocorrelation, while
the regression coefﬁcients for the US$/£ exchange market are highly signiﬁcant. These
conclusions are, however, incorrect.
The assumption that the error terms exhibit no autocorrelation was based on the
observation that (4.76) also holds for xt−1 = εt−1 such that εt+1 and εt were uncorrelated. However, this result is only valid if the frequency of the data coincides with
the maturity of the contract. In the present case, we have monthly data for 3-month
contracts. The analogue of (4.76) now is
E{(st − Et−3 {st })xt−3 } = 0 for any xt−3 known at time t − 3.

(4.80)

Consequently, this implies that εt and εt−j (j = 3, 4, 5, . . .) are uncorrelated but does
not imply that εt and εt−1 or εt−2 are uncorrelated. On the contrary, these errors are
likely to be highly correlated.
Consider an illustrative case where (log) exchange rates are generated by a socalled random walk15 process, i.e. st = st−1 + ηt where the ηt are independent and
identically distributed with mean zero and variance ση2 and where no risk premia exist,
3
i.e. ft−3
= Et−3 {st }. Then it is easily shown that
εt = st − Et−3 {st } = ηt + ηt−1 + ηt−2 .
Consequently, the error term εt is described by a moving average autocorrelation
pattern or order 2. When log exchange rates are not random walks, the error term εt
will comprise ‘news’ from periods t, t − 1 and t − 2, and therefore εt will be a moving
15

More details on random walk processes are provided in Chapter 8.

HETEROSKEDASTICITY AND AUTOCORRELATION

118

average even in the more general case. This autocorrelation problem is due to the socalled overlapping samples problem, where the frequency of observation (monthly) is
higher than the frequency of the data (quarterly). If we test whether the autocorrelation
goes beyond the ﬁrst two lags, that is whether εt is correlated with εt−3 up to εt−12 ,
we can do so by running a regression of the OLS residual et upon et−3 , . . . , et−12 and
the regressors from (4.79). This results in Breusch–Godfrey test statistics of 7.85 and
9.04, respectively, both of which are insigniﬁcant for a Chi-squared distribution with
10 degrees of freedom.
The fact that the ﬁrst two autocorrelations of the error terms in the regressions above
are nonzero implies that the regression results are not informative about the existence
of a risk premium: standard errors are computed in an incorrect way and, moreover,
the Breusch–Godfrey tests for autocorrelation may have rejected because of the ﬁrst
two autocorrelations being nonzero, which is not in conﬂict with the absence of a risk
premium. Note that the OLS estimator is still consistent, even with a moving average
error term.
One way to ‘solve’ the problem of autocorrelation is simply dropping two-thirds
of the information by using the observations from three-month intervals only. This is
unsatisfactory, because of the loss of information and therefore the potential loss of
power of the tests. Two alternatives may come to mind: (i) using GLS to (hopefully)
estimate the model more efﬁciently, and (ii) using OLS while computing corrected
(Newey–West) standard errors. Unfortunately, the ﬁrst option is not appropriate here
because the transformed data will not satisfy the conditions for consistency and GLS
3
will be inconsistent. This is due to the fact that the regressor st−3 − ft−3
is correlated
with lagged error terms.
We shall therefore consider the OLS estimation results again, but compute HAC
standard errors. Note that H = 3 is sufﬁcient. Recall that these standard errors also
allow for heteroskedasticity. The results can be summarized as follows. For the US$/£
rate we have
3
3
st − ft−3
= − 0.014 + 3.135 (st−3 − ft−3
) + et ,
[0.005] [0.663]

R 2 = 0.1146

and for the US$/¤ rate:
3
3
st − ft−3
= − 0.011 + 0.006 (st−3 − ft−3
) + et ,
[0.008] [0.523]

R 2 = 0.0000

where the standard errors within square brackets are the Newey–West standard errors
with H = 3. Qualitatively, the conclusions do not change: for the 3-month US$/£
market, uncovered interest rate parity has to be rejected. Because covered interest rate
parity implies that
∗
st − ft = Rf,t+1
− Rf,t+1 ,
where ∗ denotes the foreign country and the exchange rates are measured, as before,
in units of home currency for one unit of foreign currency, the results imply that at
times when the US interest rate is high relative to the UK one, UK investors pay a risk
premium to US traders. For the European/US market, the existence of a risk premium
was not found in the data.

EXERCISES

119

Exercises
Exercise 4.1 (Heteroskedasticity – Empirical)

The data set AIRQ contains observations for 30 standard metropolitan statistical areas
(SMSAs) in California for 1972 on the following variables:
airq:
vala:
rain:
coas:
dens:
medi :

indicator for air quality (the lower the better);
value added of companies (in 1000 US$);
amount of rain (in inches);
dummy variable, 1 for SMSAs at the coast; 0 for others;
population density (per square mile);
average income per head (in US$).

a. Estimate a linear regression model that explains airq from the other variables using
ordinary least squares. Interpret the coefﬁcient estimates.
b. Test the null hypothesis that average income does not effect the air quality. Test
the joint hypothesis that none of the variables has an effect upon air quality.
c. Test whether the variance of the error term is different for coastal and non-coastal
areas, using the test of Subsection 4.4.1. In view of the outcome of the test, comment upon the validity of the test from b.
d. Perform a Breusch–Pagan test for heteroskedasticity related to all ﬁve explanatory variables.
e. Perform a White test for heteroskedasticity. Comment upon the appropriateness of
the White test in light of the number of observations and the degrees of freedom
of the test.
f. Assuming that we have multiplicative heteroskedasticity related to coas and medi,
estimate the coefﬁcients by running a regression of log ei2 upon these two variables. Test the null hypothesis of homoskedasticity on the basis of this auxiliary regression.
g. Using the results from f, compute an EGLS estimator for the linear model. Compare
your results with those obtained under a. Redo the tests from b.
h. Comment upon the appropriateness of the R 2 in the regression of g.
Exercise 4.2 (Autocorrelation – Empirical)

Consider the data and model of Section 4.8 (the demand for ice cream). Extend the
model by including lagged consumption (rather than lagged temperature). Perform a
test for ﬁrst order autocorrelation in this extended model.
Exercise 4.3 (Autocorrelation Theory)

a. Explain what is meant by the ‘inconclusive region’ of the Durbin–Watson test.
b. Explain why autocorrelation may arise as the result of an incorrect functional
form.
c. Explain why autocorrelation may arise because of an omitted variable.

120

d.

e.
f.
g.
h.

HETEROSKEDASTICITY AND AUTOCORRELATION

Explain why adding a lagged dependent variable and lagged explanatory variables
to the model eliminates the problem of ﬁrst order autocorrelation. Give at least
two reasons why this is not necessarily a preferred solution.
Explain what is meant by an ‘overlapping samples’ problem. What is the problem?
Give an example where ﬁrst order autocorrelation leads to an inconsistent OLS
estimator.
Explain when you would use Newey–West standard errors.
Describe in steps how you would compute the feasible GLS estimator for β in
the standard model with (second order) autocorrelation of the form εt = ρ1 εt−1 +
ρ2 εt−2 + vt . (You do not have to worry about the initial observation(s).)

Exercise 4.4 (Overlapping Samples – Empirical)

The data set FORWARD2 also contains exchange rates for the £ against the euro, for the
period January 1979 to December 2001. Pre-euro exchange rates are computed from
the German Mark.
a.
b.
c.
d.

e.
f.
g.
h.
i.

Produce a graph of the £/¤ exchange rate.
Compute the 1-month and 3-month forward discount for this market and produce
a graph.
Test for the existence of a risk premium on the 1-month horizon using (4.77),
including the lagged forward discount as a regressor.
Test for autocorrelation in this model using the Breusch–Godfrey test. Use a few
different values for the maximum lag length. Why is the Durbin–Watson test not
valid in this case?
Test for the existence of a risk premium on the 3-month horizon using (4.79),
including the 3-month forward discount, lagged three months, as a regressor.
Test for autocorrelation in this model using the Breusch–Godfrey test for up to 2
lags and for up to 12 lags.
Test for autocorrelation in this model for lags 3 to 12.
Compute HAC standard errors for the 3-month risk premium regression.
Interpret your results and compare with those reported in Section 4.11.

5

Endogeneity,
Instrumental Variables
and GMM

Until now, it was assumed that the error terms in the linear regression model were
contemporaneously uncorrelated with the explanatory variables, or even that they were
independent of all explanatory variables.1 As a result, the linear model could be interpreted as describing the conditional expectation of yt given a set of variables xt . In
this chapter we shall discuss cases in which it is unrealistic or impossible to treat
the explanatory variables in a model as given or exogenous. In such cases, it can
be argued that some of the explanatory variables are correlated with the equation’s
error term, such that the OLS estimator is biased and inconsistent. There are different reasons why one may argue that error terms are contemporaneously correlated
with one or more of the explanatory variables, but the common aspect is that the
linear model no longer corresponds to a conditional expectation or a best linear
approximation.
In Section 5.1, we start with a review of the properties of the OLS estimator in the
linear model under different sets of assumptions. Section 5.2 discusses cases where the
OLS estimator cannot be shown to be unbiased or consistent. In such cases, we need
to look for alternative estimators. The instrumental variables estimator is considered in
Sections 5.3 and 5.5, while in Section 5.6 we extend this class of instrumental variables
estimators to the generalized method of moments (GMM), which also allows estimation
of nonlinear models. Empirical illustrations concerning the returns to schooling and the
estimation of intertemporal asset pricing models are provided in Sections 5.4 and 5.7,
respectively.

1

Recall that independence is stronger than uncorrelatedness (see Appendix B).

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

122

5.1

A Review of the Properties of the OLS Estimator

Let us consider the linear model again
yt = xt β + εt ,
or, in matrix notation,

t = 1, 2, . . . , T ,

y = Xβ + ε.

(5.1)
(5.2)

In Chapters 2 and 4 we saw that the OLS estimator b is unbiased for β if it can be
assumed that ε is mean zero and conditional mean independent of X, i.e. if E{ε|X} = 0
(assumption (A10) from Chapter 4). This says that knowing any of the explanatory
variables is uninformative about the expected value of any of the error terms. Independence of X and ε with E{ε} = 0 (assumptions (A1) and (A2) from Section 2.3) implies
that E{ε|X} = 0 but is stronger, as it does not allow the variance of ε to depend upon
X either.
In many cases, the assumption that ε is conditionally mean independent of X is
too strong. To illustrate this, let us start with a motivating example. The efﬁcient
market hypothesis (under constant expected returns) implies that the returns on any
asset are unpredictable from any publicly available information. Under the so-called
weak form of the efﬁcient market hypothesis asset returns cannot be predicted from
their own past (see Fama, 1991). This hypothesis can be tested statistically using a
regression model and testing whether lagged returns explain current returns. That is,
in the model
(5.3)
yt = β1 + β2 yt−1 + β3 yt−2 + εt ,
where yt denotes the return in period t, the null hypothesis of weak form efﬁciency
implies that β2 = β3 = 0. Because the explanatory variables are lagged dependent
variables (which are a function of lagged error terms), the assumption E{ε|X} = 0 is
inappropriate. Nevertheless, we can make weaker assumptions under which the OLS
estimator is consistent for β = (β1 , β2 , β3 ) .
In the notation of the more general model (5.1), consider the following set of assumptions:
xt and εt are independent (for each t)
(A8)
εt ∼ IID(0, σ 2 ),

(A11)

where the notation in (A11) is shorthand for saying that the error terms εt are independent and identically distributed (i.i.d.) with mean zero and variance σ 2 . Under
some additional regularity conditions,2 the OLS estimator b is consistent for β and
−1
asymptotically normally distributed with covariance matrix σ 2 xx
, where as before
T
1
xt xt .
T
T →∞
t=1

xx = plim

2

We shall not present any proofs or derivations here. The interested reader is referred to more advanced
textbooks, like Hamilton (1994, Chapter 8). The most important ‘regularity condition’ is that xx is ﬁnite
and invertible (compare assumption (A6) from Section 2.6).

A REVIEW OF THE PROPERTIES OF THE OLS ESTIMATOR

123

Formally it holds that
√

−1
T (b − β) → N(0, σ 2 xx
),

(5.4)

which corresponds to (2.74) from Chapter 2. In small samples, it thus holds approximately that

 T
−1 

a
.
b ∼ N β, σ 2
(5.5)
xt xt
t=1

This distributional result for the OLS estimator is the same as that obtained under the
Gauss–Markov assumptions (A1)–(A4), combined with normality of the error terms
in (A5), albeit that (5.5) only holds approximately by virtue of the asymptotic result
in (5.4). This means that all standard tests in the linear model (t-tests, F-tests, Wald
tests) are valid by approximation, provided assumptions (A8) and (A11) are satisﬁed.
For the asymptotic distribution in (5.4) to be valid we have to assume that xt and
εt are independent (for each t). This means that xs is allowed to depend upon εt
as long as s = t. The inclusion of a lagged dependent variable is the most important
example of such a situation. The current result shows that as long as the error terms are
independently and identically distributed, the presence of a lagged dependent variable in
xt only affects the small sample properties of the OLS estimator but not the asymptotic
distribution. Under assumptions (A6), (A8) and (A11), the OLS estimator is consistent
and asymptotically efﬁcient.
Assumption (A11) excludes autocorrelation and heteroskedasticity in εt . In the
example above, autocorrelation can be excluded as it is a violation of market efﬁciency (returns should be unpredictable). The homoskedasticity assumption is more
problematic. Heteroskedasticity may arise when the error term is more likely to take
on extreme values for particular values of one or more of the regressors. In this case
the variance of εt depends upon xt . Similarly, shocks in ﬁnancial time series are usually clustered over time, i.e. big shocks are likely to be followed by big shocks, in
either direction. An example of this is that after a stock market crash, it is hard to
predict whether stock prices will go up or down in subsequent periods, but it is clear
that there is much more uncertainty in the market than in other periods. In this case,
the variance of εt depends upon historical innovations εt−1 , εt−2 , . . . . Such cases are
referred to as conditional heteroskedasticity, or sometimes just as ARCH or GARCH,
which are particular speciﬁcations to model this phenomenon.3
−1
When assumption (A11) is dropped, it can no longer be claimed that σ 2 xx
is
the appropriate covariance matrix nor that (5.5) holds by approximation. In general,
however, consistency and asymptotic normality of b are not affected. Moreover, asymptotically valid inferences can be made if we estimate the covariance matrix in a different
way. Let us relax assumptions (A8) and (A11) to
E{xt εt } = 0 for each t
εt are serially uncorrelated with expectation zero.
3

(A7)
(A12)

ARCH is short for AutoRegressive Conditional Heteroskedasticity, and GARCH is a Generalized form of
that. We shall discuss this in more detail in Chapter 8.

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

124

Assumption (A7) imposes that xt is uncorrelated4 with εt , while (A12) allows for
heteroskedasticity in the error term, but excludes autocorrelation. Under some additional regularity conditions, it can be shown that the OLS estimator b is consistent for
β and asymptotically normal, according to
√

−1
−1
T (b − β) → N(0, xx
xx
),

where
 ≡ plim

(5.6)

T
1 2 
ε xx.
T t=1 t t t

In this case, the asymptotic covariance matrix can be estimated following the method
of White (see Chapter 4). Consequently,

V̂ {b} =

T

t=1

−1
xt xt

T

t=1


et2 xt xt

T


−1
xt xt

,

(5.7)

t=1

where et denotes the OLS residual is a consistent estimator for the true covariance
matrix of the OLS estimator under assumptions (A6), (A7) and (A12). Consequently,
all standard tests for the linear model are asymptotically valid in the presence of
heteroskedasticity of unknown form if the test statistics are adjusted by replacing the
standard estimate for the OLS covariance matrix with the heteroskedasticity-consistent
estimate from (5.7).
In several cases, people are interested in predictability of long-horizon returns, for
example over a horizon of several years. In principle, tests of long-term predictability
can be carried out along the same lines as short-term predictability tests. However,
for horizons of ﬁve years, say, this would imply that only a limited number of 5-year
returns can be analysed, even if the sample period covers several decades. Therefore, tests of predictability of long-horizon returns have typically tried to make more
efﬁcient use of the available information by using overlapping samples (compare Subsection 4.11.3); see Fama and French (1988) for an application. In this case, 5-year
returns are computed over all periods of ﬁve consecutive years. Ignoring second order
effects, the return over ﬁve years is simply the sum of ﬁve annual returns, so that the
return over 1990–1994 partly overlaps with, for example, the returns over 1991–1995
and 1992–1996. Denoting
the return in year t as yt , the 5-year return over the years t
to t + 4 is given by Yt = 4j =0 yt+j . To test the predictability of these 5-year returns,
suppose we estimate a model that explains Yt from its value in the previous 5-year
period (Yt−5 ) using data for every year, that is
Yt = δ5 + θ5 Yt−5 + εt ,

t = 1, . . . , T years.

(5.8)

All T annual observations in the sample on 5-year returns are regressed on a constant and the 5-year return lagged ﬁve years. In this model the error term exhibits
4

Note that E{xt zt } = cov{xt , zt } if either xt or zt has a zero mean (see Appendix B).

CASES WHERE THE OLS ESTIMATOR CANNOT BE SAVED

125

autocorrelation because of the overlapping samples problem. In order to explain this
issue, assume that the following model holds for annual returns
yt = δ1 + θ1 yt−1 + ut ,

(5.9)

where ut exhibits no autocorrelation. Under 
the null hypothesis that θ1 = 0, it can be
shown that δ5 = 5δ1 and θ5 = 0, while εt = 4j =0 ut+j . Consequently, the covariance
between εt and εt−j is nonzero as long as j < 5. From Chapter 4 we know that the
presence of autocorrelation invalidates routinely computed standard errors, including
those based on the heteroskedasticity-consistent covariance matrix in (5.7). However,
if we can still assume that the regressors are contemporaneously uncorrelated with the
error terms (condition (A7)), and the autocorrelation is zero after H periods, it can be
shown that all results based on assumptions (A7) and (A12) hold true if the covariance matrix of the OLS estimator is estimated by the Newey–West (1987) estimator
presented in Section 4.10.2

∗

V̂ {b} =

T

t=1

−1
xt xt


TS

∗

T


−1
xt xt

,

(5.10)

t=1

where
S∗ =

T
H −1
T

1 2 
1 

et xt xt +
wj
es es−j (xs xs−j
+ xs−j xs )
T t=1
T j =1
s=j +1

(5.11)

with wj = 1 − j/H . Note that in the above example H equals 5. As a consequence,
the standard tests from the linear model are asymptotically valid in the presence of
heteroskedasticity and autocorrelation (up to a ﬁnite number of lags), if we replace
the standard covariance matrix estimate by the heteroskedasticity and autocorrelation
consistent estimate from (5.10).

5.2

Cases Where the OLS Estimator Cannot be Saved

The previous section shows that we can go as far as assumption (A7) and impose
E{εt xt } = 0, essentially without affecting the consistency of the OLS estimator. If the
autocorrelation in the error term is somehow restricted, it is still possible to make
appropriate inferences in this case, using the White or Newey–West estimates for the
covariance matrix. The assumption that E{εt xt } = 0 says that error terms and explanatory variables are contemporaneously uncorrelated. Sometimes there are statistical or
economic reasons why we would not want to impose this condition. In such cases, we
can no longer argue that the OLS estimator is unbiased or consistent, and we need
to consider alternative estimators. Some examples of such situations are: the presence
of a lagged dependent variable and autocorrelation in the error term, measurement
errors in the regressors, and simultaneity or endogeneity of regressors. Let us now
consider examples of these situations in turn.

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

126

5.2.1 Autocorrelation with a Lagged Dependent Variable

Suppose the model of interest is given by
yt = β1 + β2 xt + β3 yt−1 + εt ,

(5.12)

where xt is a single variable. Recall that as long as we can assume that E{xt εt } = 0
and E{yt−1 εt } = 0 for all t, the OLS estimator for β is consistent (provided that
some regularity conditions are met). However, suppose that εt is subject to ﬁrst order
autocorrelation, that is
εt = ρεt−1 + vt .
(5.13)
Now, we can rewrite the model as
yt = β1 + β2 xt + β3 yt−1 + ρεt−1 + vt .

(5.14)

yt−1 = β1 + β2 xt−1 + β3 yt−2 + εt−1 ,

(5.15)

But it also holds that

from which it follows immediately that the error term εt is correlated with yt−1 . Thus,
if ρ = 0 OLS no longer yields consistent estimators for the regression parameters in
(5.12). A possible solution is the use of maximum likelihood or instrumental variables
techniques that will be discussed below; Stewart and Gill (1998, Section 7.4) provide
additional discussion and details. Note that the Durbin–Watson test is not valid to
test for autocorrelation in model (5.12), because the condition that the explanatory
variables can be treated as deterministic is violated. An alternative test is provided by
the Breusch–Godfrey Lagrange Multiplier test for autocorrelation (see Section 4.7, or
Chapter 6 for a general discussion on Lagrange Multiplier tests). This test statistic can
be computed as T times the R 2 of a regression of the least squares residuals et on
et−1 and all included explanatory variables (including the relevant lagged values of
yt ). Under H0 , the test statistic asymptotically has a Chi-squared distribution with 1
degree of freedom.
It can be noted that in the above example the linear regression model does not
correspond with the conditional expectation of yt given xt and yt−1 . Because knowledge
of yt−1 tells us something about the expected value of the error term εt , it will be the
case that E{εt |xt , yt−1 } is a function of yt−1 . Consequently the last term in
E{yt |xt , yt−1 } = β1 + β2 xt + β3 yt−1 + E{εt |xt , yt−1 }

(5.16)

will be nonzero. As we know that OLS is generally consistent when estimating a
conditional expectation, we may suspect that OLS is inconsistent whenever the model
we are estimating does not correspond to a conditional expectation. A lagged dependent
variable, combined with autocorrelation of the error term, is such a case.

CASES WHERE THE OLS ESTIMATOR CANNOT BE SAVED

127

5.2.2 An Example with Measurement Error

Another illustration where the OLS estimator is likely to be inconsistent arises when
an explanatory variable is subject to measurement error. Suppose that a variable yt
depends upon a variable wt according to
yt = β1 + β2 wt + vt ,

(5.17)

where vt is an error term with zero mean and variance σv2 . It is assumed that E{vt |wt } =
0, such that the model describes the expected value of yt given wt ,
E{yt |wt } = β1 + β2 wt .
As an example, we can think of yt denoting household savings and wt denoting disposable income. We shall suppose that wt cannot be measured absolutely accurately
(for example, because of misreporting) and denote the measured value for wt by xt .
For each observation, xt equals – by construction – the true value wt plus the measurement error ut , that is
xt = wt + ut .
(5.18)
Let us consider the following set of assumptions, which may be reasonable in certain
applications. First, it is assumed that the measurement error ut is mean zero with
constant variance σu2 . Second, ut is assumed to be independent of the error term vt
in the model. Third, and most importantly, the measurement error is independent of
the underlying true value wt . This means that the true level of disposable income (in
our example) does not reveal any information about the size, sign or value of the
measurement error. Substituting (5.18) into (5.17) we obtain
yt = β1 + β2 xt + εt ,

(5.19)

where εt = vt − β2 ut .
Equation (5.19) presents a linear model in terms of the observables yt and xt with an
error term εt . If we use the available data on yt and xt , and unsuspectingly regress yt
upon xt and a constant, the OLS estimator b is inconsistent for β = (β1 , β2 ) , because xt
depends on ut and so does εt . That is, E{xt εt } = 0 and one of the necessary conditions
for consistency of b is violated. Suppose that β2 > 0. When the measurement error in
an observation is positive two things happen: xt has a positive component ut , and εt
has a negative component −β2 ut . Consequently, xt and εt are negatively correlated,
E{xt εt } = cov{xt , εt } < 0, and it follows that the OLS estimator is inconsistent for β.
When β2 < 0, xt and εt are positively correlated.
To illustrate the inconsistency of the OLS estimator, write the estimator for β2 as
(compare Subsection 2.1.2)
b2 =

T

t=1 (xt − x̄)(yt −
T
2
t=1 (xt − x̄)

ȳ)

,

(5.20)

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

128

where x̄ denotes the sample mean of xt . Substituting (5.19), this can be written as
b2 = β2 +

T

t=1 (xt − x̄)(εt −

(1/T ) Tt=1 (xt − x̄)2

(1/T )

ε̄)

.

(5.21)

As the sample size increases to inﬁnity, sample moments converge to population
moments. Thus
plim b2 = β2 +

plim (1/T )

T

t=1 (xt − x̄)(εt −

plim (1/T ) Tt=1 (xt − x̄)2

ε̄)

= β2 +

E{xt εt }
.
V {xt }

(5.22)

The last term in this probability limit is not equal to zero. First,
E{xt εt } = E{(wt + ut )(vt − β2 ut )} = −β2 σu2 ,
and, second,
V {xt } = V {wt + ut } = σw2 + σu2 ,
where σw2 = V {wt }. Consequently,
plim b2 = β2 1 −

σu2
σw2 + σu2

.

(5.23)

So, b2 is consistent only if σu2 = 0, that is, if there is no measurement error. It is asymptotically biased towards zero if σu2 is positive, with a larger bias if the measurement
error is large relative to the variance in the true variable wt . The ratio σu2 /σw2 may be
referred to as a noise-to-signal ratio, because it gives the variance of the measurement
error (the noise) in relation to the variance of the true values (the signal). If this ratio
is small, we have a small bias, if it is large, the bias is also large. In general, the OLS
estimator underestimates the effect of true disposable income if reported disposable
income is subject to measurement error unrelated to the true level.
It is important to note that the inconsistency of b2 carries over to the estimator b1
for the constant term β1 = E{yt − β2 xt }. In particular,
plim (b1 − β1 ) = plim (ȳ − b2 x̄ − E{yt } + β2 E{xt })
= −plim (b2 − β2 )E{xt }.

(5.24)

So, if E{xt } > 0 an overestimation of the slope parameter corresponds to an underestimated intercept. This is a general result: inconsistency of one element in b usually
carries over to all other elements.
Again, let us stress that the model of interest in this case does not correspond to the
conditional expectation of yt given xt . From (5.19) we can derive that
E{yt |xt } = β1 + β2 xt − β2 E{ut |xt },

CASES WHERE THE OLS ESTIMATOR CANNOT BE SAVED

129

where the latter term is nonzero because of (5.18). If we assume normality of ut , wt
and xt , it follows that (see Appendix B)
E{ut |xt } =

σu2
(x − E{xt }).
σw2 + σu2 t

Combining the last two equations and using (5.23) shows that the OLS estimator,
though inconsistent for β2 , is consistent for the coefﬁcients in the conditional expectation of savings yt given reported disposable income xt , but this is not what we are
interested in!5
5.2.3 Simultaneity: the Keynesian Model

Another important situation where we are not interested in a conditional expectation
arises when the model of interest contains behavioural parameters, usually measuring the causal effects of changes in the explanatory variables, and one or more of
these explanatory variables are jointly determined with the left-hand side variable. For
example, if we write down a Keynesian consumption function
Ct = β1 + β2 Yt + εt ,

(5.25)

where Ct denotes a country’s real per capita consumption and Yt is real per capita
income, we want to interpret the coefﬁcient β2 as the marginal propensity to consume
(0 < β2 < 1). This means that β2 has a causal interpretation reﬂecting the impact
of income upon consumption: how much more will people consume if their income
increases by one unit? However, aggregate income Yt is not exogenously given as it
will be determined by the identity
Yt = Ct + It ,

(5.26)

where It is real per capita investment. This equation is a deﬁnition equation for a closed
economy without government and says that total consumption plus total investment
should equal total income. We assume that this relationship holds in the sample.
Let us assume that assumption (A11) holds, which says that εt is i.i.d. over time
with mean zero and variance σ 2 . In addition, it is assumed that
It and εt are independent (for each t)

(5.27)

This last assumption says that investment It is exogenous and determined independently of the error term (that is, determined outside the model). In contrast, both Ct
and Yt are endogenous variables, which are jointly (simultaneously) determined in the
model. The model in (5.25)–(5.26) is a very simple simultaneous equations model in
structural form (or in short: a structural model).
5

This result may be useful as it implies that we can ignore the measurement error problem if we interpret
the coefﬁcients in terms of the effects of reported variables rather than their true underlying values. This
would often not make sense economically but statistically there is no problem.

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

130

The fact that Yt is endogenous has its consequences for the estimation of the consumption function (5.25). Because Ct inﬂuences Yt through (5.26) we can no longer
argue that Yt and εt are uncorrelated. Consequently, the OLS estimator for β2 will be
biased and inconsistent. To elaborate upon this, it is useful to consider the reduced
form of this model, in which the endogenous variables Ct and Yt are expressed as a
function of the exogenous variable It and the error term. Solving (5.25)–(5.26) for Ct
and Yt we obtain the reduced form equations
Yt =

β1
1
1
+
It +
ε,
1 − β2
1 − β2
1 − β2 t

(5.28)

Ct =

β1
β2
1
+
It +
ε.
1 − β2
1 − β2
1 − β2 t

(5.29)

From the ﬁrst of these two equations it follows that
cov{Yt , εt } =

1
1
σ2
cov{It , εt } +
V {εt } =
.
1 − β2
1 − β2
1 − β2

(5.30)

Consequently, equation (5.25) presents a linear model where the regressor Yt is correlated with the error term εt . As a result, OLS applied to (5.25) will be biased and
inconsistent. Similar to the earlier derivation, it holds that
plim b2 = β2 +
where
V {Yt } = V

cov{Yt , εt }
,
V {Yt }

1
1
I +
ε
1 − β 2 t 1 − β2 t

=

(5.31)

1
(V {It } + σ 2 ),
(1 − β2 )2

so that we ﬁnally ﬁnd that
plim b2 = β2 + (1 − β2 )

σ2
.
V {It } + σ 2

(5.32)

As 0 < β2 < 1, and σ 2 > 0, the OLS estimator will overestimate the true marginal
propensity to consume β2 . While we have only shown the inconsistency of the estimator for the slope coefﬁcient, the intercept term will in general also be estimated
inconsistently (compare (5.24)).
The simple model in this subsection illustrates a common problem in macro- and
micro-economic models. If we consider an equation where one or more of the explanatory variables is jointly determined with the left-hand side variable, the OLS estimator
will typically provide inconsistent estimators for the behavioural parameters in this
equation. Statistically, this means that the equation we have written down does not
correspond to a conditional expectation so that the usual assumptions on the error term
cannot be imposed.
In the next sections we shall consider alternative approaches to estimating a single equation with endogenous regressors, using so-called instrumental variables. While
relaxing the exogeneity assumption in (A7), we shall stress that these approaches

THE INSTRUMENTAL VARIABLES ESTIMATOR

131

require the imposition of alternative assumptions, like (5.27), which may or may not
be valid in practice. Such assumptions can be motivated by presenting a complete
system of structural equations, explaining all the endogenous variables and stating all
the relevant exogenous variables. It will be shown that if there are sufﬁcient exogenous variables in the system that can act as instruments, it is possible to identify and
consistently estimate the structural parameters of interest.
The reduced form equations (5.28) and (5.29) express the two endogenous variables in terms of the exogenous variable and an error term. Consequently, we can
estimate the reduced form parameters consistently by applying ordinary least squares
to (5.28) and (5.29). The reduced form parameters are, however, nonlinear functions
of the structural form parameters (which is what we are really interested in) and it
is the question whether the reduced form parameters give us sufﬁcient information
to identify all structural parameters. This is the well-known issue of identiﬁcation
in simultaneous equations models. We shall not, in this text, discuss identiﬁcation in
the context of deriving structural parameters from the reduced form. Interested readers are referred to Hayashi (2000, Chapter 3) or Greene (2003, Chapter 15). Instead,
we consider the identiﬁcation problem as one of ﬁnding sufﬁcient instruments for
the endogenous variables in the model. Strictly speaking, this only gives necessary
conditions for identiﬁcation.

5.3

The Instrumental Variables Estimator

In macro-economics there is a wide range of models that consists of a system of
equations that simultaneously determine a number of endogenous variables. Consider,
for example, a demand and a supply equation, both depending upon prices, and an
equilibrium condition that says that demand and supply should be equal. The resulting
system simultaneously determines quantities and prices and it can typically not be
said that prices determine quantities or quantities determine prices. An even simpler
example is the Keynesian model discussed in the previous section. It is becoming more
and more common that a researcher’s interest is in just one of the equations in such a
system and that the remaining equations are not explicitly formulated. In this case there
is need for an estimator that can consistently estimate such an equation, even if one or
more of the explanatory variables are not exogenous. In this section we shall consider
such an estimator, where we use a motivating example from micro-economics.
5.3.1 Estimation with a Single Endogenous Regressor
and a Single Instrument

Suppose we explain an individual’s log wage yi by a number of personal characteristics,
x1i , as well as the number of hours person i is working (x2i ) by means of a linear model

yi = x1i
β1 + x2i β2 + εi .

(5.33)

We know from Chapter 2 that this model has no interpretation unless we make some
assumptions about εi . Otherwise, we could just set β1 and β2 to arbitrary values and
deﬁne εi such that the equality in (5.33) holds for every observation. The most common

132

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

interpretation so far is that (5.33) describes the conditional expectation or the best linear
approximation of yi given x1i and x2i . This requires us to impose that
E{εi x1i } = 0

(5.34)

E{εi x2i } = 0,

(5.35)

which are the necessary conditions for consistency of the OLS estimator. As soon as
we relax any of these conditions, the model no longer corresponds to the conditional
expectation of yi given x1i and x2i .
In the above wage equation, εi includes all unobservable factors that affect a person’s wage, including things like ‘ability’ or ‘intelligence’. Typically, it is argued that
the number of hours a person is working partly also depends on these unobserved characteristics. If this is the case, OLS is consistently estimating the conditional expected
value of a person’s wage given, among other things, how many hours he or she is
working, but not consistently estimating the causal effect of hours of work. That is, the
OLS estimate for β2 would reﬂect the difference in the expected wage of two arbitrary
persons with the same observed characteristics in x1i , but working x2 and x2 + 1 hours,
respectively. It does not, however, measure the expected wage difference if an arbitrary
person (for some exogenous reason) decides to increase his hours from x2 to x2 + 1.
The reason is that in the ﬁrst interpretation the unobservable factors affecting a person’s wage are not assumed to be constant across the two persons, while in the second
interpretation the unobservables are kept unchanged. Put differently, when we interpret
the model as a conditional expectation, the ceteris paribus condition only refers to the
included variables, while for a causal interpretation it also includes the unobservables
(omitted variables) in the error term.
Quite often, coefﬁcients in a regression model are interpreted as measuring causal
effects. In such cases, it makes sense to discuss the validity of conditions like (5.34)
and (5.35). If E{εi x2i } = 0, we say that x2i is endogenous (with respect to the causal
effect β2 ). For micro-economic wage equations, it is often argued that many explanatory
variables are potentially endogenous, including education level, union status, sickness,
industry, and marital status. To illustrate this, it is not uncommon (for USA data) to
ﬁnd that expected wages are about 10% higher if a person is married. Quite clearly, this
is not reﬂecting the causal effect of being married, but the consequence of differences
in unobservable characteristics of married and unmarried people.
If it is no longer imposed that E{εi x2i } = 0, the OLS method produces a biased
and inconsistent estimator for the parameters in the model. The solution requires an
alternative estimation method. To derive a consistent estimator, it is necessary that
we make sure that our model is statistically identiﬁed. This means that we need to
impose additional assumptions; otherwise the model is not identiﬁed and any estimator
is necessarily inconsistent. To see this, let us go back to the conditions (5.34)–(5.35).
These conditions are so-called moment conditions, conditions in terms of expectations
(moments) that are implied by the model. These conditions should be sufﬁcient to
identify the unknown parameters in the model. That is, the K parameters in β1 and β2
should be such that the following K equalities hold:

β1 − x2i β2 )x1i } = 0
E{(yi − x1i

E{(yi −


x1i
β1

− x2i β2 )x2i } = 0.

(5.36)
(5.37)

THE INSTRUMENTAL VARIABLES ESTIMATOR

133

When estimating the model by OLS we impose these conditions on the estimator
through the corresponding sample moments. That is, the OLS estimator b = (b1 , b2 )
for β = (β1 , β2 ) is solved from
N
1 

(y − x1i
b1 − x2i b2 )x1i = 0
N i=1 i

(5.38)

N
1 

(y − x1i
b1 − x2i b2 )x2i = 0.
N i=1 i

(5.39)

In fact, these are the ﬁrst order conditions for the minimization of the least squares
criterion. The number of conditions exactly equals the number of unknown parameters,
so that b1 and b2 can be solved uniquely from (5.38) and (5.39). However, as soon as
(5.35) is violated, condition (5.39) drops out and we can no longer solve for b1 and
b2 . This means that β1 and β2 are no longer identiﬁed.
To identify β1 and β2 in the more general case, we need at least one additional
moment condition. Such a moment condition is usually derived from the availability
of an instrument or instrumental variable. An instrumental variable z2i , say, is a
variable that can be assumed to be uncorrelated with the models error εi but correlated
with the endogenous regressor x2i .6 If such an instrument can be found, condition
(5.37) can be replaced by

E{(yi − x1i
β1 − x2i β2 )z2i } = 0.

(5.40)

Provided this moment condition is not a combination of the other ones (z2i is not a
linear combination of x1i s), this is sufﬁcient to identify the K parameters β1 and β2 .
The instrumental variables estimator β̂IV can then be solved from
N
1 

β̂1,I V − x2i β̂2,IV )x1i = 0
(y − x1i
N i=1 i

(5.41)

N
1 

(y − x1i
β̂1,I V − x2i β̂2,IV )z2i = 0.
N i=1 i

(5.42)

The solution can be determined analytically and leads to the following expression for
the IV estimator
 N
−1 N


zi xi
zi yi ,
(5.43)
β̂IV =
i=1

i=1



, x2i ) and zi = (x1i
, z2i ). Clearly, if z2i = x2i , this expression reduces
where xi = (x1i
to the OLS estimator.
6

The assumption that the instrument is correlated with x2i is needed for identiﬁcation. If there would be
no correlation the additional moment does not provide any (identifying) information on β2 .

134

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

Under assumptions (5.36) and (5.40) and some regularity conditions, the instrumental
variables estimator is consistent and asymptotically normal. The most important of
these regularity conditions is that the K × K matrix
plim

N
1  
z x = zx
N i=1 i i

is ﬁnite and invertible. A necessary condition for this is that the instrument z2i is
correlated with x2i and not a linear combination of the elements in x1i . The asymptotic
covariance matrix of β̂IV depends upon the assumptions we make about the distribution
of εi . In the standard case where εi is IID(0, σ 2 ), independent of zi , it can be shown that
√
−1
N (β̂IV − β) → N(0, σ 2 (xz zz
zx )−1 ),
(5.44)
where the symmetric K × K matrix
N
1  
zz ≡ plim
zz
N i=1 i i

is assumed to be invertible, and zx = xz
. Nonsingularity of zz requires that there
is no multicollinearity among the K elements in the vector zi . In ﬁnite samples we can
estimate the covariance matrix of β̂IV by

 N
−1  N
−1
N



V̂ {β̂IV } = σ̂ 2 
xi zi
zi zi
zi xi  ,
(5.45)
i=1

i=1

i=1

where σ̂ 2 is a consistent estimator for σ 2 based upon the residual sum of squares,
for example,
N
1 
σ̂ 2 =
(y − xi β̂IV )2 .
(5.46)
N i=1 i
As in the least squares case, it is possible to adjust for degrees of freedom and divide
by N − K rather than N.
The problem for the practitioner is that it is sometimes far from obvious which variables could act as appropriate instruments. In the above example we need a variable
that is correlated with hours of work x2i but uncorrelated with the unobserved ‘ability’
factors that are included in εi . It can be argued that variables relating to the composition of one’s family may serve as instrumental variables. Another problem is that
standard errors of instrumental variables estimators are typically quite high compared
to those of the OLS estimator. The most important reason for this is that instrument
and regressor have a low correlation; see Wooldridge (2002, Subsection 5.2.6) for
additional discussion.
It is important to realize that the assumptions captured in the moment conditions are
identifying. That is, they cannot be tested statistically. The only case where the moment
conditions are partially testable is when there are more conditions than actually needed
for identiﬁcation. In this case, one can test the so-called overidentifying restrictions,

THE INSTRUMENTAL VARIABLES ESTIMATOR

135

without, however, being able to specify which of the moment conditions corresponds
to these restrictions (see below).
Keeping in mind the above, the endogeneity of x2i can be tested provided we assume
that the instrument z2i is valid. Hausman (1978) proposes to compare the OLS and IV
estimators for β. Assuming E{εi zi } = 0, the IV estimator is consistent. If, in addition,
E{εi x2i } = 0, the OLS estimator is also consistent and should differ from the IV one
by sampling error only. A computationally attractive version of the Hausman test for
endogeneity (often referred to as the Durbin–Wu–Hausman test) can be based upon
a simple auxiliary regression. First, estimate a regression explaining x2i from x1i and
z2i , and save the residuals, say v̂i . This is the reduced form equation. Next, add the
residuals to the model of interest and estimate

yi = x1i
β1 + x2i β2 + v̂i γ + ei

by OLS. This reproduces7 the IV estimator for β1 and β2 , but also produces an estimate
for γ . If γ = 0, x2i is exogenous. Consequently, we can easily test the endogeneity
of x2i by performing a standard t-test on γ = 0 in the above regression. Note that the
endogeneity test requires the assumption that the instrument is valid and therefore does
not help to determine which identifying moment condition, E{εi x2i } = 0 or E{εi z2i } =
0, is appropriate.
5.3.2 Back to the Keynesian Model

The problem for the practitioner is thus to ﬁnd suitable instruments. In most cases, this
means that somehow our knowledge of economic theory has to be exploited. In a complete simultaneous equations model (which speciﬁes relationships for all endogenous
variables), this problem can be solved because any exogenous variable in the system
that is not included in the equation of interest can be used as an instrument. More
precisely, any exogenous variable that has an effect on the endogenous regressor can
be used as an instrument, provided it is excluded from the equation that is estimated.8
Information on this is obtained from the reduced form for the endogenous regressor.
For the Keynesian model, this implies that investments It provide a valid instrument
for income Yt . The resulting instrumental variable estimator is then given by
β̂IV =

T

t=1



1
It

1 Yt

−1 T


t=1

1
It

Ct ,

(5.47)

which we can solve for β̂2,IV as
T

(It − I¯)(Ct − C̄)
,
¯
t=1 (It − I )(Yt − Ȳ )

β̂2,IV = t=1
T

(5.48)

where I¯, C̄, and Ȳ denote the sample averages.
While the estimates for β1 and β2 will be identical to the IV estimates, the standard errors will not be
appropriate; see Davidson and MacKinnon (1993, Section 7.9) or Wooldridge (2002, Section 6.2).
8
This explains why choosing instruments can be interpreted as imposing exclusion restrictions.
7

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

136

An alternative way to see that the estimator (5.48) works, is to start from (5.25) and
take the covariance with our instrument It on both sides of the equality sign. This gives
cov{Ct , It } = β2 cov{Yt , It } + cov{εt , It }.

(5.49)

Because the last term in this equality is zero (It is assumed to be exogenous) and
cov{Yt , It } = 0, we can solve β2 from this as
β2 =

cov{It , Ct }
.
cov{It , Yt }

(5.50)

This relationship suggests an estimator for β2 by replacing the population covariances
by their sample counterparts. This gives the instrumental variables estimator we have
seen above:

(1/T ) Tt=1 (It − I¯)(Ct − C̄)
.
(5.51)
β̂2,IV =

(1/T ) Tt=1 (It − I¯)(Yt − Ȳ )
Consistency follows directly from the general result that sample moments converge to
population moments.

5.3.3 Back to the Measurement Error Problem

The model is given by
yt = β1 + β2 xt + εt ,
where (as an interpretation) yt denotes savings and xt denotes observed disposable
income, which equals true disposable income plus a random measurement error. The
presence of this measurement error induces correlation between xt and εt .
Given this model, no obvious instruments arise. In fact, this is a common problem
in models with measurement errors due to inaccurate recording. The task is to ﬁnd an
observed variable that is (1) correlated with income xt , but (2) not correlated with ut ,
the measurement error in income (nor with εt ). If we can ﬁnd such a variable, we can
apply instrumental variables estimation. Mainly due to the problem of ﬁnding suitable
instruments, the problem of measurement error is often ignored in empirical work.

5.3.4 Multiple Endogenous Regressors

If more than one explanatory variable is considered to be endogenous, the dimension
of x2i is increased accordingly and the model reads


yi = x1i
β1 + x2i
β2 + ε i .

To estimate this equation we need an instrument for each element in x2i . This means
that if we have ﬁve endogenous regressors we need at least ﬁve different instruments.

ILLUSTRATION: ESTIMATING THE RETURNS TO SCHOOLING

137

Denoting the instruments by the vector z2i , the instrumental variables estimator can
again be written as in (5.43),

β̂IV =

N

i=1

−1
zi xi

N


zi yi ,

i=1





where now xi = (x1i
, x2i
) and zi = (x1i
, z2i
).
It is sometimes convenient to refer to the entire vector zi as the vector of instruments.
If a variable in xi is assumed to be exogenous, we do not need to ﬁnd an instrument
for it. Alternatively and equivalently, this variable is used as its own instrument. This
means that the vector of exogenous variables x1i is included in the K-dimensional
vector of instruments zi . If all the variables are exogenous, zi = xi and we obtain the
OLS estimator, where ‘each variable is instrumented by itself’.
In a simultaneous equations context, the exogenous variables from elsewhere in the
system are candidate instrumental variables. The so-called ‘order condition for identiﬁcation’ (see Hayashi, 2000, Section 3.3, or Greene, 2003, Section 15.3) essentially
says that sufﬁcient instruments should be available in the system. If, for example,
there are ﬁve exogenous variables in the system that are not included in the equation
of interest, we can have up to ﬁve endogenous regressors. If there is only one endogenous regressor, we have ﬁve different instruments to choose from. It is also possible and advisable to estimate more efﬁciently by using all the available instruments
simultaneously. This is discussed in Section 5.5. First, however, we shall discuss
an empirical illustration concerning the estimation of the causal effect of schooling
on earnings.

5.4

Illustration: Estimating the Returns to Schooling

It is quite clear that, on average, people with more education have higher wages. It is
less clear, however, whether this positive correlation reﬂects a causal effect of schooling, or that individuals with a greater earnings capacity have chosen for more years of
schooling. If the latter possibility is true, the OLS estimates on the returns to schooling
simply reﬂect differences in unobserved characteristics of working individuals and an
increase in a person’s schooling due to an exogenous shock will have no effect on
this person’s wage. The problem of estimating the causal effect of schooling upon
earnings has therefore attracted substantive attention in the literature; see Card (1999)
for a survey.
Most studies are based upon the human capital earnings function, which says that
wi = β1 + β2 Si + β3 Ei + β4 Ei2 + εi ,
where wi denotes the log of individual earnings, Si denotes years of schooling and Ei
denotes years of experience. In the absence of information on actual experience, Ei
is sometimes replaced by ‘potential experience’, measured as age i − Si − 6, assuming
people start school at the age of 6. This speciﬁcation is usually augmented with additional explanatory variables that one wants to control for, like regional, gender and

138

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

racial dummies. In addition, it is sometimes argued that the returns to education vary
across individuals. With this in mind, let us reformulate the wage equation as
wi = zi β + γi Si + ui
= zi β + γ Si + εi ,

(5.52)

where εi = ui + (γi − γ )Si , and zi includes all observable variables (except Si ), including the experience variables and a constant. It is assumed that E{εi zi } = 0. The
coefﬁcient γ has the interpretation of the average return to (an additional year of)
schooling E{γi } = γ and is our parameter of interest. In addition, we specify a reduced
form for Si as
(5.53)
Si = zi π + vi ,
where E{vi zi } = 0. This reduced form is simply a best linear approximation of Si
and does not necessarily have an economic interpretation. OLS estimation of β and
γ in (5.52) is consistent only if E{εi Si } = E{εi vi } = 0. This means that there are no
unobservable characteristics that both affect a person’s choice of schooling and his
(later) earnings.
As discussed in Card (1995), there are different reasons why schooling may be
correlated with εi . An important one is ‘ability bias’ (see Griliches, 1977). Suppose
that some individuals have unobserved characteristics (ability) that enable them to
get higher earnings. If these individuals also have above average schooling levels,
this implies a positive correlation between εi and vi and an OLS estimator that is
upward biased. Another reason why εi and vi may be correlated is the existence of
measurement error in the schooling measure. As discussed in Subsection 5.2.2 this
induces a negative correlation between εi and vi and, consequently a downward bias
in the OLS estimator for γ . Finally, if the individual speciﬁc returns to schooling (γi )
are higher for individuals with low levels of schooling, the unobserved component
(γi − γ )Si will be negatively correlated with Si , which, again, induces a downward
bias in the OLS estimator.
In the above formulation there are no instruments available for schooling as all
potential candidates are included in the wage equation. Put differently, the number of
moment conditions in
E{εi zi } = E{(wi − zi β − γ Si )zi } = 0
is one short to identify β and γ . However, if we can think of a variable in zi (z2i ,
say) that affects schooling but not wages, this variable can be excluded from the wage
equation so as to reduce the number of unknown parameters by one, thereby making
the model exactly identiﬁed. In this case the instrumental variables estimator for9 β
and γ , using z2i as an instrument, is a consistent estimator.
A continuing discussion in labour economics is the question which variable can
legitimately serve as an instrument. Typically, an instrument is thought of as a variable
that affects the costs of schooling (and thus the choice of schooling) but not earnings.
There is a long tradition of using family background variables, e.g. parents’ education,
9

Note that z2i is excluded from the wage equation so that the element in β corresponding to z2i is set
to zero.

ILLUSTRATION: ESTIMATING THE RETURNS TO SCHOOLING

139

as instrument. As Card (1999) notes, the interest in family background is driven by the
fact that children’s schooling choices are highly correlated with the characteristics of
their parents. More recently, institutional factors of the schooling system are exploited
as potential instruments. For example, Angrist and Krueger (1991) use an individual’s
quarter of birth as an instrument for schooling. Using an extremely large data set of
men born from 1930 to 1959 they ﬁnd that people with birth dates earlier in the year
have slightly less schooling than those born later in the year. Assuming that quarter
of birth is independent of unobservable taste and ability factors, it can be used as an
instrument to estimate the returns to schooling. In a more recent paper, Card (1995)
uses the presence of a nearby college as an instrument that can validly be excluded
from the wage equation. Students who grow up in an area without a college face a
higher cost of college education, while one would expect that higher costs, on average,
reduce the years of schooling, particularly in low-income families.
In this section we use data10 on 3010 men taken from the US National Longitudinal
Survey of Young Men, also employed in Card (1995). In this panel survey, a group
of individuals is followed since 1966 when they were aged 14–24, and interviewed
in a number of consecutive years. The labour market information that we use covers
1976. In this year, the average years of schooling in this sample is somewhat more
than 13 years, with a maximum of 18. Average experience in 1976, when this group
of men was between 24 and 34 years old, is 8.86 years, while the average hourly raw
wage is $5.77.
Table 5.1 reports the results of an OLS regression of an individual’s log hourly
wage upon years of schooling, experience and experience squared and three dummy
variables indicating whether the individual was black, lived in a metropolitan area
(SMSA) and lived in the south. The OLS estimator implies estimated average returns
to schooling of approximately 7.4% per year.11 The inclusion of additional variables,
like region of residence in 1966 and family background characteristics in some cases
signiﬁcantly improved the model but hardly affected the coefﬁcients for the variables
reported in Table 5.1 (see Card, 1995), so that we shall continue with this fairly simple
speciﬁcation.
Table 5.1 Wage equation estimated by OLS
Dependent variable: log(wage)
Variable

Estimate

Standard error

t-ratio

constant
schooling
exper
exper 2
black
smsa
south

4.7337
0.0740
0.0836
−0.0022
−0.1896
0.1614
−0.1249

0.0676
0.0035
0.0066
0.0003
0.0176
0.0156
0.0151

70.022
21.113
12.575
−7.050
−10.758
10.365
−8.259

s = 0.374 R 2 = 0.2905 R̄ 2 = 0.2891 F = 204.93
10
11

Available in SCHOOLING.
Because the dependent variable is in logs, a coefﬁcient of 0.074 corresponds to a relative difference of
approximately 7.4%; see Chapter 3.

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

140

Table 5.2 Reduced form for schooling, estimated by OLS
Dependent variable: schooling
Variable

Estimate

Standard error

t-ratio

constant
age
age 2
black
smsa
south
lived near college

−1.8695
1.0614
−0.0188
−1.4684
0.8354
−0.4597
0.3471

4.2984
0.3014
0.0052
0.1154
0.1093
0.1024
0.1070

−0.435
3.522
−3.386
−12.719
7.647
−4.488
3.244

s = 2.5158 R 2 = 0.1185 R̄ 2 = 0.1168 F = 67.29

If schooling is endogenous then experience and its square are by construction also
endogenous, given that age is not a choice variable and therefore unambiguously exogenous. This means that our linear model may suffer from three endogenous regressors
so that we need (at least) three instruments. For experience and its square, age and
age squared are obvious candidates. As discussed above, for schooling the solution
is less trivial. Card (1995) argues that the presence of a nearby college in 1966 may
provide a valid instrument. A necessary (but not sufﬁcient) condition for this is that
college proximity in 1966 affects the schooling variable, conditional upon the other
exogenous variables. To see whether this is the case, we estimate a reduced form,
where schooling is explained by age and age squared, the three dummy variables from
the wage equation and a dummy indicating whether an individual lived near a college
in 1966. The results, by OLS, are reported in Table 5.2. Recall that this reduced form
is not an economic or causal model to explain schooling choice. It is just a statistical
reduced form, corresponding to the best linear approximation of schooling.
The fact that the lived near college dummy is signiﬁcant in this reduced form is
reassuring. It indicates that, ceteris paribus, students who lived near a college in 1966
have on average 0.35 years more schooling. Recall that a necessary condition for a
valid instrument was that the candidate instrument is correlated with schooling but not
a linear combination of the other variables in the model. The crucial condition for a
valid instrument, viz. that it is uncorrelated with the error term in the wage equation,
cannot be tested. It would only be possible to test for such a correlation if we have a
consistent estimator for β and γ ﬁrst, but we can only ﬁnd a consistent estimator if
we impose that our instrument is valid. The validity of instruments can only be tested,
to some extent, if the model is overidentiﬁed; see Section 5.5 below. In this case we
thus need to trust economic arguments, rather than statistical ones, to rely upon the
instrument that is chosen.
Using age, age squared and the lived near college dummy as instruments for experience, experience squared and schooling,12 we obtain the estimation results reported
in Table 5.3. The estimated returns to schooling are over 13% with a relatively large
standard error of somewhat more than 5%. While the estimate is substantially higher
than the OLS one, its inaccuracy is such that this difference could just be due to sampling error. Nevertheless, the value of the IV estimate is fairly robust to changes in the
12

Although the formulation suggests otherwise, it is not the case that instruments have a one-to-one correspondence with the endogenous regressors. Implicitly, all instruments are jointly used for all variables.

ILLUSTRATION: ESTIMATING THE RETURNS TO SCHOOLING

Table 5.3

141

Wage equation estimated by IV

Dependent variable: log(wage)
Variable

Estimate

Standard error

t-ratio

constant
schooling
exper
exper 2
black
smsa
south

4.0656
0.1329
0.0560
−0.0008
−0.1031
0.1080
−0.0982

0.6085
0.0514
0.0260
0.0013
0.0774
0.0050
0.0288

6.682
2.588
2.153
−0.594
−1.333
2.171
−3.413

Instruments: age, age 2 , lived near college
used for: exper, exper 2 and schooling

speciﬁcation (for example, the inclusion of regional indicators or family background
variables). The fact that the IV estimator suffers from such large standard errors is due
to the fairly low correlation between the instruments and the endogenous regressors.
This is reﬂected in the R 2 of the reduced form for schooling, which is only 0.1185.13
While in general the instrumental variables estimator is less accurate than the OLS
estimator (which may be inconsistent), the loss in efﬁciency is particularly large if the
instruments are only weakly correlated with the endogenous regressors.
Table 5.3 does not report any goodness-of-ﬁt statistics. The reason is that there is
no unique deﬁnition of an R 2 or adjusted R 2 if the model is not estimated by ordinary
least squares. More importantly, the fact that we estimate the model by instrumental
variables methods indicates that goodness-of-ﬁt is not what we are after. Our goal was
to consistently estimate the causal effect of schooling upon earnings and that is exactly
what instrumental variables is trying to do. Again this reﬂects that the R 2 plays no
role whatsoever in comparing alternative estimators.
If college proximity is to be a valid instrument for schooling it has to be the case
that it has no direct effect on earnings. As with most instruments, this is a point of
discussion (see Card, 1995). For example, it is possible that families that place a strong
emphasis on education choose to live near a college, while children of such families
have a higher ‘ability’ or are more motivated to achieve labour market success (as
measured by earnings). Unfortunately, as said before, the current, exactly identiﬁed,
speciﬁcation does not allow us to test the validity of the instruments.
The fact that the IV estimate of the returns to schooling is higher than the OLS
one suggests that OLS underestimates the true causal effect of schooling. This is at
odds with the most common argument against the exogeneity of schooling, namely
‘ability bias’, but in line with the more recent empirical studies on the returns to
schooling (including, for example, Angrist and Krueger, 1991). The downward bias
of OLS could be due to measurement error, or – as argued by Card (1995) – to the
possibility that the true returns to schooling vary across individuals, negatively related
to schooling.
13

The R 2 s for the reduced forms for experience and experience squared (not reported) are both larger
than 0.60.

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

142

5.5

The Generalized Instrumental Variables Estimator

In Section 5.3 we considered the linear model where for each explanatory variable
exactly one instrument was available, which could equal the variable itself if it was
assumed exogenous. In this section we generalize this by allowing the use of an
arbitrary number of instruments.
5.5.1 Multiple Endogenous Regressors with an Arbitrary Number
of Instruments

Let us, in general, consider the following model
yi = xi β + εi ,

(5.54)

where xi is of dimension K. The OLS estimator is based upon the K moment conditions
E{εi xi } = E{(yi − xi β)xi } = 0.
More generally, let us assume that there are R instruments available in the vector zi ,
which may overlap with xi . The relevant moment conditions are then given by the
following R restrictions
E{εi zi } = E{(yi − xi β)zi } = 0.

(5.55)

If R = K we are back in the previous situation and the instrumental variables estimator
can be solved from the sample moment conditions
N
1 
(y − xi β̂IV )zi = 0
N i=1 i

and we obtain


β̂IV =

N


−1
zi xi

i=1

N


zi yi .

i=1

If the model is written in matrix notation
y = Xβ + ε
and the matrix Z is the N × R matrix of values for the instruments, this instrumental
variables estimator can also be written as
β̂IV = (Z  X)−1 Z  y.

(5.56)

If R > K there are more instruments than regressors. In this case it is not possible to
solve for an estimate of β by replacing (5.55) with its sample counterpart. The reason
for this is that there would be more equations than unknowns. Instead of dropping

THE GENERALIZED INSTRUMENTAL VARIABLES ESTIMATOR

143

instruments (and losing efﬁciency), one therefore chooses β in such a way that the R
sample moments
N
1 
(y − xi β)zi
N i=1 i
are as close as possible to zero. This is done by minimizing the following quadratic form
QN (β) =

N
1 
(y − xi β)zi
N i=1 i


WN


N
1 

(y − xi β)zi ,
N i=1 i

(5.57)

where WN is an R × R positive deﬁnite symmetric matrix. This matrix is a weighting
matrix and tells us how much weight to attach to which (linear combinations of the)
sample moments. In general it may depend upon the sample size N, because it may
itself be an estimate. For the asymptotic properties of the resulting estimator for β, the
probability limit of WN , denoted W = plim WN , is important. This matrix W should
be positive deﬁnite and symmetric. Using matrix notation for convenience, we can
rewrite (5.57) as




1 
1 
Z (y − Xβ) WN
Z (y − Xβ) .
QN (β) =
(5.58)
N
N
Differentiating this with respect to β (see Appendix A) gives as ﬁrst order conditions:
−2X ZWN Z  y + 2X ZWN Z  Xβ̂IV = 0,
which in turn implies

X ZWN Z  y = X ZWN Z  Xβ̂IV .

(5.59)

This is a system with K equations and K unknown elements in β̂IV , where X Z is
of dimension K × R and Z  y is R × 1. Provided the matrix X Z is of rank K, the
solution to (5.59) is
β̂IV = (X ZWN Z  X)−1 X ZWN Z  y,
(5.60)
which, in general, depends upon the weighting matrix WN .
If R = K the matrix X Z is square and (by assumption) invertible. This allows us
to write
β̂IV = (Z  X)−1 WN−1 (X Z)−1 X ZWN Z  y
= (Z  X)−1 Z  y,
which corresponds to (5.56), the weighting matrix being irrelevant. In this situation,
the number of moment conditions is exactly equal to the number of parameters to be
estimated. One can think of this as a situation where β is ‘exactly identiﬁed’ because
we have just enough information (i.e. moment conditions) to estimate β. An immediate
consequence of this is that the minimum of (5.58) is zero, implying that all samples
moments can be set to zero by choosing β appropriately. That is, QN (β̂IV ) is equal to

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

144

zero. In this case β̂IV does not depend upon WN and the same estimator is obtained
regardless of the choice of weighting matrix.
If R < K, the number of parameters to be estimated exceeds the number of moment
conditions. In this case β is ‘underidentiﬁed’ (not identiﬁed) because there is insufﬁcient information (i.e. moment conditions) from which to estimate β uniquely. Technically, this means that the inverse in (5.60) does not exist and an inﬁnite number
of solutions satisfy the ﬁrst order conditions in (5.59). Unless we can come up with
additional moment conditions, this identiﬁcation problem is fatal in the sense that no
consistent estimator for β exists. Any estimator is necessarily inconsistent.
If R > K, then the number of moment conditions exceeds the number of parameters
to be estimated, and so β is ‘overidentiﬁed’ because there is more information than is
necessary to obtain a consistent estimate of β. In this case we have a range of estimators
for β, corresponding to alternative choices for the weighting matrix WN . As long as
the weighting matrix is (asymptotically) positive deﬁnite, the resulting estimators are
all consistent for β. The idea behind the consistency result is that we are minimizing a
quadratic loss function in a set of sample moments that asymptotically converge to the
corresponding population moments, while these population moments are equal to zero
for the true parameter values. This is the basic principle behind the so-called method
of moments, which will be discussed in more detail in the next section.
Different weighting matrices WN lead to different consistent estimators with generally different asymptotic covariance matrices. This allows us to choose an optimal
weighting matrix that leads to the most efﬁcient instrumental variables estimator. It
can be shown that the optimal weighting matrix is proportional to the inverse of the
covariance matrix of the sample moments. Intuitively, this means that sample moments
with a small variance which – consequently – provide accurate information about the
parameters in β, get more weight in estimation than the sample moments with a large
variance. Essentially, this is the same idea as the weighted least squares approach
discussed in Chapter 4, albeit that the weights now reﬂect different sample moments
rather than different observations.
Of course the covariance matrix of the sample moments
N
1 
εz
N i=1 i i

depends upon the assumptions we make about εi and zi . If, as before, we assume that
εi is IID(0, σ 2 ) and independent of zi , the asymptotic covariance matrix of the sample
moments is given by
N
1  
σ 2 zz = σ 2 plim
zz.
N i=1 i i
Consequently, an optimal weighting matrix is obtained as

opt
WN

=

N
1  
zz
N i=1 i i

−1
=

1 
ZZ
N

−1

,

THE GENERALIZED INSTRUMENTAL VARIABLES ESTIMATOR

145

and the resulting IV estimator is
β̂IV = (X Z(Z  Z)−1 Z  X)−1 X Z(Z  Z)−1 Z  y.

(5.61)

This is the expression that is found in most textbooks (see, e.g. Greene, 2003,
Section 15.5). The estimator is sometimes referred to as the generalized instrumental
variables estimator (GIVE). It is also known as the two-stage least squares or 2SLS
estimator (see below). If εi is heteroskedastic or exhibits autocorrelation, the optimal
weighting matrix should be adjusted accordingly. How this is done follows from the
general discussion in the next section.
The asymptotic distribution of β̂IV is given by
√
−1
N(β̂IV − β) → N(0, σ 2 (xz zz
zx )−1 ),
which is the same expression as given in Section 5.3. The only difference is in the
dimensions of the matrices xz and zz . An estimator for the covariance matrix is
easily obtained by replacing the asymptotic limits with their small sample counterparts.
This gives
V̂ {β̂IV } = σ̂ 2 (X Z(Z  Z)−1 Z  X)−1
(5.62)
where the estimator for σ 2 is obtained from the IV-residuals ε̂i = yi − xi β̂IV as
σ̂ 2 =

N
1  2
ε̂ .
N i=1 i

5.5.2 Two-stage Least Squares and the Keynesian Model Again

The estimator in (5.61) is often used in the context of a simultaneous equations system
and then has the name of the two-stage least squares (2SLS) estimator. Essentially,
this interpretation says that the same estimator can be obtained in two steps, both of
which can be estimated by least squares. In the ﬁrst step the reduced form is estimated
by OLS (that is: a regression of the endogenous regressors upon all instruments). In
the second step the original structural equations are estimated by OLS, while replacing
all endogenous variables on the right hand side with their predicted values from the
reduced form.
To illustrate this, let the reduced form of the k-th explanatory variable be given by
(in vector notation)
xk = Zπk + vk .
OLS in this equation produces predicted values x̂k = Z(Z  Z)−1 Z  xk . If xk is a column
in Z we will automatically have that x̂k = xk . Consequently, the matrix of explanatory
variables in the second step can be written as X̂ which has the columns x̂k , k =
1, . . . , K, where
X̂ = Z(Z  Z)−1 Z  X.
The OLS estimator in the second step is thus given by
β̂IV = (X̂ X̂)−1 X̂ y,

(5.63)

146

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

which can easily be shown to be identical to (5.61). The advantage of this approach is
that the estimator can be computed using standard OLS software. In the second step
OLS is applied to the original model where all endogenous regressors are replaced by
their predicted values on the basis of the instruments.14 It should be stressed, although
this is often overlooked, that the second stage does not automatically provide the correct
standard errors (see Maddala, 2001, Section 9.6, for details).
The use of X̂ also allows us to write the generalized instrumental variables estimator
in terms of the standard formula in (5.56) if we redeﬁne our matrix of instruments. If
we use the K columns of X̂ as instruments in the standard formula (5.56) we obtain
β̂IV = (X̂ X)−1 X̂ y,
which is identical to (5.61). It shows that one can also interpret X̂ as the matrix of
instruments (which is sometimes done).
To go back to our Keynesian model, let us now assume that the economy includes a
government and a private sector, with government expenditures Gt and private investment It , both of which are assumed exogenous. The deﬁnition equation now reads
Yt = Ct + Gt + It .
This implies that both Gt and It are now valid instruments to use for income Yt in the
consumption function. Although it is possible to deﬁne simple IV estimators similar
to (5.51) using either Gt or It as instrument, the most efﬁcient estimator uses both
instruments simultaneously. The generalized instrumental variables estimator is thus
given by
β̂IV = (X Z(Z  Z)−1 Z  X)−1 X Z(Z  Z)−1 Z  y,
where the rows in Z, X and y are given by zt = (1, Gt , It ), xt = (1, Yt ) and yt = Ct ,
respectively.
5.5.3 Speciﬁcation Tests

The results on consistency and the asymptotic distribution of the generalized instrumental variables estimator are based on the assumption that the model is correctly
speciﬁed. As the estimator is only based on the model’s moment conditions, it is
required that the moment conditions are correct. It is therefore important to test whether
the dataare consistent with these moment conditions. In the ‘exactly identiﬁed’ case,
(1/N ) i ε̂i zi = 0 by construction, regardless of whether or not the population moment
conditions are true. Consequently, one cannot derive a useful test from the corresponding sample moments. Put differently, these K = R identifying restrictions are not
testable. However, if β is overidentiﬁed
it is clear that only K (linear combinations)

of the R elements in (1/N ) i ε̂i zi are set equal to zero. If the population moment

conditions are true one would expect that the elements in the vector (1/N ) i ε̂i zi
are all sufﬁciently close to zero (as they should converge to zero asymptotically). This
14

Note that the predicted values from the instruments rather than the instruments themselves should be
included in the equation of interest to replace the endogenous regressors.

THE GENERALIZED INSTRUMENTAL VARIABLES ESTIMATOR

147

provides a basis for a test of the model speciﬁcation. It can be shown that (under
(5.55)) the statistic (based on the GIV estimator with optimal weighting matrix)15

ξ = N QN (β̂IV ) =

N

i=1

 
ε̂i zi

σ̂

2

N

i=1

−1 
zi zi

N



ε̂i zi

(5.64)

i=1

has an asymptotic Chi-squared distribution with R − K degrees of freedom. Note that
the degrees of freedom equals the number of moment conditions minus the number
of parameters to be estimated.
 This is the case because only R − K of the sample
moment conditions (1/N ) i ε̂i zi are free on account of the K restrictions implied by
the ﬁrst order conditions for β̂IV in (5.59). A test based on (5.64) is usually referred to
as an overidentifying restrictions test or Sargan test. A simple way to compute (5.64)
is by taking N times the R 2 of an auxiliary regression of IV residuals ε̂i upon the full
set of instruments zi . If the test rejects, the speciﬁcation of the model is rejected in the
sense that the sample evidence is inconsistent with the joint validity of all R moment
conditions. Without additional information it is not possible to determine which of the
moments are incorrect, i.e. which of the instruments are invalid.16
If a subset of the instruments is known to satisfy the moment conditions, it is
possible to test the validity of the remaining instruments or moments provided that
the model is identiﬁed on the basis on the non-suspect instruments. Assume that
R1 ≥ K moment conditions are non-suspect and we want to test the validity of the
remaining R − R1 moment conditions. To compute the test statistic, estimate the
model using all R instruments and compute the overidentifying restrictions test statistic ξ . Next, estimate the model using only the R1 non-suspect instruments. Typically, this will lead to a lower value for the overidentifying restrictions test, ξ1 , say.
The test statistic to test the suspect moment conditions is easily obtained as ξ − ξ1 ,
which, under the null hypothesis, has an approximate Chi-squared distribution with
R − R1 degrees of freedom (see Hayashi, 2000, Section 3.6). In the special case that
R1 = K, this test reduces to the overidentifying restrictions test in (5.64), and the test
statistic is independent of the choice of the R1 instruments that are said to be nonsuspect.

5.5.4 Weak Instruments

A problem with instrumental variables estimation that has received considerable attention recently is that of ‘weak instruments’. The problem is that the properties of the IV
estimator can be very poor, and the estimator can be severely biased, if the instruments
exhibit only weak correlation with the endogenous regressor(s). In these cases, the normal distribution provides a very poor approximation to the true distribution of the IV
estimator, even if the sample size is large. To illustrate the problem, let us consider the
IV estimator for the case of a single regressor and a constant. If x̃i = xi − x̄ denotes
15
16

Note that all terms involving N cancel out.
Suppose a pub allows you to buy three beers but pay for only two. Can you tell which of the three beers
is the free one?

148

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

the regressor values in deviation from the sample mean, and similarly for ỹi and z̃i ,
the IV estimator for β2 can be written as (compare (5.51))

(1/N ) N
i=1 z̃i ỹi
.
β̂2,IV =
N
(1/N ) i=1 z̃i x̃i
If the instrument is valid (and under weak regularity conditions), the estimator is
consistent and converges to
cov{zi , yi }
.
β2 =
cov{zi , xi }
However, if the instrument is not correlated with the regressor, the denominator of this
expression is zero. In this case, the IV estimator is inconsistent and the asymptotic
distribution of β̂2,IV deviates substantially from a normal distribution. The instrument
is weak if there is some correlation between zi and xi , but not enough to make the
asymptotic normal distribution provide a good approximation in ﬁnite (potentially very
large) samples. For example, Bound, Jaeger and Baker (1995) show that part of the
results of Angrist and Krueger (1991), who use quarter of birth to instrument for
schooling in a wage equation, suffer from the weak instruments problem. Even with
samples of more than 300 000 (!) individuals, the IV estimator appeared to be unreliable
and misleading.
To ﬁgure out whether you have weak instruments, it is useful to examine the reduced
form regression and evaluate the explanatory power of the additional instruments that
are not included in the equation of interest. Consider the linear model with one endogenous regressor

β1 + x2i β2 + εi ,
yi = x1i
where E{x1i εi } = 0 and where additional instruments z2i (for x2i ) satisfy E{z2i εi } = 0.
The appropriate reduced form is given by


π1 + z2i
π2 + vi .
x2i = x1i

If π2 = 0, the instruments in z2i are irrelevant and the IV estimator is inconsistent. If
π2 is ‘close to zero’, the instruments are weak. The value of the F-statistic for π2 = 0 is
a measure for the information content contained in the instruments. Staiger and Stock
(1997) provide a theoretical analysis of the properties of the IV estimator and provide
some guidelines about how large the F-statistic should be for the IV estimator to have
good properties. As a simple rule-of-thumb, Stock and Watson (2003, Chapter 10)
suggest that you do not have to worry about weak instruments if the F-statistic exceeds
10. In any case, it is a good practice to compute and present the F-statistic of the reduced
form in empirical work. If the instruments in z2i are insigniﬁcant in the reduced form,
you should not put much conﬁdence in the IV results. If you have many instruments
available, it may be a good strategy to use the most relevant subset and drop the
‘weak’ ones.

5.6

The Generalized Method of Moments

The approaches sketched above are special cases of an approach proposed by Hansen
(1982), usually referred to as the Generalized Method of Moments (GMM). This

THE GENERALIZED METHOD OF MOMENTS

149

approach estimates the model parameters directly from the moment conditions that
are imposed by the model. These conditions can be linear in the parameters (as in the
above examples) but quite often are nonlinear. To enable identiﬁcation, the number of
moment conditions should be at least as large as the number of unknown parameters.
The present section provides a fairly intuitive discussion of the Generalized Method of
Moments. First, in the next subsection, we start with a motivating example that illustrates how economic theory can imply nonlinear moment conditions. An extensive, not
too technical, overview of GIVE and GMM methodology is given in Hall (1993).
5.6.1 Example

The following example is based on Hansen and Singleton (1982). Consider an individual agent who maximizes the expected utility of current and future consumption
by solving
 S


s
max Et
δ U (Ct+s ) ,
(5.65)
s=0

where Ct+s denotes consumption in period t + s, U (Ct+s ) is the utility attached to
this consumption level, which is discounted by the discount factor δ (0 < δ ≤ 1), and
where Et is the expectation operator conditional upon all information available at time
t. Associated with this problem is a set of intertemporal budget constraints of the form
Ct+s + qt+s = wt+s + (1 + rt+s )qt+s−1 ,

(5.66)

where qt+s denotes ﬁnancial wealth at the end of period t + s, rt+s is the return on
ﬁnancial wealth (invested in a portfolio of assets), and wt+s denotes labour income.
The budget constraint thus says that labour income plus asset income should be spent
on consumption Ct+s or saved in qt+s . This maximization problem is hard to solve analytically. Nevertheless, it is still possible to estimate the unknown parameters involved
through the ﬁrst order conditions. The ﬁrst order conditions of (5.65) subject to (5.66)
imply that
Et {δU  (Ct+1 )(1 + rt+1 )} = U  (Ct ),
where U  is the ﬁrst derivative of U. The right-hand side of this equality denotes the
marginal utility of one additional dollar consumed today, while the left-hand side
gives the expected marginal utility of saving this dollar until the next period (so
that it becomes 1 + rt+1 dollars) and consuming it then. Optimality thus implies that
(expected) marginal utilities are equalized.
As a next step, we can rewrite this equation as
Et

δU  (Ct+1 )
(1 + rt+1 ) − 1 = 0.
U  (Ct )

(5.67)

Essentially, this is a (conditional) moment condition which can be exploited to estimate
the unknown parameters if we make some assumption about the utility function U. We
can do this by transforming (5.67) into a set of unconditional moment conditions.

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

150

Suppose zt is included in the information set. This implies that zt does not provide
any information about the expected value of
δU  (Ct+1 )
(1 + rt+1 ) − 1
U  (Ct )
so that it also holds that17
E

δU  (Ct+1 )
(1 + rt+1 ) − 1 zt
U  (Ct )

= 0.

(5.68)

Thus we can interpret zt as a vector of instruments, valid by the assumption of optimal
behaviour (rational expectations) of the agent. For simplicity, let us assume that the
utility function is of the power form, that is
U (C) =

C 1−γ
,
1−γ

where γ denotes the (constant) coefﬁcient of relative risk aversion, where higher values
of γ correspond to a more risk averse agent. Then we can write (5.68) as
 

Ct+1 −γ
(5.69)
(1 + rt+1 ) − 1 zt = 0.
E
δ
Ct
We now have a set of moment conditions which identify the unknown parameters
δ and γ , and given observations on Ct+1 /Ct , rt+1 and zt allow us to estimate them
consistently. This requires an extension of the earlier approach to nonlinear functions.
5.6.2 The Generalized Method of Moments

Let us, in general, consider a model that is characterized by a set of R moment conditions as
E{f (wt , zt , θ )} = 0,
(5.70)
where f is a vector function with R elements, θ is a K-dimensional vector containing all
unknown parameters, wt is a vector of observable variables that could be endogenous
or exogenous, and zt is the vector of instruments. In the example of the previous
subsection wt = (Ct+1 /Ct , rt+1 ); in the linear model of Section 5.5 wt = (yt , xt ).
To estimate θ we take the same approach as before and consider the sample equivalent of (5.70) given by
T
1
gT (θ ) ≡
f (wt , zt , θ ).
(5.71)
T t=1
If the number of moment conditions R equals the number of unknown parameters K,
it would be possible to set the R elements in (5.71) to zero and to solve for θ to obtain
17

We use the general result that E{x1 |x2 } = 0 implies that E{x1 g(x2 )} = 0 for any function g (see
Appendix B).

THE GENERALIZED METHOD OF MOMENTS

151

a unique consistent estimator. If f is nonlinear in θ an analytical solution may not be
available. If the number of moment conditions is less than the number of parameters,
the parameter vector θ is not identiﬁed. If the number of moment conditions is larger,
we cannot solve uniquely for the unknown parameters by setting (5.71) to zero. Instead,
we choose our estimator for θ such that the vector of sample moments is as close as
possible to zero, in the sense that a quadratic form in gT (θ ) is minimized. That is,
min QT (θ ) = min gT (θ ) WT gT (θ ),
θ

θ

(5.72)

where, as before, WT is a positive deﬁnite matrix with plim WT = W . The solution
to this problem provides the generalized method of moments or GMM estimator θ̂ .
Although we cannot obtain an analytical solution for the GMM estimator in the general
case, it can be shown that it is consistent and asymptotically normal under some weak
regularity conditions. The heuristic argument presented for the generalized instrumental
variables estimator in the linear model extends to this more general setting. Because
sample averages converge to population means, which are zero for the true parameter
values, an estimator chosen to make these sample moments as close to zero as possible
(as deﬁned by (5.72)) will converge to the true value and will thus be consistent.
In practice, the GMM estimator is obtained by numerically solving the minimization
problem in (5.72), for which a variety of algorithms is available; see Wooldridge (2002,
Section 12.7) or Greene (2003, Appendix E) for a general discussion.
As before, different weighting matrices WT lead to different consistent estimators
with different asymptotic covariance matrices. The optimal weighting matrix, which
leads to the smallest covariance matrix for the GMM estimator, is the inverse of
the covariance matrix of the sample moments. In the absence of autocorrelation it is
given by
W opt = (E{f (wt , zt , θ )f (wt , zt , θ ) })−1 .
In general this matrix depends upon the unknown parameter vector θ , which presents
a problem which we did not encounter in the linear model. The solution is to adopt
a multi-step estimation procedure. In the ﬁrst step we use a suboptimal choice of
WT which does not depend upon θ (for example the identity matrix), to obtain a ﬁrst
consistent estimator θ̂[1] , say. Then, we can consistently estimate the optimal weighting
matrix by18

−1
T
1
opt

WT =
f (wt , zt , θ̂[1] )f (wt , zt , θ̂[1] )
.
(5.73)
T t=1
In the second step one obtains the asymptotically efﬁcient (optimal) GMM estimator
θ̂GMM . Its asymptotic distribution is given by
√

18

T (θ̂GMM − θ ) → N(0, V ),

(5.74)

If there is autocorrelation in f (wt , zt , θ ) up to a limited order, the optimal weighting matrix can be
estimated using a variant of the Newey-West estimator discussed in Section 5.1; see Greene (2003, Subsection 18.3.4).

152

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

where the asymptotic covariance matrix V is given by
V = (DW opt D  )−1 ,

(5.75)

where D is the K × R derivative matrix
D=E

∂f (wt , zt , θ )
.
∂θ 

(5.76)

Intuitively, the elements in D measure how sensitive a particular moment is with
respect to small changes in θ . If the sensitivity with respect to a given element in θ
is large, small changes in this element lead to relatively large changes in the objective
function QT (θ ) and the particular element in θ is relatively accurately estimated. As
usual, the covariance matrix in (5.75) can be estimated by replacing the population
moments in D and W opt by their sample equivalents, evaluated at θ̂GMM .
The great advantage of the generalized method of moments is that (1) it does not
require distributional assumptions, like normality, (2) it can allow for heteroskedasticity
of unknown form and (3) it can estimate parameters even if the model cannot be solved
analytically from the ﬁrst order conditions. Unlike most of the cases we discussed
before, the validity of the instruments in zt is beyond doubt if the model leads to
a conditional moment restriction (as in (5.67)) and zt is in the conditioning set. For
example, if at time t the agent maximizes expected utility given all publicly available
information then any variable that is observed (to the agent) at time t provides a
valid instrument.
Unfortunately, there is considerable evidence that the asymptotic distribution in
(5.74) often provides a poor approximation to the sampling distribution of the GMM
estimator in sample sizes that are typically encountered in empirical work, see, for
example, Hansen, Heaton and Yaron (1996). In a recent paper, Stock and Wright
(2003) explore the distribution theory for GMM estimators when some or all of the
parameters are weakly identiﬁed, paying particular attention to variants of the nonlinear
model discussed in Subsection 5.6.1. The weak instruments problem, discussed in
Subsection 5.5.4, appears also relevant in a general GMM setting.
Finally, we consider the extension of the overidentifying restrictions test to nonlinear models. Following the intuition from the linear model, it would be anticipated that if
the population moment conditions E{f (wt , zt , θ )} = 0 are correct then gT (θ̂GMM ) ≈ 0.
Therefore, the sample moments provide a convenient test of the model speciﬁcation.
Provided that all moment conditions are correct, the test statistic
ξ = T gT (θ̂GMM ) WT gT (θ̂GMM ),
opt

opt

where θ̂GMM is the optimal GMM estimator and WT is the optimal weighting matrix
given in (5.73) (based upon a consistent estimator for θ ), is asymptotically Chi-squared
distributed with R − K degrees of freedom. Recall that for the exactly identiﬁed case,
there are zero degrees of freedom and there is nothing that can be tested.
In Section 5.7 we present an empirical illustration using GMM to estimate intertemporal asset pricing models. In Section 10.5 we shall consider another example of GMM,
where it is used to estimate a dynamic panel data model. First, we consider a few
simple examples.

THE GENERALIZED METHOD OF MOMENTS

153

5.6.3 Some Simple Examples

As a very simple example, assume we are interested in estimating the population mean
µ of a variable yi on the basis of a sample of N observations (i = 1, 2, . . . , N ). The
moment condition of this ‘model’ is given by
E{yi − µ} = 0,
with sample equivalent

N
1 
(y − µ).
N i=1 i

By setting this to zero and solving for µ we obtain a method of moments estimator
µ̂ =

N
1 
y,
N i=1 i

which is just the sample average.
If we consider the linear model
yi = xi β + εi
again, with instrument vector zi , the moment conditions are
E{εi zi } = E{(yi − xi β)zi } = 0.
If εi is i.i.d. the optimal GMM is the instrumental variables estimator given in (5.43)
and (5.56). More generally, the optimal weighting matrix is given by
W opt = (E{εi2 zi zi })−1 ,
which is estimated unrestrictedly as

opt
WN

=

N
1  2 
ε̂ z z
N i=1 i i i

−1
,

where ε̂i is the residual based upon an initial consistent estimator. When it is imposed
that εi is i.i.d. we can simply use

opt
WN

=

N
1  
zz
N i=1 i i

The K × R derivative matrix is given by
D = E{xi zi },

−1
.

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

154

which we can estimate consistently by
DN =

N
1  
xz.
N i=1 i i

In general, the covariance matrix of the optimal GMM or GIV estimator β̂ for β can
be estimated as

V̂ {β̂} =

N

i=1

−1
xi zi

N



ε̂i2 zi zi

i=1

N


−1
zi xi

.

(5.77)

i=1

This estimator generalizes (5.62) just as the White heteroskedasticity consistent covariance matrix generalizes the standard OLS expression. Thus, the general GMM set-up
allows for heteroskedasticity of εi automatically.

5.7

Illustration: Estimating Intertemporal Asset
Pricing Models

In the recent ﬁnance literature, the GMM framework is frequently used to estimate and
test asset pricing models. An asset pricing model, for example the CAPM discussed
in Section 2.7, should explain the variation in expected returns for different risky
investments. Because some investments are more risky than others, investors may
require compensation for bearing this risk by means of a risk premium. This leads to
variation in expected returns across different assets. An extensive treatment of asset
pricing models and their link with the generalized method of moments is provided
in Cochrane (2001).
In this section we consider the consumption-based asset pricing model. This model
is derived from the framework sketched in Subsection 5.6.1 by introducing a number
of alternative investment opportunities for ﬁnancial wealth. Assume that there are J
alternative risky assets available that the agent can invest in, with returns rj,t+1 , j =
1, . . . , J , as well as a riskless asset with certain return rf,t+1 . Assuming that the agent
optimally chooses his portfolio of assets, the ﬁrst order conditions of the problem now
imply that
Et {δU  (Ct+1 )(1 + rf,t+1 )} = U  (Ct )
Et {δU  (Ct+1 )(1 + rj,t+1 )} = U  (Ct ),

j = 1, . . . , J.

This says that the expected marginal utility of investing one additional dollar in asset j is
equal for all assets and equal to the marginal utility of consuming this additional dollar
today. Assuming power utility, as before, and restricting attention to unconditional
expectations19 the ﬁrst order conditions can be rewritten as
19

This means that we restrict attention to moments using instrument zt = 1 only.

ILLUSTRATION: ESTIMATING INTERTEMPORAL ASSET PRICING MODELS





Ct+1
E δ
Ct

Ct+1
E δ
Ct

−γ

−γ

155


(1 + rf,t+1 ) = 1

(5.78)


(rj,t+1 − rf,t+1 ) = 0,

j = 1, . . . , J,

(5.79)

where the second set of conditions is written in terms of excess returns, i.e. returns in
excess of the riskfree rate.
Let us, for convenience, deﬁne the intertemporal marginal rate of substitution
mt+1 (θ ) ≡ δ

Ct+1
Ct

−γ

,

where θ contains all unknown parameters. In ﬁnance, mt+1 (θ ) is often referred to as
a stochastic discount factor or a pricing kernel (see Campbell, Lo and MacKinlay,
1997, Chapter 8, or Cochrane, 2001). Alternative asset pricing models are described
by alternative speciﬁcations for the pricing kernel mt+1 (θ ). To see how a choice for
mt+1 (θ ) provides a model that describes expected returns, we use that for two arbitrary
random variables E{xy} = cov{x, y} + E{x}E{y} (see Appendix B), from which it
follows that
cov{mt+1 (θ ), rj,t+1 − rf,t+1 } + E{mt+1 (θ )}E{rj,t+1 − rf,t+1 } = 0.
This allows us to write
E{rj,t+1 − rf,t+1 } = −

cov{mt+1 (θ ), rj,t+1 − rf,t+1 }
E{mt+1 (θ )}

,

(5.80)

which says that the expected excess return on any asset j is equal to a risk premium
that depends linearly upon the covariance between the asset’s excess return and the
stochastic discount factor. Knowledge of mt+1 (θ ) allows us to describe or explain the
cross-sectional variation of expected returns across different assets. In the consumptionbased model, this tells us that assets that have a positive covariance with consumption
growth (and thus make future consumption more volatile) must promise higher expected
returns to induce investors to hold them. Conversely, assets that covary negatively with
consumption growth can offer expected returns that are lower than the riskfree rate.20
The moment conditions in (5.78)–(5.79) can be used to estimate the unknown parameters δ and γ . In this section we use data21 that cover monthly returns over the period
February 1959–November 1993. The basic assets we consider are ten portfolios of
stocks, maintained by the Center for Research in Security Prices at the University of
Chicago. These portfolios are size-based, which means that portfolio 1 contains the
10% smallest ﬁrms listed at the New York Stock Exchange, while portfolio 10 contains the 10% largest ﬁrms that are listed. The riskless return is approximated by the
20

For example, you may reward a particular asset if it delivers a high return in the situation where you
happen to get unemployed.
21
The data are available in PRICING.

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

156

monthly return on a 3 month US Treasury Bill, which does not vary much over time.
For consumption we use total US personal consumption expenditures on nondurables
and services. It is assumed that the model is valid for a representative agent, whose
consumption corresponds to this measure of aggregate per capita consumption. Data
on size-based portfolios are used because most asset pricing models tend to underpredict the returns on the stocks of small ﬁrms. This is the so-called small ﬁrm effect
(see Banz, 1981; or Campbell, Lo and MacKinlay, 1997, p. 211).
With one riskless asset and ten risky portfolios, (5.78)–(5.79) provide 11 moment
conditions with only two parameters to estimate. These parameters can be estimated
using the identity matrix as a suboptimal weighting matrix, using the efﬁcient two-step
GMM estimator that was presented above, or using a so-called iterated GMM estimator.
This estimator has the same asymptotic properties as the two-step one, but is sometimes
argued to have a better small-sample performance. It is obtained by computing a new
optimal weighting matrix using the two-step estimator, and using this to obtain a next
estimator, θ̂[3] , say, which in turn is used in a weighting matrix to obtain θ̂[4] . This
procedure is repeated until convergence.
Table 5.4 presents the estimation results on the basis of the monthly returns from
February 1959–November 1993, using one-step GMM (using the identity matrix as
weighting matrix) and iterated GMM.22 The γ estimates are huge and rather imprecise. For the iterated GMM procedure, for example, a 95% conﬁdence interval for γ
based upon the approximate normal distribution is as large as (−9.67, 124.47). The
estimated risk aversion coefﬁcients of 57.4 and 91.4 are much higher than what is
considered economically plausible. This ﬁnding illustrates the so-called equity premium puzzle (see Mehra and Prescott, 1985), which reﬂects that the high risk premia
on risky assets (equity) can only be explained in this model if agents are extremely
risk averse (compare Campbell, Lo and MacKinlay, 1997, Section 8.2). If we look
at the overidentifying restrictions tests, we see, somewhat surprisingly, that they do
not reject the joint validity of the imposed moment conditions. This means that the
consumption-based asset pricing model is statistically not rejected by the data. This
is solely due to the high imprecision of the estimates. Unfortunately this is only a
statistical satisfaction and certainly does not imply that the model is economically
valuable. The gain in efﬁciency from the use of the optimal weighting matrix appears
to be fairly limited with standard errors that are only up to 20% smaller than for the
one-step method.
Table 5.4 GMM estimation results consumption-based asset pricing model

δ
γ
ξ (df = 9)
22

One-step GMM
Estimate
s.e.

Iterated GMM
Estimate
s.e.

0.6996
91.4097

(0.1436)
(38.1178)

0.8273
57.3992

(0.1162)
(34.2203)

4.401

(p = 0.88)

5.685

(p = 0.77)

For the one-step GMM estimator the standard errors and the overidentifying restrictions test are computed
in a non-standard way. The formulae given in the text do not apply because the optimal weighting matrix
is not used. See Cochrane (2001, Chapter 11) for the appropriate expressions. The estimation results in
Table 5.4 were obtained using RATS 5.1.

CONCLUDING REMARKS

157

0.14
0.12

Mean excess return

0.10
0.08
0.06
0.04
0.02

−0.02

0.00
0.00

−0.02

Figure 5.1

0.02

0.04

0.06

0.08

0.10

0.12

0.14

Predicted mean excess return

Actual versus predicted mean excess returns of size-based portfolios

To investigate the economic value of the above model, it is possible to compute so-called pricing errors (compare Cochrane, 1996). One can directly compute
the average expected excess return according to the model, simply by replacing the
population moments in (5.80) by the corresponding sample moments and using the
estimated values for δ and γ . On the other hand the average excess returns on
asset j can be directly computed from the data. In Figure 5.1, we plot the average
excess returns against the predicted average excess returns, as well as a 45◦ line.
We do this for the one-step estimator only because, as argued by Cochrane (1996),
this estimator minimizes the vector of pricing errors of the 11 assets. Points on the
45◦ line indicate that the average pricing error is zero. Points above this line indicate that the return of the corresponding asset is underpredicted by the model. The
ﬁgure conﬁrms our idea that the economic performance of the model is somewhat
disappointing. Clearly, the model is unable to fully capture the cross-sectional variation in expected excess returns. The two portfolios with the smallest ﬁrms have the
highest mean excess return and are both above the 45◦ line. The model apparently
does not solve the small-ﬁrm effect as the returns on these portfolios are underpredicted.
Cochrane (1996) also presents a range of alternative asset pricing models that are
estimated by GMM, which, in a number of cases, perform much better than the
simple consumption-based model discussed here. Marquering and Verbeek (1999)
extend the above model by including transaction costs and habit persistence in the
utility function.

5.8

Concluding Remarks

This chapter has discussed a variety of models that can be headed under the term
‘stochastic regressors’. Starting from a linear model with an endogenous regressor,

ENDOGENEITY, INSTRUMENTAL VARIABLES AND GMM

158

we discussed instrumental variables estimation. It was shown how instrumental variables estimation exploits different moment conditions compared to the OLS estimator.
If more moment conditions are imposed than unknown parameters, we can use a
generalized instrumental variables estimator, which can also be derived in a GMM
framework with an optimal weighting matrix. GMM was discussed in detail, with an
application to intertemporal asset pricing models. In dynamic models one usually has
the advantage that the choice of instruments is less suspect: lagged values can often be
assumed to be uncorrelated with current innovations. The big advantage of GMM is
that it can estimate the parameters in a model without having to solve the model analytically. That is, there is no need to write the model as y = something + error term.
All one needs is conditions in terms of expectations, which are often derived directly
from economic theory.

Exercises
Exercise 5.1 (Instrumental Variables)

Consider the following model
yi = β1 + β2 xi2 + β3 xi3 + εi ,

i = 1, . . . , N,

(5.81)

where (yi , xi2 , xi3 ) are observed and have ﬁnite moments, and εi is an unobserved
error term. Suppose this model is estimated by ordinary least squares. Denote the OLS
estimator by b.
a.

b.

What are the essential conditions required for unbiasedness of b? What are the
essential conditions required for consistency of b? Explain the difference between
unbiasedness and consistency.
Show how the conditions for consistency can be written as moment conditions (if
you have not done so already). Explain how a method of moments estimator can
be derived from these moment conditions. Is the resulting estimator any different
from the OLS one?

Now suppose that cov{εi , xi3 } = 0.
c. Give two examples of cases where one can expect a nonzero correlation between
a regressor, xi3 , and the error εi .
d. In this case, is it possible to still make appropriate inferences based on the OLS
estimator, while adjusting the standard errors appropriately?
e. Explain how an instrumental variable, zi , say, leads to a new moment condition
and, consequently, an alternative estimator for β.
f. Why does this alternative estimator lead to a smaller R 2 than the OLS one? What
does this say of the R 2 as a measure for the adequacy of the model?
g. Why can we not choose zi = xi2 as an instrument for xi3 , even if E{xi2 εi } = 0?
2
Would it be possible to use xi2
as an instrument for xi3 ?

EXERCISES

159

Exercise 5.2 (Returns to Schooling – Empirical)

Consider the data used in Section 5.4, as available in SCHOOLING. The purpose of this
exercise is to explore the role of parents’ education as instruments to estimate the
returns to schooling.
a. Estimate a reduced form for schooling, as reported in Table 5.2, but include
mother’s and father’s education levels, instead of the lived near college dummy.
What do these results indicate about the possibility of using parents’ education as
instruments?
b. Estimate the returns to schooling, on the basis of the same speciﬁcation as in
Section 5.4, using mother’s and father’s education as instruments (and age and
age-squared as instruments for experience and its square).
c. Test the overidentifying restriction.
d. Re-estimate the model using also the lived near college dummy and test the two
overidentifying restrictions.
e. Compare and interpret the different estimates on the returns to schooling from
Table 5.3, and parts b and d of this exercise.
Exercise 5.3 (GMM)

An intertemporal utility maximization problem gives the following ﬁrst order condition


Ct+1 −γ
Et δ
(1 + rt+1 ) = 1,
Ct
where Et denotes the expectation operator conditional upon time t information, Ct
denotes consumption in period t, rt+1 is the return on ﬁnancial wealth, δ is a discount
rate and γ is the coefﬁcient of relative risk aversion. Assume that we have a time
series of observations on consumption levels, returns and instrumental variables zt .
a. Show how the above condition can be written as a set of unconditional moment
conditions. Explain how we can estimate δ and γ consistently from these
moment conditions.
b. What is the minimum number of moment conditions that is required? What do we
(potentially) gain by having more moment conditions?
c. How can we improve the efﬁciency of the estimator for a given set of moment
conditions? In which case does this not work?
d. Explain what we mean by ‘overidentifying restrictions’. Is this a good or a
bad thing?
e. Explain how the overidentifying restrictions test is performed. What is the null
hypothesis that is tested? What do you conclude if the test rejects?

6

Maximum Likelihood
Estimation and
Speciﬁcation Tests

In the previous chapter we paid attention to the generalized method of moments. In
the GMM approach the model imposes assumptions about a number of expectations
(moments) that involve observable data and unknown coefﬁcients, which are exploited
in estimation. In this chapter we consider an estimation approach that typically makes
stronger assumptions, because it assumes knowledge of the entire distribution, not just
of a number of its moments. If the distribution of a variable yi conditional upon a
number of variables xi is known up to a small number of unknown coefﬁcients, we
can use this to estimate these unknown parameters by choosing them in such a way
that the resulting distribution corresponds as well as possible, in a way to be deﬁned
more precisely below, to the observed data. This is, somewhat loosely formulated, the
method of maximum likelihood.
In certain applications and models, distributional assumptions like normality are
commonly imposed because estimation strategies that do not require such assumptions
are complex or unavailable. If the distributional assumptions are correct, the maximum
likelihood estimator is, under weak regularity conditions, consistent and asymptotically
normal. Moreover, it fully exploits the assumptions about the distribution so that the
estimator is asymptotically efﬁcient. That is, alternative consistent estimators will have
an asymptotic covariance matrix that is at least as large (in a matrix sense) as that of
the maximum likelihood estimator.
This chapter starts with an introduction to maximum likelihood estimation.
Section 6.1 describes the approach starting with some simple examples and concluding
with some general results and discussion. Because the distributional assumptions
are typically crucial for the consistency and efﬁciency of the maximum likelihood
estimator, it is important to be able to test these assumptions. This is discussed
in Section 6.2, while Section 6.3 focuses on the implementation of the Lagrange

162

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

multiplier tests for particular hypotheses, mostly in the context in the linear regression
model. Section 6.4 explores the link with the generalized method of moments (GMM)
to introduce quasi-maximum likelihood estimation and to extend the class of Lagrange
multiplier tests to moment conditions tests. Knowledge of the issues in Section 6.1 is
crucial for understanding Chapter 7 and some speciﬁc sections of Chapters 8, 9 and 10.
The remaining sections of this chapter cover issues relating to speciﬁcation tests and
are somewhat more technical. They are a prerequisite for some speciﬁc sections of
Chapter 7 that can be skipped without loss of continuity. The material in Section 6.4
is used in Section 7.3 (count data models) and Section 8.10 (GARCH models).

6.1

An Introduction to Maximum Likelihood

The starting point of maximum likelihood estimation is the assumption that the (conditional) distribution of an observed phenomenon (the endogenous variable) is known,
except for a ﬁnite number of unknown parameters. These parameters will be estimated
by taking those values for them that give the observed values the highest probability,
the highest likelihood. The maximum likelihood method thus provides a means of
estimating a set of parameters characterizing a distribution, if we know, or assume we
know, the form of this distribution. For example, we could characterize the distribution
of some variable yi (for given xi ) as normal with mean β1 + β2 xi and variance σ 2 .
This would represent the simple linear regression model with normal error terms.
6.1.1 Some Examples

The principle of maximum likelihood is most easily introduced in a discrete setting
where yi only has a ﬁnite number of outcomes. As an example, consider a large pool,
ﬁlled with red and yellow balls. We are interested in the fraction p of red balls in
this pool. To obtain information on p, we take a random sample of N balls (and do
not look at all the other balls). Let us denote yi = 1 if ball i is red and yi = 0 if it is
1
not. Then
 it holds by assumption that P {yi = 1} = p. Suppose our sample contains
N1 = i yi red and N − N1 yellow balls. The probability of obtaining such a sample
(in a given order) is given by
P {N1 red balls, N − N1 yellow balls} = p N1 (1 − p)N−N1 .

(6.1)

The expression in (6.1), interpreted as a function of the unknown parameter p, is
referred to as the likelihood function. Maximum likelihood estimation for p implies
that we choose a value for p such that (6.1) is maximal. This gives the maximum
likelihood estimator p̂. For computational purposes it is often more convenient to
maximize the (natural) logarithm of (6.1), which is a monotone transformation. This
gives the loglikelihood function
log L(p) = N1 log(p) + (N − N1 ) log(1 − p).
1

(6.2)

We assume that sampling takes place with replacement. Alternatively, one can assume that the number
of balls in the pool is inﬁnitely large, such that previous draws do not affect the probability of drawing a
red ball.

AN INTRODUCTION TO MAXIMUM LIKELIHOOD

163

logl

−68.597

−133.067
.1

Figure 6.1

.2

.3

.4

.5
p

.6

.7

.8

.9

Sample loglikelihood function for N = 100 and N1 = 44

For a sample of size 100 with 44 red balls (N1 = 44), Figure 6.1 displays the loglikelihood function for values of p between 0.1 and 0.9. Maximizing (6.2) gives as ﬁrst
order condition
d log L(p)
N
N − N1
= 1−
= 0,
(6.3)
dp
p
1−p
which, solving for p, gives the maximum likelihood (ML) estimator
p̂ = N1 /N.

(6.4)

The ML estimator thus corresponds with the sample proportion of red balls, and probably also corresponds with your best guess for p based on the sample that was drawn.
In principle, we also need to check the second order condition to make sure that the
solution we have corresponds to a maximum, although in this case it is obvious from
Figure 6.1. This gives
d 2 log L(p)
N
N − N1
= − 21 −
< 0,
dp 2
p
(1 − p)2

(6.5)

showing, indeed, that we have found a maximum.
So the intuition of the maximum likelihood principle is as follows. From the
(assumed) distribution of the data (e.g. yi ), we determine the likelihood of observing
the sample that we happen to observe as a function of the unknown parameters that
characterize the distribution. Next, we choose as our maximum likelihood estimates
those values for the unknown parameters that give us the highest likelihood. It is
clear that this approach makes sense in the above example. The usefulness of the
maximum likelihood method is more general, as it can be shown that – under suitable

164

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

regularity conditions – the maximum likelihood estimator is generally consistent for the
true underlying parameters. The ML estimator has several other attractive properties,
which we shall discuss below.
As a next illustration, consider the simple regression model
yi = β1 + β2 xi + εi ,

(6.6)

where we make assumptions (A1)–(A4) from Chapter 2. These assumptions state that
εi has mean zero, is homoskedastic, has no autocorrelation and is independent of
all xi (i = 1, . . . , N ). While these assumptions imply that E{yi |xi } = β1 + β2 xi and
V {yi |xi } = σ 2 , they do not impose a particular distribution. To enable maximum likelihood estimation we thus need to augment the above assumptions with an assumption
about the shape of the distribution. The most common assumption is that εi is normal,
as in assumption (A5) from Chapter 2. We can summarize these assumptions by saying
that the error terms εi are normally and independently distributed (n.i.d.) with mean
zero and variance σ 2 , or εi ∼ NID(0, σ 2 ).
The probability of observing a particular outcome y for yi is, however, zero for any
y, because yi has a continuous distribution. Therefore the contribution of observation
i to the likelihood function is the value of the density function at the observed point
yi . For the normal distribution (see Appendix B) this gives
f (yi |xi ; β, σ 2 ) = √



1 (yi − β1 − β2 xi )2
exp −
,
2
σ2
2πσ 2
1

(6.7)

where β = (β1 , β2 ) . Because of the independence assumption, the joint density of
y1 , . . . , yN (conditional on X = (x1 , . . . , xN ) ) is given by
f (y1 , . . . , yN |X; β, σ 2 ) =

N


f (yi |xi ; β, σ 2 )

i=1


=

√

1
2πσ 2

N 
N



1 (yi − β1 − β2 xi )2
exp −
. (6.8)
2
σ2
i=1

The likelihood function is identical to the joint density function of y1 , . . . , yN but it
is considered as a function of the unknown parameters β, σ 2 . Consequently, we can
write down the loglikelihood function as
N

log L(β, σ 2 ) = −

N
1  (yi − β1 − β2 xi )2
log(2πσ 2 ) −
.
2
2 i=1
σ2

(6.9)

As the ﬁrst term in this expression does not depend upon β, it is easily seen that
maximizing (6.9) with respect to β1 and β2 corresponds to minimizing the residual sum
of squares S(β), as deﬁned in Section 2.1. That is, the maximum likelihood estimators
for β1 and β2 are identical to the OLS estimators. Denoting these estimators by β̂1 and
β̂2 , and deﬁning the residuals ei = yi − β̂1 − β̂2 xi , we can go on and maximize (6.9)

AN INTRODUCTION TO MAXIMUM LIKELIHOOD

165

with respect to σ 2 . Substituting the ML solutions for β1 and β2 and differentiating2
with respect to σ 2 we obtain the ﬁrst order condition
N

N 2π
1  ei2
−
+
= 0.
2 2πσ 2
2 i=1 σ 4

(6.10)

Solving this for σ 2 gives the maximum likelihood estimator for σ 2 as
σ̂ 2 =

N
1  2
e .
N i=1 i

(6.11)

This estimator is a consistent estimator for σ 2 . It does not, however, correspond to the
unbiased estimator for σ 2 that was derived from the OLS estimator (see Chapter 2),
given by
N

1
e2 ,
s2 =
N − K i=1 i
where K is the number of regressors (including the intercept). The difference lies in the
degrees of freedom correction in s 2 . Because s 2 is unbiased, the ML estimator σ̂ 2 will
be biased in ﬁnite samples. Asymptotically, (N − K)/N converges to 1 and the bias
disappears, so that the ML estimator is consistent, the degrees of freedom correction
being a small sample issue.
In this particular example the maximum likelihood estimator for β happens to reproduce the OLS estimator and consequently has the small sample properties of the OLS
estimator. The fact that the ML estimator for σ 2 deviates from the unbiased estimator
s 2 indicates that this is not a general result. In small samples the latter estimator has
better properties than the ML estimator. In many relevant cases, the ML estimator
cannot be shown to be unbiased and its small sample properties are unknown. This
means that in general the maximum likelihood approach can be defended only on
asymptotic grounds, the ML estimator being consistent and asymptotically efﬁcient.
Furthermore, it is typically not possible to analytically solve for the ML estimator,
except in a number of special cases (like those considered above).
If the error terms εi in this example are non-normal or heteroskedastic, the loglikelihood function given in (6.9) is incorrect, that is, does not correspond to the true
distribution of yi given xi . In such a case the estimator derived from maximizing the
incorrect loglikelihood function (6.9) is not the maximum likelihood estimator in a
strict sense, and there is no guarantee that it will have good properties. In some particular cases consistency can still be achieved by maximizing an incorrect likelihood
function, in which case it is common to refer to the estimator as a quasi-ML estimator.
This example illustrates this point, because the (quasi-)ML estimator for β equals the
OLS estimator b, which is consistent under much weaker conditions. Again this is not
a general result and it is not appropriate in general to rely upon such an argument to
2

We shall consider σ 2 as an unknown parameter, so that we differentiate with respect to σ 2 rather than σ .
The resulting estimator is invariant to this choice.

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

166

defend the use of maximum likelihood. Section 6.4 presents some additional discussion
on this issue.
6.1.2 General Properties

To deﬁne the maximum likelihood estimator in a more general situation, suppose that
interest lies in the conditional distribution of yi given xi . Let the density or probability
mass function be given by f (yi |xi ; θ ), where θ is a K-dimensional vector of unknown
parameters and assume that observations are mutually independent. In this situation
the joint density or probability mass function of the sample y1 , . . . , yN (conditional
upon X = (x1 , . . . , xN ) ) is given by3
f (y1 , . . . , yN |X; θ ) =

N


f (yi |xi ; θ ).

i=1

The likelihood function for the available sample is then given by
L(θ |y, X) =

N


Li (θ |yi , xi ) =

i=1

N


f (yi |xi ; θ ),

i=1

which is a function of θ . For several purposes it is convenient to employ the likelihood contributions, denoted by Li (θ |yi , xi ), which reﬂect how much observation i
contributes to the likelihood function. The maximum likelihood estimator θ̂ for θ is
the solution to
N

log Li (θ ),
(6.12)
max log L(θ ) = max
θ

θ

i=1

where log L(θ ) is the loglikelihood function, and for simplicity we dropped the other
arguments. The ﬁrst order conditions of this problem imply that


N

∂ log L(θ ) 
∂ log Li (θ ) 
(6.13)
 =
 = 0,


∂θ
∂θ
θ̂

i=1

θ̂

where |θ̂ indicates that the expression is evaluated at θ = θ̂ . If the loglikelihood function
is globally concave there is a unique global maximum and the maximum likelihood
estimator is uniquely determined by these ﬁrst order conditions. Only in special cases
the ML estimator can be determined analytically. In general, numerical optimization
is required (see Greene, 2003, Appendix E, for a discussion). Fortunately, for many
standard models efﬁcient algorithms are available in recent software packages.
For notational convenience, we shall denote the vector of ﬁrst derivatives of the
loglikelihood function, also known as the score vector, as
N

s(θ ) ≡
3

N

∂ log L(θ )  ∂ log Li (θ ) 
si (θ ),
=
≡
∂θ
∂θ
i=1
i=1

We use f (.) as generic notation for a (multivariate) density or probability mass function.

(6.14)

AN INTRODUCTION TO MAXIMUM LIKELIHOOD

167

which also deﬁnes the individual score contributions si (θ ). The ﬁrst order conditions
s(θ̂ ) =

N


si (θ̂ ) = 0

i=1

thus say that the K sample averages of the score contributions, evaluated at the ML
estimate θ̂ , should be zero.
Provided that the likelihood function is correctly speciﬁed, it can be shown under
weak regularity conditions that the maximum likelihood estimator:
1. is consistent for θ (plim θ̂ = θ );
2. is asymptotically efﬁcient (that is, asymptotically the ML estimator has the ‘smallest’ variance among all consistent asymptotically normal estimators);
3. is asymptotically normally distributed, according to
√
N (θ̂ − θ ) → N(0, V ),
(6.15)
where V is the asymptotic covariance matrix.
The covariance matrix V is determined by the shape of the loglikelihood function.
To describe it in the general case, we deﬁne the information in observation i as
 2

∂ log Li (θ )
¯
,
(6.16)
Ii (θ ) ≡ −E
∂θ ∂θ 
which is a symmetric K × K matrix. Loosely speaking, this matrix summarizes the
expected amount of information about θ contained in observation i. The average information matrix for a sample of size N is deﬁned as


N
1 
1 ∂ 2 log L(θ )
¯
IN (θ ) ≡
I (θ ) = −E
,
N i=1 i
N ∂θ ∂θ 

(6.17)

while the limiting information matrix is deﬁned as I (θ ) ≡ limN→∞ I¯N (θ ). In the
special case where the observations are independently and identically distributed, it
follows that I¯i (θ ) = IN (θ ) = I (θ ). Under appropriate regularity conditions, the asymptotic covariance matrix of the maximum likelihood estimator can be shown to equal
the inverse of the information matrix, that is
V = I (θ )−1 .

(6.18)

The term on the right-hand side of (6.17) is the expected value of the matrix of second
order derivatives, scaled by the number of observations, and reﬂects the curvature of the
loglikelihood function. Clearly, if the loglikelihood function is highly curved around
its maximum, the second derivative is large, the variance is small, and the maximum
likelihood estimator is relatively accurate. If the function is less curved, the variance
will be larger. Given the asymptotic efﬁciency of the maximum likelihood estimator,
the inverse of the information matrix I (θ )−1 provides a lower bound on the asymptotic

168

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

covariance matrix for any consistent asymptotically normal estimator for θ . The ML
estimator is asymptotically efﬁcient because it attains this bound, often referred to as
the Cramèr–Rao lower bound.
In practice the covariance matrix V can be estimated consistently by replacing the
expectations operator by a sample average and replacing the unknown coefﬁcients by
the maximum likelihood estimates. That is,

N
1  ∂ 2 log Li (θ ) 
V̂H = −
N i=1
∂θ ∂θ  θ̂

−1

,

(6.19)

where we take derivatives ﬁrst and in the result replace the unknown θ by θ̂ . The
sufﬁx H is used to stress that the estimator for V is based upon the Hessian matrix,
the matrix of second derivatives.
An alternative expression for the information matrix can be obtained from the result
that the matrix
Ji (θ ) ≡ E{si (θ )si (θ ) },
(6.20)
with si (θ ) deﬁned in (6.14), is identical to Ii (θ ), provided that the likelihood function is correctly speciﬁed. In Section 6.4, we shall return to the possibility that the
likelihood function is misspeciﬁed and that the matrices Ii (θ ) and Ji (θ ) are different.
For the moment, we shall use I (θ ) to denote the information matrix based on either
deﬁnition. The result in (6.20) indicates that V can also be estimated from the ﬁrst
order derivatives of the loglikelihood function, as

V̂G =

N
1 
s (θ̂ )si (θ̂ )
N i=1 i

−1

,

(6.21)

where the sufﬁx G reﬂects that the estimator employs the outer product of the gradients (ﬁrst derivatives). This estimator for V was suggested by Berndt, Hall, Hall and
Hausman (1974) and is sometimes referred to as the BHHH estimator. It is important to note that computation of the latter expression requires the individual likelihood
contributions. In general the two covariance matrix estimates V̂H and V̂G will not be
identical. The ﬁrst estimator typically has somewhat better properties in small samples.
To illustrate the maximum likelihood principle, Subsection 6.1.3 again considers
the simple example of the pool with balls, while Subsection 6.1.4 treats the linear
regression model with normal errors. Chapter 7 provides more interesting models that
typically require maximum likelihood estimation. The remainder of this chapter discusses issues relating to speciﬁcation and misspeciﬁcation tests. While this is not
without importance, it is somewhat more technical and some readers may prefer to
skip these sections on ﬁrst reading and continue with Chapter 7. Section 6.4 also discusses the relationship between GMM estimation and maximum likelihood estimation
in more detail and explains quasi-maximum likelihood estimation. This is relevant
for Section 7.3, where count data models are discussed, and for Section 8.10, where
models for conditional heteroskedasticity are presented.

AN INTRODUCTION TO MAXIMUM LIKELIHOOD

169

6.1.3 An Example (Continued)

To clarify the general formulae in the previous subsection let us reconsider the example
concerning the pool of red and yellow balls. In this model, the loglikelihood contribution of observation i can be written as
log Li (p) = yi log p + (1 − yi ) log(1 − p),
with a ﬁrst derivative

∂ log Li (p)
y
1 − yi
= i −
.
∂p
p
1−p

Note that the expected value of the ﬁrst derivative is zero, using E{yi } = p. The
negative of the second derivative is
−

y
1 − yi
∂ 2 log Li (p)
= i2 +
,
2
∂p
p
(1 − p)2

which has an expected value of


∂ 2 log Li (p)
E −
∂p 2


=

E{yi } 1 − E{yi }
1
1
1
+
= +
=
.
p2
(1 − p)2
p 1−p
p(1 − p)

From this it follows that the asymptotic variance of the maximum likelihood estimator
p̂ is given by V = p(1 − p) and we have that
√
N(p̂ − p) → N(0, p(1 − p)).
This result can be used to construct conﬁdence intervals or to test hypotheses. For
example, the hypothesis H0 : p = p0 can be tested using the test statistic
p̂ − p0
,
se(p̂)
where se(p̂) = p̂(1 − p̂)/N . Under the null hypothesis the test statistic has an asymptotic standard normal distribution. This is similar to the usual t-tests discussed in the
context of the linear model. A 95% conﬁdence interval is given by
p̂ − 1.96 se(p̂), p̂ + 1.96 se(p̂)
so that with a sample of 100 balls of which 44 are red (p̂ = 0.44), we can conclude
with 95% conﬁdence that p is between 0.343 and 0.537. When N = 1000 with 440
red balls, the interval reduces to (0.409, 0.471). In this particular application it is clear
that the normal distribution is an approximation based on large sample theory and will
never hold in small samples. In any ﬁnite sample, p̂ can only take a ﬁnite number
of different outcomes in the range [0, 1]. In fact, in this example the small sample
distribution of N1 = N p̂ is known to be binomial with parameters N and p, and this
result could be employed instead.

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

170

6.1.4 The Normal Linear Regression Model

In this subsection we consider the linear regression model with normal i.i.d. errors
(independent of all xi ). This is the model considered in Chapter 2 combined with
assumptions (A1)–(A5). Writing
yi = xi β + εi ,

εi ∼ NID(0, σ 2 ),

this imposes that (conditional upon the exogenous variables), yi is normal with mean
xi β and a constant variance σ 2 . Generalizing (6.9), the loglikelihood function for this
model can be written as
log L(β, σ 2 ) =

N


N

log Li (β, σ 2 ) = −

i=1

N
1  (yi − xi β)2
.
log(2πσ 2 ) −
2
2 i=1
σ2

(6.22)

The score contributions are given by


 
∂ log Li (β, σ 2 )

 
∂β

 
si (β, σ 2 ) = 
=
 ∂ log L (β, σ 2 )  
i

∂σ 2

(yi − xi β)
xi
σ2




,
1
1 (yi − xi β)2 
− 2+
2σ
2
σ4

while the maximum likelihood estimates β̂, σ̂ 2 will satisfy the ﬁrst order conditions
N

(yi − xi β̂)
xi = 0,
σ̂ 2
i=1

and

N

−

N
1  (yi − xi β̂)2
+
= 0.
2
2σ̂
2 i=1
σ̂ 4

It is easily veriﬁed that the solutions to these equations are given by
β̂ =

N

i=1

xi xi

−1 N

i=1

xi yi

and σ̂ 2 =

N
1 
(y − xi β̂)2 .
N i=1 i

The estimator for the vector of slope coefﬁcients is identical to the familiar OLS
estimator, while the estimator for the variance differs from the OLS value s 2 by dividing
through N rather than N − K.
To obtain the asymptotic covariance matrix of the maximum likelihood estimator
for β and σ 2 , we use that
Ii (β, σ 2 ) = E{si (β, σ 2 )si (β, σ 2 ) }.

SPECIFICATION TESTS

171

Using that for a normal distribution E{εi } = 0, E{εi2 } = σ 2 , E{εi3 } = 0 and E{εi4 } =
3σ 4 (see Appendix B), this expression can be shown to equal
Ii (β, σ 2 ) =

1
x x
σ2 i i

0

0

1
2σ 4

,

if we take expectations conditional upon xi . Using this, the asymptotic covariance
matrix is given by
−1
σ 2 xx
0
V = I (β, σ 2 )−1 =
,
0
2σ 4
where

N
1 
xi xi .
N→∞ N
i=1

xx = lim

From this it follows that β̂ and σ̂ 2 are asymptotically normally distributed according to
√

√
−1
N (β̂ − β) → N(0, σ 2 xx
)
N (σ̂ 2 − σ 2 ) → N(0, 2σ 4 ).

The fact that the information matrix is block diagonal implies that the ML estimators
for β and σ 2 are (asymptotically) independent. In ﬁnite samples, β̂ is approximately
normally distributed, with mean β, and with a covariance matrix that can be estimated as
−1
N

2

σ̂
xi xi
.
i=1

Note that this corresponds quite closely to the results that are familiar for the OLS
estimator.

6.2

Speciﬁcation Tests

6.2.1 Three Test Principles

On the basis of the maximum likelihood estimator a large number of alternative tests
can be constructed. Such tests are typically based upon one out of three different
principles: the Wald, the likelihood ratio or the Lagrange multiplier principle. Although
any of the three principles can be used to construct a test for a given hypothesis, each
of them has its own merits and advantages. The Wald test is used a number of times
in the previous chapters and is generally applicable to any estimator that is consistent
and asymptotically normal. The likelihood ratio (LR) principle provides an easy way to
compare two alternative nested models, while the Lagrange multiplier (LM) tests allow
one to test restrictions that are imposed in estimation. This makes the LM approach
particularly suited for misspeciﬁcation tests where a chosen speciﬁcation of the model is

172

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

tested for misspeciﬁcation in several directions (like heteroskedasticity, non-normality,
or omitted variables).
Consider again the general problem where we estimate a K-dimensional parameter
vector θ by maximizing the loglikelihood function, i.e.
max log L(θ ) = max
θ

θ

N


log Li (θ ).

i=1

Suppose that we are interested in testing one or more linear restrictions on the parameter
vector θ = (θ1 , . . . , θK ) . These restrictions can be summarized as H0 : Rθ = q for some
ﬁxed J -dimensional vector q, where R is a J × K matrix. It is assumed that the J
rows of R are linearly independent, so that the restrictions are not in conﬂict with each
other nor redundant. The three test principles can be summarized as follows:
1. Wald test. Estimate θ by maximum likelihood and check whether the difference
R θ̂ − q is close to zero, using its (asymptotic) covariance matrix. This is the idea
that underlies the well-known t- and F -tests.
2. Likelihood ratio test. Estimate the model twice: once without the restriction
imposed (giving θ̂ ) and once with the null hypothesis imposed (giving the
constrained maximum likelihood estimator θ̃ , where R θ̃ = q) and check whether
the difference in loglikelihood values log L(θ̂ ) − log L(θ̃ ) is signiﬁcantly different
from zero. This implies the comparison of an unrestricted and a restricted maximum
of log L(θ ).
3. Lagrange multiplier test. Estimate the model with the restriction from the null
hypothesis imposed (giving θ̃) and check whether the ﬁrst order conditions from
the general model are signiﬁcantly violated. That is, check whether ∂ log L(θ )/∂θ|θ̃
is signiﬁcantly different from zero.
While the three tests look at different aspects of the likelihood function, they are, in
general, asymptotically equivalent (that is: the test statistics have the same asymptotic
distribution, even if the null hypothesis is violated) and in a few very special cases
they even give the same numerical outcomes. In ﬁnite samples, the (actual) size and
power of the tests may differ (see Exercise 6.1). Most of the time, however, we will
choose the test that is most easily computed from the results that we have. For example,
the Wald test requires estimating the model without the restriction imposed, while the
Lagrange multiplier (LM) test requires only that the model is estimated under the null
hypothesis. As a result, the LM test may be particularly attractive when relaxing the
null hypothesis substantially complicates model estimation. It is also attractive when
the number of different hypotheses one wants to test is large, as the model has to
be estimated only once. The likelihood ratio test requires the model to be estimated
with and without the restriction, but, as we shall see, is easily computed from the
loglikelihood values.
The Wald test starts from the result that
√
N (θ̂ − θ ) → N(0, V ),

(6.23)

SPECIFICATION TESTS

173

from which it follows that the J -dimensional vector R θ̂ also has an asymptotic normal
distribution, given by (see Appendix B)
√
N (R θ̂ − Rθ ) → N(0, RV R  ).

(6.24)

Under the null hypothesis Rθ equals the known vector q, so that we can construct a
test statistic by forming the quadratic form
ξW = N (R θ̂ − q) [R V̂ R  ]−1 (R θ̂ − q),

(6.25)

where V̂ is a consistent estimator for V (see above). Under H0 this test statistic has a
Chi-squared distribution with J degrees of freedom, so that large values for ξW lead
us to reject the null hypothesis.
The likelihood ratio test is even simpler to compute, provided the model is estimated with and without the restrictions imposed. This means that we have two different
estimators: the unrestricted ML estimator θ̂ and the constrained ML estimator θ̃ ,
obtained by maximizing the loglikelihood function log L(θ ) subject to the restrictions Rθ = q. Clearly, maximizing a function subject to a restriction will not lead
to a larger maximum compared to the case without the restriction. Thus it follows
that log L(θ̂ ) − log L(θ̃ ) ≥ 0. If this difference is small, the consequences of imposing
the restrictions Rθ = q are limited, suggesting that the restrictions are correct. If the
difference is large, the restrictions are likely to be incorrect. The LR test statistic is
simply computed as
ξLR = 2[log L(θ̂ ) − log L(θ̃ )],
which, under the null hypothesis, has a Chi-squared distribution with J degrees of
freedom. This shows that if we have estimated two speciﬁcations of a model we
can easily test the restrictive speciﬁcation against the more general one by comparing
loglikelihood values. It is important to stress that the use of this test is only appropriate
if the two models are nested (see Chapter 3). An attractive feature of the test is that
it is easily employed when testing nonlinear restrictions and that the result is not
sensitive to the way in which we formulate these restrictions. In contrast, the Wald test
can handle nonlinear restrictions but is sensitive to the way they are formulated. For
example, it will matter whether we test θk = 1 or log θk = 0. See Gregory and Veal
(1985), Lafontaine and White (1986) or Phillips and Park (1988) for a discussion.

6.2.2 Lagrange Multiplier Tests

Some of the tests discussed in the previous chapters, like the Breusch–Pagan test
for heteroskedasticity, are Lagrange multiplier tests (LM tests). To introduce the
general idea of an LM test, suppose the null hypothesis restricts some elements in
the parameter vector θ to equal a set of given values. To stress this, let us write
θ  = (θ1 , θ2 ), where the null hypothesis now says that θ2 = q, where θ2 has dimension
J . The term ‘Lagrange multiplier’ comes from the fact that it is implicitly based upon

174

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

the value of the Lagrange multiplier in the constrained maximization problem. The
ﬁrst order conditions of the Lagrangian

H (θ, λ) =

N





log Li (θ ) − λ (θ2 − q) ,

(6.26)

i=1

yield the constrained ML estimator θ̃ = (θ̃1 , q  ) and λ̃. The vector λ̃ can be interpreted
as a vector of shadow prices of the restrictions θ2 = q. If the shadow prices are high,
we would like to reject the restrictions. If they are close to zero, the restrictions are
relatively ‘innocent’. To derive a test statistic we would therefore like to consider the
distribution of λ̃. From the ﬁrst order conditions of (6.26) it follows that

N
N


∂ log Li (θ ) 
si1 (θ̃ ) = 0
 =
∂θ1
θ̃
i=1
i=1

(6.27)


N
N


∂ log Li (θ ) 
=
si2 (θ̃ ),

∂θ
2
θ̃
i=1
i=1

(6.28)

and
λ̃ =

where the vector of score contributions si (θ ) is decomposed into the subvectors si1 (θ )
and si2 (θ ), corresponding to θ1 and θ2 , respectively. The result in (6.28) shows that the
vector of Lagrange multipliers λ̃ equals the vector of ﬁrst derivatives with respect to
the restricted parameters θ2 , evaluated at the constrained estimator θ̃ . Consequently,
the vector of shadow prices of the restrictions θ2 = q also has the interpretation of
measuring the extent to which the ﬁrst order conditions with respect to θ2 are violated,
if we evaluate them at the constrained estimates θ̃ = (θ̃1 , q  ) . As the ﬁrst derivatives are
also referred to as scores, the Lagrange multiplier test is also known as the score test.
To determine an appropriate test statistic, we exploit that it can be shown that the
sample average N −1 λ̃ is asymptotically normal with covariance matrix
Vλ = I22 (θ ) − I21 (θ )I11 (θ )−1 I12 (θ ),

(6.29)

where Ij k (θ ) are blocks in the information matrix I (θ ), deﬁned below (6.17), that is
I (θ ) =

I11 (θ )

I12 (θ )

I21 (θ )

I22 (θ )

,

where I22 (θ ) is of dimension J × J . Computationally, we can make use of the fact4
that (6.29) is the inverse of the lower right J × J block of the inverse of I (θ ),
I (θ )−1 =
4

I 11 (θ )

I 12 (θ )

I 21 (θ )

I 22 (θ )

,

This result is generally true and follows using partitioned inverses (see Davidson and MacKinnon, 1993,
Appendix A; or Greene, 2003, Appendix A).

SPECIFICATION TESTS

175

that is Vλ = I 22 (θ )−1 . The Lagrange multiplier test statistic can be derived as
ξLM = N −1 λ̃ Iˆ22 (θ̃ )λ̃,

(6.30)

which under the null hypothesis has an asymptotic Chi-squared distribution with J
degrees of freedom, and where Iˆ(θ̃ ) denotes an estimate of the information matrix
based upon the constrained estimator θ̃ . Only if I12 (θ ) = 0 and the information matrix
is block diagonal it holds that I 22 (θ ) = I (θ )−1
22 . In general the other blocks of the
information matrix are required to compute the appropriate covariance matrix of N −1 λ̃.
Computation of the LM test statistic is particularly attractive if the information matrix
is estimated on the basis of the ﬁrst derivatives of the loglikelihood function, as
N
1 
ˆ
IG (θ̃ ) =
s (θ̃ )si (θ̃ ) ,
N i=1 i

(6.31)

i.e. the average outer product of the vector of ﬁrst derivatives, evaluated under the
constrained ML estimates θ̃ . Using (6.27) and (6.28) we can write an LM test statistic as
ξLM =

N




si (θ̃ )

N


i=1

−1 N




si (θ̃ )si (θ̃ )

i=1

si (θ̃ ).

(6.32)

i=1

Note that the ﬁrst K − J elements in the vector of score contributions si (θ̃ ) sum
to zero because of (6.27). Nevertheless, these elements are generally important for
computing the correct covariance matrix. Only in the case of block diagonality it
holds that I12 (θ ) = 0 and the other block of the information matrix is irrelevant. An
asymptotically equivalent version of the LM test statistic in the block diagonal case
can be written as
ξLM =

N

i=1



si2 (θ̃ )

N




si2 (θ̃ )si2 (θ̃ )

i=1

−1 N


si2 (θ̃ ).

(6.33)

i=1

The expression in (6.32) suggests an easy way to compute a Lagrange multiplier
test statistic. Let us denote the N × K matrix of ﬁrst derivatives as S, such that


s1 (θ̃ )


 s2 (θ̃ ) 

(6.34)
S= . 
.
 .. 
sN (θ̃ )

In the matrix S each row corresponds to an observation and each column corresponds
to the derivative with respect to one of the parameters. Consequently, we can write
N

i=1

si (θ̃ ) = S  ι,

176

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

where ι = (1, 1, 1, . . . , 1) of dimension N . Moreover
N


si (θ̃ )si (θ̃ ) = S  S.

i=1

This allows us to rewrite (6.32) as
ξLM = ι S(S  S)−1 S  ι = N

ι S(S  S)−1 S  ι
.
ι ι

(6.35)

Now, consider an auxiliary regression of a column of ones upon the columns of the
matrix S. From the standard expression for the OLS estimator, (S  S)−1 S  ι, we obtain
predicted values of this regression as S(S  S)−1 S  ι. The explained sum of squares,
therefore, is given by
ι S(S  S)−1 S  S(S  S)−1 S  ι = ι S(S  S)−1 S  ι,
while the total (uncentred) sum of squares of this regression is ι ι. Consequently, it
follows that one version of the Lagrange multiplier test statistic can be computed as
ξLM = NR 2 ,

(6.36)

where R 2 is the uncentred R 2 (see Section 2.4) of an auxiliary regression of a vector of
ones upon the score contributions (in S).5 Under the null hypothesis, the test statistic
is asymptotically χ 2 distributed with J degrees of freedom, where J is the number of
restrictions imposed upon θ . Note that the auxiliary regression should not include an
intercept term.
The formulae in (6.32) or (6.36) provide one way of computing the Lagrange multiplier test statistic, often referred to as the outer product gradient (OPG) version
of the LM test statistic (see Godfrey, 1988, p. 15). Unfortunately, tests based on the
OPG estimate of the covariance matrix typically have small sample properties that
are quite different from those asymptotic theory predicts. Several Monte Carlo experiments suggest that the OPG-based tests tend to reject the null hypothesis too often in
cases where it happens to be true. That is, the actual size of the tests may be much
larger than the asymptotic size (typically 5%). This means that one has to be careful
in rejecting the null hypothesis when the test statistic exceeds the asymptotic critical
value. See Davidson and MacKinnon (1993, p. 477) for additional discussion. Alternative ways are available to compute LM test statistics, for example using (6.30) and
the matrix of second derivatives of the loglikelihood function, or on the basis of other
auxiliary regressions. Some of these will be discussed in the next section.
Despite the above reservations, we shall focus our discussion mostly upon the NR 2
approach of the LM test. This is because computation is convenient as it requires
only the ﬁrst derivatives. A test for any hypothesis can easily be constructed in this
approach, while the columns of S are often determined fairly easily on the basis of the
5

If your software does not report uncentred R 2 s, the same result is obtained by computing N − RSS, where
RSS denotes the residual sum of squares.

SPECIFICATION TESTS

177

estimation results. When implementing the OPG version of the test, it is recommended
to check your programming by also running a regression of a vector of ones upon the
columns in S that correspond to the unconstrained parameters. This should result in
an R 2 of zero.
In Section 6.3 we discuss the implementation of the Lagrange multiplier principle
to test for omitted variables, heteroskedasticity, autocorrelation and non-normality, all
in the context of the linear regression model with normal errors. Chapter 7 will cover
several applications of LM tests in different types of models. First, however, we shall
consider our simple example again.
6.2.3 An Example (Continued)

Let us again consider the simple example concerning the pool of red and yellow
balls. This example is particularly simple as it involves only one unknown coefﬁcient.
Suppose we are interested in testing the hypothesis H0 : p = p0 for a given value p0 .
The (unrestricted) maximum likelihood estimator was seen to equal
p̂ =

N
1 
N
yi = 1 ,
N i=1
N

while the constrained ‘estimator’ is simply p̃ = p0 . The Wald test for H0 , in its
quadratic form, is based upon the test statistic
ξW = N (p̂ − p0 )[p̂(1 − p̂)]−1 (p̂ − p0 ),
which is simply the square of the test statistic presented in Subsection 6.1.3
For the likelihood ratio test we need to compare the maximum loglikelihood values
of the unrestricted and the restricted model, that is
log L(p̂) = N1 log(N1 /N ) + (N − N1 ) log(1 − N1 /N ),

(6.37)

and
log L(p̃) = N1 log(p0 ) + (N − N1 ) log(1 − p0 ).
The test statistic is simply computed as
ξLR = 2(log L(p̂) − log L(p̃)).
Finally, we consider the Lagrange multiplier test. With a single coefﬁcient we obtain
that the Lagrange multiplier N −1 λ̃ (expressed as a sample average) is asymptotically
normal with variance I (p) = [p(1 − p)]−1 . Furthermore,

N

∂ log Li (p) 
N
N − N1
= 1−
.
λ̃ =

∂p
p0
1 − p0
p0
i=1

178

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

We can thus compute the LM test statistic as
ξLM = N −1 λ̃[p0 (1 − p0 )]λ̃
= N −1 (N1 − Np0 )[p0 (1 − p0 )]−1 (N1 − Np0 )
= N (p̂ − p0 )[p0 (1 − p0 )]−1 (p̂ − p0 ).
This shows that in this case the LM test statistic is very similar to the Wald test statistic:
the only difference is that the information matrix is estimated using the restricted
estimator p0 rather than the unrestricted estimator p̂.
As an illustration, suppose that we have a sample of N = 100 balls, of which 44%
are red. If we test the hypothesis that p = 0.5, we obtain Wald, LR and LM test
statistics of 1.46, 1.44 and 1.44, respectively. The 5% critical value taken from the
asymptotic Chi-squared distribution with one degree of freedom is 3.84, so that the
null hypothesis is not rejected at the 5% level with each of the three tests.

6.3

Tests in the Normal Linear Regression Model

Let us again consider the normal linear regression model, as discussed in Subsection 6.1.4,
yi = xi β + εi , εi ∼ NID(0, σ 2 ),
where εi is independent of xi . Suppose we are interested in testing whether the current
speciﬁcation is misspeciﬁed. Misspeciﬁcation could reﬂect the omission of relevant
variables, the presence of heteroskedasticity or autocorrelation, or non-normality of the
error terms. It is relatively easy to test for such misspeciﬁcations using the Lagrange
multiplier framework, where the current model is considered to be the restricted model
and the ML estimates are the constrained ML estimates. We then consider more general
models, that allow, e.g. for heteroskedasticity, and test whether the current estimates
signiﬁcantly violate the ﬁrst order conditions of the more general model.

6.3.1 Testing for Omitted Variables

The ﬁrst speciﬁcation test that we consider is testing for omitted variables. In this case,
the more general model is
yi = xi β + zi γ + εi ,
where the same assumptions are made about εi as before, and zi is a J -dimensional
vector of explanatory variables, independent of εi . The null hypothesis states H0 : γ =
0. Note that under the assumptions above, the F -test discussed in Subsection 2.5.4
provides an exact test for γ = 0 and there is no real need to look at asymptotic
tests. We discuss the Lagrange multiplier test for γ = 0 for illustrative purposes as
it can be readily extended to nonlinear models in which the F -test is not available

TESTS IN THE NORMAL LINEAR REGRESSION MODEL

179

(see Chapter 7). The ﬁrst order conditions for the more general model imply that the
following derivatives, when evaluated in the restricted estimates, are all equal to zero:
N

(yi − xi β − zi γ )
xi ,
σ2
i=1
N

(yi − xi β − zi γ )
zi ,
σ2
i=1

and

N

−

N
1  (yi − xi β − zi γ )2
+
.
2σ 2
2 i=1
σ4

Evaluating these derivatives at the (constrained) maximum likelihood estimates β̂, σ̂ 2
(and γ = 0), while deﬁning residuals ε̂i = yi − xi β̂ we can write the derivatives as
N

ε̂i
x;
σ̂ 2 i
i=1

N

ε̂i
z;
σ̂ 2 i
i=1

N

−

N
1  ε̂i2
+
,
2σ̂ 2
2 i=1 σ̂ 4

6
where the ﬁrst and third expressions
N are zero 2by construction. The Lagrange multiplier
test should thus check whether i=1 ε̂i zi /σ̂ differs signiﬁcantly from zero. The LM
test statistic can be computed as (6.35), where S has typical row

[ε̂i xi

ε̂i zi ].

(6.38)

Because of the block diagonality of the information matrix, the derivatives with respect
to σ 2 can be omitted here, although it would not be incorrect to include them in the
matrix S as well. Furthermore, irrelevant proportionality factors are eliminated in S.
This is allowed because such constants do not affect the outcome of (6.35). In summary,
we compute the LM test statistic by regressing a vector of ones upon the (ML or
OLS) residuals interacted with the included explanatory variables xi and the omitted
variables zi , and multiplying the uncentred R 2 by the sample size N . Under the null
hypothesis, the resulting test statistic NR 2 has an asymptotic Chi-squared distribution
with J degrees of freedom. An asymptotically equivalent version of the test statistic
can be obtained as NR 2 , where R 2 is the R 2 of an auxiliary regression of the ML
(or OLS) residuals upon the complete set of regressors, xi and zi . If zi is taken to
be a nonlinear function of xi , this approach can straightforwardly be used to test the
functional form of the model (against a well-deﬁned alternative).
6.3.2 Testing for Heteroskedasticity

Now suppose that the variance of εi may not be constant, but a function of some
variables zi , typically a subset or function of xi . This is formalized in equation (4.44)
6

These two expressions correspond to the ﬁrst order conditions of the restricted model, and deﬁne β̂
and σ̂ 2 .

180

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

from Chapter 4 which says that
V {εi } = σi2 = σ 2 h(zi α),

(6.39)

where h is an unknown, continuously differentiable function (that does not depend on
i), such that h(.) > 0, h (.) = 0 and h(0) = 1, and where zi is a J -dimensional vector of
explanatory variables (not including a constant). The null hypothesis of homoskedastic errors corresponds to H0 : α = 0 (and we have V {εi } = σ 2 ). The loglikelihood
contribution for observation i in the more general model is given by
1
1
1 (yi − xi β)2
.
log Li (β, α) = − log(2π) − log σ 2 h(zi α) −
2
2
2 σ 2 h(zi α)

(6.40)

The score with respect to α is given by


1 1
1 (yi − xi β)2 ∂h(zi α)
∂ log Li (β, α)
= −
+
,
∂α
2 h(zi α) 2 σ 2 h(zi α)2
∂α
where

∂h(zi α)
= h (zi α)zi ,
∂α

where h is the derivative of h. If we evaluate this under the constrained ML estimates
β̂ and σ̂ 2 this reduces to


1 1 (yi − xi β̂)2
− +
κzi ,
2 2
σ̂ 2
where κ = h (0) = 0 is an irrelevant constant. This explains the surprising result that
the test does not require us to specify the function h.
Because the information matrix is block diagonal with respect to β and (σ 2 , α),
the OPG version of the Lagrange multiplier test for heteroskedasticity is obtained by
computing (6.35), where S has typical row
[ε̂i2 − σ̂ 2

(ε̂i2 − σ̂ 2 )zi ],

where irrelevant proportionality factors are again eliminated. In the auxiliary regression
we thus include the variables that we suspect to affect heteroskedasticity interacted with
the squared residuals in deviation from the error variance estimated under the null
hypothesis. With J variables in zi the resulting test statistic NR 2 has an asymptotic
Chi-squared distribution with J degrees of freedom (under the null hypothesis).
The above approach presents a way to compute the Breusch–Pagan test for heteroskedasticity corresponding to our general computation rule given in (6.35). There are
alternative ways to compute (asymptotically equivalent) versions of the Breusch–Pagan
test statistic, for example by computing N times the R 2 of an auxiliary regression of
ε̂i2 (the squared OLS or maximum likelihood residuals) on zi and a constant. This was
discussed in Chapter 4. See Engle (1984) or Godfrey (1988, Section 4.5) for additional discussion.

TESTS IN THE NORMAL LINEAR REGRESSION MODEL

181

If the null hypothesis of homoskedasticity is rejected, one option is to estimate a more
general model that allows for heteroskedasticity. This can be based upon (6.40), with a
particular choice for h(.), for example the exponential function. As heteroskedasticity,
in this particular model, does not result in an inconsistent maximum likelihood (OLS)
estimator for β, it is also appropriate to compute heteroskedasticity-consistent standard
errors; see Chapter 4 and Section 6.4 below.
6.3.3 Testing for Autocorrelation

In a time series context, the error term in a regression model may suffer from autocorrelation. Consider the linear model
yt = xt β + εt ,

t = 1, 2, . . . , T ,

with assumptions as stated above. The alternative hypothesis of ﬁrst order autocorrelation states that
εt = ρεt−1 + vt ,
such that the null hypothesis corresponds to H0 : ρ = 0. If we rewrite the model as
yt = xt β + ρεt−1 + vt
it follows that testing for autocorrelation is similar to testing for an omitted variable,

namely εt−1 = yt−1 − xt−1
β. Consequently, one can compute a version of the Lagrange
multiplier test for autocorrelation using (6.35), where S has typical row
[ε̂t xt

ε̂t ε̂t−1 ]

and the number of observations is T − 1. If xt does not contain a lagged dependent
variable the information matrix is block diagonal with respect to β and (σ 2 , ρ), and
the scores with respect to β, corresponding to ε̂t xt may be dropped from S. This gives
a test statistic of
ξLM =

T


T


ε̂t ε̂t−1

t=2

2
ε̂t2 ε̂t−1

−1 T


t=2

ε̂t ε̂t−1 .

t=2

2
Because under the null hypothesis εt and εt−1 are independent,7 it holds that E{εt2 εt−1
}
2
2
equivalent
test
statistic
is
= E{εt }E{εt−1 }. This indicatesthat an asymptotically

 
 2 
2
by 1/(T − 1) t ε̂t2 1/(T − 1) t ε̂t−1
obtained by replacing 1/(T − 1) t ε̂t2 ε̂t−1
.
This gives

T
ξLM = (T − 1)
7

t=2 ε̂t ε̂t−1



T
2
t=2 ε̂t−1

T

−1 

2
t=2 ε̂t

T
t=2 ε̂t ε̂t−1

= (T − 1)R 2 ,

Recall that under normality zero correlation implies independence (see Appendix B).

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

182

where R 2 is the R 2 of an auxiliary regression of the OLS or ML residual ε̂t upon its lag
ε̂t−1 . This corresponds to the Breusch–Godfrey test for autocorrelation as discussed
in Chapter 4. If xt contains a lagged dependent variable, the appropriate auxiliary
regression is of ε̂t upon ε̂t−1 and xt . Tests for p-th order autocorrelation are obtained
by augmenting the rows of S with ε̂t ε̂t−2 up to ε̂t ε̂t−p , or – for the latter computation – by adding ε̂t−2 up to ε̂t−p in the auxiliary regression explaining ε̂t . Engle
(1984) and Godfrey (1988, Section 4.4) provide additional discussion.

6.4

Quasi-maximum Likelihood and Moment
Conditions Tests

It is typically the case that maximum likelihood requires researchers to make full
distributional assumptions, while the generalized method of moments (GMM) discussed
in the previous chapter only makes assumptions about moments of the distribution.
However, it is possible that the moment conditions employed in a GMM approach
are based upon assumptions about the shape of the distribution as well. This allows
us to re-derive the maximum likelihood estimator as a GMM estimator with moment
conditions corresponding to the ﬁrst order conditions of maximum likelihood. This is a
useful generalization as it allows us to argue that in some cases the maximum likelihood
estimator is consistent, even if the likelihood function is not entirely correct (but the
ﬁrst order conditions are). Moreover, it allows us to extend the class of Lagrange
multiplier tests to (conditional) moment tests.
6.4.1 Quasi-maximum Likelihood

In this subsection we shall see that the maximum likelihood estimator can be interpreted
as a GMM estimator by noting that the ﬁrst order conditions of the maximum likelihood
problem correspond to sample averages based upon theoretical moment conditions. The
starting point is that it holds that
E{si (θ )} = 0

(6.41)

for the true K-dimensional parameter vector θ , under the assumption that the likelihood function is correct. The proof of this is relatively easy and instructive. If we
consider the density function of yi given xi , f (yi |xi ; θ ), it holds by construction that
(see Appendix B)

f (yi |xi ; θ ) dyi = 1,
where integration is over the support of yi . Differentiating this with respect to θ gives

∂f (yi |xi ; θ )
dyi = 0.
∂θ
Because
∂ log f (yi |xi ; θ )
∂f (yi |xi ; θ )
=
f (yi |xi ; θ ) = si (θ )f (yi |xi ; θ )
∂θ
∂θ

QUASI-MAXIMUM LIKELIHOOD AND MOMENT CONDITIONS TESTS

it follows that

183


si (θ )f (yi |xi ; θ )dyi = E{si (θ )} = 0,

where the ﬁrst equality follows from the deﬁnition of the expectation operator.
Let us assume that θ is uniquely deﬁned by these conditions. That is, there is only
one vector θ that satisﬁes (6.41). Then (6.41) is a set of valid moment conditions and we
can use the GMM approach to estimate θ . Because the number of parameters is equal
to the number of moment conditions, this involves solving the ﬁrst order conditions
N
1 
s (θ ) = 0.
N i=1 i

Of course this reproduces the maximum likelihood estimator θ̂ . However, it shows that
the resulting estimator is consistent for θ provided that (6.41) is correct, which may
be weaker than the requirement that the entire distribution is correctly speciﬁed. In the
linear regression model with normal errors, the ﬁrst order conditions with respect to β
are easily seen to correspond to
E{(yi − xi β)xi } = 0,
which corresponds to the set of moment conditions imposed by the OLS estimator.
This explains why the maximum likelihood estimator in the normal linear regression
model is consistent even if the distribution of εi is not normal.
If the maximum likelihood estimator is based upon the wrong likelihood function,
but can be argued to be consistent on the basis of the validity of (6.41), the estimator is
sometimes referred to as a quasi-maximum likelihood estimator or pseudo-maximum
likelihood estimator (see White, 1982; or Gouriéroux, Monfort and Trognon, 1984).
The asymptotic distribution of the quasi-ML estimator may differ from that of the
ML estimator. In particular, the result in (6.18) may no longer be valid. Using our
general formulae for the GMM estimator it is possible to derive the asymptotic covariance matrix of the quasi-ML estimator for θ , assuming that (6.41) is correct. Using
(5.74)–(5.76), it follows that the quasi-maximum likelihood estimator θ̂ satisﬁes
√
N (θ̂ − θ ) → N(0, V )
where8
with

V = I (θ )−1 J (θ )I (θ )−1 ,
N
1 
Ii (θ )
N→∞ N
i=1

I (θ ) ≡ lim
where

8



∂s (θ )
Ii (θ ) = E − i 
∂θ

and


(6.42)
N
1 
Ji (θ ),
N→∞ N
i=1

J (θ ) ≡ lim

 2

∂ log Li (θ )
=E −
,
∂θ ∂θ 

The covariance matrix maintains the assumption that observations are mutually independent.

184

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

as deﬁned in (6.16), and
Ji (θ ) = E{si (θ )si (θ )},
as deﬁned in (6.20). The covariance matrix in (6.42) generalizes the one in (6.18)
and is correct whenever the quasi-ML estimator θ̂ is consistent. Its expression is
popularly referred to as the ‘sandwich formula’. In the case of the linear regression model estimating the covariance matrix on the basis of (6.42) would reproduce
the heteroskedasticity-consistent covariance matrix as discussed in Subsection 4.3.4.
Several software packages have the option to compute robust standard errors for the
(quasi-)maximum likelihood estimator, based on the covariance matrix in (6.42).
The information matrix test (IM test) suggested by White (1982) tests the equality
of the two K × K matrices I (θ ) and J (θ ) by comparing their sample counterparts.
Because of the symmetry a maximum of K(K + 1)/2 elements have to be compared,
so that the degrees of freedom for the IM test is potentially very large. Depending on the shape of the likelihood function, the information matrix test checks for
misspeciﬁcation in a number of directions simultaneously (like functional form, heteroskedasticity, skewness and kurtosis). For additional discussion and computational
issues, see Davidson and MacKinnon (1993, Section 16.9).
6.4.2 Conditional Moment Tests

The analysis in the previous subsection allows us to generalize the class of Lagrange
multiplier tests to so-called conditional moment tests (CM tests), as suggested
by Newey (1985) and Tauchen (1985). Consider a model characterized by (6.41)
E{si (θ )} = 0,
where the (quasi-)ML estimator θ̂ satisﬁes
N
1 
s (θ̂ ) = 0.
N i=1 i

Now consider a hypothesis characterized by
E{mi (θ )} = 0,

(6.43)

where mi (θ ) is a J -dimensional function of the data and the unknown parameters in θ ,
just like si (θ ). The difference is that (6.43) is not imposed in estimation. It is possible
to test the validity of (6.43) by testing whether its sample counterpart
N
1 
m (θ̂ )
N i=1 i

(6.44)

is close to zero. This can be done fairly easily by noting the resemblance between
(6.44) and the scores of a more general likelihood function. Consequently, the OPG
version of a moment conditions test for (6.43) can be computed by taking N times

QUASI-MAXIMUM LIKELIHOOD AND MOMENT CONDITIONS TESTS

185

the uncentred R 2 of a regression of a vector of ones upon the columns of a matrix S,
where S now has typical row
[si (θ̂ ) mi (θ̂ ) ].
Under the null hypothesis that (6.43) is correct, the resulting test statistic has an asymptotic Chi-squared distribution with J degrees of freedom.
The above approach shows that the additional conditions that are tested do not necessarily have to correspond to scores of a more general likelihood function. A particular
area where this approach is useful is when testing the hypothesis of normality.

6.4.3 Testing for Normality

Let us consider the linear regression model again with, under the null hypothesis,
normal errors. For a continuously observed variable, normality tests usually check for
skewness (third moment) and excess kurtosis (fourth moment), because the normal distribution implies that E{εi3 } = 0 and E{εi4 − 3σ 4 } = 0 (see Appendix B). If E{εi3 } = 0
the distribution of εi is not symmetric around zero. If E{εi4 − 3σ 4 } > 0 the distribution
of εi is said to display excess kurtosis. This means that it has fatter tails than the normal
distribution. Davidson and MacKinnon (1993, p. 63) provide graphical illustrations of
these situations.
Given the discussion in the previous subsection, a test for normality can be obtained
by running a regression of a vector of ones upon the columns of the matrix S, which
now has typical row
[ε̂i xi ε̂i2 − σ̂ 2 ε̂i3 ε̂i4 − 3σ̂ 4 ],
where ε̂i denotes the maximum likelihood (or OLS) residual, and then computing N
times the uncentred R 2 . Although non-normality of εi does not invalidate consistency
of the OLS estimator nor its asymptotic normality, the above test is occasionally of
interest. Finding that εi has a severely skewed distribution may indicate that it may
be advisable to transform the dependent variable prior to estimation (for example, by
considering log wages rather than wages itself). In Chapter 7 we shall see classes of
models where normality is far more crucial.
A popular variant of the LM test for normality is the Jarque–Bera test (Jarque and
Bera, 1980). The test statistic is computed as

ξLM = N 

1
6

N
1  3 3
ε̂ /σ̂
N i=1 i

2

+

1
24

N
1  4 4
ε̂ /σ̂ − 3
N i=1 i

2


,

(6.45)

which is a weighted average of the squared sample moments corresponding to skewness
and excess kurtosis, respectively. Under the null hypothesis, it is asymptotically distributed as a Chi-squared with two degrees of freedom; see Godfrey (1988, Section 4.7)
for more details.

MAXIMUM LIKELIHOOD ESTIMATION AND SPECIFICATION TESTS

186

Exercises
Exercise 6.1 (The Normal Linear Regression Model)

Consider the following linear regression model
yi = β1 + β2 xi + εi ,
where β = (β1 , β2 ) is a vector of unknown parameters, and xi is a one-dimensional
observable variable. We have a sample of i = 1, . . . , N independent observations and
assume that the error terms εi are NID(0, σ 2 ), independent of all xi . The density
function of yi (for a given xi ) is then given by


1 (yi − β1 − β2 xi )2
f (yi |β, σ ) = √
exp −
.
2
σ2
2πσ 2
2

a.

1

Give an expression for the loglikelihood contribution of observation
i, log Li (β, σ 2 ). Explain why the loglikelihood function of the entire sample is
given by
N

log L(β, σ 2 ) =
log Li (β, σ 2 ).
i=1

b. Determine expressions for the two elements in ∂ log Li (β, σ 2 )/∂β and show that
both have expectation zero for the true parameter values.
c. Derive an expression for ∂ log Li (β, σ 2 )/∂σ 2 and show that it also has expectation
zero for the true parameter values.
Suppose that xi is a dummy variable equal to 1 for males and 0 for females, such that
xi = 1 for i = 1, . . . , N1 (the ﬁrst N1 observations) and xi = 0 for i = N1 + 1, . . . , N .
d. Derive the ﬁrst order conditions for maximum likelihood. Show that the maximum
likelihood estimators for β are given by
β̂1 =

N

1
y,
N − N1 i=N +1 i
1

β̂2 =

N1
1 
y − β̂1 .
N1 i=1 i

What is the interpretation of these two estimators? What is the interpretation of
the true parameter values β1 and β2 ?
e. Show that
∂ 2 log Li (β, σ 2 )/∂β∂σ 2 = ∂ 2 log Li (β, σ 2 )/∂σ 2 ∂β,
and show that it has expectation zero. What are the implications of this for the
asymptotic covariance matrix of the ML estimator (β̂1 , β̂2 , σ̂ 2 )?
f. Present two ways to estimate the asymptotic covariance matrix of (β̂1 , β̂2 ) and
compare the results.
g. Present an alternative way to estimate the asymptotic covariance matrix of (β̂1 , β̂2 )
that allows εi to be heteroskedastic.

EXERCISES

187

Suppose that we are interested in the hypothesis H0 : β2 = 0 with alternative H1 : β2 = 0.
Tests can be based upon the likelihood ratio, Lagrange multiplier or Wald principle.
h. Explain what these three principles are.
i. Discuss for each of the three tests what is required to compute them.
Although the three test statistics have the same asymptotic Chi-squared distribution, it
can be shown (see, e.g. Godfrey, 1988, Section 2.3) that in the above model it holds
for any ﬁnite sample that
ξW ≥ ξLR ≥ ξLM .
j. Explain what is meant by the power of a test. What does this inequality tell us
about the powers of the three tests? (Hint: if needed consult Chapter 2.)
k. Explain what is meant by the (actual) size of a test. What does the inequality tell
us about the sizes of the three tests?
l. Would you prefer one of the three tests, knowing the above inequality?
Exercise 6.2 (The Poisson Regression Model)

Let yi denote the number of times individual i buys tobacco in a given month. Suppose a random sample of N individuals is available, for which we observe values
0, 1, 2, 3, . . . . Let xi be an observed characteristic of these individuals (e.g. gender). If we assume that, for given xi , yi has a Poisson distribution with parameter λi = exp{β1 + β2 xi }, the probability mass function of yi conditional upon xi is
given by
y
e−λi λi
P {yi = y|xi } =
.
y!
a. Write down the loglikelihood function for this so-called Poisson regression model.
b. Derive the score contributions. Using that the Poisson distribution implies that
E{yi |xi } = λi , show that the score contributions have expectation zero.
c. Derive an expression for the information matrix I (β1 , β2 ). Use this to determine
the asymptotic covariance matrix of the ML estimator and a consistent estimator
for this matrix.
d. Describe how one can test for an omitted variable using the Lagrange multiplier
framework. Which auxiliary regression is needed?
More details about the Poisson regression model can be found in Section 7.3.

7

Models with Limited
Dependent Variables

In practical applications one often has to cope with phenomena that are of a discrete or
mixed discrete continuous nature. For example, one could be interested in explaining
whether married women have a paid job (yes or no), or how many hours they work
(zero or positive). If this type of variable has to be explained a linear regression
model is generally inappropriate. In this chapter we consider alternative models that
can be used to model discrete and discrete/continuous variables and pay attention to
the estimation and interpretation of their parameters.
Although not exclusively, in many cases the problems analysed with this type of
model are of a micro-economic nature, thus requiring data on individuals, households
or ﬁrms. To stress this, we shall index all variables by i, running from 1 to sample
size N . Section 7.1 starts with probably the simplest case of a limited dependent
variable model, viz. a binary choice model. Extensions to multiple discrete outcomes
are discussed in Section 7.2. When the endogenous variable is the frequency of a
certain event, for example the number of patents in a given year, count data models are
often employed. Section 7.3 introduces several models for count data and presents an
empirical illustration. If the distribution of the endogenous variable is continuous with a
probability mass at one or more discrete points, the use of tobit models is recommended.
The standard tobit model is discussed in Section 7.4, while some extensions, including
models with sample selection where a nonrandom proportion of the outcomes is not
observed, are contained in Section 7.5. Because sample selection is a problem that
often arises with micro data, Section 7.6 contains some additional discussion of the
sample selection problem, mainly focusing on the identiﬁcation problem and under
what assumptions it can be solved. An area that has gained interest recently is the
estimation of treatment effects, and we discuss this in Section 7.7. Finally, Section 7.8
discusses models in which the dependent variable is a duration. Throughout, a number
of empirical illustrations are provided in subsections. Additional discussion of limited
dependent variable models in econometrics can be found in two surveys by Amemiya

MODELS WITH LIMITED DEPENDENT VARIABLES

190

(1981, 1984) and the monographs by Maddala (1983), Lee (1996), Franses and Paap
(2001) and Wooldridge (2002).

7.1

Binary Choice Models

7.1.1 Using Linear Regression?

Suppose we want to explain whether a family possesses a car or not. Let the sole
explanatory variable be the family income. We have data on N families (i = 1, . . . , N ),
with observations on their income, xi2 , and whether or not they own a car. This latter
element is described by the binary variable yi , deﬁned as
yi = 1 if family i owns a car
yi = 0 if family i does not own a car.
Suppose we would use a regression model to explain yi from xi2 and an intercept term
(xi1 ≡ 1). This linear model would be given by
yi = β1 + β2 xi2 + εi = xi β + εi ,

(7.1)

where xi = (xi1 , xi2 ) . It seems reasonable to make the standard assumption that
E{εi |xi } = 0 such that E{yi |xi } = xi β. This implies that
E{yi |xi } = 1.P {yi = 1|xi } + 0.P {yi = 0|xi }
= P {yi = 1|xi } = xi β.

(7.2)

Thus, the linear model implies that xi β is a probability and should therefore lie between
0 and 1. This is only possible if the xi values are bounded and if certain restrictions on β
are satisﬁed. Usually this is hard to achieve in practice. In addition to this fundamental
problem, the error term in (7.1) has a highly non-normal distribution and suffers from
heteroskedasticity. Because yi has only two possible outcomes (0 or 1), the error term,
for a given value of xi , has two possible outcomes as well. In particular, the distribution
of εi can be summarized as
P {εi = −xi β|xi } = P {yi = 0|xi } = 1 − xi β
P {εi = 1 − xi β|xi } = P {yi = 1|xi } = xi β.

(7.3)

This implies that the variance of the error term is not constant but dependent upon
the explanatory variables according to V {εi |xi } = xi β(1 − xi β). Note that the error
variance also depends upon the model parameters β.
7.1.2 Introducing Binary Choice Models

To overcome the problems with the linear model, there exists a class of binary choice
models (or univariate dichotomous models), designed to model the ‘choice’ between
two discrete alternatives. These models essentially describe the probability that yi = 1

BINARY CHOICE MODELS

191

directly, although they are often derived from an underlying latent variable model (see
below). In general, we have
P {yi = 1|xi } = G(xi , β)

(7.4)

for some function G(.). This equation says that the probability of having yi = 1 depends
on the vector xi containing individual characteristics. So, for example, the probability
that a person owns a house depends on his income, education level, age and marital
status. Or, from a different ﬁeld, the probability that an insect survives a dose of
poisonous insecticide depends upon the quantity xi of the dose, and possibly some other
characteristics. Clearly, the function G(.) in (7.4) should take on values in the interval
[0, 1] only. Usually, one restricts attention to functions of the form G(xi , β) = F (xi β).
As F (.) also has to be between 0 and 1, it seems natural to choose F to be some
distribution function. Common choices are the standard normal distribution function


 w
1 2
1
F (w ) = (w ) =
(7.5)
√ exp − t dt,
2
2π
−∞
leading to the so-called probit model, and the standard logistic distribution function,
given by
ew
F (w ) = L(w ) =
,
(7.6)
1 + ew
which results in the logit model. A third choice corresponds to a uniform distribution
over the interval [0, 1] with distribution function
F (w ) = 0, w < 0;
F (w ) = w , 0 ≤ w ≤ 1;

(7.7)

F (w ) = 1, w > 1.
This results in the so-called linear probability model, which is similar to the regression
model in (7.1), but the probabilities are set to 0 or 1 if xi β exceeds the lower or upper
limit, respectively. In fact, the ﬁrst two models (probit and logit) are more common
in applied work. Both a standard normal and a standard logistic random variable have
an expectation of zero, while the latter has a variance of π 2 /3 instead of 1. These two
distribution functions are very similar if one corrects for this difference in scaling; the
logistic distribution has slightly heavier tails. Accordingly, the probit and logit model
typically yield very similar results in empirical work.
Apart from their signs, the coefﬁcients in these binary choice models are not easy to
interpret directly. One way to interpret the parameters (and to ease comparison across
different models) is to consider the partial derivative of the probability that yi equals
one with respect to a continuous explanatory variable, xik , say. For the three models
above, we obtain:
∂(xi β)
= φ(xi β)βk ;
∂xik

MODELS WITH LIMITED DEPENDENT VARIABLES

192



∂L(xi β)
e xi β
=
βk ;

∂xik
(1 + exi β )2
∂xi β
= βk ; (or 0),
∂xik
where φ(.) denotes the standard normal density function. Except for the last model,
the effect of a change in xik depends upon the values of xi . In all cases, however, the
sign of the effect of a change in xik corresponds to the sign of its coefﬁcient βk . For
a discrete explanatory variable, for example a dummy, the effect of a change can be
determined from computing the implied probabilities for the two different outcomes,
ﬁxing the values of all other explanatory variables.

7.1.3 An Underlying Latent Model

It is possible (but not necessary) to derive a binary choice model from underlying
behavioural assumptions. This leads to a latent variable representation of the model,
which is in common use even when such behavioural assumptions are not made. Let us
look at the decision of a married female to have a paid job or not. The utility difference
between having a paid job and not having one depends upon the wage that could be
earned but also on other personal characteristics, like the woman’s age and education,
whether there are young children in the family, etc. Thus, for each person i we can
write the utility difference between having a job and not having one as a function of
observed characteristics, xi say, and unobserved characteristics, εi say.1 Assuming a
linear additive relationship we obtain for the utility difference, denoted yi∗ ,
yi∗ = xi β + εi .

(7.8)

Because yi∗ is unobserved, it is referred to as a latent variable. In this chapter, latent
variables are indicated by an asterisk. Our assumption is that an individual chooses
to work if the utility difference exceeds a certain threshold level, which can be set to
zero without loss of generality. Consequently, we observe yi = 1 (job) if and only if
yi∗ > 0 and yi = 0 (no job) otherwise. Thus we have
P {yi = 1} = P {yi∗ > 0} = P {xi β + εi > 0} = P {−εi ≤ xi β} = F (xi β),

(7.9)

where F denotes the distribution function of −εi , or, in the common case of a symmetric distribution, the distribution function of εi . Consequently, we have obtained a
binary choice model, the form of which depends upon the distribution that is assumed
for εi . As the scale of utility is not identiﬁed, a normalization on the distribution of εi
is required. Usually this means that its variance is ﬁxed at a given value. If a standard
normal distribution is chosen one obtains the probit model, for the logistic one the
logit model is obtained.
1

The error term εi is not to be confused with the one in the linear model (7.1).

BINARY CHOICE MODELS

193

Although binary choice models in economics can often be interpreted as being
derived from an underlying utility maximization problem, this is certainly not necessary. Usually, one deﬁnes the latent variable yi∗ directly, such that the probit model is
fully described by
yi∗ = xi β + εi , εi ∼ NID(0, 1)
yi = 1
=0

if yi∗ > 0

(7.10)

if yi∗ ≤ 0,

where the εi s are independent of all xi . For the logit model, the normal distribution is
replaced by the standard logistic one. Most commonly, the parameters in binary choice
models (or limited dependent variable models in general) are estimated by the method
of maximum likelihood.
7.1.4 Estimation

Given our general discussion of maximum likelihood estimation in Chapter 6, we
can restrict attention to the form of the likelihood function here. In fact, this form
is rather simple as it follows immediately from the models given above. In general,
the likelihood contribution of observation i with yi = 1 is given by P {yi = 1|xi } as a
function of the unknown parameter vector β, and similarly for yi = 0. The likelihood
function for the entire sample is thus given by
L(β) =

N


P {yi = 1|xi ; β}yi P {yi = 0|xi ; β}1−yi ,

(7.11)

i=1

where we included β in the expressions for the probabilities to stress that the likelihood
function is a function of β. As usual we prefer to work with the loglikelihood function.
Substituting P {yi = 1|xi ; β} = F (xi β) we obtain
log L(β) =

N

i=1

yi log F (xi β) +

N


(1 − yi ) log(1 − F (xi β)).

(7.12)

i=1

Substituting the appropriate form for F gives an expression that can be maximized
with respect to β. As indicated above, the values of β and their interpretation depend
upon the distribution function that is chosen. An empirical example in Subsection 7.1.6
will illustrate this.
It is instructive to consider the ﬁrst order conditions of the maximum likelihood
problem. Differentiating (7.12) with respect to β yields

N 
yi − F (xi β)
∂ log L(β) 

=
f (xi β) xi = 0,
∂β
F (xi β)(1 − F (xi β))
i=1

(7.13)

where f = F  is the derivative of the distribution function (so f is the density function).
The term in square brackets is often referred to as the generalized residual of the

MODELS WITH LIMITED DEPENDENT VARIABLES

194

model, and we shall see it reappearing when discussing speciﬁcation tests. It equals
f (xi β)/F (xi β) for the positive observations (yi = 1) and −f (xi β)/(1 − F (xi β)) for
the zero observations (yi = 0). The ﬁrst order conditions thus say that each explanatory
variable should be orthogonal to the generalized residual (over the whole sample). This
is comparable to the OLS ﬁrst order conditions in (2.10), which state that the least
squares residuals are orthogonal to each variable in xi .
For the logit model we can simplify (7.13) to

N 
exp(xi β)
∂ log L(β) 
x = 0.
yi −
=
∂β
1 + exp(xi β) i
i=1

(7.14)

The solution of (7.14) is the maximum likelihood estimator β̂. From this estimate we
can estimate the probability that yi = 1 for a given xi as
p̂i =

exp(xi β̂)
1 + exp(xi β̂)

.

(7.15)

Consequently, the ﬁrst order conditions for the logit model imply that
N

i=1

p̂i xi =

N


yi xi .

(7.16)

i=1

Thus, if xi contains a constant term (and there is
no reason why it should not), then
the sum of the estimated probabilities is equal to i yi or the number of observations
in the sample for which yi = 1. In other words, the predicted frequency is equal to
the actual frequency. Similarly, if xi includes a dummy variable, say 1 for females, 0
for males, then the predicted frequency will be equal to the actual frequency for each
gender group. Although a similar result does not hold exactly for the probit model, it
does hold approximately by virtue of the similarity of the logit and probit model.
A look at the second order conditions of the ML problem reveals that the matrix of
second order derivatives is negative deﬁnite (assuming that the xs are not collinear).
Consequently, the loglikelihood function is globally concave and convergence of the
iterative maximum likelihood algorithm is guaranteed (and usually quite fast).
7.1.5 Goodness-of-ﬁt

A goodness-of-ﬁt measure is a summary statistic indicating the accuracy with which
the model approximates the observed data, like the R 2 measure in the linear regression
model. In the case in which the dependent variable is qualitative, accuracy can be
judged either in terms of the ﬁt between the calculated probabilities and observed
response frequencies or in terms of the model’s ability to forecast observed responses.
Contrary to the linear regression model, there is no single measure for the goodnessof-ﬁt in binary choice models and a variety of measures exists.
Often, goodness-of-ﬁt measures are implicitly or explicitly based on comparison
with a model that contains only a constant as explanatory variable. Let log L1 denote
the maximum loglikelihood value of the model of interest and let log L0 denote the

BINARY CHOICE MODELS

195

maximum value of the loglikelihood function when all parameters, except the intercept,
are set to zero. Clearly, log L1 ≥ log L0 . The larger the difference between the two
loglikelihood values, the more the extended model adds to the very restrictive model.
(Indeed, a formal likelihood ratio test can be based on the difference between the
two values.) A ﬁrst goodness-of-ﬁt measure is deﬁned as (see Amemiya, 1981, for an
extensive list)
1
pseudoR 2 = 1 −
,
(7.17)
1 + 2(log L1 − log L0 )/N
where N denotes the number of observations. An alternative measure is suggested
by McFadden (1974),
McFaddenR 2 = 1 − log L1 / log L0 ,

(7.18)

sometimes referred to as the likelihood ratio index. Because the loglikelihood is the
sum of log probabilities, it follows that log L0 ≤ log L1 < 0, from which it is straightforward to show that both measures take on values in the interval [0, 1] only. If all
estimated slope coefﬁcients are equal to zero we have log L0 = log L1 , such that both
R 2 s are equal to zero. If the model would be able to generate (estimated) probabilities
that correspond exactly to the observed values (that is p̂i = yi for all i), all probabilities
in the loglikelihood would be equal to one, such that the loglikelihood would be exactly
equal to zero. Consequently, the upper limit for the two measures above is obtained
for log L1 = 0. The upper bound of 1 can therefore, in theory, only be attained by
McFadden’s measure; see Cameron and Windmeijer (1997) for a discussion of the
properties of this and alternative measures.
To compute log L0 it is not necessary to estimate a probit or logit model with an intercept term only. If there is only a constant term in the model, the distribution function is
irrelevant for the implied probabilities and the model essentially says P {yi = 1} = p
for some unknown p. The ML estimator for p can easily be shown to be (see (6.4))
p̂ = N1 /N,

where N1 = i yi . That is, the estimated probability is equal to the proportion of ones
in the sample. The maximum loglikelihood value is therefore given by (compare (6.37))
log L0 =

N

i=1

yi log(N1 /N ) +

N


(1 − yi ) log(1 − N1 /N )

i=1

= N1 log(N1 /N ) + N0 log(N0 /N ),

(7.19)

where N0 = N − N1 denotes the number of zeros in the sample. It can be directly
computed from the sample size N and the sample frequencies N0 and N1 . The value
of log L1 should be given by your computer package.
An alternative way to evaluate the goodness-of-ﬁt is comparing correct and incorrect
predictions. To predict whether yi = 1 or not, it seems natural to look at the estimated
probability that follows from the model, which is given by F (xi β̂). In general, one

MODELS WITH LIMITED DEPENDENT VARIABLES

196

Table 7.1
outcomes

Cross-tabulation of actual and predicted
ŷi

yi

0
1
Total

0

1

Total

n00
n10
n0

n01
n11
n1

N0
N1
N

predicts that yi = 1 if F (xi β̂) > 1/2. Because F (0) = 1/2 for distributions that are
symmetric around 0, this corresponds to xi β̂ > 0. Thus, the implied predictions are
yˆi = 1 if xi β̂ > 0
yˆi = 0 if xi β̂ ≤ 0.

(7.20)

Now it is possible to construct a cross-tabulation of predictions and actual observations.
In Table 7.1, n11 denotes the number of correct predictions when the actual outcome is
1, and n10 denotes the number of times we predict a zero, while the actual value is 1.
Note that N1 = n11 + n10 and n1 = n11 + n01 . Several goodness-of-ﬁt measures can be
computed on the basis of this table. Overall, the proportion of incorrect predictions is
wr 1 =

n01 + n10
,
N

which can be compared with the proportion of incorrect predictions based on the model
with an intercept term only. It is easily seen that for this latter model we will predict
a one for all observation if p̂ = N1 /N > 1/2 and a zero otherwise. The proportion of
incorrect predictions is thus given by
w r0

= 1 − p̂
= p̂

if p̂ > 0.5,
if p̂ ≤ 0.5.

A goodness-of-ﬁt measure is ﬁnally obtained as
Rp2 = 1 −

wr 1
.
wr 0

(7.21)

Because it is possible that the model predicts worse than the simple model one can
have wr 1 > wr 0 , in which case the Rp2 becomes negative. Of course, this is not a
good sign for the predictive quality of the model. Also note that wr 0 ≤ 1/2, that is
even the simplest model will predict at most half of the observations incorrectly. If
in the sample 90% corresponds to yi = 1, we even have wr 0 = 0.1. Consequently, in
this case any binary choice model needs more than 90% correct predictions to beat
the simple model. As a consequence, the overall proportion of correct predictions,
1 − wr 1 = (n00 + n11 )/N , also known as the hit rate, does not give much information
about the quality of the model. It may be more informative to consider the proportions
of correct predictions for the two subsamples. From Table 7.1, the proportions of correct

BINARY CHOICE MODELS

197

predictions for the subsamples with yi = 0 and yi = 1 are given by p00 = n00 /N0 and
p11 = n11 /N1 respectively. Their sum, p00 + p11 , should be larger than 1 for a good
model. Unlike the pseudo R 2 measures based on the loglikelihood function, the last
two measures, based on the cross-tabulation of yi and ŷi can also be used to evaluate
out of sample forecasts.2
7.1.6 Illustration: the Impact of Unemployment Beneﬁts on Recipiency

As an illustration we consider a sample3 of 4877 blue collar workers who lost their
jobs in the US between 1982 and 1991, taken from a study by McCall (1995). Not all
unemployed workers eligible for unemployment insurance (UI) beneﬁts apply for it,
probably due to the associated pecuniary and psychological costs. The percentage of
eligible unemployed blue collar workers that actually applies for UI beneﬁts is called
the take-up rate, and it was only 68% in the available sample. It is therefore interesting
to investigate what makes people decide not to apply.
The amount of UI beneﬁts a person can receive depends upon the state of residence,
the year of becoming unemployed, and his or her previous earnings. The replacement
rate, deﬁned as the ratio of weekly UI beneﬁts to previous weekly earnings, varies from
33% to 54% with a sample average of 44%, and is potentially an important factor for
an unemployed worker’s choice to apply for unemployment beneﬁts. Of course, other
variables may inﬂuence the take-up rate as well. Due to personal characteristics, some
people are more able than others to ﬁnd a new job in a short period of time and
will therefore not apply for UI beneﬁts. Indicators of such personal characteristics are
schooling, age, and, due to potential (positive or negative) discrimination in the labour
market, racial and gender dummies. In addition, preferences and budgetary reasons, as
reﬂected in the family situation, may be of importance. Due to the important differences
in the state unemployment rates, the probability of ﬁnding a new job varies across
states and we will therefore include the state unemployment rate in the analysis. The
last type of variables that could be relevant relates to the reason why the job was lost.
In the analysis we will include dummy variables for the reasons: slack work, position
abolished, and end of seasonal work.
We estimate three different models, the results of which are presented in Table 7.2.
The linear probability model is estimated by ordinary least squares, so no corrections
for heteroskedasticity are made and no attempt is made to keep the implied probabilities between 0 and 1. The logit and probit model are both estimated by maximum
2
likelihood. Because the logistic distribution has a variance
√ of π /3, the estimates of β
obtained from the logit model are roughly a factor π/ 3 larger than those obtained
from the probit model, acknowledging the small differences in the shape of the distributions. Similarly, the estimates for the linear probability model are quite different in
magnitude and approximately four times as small as those for the logit model (except
for the intercept term). Looking at the results in Table 7.2, we see that the signs of the
coefﬁcients are identical across the different speciﬁcations, while the statistical significance of the explanatory variables is also comparable. This is not an unusual ﬁnding.
Qualitatively, the different models typically do not provide different answers.
2

Henriksson and Merton (1981) describe a test for forecasting performance based upon the null hypothesis
that sum of the conditional forecast probabilities is equal to one.
3
The data for this illustration are available as BENEFITS.

MODELS WITH LIMITED DEPENDENT VARIABLES

198

Table 7.2

Binary choice models for applying for unemployment beneﬁts (blue collar workers)
LPM

Logit

Probit

Variable

Estimate

s.e.

Estimate

s.e.

Estimate

s.e.

constant
replacement rate
replacement rate2
age
age 2 /10
tenure
slack work
abolished position
seasonal work
head of household
married
children
young children
live in SMSA
non-white
year of displacement
>12 years of school
male
state max. beneﬁts
state unempl. rate

−0.077
0.629
−1.019
0.0157
−0.0015
0.0057
0.128
−0.0065
0.058
−0.044
0.049
−0.031
0.043
−0.035
0.017
−0.013
−0.014
−0.036
0.0012
0.018

(0.122)
(0.384)
(0.481)
(0.0047)
(0.0006)
(0.0012)
(0.014)
(0.0248)
(0.036)
(0.017)
(0.016)
(0.017)
(0.020)
(0.014)
(0.019)
(0.008)
(0.016)
(0.018)
(0.0002)
(0.003)

−2.800
3.068
−4.891
0.068
−0.0060
0.0312
0.625
−0.0362
0.271
−0.211
0.242
−0.158
0.206
−0.170
0.074
−0.064
−0.065
−0.180
0.0060
0.096

(0.604)
(1.868)
(2.334)
(0.024)
(0.0030)
(0.0066)
(0.071)
(0.1178)
(0.171)
(0.081)
(0.079)
(0.086)
(0.097)
(0.070)
(0.093)
(0.015)
(0.082)
(0.088)
(0.0010)
(0.016)

−1.700
1.863
−2.980
0.042
−0.0038
0.0177
0.375
−0.0223
0.161
−0.125
0.145
−0.097
0.124
−0.100
0.052
−0.038
−0.042
−0.107
0.0036
0.057

(0.363)
(1.127)
(1.411)
(0.014)
(0.0018)
(0.0038)
(0.042)
(0.0718)
(0.104)
(0.049)
(0.048)
(0.052)
(0.059)
(0.042)
(0.056)
(0.009)
(0.050)
(0.053)
(0.0006)
(0.009)

Loglikelihood
Pseudo R 2
McFadden R 2
Rp2

0.035

−2873.197
0.066
0.057
0.046

−2874.071
0.066
0.057
0.045

For all speciﬁcations, the replacement rate has an insigniﬁcant positive coefﬁcient,
while its square is signiﬁcantly negative. The ceteris paribus effect of the replacement
rate will thus depend upon its value. For the probit model, for example, we can derive
that the estimated marginal effect4 of a change in the replacement rate (rr) equals
the value of the normal density function multiplied by 1.863 − 2 × 2.980rr, which
is negative for 85% of the observations in the sample. This is counterintuitive and
suggests that other variables might be more important in explaining the take-up rate.
The dummy variable which indicates whether the job was lost because of slack
work is highly signiﬁcant in all speciﬁcations, which is not surprising given that these
workers typically will ﬁnd it hard to get a new job. Many other variables are statistically
insigniﬁcant or only marginally signiﬁcant. This is particularly troublesome as with
this large number of observations a signiﬁcance level of 1% or less may be more
appropriate5 than the traditional 5%. The two variables relating to the state of residence
are statistically signiﬁcant. The higher the state unemployment rate and the higher the
maximum beneﬁt level, the more likely it is that individuals apply for beneﬁts, which
is intuitively reasonable. The ceteris paribus effect of being married is estimated to
be positive, while, somewhat surprisingly, being head of the household has a negative
effect on the probability of take-up.
4
5

See Section 3.1 for the computation of marginal effects in the linear model.
See the discussion on this issue in Section 2.5.7.

BINARY CHOICE MODELS

199

Table 7.3 Cross-tabulation of actual and predicted
outcomes (logit model)
ŷi
yi

0
1
Total

0

1

Total

242
171
413

1300
3164
4464

1542
3335
4877

The fact that the models do not do a very good job in explaining the probability
that someone applies for UI beneﬁts is reﬂected in the goodness-of-ﬁt measures that
are computed. Usually, goodness-of-ﬁt is fairly low for discrete choice models. In this
application, the alternative goodness-of-ﬁt measures indicate that the speciﬁed models
perform between 3.5 and 6.6% better than a model that speciﬁes the probability of takeup to be constant. To elaborate upon this, let us consider the Rp2 criterion for the logit
model. If we generate predictions ŷi on the basis of the estimated logit probabilities by
predicting a one if the estimated probability is larger than 0.5 and a zero otherwise, we
can produce the cross-tabulation in Table 7.3. The off-diagonal elements in this table
indicate the number of observations for which the model’s prediction is incorrect. It is
clear that for the majority of individuals we predict that they will apply for UI beneﬁts,
while for 171 individuals we predict that they do not apply while in fact they do. The
Rp2 criterion can be computed directly from this table as
Rp2 = 1 −

171 + 1300
,
1542

where 1542 corresponds to the number of incorrect predictions from the naive model
where the probability of take-up is constant (p̂ = 3335/4877). The loglikelihood value
for the latter model is given by
log L0 = 3335 log

3335
1542
+ 1542 log
= −3046.187,
4877
4877

which allows us to compute the pseudo and McFadden R 2 measures. Finally, we note
that p00 + p11 for this logit model is
3164
242
+
= 1.106,
1542 3335
while it is 1 for the naive model by construction.

7.1.7 Speciﬁcation Tests in Binary Choice Models

Although maximum likelihood estimators have the property of being consistent, there
is one important condition for this to hold: the likelihood function has to be correctly

MODELS WITH LIMITED DEPENDENT VARIABLES

200

speciﬁed.6 This means that we must be sure about the entire distribution that we
impose upon our data. Deviations will cause inconsistent estimators and in binary
choice models this typically arises when the probability that yi = 1 is misspeciﬁed as
a function of xi . Usually, such misspeciﬁcations are motivated from the latent variable
model and reﬂect heteroskedasticity or non-normality (in the probit case) of εi . In
addition, we may want to test for omitted variables without having to re-estimate the
model. The most convenient framework for such tests is the Lagrange multiplier (LM)
framework as discussed in Section 6.2.
LM tests are based on the ﬁrst order conditions from a more general model that
speciﬁes the alternative hypothesis, and check whether these are violated if we evaluate
them at the parameter estimates of the current, restricted, model. Thus, if we want to
test for J omitted variables zi , we should evaluate whether
N

i=1

yi − F (xi β̂)
F (xi β̂)(1 − F (xi β̂))

f (xi β̂) zi

(7.22)

is signiﬁcantly different from zero. Denoting the term in square brackets as the generalized residual, ε̂iG , this means checking whether ε̂iG and zi are correlated. As we have
seen in Section 6.2, a simple way of computing the LM test statistic is obtained from
a regression of a vector of ones upon the K + J variables ε̂iG xi and ε̂iG zi and computing N times the uncentred R 2 (see Section 2.4) of this auxiliary regression. Under
the null hypothesis that zi enters the model with zero coefﬁcients, the test statistic is
asymptotically Chi-squared distributed with J degrees of freedom.
Heteroskedasticity of εi will cause the maximum likelihood estimators to be inconsistent and we can test for it fairly easily. Consider the alternative that the variance of
εi depends upon exogenous variables7 zi as
V {εi } = kh(zi α)

(7.23)

for some function h > 0 with h(0) = 1, k = 1 or π 2 /3 (depending on whether we have
a probit or logit model) and h (0) = 0. The loglikelihood function would generalize to
log L(β, α) =

N

i=1

yi log F

xi β

h(zi α)

+

N

i=1

(1 − yi ) log 1 − F

xi β

h(zi α)

.

(7.24)
The derivatives with respect to α, evaluated under the null hypothesis that α = 0 are
given by
N

yi − F (xi β̂)
(7.25)
f (xi β̂) (xi β̂) κzi ,


F
(x
β̂)(1
−
F
(x
β̂))
i
i
i=1
6

We can relax this requirement somewhat to say that the ﬁrst order conditions of the maximum likelihood
problem should be valid (in the population). If this is the case, we can obtain consistent estimators even
with the incorrect likelihood function. This is referred to as quasi-maximum likelihood estimation (see
Section 6.4).
7
As the model describes the probability of yi = 1 for a given set of xi variables, the variables determining
the variance of εi should be in this conditioning set as well. This means that zi is a subset of (functions
of) xi . Note that it is possible that a priori restrictions on β are imposed to exclude some xi variables
from the ‘mean’ function xi β.

BINARY CHOICE MODELS

201

where κ is a constant that depends upon the form of h. Consequently, it is easy to
test H0 : α = 0 using the LM test by taking N times the uncentred R 2 of a regression
of ones upon ε̂iG xi and (ε̂iG · xi β̂)zi . Again, the test statistic is Chi-squared with J
degrees of freedom (the dimension of zi ). Because of the normalization
(the variance

is not estimated), zi should not include a constant. Also note that i ε̂iG · xi β̂ = 0 by
construction because of the ﬁrst order conditions. Although κ appears in the derivatives
in (7.25), it is just a constant and therefore irrelevant in the computation of the test
statistic. Consequently, the test for heteroskedasticity does not depend upon the form of
the function h(.), only upon the variables zi that affect the variance (compare Newey,
1985). This is similar to the Breusch–Pagan test for heteroskedasticity in the linear
regression model, as discussed in Subsections 4.4.3 and 6.3.2.
Finally, we discuss a normality test for the probit model. For a continuously observed
variable, normality tests usually check for skewness (third moment) and excess kurtosis
(fourth moment), that is they check whether E{εi3 } = 0 and E{εi4 − 3σ 4 } = 0 (compare Pagan and Vella, 1989). It is possible to derive tests for normality in the case
with non-continuous observations in this way. Alternatively, and often equivalently,
we can remain within the Lagrange multiplier framework and specify an alternative
distribution that is more general than the normal, and test the restrictions implied by the
latter. A parametrization of non-normality is obtained by stating that εi has distribution
function (compare Bera, Jarque and Lee, 1984, Ruud, 1984, or Newey, 1985)
P {εi ≤ t} = (t + γ1 t 2 + γ2 t 3 )

(7.26)

which characterizes the Pearson family of distributions (some restrictions on γ1 and γ2
apply). This class of distributions allows for skewness (γ1 = 0) and excess kurtosis (fat
tails) (γ2 = 0) and reduces to the normal distribution if γ1 = γ2 = 0. Consequently, a
test for normality is simply a test of two parametric restrictions. In the probit model
the probability that yi = 1 would more generally be described by
P {yi = 1|xi } = (xi β + γ1 (xi β)2 + γ2 (xi β)3 ).

(7.27)

This shows that a test for normality, in this case, corresponds to a test for the omitted
variables (xi β)2 and (xi β)3 . Consequently, the test statistic for the null hypothesis γ1 = γ2 = 0 is easily obtained by running an auxiliary regression of ones upon
ε̂iG xi , ε̂iG (xi β̂)2 and ε̂iG (xi β̂)3 and computing N times R 2 . Under the null, the test
statistic is Chi-squared distributed with two degrees of freedom. The two additional
terms in the regression correspond to skewness and kurtosis, respectively.
7.1.8 Relaxing Some Assumptions in Binary Choice Models

For a given set of xi variables a binary choice model describes the probability that
yi = 1 as a function of these variables. There are several ways in which the restrictions
imposed by the model can be relaxed. Almost without exception, these extensions are
within the class of single index models in which there is one function of xi that
determines all probabilities (like xi β). First, it is straightforward, using the results of
the previous subsection and analogous to linear regression models, to include nonlinear
functions of xi as additional explanatory variables. For example, if age is included in
xi , you could include age-squared as well.

202

MODELS WITH LIMITED DEPENDENT VARIABLES

Most extensions of binary choice models are motivated by the latent variable framework and involve relaxation of the distributional assumptions on the error term. For
example, one could allow that the error term εi in (7.8) is heteroskedastic. If the form
of heteroskedasticity is known, say V {εi } = exp{zi α}, where zi contains (functions of)
elements in xi and α is an unknown parameter vector, the essential change is that the
probability that yi = 1 also depends upon the error variance, that is



P {yi = 1|xi } = F xi β/ exp{zi α} ,
The parameters in β and α can be estimated simultaneously by maximizing the loglikelihood function, as given in (7.24), with h(.) as the exponential function. As in the
standard homoskedastic case we have to impose a normalization restriction, which is
done most easily by not including an intercept term in zi . In this case α = 0 corresponds to V {εi } = 1. Alternatively, one can set one of the β coefﬁcients equal to 1
or −1, preferably one corresponding to a variable that is ‘known’ to have a nonzero
effect on yi , while not imposing a restriction on the variance of εi . This is a common
normalization constraint when a semi-parametric estimator is employed.
It is also possible to estimate the parameter vector β semi-parametrically, that is
without imposing distributional assumptions on the error εi , except that it has a median
of zero and is independent of xi . Although the interpretation of the β coefﬁcients
without a distribution function F is hard (if not impossible), their signs and signiﬁcance
are of interest. A well known method is referred to as Manski’s maximum score
estimator (Manski, 1975, 1985). Essentially, it tries to maximize the number of correct
predictions based
on (7.20). This is equivalent to minimizing the number of incorrect

predictions i (yi − ŷi )2 with respect to β, where ŷi is deﬁned from (7.20). Because
this objective function is not differentiable with respect to β, Manski describes a
numerical algorithm to solve the maximization√problem. Another problem is that the
rate of convergence (to get consistency) is not N , as usual, but less (N 1/3 ). To some
extent, both problems are solved in Horowitz’s smooth maximum score estimator
(Horowitz, 1992), which is based on a smoothed version of the objective function
above. Additional details and discussion can be found in Horowitz (1998), Lee (1996,
Section 9.2), and Pagan and Ullah (1999, Chapter 7).

7.2

Multi-response Models

In many applications, the number of alternatives that can be chosen is larger than two.
For example, we can distinguish the choice between full-time work, part-time work or
not working, or the choice of a company to invest in Europe, Asia or the USA. Some
quantitative variables can only be observed to lie in certain ranges. This may be because
questionnaire respondents are unwilling to give precise answers, or are unable to do so,
perhaps because of conceptual difﬁculties in answering the question. Examples of this
are questions about income, the value of a house, or about job or income satisfaction.
Multi-response models are developed to describe the probability of each of the possible
outcomes as a function of personal or alternative speciﬁc characteristics. An important
goal is to describe these probabilities with a limited number of unknown parameters

MULTI-RESPONSE MODELS

203

and in a logically consistent way. For example, probabilities should lie between 0 and
1 and, over all alternatives, add up to one.
An important distinction exists between ordered response models and unordered
models. An ordered response model is generally more parsimonious but can only
be applied if there exists a logical ordering of the alternatives. The reason is that
there is assumed to exist one underlying latent variable that drives the choice between
the alternatives. In other words, the results will be sensitive to the ordering of the
alternatives, so this ordering should make sense. Unordered models are not sensitive
to the way in which the alternatives are numbered. In many cases, they can be based
upon the assumption that each alternative has a random utility level and that individuals
choose the alternative that yields highest utility.
7.2.1 Ordered Response Models

Let us consider the choice between M alternatives, numbered from 1 to M. If there
is a logical ordering in these alternatives (for example, no car, 1 car, more than one
car), a so-called ordered response model can be used. This model is also based on
one underlying latent variable but with a different match from the latent variable, yi∗ ,
to the observed one (yi = 1, 2, . . . , M). Usually, one says that
yi∗ = xi β + εi
yi = j

if γj −1 <

(7.28)
yi∗

≤ γj ,

(7.29)

for unknown γj s with γ0 = −∞, γ1 = 0 and γM = ∞. Consequently, the probability
that alternative j is chosen is the probability that the latent variable yi∗ is between
two boundaries γj −1 and γj . Assuming that εi is i.i.d. standard normal results in the
ordered probit model. The logistic distribution gives the ordered logit model. For
M = 2 we are back at the binary choice model.
Consider an example from the labour supply literature. Suppose married females
answer the question ‘How much would you like to work?’ in three categories ‘not’,
‘part-time’ and ‘full-time’. According to neo-classical theory, desired labour supply,
as measured by these answers, will depend upon preferences and a budget constraint.
So variables related to age, family composition, husband’s income and education level
could be of importance. To model the outcomes, yi = 1 (not working), yi = 2 (parttime working) and yi = 3 (full-time working), we note that there appears to be a logical
ordering in these answers. To be precise, the question is whether it is reasonable to
assume that there exists a single index xi β such that higher values for this index
correspond with, on average, larger values for yi . If this is the case, we can write an
ordered response model as
yi∗ = xi β + εi
(7.30)
yi = 1 if yi∗ ≤ 0,
= 2 if 0 < yi∗ ≤ γ ,
= 3 if yi∗ > γ ,

(7.31)

MODELS WITH LIMITED DEPENDENT VARIABLES

204

where we can loosely interpret yi∗ as ‘willingness to work’ or ‘desired hours of work’.
One of the boundaries is normalized to zero, which ﬁxes the location, but we also need
a normalization on the scale of yi∗ . The most natural one is that εi has a ﬁxed variance.
In the ordered probit model this means that εi is NID(0, 1). The implied probabilities
are obtained as
P {yi = 1|xi } = P {yi∗ ≤ 0|xi } = (−xi β),
P {yi = 3|xi } = P {yi∗ > γ |xi } = 1 − (γ − xi β)
and

P {yi = 2|xi } = (γ − xi β) − (−xi β),

where γ is an unknown parameter that is estimated jointly with β. Estimation is based
upon maximum likelihood, where the above probabilities enter the likelihood function.
The interpretation of the β coefﬁcients is in terms of the underlying latent variable
model (for example, a positive β means that the corresponding variable increases a
woman’s willingness to work), or in terms of the effects on the respective probabilities,
as we have seen above for the binary choice model. Suppose in the above model that
the k-th coefﬁcient, βk , is positive. This means that the latent variable yi∗ increases if
xik increases. Accordingly, the probability that yi = 3 will increase, while the probability that yi = 1 will decrease. The effect on the intermediate categories, however, is
ambiguous; the probability that yi = 2 may increase or decrease.
7.2.2 About Normalization

To illustrate the different normalization constraints that are required, let us consider a
model where such constraints are not imposed. That is,
yi∗ = β1 + xi β + εi ,
yi = 1

εi ∼ NID(0, σ 2 ).

if yi∗ ≤ γ1 ,

=2

if γ1 < yi∗ ≤ γ2 ,

=3

if yi∗ > γ2 ,

where the constant is taken out of the xi vector. As we can only observe whether yi
is 1, 2 or 3, the only elements that the data can identify are the probabilities of these
three events, for given values of xi . Not accidentally, these are exactly the probabilities
that enter the likelihood function. To illustrate this, consider the probability that yi = 1
(given xi ), given by

 
β
γ − β1
P {yi = 1|xi } = P {β1 + xi β + εi ≤ γ1 |xi } =  1
− xi
,
σ
σ
which shows that varying β, β1 , σ and γ1 does not lead a different probability as long as
β/σ and (γ1 − β1 )/σ remain the same. This reﬂects an identiﬁcation problem: different
combinations of parameter values lead to the same loglikelihood value and there is no
unique maximum. To circumvent this problem, normalization constraints are imposed.

MULTI-RESPONSE MODELS

205

The standard model imposes that σ = 1 and γ1 = 0, but it would also be possible, for
example, to set σ = 1 and β1 = 0. The interpretation of the coefﬁcients is conditional
upon a particular normalization constraint, but the probabilities are insensitive to it. In
some applications, the boundaries correspond to observed values rather than unknown
parameters and it is possible to estimate the variance of εi . This is illustrated in the
next subsection.
7.2.3 Illustration: Willingness to Pay for Natural Areas

An interesting problem in public economics is how to determine the value of a good
which is not traded. For example, what is the economic value of a public good like a
forest or ‘clean air’? In this subsection we consider an example from the contingent
valuation literature. In this ﬁeld surveys are used to elicit willingness to pay (WTP)
values for a hypothetical change in the availability of some non-market good, e.g. a
forest. Since the extensive study to measure the welfare loss to US citizens as a result
of the massive oil spill due to the grounding of the oil tanker Exxon Valdez in the Gulf
of Alaska (March 1989), the contingent valuation method plays an important role in
measuring the beneﬁts of a wide range of environmental goods.8
In this subsection, we consider a survey which has been conducted in 1997 in Portugal. The survey responses capture how much individuals are willing to pay to avoid
the commercial and tourism development of the Alentejo Natural Park in southwest
Portugal.9 To ﬁnd out what an individual’s WTP is, it is not directly asked what amount
a person would be willing to pay to preserve the park. Instead, each individual i in the
sample is faced with a (potentially) different initial bid amount BiI and asked whether
he would be willing to pay this amount or not. The interviewers used a so-called
double-bounded procedure: each person is asked on a follow-up bid which is higher
(lower) if the initial bid was accepted (rejected). For each respondent we thus have an
initial bid BiI and one of the follow-up bids BiL or BiU , where BiL < BiI < BiU . Each
person in the sample faced a random initial bid and the follow-up bid was dependent
on this amount according to the following scheme (in euro):10

Scheme
Scheme
Scheme
Scheme

1
2
3
4

Initial bid

Increased bid

Decreased bid

6
12
24
48

18
24
48
120

3
6
12
24

A person’s willingness to pay is unobserved and will be denoted by the latent variable
Bi∗ . To model how Bi∗ varies with personal characteristics xi we may want to specify
a linear relationship
Bi∗ = xi β + εi ,
(7.32)
8

A non-technical discussion of contingent valuation is given in Portney (1994), Hanemann (1994) and
Diamond and Hausman (1994).
9
I am grateful to Paulo Nunes for providing the data used in this subsection. The data set employed here
is available as WTP.
10
The original amounts are in escudos. One euro is approximately 200 escudos.

MODELS WITH LIMITED DEPENDENT VARIABLES

206

where εi is an unobserved error term, independent of xi . Four possible outcomes can
be observed, indexed by yi = 1, 2, 3, 4. In particular,
yi = 1 if both bids get rejected (Bi∗ < BiL );
yi = 2 if the ﬁrst bid gets rejected and the second gets accepted (BiL ≤ Bi∗ < BiI );
yi = 3 if the ﬁrst gets accepted while the second gets rejected (BiI ≤ Bi∗ < BiU );
yi = 4 if both bids get accepted (Bi∗ ≥ BiU ).
If we assume that εi is NID(0, σ 2 ) the above setting corresponds to an ordered probit
model. Because the boundaries BiL , BiI and BiU are observed, no normalization is
needed on σ 2 and it can be estimated. Note that in this application the latent variable
Bi∗ has the clear interpretation of a person’s willingness to pay, measured in euro.
Under the above assumptions, the probability of observing the last outcome (yi = 4)
is given by11
P {yi = 4|xi } =

P {xi β

+ εi ≥

BiU |xi }



BiU − xi β
=1−
σ


.

(7.33)

Similarly, the probability of observing the second outcome is
P {yi = 2|xi } = P {BiL ≤ xi β + εi < BiI |xi }
 I

 L

Bi − xi β
Bi − xi β
=
−
.
σ
σ

(7.34)

The other two probabilities can be derived along the same lines. These probabilities directly enter the loglikelihood function, maximization of which yields consistent
estimators for β and σ 2 .
The ﬁrst model we estimate contains an intercept only. This is of interest as it can
be interpreted as describing the (unconditional) distribution of the willingness to pay in
the population. The second model includes three explanatory variables that may affect
people’s WTP, corresponding to age, gender and income. Consequently, we estimate
two different models using maximum likelihood, one with an intercept only and one
which includes age class (from 1 to 6), a female dummy and income class (ranging
from 1 to 8). The results are presented in Table 7.4.12 In the subsample that we use, a
total of N = 312 people were interviewed, of which 123 (39%) answered no to both
bids, 18 answered no–yes, 113 yes–no and 58 answered yes to both questions.
From the model with an intercept only we see that the estimated average WTP
is almost 19 euro, with a fairly large standard deviation of 38.6 euro. Because we
assumed that the distribution of Bi∗ is normal, this implies that 31% of the population
has a negative willingness to pay.13 As this is not possible, we will reinterpret the
As Bi∗ is continuously distributed, the probability of each outcome is zero. This implies that the places
of the equality signs in the inequalities are irrelevant.
12
The results are obtained using RATS 6.0.
13
Note that P {Bi∗ < 0} = (−µ/σ ) if Bi∗ is normally distributed with mean µ and standard deviation σ .
Substituting the estimated values gives a probability of 0.31.
11

MULTI-RESPONSE MODELS

Table 7.4

207

Ordered probit model for willingness to pay
I: intercept only

Variable
constant
age class
female
income class
σ̂
Loglikelihood
Normality test (χ22 )

Estimate

s.e.

18.74
–
–
–
38.61

(2.77)

6.326

(2.11)
−409.00
(p = 0.042)

II: with characteristics
Estimate
30.55
−6.93
−5.88
4.86
36.47
2.419

s.e.
(8.59)
(1.64)
(5.07)
(1.87)
(1.89)
−391.25
(p = 0.298)

latent variable as ‘desired WTP’, the actual WTP being the maximum of zero and the
desired amount.14 In this case, actual willingness to pay, given that it is positive, is
described by a truncated normal distribution the expected value of which is estimated
to be ¤38.69.15 The estimate for the expected WTP over the entire sample is then
38.69 × 0.69 = 26.55 euro, because 31% has a zero willingness to pay. Multiplying
this by the total number of households in the population (about 3 million) produces an
estimated total willingness to pay of about 80 million euro.
The inclusion of personal characteristics is not very helpful in eliminating the problem of negative values for Bi∗ . Apparently, there is a relatively large group of people
that says no to both bids, such that the imposed normal distribution generates substantial
probability mass in the negative region. The explanatory variables that are included
are age, in six brackets (<29, 29–39, . . . , >69), a female dummy, and income (in
8 brackets). With the inclusion of these variables, the intercept term no longer has
the same interpretation as before. Now, for example, the expected willingness to pay
for a male in income class 1 (<¤375 per month) and age between 20 and 29 is
30.55 − 6.93 + 4.86 = 28.48 euro, or, taking into account the censoring, 33.01 euro.
We see that the WTP signiﬁcantly decreases with age and increases with income, while
there is no statistical evidence of a gender effect.
As in the binary probit model, the assumption of normality is crucial here for consistency of the estimators as well as the interpretation of the parameter estimates (in
terms of expected WTP). A test for normality can be computed within the Lagrange
multiplier framework discussed in Section 6.2. As before, the alternative is that the
appropriate distribution is within the Pearson family of distributions and a test for
normality tests two parametric restrictions. Unfortunately, the analytical expressions
are rather complicated and will not be presented here (see Glewwe, 1997). Under the
null hypothesis of normality, the test statistics have a Chi-squared distribution with
two degrees of freedom. The two statistics in the table indicate a marginal rejection of
normality in the simple model with an intercept only, but do not lead to rejection of
the model with individual characteristics.
14
15

This interpretation is similar to the one employed in tobit models. See below.
If y ∼ N(µ, σ 2 ) we have that E{y|y > c} = µ + σ λ([c − µ]/σ ), where λ(t) = φ(−t)/(−t) ≥ 0. See
Appendix B for details.

MODELS WITH LIMITED DEPENDENT VARIABLES

208

7.2.4 Multinomial Models
In several cases, there is no natural ordering in the alternatives and it is not realistic to
assume that there is a monotonic relationship between one underlying latent variable
and the observed outcomes. Consider, for example, modelling the mode of transportation (bus, train, car, bicycle, walking). In such cases, an alternative framework has to
be used to put some structure on the different probabilities. A common starting point
is a random utility framework, in which the utility of each alternative is a linear function of observed characteristics (individual and/or alternative speciﬁc) plus an additive
error term. Individuals are assumed to choose the alternative that has the highest utility.
With appropriate distributional assumptions on these error terms, this approach leads
to manageable expressions for the probabilities implied by the model.
To formalize this, suppose that there is a choice between M alternatives, indexed
j = 1, 2, . . . , M, noting that the order is arbitrary. Next, assume that the utility level
that individual i attaches to each of the alternatives is given by Uij , j = 1, 2, . . . , M.
Then alternative j is chosen by individual i if it gives highest utility, that is if Uij =
max{Ui1 , . . . , UiM }. Of course, these utility levels are not observed and we need to
make some additional assumptions to make this set-up operational. Let us assume that
Uij = µij + εij , where µij is a non-stochastic function of observables and a small
number of unknown parameters, and εij is an unobservable error term. From this, it
follows that

P {yi = j } = P {Uij = max{Ui1 , . . . , UiM }}


max {µik + εik } .
= P µij + εij >
k=1,...,J,k =j

(7.35)

To evaluate this probability, we need to be able to say something about the maximum
of a number of random variables. In general, this is complicated, but a very convenient
result arises if we can assume that all εij are mutually independent with a so-called
log Weibull distribution (also known as a Type I extreme value distribution). In this
case, the distribution function of each εij is given by
F (t) = exp{−e−t },

(7.36)

which does not involve unknown parameters. Under these assumptions, it can be
shown that
exp{µij }
P {yi = j } =
.
(7.37)
exp{µi1 } + exp{µi2 } + · · · + exp{µiM }
Notice
that this structure automatically implies that 0 ≤ P {yi = j } ≤ 1 and that
M
P
j =1 {yi = j } = 1.
The distribution of εij sets the scaling of utility (which is undeﬁned) but not the
location. To solve this, it is common to normalize one of the deterministic utility levels
to zero, say µi1 = 0. Usually, µij is assumed to be a linear function of observable
variables, which may depend upon the individual (i), the alternative (j ), or both. Thus
we write µij = xij β. With this we obtain
P {yi = j } =

exp{xij β}



1 + exp{xi2
β} + · · · + exp{xiM
β}

,

j = 1, 2, . . . , M.

(7.38)

MULTI-RESPONSE MODELS

209

This constitutes the so-called multinomial logit model or independent logit model; for
details of the genesis of this model, see Greene (2003, Section 21.7). If there are only
two alternatives (M = 2), it reduces to the standard binary logit model. The probability
of an individual choosing alternative j is a simple expression of explanatory variables
and coefﬁcients β because of the convenient assumption made about the distribution
of the unobserved errors. As before, the multinomial model is estimated by maximum
likelihood, where the above probabilities enter the likelihood function.
Typical things to include in xij β are alternative-speciﬁc characteristics. When
explaining the mode of transportation, it may include variables like travelling time and
costs, which may vary from one person to another. A negative β coefﬁcient then means
that the utility of an alternative is reduced if travelling time is increased. Consequently,
if travelling time of one of the alternatives is reduced (while the other alternatives are
not affected), this alternative will get a higher probability of being picked. Other things
to include in xij β are personal characteristics (like age and gender), with coefﬁcients
that are alternative-speciﬁc. This could reveal, for example, that, ceteris paribus, men
are more likely to travel by car than women.
Despite the attractiveness of the analytical expressions in the multinomial logit
model, it has one big drawback, which is due to the assumption that all εij s are
independent. This implies that (conditional upon observed characteristics) utility levels
of any two alternatives are independent. This is particularly troublesome if two or more
alternatives are very similar. A typical example would be to decompose the category
‘travel by bus’, into ‘travel by blue bus’ and ‘travel by red bus’. Clearly, we would
expect that a high utility for a red bus implies a high utility for a blue bus. Another
way to see the problem is to note that the probability ratio of two alternatives does
not depend upon the nature of any of the other alternatives. Suppose that alternative 1
denotes travel by car and alternative 2 denotes travel by (blue) bus. Then the probability
ratio (or odds ratio) is given by
P {yi = 2}

= exp{xi2
β}
P {yi = 1}

(7.39)

irrespective of whether the third alternative is a red bus or a train. Clearly, this is
something undesirable. McFadden (1974) called this property of the multinomial logit
model independence of irrelevant alternatives (IIA). Hausman and McFadden (1984)
propose a test for the IIA restriction based on the result that the model parameters
can be estimated consistently by applying a multinomial logit model to any subset
of alternatives (see Franses and Paap, 2001, Section 5.3, for details). The test compares the estimates from the model with all alternatives to estimates using a subset of
alternatives.
Let us consider a simple example from marketing, which involves stated preferences (rather than observed choices). Suppose that a number of respondents are asked
to pick their preferred coffee-maker from a set of ﬁve, say, alternative combinations
of characteristics (capacity, price, special ﬁlter (yes/no) and thermos ﬂask (yes/no)).
Typically, the combinations are not the same for all respondents. Let us refer to these
characteristics as xij . To make sure that µi1 = 0, the xij are measured in differences
from a reference coffee-maker, which without loss of generality corresponds to alternative 1. The probability that a respondent selects alternative j can be (assumed to be)

MODELS WITH LIMITED DEPENDENT VARIABLES

210

described by a multinomial logit model, with
P {yi = j } =

exp{xij β}



1 + exp{xi2
β} + · · · + exp{xi5
β}

.

(7.40)

A positive β coefﬁcient implies that people attach positive utility to the corresponding
characteristic.
Under appropriate assumptions, the estimated model can be used to predict the
probability of an individual choosing an alternative that is not yet on the market,
provided this alternative is a (new) combination of existing characteristics. To illustrate
this, suppose the current market for coffee-makers consists of two products: a machine
for 10 cups without ﬁlter and thermos for 25 euro (z1 ) and a machine for 15 cups
with ﬁlter for 35 euro (z2 ), while brand X is considering to introduce a new product:
a machine for 12 cups with ﬁlter and thermos for 33 euro (z3 ). If the respondents
are representative for those who buy coffee-makers, the expected market share of this
new product corresponds to the probability of preferring the new machine to the two
existing ones, and could be estimated as
exp{(z3 − z1 ) β̂}
1 + exp{(z2 − z1 ) β̂} + exp{(z3 − z1 ) β̂}

,

where β̂ is the maximum likelihood estimate for β. In fact, it would be possible to
select an optimal combination of characteristics in z3 so as to maximize this estimated
market share.16
While it is possible to relax the IIA property, this generally leads to (conceptually
and computationally) more complicated models (see, for example, Amemiya, 1981,
or Maddala, 1983). The multinomial probit model is obtained if it is assumed that
the εij are jointly normally distributed with arbitrary correlations. For this model, the
response probabilities are rather complicated expressions involving an M − 1 dimensional integral, so that maximum likelihood estimation using numerical integration is
only possible with a limited number of alternatives. Estimation techniques based on
simulation make estimation with more than four alternatives feasible; see Keane (1993)
for a survey.
In some applications, the choice between M alternatives can be decomposed into
two or more sequential choices. A popular speciﬁcation is the nested logit model,
which is appropriate if the alternatives can be divided into S groups, where the IIA
assumption holds within each group but not across groups. To illustrate this, suppose
the three relevant alternatives in the mode of transportation example are: travel by
car, train or bus. We may divide these alternatives into private and public modes of
transportation. Then the ﬁrst choice is between private and public, while the second
one is between train and bus, conditional upon the ﬁrst choice being public transport.
It is possible to model these two choices by two (bivariate) logit models; see Franses
and Paap (2001, Section 5.1) or Wooldridge (2002, Section 15.9) for more details.
16

This example is clearly oversimpliﬁed. In marketing applications the property of independence of irrelevant alternatives is often considered unsatisfactory. Moreover, the model does not take into account
observed and unobserved heterogeneity across consumers. See Louviere (1988) or Caroll and Green
(1995) for some additional discussion.

MODELS FOR COUNT DATA

7.3

211

Models for Count Data

In certain applications, we would like to explain the number of times a given event
occurs, for example, how often a consumer visits a supermarket in a given week, or
the number of patents a ﬁrm has obtained in a given year. Clearly, the outcome might
be zero for a substantial part of the population. While the outcomes are discrete and
ordered, there are two important differences with ordered response outcomes. First,
the values of the outcome have a cardinal rather than just an ordinal meaning (4 is
twice as much as 2 and 2 is twice as much as 1). Second, there (often) is no natural
upper bound to the outcomes. As a result, models for count data are very different
from ordered response models.
7.3.1 The Poisson and Negative Binomial Models

Let us denote the outcome variable by yi , taking values 0, 1, 2, . . . . Our goal is to
explain the distribution of yi , or the expected value of yi , given a set of characteristics
xi . Let us assume that the expected value of yi , given xi , is given by
E{yi |xi } = exp{xi β},

(7.41)

where β is a set of unknown parameters. Because yi is non-negative, we choose a functional form that produces non-negative conditional expectations. The above assumption
relates the expected outcome of yi to the individual characteristics in xi , but does not
fully describe the distribution. If we want to determine the probability of a given
outcome (e.g. P {yi = 1|xi }), additional assumptions are necessary.
A common assumption in count data models is that, for given xi , the count variable
yi has a Poisson distribution with expectation λi ≡ exp{xi β}. This implies that the
probability mass function of yi conditional upon xi is given by
y

P {yi = y|xi } =

exp{−λi }λi
,
y!

y = 0, 1, 2, . . . ,

(7.42)

where y! is short hand notation for y × (y − 1) × · · · × 2 × 1 (referred to as ‘y factorial’), with 0! = 1. Substituting the appropriate functional form for λi produces
expressions for the probabilities that can be used to construct the loglikelihood function
for this model, referred to as the Poisson regression model. Assuming that observations on different individuals are mutually independent, estimation of β by means of
maximum likelihood is therefore reasonably simple: the loglikelihood function is the
sum of the appropriate log probabilities, interpreted as a function of β. If the Poisson distribution is correct, and assuming we have a random sample of yi and xi , this
produces a consistent, asymptotically efﬁcient and asymptotically normal estimator
for β.
To illustrate the above probabilities, consider an individual characterized by λi = 2.
For this person, the probabilities of observing yi = 0, 1, 2, 3 are given by 0.135, 0.271,
0.271 and 0.180, respectively (such that the probability of observing 4 or more events is
0.143). The expected value of yi corresponds to the weighted average of all outcomes,
weighted by their respective probabilities, and is equal to λi = 2. The speciﬁcation
in (7.41) and (7.42) allows λi and the probabilities to vary with xi . In particular, the

MODELS WITH LIMITED DEPENDENT VARIABLES

212

parameters in β indicate how the expected value of yi varies with xi (taking into
account the exponential function). For an individual with expected value λi = 3, the
probabilities of yi = 0, 1, 2, 3 change to 0.050, 0.149, 0.224, 0.224, respectively (with
a probability of 0.353 of observing 4 or more).
An important drawback of the Poisson distribution is that it automatically implies
that the conditional variance of yi is also equal to λi . That is, in addition to (7.41), the
assumption in (7.42) implies that
V {yi |xi } = exp{xi β}.

(7.43)

This condition is referred to as equidispersion and illustrates the restrictive nature of
the Poisson distribution. In many applications, the equality of the conditional mean
and variance of the distribution has been rejected. A wide range of alternative count
distributions have been proposed that do not impose (7.43); see Cameron and Trivedi
(1998) or Winkelmann (2003) for an overview. Alternatively, it is possible to consistently estimate the conditional mean in (7.41), without specifying the conditional
distribution, like we did in (7.42). In fact, the Poisson regression model is able to do
so even if the Poisson distribution is invalid. This is because the ﬁrst order conditions of the maximum likelihood problem are valid more generally, so that we can
obtain a consistent estimator for β using the quasi-maximum likelihood approach, as
discussed in Section 6.4 (see Wooldridge, 2002, Section 19.2). This means that we
solve the usual maximum likelihood problem, but adjust the way in which standard
errors are computed. Several software packages provide computation of such ‘robust’
standard errors.
To illustrate the (quasi-) maximum likelihood approach, consider the loglikelihood
function of the Poisson regression model (assuming a random sample of size N ),
given by
N

log L(β) =
[−λi + yi log λi − log yi !]
(7.44)
i=1

=

N


[− exp{xi β} + yi (xi β) − log yi !].

i=1

The last term in this expression is typically dropped, because it does not depend upon
the unknown parameters. For computational reasons, this may be a good idea too, as
yi ! may become very large. The ﬁrst order conditions of maximizing log L(β) with
respect to β are given by
N

i=1

(yi −

exp{xi β})xi

=

N


εi xi = 0,

(7.45)

i=1

where the ﬁrst equality deﬁnes the error term εi = yi − exp{xi β}. Because (7.41)
implies that E{εi |xi } = 0, we can interpret (7.45) as the sample moment conditions
corresponding to the set of orthogonality conditions E{εi xi } = 0. As a result, the estimator that maximizes (7.44) is generally consistent under condition (7.41), even if yi

MODELS FOR COUNT DATA

213

given xi does not have a Poisson distribution. In this case, we refer to the estimator as
a quasi-maximum likelihood estimator (QMLE).
Using (7.44) and from the general discussion on maximum likelihood estimation in
Section 6.1, we can easily derive the asymptotic covariance matrix of the ML estimator.
In the i.i.d. case, it is given by
VMLE = I (β)−1 = (E{exp{xi β}xi xi })−1 .

(7.46)

However, for the quasi-maximum likelihood estimator β̂QMLE , it follows from the
results in Section 6.4 that the appropriate asymptotic covariance matrix is

where

VQMLE = I (β)−1 J (β)I (β)−1 ,

(7.47)

J (β) = E{[yi − exp{xi β}]2 xi xi } = E{εi2 xi xi }.

(7.48)

These covariance matrices can easily be estimated by replacing expectations by sample averages and unknown parameters by their ML estimates. The QMLE covariance
matrix is similar to the White covariance matrix used for the linear regression model. If
V {yi |xi } = E{εi2 |xi } > exp{xi β}
we have a case of overdispersion. In such a case, it follows from (7.47) and (7.48)
that the variance of the quasi-maximum likelihood estimator may be much larger than
suggested by (7.46).
Despite its robustness, a disadvantage of the quasi-maximum likelihood approach is
that it does not allow to compute conditional probabilities, as in (7.42). All we impose
and estimate is (7.41). Consequently, it is not possible to determine, for example, what
the probability is that a given ﬁrm has zero patents in a given year, conditional upon its
characteristics, unless we are willing to make additional assumptions. Of course, from
(7.41) we can determine what the expected number of patents is for the above ﬁrm.
Alternative more general count data models are therefore useful. One alternative is the
application of a full maximum likelihood analysis of the NegBin I model of Cameron
and Trivedi (1986). NegBin I is a special case of the negative binomial distribution. It
imposes that
V {yi |xi } = (1 + δ 2 ) exp{xi β}
(7.49)
for some δ 2 > 0 to be estimated. As a result, the NegBin I model allows for overdispersion (relative to the Poisson regression model). Unfortunately, the NegBin I maximum
likelihood estimators are consistent only if (7.49) is valid and thus do not have the
robustness property of the (quasi-) maximum likelihood estimators of the Poisson
model. If (7.49) is valid, the NegBin I estimates are more efﬁcient than the Poisson
estimates. A further generalization is the NegBin II model, which assumes
V {yi |xi } = (1 + α 2 exp{xi β}) exp{xi β},

(7.50)

for some α 2 > 0, where the amount of overdispersion is increasing with the conditional
mean E{yi |xi } = exp{xi β}; see Cameron and Trivedi (1986) for more details. In many

MODELS WITH LIMITED DEPENDENT VARIABLES

214

software packages, the NegBin II model is referred to as the ‘negative binomial model’.
The NegBin I model is quite popular in the statistics literature, because it is a special
case of the ‘generalized linear model’ (see Cameron and Trivedi, 1997, Section 2.4.4).
Unlike the NegBin I model, the maximum likelihood estimator for the NegBin II
model is robust to distributional misspeciﬁcation. Thus, provided the conditional mean
is correctly speciﬁed, the NegBin II maximum likelihood estimator is consistent for β.
The associated maximum likelihood standard errors, however, will only be correct if
the distribution is correctly speciﬁed (see Cameron and Trivedi, 1997, Section 3.3.4).
Given that maximum likelihood estimation of the negative binomial models is fairly
easy using standard software, a test of the Poisson distribution is often carried out by
testing δ 2 = 0 or α 2 = 0 using a Wald or likelihood ratio test. Rejection is an indication
of overdispersion. The alternative hypotheses are one-sided and given by δ 2 > 0 and
α 2 > 0, respectively. Because δ 2 and α 2 cannot be negative, the distribution of the Wald
and LR test statistics is nonstandard (see Cameron and Trivedi, 1997, Section 3.4). In
practice, this problem only affects the appropriate critical values. Rather than using the
95% critical value for the Chi-squared distribution, one should use the 90% percentile
to test with 95% conﬁdence. That is, the null hypothesis of no overdispersion is rejected
with 95% conﬁdence if the test statistic exceeds 2.71 (rather than 3.84).
All three models presented above state that the variance of yi is larger if the expected
value of yi is larger. The Poisson model is very restrictive in that it imposes that the
variance and the mean are equal. The NegBin I model allows the variance to exceed the
mean, but imposes that their ratio is the same for all observations (and equals 1 + δ 2 ).
The NegBin II model allows the variance to exceed the mean, their ratio being larger
for units that have a high mean. In this case, the amount of overdispersion increases
with the conditional mean.
The easiest way to interpret the coefﬁcients in count data models is through the
conditional expectation in (7.41). Suppose that xik is a continuous explanatory variable.
The impact of a marginal change in xik upon the expected value of yi (keeping all
other variables ﬁxed) is given by
∂E{yi |xi }
= exp{xi β}βk ,
∂xik

(7.51)

which has the same sign as the coefﬁcient βk . The exact response depends upon the
values of xi through the conditional expectation of yi . This expression can be evaluated
for the ‘average’ individual in the sample, using sample averages of xi , or for each
individual in the sample separately. A more attractive approach is to convert this
response into a semi-elasticity. Computing
βk =

∂E{yi |xi }
1
∂xik E{yi |xi }

(7.52)

provides the relative change in the conditional mean if the k-th regressor changes by
one unit (ceteris paribus). If xik is a logarithm of an explanatory variable, say Xik , then
βk measures the elasticity of yi with respect to Xik . That is, it measures the relative
change in the expected value of yi if Xik changes by 1%.
For a discrete variable, the above calculus methods are inappropriate. Consider a
binary variable xik that only takes the values 0 and 1. Then, we can compare the

MODELS FOR COUNT DATA

215

conditional means of yi , given xik = 0 and given xik = 1, keeping the other variables
in xi ﬁxed. It is easily veriﬁed that
E{yi |xik = 1, xi∗ }
= exp{βk },
E{yi |xik = 0, xi∗ }

(7.53)

where xi∗ denotes the vector xi , excluding its k-th element. Thus, the conditional mean is
exp{βk } times larger if the binary indicator is equal to one rather than zero, irrespective
of the values of the other explanatory variables. For small values of βk , we have that
exp{βk } ≈ 1 + βk . For example, a value of βk = 0.05 indicates that the expected value
of yi increases by approximately 5% if the indicator variable changes from 0 to 1.
7.3.2 Illustration: Patents and R&D Expenditures

The relationship between research and development expenditures of ﬁrms and the
number of patents applied and received by them has received substantial attention in
the literature; see Hausman, Hall and Griliches (1984). Because the number of patents
is a count variable, ranging from zero to many, count data models are commonly
applied to this problem. In this subsection, we consider a sample of 181 international
manufacturing ﬁrms, taken from Cincera (1997). For each ﬁrm, we observe annual
expenditures on research and development (R&D), the industrial sector it operates in,
the country of its registered ofﬁce and the total number of patent applications for a
number of consecutive years. We shall use the information on 1991 only.17
The average number of patent applications in 1991 was 73.6, with a minimum of 0
and a maximum of 925. About 10% of the ﬁrms in the sample have zero applications,
while the median number is 20. Given the large spread in the number of applications,
with a sample standard deviation of 151, the unconditional count distribution is clearly
overdispersed and far away from a Poisson distribution. The inclusion of conditioning
explanatory variables may reduce the amount of overdispersion. However, given the
descriptive statistics it seems unlikely that it would eliminate it completely.
Each of the models we consider states that the expected number of patents yi is
given by
E{yi |xi } = exp{xi β},
(7.54)
where xi contains a constant, the log of R&D expenditures, industry and geographical dummies. Despite its restrictive nature, the ﬁrst model we consider is the Poisson
regression model, which assumes that yi conditional upon xi follows a Poisson distribution. The maximum likelihood estimation results are presented in Table 7.5. The sector
dummies refer to aerospace, chemistry, computers (hardware and software), machinery
and instruments and motor vehicles. These estimates suggest that aerospace and motor
vehicles are sectors with a relatively low number of patent applications, whereas the
chemistry, computers and machines sectors have relatively high numbers of applications. The reference category for the geographical dummies is Europe, although there
is one ﬁrm located in ‘the rest of the world’. The estimates indicate clear differences
between Japan, Europe and the United States of America in terms of the expected
17

The data for this illustration are available as PATENTS.

MODELS WITH LIMITED DEPENDENT VARIABLES

216

Table 7.5

Estimation results Poisson model, MLE and QMLE
MLE

constant
log (R&D)
aerospace
chemistry
computers
machines
vehicles
Japan
USA
Loglikelihood
Pseudo R 2
LR test (χ82 )
Wald test (χ82 )

QMLE

Estimate

Standard error

Robust s.e.

−0.8737
0.8545
−1.4218
0.6363
0.5953
0.6890
−1.5297
0.2222
−0.2995

0.0659
0.0084
0.0956
0.0255
0.0233
0.0383
0.0419
0.0275
0.0253

0.7429
0.0937
0.3802
0.2254
0.3008
0.4147
0.2807
0.3528
0.2736

−4950.789
0.675
20587.54 (p = 0.000)

338.9 (p = 0.000)

number of applications. The high levels of signiﬁcance are striking and somewhat suspicious. However, one should keep in mind that the standard errors are only valid if
the Poisson distribution is correct, which seems unlikely given the amount of overdispersion in the count variable. Nevertheless, the estimator is consistent as long as (7.54)
is correct, even if the Poisson distribution is invalid. In this case, we need to compute standard errors using the more general expression for the covariance matrix (see
(7.47)). Such standard errors of the quasi-maximum likelihood estimator are provided
in the third column of Table 7.5, and are substantially higher than those in column 2.
As a result, statistical signiﬁcance is reduced considerably. For example, we no longer
ﬁnd that the Japanese and US ﬁrms are signiﬁcantly different from European ones. The
huge difference between the alternative standard errors is a strong indication of model
misspeciﬁcation. That is, the Poisson distribution has to be rejected (even though we
did not perform a formal misspeciﬁcation test). Nevertheless, the conditional mean in
(7.54) may still be correctly speciﬁed.
The likelihood ratio and Wald test statistics reported in Table 7.5 provide tests for
the hypothesis that all coefﬁcients in the model except the intercept term are equal
to zero. The Wald test is based on the robust covariance matrix and therefore more
appropriate than the likelihood ratio test, which assumes that the Poisson distribution
is correct. The Wald test strongly rejects the hypothesis that the conditional mean
is constant and independent of the explanatory variables. The pseudo R 2 reported in
the table is the likelihood ratio index (see Subsection 7.1.5), as it is computed by
many software packages. As in all nonlinear models, there is no universal deﬁnition of a goodness-of-ﬁt measure in models for count data. Cameron and Windmeijer
(1996) discuss several alternative measures, which are typically considered more appropriate.
Because the coefﬁcient for the log of R&D expenditures has the interpretation of an
elasticity, the estimated value of 0.85 implies that the expected number of patents
increases by 0.85% if R&D expenditures (ceteris paribus) increase by 1%.18 The
18

A speciﬁcation including R&D expenditures from the previous year (1990) yields almost identical results.

MODELS FOR COUNT DATA

Table 7.6

217

Estimation results NegBin I and NegBin II model, MLE
NegBin I (MLE)

constant
log (R&D)
aerospace
chemistry
computers
machines
vehicles
Japan
USA
δ2
Loglikelihood
Pseudo R 2
LR test (χ82 )

Estimate

Standard error

0.6899
0.5784
−0.7865
0.7333
0.1450
0.1559
−0.8176
0.4005
0.1588
95.2437

0.5069
0.0676
0.3368
0.1852
0.2063
0.2550
0.2686
0.2573
0.1984
14.0069

−848.195
0.944
88.55 (p = 0.000)

NegBin II (MLE)
Estimate

α2

−0.3246
0.8315
−1.4975
0.4886
−0.1736
0.0593
−1.5306
0.2522
−0.5905
1.3009

Standard error
0.4982
0.0766
0.3772
0.2568
0.2988
0.2793
0.3739
0.4264
0.2788
0.1375

−819.596
0.946
145.75 (p = 0.000)

estimated coefﬁcient of −1.42 for aerospace indicates that, ceteris paribus, the average number of patents in the aerospace industry is 100[exp(−1.4218) − 1] = −75.9%
less than in the reference industries (food, fuel, metal and others). The hardware and
software industry has an expected number of patents that are 100[exp(0.5953) − 1] =
81.3% higher. These numbers are statistically signiﬁcant at the 95% level when using
the robust standard errors.
In Table 7.6, we present the estimation results for two alternative models: the NegBin
I and the NegBin II model. These two models specify a negative binomial distribution
for the number of patents and differ in the speciﬁcation for the conditional variance.
The NegBin I model implies a constant dispersion, according to (7.49), whereas the
NegBin II allows the dispersion to depend upon the conditional mean according to
(7.50). The two models reduce to the Poisson regression model when δ 2 = 0 or α 2 = 0,
respectively. The two Wald tests for overdispersion, based on δ 2 and α 2 , strongly
reject the null hypothesis. Again, these results indicate that the Poisson model should
be rejected.
Within the maximum likelihood framework, the NegBin II model is preferred here
to NegBin I because it has a higher loglikelihood value, with the same number of
parameters. Note that the loglikelihood values are substantially larger (less negative)
than the −4950.789 reported for the Poisson regression model. Interestingly, the estimated coefﬁcients for the NegBin I speciﬁcation are quite different from those for the
NegBin II model, as well as from the Poisson quasi-maximum likelihood estimates.
For example, the estimated elasticity of R&D expenditures is as low as 0.58 for the
NegBin I model. Given that the NegBin II estimates, unlike the NegBin I estimates, are
robust to misspeciﬁcation of the conditional variance, this ﬁnding is also unfavourable
to the NegBin I model. If the NegBin II model is correctly speciﬁed we expect that
estimation by maximum likelihood is more efﬁcient than the robust quasi-maximum
likelihood estimator based upon the Poisson loglikelihood function. The standard errors
in Tables 7.5 and 7.6 are consistent with this suggestion.

MODELS WITH LIMITED DEPENDENT VARIABLES

218

7.4

Tobit Models

In certain applications, the dependent variable is continuous, but its range may be
constrained. Most commonly this occurs when the dependent variable is zero for a
substantial part of the population but positive (with many different outcomes) for the
rest of the population. Examples are: expenditures on durable goods, hours of work,
and the amount of foreign direct investment of a ﬁrm. Tobit models are particularly
suited to model this type of variable. The original tobit model was suggested by James
Tobin (Tobin, 1958), who analysed household expenditures on durable goods taking
into account their non-negativity, while only in 1964 Arthur Goldberger referred to
this model as a tobit model, because of its similarity to probit models. Since then,
the original model has been generalized in many ways. In particular since the survey
by Amemiya (1984), economists also refer to these generalizations as tobit models. In
this section and the next we present the original tobit model and some of its extensions. More details can be found in Maddala (1983), Amemiya (1984), Lee (1996)
and Wooldridge (2002).
7.4.1 The Standard Tobit Model

Suppose that we are interested in explaining the expenditures on tobacco of US households in a given year. Let y denote the expenditures on tobacco, while z denotes all
other expenditures (both in US$). Total disposable income (or total expenditures) is
denoted by x. We can think of a simple utility maximization problem, describing the
household’s decision problem:
max U (y, z)
(7.55)
y,z

y+z≤x

(7.56)

y, z ≥ 0.

(7.57)

The solution to this problem depends, of course, on the form of the utility function
U . As it is unrealistic to assume that some households would spend all their money on
tobacco, the corner solution z = 0 can be excluded a priori. However, the solution for
y will be zero or positive and we can expect a corner solution for a large proportion
of households. Let us denote the solution to (7.55)–(7.56) without the constraint in
(7.57) as y ∗ . Under appropriate assumptions on U this solution will be linear in x. As
economists we do not observe everything that determines the utility that a household
attaches to tobacco. We account for this by allowing for unobserved heterogeneity in
the utility function and thus for unobserved heterogeneity in the solution as well. Thus
we write
y ∗ = β1 + β2 x + ε,
(7.58)
where ε corresponds to unobserved heterogeneity.19 So, if there were no restrictions
on y and consumers could spend any amount on tobacco, they would choose to spend
19

Alternative interpretations of ε are possible. These may involve optimization errors of the household or
measurement errors.

TOBIT MODELS

219

y ∗ . The solution to the original, constrained problem, will therefore be given by
y = y∗

if y ∗ > 0

y=0

if y ∗ ≤ 0

(7.59)

So if a household would like to spend a negative amount y ∗ , it will spend nothing
on tobacco. In essence, this gives us the standard tobit model, which we formalize
as follows.
yi∗ = xi β + εi , i = 1, 2, . . . , N,
yi = yi∗
=0

if yi∗ > 0
if

yi∗

(7.60)

≤ 0,

where εi is assumed to be NID(0, σ 2 ) and independent of xi . Notice the similarity
of this model with the standard probit model as given in (7.10); the difference is in
the mapping from the latent variable to the observed variable. (Also note that we can
identify the scaling here, so that we do not have to impose a normalization restriction.)
The model in (7.60) is also referred to as the censored regression model. It is a
standard regression model, where all negative values are mapped to zeros. That is,
observations are censored (from below) at zero. The model thus describes two things.
One is the probability that yi = 0 (given xi ), given by
P {yi = 0} = P {yi∗ ≤ 0} = P {εi ≤ −xi β}


  
  
xβ
xβ
xβ
εi
=P
≤− i
= − i
=1− i
.
σ
σ
σ
σ

(7.61)

The other is the distribution of yi given that it is positive. This is a truncated normal
distribution with expectation
E{yi |yi > 0} = xi β + E{εi |εi > −xi β} = xi β + σ

φ(xi β/σ )
.
(xi β/σ )

(7.62)

The last term in this expression denotes the conditional expectation of a mean-zero
normal variable given that it is larger than −xi β (see Appendix B). Obviously, this
expectation is larger than zero. The result in (7.62) also shows why it is inappropriate
to restrict attention to the positive observations only and estimate a linear model from
this subsample: the conditional expectation of yi no longer equals xi β, but also depends
nonlinearly on xi through φ(.)/(.).
The coefﬁcients in the tobit model can be interpreted in a number of ways, depending
upon one’s interest. For example, the tobit model describes the probability of a zero
outcome as
P {yi = 0} = 1 − (xi β/σ ).
This means that β/σ can be interpreted in a similar fashion as β in the probit model
to determine the marginal effect of a change in xik upon the probability of observing

MODELS WITH LIMITED DEPENDENT VARIABLES

220

a zero outcome (compare Subsection 7.1.2). That is,
β
∂P {yi = 0}
= −φ(xi β/σ ) k .
∂xik
σ

(7.63)

Moreover, as shown in (7.62), the tobit model describes the expected value of yi given
that it is positive. This shows that the marginal effect of a change in xik upon the value
of yi , given the censoring, will be different from βk . It will also involve the marginal
change in the second term of (7.62), corresponding to the censoring. From (7.62) it
follows that the expected value of yi is given by20
E{yi } = xi β(xi β/σ ) + σ φ(xi β/σ ).

(7.64)

From this it follows that the marginal effect on the expected value of yi of a change
in xik is given by21
∂E{yi }
= βk (xi β/σ ).
(7.65)
∂xik
This tells us that the marginal effect of a change in xik upon the expected outcome yi
is given by the model’s coefﬁcient multiplied by the probability of having a positive
outcome. If this probability is one for a particular individual, the marginal effect is
simply βk , as in the linear model. Finally, the marginal effect upon the latent variable
is easily obtained as
∂E{yi∗ }
= βk .
(7.66)
∂xik
Unless the latent variable has a direct interpretation, which is typically not the case, it
seems most natural to be interested in (7.65).
7.4.2 Estimation

Estimation of the tobit model is usually done through maximum likelihood. The contribution to the likelihood function of an observation either equals the probability mass
(at the observed point yi = 0) or the conditional density of yi , given that it is positive,
times the probability mass of observing yi > 0. The loglikelihood function can thus
be written as


log P {yi = 0} +
[log f (yi |yi > 0) + log P {yi > 0}]
log L1 (β, σ 2 ) =
i∈I0

=


i∈I0

i∈I1

log P {yi = 0} +



log f (yi ),

(7.67)

i∈I1

where f (.) is generic notation for a density function and the last equality follows from
the deﬁnition of a conditional density.22 The index sets I0 and I1 are deﬁned as the sets
Use that E{y} = E{y|y > 0}P {y > 0} + 0.
This is obtained by differentiating with respect to xik , using the chain rule and using the functional form
of φ. Several terms cancel out (compare Greene, 2003, Section 22.3).
22
Recall that f (y|y > c) = f (y)/P {y > c} for y > c and 0 otherwise (see Appendix B).
20
21

TOBIT MODELS

221

of those indices corresponding to the zero and the positive observations, respectively.
That is, I0 = {i = 1, . . . , N : yi = 0}. Using the appropriate expressions for the normal
distribution, we obtain
log L1 (β, σ 2 ) =


  
xβ
log 1 −  i
σ
i∈I0




1 (yi − xi β)2
1
+
.
log √
exp −
2
σ2
2πσ 2
i∈I1


(7.68)

Maximization of (7.68) with respect to β and σ 2 yields the maximum likelihood
estimates, as usual. Assuming that the model is correctly speciﬁed, this gives us
consistent and asymptotically efﬁcient estimators for both β and σ 2 (under mild regularity conditions).
The parameters in β have a double interpretation: one as the impact of a change in
xi on the probability of a non-zero expenditure, and one as the impact of a change
in xi on the level of this expenditure. Both effects thus automatically have the same
sign. Although we motivated the tobit model above through a utility maximization
framework, this is usually not the starting point in applied work: yi∗ could simply be
interpreted as ‘desired expenditures’, with actual expenditures being equal to zero if
the desired quantity is negative.
In some applications, observations are completely missing if yi∗ ≤ 0. For example,
our sample may be restricted to households with positive expenditures on tobacco
only. In this case, we can still assume the same underlying structure but with a slightly
different observation rule. This leads to the so-called truncated regression model.
Formally, it is given by
yi∗ = xi β + εi , i = 1, 2, . . . , N,
(7.69)
yi = yi∗

if yi∗ > 0

(yi , xi ) not observed if yi∗ ≤ 0,
where, as before, εi is assumed to be NID(0, σ 2 ) and independent of xi . In this case we
no longer have a random sample and we have to take this into account when making
inferences (e.g. estimating β, σ 2 ). The likelihood contribution of an observation i is
not just the density evaluated at the observed point yi but the density at yi conditional
upon selection into the sample, i.e. conditional upon yi > 0. The loglikelihood function
for the truncated regression model is thus given by
log L2 (β, σ 2 ) =


i∈I1

log f (yi |yi > 0) =



[log f (yi ) − log P {yi > 0}],

(7.70)

i∈I1

which, for the normal distribution, reduces to


  
  1
xβ
1 (yi − xi β)2
exp −
log √
− log  i
.
log L2 (β, σ ) =
2
2
σ
σ
2πσ 2
i∈I1
(7.71)
2

MODELS WITH LIMITED DEPENDENT VARIABLES

222

Although there is no need to observe what the characteristics of the individuals with
yi = 0 are, nor to know how many individuals are ‘missing’, we have to assume that
they are unobserved only because their characteristics are such that yi∗ ≤ 0. Maximizing
log L2 with respect to β and σ 2 again gives consistent estimators. If observations with
yi = 0 are really missing it is the best one can do. However, even if observations with
yi = 0 are available, one could still maximize log L2 instead of log L1 , that is, one
could estimate a truncated regression model even if a tobit model would be applicable.
It is intuitively obvious that the latter (tobit) approach uses more information and
therefore will generally lead to more efﬁcient estimators. In fact, it can be shown that
the information contained in the tobit model combines that contained in the truncated
regression model with that of the probit model describing the zero/non-zero decision.
This fact follows easily from the result that the tobit loglikelihood function is the sum
of the truncated regression and probit loglikelihood functions.
7.4.3 Illustration: Expenditures on Alcohol and Tobacco (Part 1)

In economics, (systems of) demand equations are often used to analyse the effect of,
for example, income, tax or price changes on consumer demand. A practical problem
that emerges is that expenditures on particular commodities may be zero, particularly
if the goods are not aggregated into broad categories. While this typically occurs with
durable goods, we shall concentrate on a different type of commodity here: alcoholic
beverages and tobacco.
Starting from the assumption that a consumer maximizes his utility as a function of
the quantities of the goods consumed, it is possible to derive (Marshallian) demand
functions for each good as
qj = gj (x, p),
where qj denotes the quantity of good j, x denotes total expenditures and p is a
vector of prices of all relevant goods. The function gj depends upon the consumer’s
preferences. In the empirical application we shall consider cross-sectional data where
prices do not vary across observations. Therefore, p can be absorbed into the functional
form to get
qj = gj∗ (x).
This relationship is commonly referred to as an Engel curve (see, e.g. Deaton and
Muellbauer, 1980, Chapter 1). From this, one can deﬁne the total expenditure elasticity
of qj , the quantity of good j that is consumed, as
j =

∂gj∗ (x) x
.
∂x qj

This elasticity measures the relative effect of a 1% increase in total expenditures and
can be used to classify goods into luxuries, necessities and inferior goods. A good
is referred to as a luxury good if the quantity that is consumed increases more than
proportionally with total expenditures (j > 1), while it is a necessity if j < 1. If the
quantity of a good’s purchase decreases when total expenditure increases, the good is
said to be inferior, which implies that the elasticity j is negative.

TOBIT MODELS

223

A convenient parametrization of the Engel curve is
wj = αj + βj log x,
where wj = pj qj /x denotes the budget share of good j . It is a simple exercise to
derive that the total expenditure elasticities for this functional form are given by
j = 1 + βj /wj .

(7.72)

Recall that good j is a necessity if j < 1 or βj < 0, while a luxury good corresponds
to βj > 0.
Below we shall focus on two particular goods, alcoholic beverages and tobacco.
Moreover, we explicitly focus on heterogeneity across households and the sufﬁx i will
be used to index observations on individual households. The Almost Ideal Demand
System of Deaton and Muellbauer (1980, Section 3.4) implies Engel curves of the form
wj i = αj i + βj i log xi + εj i ,
where wj i is household i’s budget share of commodity j , and xi denotes total expenditures. The parameters αj i and βj i may depend upon household characteristics, like
family composition, age and education of the household head. The random terms εj i
capture unobservable differences between households. Because βj i varies over households, the functional form of the above Engel curve permits goods to be luxuries or
necessities depending upon household characteristics.
When we consider expenditures on alcohol or tobacco, the number of zeroes is
expected to be substantial. A ﬁrst way to explain these zeroes is that they arise from
corner solutions when the non-negativity constraint of the budget share (wj i ≥ 0)
becomes binding. This means that households prefer not to buy alcoholic beverages or
tobacco at current prices and income, but that a price decrease or income increase would
(ultimately) change this. The discussion whether or not this is a realistic assumption
is deferred to Subsection 7.5.4. As the corner solutions do not satisfy the ﬁrst order
conditions for an interior optimum of the underlying utility maximization problem, the
Engel curve does not apply to observations with wj i = 0. Instead, the Engel curve is
assumed to describe the solution to the household’s utility maximization problem if
the non-negativity constraint is not imposed, a negative solution corresponding with
zero expenditures on the particular good. This way, we can adjust the model to read
wj∗i = αj i + βj i log xi + εj i ,
wj i = wj∗i
=0

if wj∗i > 0
otherwise,

which corresponds to a standard tobit model if it is assumed that εj i ∼ NID(0, σ 2 )
for a given good j . Atkinson, Gomulka and Stern (1990) use a similar approach to
estimate an Engel curve for alcohol, but assume that εj i has a non-normal skewed
distribution.

MODELS WITH LIMITED DEPENDENT VARIABLES

224

To estimate the above model, we employ data23 from the Belgian household budget survey of 1995–1996, supplied by the National Institute of Statistics (NIS). The
sample contains 2724 households for which expenditures on a broad range of goods
are observed as well as a number of background variables, relating to, e.g. family
composition and occupational status. In this sample, 62% of the households has zero
expenditures on tobacco, while 17% does not spend anything on alcoholic beverages.
The average budget shares, for the respective subsamples of positive expenditures, are
3.22% and 2.15%.
Below we shall estimate the two Engel curves for alcohol and tobacco separately.
This means that we do not take into account the possibility that a binding non-negativity
constraint on tobacco may also affect expenditures on alcohol, or vice versa. We shall
assume that αj i is a linear function of the age of the household head,24 the number of
adults in the household, and the numbers of children younger than 2 and 2 or older,
while βj i is taken to be a linear function of age and the number of adults. This implies
that the products of log total expenditures with age and number of adults are included
as explanatory variables in the tobit model. The estimation results for the standard tobit
models are presented in Table 7.7.
For tobacco, there is substantial evidence that age is an important factor in explaining
the budget share, both separately and in combination with total expenditures. For
alcoholic beverages only the number of children and total expenditures are individually
signiﬁcant. As reported in the table, Wald tests for the hypothesis that all coefﬁcients,
except the intercept term, are equal to zero, produce highly signiﬁcant values for
both goods. Under the null hypothesis, these test statistics, comparable with the F statistic that is typically computed for the linear model (see Subsection 2.5.4), have an
asymptotic Chi-squared distribution with 7 degrees of freedom.
If we assume that a household under consideration has a sufﬁciently large budget
share to ignore changes in the second term of (7.62), the total expenditure elasticity can
be computed on the basis of (7.72) as 1 + βj i /wj i . It measures the total elasticity for
Table 7.7 Tobit models for budget shares alcohol and tobacco
Alcoholic beverages

23
24

Tobacco

Variable

Estimate

s.e.

Estimate

s.e.

constant
age class
nadults
nkids ≥ 2
nkids < 2
log(x)
age × log(x)
nadults × log(x)
σ̂

−0.1592
0.0135
0.0292
−0.0026
−0.0039
0.0127
−0.0008
−0.0022
0.0244

(0.0438)
(0.0109)
(0.0169)
(0.0006)
(0.0024)
(0.0032)
(0.0088)
(0.0012)
(0.0004)

0.5900
−0.1259
0.0154
0.0043
−0.0100
−0.0444
0.0088
−0.0006
0.0480

(0.0934)
(0.0242)
(0.0380)
(0.0013)
(0.0055)
(0.0069)
(0.0018)
(0.0028)
(0.0012)

Loglikelihood
Wald test (χ72 )

117.86

4755.371
(p = 0.000)

170.18

758.701
(p = 0.000)

I am grateful to the NIS for permission to use these data; available as TOBACCO.
Age is measured in 10-year interval classes ranging from 0 (younger than 30) to 4 (60 or older).

TOBIT MODELS

225

those that consume alcohol and those that smoke, respectively. If we evaluate the above
elasticities at the sample averages of those households that have positive expenditures,
we obtain estimated elasticities25 of 1.294 and 0.180, respectively. This indicates that
alcoholic beverages are a luxury good, while tobacco is a necessity. In fact, the total
expenditure elasticity of tobacco expenditures is fairly close to zero.
In this application the tobit model assumes that all zero expenditures are the result of
corner solutions and that a sufﬁciently large change in income or relative prices would
ultimately create positive expenditures for any household. In particular for tobacco
this seems not really appropriate. Many people do not smoke because of, e.g. health
or social reasons and would not smoke even if cigarettes were free. If this is the
case, it seems more appropriate to model the decision to smoke or not as a process
separate from the decision of how much to spend on it. The so-called tobit II model,
one of the extensions of the standard tobit model that will be discussed below, could
be appropriate for this situation. Therefore, we shall come back to this example in
Subsection 7.5.4.
7.4.4 Speciﬁcation Tests in the Tobit Model

A violation of the distributional assumptions on εi will generally lead to inconsistent
maximum likelihood estimators for β and σ 2 . In particular non-normality and heteroskedasticity are a concern. We can test for these alternatives, as well as for omitted
variables, within the Lagrange multiplier framework. To start the discussion, ﬁrst note
that the ﬁrst order conditions of the loglikelihood log L1 with respect to β are given by
 −φ(x  β̂/σ̂ )
i
i∈I0

1 − (xi β̂/σ̂ )

xi +

 ε̂
i∈I1

i

σ̂

xi =

N


ε̂iG xi = 0,

(7.73)

i=1

where we deﬁne the generalized residual ε̂iG as the scaled residual ε̂i /σ̂ = (yi − xi β̂)/σ̂
for the positive observations and as −φ(.)/(1 − (.)), evaluated at xi β̂/σ̂ , for the zero
observations. Thus we obtain ﬁrst order conditions that are of the same form as in the
probit model or the linear regression model. The only difference is the deﬁnition of
the appropriate (generalized) residual.
Because σ 2 is also a parameter that is estimated, we also need the ﬁrst order condition
with respect to σ 2 to derive the speciﬁcation tests. Apart from an irrelevant scaling
factor, this is given by
 x  β̂
i

i∈I0

φ(xi β̂/σ̂ )

σ̂ 1 − (xi β̂/σ̂ )

+

  ε̂2
i∈I1

i
σ̂ 2

 
N
−1 =
ε̂iG(2) = 0,

(7.74)

i=1

where we deﬁned ε̂iG(2) , a second order generalized residual. The ﬁrst order condition
with respect to σ 2 says that the sample average of ε̂iG(2) should be zero. It can be
shown (see Gouriéroux et al., 1987) that the second order generalized residual is an
estimate for E{εi2 /σ 2 − 1|yi , xi }, just like the (ﬁrst order) generalized residual ε̂iG is
an estimate for E{εi /σ |yi , xi }. While it is beyond the scope of this text to derive this,
25

We ﬁrst take averages and then compute the ratio.

MODELS WITH LIMITED DEPENDENT VARIABLES

226

it is intuitively reasonable: if εi cannot be determined from yi , xi and β, we replace
the expressions by the conditional expected values given all we know about yi∗ , as
reﬂected in yi . This is simply the best guess of what we think the residual should be,
given that we only know that it satisﬁes εi < −xi β.
From (7.73) it is immediately clear how we would test for J omitted variables zi .
As the additional ﬁrst order conditions would imply that
N


ε̂iG zi = 0,

i=1

we can simply do a regression of ones upon the K + 1 + J variables ε̂iG xi , ε̂iG(2) ,
and ε̂iG zi and compute the test statistic as N times the uncentred R 2 . The appropriate
asymptotic distribution under the null hypothesis is a Chi-squared with J degrees
of freedom.
A test for heteroskedasticity can be based upon the alternative that
V {εi } = σ 2 h(zi α),

(7.75)

where h(.) is an unknown differentiable function with h(0) = 1 and h(.) > 0, and zi
is a J -dimensional vector of explanatory variables, not including an intercept term.
The null hypothesis corresponds to α = 0, implying that V {εi } = σ 2 . The additional
scores with respect to α, evaluated under the current set of parameters estimates β̂, σ̂ 2
are easily obtained as κ ε̂iG(2) zi , where κ is an irrelevant constant that depends upon h.
Consequently, the LM test statistic for heteroskedasticity is easily obtained as N times
the uncentred R 2 of a regression of ones upon the K + 1 + J variables ε̂iG xi , ε̂iG(2) ,
and ε̂iG(2) zi . Note that also in this case the test statistic does not depend upon the form
of h, only upon zi .
If homoskedasticity is rejected, we can estimate the model with heteroskedastic
errors if we specify a functional form for h, for example, h(zi α) = exp{zi α}. In the
loglikelihood function, we simply replace σ 2 by σ 2 exp{zi α} and we estimate α jointly
with the parameters β and σ 2 . Alternatively, it is possible that heteroskedasticity is
found because something else is wrong with the model. For example, the functional
form may not be appropriate and nonlinear functions of xi should be included. Also a
transformation of the dependent variable could eliminate the heteroskedasticity problem. This explains, for example, why in many cases people specify a model for log
wages rather than wages themselves.
Finally, we discuss a test for non-normality. This test can be based upon the framework of Pagan and Vella (1989) and implies a test of the following two conditional
moment conditions that are implied by normality: E{εi3 /σ 3 |xi } = 0 and E{εi4 /σ 4 −
3|xi } = 0, corresponding to the absence of skewness and excess kurtosis, respectively
(see Section 6.4). Let us ﬁrst consider the quantities E{εi3 /σ 3 |yi , xi } and E{εi4 /σ 4 −
3|yi , xi }, noting that taking expectations over yi (given xi ) produces the two moments
of interest. If yi > 0 we can simply estimate the sample equivalents as ε̂i3 /σ̂ 3 and
ε̂i4 /σ̂ 4 − 3, respectively, where ε̂i = yi − xi β̂. For yi = 0 the conditional expectations
are more complicated, but they can be computed using the following formulae (Lee

EXTENSIONS OF TOBIT MODELS

227

and Maddala, 1985):


  2
ε 

εi3 
xi β
i
x
E
x
,
y
=
0
=
2
+
,
y
=
0
(7.76)
E

i
i
i
i

σ3
σ
σ


 4

 2
   3  



εi
 x , y = 0 = 3E εi − 1 x , y = 0 + xi β E εi  x , y = 0 .
E
−
3
i
i
i
i
i
i


σ4
σ2
σ
σ
(7.77)


These two quantities can easily be estimated from the ML estimates β̂ and σ̂ 2 and the
generalized residuals ε̂iG and ε̂iG(2) . Let us denote the resulting estimates as ε̂iG(3) and
ε̂iG(4) , respectively, such that
ε̂iG(3) = ε̂i3 /σ̂ 3
= [2 +

(xi β̂/σ̂ )2 ]ε̂iG

if yi > 0
otherwise,

(7.78)

and
ε̂iG(4) = ε̂i4 /σ̂ 4 − 3
= 3ε̂iG(2) + (xi β̂/σ̂ )3 ε̂iG

if yi > 0
otherwise.

(7.79)

By the law of iterated expectations the null hypothesis of normality implies that (asymptotically) E{ε̂iG(3) |xi } = 0 and E{ε̂iG(4) |xi } = 0. Consequently, the conditional moment
test for non-normality can be obtained by running a regression of a vector of ones upon
the K + 3 variables ε̂iG xi , ε̂iG(2) , ε̂iG(3) and ε̂iG(4) and computing N times the uncentred
R 2 . Under the null hypothesis, the asymptotic distribution of the resulting test statistic
is Chi-squared with 2 degrees of freedom.
Although the derivation of the different test statistics may seem complicated, their
computation is relatively easy. They can be computed using an auxiliary regression
after some straightforward computations involving the maximum likelihood estimates
and the data. As consistency of the ML estimators crucially depends upon a correct
speciﬁcation of the likelihood function, testing for misspeciﬁcation should be a standard
routine in empirical work.

7.5

Extensions of Tobit Models

The standard tobit model imposes a structure which is often too restrictive: exactly
the same variables affecting the probability of a nonzero observation determine the
level of a positive observation and, moreover, with the same sign. This implies, for
example, that those who are more likely to spend a positive amount are, on average,
also those that spend more on a durable good. In this section, we shall discuss models
that relax this restriction. Taking the speciﬁc example of holiday expenditures, it is
conceivable that households with many children are less likely to have positive expenditures, while if a holiday is taken up, the expected level of expenditures for such
households is higher.

MODELS WITH LIMITED DEPENDENT VARIABLES

228

Suppose that we are interested in explaining wages. Obviously, wages are only
observed for people that are actually working, but for economic purposes we are often
interested in (potential) wages not conditional upon this selection. For example: a
change in some x variable may lower someone’s wage such that he decides to stop
working. Consequently, his wage would no longer be observed and the effect of this
x variable could be underestimated from the available data. Because the sample of
workers may not be a random sample of the population (of potential workers) – in
particular one can expect that people with lower (potential) wages are more likely to
be unemployed – this problem is often referred to as a sample selection problem.

7.5.1 The Tobit II Model

The traditional model to describe sample selection problems is the tobit II model,26
also referred to as the sample selection model. In this context, it consists of a linear
wage equation

β1 + ε1i ,
(7.80)
wi∗ = x1i
where x1i denotes a vector of exogenous characteristics (age, education, gender, . . .)
and wi∗ denotes person i’s wage. The wage wi∗ is not observed for people that are
not working (which explains the ∗ ). To describe whether a person is working or not a
second equation is speciﬁed, which is of the binary choice type. That is,

h∗i = x2i
β2 + ε2i ,

(7.81)

where we have the following observation rule:
wi = wi∗ , hi = 1 if h∗i > 0
wi not observed, hi = 0 if h∗i ≤ 0,

(7.82)
(7.83)

where wi denotes person i’s actual wage.27 The binary variable hi simply indicates
working or not working. The model is completed by a distributional assumption on
the unobserved errors (ε1i , ε2i ), usually a bivariate normal distribution with expectations zero, variances σ12 , σ22 , respectively, and a covariance σ12 . The model in (7.81)
is, in fact, a standard probit model, describing the choice working or not working.
Therefore, a normalization restriction is required and, as before, one usually sets
σ22 = 1. The choice to work is affected by the variables in x2i with coefﬁcients β2 . The
equation (7.80) describes (potential) wages as a function of the variables in x1i with
coefﬁcients β1 . The signs and magnitude of the β coefﬁcients may differ across the two
equations. In principle, the variables in x1 and x2 can be different, although one has


to be very careful in this respect (see below). If we would impose that x1i
β1 = x2i
β2
and ε1i = ε2i , it is easily seen that we are back at the standard tobit model (tobit I).
26

This classiﬁcation of tobit models is due to Amemiya (1984). The standard tobit model of Section 7.4 is
then referred to as tobit I.
27
In most applications the model is formulated in terms of log(wages).

EXTENSIONS OF TOBIT MODELS

229

The conditional expected wage, given that a person is working, is given by

E{wi |hi = 1} = x1i
β1 + E{ε1i |hi = 1}


β1 + E{ε1i |ε2i > −x2i
β2 }
= x1i
σ


β1 + 122 E{ε2i |ε2i > −x2i
β2 }
= x1i
σ2

β1 + σ12
= x1i


φ(x2i
β2 )
,

(x2i β2 )

(7.84)

where the last equality uses σ22 = 1 and the expression for the expectation of a truncated standard normal distribution, similar to that used in (7.63). The third equality
uses that for two normal random variables E{ε1 |ε2 } = (σ12 /σ22 )ε2 . Appendix B provides more details on these results. Note that we can write σ12 = ρ12 σ1 , where ρ12
is the correlation coefﬁcient between the two errors. Again, this shows the generality of the model in comparison with (7.63). It follows directly from (7.84) that

the conditional expected wage equals x1i
β1 only if σ12 = ρ12 = 0. So, if the error
terms from the two equations are uncorrelated, the wage equation can be estimated
consistently by ordinary least squares. A sample selection bias in the OLS estimator


arises if σ12 = 0. The term φ(x2i
β2 )/(x2i
β2 ) is known as the inverse Mill’s ratio.

Because it is denoted λ(x2i β2 ) by Heckman (1979) it is also referred to as Heckman’s lambda.
The crucial parameter which makes the sample selection model different from just
a regression model and a probit model is the correlation coefﬁcient (or covariance)
between the two equations’ error terms. If the errors are uncorrelated we could simply
estimate the wage equation by OLS and ignore the selection equation (unless we are
interested in it). Now, why can we expect correlation between the two error terms?
Although the tobit II model can be motivated in different ways, we shall more or less
follow Gronau (1974) in his reasoning. Assume that the utility maximization problem
of the individual (in Gronau’s case: housewives) can be characterized by a reservation wage wir (the value of time). An individual will work if the actual wage she is
offered exceeds this reservation wage. The reservation wage of course depends upon
personal characteristics, via the utility function and the budget constraint, so that we
write (assume)
wir = zi γ + ηi ,
where zi is a vector of characteristics and ηi is unobserved. Usually the reservation
wage is not observed.
Now assume that the wage a person is offered depends on her personal characteristics
(and some job characteristics) as in (7.80), i.e.

wi∗ = x1i
β1 + ε1i .

If this wage is below wir individual i is assumed not to work. We can thus write her
labour supply decision as
hi = 1 if wi∗ − wir > 0
= 0 if wi∗ − wir ≤ 0

MODELS WITH LIMITED DEPENDENT VARIABLES

230

The inequality can be written in terms of observed characteristics and unobserved
errors as


β1 − zi γ + (ε1i − ηi ) = x2i
β2 + ε2i ,
(7.85)
h∗i ≡ wi∗ − wir = x1i
by appropriately deﬁning x2i and ε2i . Consequently, our simple economic model where
labour supply is based on a reservation wage leads to a model of the tobit II form.
A few things are worth noticing from (7.85). First, the offered wage inﬂuences the
decision to work or not. This implies that the error term ε2i involves the unobserved
heterogeneity inﬂuencing the wage offer, i.e. involves ε1i . If ηi is uncorrelated with ε1i ,
the correlation between ε2i and ε1i is expected to be positive. Consequently, we can
expect a sample selection bias in the least squares estimator from economic arguments.
Second, the variables in x1i are all included in x2i , plus all variables in zi that are not
contained in x1i . Economic arguments thus indicate that we should include in x2i at
least those variables which are contained in x1i .
Let us repeat the statistical model, the tobit II model, for convenience, substituting
y for w to stress generality.

β1 + ε1i
(7.86)
yi∗ = x1i

β2 + ε2i
h∗i = x2i

(7.87)

yi = yi∗ , hi = 1 if h∗i > 0
yi not observed, hi = 0 if
where



ε1i
ε2i


∼ NID

   2
0
σ1
,
0
σ12

h∗i

(7.88)

≤ 0,

σ12
1

(7.89)


.

(7.90)

This model has two observed endogenous variables yi and hi . Statistically, it describes
the joint distribution of yi and hi conditional upon the variables in both x1i and x2i .
That is, (7.86) should describe the conditional distribution of yi∗ conditional upon both
x1i and x2i . The only reason not to include a certain variable in x1i which is included
in x2i is that we are conﬁdent that it has a zero coefﬁcient in the wage equation.
For example, there could be variables which affect reservation wages only but not the
wage itself. Incorrectly omitting a variable from (7.86), while including it in (7.87),
may seriously affect the estimation results and may lead to spurious conclusions of the
existence of sample selection bias.
7.5.2 Estimation

For estimation purposes, the model can be thought of as consisting of two parts.
The ﬁrst part describes the binary choice problem. The contribution to the likelihood
function is simply the probability of observing hi = 1 or hi = 0. The second part
describes the distribution of the wage for those actually working, so that the likelihood
contribution is f (yi |hi = 1). We thus have for the loglikelihood function

log P {hi = 0}
log L3 (β, σ12 , σ12 ) =
i∈I0

+


i∈I1

[log f (yi |hi = 1) + log P {hi = 1}].

(7.91)

EXTENSIONS OF TOBIT MODELS

231

The binary choice part is standard; the only complicated part is the conditional distribution of yi given hi = 1. Therefore, it is more common to decompose the joint
distribution of yi and hi differently, by using that
f (yi |hi = 1)P {hi = 1} = P {hi = 1|yi }f (yi ).

(7.92)

The last term on the right-hand side is simply the normal density function, while the
ﬁrst term is a probability from a conditional normal density function, characterized by
(see Appendix B)
σ


E{h∗i |yi } = x2i
β2 + 122 (yi − x1i
β1 )
σ1
2
V {h∗i |yi } = 1 − σ12
/σ12

where the latter equality denotes the variance of h∗i conditional upon yi and given the
exogenous variables. We thus write the loglikelihood as
log L3 (β, σ12 , σ12 ) =



log P {hi = 0} +

i∈I0



[log f (yi ) + log P {hi = 1|yi }] (7.93)

i∈I1

with the following equalities

P {hi = 0} = 1 − (x2i
β2 )






 x β + (σ /σ )(y − x1i β1 ) 
P {hi = 1|yi } =   2i 2  12 1 i

2
1 − σ12
/σ12


1
1

f (yi ) = 
exp − (yi − x1i
β1 )2 /σ12 .
2
2πσ12
2

(7.94)
(7.95)

(7.96)

Maximization of log L3 (β, σ12 , σ12 ) with respect to the unknown parameters leads
(under mild regularity conditions) to consistent and asymptotically efﬁcient estimators,
that have an asymptotic normal distribution.
In empirical work, the sample selection model is often estimated in a two-step way.
This is computationally simpler and it will also provide good starting values for the
maximum likelihood procedure. The two-step procedure is due to Heckman (1979) and
is based on the following regression (compare (7.84) above)

yi = x1i
β1 + σ12 λi + ηi ,

where
λi =

(7.97)


β2 )
φ(x2i
.

(x2i β2 )

The error term in this model equals ηi = ε1i − E{ε1i |xi , hi = 1}. Given the assumption
that the distribution of ε1i is independent of xi (but not of hi ), ηi is uncorrelated with

232

MODELS WITH LIMITED DEPENDENT VARIABLES

x1i and λi by construction. This means that we could estimate β1 and σ12 by running
a least squares regression of yi upon the original regressors x1i and the additional
variable λi . The fact that λi is not observed is not a real problem because the only
unknown element in λi is β2 , which can be estimated consistently by probit maximum
likelihood applied to the selection model. This means that in the regression (7.97) we
replace λi by its estimate λ̂i and OLS will still produce consistent estimators of β1 and
σ12 . In general, this two-step estimator will not be efﬁcient, but it is computationally
simple and consistent.
One problem with the two-step estimator is that routinely computed OLS standard
errors are incorrect, unless σ12 = 0. This problem is often ignored because it is
still possible to validly test the null hypothesis of no sample selection bias using a
standard t-test on σ12 = 0. In general, however, standard errors will have to be adjusted
because ηi in (7.97) is heteroskedastic and because β2 is estimated. See Greene (2003,
Section 22.4) for details. If x1i and x2i are identical, the model is only identiﬁed through
the fact that λi is a nonlinear function. Empirically, the two-step approach will therefore
not work very well if there is little variation in λi and λi is close to being linear in
x2i . This is the subject of many Monte Carlo studies, a recent one being Leung and Yu
(1996). The inclusion of variables in x2i in addition to those in x1i can be important
for identiﬁcation in the second step, although often there are no natural candidates and
any choice is easily criticized. At the very least, some sensitivity analysis with respect
to the imposed exclusion restrictions should be performed, to make sure that the λ
term is not incorrectly picking up the effect of omitted variables.
The model that is estimated in the second step describes the conditional expected
value of yi given xi and given that hi = 1, for example the expected wage given that a
person is working. This is information that is not directly provided if you estimate the
model by maximum likelihood, although it can easily be computed from the estimates.
Often, the expected value of yi given xi , not conditional upon hi = 1, is the focus

of interest and this is given by x1i
β1 , which is also provided by the last regression.
Predicting wages for an arbitrary person can thus be based upon (7.97), but should

not include σ12 λ(x2i
β2 ). A positive covariance σ12 indicates that there is unobserved
heterogeneity that positively affects both wages and the probability of working. That
is, those with a wage that is higher than expected are more likely to be working
(conditional on a given set of xi values).
The two-step estimator of the sample selection model is one of the most often used
estimators in empirical micro-econometric work. There seems to be a strong belief that
the inclusion of a λ correction term in a model eliminates all problems of selection
bias. This is certainly not generally true. The presence of nonrandom selection induces
a fundamental identiﬁcation problem and, consequently, the validity of any solution
will depend upon the validity of the assumptions that are made, which can only be
partly tested. Section 7.6 below will pay more attention to sample selection bias and
the implied identiﬁcation problem.
7.5.3 Further Extensions

The structure of a model with one or more latent variables, normal errors and an
observation rule mapping the unobserved endogenous variables into observed ones,
can be used in a variety of applications. Amemiya (1984) characterizes several tobit

EXTENSIONS OF TOBIT MODELS

233

models by the form of the likelihood function, because different structures may lead
to models that are statistically the same. An obvious extension, resulting in the tobit
III model, is the one where h∗i in the above labour supply/wage equation model is
partially observed as hours of work. In that case we observe
yi = yi∗ , hi = h∗i
yi not observed, hi = 0

if h∗i > 0
if

h∗i

≤ 0,

(7.98)
(7.99)

with the same underlying latent structure. Essentially, this says that the selection model
is not of the probit type but of the standard tobit type. Applications using models of
this and more complicated structures can often be found in labour economics, where
one explains wages for different sectors, union/non-union members, etc. taking into
account that sectoral choice is probably not exogenous but based upon potential wages
in the two sectors, that labour supply is not exogenous, or both. Other types of selection
models are also possible, including, for example, an ordered response model. See Vella
(1998) for more discussion on this topic.
7.5.4 Illustration: Expenditures on Alcohol and Tobacco (Part 2)

In Subsection 7.4.3 we considered the estimation of Engel curves for alcoholic beverages and tobacco taking into account the problem of zero expenditures. The standard
tobit model assumes that these zero expenditures are the result of corner solutions. That
is, a household’s budget constraint and preferences are such that the optimal budget
shares of alcohol and tobacco, as determined by the ﬁrst order conditions and in the
absence of a non-negativity constraint, would be negative. As a consequence, the optimal allocation for the household is zero expenditures, which corresponds to a corner
solution that is not characterized by the usual ﬁrst order conditions. It can be disputed
that this is a realistic assumption and this subsection considers some alternatives to the
tobit I model. The alternatives are a simple OLS for the positive observations, possibly
combined with a binary choice model that explains whether expenditures are positive
or not, and a combined tobit II model that models budget shares jointly with the binary
decision to consume or not.
Obviously one can think of reasons other than those implicit in the tobit model why
households do not consume tobacco or alcohol. Because of social or health reasons, for
example, many non-smokers would not smoke even if tobacco were available for free.
This implies that whether or not we observe zero expenditures may be determined
quite differently from the amount of expenditures for those that consume the good.
Some commodities are possibly subject to abstention.28 Keeping this in mind, we can
consider alternative speciﬁcations to the tobit model. A ﬁrst alternative is very simple
and assumes that abstention is determined randomly in the sense that the unobservables
that determine budget shares are independent of the decision to consume or not. If this
is the case, we can simply specify an Engel curve that is valid for people that do not
abstain and ignore the abstention decision. This would allow us to estimate the total
expenditure elasticity for people that have a positive budget share, but would not allow
us to analyse possible effects arising through a changing composition of the population
28

Some authors refer to these goods as ‘bads’.

MODELS WITH LIMITED DEPENDENT VARIABLES

234

with positive values. Statistically, this means that we can estimate the Engel curve
simply by ordinary least squares but using only those observations that have positive
expenditures. The results of this exercise are reported in Table 7.8. In comparison with
the results for the tobit model, reported in Table 7.7, it is surprising that the coefﬁcient
for log total expenditures in the Engel curve for alcohol is negative and statistically not
signiﬁcantly different from zero. Estimating total expenditure elasticities, as deﬁned in
(7.72), on the basis of the OLS estimation results leads to values of 0.923 and 0.177,
for alcohol and tobacco, respectively.
The elasticities based on the OLS estimates are valid if abstention is determined
on the basis of the observables in the model but not on the basis of the unobservables that are collected in the error term. Moreover they are conditional upon the
fact that the household has positive expenditures. To obtain insight in what causes
households to consume these two goods or not we can use a binary choice model, the
most obvious choice being a probit model. If all zero expenditures are explained by
abstention rather than by corner solutions, the probit model should include variables
that determine preferences and should not include variables that determine the household’s budget constraint. This is because in this case a changing budget constraint
will never induce a household to start consuming alcohol or tobacco. This would
imply that total expenditures and relative prices should not be included in the probit
model. In the absence of price variation across households, total expenditures are an
obvious candidate for exclusion from the probit model. However, it is conceivable
that education level is an important determinant of abstention of alcohol or tobacco,
while – unfortunately – no information about education is available in our sample. This
is why we include total expenditures in the probit model, despite our reservations, but
think of total expenditures as a proxy for education level, social status or other variables that affect household preferences. In addition to variables included in the Engel
curve, the model for abstention also includes two dummy variables for blue and white
collar workers.29 It is assumed that these two variables do not affect the budget shares
of alcohol and tobacco but only the decision to consume or not. As any exclusion
Table 7.8 Models for budget shares alcohol and tobacco, estimated by OLS using positive
observations only
Alcoholic beverages

Tobacco

Variable

Estimate

s.e.

Estimate

s.e.

constant
age class
nadults
nkids ≥ 2
nkids < 2
log(x)
age × log(x)
nadults × log(x)

0.0527
0.0078
−0.0131
−0.0020
−0.0024
−0.0023
−0.0004
0.0008

(0.0439)
(0.0110)
(0.0163)
(0.0006)
(0.0023)
(0.0032)
(0.0008)
(0.0012)

0.4897
−0.0315
−0.0130
0.0013
−0.0034
−0.0336
0.0022
0.0011

(0.0741)
(0.0206)
(0.0324)
(0.0011)
(0.0045)
(0.0055)
(0.0015)
(0.0023)

R 2 = 0.051
s = 0.0215
N = 2258
29

R 2 = 0.154
s = 0.0291
N = 1036

The excluded category (reference group) includes inactive and self-employed people.

EXTENSIONS OF TOBIT MODELS

235

restriction, this one can also be disputed and we shall return to this issue below when
estimating a joint model for budget shares and abstention.
The estimation results for the two probit models are given in Table 7.6. For alcoholic
beverages it appears that total expenditures, the number of adults in the household as
well as the number of children older than two are statistically signiﬁcant in explaining
abstention. For tobacco, total expenditures, number of children older than two, age
and being a blue collar worker are statistically important explanators for abstention. To
illustrate the estimation results, consider a household consisting of two adults, the head
being a 35-year-old blue collar worker, and two children older than two. If the total
expenditures of this artiﬁcial household are equal to the overall sample average, the
implied estimated probabilities of a positive budget share of alcohol and tobacco are
given by 86.8% and 51.7% respectively. A 10% increase in total expenditures changes
these probabilities only marginally to 88.5% and 50.4%.
Assuming that the speciﬁcation of the Engel curve and the abstention model are
correct, the estimation results in Tables 7.8 and 7.9 are appropriate provided that the
error term in the probit model is independent of the error term in the Engel curve.
Correlation between these error terms invalidates the OLS results and would make
a tobit II model more appropriate. Put differently, the two equation model that was
estimated is a special case of a tobit II model in which the error terms in the respective
equations are uncorrelated. It is possible to test for a nonzero correlation if we estimate
the more general model. As discussed above, in the tobit II model it is very important
which variables are included in which of the two equations. If the same variables
are included in both equations, the model is only identiﬁed through the normality
assumption that was imposed upon the error terms.30 This is typically considered to be
an undesirable situation. The exclusion of variables from the abstention model does not
solve this problem. Instead, it is desirable to include variables in the abstention model
of which we are conﬁdent that they do not determine the budget shares directly. The
problem of ﬁnding such variables is similar to ﬁnding appropriate instruments with
Table 7.9

Probit models for abstention of alcohol and tobacco
Alcoholic beverages

Variable

30

Tobacco

Estimate

s.e.

Estimate

s.e.

constant
age
nadults
nkids ≥ 2
nkids < 2
log(x)
age × log(x)
nadults × log(x)
blue collar
white collar

−15.882
0.6679
2.2554
−0.0770
−0.1857
1.2355
−0.0448
−0.1688
−0.0612
0.0506

(2.574)
(0.6520)
(1.0250)
(0.0372)
(0.1408)
(0.1913)
(0.0485)
(0.0743)
(0.0978)
(0.0847)

8.244
−2.4830
0.4852
0.0813
−0.2117
−0.6321
0.1747
−0.0253
0.2064
0.0215

(2.211)
(0.5596)
(0.8717)
(0.0308)
(0.1230)
(0.1632)
(0.0413)
(0.0629)
(0.0834)
(0.0694)

Loglikelihood
Wald test (χ92 )

173.18

−1159.865
(p = 0.000)

108.91

−1754.886
(p = 0.000)

To see this, note that the functional form of λ is determined by the distributional assumptions of the error
term. See the discussion in Section 7.6 below.

MODELS WITH LIMITED DEPENDENT VARIABLES

236

endogenous regressors (see Chapter 5) and we should be equally critical and careful
in choosing them; our estimation results will critically depend upon the choice that
we make. In the above abstention model the dummies for being a blue or white collar
worker are included to take up this role. If we are conﬁdent that these variables do not
affect budget shares directly, estimation of the tobit II model may be appropriate.
Using the two-step estimation procedure, as proposed by Heckman (1979), we can
re-estimate the two Engel curves taking into account the sample selection problem due
to possible endogeneity of the abstention decision. The results of this are presented in
Table 7.10, where OLS is used but standard errors are adjusted to take into account
heteroskedasticity and the estimation error in λ. For alcoholic beverages the inclusion
of λ̂ does not affect the results very much and we obtain estimates that are pretty
close to those reported in Table 7.8. The t-statistic on the coefﬁcient for λ̂ does not
allow us to reject the null hypothesis of no correlation, while the estimation results
imply an estimated correlation coefﬁcient (computed as the ratio of the coefﬁcient for
λ̂ and the standard deviation of the error term σ̂1 ) of only −0.01. Computation of these
correlation coefﬁcients is important because the two-step approach may easily imply
correlations outside the [−1, 1] range, indicating that the tobit II model may not be
appropriate, or indicating that some exclusion restrictions are not appropriate. Note
that these estimation results imply that total expenditures have a signiﬁcant impact
on the probability of having positive expenditures on alcohol, but do not signiﬁcantly
affect the budget share of alcohol. For tobacco, on the other hand, we do ﬁnd a
signiﬁcant impact of the sample selection term λ, with an implied estimated correlation
coefﬁcient of −0.31. Qualitatively, however, the results do not appear to be very
different from those in Table 7.8. The negative correlation coefﬁcient indicates the
existence of unobservable characteristics that positively affect the decision to smoke
but negatively affect the budget share of tobacco. Let us, ﬁnally, compute the total
expenditure elasticities of alcohol and tobacco on the basis of the estimation results in
Table 7.10. Using similar computations as before, we obtain estimated elasticities of
0.920 and 0.243, respectively. Apparently and not surprisingly, tobacco is a necessary
good for those that smoke. In fact, tobacco expenditures are close to being inelastic.
Table 7.10 Two-step estimation of Engel curves for alcohol and tobacco
(tobit II model)
Alcoholic beverages

Tobacco

Variable

Estimate

s.e.

Estimate

s.e.

constant
age class
nadults
nkids ≥ 2
nkids < 2
log(x)
age × log(x)
nadults × log(x)
λ
σ̂1
Implied ρ

0.0543
0.0077
−0.0133
−0.0020
−0.0024
−0.0024
−0.0004
0.0008
−0.0002
0.0215
−0.01

(0.0487)
(0.0110)
(0.0166)
(0.0006)
(0.0023)
(0.0035)
(0.0008)
(0.0012)
(0.0028)
n.c.
n.c.

0.4516
−0.0173
−0.0174
0.0008
−0.0021
−0.0301
0.0012
−0.0041
−0.0090
0.0291
−0.31

(0.0735)
(0.0206)
(0.0318)
(0.0010)
(0.0045)
(0.0055)
(0.0015)
(0.0023)
(0.0026)
n.c.
n.c.

N = 2258

N = 1036

SAMPLE SELECTION BIAS

7.6

237

Sample Selection Bias

When the sample used in a statistical analysis is not randomly drawn from a larger
population, selection bias may arise. That is, standard estimators and tests may result
in misleading inferences. Because there are many situations where this may be the
case, and the tobit II model not necessarily provides an adequate solution to it, some
additional discussion of this problem is warranted.
At the general level, we can say that selection bias arises if the probability of a
particular observation to be included in the sample depends upon the phenomenon we
are explaining. There are a number of reasons why this may occur. First, it could be due
to the sampling frame. For example, if you would interview people in the university
restaurant and ask how often they visit it, those that go there every day are much
more likely to end up in the sample than those that visit every two weeks. Second,
nonresponse may result in selection bias. For example, people that refuse to report
their income are typically those with relatively high or relatively low income levels.
Third, it could be due to self-selection of economic agents. That is, individuals select
themselves into a certain state, e.g. working, union member, public sector employment,
in a nonrandom way on the basis of economic arguments. In general, those who beneﬁt
most from being in a certain state will be more likely to be in this state.
7.6.1 The Nature of the Selection Problem

Suppose we are interested in the conditional distribution of a variable yi given a set
of other (exogenous) variables xi , that is f (yi |xi ). Usually, we will formulate this as
a function of a limited number of parameters and interest lies in these parameters.
Selection is indicated by a dummy variable ri such that both yi and xi are observed if
ri = 1 and that either yi is unobserved if ri = 0, or both yi and xi are unobserved if
ri = 0.
All inferences ignoring the selection rule are (implicitly) conditional upon ri = 1.
Interest, however, lies in the conditional distribution of yi given xi but not given ri = 1.
We can thus say that the selection rule is ignorable (Rubin, 1976, Little and Rubin,
1987) if conditioning upon the outcome of the selection process has no effect. That
is, if
f (yi |xi , ri = 1) = f (yi |xi ).
(7.100)
If we are only interested in the conditional expectation of yi given xi , we can relax
this to
E{yi |xi , ri = 1} = E{yi |xi }.
(7.101)
A statement that is equivalent to (7.100) is that
P {ri = 1|xi , yi } = P {ri = 1|xi },

(7.102)

which says that the probability of selection into the sample should not depend upon
yi , given that it is allowed to depend upon the variables in xi . This already shows
some important results. First, selection bias does not arise if selection depends upon
the exogenous variables only. Thus, if we are estimating a wage equation that has
marital status on the right-hand side, it does not matter if married people are more

MODELS WITH LIMITED DEPENDENT VARIABLES

238

likely to end up in the sample than those who are not married. At a more general level,
it follows that whether or not selection bias is a problem depends upon the distribution
of interest.
If the selection rule is not ignorable, it should be taken into account when making
inferences. As stressed by Manski (1989), a fundamental identiﬁcation problem arises
in this case. To see this, note that
E{yi |xi } = E{yi |xi , ri = 1}P {ri = 1|xi } + E{yi |xi , ri = 0}P {ri = 0|xi }.

(7.103)

If xi is observed irrespective of ri , it is possible to identify the probability that ri = 1
as a function of xi (using a binary choice model, for example). Thus, it is possible to
identify P {ri = 1|xi } and P {ri = 0|xi }, while E{yi |xi , ri = 1} is also identiﬁed (from
the selected sample). However, since no information on E{yi |xi , ri = 0} is provided
by the data, it is not possible to identify E{yi |xi } without additional information or
making additional (non-testable) assumptions. As Manski (1989) notes, in the absence
of prior information, the selection problem is fatal for inference on E{yi |xi }.
If it is possible to restrict the range of possible values of E{yi |xi , ri = 0}, it is
possible to determine bounds on E{yi |xi } that may be useful. To illustrate this, suppose
we are interested in the unconditional distribution of yi (so no xi variables appear) and
we happen to know that this distribution is normal with unknown mean µ and unit
variance. If 10% is missing, the most extreme cases arise where these 10% are all in
the left or all in the right tail of the distribution. Using properties of a truncated normal
distribution,31 one can derive that
−1.75 ≤ E{yi |ri = 0} ≤ 1.75,
so that
0.9E{yi |ri = 1} − 0.175 ≤ E{yi } ≤ 0.9E{yi |ri = 1} + 0.175,
where E{yi |ri = 1} can be estimated by the sample average in the selected sample.
In this way, we can estimate an upper and lower bound for the unconditional mean
of yi , not making any assumptions about the selection rule. The price that we pay
for this is that we need to make assumptions about the form of the distribution of yi ,
which are not testable. If we shift interest in other aspects of the distribution of yi ,
given xi , rather than its mean, such assumptions may not be needed. For example,
if we are interested in the median of the distribution we can derive upper and lower
bounds from the probability of selection without assuming anything about the shape
of the distribution.32 Manski (1989, 1994) provides additional details and discussion
of these issues.
For a standard normal variable y it holds that P {y > 1.28} = 0.10 and E{y|y > 1.28} = φ(1.28)/0.10 =
1.75 (see Appendix B).
32
Recall that the median of a random variable y is deﬁned as the value m for which P {y ≤ m} = 0.5 (see
Appendix B). If 10% of the observations is missing we know that m is between the (theoretical) 40%
and 60% quantiles of the observed distribution. That is, m1 ≤ m ≤ m2 , with P {y ≤ m1 |r = 1} = 0.4 and
P {y ≤ m2 |r = 1} = 0.6.
31

SAMPLE SELECTION BIAS

239

A more common approach in applied work imposes additional structure on the
problem to identify the quantities of interest. Let
E{yi |xi } = g1 (xi )

(7.104)

E{yi |xi , ri = 1} = g1 (xi ) + g2 (xi ),

(7.105)

and
which, as long as we do not make any assumptions about the functions g1 and g2 , is
not restrictive. Assumptions about the form of g1 and g2 are required to identify g1 ,
which is what we are interested in. The most common assumption is the single index
assumption, which says that g2 depends upon xi only through a single index, xi β2 ,
say. This assumption is often motivated from a latent variable model:
yi = g1 (xi ) + ε1i

(7.106)

ri∗

(7.107)

=

xi β2

ri = 1

+ ε2i

if ri∗ > 0,

0 otherwise,

(7.108)

where E{ε1i |xi } = 0 and ε2i is independent of xi . Then it holds that
E{yi |xi , ri = 1} = g1 (xi ) + E{ε1i |ε2i > −xi β2 },

(7.109)

where the latter term depends upon xi only through the single index xi β2 . Thus we
can write
E{yi |xi , ri = 1} = g1 (xi ) + g2∗ (xi β2 ),
(7.110)
for some function g2∗ . Because β2 can be identiﬁed from the selection process, provided
observations on xi are available irrespective of ri , identiﬁcation of g1 is achieved by
assuming that it does not depend upon one or more variables in xi (while these variables
have a nonzero coefﬁcient in β2 ). This means that exclusion restrictions are imposed
upon g1 .
From (7.84), it is easily seen that the tobit II model constitutes a special case of
the above framework, where g1 (xi ) = xi β1 and g2∗ is given by σ12 φ(xi β2 )/(xi β2 ).
The assumption that ε1i and ε2i are i.i.d. jointly normal produces the functional form
of g2∗ . Moreover, the restriction that g1 is linear (while g2∗ is not) implies that the
model is identiﬁed even in the absence of exclusion restrictions in g1 (xi ). In practice
though, empirical identiﬁcation may beneﬁt from imposing zero restrictions on β1 .
When the distribution of ε1i and ε2i is not normal, (7.101) is still valid and this is what
is exploited in many semi-parametric estimators of the sample selection model.
7.6.2 Semi-parametric Estimation of the Sample Selection Model

Although it is beyond the scope of this text to fully discuss semi-parametric estimators
for limited dependent variable models, some intuitive discussion will be provided here.
While semi-parametric estimators relax the joint normality assumption of ε1i and ε2i
they generally maintain the single index assumption. That is, the conditional expectation

240

MODELS WITH LIMITED DEPENDENT VARIABLES

of ε1i given selection into the sample (and given the exogenous variables) depends upon
xi only through xi β2 . This requires that we can model the selection process in a fairly
homogeneous way. If observations are missing for a variety of reasons, the single index
assumption may no longer be appropriate. For example, individuals that do not have
a job, may not be working because their reservation wage is too high (a supply side
argument), as in the standard model, but also because employers are not interested in
hiring them (a demand side argument). These two processes are not necessarily well
described by a single index model.
The other crucial assumption in all semi-parametric approaches is that there is at least
one variable that enters the selection equation (xi β2 ) that does not enter the equation
of interest g1 (xi ). This means that we need an exclusion restriction in g1 in order to
identify the model. This is obvious as we would never be able to separate g1 from g2∗
if both depend upon the same set of variables and no functional form restrictions are
imposed. Because a constant in g1 cannot be distinguished from a constant in g2∗ , the
constant term in the model will not be identiﬁed, which is not a problem if we are
not interested in the intercept. If the intercept in g1 is of interest, it can be estimated
(Heckman, 1990, Andrews and Schafgans, 1998) from observations that are known to
have values of g2∗ close to zero (individuals that have high values for xi β2 ).
Most semi-parametric estimators are two-step estimators, just like Heckman’s (1979).
In the ﬁrst step, the single index parameter β2 is estimated semi-parametrically, that
is, without imposing a particular distribution upon ε2i . From this an estimate for the
single index is constructed, so that in the second step the unknown function g2∗ is estimated jointly with g1 (usually imposing some functional form upon g1 , like linearity).
A simple way to approximate the unknown function g2∗ (xi β2 ) is the use of a series
approximation, for example a polynomial in xi β2 . An alternative approach is based
on the elimination of g2∗ (xi β2 ) from the model by considering differences between
observations that have values of xi βˆ2 that are similar.
All semi-parametric methods involve some additional regularity conditions and
assumptions. An intuitive survey of alternative estimation methods for the sample
selection model is given in Vella (1998). Pagan and Ullah (1999) provide more
details. Empirical implementation is usually not straightforward; see Newey, Powell
and Walker (1990) or Martins (2001) for some applications.

7.7

Estimating Treatment Effects

Another area where sample selection plays an important role is in the estimation of
treatment effects. A treatment effect refers to the impact of receiving a certain treatment upon a particular outcome variable, for example the effect of participating in a
job training programme upon future earnings.33 Because this effect may be different
across individuals and selection into the training programme may be nonrandom, the
estimation of treatment effects has received much attention in the recent literature (see
e.g., Imbens and Angrist, 1994, Angrist, Imbens and Rubin, 1996, and Heckman, 1997,
2001). In the simplest case, the treatment effect is simply the coefﬁcient for a treatment dummy variable in a regression model. Because interest is in the causal effect
33

Below, we use the terms ‘participating in a programme’ and ‘receiving treatment’ as being equivalent.

ESTIMATING TREATMENT EFFECTS

241

of the treatment, we need to worry about the potential endogeneity of the treatment
dummy. Or, to be more precise, we need to worry about selection into treatment. In
this section, we consider the problem of estimating treatment effects in a more general
context, where the effect of treatment may differ across individuals and may affect the
probability of individuals to choose for treatment. A more extensive discussion can be
found in Wooldridge (2002, Chapter 18).
Let us start by deﬁning the two potential outcomes for an individual as y0i and
y1i , corresponding to the outcome without and with treatment, respectively. At this
stage, we think of y0i and y1i as having a continuous distribution (e.g. earnings).
The individual speciﬁc gains to treatment are given by y1i − y0i . There are several
important problems in estimating treatment effects. First, only one of the two potential
outcomes is observed depending upon the decision of the individual to participate in
the programme or not. In particular, if ri is a binary variable indicating treatment, we
only observe
yi ≡ (1 − ri )y0i + ri y1i .
(7.111)
Second, the gains to treatment are typically different across individuals and several
alternative population parameters are proposed to summarize the effect of treatment
for a particular group of individuals. A standard one is the average treatment effect,34
deﬁned as
(7.112)
ATE ≡ E{y1i − y0i }
or, conditional upon one or more covariates, E{y1i − y0i |xi }. The average treatment
effect describes the expected effect of treatment for an arbitrary person (with characteristics xi ). That is, it measures the effect of randomly assigning a person in the population
to the programme. While Heckman (1997) criticizes this parameter of interest by stating that ‘picking a millionaire at random to participate in a training programme for low
skilled workers’ is not policy relevant or feasible, it may be of interest if the population
of interest is appropriately deﬁned.
Also of interest is the average treatment effect for the treated, deﬁned as
TT ≡ E{y1i − y0i |ri = 1}

(7.113)

or, conditional upon one or more covariates, E{y1i − y0i |xi , ri = 1}. Thus, TT is the
mean effect for those that actually participate in the programme.
In special cases, ATE and TT are identical, but in general they could be different
because selection into the programme might be nonrandom and related to the expected
gains from the treatment. That is, one might expect that the average treatment effect
for those who choose to participate in the programme is somewhat larger than the
average treatment effect for the entire population. On the other hand, if people are
randomly selected into the programme (in a way deﬁned more precisely below), we
expect no difference between ATE and TT. A third parameter of interest is the local
average treatment effect deﬁned by Imbens and Angrist (1994). This fairly complicated
concept considers the effect of treatment upon persons at the margin of being treated
and we shall not discuss it here.
34

Because the expectation refers to the population of interest, it would be more appropriate to refer to this
quantity as the expected treatment effect. The current terminology follows the convention in the literature.

242

MODELS WITH LIMITED DEPENDENT VARIABLES

The econometric problem is to identify ATE or TT from observations on yi , ri and
xi . Note that it is easy to identify E{yi |ri = 1} = E{y1i |ri = 1} and E{yi |ri = 0} =
E{y0i |ri = 0}, but in general this is insufﬁcient to identify either ATE or TT. To illustrate the issues, let us assume that both y0i and y1i can be related to xi by means of a
linear model, that is
y0i = α0 + xi β0 + ε0i ,
(7.114)
y1i = α1 + xi β1 + ε1i ,

(7.115)

where the constant is eliminated from xi , and where ε0i and ε1i are zero mean error
terms satisfying E{εj i |xi } = 0 for j = 0, 1. The linearity assumption is not crucial.
With this, the observed outcome is given by
yi = α0 + xi β0 + ε0i + ri [(α1 − α0 ) + xi (β1 − β0 ) + (ε1i − ε0i )],

(7.116)

where the term in square brackets denotes the gain from the programme. This is
an example of a switching regression model where the outcome equation depends
upon the regime (ri = 0 or ri = 1). The individual speciﬁc gain from the programme
consists of three components: a constant, a component related to observable characteristics and an idiosyncratic component related to unobservables.35 We can rewrite
(7.116) as
yi = α0 + xi β + δri + ri xi γ + εi ,
(7.117)
where δ ≡ α1 − α0 , γ ≡ β1 − β0 and εi ≡ (1 − ri )ε0i + ri ε1i . In this model, the average treatment effect for individuals with characteristics xi is given by
ATE (xi ) = δ + xi γ ,

(7.118)

while the average treatment effect upon the treated is given by
TT (xi ) = δ + xi γ + E{ε1i − ε0i |xi , ri = 1}.

(7.119)

The two concepts are equivalent if the last term in this expression is zero, which
happens in two important special cases. The ﬁrst case arises when there are no unobservable components of the gain from treatment and we have that ε0i = ε1i . The second
case arises when the treatment decision is independent of the unobservable gains from
the treatment. In this case, E{ε1i − ε0i |xi , ri = 1} = E{ε1i − ε0i |xi } = 0. This implies
that individuals, at the time they make their participation decision, are unaware of
ε1i − ε0i (or simply ignore it).
Below, we shall focus on estimating the average treatment effect, ATE, noting that
in special cases it may be identical to the average treatment effect on the treated, TT.
To estimate ATE consistently, we thus need to ﬁnd consistent estimators for δ and γ .
35

While the unobservable components are not observed by the researcher, they may be (partially) known
to the individual.

ESTIMATING TREATMENT EFFECTS

243

First, note that we can estimate the parameters in (7.114) and (7.115) consistently by
applying OLS on the appropriate subsamples if
E{ε0i |xi , ri = 0} = 0
and
E{ε1i |xi , ri = 1} = 0,
respectively. If the slope coefﬁcients in (7.114) and (7.115) are identical (β0 = β1 = β),
the average treatment effect reduces to a constant and can also be estimated from
OLS in
yi = α0 + xi β + δri + εi ,
(7.120)
where δ denotes the average treatment effect, and εi = (1 − ri )ε0i + ri ε1i , as before.
This error term satisﬁes E{εi |xi , ri } = 0 by virtue of the previous two assumptions.
However, these assumptions are very restrictive and require that there are no unobservable components to y0i and y1i that also affect a person’s decision to participate
in the programme. That is, individuals may decide to participate in the programme
on the basis of xi (e.g. previous education or gender), but not on the basis of unobservables affecting either y0i or y1i . This is similar to condition (7.102) in the previous section.
To discuss the more general case, let us assume that the treatment decision can be
described by a probit equation
ri∗ = xi β2 + ηi ,

(7.121)

with ri = 1 if ri∗ > 0 and 0 otherwise, where ηi is assumed to be NID(0, 1), independent
of xi . Further, assume that the error terms in (7.114) and (7.115) are also normal, with
variances σ02 and σ12 and covariances σ02 and σ12 with ηi . Then, we can write (see Vella
and Verbeek, 1999b)
E{ε0i |xi , ri = 0} = σ02 E{ηi |xi , ηi ≤ −xi β2 } = σ02 λi (xi β2 )
E{ε1i |xi , ri = 1} = σ12 E{ηi |xi , ηi > −xi β2 } = σ12 λi (xi β2 ),
where
λi (xi β2 ) = E{ηi |xi , ri } =

ri − (xi β2 )
φ(xi β2 ),
(xi β2 )(1 − (xi β2 ))

(7.122)

which corresponds to the generalized residual of the probit model (see Subsection 7.1.4).
For ri = 1, it also corresponds to the deﬁnition of Heckman’s lambda given in Subsection 7.5.1. In the general case where σ02 or σ12 may be nonzero, these results indicate
that the parameters in (7.114) and (7.115) can be estimated consistently by a variant of
the two-step approach discussed for the sample selection model, including λi (xi β̂2 ) as
additional variable. The identiﬁcation strongly rests upon distributional assumptions,36
36

Heckman, Tobias and Vytlacil (2003) extend the above latent variable model to cases where the error
terms are not jointly normal.

MODELS WITH LIMITED DEPENDENT VARIABLES

244

and it is advisable to have exclusion restrictions in (7.114) and (7.115). That is, ideally
an instrumental variable can be found that affects the decision whether to participate
in the programme, but not the actual and counterfactual outcomes of yi . Under these
assumptions, the average treatment effect on the treated from (7.119) equals
TT (xi ) = δ + xi γ + (σ12 − σ02 )λi (xi β2 ).
If it is imposed that β0 = β1 = β, it follows that
E{yi |xi , ri } = α0 + xi β + δri + E{εi |xi , ri }
= α0 +

xi β

+ δri +

σ12 ri λi (xi β2 )

(7.123)
+ σ02 (1 −

ri )λi (xi β2 ),

which shows that we can consistently estimate α0 , β and δ from a single regression
provided that we include the generalized residual interacted with the treatment dummy.
If it can be assumed that σ02 = σ12 , in which case ATE (xi ) and TT (xi ) are identical, simpler alternative estimation techniques are available. For example, the two-step
approach reduces to the standard approach described in (7.97), provided that we extend
the deﬁnition of λi to the ri = 0 cases. This is the dummy endogenous variable model
of Heckman (1978b). Alternatively, the model parameters can also be estimated consistently by instrumental variables techniques, as discussed in Chapter 5, provided that
there is a valid exclusion restriction in (7.120). A sufﬁcient condition for σ02 = σ12
is that ε1i − ε0i = 0, which implies that there are no unobservable components to the
individual speciﬁc gains of the programme. Heckman (1997) and Vella and Verbeek
(1999), among others, stress the behavioural assumptions that are implicitly made in
an instrumental variables context. If responses to treatment vary across individuals, the
instrumental variables estimator is only consistent for ATE if individuals do not select
into the programme on the basis of the idiosyncratic component of their response to the
programme. Similar arguments can be made in cases where treatment is a multi-valued
or continuous variable, like schooling; see Angrist and Imbens (1995) or Card (1999)
for examples and discussion.

7.8

Duration Models

In some applications, we are interested in explaining the duration of a certain event.
For example, we may be interested in explaining the time it takes for an unemployed
person to ﬁnd a job, the time that elapses between two purchases of the same product, the duration of a strike or the duration of a ﬁrm’s bank relationships. The data
we have contain duration spells, that is, we observe the time elapsed until a certain
event (e.g. ﬁnding a job) occurs. Usually, duration data are censored in the sense that
the event of interest has not occurred for some individuals at the time the data are
analysed. Duration models have their origin in survival analysis, where the duration of
interest is the survival of a given subject, for example an insect. In economics, duration models are often used in labour market studies, where unemployment spells are
analysed. In this section, we will brieﬂy touch upon duration modelling. More details
can be found in Wooldridge (2002, Chapter 20), Jenkins (2003), or, more extensively,
in Lancaster (1990).

DURATION MODELS

245

7.8.1 Hazard Rates and Survival Functions

Let T denote the time spent in the initial state. For example, if the initial state is
unemployment, T may denote the number of weeks until a person becomes employed.
It is most convenient to treat T as a continuous variable. The distribution of T is
deﬁned by the cumulative density function
F (t) = P {T ≤ t},

(7.124)

which denotes the probability that the event has occurred by duration t. It is typically
assumed that F (t) is differentiable, so that the density function of T can be written
as f (t) = F  (t). Later on, we will allow the distribution of T to depend upon personal characteristics. The survivor function is the probability surviving past t and is
deﬁned as
S(t) ≡ 1 − F (t) = P {T > t}.
The conditional probability of leaving the initial state within the time interval t until
t + h, given survival up to time t, can be written as
P {t ≤ T < t + h|T ≥ t}.
If we divide this probability by h, we obtain the average probability of leaving per
unit time period over the interval t until t + h. Consideration of shorter and shorter
intervals results in the so-called hazard function that is formally deﬁned as
λ(t) = lim
h↓0

P {t ≤ T < t + h|T ≥ t}
.
h

(7.125)

At each time t, the hazard function is the instantaneous rate of leaving the initial state
per unit of time. The hazard function can be expressed as a function of the (cumulative)
density function of T in a straightforward way. First, write
P {t ≤ T < t + h|T ≥ t} =
Because
lim
h↓0

F (t + h) − F (t)
P {t ≤ T < t + h}
=
.
P {T ≥ t}
1 − F (t)

F (t + h) − F (t)
= F  (t) = f (t),
h

it follows directly that
λ(t) =

f (t)
f (t)
=
.
1 − F (t)
S(t)

(7.126)

The hazard and survival functions provide alternative but equivalent characterizations
of the distributions of T , while most duration models are based on making particular
assumptions about the hazard function.

MODELS WITH LIMITED DEPENDENT VARIABLES

246

There is a one-to-one relation between a speciﬁcation for the hazard function and
a speciﬁcation for the cumulative density function of T . To see this, ﬁrst note that
∂ log[1 − F (t)]/∂t = −F  (t)/[1 − F (t)], where F  (t) = f (t). So we can write
λ(t) = −

∂ log[1 − F (t)]
.
∂t

Now integrate both sides over the interval [0, s]. This gives
 s
λ(t)dt = − log[1 − F (s)] + log[1 − F (0)]
0

= − log[1 − F (s)],
because F (0) = 0. Consequently, it follows that

  s
λ(t)dt .
F (s) = 1 − exp −

(7.127)

0

The important result is that whatever functional form we choose for λ(t), we can derive
F (t) from it and vice versa. While most implementations start from a speciﬁcation of
the hazard function, the cumulative density function and survival function are important
for constructing the likelihood function of the model.
As a simple case, assume that the hazard rate is constant, that is, λ(t) = λ. This
implies that the probability of leaving during the next time interval does not depend
upon the duration spent in the initial state. A constant hazard implies
F (t) = 1 − exp(−λt),
corresponding to the exponential distribution. In most cases, researchers work with
a convenient speciﬁcation for the hazard function, for example one that leads to
closed-form expressions for the survival function S(t). Moreover, the hazard function is typically allowed to depend upon personal characteristics, xi , say. Let us, in
general, denote the hazard function of an individual i with characteristics xi as λ(t, xi ).
For the moment, we assume that these characteristics do not vary with survival or calendar time. A popular class of models are the so-called proportional hazards models,
in which the hazard function can be written as the product of a baseline hazard function that does not depend upon xi , and a person-speciﬁc non-negative function that
describes the effect of the characteristics xi . In particular,
λ(t, xi ) = λ0 (t) exp{xi β}.

(7.128)

In this model, λ0 (t) is a baseline hazard function that describes the risk of leaving the
initial state for (hypothetical) individuals with xi = 0, who serve as a reference group,
and exp{xi β} is an adjustment factor that depends upon the set of characteristics xi .
Note that the adjustment is the same at all durations t. To identify the baseline hazard, xi
should not include an intercept term. If xik is a continuous variable, we can derive that
∂ log λ(t, xi )
= βk .
∂xik

(7.129)

DURATION MODELS

247

Consequently, the coefﬁcient βk measures the proportional change in the hazard rate
that can be attributed to an absolute change in xik . Note that this effect does not depend
upon duration time t. If λ0 (t) is not constant, the model exhibits duration dependence.
There is positive duration dependence if the hazard rate increases with the duration.
In this case, the probability of leaving the initial state increases (ceteris paribus) the
longer one is in the initial state.
A wide range of possible functional forms can be chosen for the baseline hazard
λ0 (t). Some of them impose either positive or negative duration dependence at all
durations, while others allow the baseline hazard to increase for short durations and to
decrease for longer durations. A relatively simple speciﬁcation is the Weibull model,
which states that37
λ0 (t) = γ αt α−1 ,
where α > 0 and γ > 0 are unknown parameters. When α = 1 we obtain the exponential distribution with γ = λ. If α > 1, the hazard rate is monotonically increasing,
for α < 1 it is monotonically decreasing. The log-logistic hazard function is given by
λ0 (t) =

γ αt α−1
,
1 + γ tα

where, again, α > 0 and γ > 0 are unknown parameters. When α ≤ 1, the hazard rate is
monotonically decreasing to zero as t increases. If α > 1, the hazard is increasing until
t = [(α − 1)/γ ]1−α and then it decreases to zero. With a log-logistic hazard function,
it can be shown that the log duration, log(T ), has a logistic distribution. See Franses
and Paap (2001, Chapter 8) or Greene (2003, Section 22.5) for a graphical illustration
of these hazard functions.
7.8.2 Samples and Model Estimation

Before turning to estimation, it is important to consider the types of data that are used
for estimation. We assume that the population of interest consists of all individuals
that enter the initial state between time 0 and time t0 (e.g. a given calendar year),
where t0 is a known constant. Two sampling schemes are typically encountered in
duration analysis. With stock sampling, we randomly sample individuals that are in
the initial state at time t0 , while with ﬂow sampling we sample individuals who enter
the initial state between time 0 and t0 . In both cases, we record the length of time
each individual is in the initial state. Because after a certain amount of time we stop
following individuals in the sample (and start analysing our data), both types of data are
typically right-censored. That is, for those individuals that are still in the initial state
we only know that the duration lasted at least as long as the tracking period. With stock
sampling, the data may also be left-censored if some or all of the starting times in the
initial state are not observed. Moreover, stock sampling introduces a sample selection
problem. As we shall see below, the censoring and the sample selection problem can
be handled by appropriately adjusting the likelihood function.
Let us, ﬁrst of all, consider maximum likelihood estimation with right-censored ﬂow
data. Assume that we randomly sample individuals that become unemployed (enter the
37

Different authors may use different (but equivalent) normalizations and notations.

MODELS WITH LIMITED DEPENDENT VARIABLES

248

initial state) between time 0 and t0 . Let ai denote the time at which individual i becomes
unemployed, and let ti∗ denote the total unemployment duration. For some individuals,
ti∗ will not be observed because of right-censoring (the unemployment duration exceeds
the period over which we track the individuals). If ci denotes the censoring time for
individual i, we observe
ti = min{ti∗ , ci }.
That is, for some individuals we observe the exact unemployment duration, while for
others we only know it exceeds ci . The censoring time may vary across individuals
because censoring often takes place at a ﬁxed calendar date. If, for example, we
sample from individuals that become unemployed during 2001 and we stop tracking
those individuals by the end of 2002, the censoring time may vary between one and
two years depending upon the moment in 2001 the individual became unemployed.
The contribution to the likelihood function of individual i is given by the conditional
density of ti if the observation is not censored, or the conditional probability that
ti∗ > ci (i.e. ti = ci ) in case of censoring, in each case, conditional upon the observed
characteristics xi . We assume that the distribution of ti , given xi , does not depend
upon the starting time ai . This implies, for example, that unemployment durations
that start in summer have the same expected length as those that start in winter. If
there are seasonal effects, we may capture them by including calendar dummies in xi
corresponding to different values of ai (see Wooldridge, 2002, Chapter 20). Thus, the
likelihood contribution of individual i is given by
f (ti |xi ; θ ),
if the duration is uncensored, where θ denotes the vector of unknown parameters that
characterize the distribution. For right-censored observations, the likelihood contribution is
P {ti = ci |xi ; θ } = P {ti∗ > ci |xi ; θ } = 1 − F (ci |xi ; θ ).
Given a random sample of size N , the maximum likelihood estimator is obtained
by maximizing
log L1 (θ ) =

N


[di log f (ti |xi ; θ ) + (1 − di ) log[1 − F (ci |xi ; θ )]],

(7.130)

i=1

where di is a dummy variable indicating censoring (di = 1 if uncensored, di = 0 if
censored). The functional form of f and F depend upon the speciﬁcation of the
hazard function.
With stock sampling, the loglikelihood function is slightly more complicated because
of the sample selection problem. Suppose our population of interest consists of all
individuals that became unemployed during 2001, while we sample from all those that
are unemployed by the end of the year. In this case, anyone whose unemployment
spell ended before the end of 2001 will not be included in the sample. Because this
spell is necessarily less than one year, we cannot assume that this observation is
missing randomly. Kiefer (1988) refers to this sample selection problem as lengthbiased sampling. This sample selection problem is similar to the one in the truncated

DURATION MODELS

249

regression model that was discussed in Section 7.4, and we can correct for it in a
similar fashion. The likelihood contribution for individual i in the absence of censoring
is changed into
f (ti |xi ; θ )
f (ti |xi ; θ, ti ≥ t0 − ai ) =
.
1 − F (t0 − ai |xi ; θ )
With right-censoring, the likelihood contribution is the conditional probability that ti∗
exceeds ci given by
P {ti∗ > ci |xi ; θ, ti ≥ t0 − ai } =

1 − F (ci |xi ; θ )
.
1 − F (t0 − ai |xi ; θ )

From this, it follows directly that the loglikelihood function with stock sampling can
be written as
log L2 (θ ) = log L1 (θ ) −

N


log[1 − F (t0 − ai |xi ; θ )],

(7.131)

i=1

where the additional term takes account of the sample selection problem. Unlike in
the case of ﬂow sampling, both the starting dates ai and the length of the sampling
interval t0 appear in the loglikelihood. The exact functional form of the loglikelihood
function depends upon the assumptions that we are making about the distribution of
the duration variable. As mentioned above, these assumptions are typically stated by
specifying a functional form for the hazard function.
When the explanatory variables are time varying, things are a bit more complicated,
because it does not make sense to study the distribution of a duration conditional upon
the values of the explanatory variables at one point in time. Another extension is the
inclusion of unobserved heterogeneity in the model, because the explanatory variables
that are included in the model may be insufﬁcient to capture all heterogeneity across
individuals. In the proportional hazards model, this implies that the speciﬁcation for
the hazard rate is extended to
λ(t, xi , vi ) = vi λ0 (t) exp{xi β},

(7.132)

where vi is an unobservable positive random variable with E{vi } = 1. This expression
describes the hazard rate for individual i given his or her characteristics in xi and
given his unobserved heterogeneity vi . Because vi is unobserved, it is integrated
out of the likelihood function by assuming an appropriate parametric distribution.38
See Wooldridge (2002, Chapter 20) for more details on these extensions.
7.8.3 Illustration: Duration of Bank Relationships

In this subsection, we consider an example from ﬁnancial economics concerning the
duration of ﬁrm-bank relationships. A strong bank relationship is typically considered
38

This approach is similar to using a random effects speciﬁcation in panel data models with limited dependent variables; see Section 10.7 below.

MODELS WITH LIMITED DEPENDENT VARIABLES

250

valuable to a ﬁrm because it decreases the costs of loans and increases the availability
of credit. On the other hand, however, the bank’s acquisition of private information
during a relationship may have undesirable consequences. For example, banks may be
able to extract monopoly rents from the relationship. In a recent paper, Ongena and
Smith (2001) examine the duration of 383 ﬁrm-bank relationships and investigate the
presence of positive or negative duration dependence. Moreover, they relate relationship
durations to observed ﬁrm-speciﬁc characteristics, such as size and age. The sample is
based upon annual data on bank relationships of Norwegian ﬁrms, listed on the Oslo
Stock Exchange, for the years 1979 to 1995, which corresponds to ﬂow sampling as
described above. A bank relationship is ended when the ﬁrm drops a bank from its list
of primary bank relationships or replaces one bank by another. The average duration
in the sample is 4.1 years. The data are right-censored, because a number of durations
are not completed by 1995.
We consider a small subset of the results from Ongena and Smith (2001), corresponding to the proportional hazards model in (7.128), where the baseline hazard
function is of the Weibull type. As a special case, the exponential baseline hazard is
obtained by imposing α = 1. The ﬁrm-speciﬁc characteristics that are included are:
logarithm of year-end sales, time elapsed since the ﬁrm’s founding date (age at start),
proﬁtability, as measured by the ratio of operating income to book value of assets,
Tobin’s Q, leverage and a dummy for multiple bank relationships. Tobin’s Q, deﬁned
as the ratio of the value of equity and debt to the book value of assets, is typically
interpreted as an indicator for management quality and/or the presence of proﬁtable
investment opportunities. Leverage is the book value of debt divided by the sum of
market value of equity and book value of debt. Highly leveraged ﬁrms are expected
to be more dependent on banks.
The maximum likelihood estimation results for the two different models, both
adjusted for right-censoring, are presented in Table 7.11. The results are reasonably
similar for the exponential and Weibull baseline hazard. The estimated value for α
in the latter model is 1.351 and signiﬁcantly larger than unity. This indicates that
the Weibull model is preferred to the exponential one, which is conﬁrmed by the
difference in loglikelihood values. Moreover, it implies that bank relationships exhibit
positive duration dependence. That is, the probability of ending a bank relationship,
ceteris paribus, increases as the duration lengthens. The results for the ﬁrm-speciﬁc
Table 7.11

Estimation results proportional hazards model (Ongena and Smith, 2001)
Exponential (MLE)

constant
log (sales)
age at start
proﬁtability
Tobin’s Q
leverage
multiple relationships
α
Loglikelihood

Weibull (MLE)

Estimate

Standard error

Estimate

Standard error

−3.601
−0.218
−0.00352
2.124
0.268
2.281
0.659
1

0.561
0.053
0.00259
0.998
0.195
0.628
0.231
(ﬁxed)

−3.260
−0.178
−0.00344
1.752
0.238
1.933
0.491
1.351

0.408
0.038
0.00183
0.717
0.141
0.444
0.168
0.135

−259.1469

−253.5265

EXERCISES

251

variables indicate that proﬁtable ﬁrms end bank relationships earlier, consistent with
the idea that such ﬁrms are less dependent on bank ﬁnancing. In particular, ﬁrms with
10% higher sales are associated with an approximately 2% lower hazard rate. Further,
the probability of ending a bank relationship decreases in ﬁrm size and increases in
ﬁrm leverage and when ﬁrms maintain multiple bank relationships. Using (7.129),
the coefﬁcient estimate of the dummy for multiple relationships in the Weibull model
indicates that the hazard rate is about 100[exp(0.491) − 1] = 63.4% greater for ﬁrms
that have more than one bank relationship.

Exercises
Exercise 7.1 (Binary Choice Models)
For a sample of 600 married females, we are interested in explaining participation
in market employment from exogenous characteristics in xi (age, family composition,
education). Let yi = 1 if person i has a paid job and 0 otherwise. Suppose we estimate
a linear regression model
yi = xi β + εi

by ordinary least squares.
a. Give two reasons why this is not really an appropriate model.
As an alternative, we could model the participation decision by a probit model.
b.
c.
d.
e.

Explain the probit model.
Give an expression for the loglikelihood function of the probit model.
How would you interpret a positive β coefﬁcient for education in the probit model?
Suppose you have a person with xi β = 2. What is your prediction for her labour
market status yi ? Why?
f. To what extent is a logit model different from a probit model?

Now assume that we have a sample of women that is not working (yi = 0), part-time
working (yi = 1) or full-time working (yi = 2).
g. Is it appropriate, in this case, to specify a linear model as yi = xi β + εi ?
h. What alternative model could be used instead, which exploits the information
contained in part-time versus full-time working?
i. How would you interpret a positive β coefﬁcient for education in this latter model?
j. Would it be appropriate to pool the two outcomes yi = 1 and yi = 2 and estimate
a binary choice model? Why or why not?
Exercise 7.2 (Probit and Tobit Models)
To predict the demand for its new investment fund, a bank is interested in the question
whether people invest part of their savings in risky assets. To this end, a tobit model
is formulated of the following form

yi∗ = β1 + β2 xi2 + β3 xi3 + εi ,

MODELS WITH LIMITED DEPENDENT VARIABLES

252

where xi2 denotes a person’s age, xi3 denotes income and the amount of savings
invested in risky assets is given by
yi = yi∗
=0

if yi∗ > 0
otherwise.

It is assumed that εi is NID(0, σ 2 ), independent of all explanatory variables.
Initially, the bank is only interested in the question whether a person is investing in
risky assets, which is indicated by a discrete variable di , that satisﬁes
di = 1 if yi∗ > 0
= 0 otherwise.
a.
b.
c.

d.
e.
f.

g.
h.
i.

j.

Derive the probability that di = 1 as a function of xi = (1, xi2 , xi3 ) , according to
the above model.
Show that the model that describes di is a probit model with coefﬁcients γ1 =
β1 /σ, γ2 = β2 /σ, γ3 = β3 /σ .
Write down the loglikelihood function log L(γ ) of the probit model for di . What
are, in general, the properties for the maximum likelihood estimator γ̂ for γ =
(γ1 , γ2 , γ3 ) ?
Give a general expression for the asymptotic covariance matrix of the ML estimator. Describe how it can be estimated in a given application.
Write down the ﬁrst order condition with respect to γ1 and use this to deﬁne the
generalized residual of the probit model.
Describe how the generalized residual can be used to test the hypothesis that
gender does not affect the probability of investing in risky assets. (Formulate
the hypothesis ﬁrst, describe how a test statistic can be computed and what the
appropriate distribution or critical values are.) To what class does this test belong?
Explain why it is not possible to identify σ 2 using information on di and xi only
(as in the probit model).
It is possible to estimate β = (β1 , β2 , β3 ) and σ 2 from the tobit model (using
information on yi ). Write down the loglikelihood function of this model.
Suppose we are interested in the hypothesis that age does not affect the amount
of risky savings. Formulate this hypothesis. Explain how this hypothesis can be
tested using a likelihood ratio test.
It is also possible to test the hypothesis from i on the basis of the results of the
probit model. Why would you prefer the test using the tobit results?

Exercise 7.3 (Tobit Models – Empirical)

Consider the data used in Subsections 7.4.3 and 7.5.4 to estimate Engel curves for
alcoholic beverages and tobacco. In a recent paper, Banks, Blundell and Lewbel (1997)
propose the Quadratic Almost Ideal Demand System, which implies quadratic Engel
curves of the form
wj i = αj i + βj i log xi + γj i log2 xi + εj i .

EXERCISES

253

This form has the nice property that it allows goods to be luxuries at low income levels,
while they can become necessities at higher levels of income (total expenditures). When
answering the following questions, use the data from TOBACCO.
a. Re-estimate the standard tobit model for alcohol from Subsection 7.4.3. Refer to
this as model A. Check that your results are the same as those in the text.
b. Extend model A by including the square of log total expenditures, and estimate it
by maximum likelihood.
c. Test whether the quadratic term is relevant using a Wald test and a likelihood
ratio test.
d. Compute the generalized residual for model A. Check that it has mean zero.
e. Compute the second order generalized residual for model A, as deﬁned in (7.74).
Check that is has mean zero too.
f. Perform a Lagrange multiplier test in model A for the hypothesis that the quadratic
term log2 x is irrelevant.
g. Perform an LM test for heteroskedasticity in model A related to age and the number
of adults.
h. Test for normality in model A.
Exercise 7.4 (Tobit Models)

A top university requires all students that apply to do an entry exam. Students that
obtain a score of less than 100 are not admitted. For students that score above 100,
the scores are registered, after which the university selects students from this group
for admittance. We have a sample of 500 potential students that did their entry exam
in 1996. For each student, we observe the result of the exam being:
– ‘rejected’, if the score is less than 100, or
– the score, if it is 100 or more.
In addition, we observe background characteristics of each candidate, including parents’
education, gender and the average grade at high school.
The dean is interested in the relationship between these background characteristics
and the score for the entry exam. He speciﬁes the following model
yi∗ = β0 + xi β1 + εi ,
yi = yi∗
= ‘rejected’

εi ∼ NID(0, σ 2 )

if yi∗ ≥ 100
if yi∗ < 100,

where yi is the observed score of student i and xi the vector of background characteristics (excluding an intercept).
a. Show that the above model can be written as the standard tobit model (tobit I).
b. First, the dean does a regression of yi upon xi and a constant (by OLS), using the
observed scores of 100 and more (yi ≥ 100). Show that this approach does not
lead to consistent or unbiased estimators for β1 .

254

c.
d.
e.

MODELS WITH LIMITED DEPENDENT VARIABLES

Explain in detail how the parameter vector β = (β0 , β1 ) can be estimated consistently, using the observed scores only.
Explain how you would estimate this model using all observations. Why is this
estimator preferable to the one of c? (No proof or derivations are required.)
The dean considers specifying a tobit II model (a sample-selection model). Describe
this model. Is this model adequate for the above problem?

8

Univariate Time Series
Models

One objective of analysing economic data is to predict or forecast the future values
of economic variables. One approach to do this is to build a more or less structural
econometric model, describing the relationship between the variable of interest with
other economic quantities, to estimate this model using a sample of data, and to use
it as the basis for forecasting and inference. Although this approach has the advantage
of giving economic content to one’s predictions, it is not always very useful. For
example, it may be possible to adequately model the contemporaneous relationship
between unemployment and the inﬂation rate, but as long as we cannot predict future
inﬂation rates we are also unable to forecast future unemployment.
In this chapter we follow a different route: a pure time series approach. In this
approach the current values of an economic variable are related to past values (either
directly or indirectly). The emphasis is purely on making use of the information in
past values of a variable for forecasting its future. In addition to producing forecasts,
time series models also produce the distribution of future values, conditional upon the
past, and can thus be used to evaluate the likelihood of certain events.
In this chapter we discuss the class of so-called ARIMA models that is developed to
model time series processes. In Sections 8.1 and 8.2, we analyse the properties of these
models and how they are related. An important issue is whether a time series process is
stationary, which implies that the distribution of the variable of interest does not depend
upon time. Nonstationarity can arise from different sources but an important one is the
presence of so-called unit roots. Sections 8.3 and 8.4 discuss this problem and how
one can test for this type of nonstationarity, while an empirical example concerning
exchange rates and prices is provided in Section 8.5. In Section 8.6, we discuss how the
parameters in the statistical models can be estimated, while Section 8.7 explains how
an appropriate ARIMA model is chosen. Section 8.8 demonstrates how the resulting
estimated univariate time series model can be used to forecast future values of an
economic variable. To illustrate the use of such forecasts in an economic context,

UNIVARIATE TIME SERIES MODELS

256

Section 8.9 analyses the expectations theory of the term structure of interest rates.
Finally, Section 8.10 presents autoregressive conditional heteroskedasticity models that
explain the variance of a series (of error terms) from its history.
The seminal work on the estimation and identiﬁcation of ARIMA models is the
monograph by Box and Jenkins (1976). Additional details and a discussion of more
recent topics can be found in many textbooks on time series analysis. Mills (1990),
Diebold (1998) and Enders (2004) are particularly suited for economists. At a more
advanced level, Hamilton (1994) provides an excellent exposition.

8.1

Introduction

In general we consider a time series of observations on some variable, e.g. the unemployment rate, denoted as Y1 , . . . , YT . These observations will be considered realizations of random variables that can be described by some stochastic process. It is the
properties of this stochastic process that we try to describe by a relatively simple
model. It will be of particular importance how observations corresponding to different
time periods are related, so that we can exploit the dynamic properties of the series to
generate predictions for future periods.
8.1.1 Some Examples

A simple way to model dependence between consecutive observations states that Yt
depends linearly upon its previous value Yt−1 . That is,
Yt = δ + θ Yt−1 + εt ,

(8.1)

where εt denotes a serially uncorrelated innovation with a mean of zero and a constant
variance. The process in (8.1) is referred to as a ﬁrst order autoregressive process
or AR(1) process. It says that the current value Yt equals a constant δ plus θ times
its previous value plus an unpredictable component εt . We have seen processes like
this before when discussing (ﬁrst order) autocorrelation in the linear regression model.
For the moment, we shall assume that |θ | < 1. The process for εt is an important
building block of time series models and is referred to as a white noise process. In
this chapter, εt will always denote such a process that is homoskedastic and exhibits
no autocorrelation.
The expected value of Yt can be solved from
E{Yt } = δ + θ E{Yt−1 },
which, assuming that E{Yt } does not depend upon t, allows us to write
µ ≡ E{Yt } =

δ
.
1−θ

(8.2)

Deﬁning yt ≡ Yt − µ, we can write (8.1) as
yt = θyt−1 + εt .

(8.3)

INTRODUCTION

257

Writing time series models in terms of yt rather than Yt is often notationally more
convenient, and we shall do so frequently in the rest of this chapter. One can allow
for nonzero means by adding an intercept term to the model. While Yt is observable,
yt is only observed if the mean of the series is known. Note that V {yt } = V {Yt }.
The model in (8.1) is a parsimonious way of describing a process for the Yt series
with certain properties. That is, the model in (8.1) implies restrictions on the time
series properties of the process that generates Yt . In general, the joint distribution of
all values of Yt is characterized by the so-called autocovariances, the covariances
between Yt and one of its lags, Yt−k . For the AR(1) model, the dynamic properties of
the Yt series can easily be determined using (8.1) or (8.3) if we impose that variances
and autocovariances do not depend upon the index t. This is a so-called stationarity
assumption and we return to it below. Writing
V {Yt } = V {θ Yt−1 + εt } = θ 2 V {Yt−1 } + V {εt }
and imposing that V {Yt } = V {Yt−1 }, we obtain
V {Yt } =

σ2
.
1 − θ2

(8.4)

It is clear from the resulting expression that we can only impose V {Yt } = V {Yt−1 } if
|θ | < 1, as was assumed before. Furthermore, we can determine that
cov{Yt , Yt−1 } = E{yt yt−1 } = E{(θyt−1 + εt )yt−1 } = θ V {yt−1 } = θ

σ2
1 − θ2

(8.5)

and, generally (for k = 1, 2, 3, . . .),
cov{Yt , Yt−k } = θ k

σ2
.
1 − θ2

(8.6)

As long as θ is nonzero, any two observations on Yt have a nonzero correlation, while
this dependence is smaller (and potentially arbitrary close to zero) if the observations
are further apart. Note that the covariance between Yt and Yt−k depends on k only, not
on t. This reﬂects the stationarity of the process.
Another simple time series model is the ﬁrst order moving average process or MA(1)
process, given by
Yt = µ + εt + αεt−1 .
(8.7)
Apart from the mean µ, this says that Y1 is a weighted average of ε1 and ε0 , Y2 is a
weighted average of ε2 and ε1 , etc. The values of Yt are deﬁned in terms of drawings
from the white noise process εt . The variances and autocovariances in the MA(1) case
are given by
2
V {Yt } = E{(εt + αεt−1 )2 } = E{εt2 } + α 2 E{εt−1
} = (1 + α 2 )σ 2
2
cov{Yt , Yt−1 } = E{(εt + αεt−1 )(εt−1 + αεt−2 )} = αE{εt−1
} = ασ 2

cov{Yt , Yt−2 } = E{(εt + αεt−1 )(εt−2 + αεt−3 )} = 0

UNIVARIATE TIME SERIES MODELS

258

or, in general,
cov{Yt , Yt−k } = 0,

for k = 2, 3, 4, . . .

Consequently, the simple moving average structure implies that observations that are
two or more periods apart are uncorrelated. Clearly, the AR(1) and MA(1) processes
imply very different autocovariances for Yt .
As we shall see below, both the autoregressive model and the moving average
model can be generalized by including additional lags in (8.1) or (8.7), respectively.
Apart from a few exceptions that we shall touch upon below, there are no fundamental
differences between autoregressive and moving average processes. The choice is simply
a matter of parsimony. For example, we can rewrite the AR(1) model as an inﬁnite
order moving average process, provided that |θ | < 1. To see this, substitute Yt−1 =
δ + θ Yt−2 + εt−1 into (8.1) to obtain
Yt = µ + θ 2 (Yt−2 − µ) + εt + θ εt−1 ,
which, after repeated substitution, results in
Yt = µ + θ n (Yt−2 − µ) +

n−1


θ j εt−j .

(8.8)

j =0

If we allow n → ∞, the second term of the right-hand side will converge to zero
(because |θ | < 1) and we obtain
Yt = µ +

∞


θ j εt−j .

(8.9)

j =0

This expression is referred to as the moving average representation of the autoregressive
process: the AR process in (8.1) is written as an inﬁnite order moving average processes. We can do so provided that |θ | < 1. As we shall see below, for some purposes
a moving average representation is more convenient than an autoregressive one.
In the discussion above, we assumed that the process for Yt is stationary. Before
discussing general autoregressive and moving average processes, the next subsection
pays attention to the important concept of stationarity.
8.1.2 Stationarity and the Autocorrelation Function

A stochastic process is said to be strictly stationary if its properties are unaffected
by a change of time origin; in other words, the joint probability distribution at any
set of times is not affected by an arbitrary shift along the time axis. This implies
that the distribution of Y1 is the same as that of any other Yt , and also, e.g. that the
covariances between Yt and Yt−k for any k do not depend upon t. Usually, we will
only be concerned with the means, variances and covariances of the series, and it is
sufﬁcient to impose that these moments are independent of time, rather than the entire
distribution. This is referred to as weak stationarity or covariance stationarity.

INTRODUCTION

259

Formally, a process {Yt } is deﬁned to be weakly stationary if for all t it holds that
E{Yt } = µ < ∞

(8.10)

V {Yt } = E{(Yt − µ)2 } = γ0 < ∞

(8.11)

cov{Yt , Yt−k } = E{(Yt − µ)(Yt−k − µ)} = γk ,

k = 1, 2, 3, . . .

(8.12)

In the sequel the term ‘stationary’ is taken to mean ‘weakly stationary’. Conditions
(8.10) and (8.11) require the process to have a constant ﬁnite mean and variance,
while (8.12) states that the autocovariances of Yt depend only upon the distance in
time between the two observations. The mean, variances and autocovariances are thus
independent of time. Strict stationarity is stronger1 as it requires that the whole distribution is unaffected by a change in time horizon, not just the ﬁrst and second order
moments. Obviously, under joint normality the distribution is completely characterized by ﬁrst and second order moments, and strict stationarity and weak stationarity
are equivalent.
Under covariance stationarity, we can deﬁne the k-th order autocovariance γk as
γk = cov{Yt , Yt−k } = cov{Yt−k , Yt },

(8.13)

which, for k = 0, gives the variance of Yt . As the autocovariances are not independent
of the units in which the variables are measured, it is common to standardize by
deﬁning autocorrelations ρk as
ρk =

cov{Yt , Yt−k }
γ
= k.
V {Yt }
γ0

(8.14)

Note that ρ0 = 1, while −1 ≤ ρk ≤ 1. The autocorrelations considered as a function
of k are referred to as the autocorrelation function (ACF) or, sometimes, the correlogram of the series Yt . The ACF plays a major role in modelling the dependencies
among observations, because it characterizes the process describing the evolution of
Yt over time. In addition to ρk , the process of Yt is described by its mean and its
variance γ0 .
From the ACF we can infer the extent to which one value of the process is correlated
with previous values and thus the length and strength of the memory of the process.
It indicates how long (and how strongly) a shock in the process (εt ) affects the values
of Yt . For the two processes we have seen above, we have the following. For the
AR(1) process
Yt = δ + θ Yt−1 + εt
we have autocorrelation coefﬁcients
ρk = θ k ,
while for the MA(1) process
Yt = µ + εt + αεt−1 ,
1

Strict stationarity does not necessarily imply that ﬁrst and second moments are ﬁnite.

(8.15)

UNIVARIATE TIME SERIES MODELS

260

we have
ρ1 =

α
and ρk = 0,
1 + α2

k = 2, 3, 4, . . .

(8.16)

90

80

70

100

14

12

10

8

6

Theoretical autocorrelation
function: AR (1), q = 0.9

4

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

2

14

12

rk

Figure 8.1

10

8

6

k

60

Time

Theoretical autocorrelation
function: AR (1), q = 0.5

4

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

2

rk

Time

50

40

30

0

20

Yt

−4

100

90

80

70

60

50

40

30

20

0

−4

AR (1) with q = 0.9

4
3
2
1
0
−1
−2
−3
10

AR (1) with q = 0.5

4
3
2
1
0
−1
−2
−3
10

Yt

Consequently, a shock in an MA(1) process affects Yt in two periods only, while a
shock in the AR(1) process affects all future observations with a decreasing effect.
As an illustration, we generated several artiﬁcial time series according to a ﬁrst order
autoregressive process as well as a ﬁrst order moving average process. The data for
the simulated AR(1) processes with parameter θ equal to 0.5 and 0.9 are depicted in
Figure 8.1, combined with their autocorrelation functions. All series are standardized
to have unit variance and zero mean. If we compare the AR series with θ = 0.5 and
θ = 0.9, it appears that the latter process is smoother, that is, has a higher degree
of persistence. This means that, after a shock, it takes longer for the series to return
to its mean. The autocorrelation functions show an exponential decay in both cases,
although it takes large lags for the ACF of the θ = 0.9 series to become close to zero.
For example, after 15 periods, the effect of a shock is still 0.915 = 0.21 of its original
effect. For the θ = 0.5 series, the effect at lag 15 is virtually zero.
The data and ACF for two simulated moving average processes, with α = 0.5 and
α = 0.9, are displayed in Figure 8.2. The difference between the two is less pronounced
than in the AR case. For both series, shocks only have an effect in two consecutive
periods. This means that in the absence of new shocks, the series are back at their
mean after two periods. The ﬁrst order autocorrelation coefﬁcients do not differ much,
and are 0.40 and 0.50, respectively.

k

First order autoregressive processes: data series and autocorrelation functions

GENERAL ARMA PROCESSES

261

MA (1) with a = 0.5

4
3

k

100

80

70

90
14

12

10

8

6

Theoretical autocorrelation
function: MA (1), a = 0.9

4

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

2

rk
14

12

10

8

6

60

k

Theoretical autocorrelation
function: MA (1), a = 0.5

4

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

2

rk

k

50

40

30

0

20

0
−1
−2
−3
−4

100

90

80

70

60

50

40

30

0

20

0
−1
−2

10

Yt

2
1

10

Yt

2
1

−3
−4

MA (1) with a = 0.9

4
3

k

Figure 8.2 First order moving average processes: data series and autocorrelation functions

8.2

General ARMA Processes

8.2.1 Formulating ARMA Processes

In this section, we deﬁne more general autoregressive and moving average processes.
First, we deﬁne a moving average process of order q, or in short an MA(q) process, as
yt = εt + α1 εt−1 + · · · + αq εt−q ,

(8.17)

where εt is a white noise process and yt = Yt − µ. That is, the demeaned series yt is a
weighted combination of q + 1 white noise terms. An autoregressive process of order
p, an AR(p) process, is given by
yt = θ1 yt−1 + θ2 yt−2 + · · · + θp yt−p + εt .

(8.18)

Obviously, it is possible to combine the autoregressive and moving average speciﬁcation into an ARMA(p, q) model, which consists of an AR part of order p and an MA
part of order q,
yt = θ1 yt−1 + · · · + θp yt−p + εt + α1 εt−1 + · · · + αq εt−q .

(8.19)

In fact, there is no fundamental difference between moving average and autoregressive
processes. Under suitable conditions (see below) an AR model can be written as an

UNIVARIATE TIME SERIES MODELS

262

MA model and vice versa. The order of one of these is usually quite long and the
choice for an MA, AR or a combined ARMA representation is a matter of parsimony.
For example, we have seen above that an AR(1) model can be written as an MA(∞),
a moving average model of inﬁnite order. For certain purposes, the AR representation
of the model is convenient, while for other purposes the MA representation is. This
will become clear below.
Often it is convenient to use the lag operator, denoted by L (some authors use B,
backshift operator). It is deﬁned by
Lyt = yt−1 .

(8.20)

Most of the time, the lag operator can be manipulated just as if it were a constant.
For example,
L2 yt = L(Lyt ) = Lyt−1 = yt−2 ,
so that, more generally, Lp yt = yt−p with L0 ≡ 1. Also, L−1 yt = yt+1 . Operating L
on a constant leaves the constant unaffected, e.g. Lµ = µ. Using this lag operator
allows us to write ARMA models in a concise way. For an AR(1) model we can write
yt = θ Lyt + εt

(8.21)

(1 − θ L)yt = εt .

(8.22)

or
This says that a combination of yt and its lag, with weights 1 and −θ , equals a white
noise process. Similarly, we can write a general AR(p) as
θ (L)yt = εt ,

(8.23)

where θ (L) is a polynomial of order p in the lag operator L, usually referred to as a
lag polynomial, given by
θ (L) = 1 − θ1 L − θ2 L2 − · · · − θp Lp .

(8.24)

We can interpret a lag polynomial as a ﬁlter that, if applied to a time series, produces
a new series. So the ﬁlter θ (L) applied to an AR(p) process yt produces a white
noise process εt . It is relatively easy to manipulate lag polynomials. For example,
transforming a series by two such polynomials one after the other, is the same as
transforming the series once by a polynomial that is the product of the two original
ones. This way, we can deﬁne the inverse of a ﬁlter, which is naturally given by the
inverse of the polynomial. Thus the inverse of θ (L), denoted as θ −1 (L), is deﬁned so
as to satisfy θ −1 (L)θ (L) = 1. If θ (L) is a ﬁnite order polynomial in L, its inverse will
be one of inﬁnite order. For the AR(1) case we ﬁnd
−1

(1 − θ L)

=

∞

j =0

θ j Lj

(8.25)

GENERAL ARMA PROCESSES

263


j
provided that |θ | < 1. This is similar to the result that the inﬁnite sum ∞
j =0 θ equals
−1
(1 − θ ) if |θ | < 1, while it does not converge for |θ | ≥ 1. In general, the inverse of
a polynomial θ (L) exists if it satisﬁes certain conditions on its parameters, in which
case we call θ (L) invertible. This is discussed in the next subsection. With (8.25) we
can write the AR(1) model as
(1 − θ L)−1 (1 − θ L)yt = (1 − θ L)−1 εt
or
yt =

∞


θ j Lj εt =

j =0

∞


θ j εt−j ,

(8.26)

j =0

which corresponds to (8.9) above.
Under appropriate conditions, the converse is also possible and we can write a
moving average model in autoregressive form. Using the lag operator we can write the
MA(1) process as
yt = (1 + αL)εt
and the general MA(q) process as
yt = α(L)εt
where

α(L) = 1 + α1 L + α2 L2 + · · · + αq Lq .

(8.27)

Note that we have deﬁned the polynomials such that the MA polynomial has plus signs,
while the AR polynomial has minus signs. Now, if α −1 (L) exists, we can write that
α −1 (L)yt = εt ,

(8.28)

which, in general, will be an AR model of inﬁnite order. For the MA(1) case, we use,
similar to (8.25), that
∞

(1 + αL)−1 =
(−α)j Lj ,
(8.29)
j =0

provided that |α| < 1. Consequently, an MA(1) model can be written as
yt = α

∞

(−α)j yt−j −1 + εt .

(8.30)

j =0

A necessary condition for the inﬁnite AR representation (AR(∞)) to exist is that
the MA polynomial is invertible, which, in the MA(1) case, requires that |α| < 1.
Particularly for making predictions conditional upon an observed past, the AR representations are very convenient (see Section 8.8 below). The MA representations are
often convenient to determine variances and covariances.

UNIVARIATE TIME SERIES MODELS

264

For a more parsimonious representation, we may want to work with an ARMA
model that contains both an autoregressive and a moving average part. The general
ARMA model can be written as
θ (L)yt = α(L)εt ,

(8.31)

which (if the AR lag polynomial is invertible) can be written in MA(∞) representation as
(8.32)
yt = θ −1 (L)α(L)εt ,
or (if the MA lag polynomial is invertible) in inﬁnite AR form as
α −1 (L)θ (L)yt = εt .

(8.33)

Both θ −1 (L)α(L) and α −1 (L)θ (L) are lag polynomials of inﬁnite length, with restrictions on the coefﬁcients.
8.2.2 Invertibility of Lag Polynomials

As we have seen above, the ﬁrst order lag polynomial 1 − θ L is invertible if |θ | < 1.
In this section, we shall generalize this condition to higher order lag polynomials. Let
us ﬁrst consider the case of a second order polynomial, given by 1 − θ1 L − θ2 L2 .
Generally, we can ﬁnd values φ1 and φ2 such that the polynomial can be written as
1 − θ1 L − θ2 L2 = (1 − φ1 L)(1 − φ2 L).

(8.34)

It is easily veriﬁed that φ1 and φ2 can be solved for from2 φ1 + φ2 = θ1 and −φ1 φ2 =
θ2 . The conditions for invertibility of the second order polynomial are just the conditions that both the ﬁrst order polynomials 1 − φ1 L and 1 − φ2 L are invertible. Thus,
the requirement for invertibility is that both |φ1 | < 1 and |φ2 | < 1.
These requirements can also be formulated in terms of the so-called characteristic equation
(1 − φ1 z)(1 − φ2 z) = 0.
(8.35)
This equation has two solutions, z1 and z2 say, referred to as the characteristic roots.
The requirement |φi | < 1 corresponds to |zi | > 1. If any solution satisﬁes |zi | ≤ 1 the
corresponding polynomial is non-invertible. A solution that equals unity is referred to
as a unit root.
The presence of a unit root in the lag polynomial θ (L) can be detected relatively
easily, without solving the characteristic
equation, by noting that the polynomial θ (z)
p
evaluated at z = 1 is zero if j =1 θj = 1. Thus, the presence of a ﬁrst unit root can
be veriﬁed by checking whether the sum of the polynomial coefﬁcients equals one. If
the sum exceeds one, the polynomial is not invertible.
2

It is possible that φ1 , φ2 is a pair of complex numbers, for example if θ1 = 0 and θ2 < 0. In the text we
shall ignore this possibility.

GENERAL ARMA PROCESSES

265

As an example, consider the AR(2) model
yt = 1.2yt−1 − 0.32yt−2 + εt .

(8.36)

(1 − 0.8L)(1 − 0.4L)yt = εt ,

(8.37)

1 − 1.2z + 0.32z2 = (1 − 0.8z)(1 − 0.4z) = 0.

(8.38)

We can write this as

with characteristic equation

The solutions (characteristic roots) are 1/0.8 and 1/0.4, which are both larger than one.
Consequently, the AR polynomial in (8.36) is invertible. Note that the AR(1) model
yt = 1.2yt−1 + εt

(8.39)

describes a non-invertible AR process.
The issue whether or not the lag polynomials are invertible is important for several
reasons. For moving average models, or more generally, models with a moving average component, invertibility of the MA polynomial is important for estimation and
prediction. For models with an autoregressive part, the AR polynomial is invertible if
and only if the process is stationary. Section 8.3 explores this last issue.
8.2.3 Common Roots

Decomposing the moving average and autoregressive polynomials into products of
linear functions in L also shows the problem of common roots or cancelling roots.
This means that the AR and the MA part of the model have a root that is identical
and the corresponding linear functions in L cancel out. To illustrate this, let the true
model be an ARMA(2, 1) process, described by
(1 − θ1 L − θ2 L2 )yt = (1 + αL)εt .
Then, we can write this as
(1 − φ1 L)(1 − φ2 L)yt = (1 + αL)εt .

(8.40)

Now, if α = −φ1 , we can divide both sides by (1 + αL) to obtain
(1 − φ2 L)yt = εt ,
which is exactly the same as (8.40). Thus, in the case of one cancelling root, an
ARMA(p, q) model can be written equivalently as an ARMA(p − 1, q − 1) model.
As an example, consider the model
yt = yt−1 − 0.25yt−2 + εt − 0.5εt−1 ,

(8.41)

UNIVARIATE TIME SERIES MODELS

266

which can be rewritten as
(1 − 0.5L)(1 − 0.5L)yt = (1 − 0.5L)εt .
Clearly, this can be reduced to an AR(1) model as
(1 − 0.5L)yt = εt
or
yt = 0.5yt−1 + εt ,
which describes exactly the same process as (8.41).
The problem of common roots illustrates why it may be problematic, in practice, to
estimate an ARMA model with an AR and an MA part of a high order. The reason
is that identiﬁcation and estimation are hard if roots of the MA and AR polynomial
are almost identical. In this case, a simpliﬁed ARMA(p − 1, q − 1) model will yield
an almost equivalent representation.
In Section 8.6, we shall discuss estimation of ARMA models. First, however, we
pay more attention to stationarity and unit roots in Section 8.3 and discuss several
tests for the presence of a unit root in Section 8.4. An empirical illustration concerning
long-run purchasing power parity is provided in Section 8.5.

8.3

Stationarity and Unit Roots

Stationarity of a stochastic process requires that the variances and autocovariances
are ﬁnite and independent of time. It is easily veriﬁed that ﬁnite order MA processes
are stationary by construction, because they correspond to a weighted sum of a ﬁxed
number of stationary white noise processes. Of course, this result breaks down if we
would allow the MA coefﬁcients to vary over time, as in
yt = εt + g(t)εt−1 ,

(8.42)

where g(t) is some deterministic function of t. Now we have
E{yt2 } = σ 2 + g 2 (t)σ 2 ,
which is not independent of t. Consequently, the process in (8.42) is nonstationary.
Stationarity of autoregressive or ARMA processes is less trivial. Consider, for
example, the AR(1) process
yt = θyt−1 + εt ,
(8.43)
with θ = 1. Taking variances on both sides gives V {yt } = V {yt−1 } + σ 2 , which has no
solution for the variance of the process consistent with stationarity, unless σ 2 = 0, in
which case an inﬁnity of solutions exists. The process in (8.43) is a ﬁrst order autoregressive process with a unit root (θ = 1), usually referred to as a random walk. The
unconditional variance of yt does not exist, i.e. is inﬁnite and the process is nonstationary. In fact, for any value of θ with |θ | ≥ 1, (8.43) describes a nonstationary process.

STATIONARITY AND UNIT ROOTS

267

We can formalize the above results as follows. The AR(1) process is stationary if
and only if the polynomial 1 − θ L is invertible, that is, if the root of the characteristic
equation 1 − θ z = 0 is larger than unity. This result is straightforwardly generalized
to arbitrary ARMA models. The ARMA(p, q) model
θ (L)yt = α(L)εt

(8.44)

corresponds to a stationary process if and only if the solutions z1 , . . . , zp to θ (z) = 0
are larger than one (in absolute value), that is when the AR polynomial is invertible.
For example, the ARMA(2, 1) process given by
yt = 1.2yt−1 − 0.2yt−2 + εt − 0.5εt−1

(8.45)

is nonstationary because z = 1 is a solution to 1 − 1.2z + 0.2z2 = 0.
A special case that is of particular interest arises when one root is exactly equal to
one, while the other roots are larger than one. If this arises we can write the process
for yt as
(8.46)
θ ∗ (L)(1 − L)yt = θ ∗ (L) yt = α(L)εt ,
where θ ∗ (L) is an invertible polynomial in L of order p − 1, and ≡ 1 − L is the
ﬁrst difference operator. Because the roots of the AR polynomial are the solutions
to θ ∗ (z)(1 − z) = 0 there is one solution z = 1, or in other words a single unit root.
Equation (8.46) thus shows that yt can be described by a stationary ARMA model
if the process for yt has one unit root. Consequently, we can eliminate the nonstationarity by transforming the series into ﬁrst differences (changes). Writing the process in
(8.45) as
(1 − 0.2L)(1 − L)yt = (1 − 0.5L)εt
shows that it implies that

yt is described by a stationary ARMA(1, 1) process given by
yt = 0.2 yt−1 + εt − 0.5εt−1 .

A series which becomes stationary after ﬁrst differencing is said to be integrated
of order one, denoted I (1). If yt is described by a stationary ARMA(p, q) model,
we say that yt is described by an autoregressive integrated moving average (ARIMA)
model of order p, 1, q, or in short an ARIMA(p, 1, q) model.
First differencing quite often can transform a nonstationary series into a stationary
one. In particular this may be the case for aggregate economic series or their natural logarithms. Note that when Yt is, for example, the log of national income, Yt
corresponds to the income growth rate, which is not unlikely to be stationary. Note
that the AR polynomial is required to have an exact unit root. If the true model is
an AR(1) with θ = 1.01, we have that yt = 0.01yt−1 + εt , which is nonstationary,
as it depends upon the nonstationary process yt . Consequently, an AR(1) process with
θ = 1.01 is not integrated of order one.
In some cases, taking ﬁrst differences is insufﬁcient to obtain stationarity and another
differencing step is required. In this case the stationary series is given by ( Yt ) =
Yt − Yt−1 , corresponding to the change in the growth rate for logarithmic variables.
If a series must be differenced twice before it becomes stationary, then it is said to be

UNIVARIATE TIME SERIES MODELS

268

integrated of order two, denoted I (2), and it must have two unit roots. Thus, a series
Yt is I (2) if Yt is nonstationary but 2 Yt is stationary. A more formal deﬁnition of
integration is given in Engle and Granger (1987), who also deﬁne higher orders of
integration, which are not very relevant in economic applications. Thus, a time series
integrated of order zero is stationary in levels, while for a time series integrated of
order one, the ﬁrst difference is stationary. A white noise series and a stable AR(1)
process are examples of I (0) series, while a random walk process, as described by
(8.43) with θ = 1, is an example of an I (1) series.
In the long run, it can make a surprising amount of difference whether the series has
an exact unit root or whether the root is slightly larger than one. It is the difference
between being I (0) and being I (1). In general, the main differences between processes
that are I (0) and I (1) can be summarized as follows. An I (0) series ﬂuctuates around its
mean with a ﬁnite variance that does not depend on time, while an I (1) series wanders
widely. Typically, it is said that an I (0) series is mean reverting, as there is a tendency
in the long run to return to its mean. Furthermore, an I (0) series has a limited memory
of its past behaviour (implying that the effects of a particular random innovation are
only transitory), while an I (1) process has an inﬁnitely long memory (implying that an
innovation will permanently affect the process). This last aspect becomes clear from
the autocorrelation functions: for an I (0) series the autocorrelations decline rapidly as
the lag increases, while for the I (1) process the estimated autocorrelation coefﬁcients
decay to zero only very slowly.
The last property makes the presence of a unit root an interesting question from
an economic point of view. In models with unit roots, shocks (which may be due
to policy interventions) have persistent effects that last forever, while in the case of
stationary models, shocks can only have a temporary effect. Of course, the long-run
effect of a shock is not necessarily of the same magnitude as the short-run effect.
Consequently, starting in the early 1980s a vast amount of literature has appeared3
on the presence of unit roots in many macro-economic time series, with – depending
upon the particular technique applied – sometimes conﬂicting conclusions. The fact
that the autocorrelations of a stationary series taper off or die out rapidly may help in
determining the degree of differencing needed to achieve stationarity (usually referred
to as d). In addition, a number of formal unit root tests has been proposed in the
literature, some of which we shall discuss in Section 8.4 below.
Empirical series where the choice between a unit root (nonstationarity) and a ‘near
unit root’ (stationarity) is particularly ambiguous are interest rate series. The high
degree of persistence in interest rates quite often makes the unit root hypothesis statistically not rejectable, although nonstationary interest rates do not seem to be very
plausible from an economic point of view. The empirical example in Section 8.9 illustrates this issue.

8.4

Testing for Unit Roots

To introduce the testing procedures for a unit root we concentrate on autoregressive
models. This may not be particularly restrictive since any ARMA model will always
have an AR representation (provided the MA polynomial α(L) is invertible).
3

The most inﬂuential study is Nelson and Plosser (1982), which argues that many economic time series
are better characterized by unit roots than by deterministic trends.

TESTING FOR UNIT ROOTS

269

8.4.1 Testing for Unit Roots in a First Order Autoregressive Model

Let us ﬁrst consider the AR(1) process
Yt = δ + θ Yt−1 + εt ,

(8.47)

where θ = 1 corresponds to a unit root. As the constant in a stationary AR(1) model
satisﬁes δ = (1 − θ )µ, where µ is the mean of the series, the null hypothesis of a
unit root also implies that the intercept term should be zero. Although it is possible
to jointly test the two restrictions δ = 0 and θ = 1, it is easier (and more common)
to test only that θ = 1. It seems obvious to use the estimate θ̂ for θ from an ordinary
least squares procedure (which is consistent, irrespective of the true value of θ ) and
the corresponding standard error to test the null hypothesis. However, as was shown in
the seminal paper of Dickey and Fuller (1979), under the null that θ = 1 the standard
t-ratio does not have a t distribution, not even asymptotically. The reason for this is
the nonstationarity of the process invalidating standard results on the distribution of
the OLS estimator θ̂ (as discussed in Chapter 2). For example, if θ = 1 the variance
of Yt , denoted by γ0 , is not deﬁned (or, if you want, is inﬁnitely large). For any ﬁnite
sample size, however, a ﬁnite estimate of the variance for Yt will be obtained.
To test the null hypothesis that θ = 1, it is possible to use the standard t-statistic
given by
θ̂ − 1
DF =
,
(8.48)
se(θ̂ )
where se(θ̂ ) denotes the usual OLS standard error. Critical values, however, have to
be taken from the appropriate distribution, which under the null hypothesis of nonstationarity is nonstandard. In particular, the distribution is skewed to the right so that
critical values are smaller than those for (the normal approximation of) the t distribution. Using a 5% signiﬁcance level in a one-tailed test of H0 : θ = 1 (a unit root)
against H1 : |θ | < 1 (stationarity), the correct critical value in large samples is −2.86
rather than −1.65 for the normal approximation. Consequently, if you use the standard
t tables you may reject a unit root too often. Selected percentiles of the appropriate distribution are published in several works by Dickey and Fuller. In Table 8.1 we present
1% and 5% critical values for this test, usually referred to as the Dickey–Fuller test,
for a range of different samples sizes.
Table 8.1 1% and 5% critical values for Dickey–Fuller tests
(Fuller, 1976, p. 373)
Without trend
Sample size
T
T
T
T
T
T

= 25
= 50
= 100
= 250
= 500
=∞

With trend

1%

5%

1%

5%

−3.75
−3.58
−3.51
−3.46
−3.44
−3.43

−3.00
−2.93
−2.89
−2.88
−2.87
−2.86

−4.38
−4.15
−4.04
−3.99
−3.98
−3.96

−3.60
−3.50
−3.45
−3.43
−3.42
−3.41

UNIVARIATE TIME SERIES MODELS

270

Usually, a slightly more convenient regression procedure is used. In this case, the
model is rewritten as
Yt = δ + (θ − 1)Yt−1 + εt ,
(8.49)
from which the t-statistic for θ − 1 = 0 is identical to DF above. The reason for this
is that the least squares method is invariant to linear transformations of the model.
It is possible that (8.49) holds with θ = 1 and a nonzero intercept δ = 0. Because in
this case δ cannot equal (1 − θ )µ, (8.49) cannot be derived from a pure AR(1) model.
This is seen by considering the resulting process
Y t = δ + εt ,

(8.50)

which is known as a random walk with drift, where δ is the drift parameter. In
the model for the level variable Yt , δ corresponds to a linear time trend. Because
(8.50) implies that E{ Yt } = δ, it is the case that (for a given starting value Y0 )
E{Yt } = Y0 + δt. This shows that the interpretation of the intercept term in (8.49)
depends heavily upon the presence of a unit root. In the stationary case, δ reﬂects the
non-zero mean of the series; in the unit root case, it reﬂects a deterministic trend in
Yt . Because in the latter case ﬁrst differencing produces a stationary time series, the
process for Yt is referred to as difference stationary. In general, a difference stationary
process is a process that can be made stationary by differencing.
It is also possible that nonstationarity is caused by the presence of a deterministic
time trend in the process, rather than by the presence of a unit root. This happens when
the AR(1) model is extended to
Yt = δ + θ Yt−1 + γ t + εt ,

(8.51)

with |θ | < 1 and γ = 0. In this case, we have a nonstationary process because of the
linear trend γ t. This nonstationarity can be removed by regressing Yt upon a constant
and t, and then considering the residuals of this regression, or by simply including t as
additional variable in the model. The process for Yt in this case is referred to as being
trend stationary. Nonstationary processes may thus be characterized by the presence
of a deterministic trend, like γ t, a stochastic trend implied by the presence of a unit
root, or both.
It is possible to test whether Yt follows a random walk against the alternative
that it follows the trend stationary process in (8.51). This can be tested by running
the regression
Yt = δ + (θ − 1)Yt−1 + γ t + εt .
(8.52)
The null hypothesis one would like to test is that the process is a random walk rather
than trend stationary and corresponds to H0 : δ = γ = θ − 1 = 0. Instead of testing this
joint hypothesis, it is quite common to use the t-ratio corresponding to θ̂ − 1, denoted
DFτ , assuming that the other restrictions in the null hypotheses are satisﬁed. Although
the null hypothesis is still the same as in the previous unit root test, the testing regression is different and thus we have a different distribution of the test statistic. The critical
values for DFτ , given in the last two columns of Table 8.1, are still smaller than those
for DF. In fact, with an intercept and a deterministic trend included the probability

TESTING FOR UNIT ROOTS

271

that θ̂ − 1 is positive (given that the true value θ − 1 equals zero) is negligibly small.
It should be noted, however, that if the unit root hypothesis θ − 1 = 0 is rejected, we
cannot conclude that the process for Yt is likely to be stationary. Under the alternative
hypothesis γ may be nonzero so that the process for Yt is not stationary (but only
trend stationary).
The phrase Dickey–Fuller test, or simply DF test is used for any of the tests described
above and can thus be based upon a regression with or without a trend.4 If a graphical
inspection of the series indicates a clear positive or negative trend, it is most appropriate to perform the Dickey–Fuller test with a trend. This implies that the alternative
hypothesis allows the process to exhibit a linear deterministic trend. It is important
to stress that the unit root hypothesis corresponds to the null hypothesis. If we are
unable to reject the presence of a unit root it does not necessarily mean that it is true.
It could just be that there is insufﬁcient information in the data to reject it. Of course,
this is simply the general difference between accepting a hypothesis and not rejecting
it. Because the long-run properties of the process depend crucially upon the imposition
of a unit root or not, this is something to be aware of. Not all series for which we
cannot reject the unit root hypothesis are necessarily integrated of order one.
To circumvent the problem that unit root tests often have low power, Kwiatkowski,
Phillips, Schmidt and Shin (1992) propose an alternative test where stationarity is the
null hypothesis and the existence of a unit root is the alternative. This test is usually
referred to as the KPSS test. The basic idea is that a time series is decomposed into the
sum of a deterministic time trend, a random walk and a stationary error term (typically
not white noise). The null hypothesis (of trend stationarity) speciﬁes that the variance
of the random walk component is zero. The test is actually a Lagrange multiplier
test (see Chapter 6) and computation of the test statistic is fairly simple. First, run an
auxiliary regression of Yt upon an intercept and
 a time trend t. Next, save the OLS
residuals et and compute the partial sums St = ts=1 es for all t. Then the test statistic
is given by
T

KPSS =
St2 /σ̂ 2 ,
(8.53)
t=1

where σ̂ 2 is an estimator for the error variance. This latter estimator σ̂ 2 may involve
corrections for autocorrelation based on the Newey–West formula (see Chapter 4).
The asymptotic distribution is non-standard, and Kwiatkowski et al. (1992) report
a 5% critical value of 0.146. If the null hypothesis is stationarity rather than trend
stationarity the trend term should be omitted from the auxiliary regression. The test
statistic is then computed in the same fashion, but the 5% critical value is 0.463.
8.4.2 Testing for Unit Roots in Higher Order Autoregressive Models

A test for a single unit root in higher order AR processes can easily be obtained
by extending the Dickey–Fuller test procedure. The general strategy is that lagged
differences, such as Yt−1 , Yt−2 , . . . , are included in the regression, such that its error
term corresponds to white noise. This leads to the so-called augmented Dickey–Fuller
4

If the mean of the series is known to be zero, the intercept term may be dropped from the regressions
leading to a third variant of the Dickey–Fuller test. This test is rarely used in practice.

UNIVARIATE TIME SERIES MODELS

272

tests (ADF tests), for which the same asymptotic critical values hold as those shown
in Table 8.1.
Consider the AR(2) model
Yt = δ + θ1 Yt−1 + θ2 Yt−2 + εt

(8.54)

which can be written in factorized form as
(1 − φ1 L)(1 − φ2 L)(Yt − µ) = εt .

(8.55)

The stationarity condition requires that φ1 and φ2 are both less than one in absolute
value, but if φ1 = 1 and |φ2 | < 1, we have a single unit root, θ1 + θ2 = 1 and θ2 = −φ2 .
Equation (8.54) can be used to test the unit root hypothesis by testing θ1 + θ2 = 1,
given |θ2 | < 1. This is conveniently done be rewriting (8.54) as
Yt = δ + (θ1 + θ2 − 1)Yt−1 − θ2 Yt−1 + εt .

(8.56)

The coefﬁcients in (8.56) can be consistently estimated by ordinary least squares and
the estimate of the coefﬁcient for Yt−1 provides a means for testing the null hypothesis
π ≡ θ1 + θ2 − 1 = 0. The resulting t-ratio, π̂/se(π̂ ), has the same distribution as DF
above. In the spirit of the Dickey–Fuller procedure, one might add a time trend to the
test regression. Depending on which variant is used, the resulting test statistic has to
be compared with a critical value taken from the appropriate row of Table 8.1.
This procedure can easily be generalized to the testing of a single unit root in an
AR(p) process. The trick is that any AR(p) process can be written as
Yt = δ + πYt−1 + c1 Yt−1 + · · · + cp−1 Yt−p+1 + εt ,

(8.57)

with π = θ1 + · · · + θp − 1 and suitably chosen constants c1 , . . . , cp−1 . As π = 0
implies θ (1) = 0 it also implies that z = 1 is a solution to the characteristic equation
θ (z) = 0. Thus, as before, the hypothesis that π = 0 corresponds to a unit root and
we can test it using the corresponding t-ratio. If the AR(p) assumption is correct and
under the null hypothesis of a unit root, the asymptotic distributions of the DF or
DFτ statistics, calculated from (8.57) (including a time trend, where appropriate) are
the same as before. The small sample critical values are somewhat different from the
tabulated ones and are provided by, for example, MacKinnon (1991).
Thus, when Yt follows an AR(p) process, a test for a single unit root can be constructed from a regression of Yt on Yt−1 and Yt−1 , . . . , Yt−p+1 by testing the
signiﬁcance of the ‘level’ variable Yt−1 (using the one-sided appropriate critical values). It is interesting to note that under the null hypothesis of a single unit root, all
variables in (8.57) are stationary, except Yt−1 . Therefore, the equality in (8.57) can
only make sense if Yt−1 does not appear and π = 0, which explains intuitively why
the unit root hypothesis corresponds to π = 0. The inclusion of the additional lags,
in comparison to the standard Dickey–Fuller test, is done to make the error term in
(8.57) asymptotically a white noise process, which is required for the distributional
results to be valid. As it will generally be the case that p is unknown it is advisable
to choose p fairly large. If too many lags are included this will somewhat reduce the

TESTING FOR UNIT ROOTS

273

power of the tests, but if too few lags are included the asymptotic distributions from
the table are simply not valid (because of autocorrelation in the residuals), and the
tests may lead to seriously biased conclusions. It is possible to use the model selection
criteria discussed in Subsection 8.7.4 below, or statistical signiﬁcance of the additional
variables to select the lag length in the ADF tests.
A regression of the form (8.57) can also be used to test for a unit root in a general
(invertible) ARMA model. Said and Dickey (1984) argue that when, theoretically, one
lets the number of lags in the regression grow with the sample size (at a cleverly
chosen rate), the same asymptotic distributions hold and the ADF tests are also valid
for an ARMA model with a moving average component. The argument essentially is,
as we have seen before, that any ARMA model (with invertible MA polynomial) can
be written as an inﬁnite autoregressive process. This explains why, when testing for
unit roots, people usually do not worry about MA components.
Phillips and Perron (1988) have suggested an alternative to the augmented
Dickey–Fuller tests. Instead of adding additional lags in the regressions to obtain
an error term that has no autocorrelation, they stick to the original Dickey–Fuller
regressions, but adjust the DF-statistics to take into account the (potential)
autocorrelation pattern in the errors. These adjustments, based on corrections similar
to those applied to compute Newey–West (HAC) standard errors (see Chapter 4), are
quite complicated and will not be discussed here. The (asymptotic) critical values are
again the same as those reported in Table 8.1. The Phillips–Perron test, sometimes
referred to as a nonparametric test for a unit root, is, like the Said–Dickey (or ADF)
test, applicable for general ARMA models (see Hamilton, 1994, pp. 506–515, for more
details). Monte Carlo studies do not show a clear ranking of the two tests regarding
their power (probability to reject the null if it is false) in ﬁnite samples.
If the ADF test does not allow rejection of the null hypothesis of one unit root,
the presence of a second unit root may be tested by estimating the regression of
2
Yt on Yt−1 , 2 Yt−1 , . . . , 2 Yt−p+1 , and comparing the t-ratio of the coefﬁcient
on Yt−1 with the appropriate critical value from Table 8.1. Alternatively, the presence of two unit roots may be tested jointly by estimating the regression of 2 Yt on
Yt−1 , Yt−1 , 2 Yt−1 , . . . , 2 Yt−p+1 , and computing the usual F -statistic for testing the
joint signiﬁcance of Yt−1 and Yt−1 . Again, though, this test statistic has a distribution
under the null hypothesis of a double unit root that is not the usual F -distribution.
Percentiles of this distribution are given by Hasza and Fuller (1979).
8.4.3 Extensions

Before moving to an illustration, let us stress that a stochastic process may be nonstationary for other reasons than the presence of one or two unit roots. A linear
deterministic trend is one example, but many other forms of nonstationarity are possible. To illustrate this, note that if the process for Yt is nonstationary so will the process
for log Yt . However, at most one of these process will be characterized by a unit root.
Without going into details it may be mentioned that the recent literature on unit roots
also includes discussions of stochastic unit roots, seasonal unit roots, fractional integration and panel data unit root tests. A stochastic unit root implies that a process is
characterized by a root that is not constant, but stochastic and varying around unity.
Such a process can be stationary for some periods and mildly explosive for others

UNIVARIATE TIME SERIES MODELS

274

(see Granger and Swanson, 1997). A seasonal unit root arises if a series becomes
stationary after seasonal differencing. For example, if the monthly series Yt − Yt−12
is stationary while Yt is not (see Patterson, 2000, Section 7.7, for an intuitive discussion). Fractional integration starts from the idea that a series may be integrated of
order d, where d is not an integer. If d ≥ 0.5, the process is nonstationary and said to
be fractionally integrated. By allowing d to take any value between 0 and 1, the gap
between stationary and nonstationary processes is closed; see Gourieroux and Jasiak
(2001, Chapter 5). Finally, panel data unit root tests involve tests for unit roots in multiple series, for example GDP in ten different countries. This extension is discussed in
Chapter 10 below.
8.4.4 Illustration: Annual Price/Earnings Ratio

In this subsection, we consider annual data on the ratio of the S&P Composite Stock
Price Index and S&P Composite Earnings, popularly referred to as the price/earnings
ratio over the period 1871–2002 (T = 132).5 The question whether valuation ratios,
like the price/earnings ratio, are mean reverting has received considerable attention in
the literature and has interesting implications for forecasting future stock prices. For
example, Campbell and Shiller (1998) argue that the high price/earnings ratios observed
in the late 1990s imply a decline in future stock prices to bring the ratio into line with
its historical level. First, we plot the log of the series in Figure 8.3. Seemingly, the
series ﬂuctuates around a long-run average, although it sometimes takes many years
for the series to revert to its mean. Using the above methodology, we test for the
presence of one or two unit roots in the log price/earnings ratio, denoted by Yt . First,
we estimate the standard Dickey–Fuller regression, which gives
Yt = 0.335 − 0.125Yt−1 + et ,
(0.128) (0.048)

(8.58)

4.0

3.5

3.0

2.5

2.0

1.5
1880

1900

1920

1940

1960

1980

2000

Figure 8.3 Annual log price/earnings ratio, 1871–2002
5

The data are available in the ﬁles PE.

TESTING FOR UNIT ROOTS

275

resulting in a DF test statistic of −2.569. As the appropriate critical value at the 5%
level is −2.88, this does not allow us to reject the null hypothesis of a ﬁrst unit root.
However, we need to be sure that we included sufﬁcient lags in this testing regression to
make the error term white noise. Thus, it is advisable to perform a range of augmented
Dickey–Fuller tests as well, implying that we add additional lags of Yt to the righthand side. Restricting attention to the test statistics, the results with up to 6 additional
lags are as follows:
DF
ADF (1)
−2.569 −3.000

ADF (2)
−2.487

ADF (3)
−2.503

ADF (4)
−1.778

ADF (5)
−1.627

ADF (6)
−1.825

With the exception of the ADF test with one lag, which implies a marginal rejection,
none of the (augmented) Dickey–Fuller tests reject the presence of a ﬁrst unit root
at the 5% level. Perhaps somewhat surprisingly, considering Figure 8.3, the statistical
tests do not allow us to reject the null hypothesis that the series is nonstationary.
If we impose a ﬁrst unit root, we can test for the presence of a second unit root.
This involves regressions of the form
2

Yt = δ + π Yt−1 + c1

2

Yt−1 + · · · + εt

and the null hypothesis corresponds to π = 0. The results are as follows:
DF
−10.588

ADF (1)
−9.109

ADF (2)
−7.310

ADF (3)
−7.481

ADF (4)
−6.431

ADF (5)
−5.315

ADF (6)
−4.210

For all of the tests, the null hypothesis of a second unit root is soundly rejected.
The graph of the ﬁrst differenced series is depicted in Figure 8.4, and also provides
no indication for the presence of a unit root; clearly, the series ﬂuctuates around a
long-run mean that is close to zero. Below, in Subsection 8.7.5, we shall consider the
problem of ﬁnding an appropriate ARMA model for Yt . Note that the change in the
log price/earnings ratio corresponds to the growth rate of the price/earnings ratio.
0.6
0.4
0.2
0.0
−0.2
−0.4
−0.6
1880

Figure 8.4

1900

1920

1940

1960

1980

2000

Annual change in log price/earnings ratio, 1872–2002

UNIVARIATE TIME SERIES MODELS

276

8.5

Illustration: Long-run Purchasing Power
Parity (Part 1)

To illustrate the discussion above, we shall in this section pay attention to an empirical example concerning prices in two countries and the exchange rate between these
countries. If two countries produce tradeable goods, and there are no impediments to
international trade, such as tariffs or transaction costs, then the law of one price should
hold, i.e.
St = Pt /Pt∗ ,
(8.59)
where St is the spot exchange rate (home currency price of a unit of foreign exchange),
Pt is the (aggregate) price in the domestic country, and Pt∗ is the price in the foreign
country. In logarithms, we can write
st = pt − pt∗ ,

(8.60)

(where lower case letters denote natural logarithms). Condition (8.60), which is referred
to as absolute purchasing power parity (absolute PPP), implies that an increase in
the home price level should result in an equiproportionate rise in the exchange rate.
Obviously this condition will never be satisﬁed in practice. Usually, PPP is seen as
determining the exchange rate in the long run. Below we shall analyse the question
whether (8.60) is ‘valid’ in the long run. A ﬁrst necessary step for that is an analysis
of the properties of the variables involved in (8.60).
Our empirical example concerns France and Italy over the period January 1981 until
June 1996 (T = 186).6 First we plot the two series for the log consumer price index
in Figure 8.5. Clearly, this ﬁgure indicates non-stationarity of the two series, while it
is also clear that the two series have a different growth rate. Formal unit root tests can
5.0
Italy
4.8
France

4.6
4.4
4.2
4.0
3.8
3.6
82

Figure 8.5
6

84

86

88

90

92

94

96

Log consumer price index France and Italy, 1981:1–1996:6

Data available in the ﬁles PPP.

ILLUSTRATION: LONG-RUN PURCHASING POWER PARITY (PART 1)

277

of course be obtained from regressions like (8.56) or (8.57). For pt∗ , the log of the
French consumer price index, we obtain the following results, including a constant but
no lagged differences in the model:
∗
pt∗ = 0.0694 − 0.0146pt−1
+ et .
(0.0042) (0.0009)

The Dickey–Fuller test statistic is −15.67, while the 5% critical value is −2.87,
suggesting that the null hypothesis of a unit root should be rejected at any reasonable
level of signiﬁcance. However, it is quite likely that the simple AR(1) model employed
in this regression is too restrictive. Some software packages (like MicroFit) have the
option of running a range of ADF tests simultaneously. This gives the results presented
in the ﬁrst two columns of Table 8.2. The critical values for the tests without trend are
−2.877 and −3.435 for those with trend.7
The results clearly show the danger of testing for a unit root in a too restrictive
model. Apparently, the 12th lag is important to include in the ADF regressions, which
is not surprising given that we have monthly data and that seasonal patterns in prices
are not uncommon. Thus, despite the fact that the majority of tests in the above table
suggest rejection, we will not reject the hypothesis of a unit root when we consider the
appropriate ADF test corresponding to 12 lags included in the regression. This choice
can also be defended by looking at the graphs, which clearly show some source of
nonstationarity.
For the log of the consumer price index in Italy, pt , we ﬁnd a rather similar set of
results, as shown in the last two columns of Table 8.2. The conclusion is the same:
we do not reject the null hypothesis that the log price series contains a unit root. For
the log of the exchange rate, st , measured as Liras per Franc, the Dickey–Fuller and
augmented Dickey–Fuller tests give the results in Table 8.3, where we only report the
Table 8.2

Unit root tests for log price index France and Italy
France (pt∗ )

Statistic
DF
ADF(1)
ADF(2)
ADF(3)
ADF(4)
ADF(5)
ADF(6)
ADF(7)
ADF(8)
ADF(9)
ADF(10)
ADF(11)
ADF(12)
7

Without trend
−15.670
−7.147
−7.003
−4.964
−5.118
−4.115
−3.019
−3.183
−2.878
−2.688
−2.655
−2.408
−1.763

With trend
−9.462
−6.272
−6.933
−5.294
−6.077
−5.382
−3.919
−4.146
−3.728
−3.451
−3.591
−3.691
−2.908

Italy (pt )
Without trend
−13.159
−6.378
−5.479
−4.407
−3.880
−3.692
−3.771
−3.260
−2.344
−2.039
−2.113
−1.687
−0.866

With trend
−8.403
−5.389
−5.131
−4.644
−4.289
−4.580
−5.474
−5.525
−4.529
−4.064
−3.742
−3.797
−2.997

The critical values change slightly from one row to the other. This is due to the change in the number of
observations that is available to estimate the ADF regressions.

UNIVARIATE TIME SERIES MODELS

278

Table 8.3 Unit root tests for log exchange
rate Italy–France
Statistic

Without trend

With trend

DF
ADF(1)
ADF(2)
ADF(3)
ADF(4)
ADF(5)
ADF(6)

−0.328
−0.361
−0.160
−0.291
−0.366
−0.463
−0.643

−1.900
−1.884
−1.925
−2.012
−2.026
−2.032
−2.262

ADF tests up to lag 6. The results here are quite clear. In none of the cases can we
reject the null hypothesis of a unit root.
If purchasing power parity between France and Italy holds in the long run, one can
expect that short-run deviations, st − (pt − pt∗ ), corresponding to the real exchange
rate, are limited and do not wander widely. In other words, one can expect st − (pt −
pt∗ ) to be stationary. A test for PPP can thus be based on the analysis of the log
real exchange rate rst ≡ st − (pt − pt∗ ). The series is plotted in Figure 8.6, while the
results for the augmented Dickey–Fuller tests for this variable are given in Table 8.4.
The results show that the null hypothesis of a unit root in rst (corresponding to
non-stationarity) cannot be rejected. Consequently, there is no evidence for PPP to
hold in this form. One reason why we may not be able to reject the null hypothesis
is simply that our sample contains insufﬁcient information, that is: our sample is too
short and standard errors are simply too high to reject the unit root hypothesis. This is
a problem often found in tests for purchasing power parity. A critical survey of this
literature can be found in Froot and Rogoff (1996). In the next chapter, we shall also
analyse whether some weaker form of PPP holds.
5.8

5.7

5.6

5.5

5.4

5.3
82

84

86

88

90

92

94

96

Figure 8.6 Log real exchange rate Italy–France, 1981:1–1996:6

ESTIMATION OF ARMA MODELS

279

Table 8.4 Unit root tests for log real exchange
rate Italy–France

8.6

Statistic

Without trend

With trend

DF
ADF(1)
ADF(2)
ADF(3)
ADF(4)
ADF(5)
ADF(6)

−1.930
−1.874
−1.930
−1.987
−1.942
−1.966
−2.287

−1.942
−1.892
−1.961
−2.022
−1.981
−2.005
−2.326

Estimation of ARMA Models

Suppose that we know that the data series Y1 , Y2 , . . . , YT is generated by an ARMA
process of order p, q. Depending upon the speciﬁcation of the model, and the distributional assumptions we are willing to make, we can estimate the unknown parameters
by ordinary or nonlinear least squares, or by maximum likelihood.
8.6.1 Least Squares

The least squares approach chooses the model parameters such that the residual sum
of squares is minimal. This is particularly easy for models in autoregressive form.
Consider the AR(p) model
Yt = δ + θ1 Yt−1 + θ2 Yt−2 + · · · + θp Yt−p + εt ,

(8.61)

where εt is a white noise error term that is uncorrelated with anything dated t − 1 or
before. Consequently, we have that
E{Yt−j εt } = 0

for j = 1, 2, 3, . . . , p,

that is, error terms and explanatory variables are contemporaneously uncorrelated and
OLS applied to (8.61) provides consistent estimators. Estimation of an autoregressive model is thus no different than that of a linear regression model with a lagged
dependent variable.
For moving average models, estimation is somewhat more complicated. Suppose
that we have an MA(1) model
Yt = µ + εt + αεt−1 .
Because εt−1 is not observed, we cannot apply regression techniques here. In theory,
ordinary least squares would minimize
S(α, µ) =

T

t=2

(Yt − µ − αεt−1 )2 .

UNIVARIATE TIME SERIES MODELS

280

A possible solution arises if we write εt−1 in this expression as a function of observed
Yt s. This is possible only if the MA polynomial is invertible. In this case we can
use that
∞

(−α)j (Yt−j −1 − µ)
εt−1 =
j =0

(see above) and write
S(α, µ) =

T



Y t − µ − α

t=2

∞


2
(−α)j (Yt−j −1 − µ) .

j =0

In practice, Yt is not observed for t = 0, −1, . . . , so we have to cut off the inﬁnite
sum in this expression to obtain an approximate sum of squares
S̃(α, µ) =

T



Y t − µ − α

t=2

t−2


2
(−α)j (Yt−j −1 − µ) .

(8.62)

j =0

Because, asymptotically, if T goes to inﬁnity the difference between S(α, µ) and
S̃(α, µ) disappears minimizing (8.62) with respect to α and µ gives consistent
estimators α̂ and µ̂. Unfortunately, (8.62) is a high order polynomial in α
and thus has very many local minima. Therefore, numerically minimizing (8.62)
is complicated. However, as we know that −1 < α < 1, a grid search (e.g.
−0.99, −0.98, −0.97, . . . , 0.98, 0.99) can be performed. The resulting nonlinear least
squares estimators for α and µ are consistent and asymptotically normal.
8.6.2 Maximum Likelihood

An alternative estimator for ARMA models is provided by maximum likelihood. This
requires that an assumption is made about the distribution of εt , most commonly
normality. Although the normality assumption is strong, the ML estimators are very
often consistent even in cases where εt is not normal. Conditional upon an initial value
the loglikelihood function can be written as
T

log L(α, θ, µ, σ 2 ) = −

T −1
1 2 2
ε /σ ,
log(2πσ 2 ) −
2
2 t=2 t

where εt is a function of the coefﬁcients α, θ and µ, yt and its history. For an AR(1)
model it is εt = yt − θyt−1 , where yt = Yt − µ, and for the MA(1) model we have
εt = yt − α

t−2

j =0

(−α)j yt−j −1 =

t−1


(−α)j yt−j .

j =0

Both of the implied loglikelihood functions are conditional upon an initial value. For
the AR(1) case y1 is treated as given, while for the MA(1) case the initial condition
is ε0 = 0. The resulting estimators are therefore referred to as conditional maximum

CHOOSING A MODEL

281

likelihood estimators. The conditional ML estimators for α, θ and µ are easily seen
to be identical to the least squares estimators.
The exact maximum likelihood estimator combines the conditional likelihood with
the likelihood from the initial observations. In the AR(1) case, for example, the following term is added to the loglikelihood:
y12
1
1
1
− log(2π) − log[σ 2 /(1 − θ 2 )] −
,
2
2
2 σ 2 /(1 − θ 2 )
which follows from the fact that the marginal density of y1 is normal with mean zero
and variance σ 2 /(1 − θ 2 ). For a moving average process the exact likelihood function
is somewhat more complex. If T is large the way we treat the initial values has
negligible impact, so that the conditional and exact maximum likelihood estimators are
asymptotically equivalent in cases where the AR and MA polynomials are invertible.
More details can be found in Hamilton (1994, Chapter 5).
It will be clear from the results above that estimating autoregressive models is simpler
than estimating moving average models. Estimating ARMA models, which combine
an autoregressive part with a moving average part, closely follows the lines of ML
estimation of the MA parameters. As any (invertible) ARMA model can be approximated by an autoregressive model of inﬁnite order, it has become more and more
common practice to use autoregressive speciﬁcations instead of MA or ARMA ones,
and allowing for a sufﬁcient number of lags. Particularly if the number of observations
is not too small this approach may work pretty well in practice. Of course, an MA
representation of the same process may be more parsimonious. Another advantage of
autoregressive models is that they are easily generalized to multivariate time series,
where one wants to model a set of economic variables jointly. This leads to so-called
vector autoregressive models (VAR models), which are discussed in the next chapter.

8.7

Choosing a Model

Most of the time there are no economic reasons to choose a particular speciﬁcation
of the model. Consequently, to a large extent the data will determine which time
series model is appropriate. Before estimating any model, it is common to estimate
autocorrelation and partial autocorrelation coefﬁcients directly from the data. Often
this gives some idea about which model might be appropriate. After one or more
models are estimated, their quality can be judged by checking whether the residuals
are more or less white noise, and by comparing them with alternative speciﬁcations.
These comparisons can be based on statistical signiﬁcance tests or the use of particular
model selection criteria.
8.7.1 The Autocorrelation Function
The autocorrelation function (ACF) describes the correlation between Yt and its lag Yt−k
as a function of k. Recall that the k-th order autocorrelation coefﬁcient is deﬁned as

ρk =

cov{Yt , Yt−k }
γ
= k,
V {Yt }
γ0

noting that cov{Yt , Yt−k } = E{yt yt−k }.

UNIVARIATE TIME SERIES MODELS

282

For the MA(1) model we have seen that
ρ1 =

α
,
1 + α2

ρ2 = 0,

ρ3 = 0, . . . ,

that is, only the ﬁrst autocorrelation coefﬁcient is nonzero. For the MA(2) model
yt = εt + α1 εt−1 + α2 εt−2
we have

E{yt2 } = (1 + α12 + α22 )σ 2
E{yt yt−1 } = (α1 + α1 α2 )σ 2
E{yt yt−2 } = α2 σ 2
E{yt yt−k } = 0,

k = 3, 4, 5, . . .

It follows directly from this that the ACF is zero after two lags. This is a general result
for moving average models: for an MA(q) model the ACF is zero after q lags.
The sample autocorrelation function gives the estimated autocorrelation coefﬁcients
as a function of k. The coefﬁcient ρk can be estimated by8
1 T
(Y − Ȳ )(Yt−k − Ȳ )
T
−
k t=k+1 t
ρ̂k =
,
1 T
2
(Y
−
Ȳ
)
t=1 t
T

(8.63)


where Ȳ = (1/T ) Tt=1 Yt denotes the sample average. That is, the population covariances in the ratio are replaced by their sample estimates. Alternatively, it can be
estimated by regressing Yt upon Yt−k and a constant, which will give a slightly different estimator, because the summation in numerator and denominator will be over the
same set of observations. Of course, it will usually not be the case that ρ̂k is zero for
an MA model of order q < k. But we can use ρ̂k to test the hypothesis that ρk = 0.
To do this, we can use the result that asymptotically
√
T (ρ̂k − ρk ) → N(0, vk ),
where
vk = 1 + 2ρ12 + 2ρ22 + · · · + 2ρq2

if q < k.

So, to test the hypothesis that the true model is MA(0) versus√the alternative that it
is MA(1), we can test ρ1 = 0 by comparing the test statistic T ρ̂1 with the critical
values of a standard normal distribution. Testing MA(k − 1) versus MA(k) is done by
testing ρk = 0 and comparing the test statistic
√

8

ρ̂k
T
2
1 + 2ρ̂12 + · · · + 2ρ̂k−1

(8.64)

Alternative consistent estimators are possible that have slightly different degrees of freedom corrections.

CHOOSING A MODEL

283

with critical values from the standard normal distribution. Typically, two-standard error
2
bounds for ρ̂k based on the estimated variance 1 + 2ρ̂12 + · · · + 2ρ̂k−1
are graphically
displayed in the plot of the sample autocorrelation function (see the example in Subsection 8.7.5 below). The order of a moving average model can in this way be determined
from an inspection of the sample ACF. At least it will give us a reasonable value for
q to start with and diagnostic checking, as discussed below, should indicate whether
it is appropriate or not.
For autoregressive models the ACF is less helpful. For the AR(1) model we have seen
that the autocorrelation coefﬁcients do not cut off at a ﬁnite lag length. Instead, they go
to zero exponentially corresponding to ρk = θ k . For higher order autoregressive models,
the autocorrelation function is more complex. Consider the general AR(2) model
Yt = δ + θ1 Yt−1 + θ2 Yt−2 + εt .
To derive the autocovariances, it is convenient to take the covariance of both sides
with Yt−k to obtain
cov{Yt , Yt−k } = θ1 cov{Yt−1 , Yt−k } + θ2 cov{Yt−2 , Yt−k } + cov{εt , Yt−k }.
For k = 0, 1, 2, this gives

γ0 = θ1 γ1 + θ2 γ2 + σ 2
γ1 = θ1 γ0 + θ2 γ1
γ2 = θ1 γ1 + θ2 γ0 .

This set of equations, known as the Yule–Walker equations, can be solved for the
autocovariances γ0 , γ1 and γ2 as a function of the model parameters θ1 , θ2 and σ 2 . The
higher order covariances can be determined recursively from
γk = θ1 γk−1 + θ2 γk−2

(k = 2, 3, . . .),

which corresponds to a second order differential equation. Depending on θ1 and θ2 the
patterns of the ACF can be very different. Consequently, in general only a real expert
may be able to identify an AR(2) process from the ACF pattern, let alone from the
sample ACF pattern. An alternative source of information that is helpful is provided
by the partial autocorrelation function, discussed in the next subsection.
8.7.2 The Partial Autocorrelation Function

We now deﬁne the k-th order sample partial autocorrelation coefﬁcient as the estimate
for θk in an AR(k) model. We denote this by θ̂kk . So estimating
Yt = δ + θ1 Yt−1 + εt
gives us θ̂11 , while estimating
Yt = δ + θ1 Yt−1 + θ2 Yt−2 + εt

284

UNIVARIATE TIME SERIES MODELS

yields θ̂22 , the estimated coefﬁcient for Yt−2 in the AR(2) model. The partial autocorrelation θ̂kk measures the additional correlation between Yt and Yt−k after adjustments
have been made for the intermediate values Yt−1 , . . . , Yt−k+1 .
Obviously, if the true model is an AR(p) process then estimating an AR(k) model
by OLS gives consistent estimators for the model parameters if k ≥ p. Consequently,
we have
plim θ̂kk = 0 if k > p.
(8.65)
Moreover, it can be shown that the asymptotic distribution is standard normal, i.e.
√
T (θ̂kk − 0) → N(0, 1) if k > p.
(8.66)
Consequently, the partial autocorrelation coefﬁcients (or the partial autocorrelation
function (PACF)) can be used to determine the order of an AR process. Testing an
AR(k − 1) model versus an AR(k) model implies testing the null hypothesis that θkk =
0. Under the null hypothesis that √
the model is AR(k − 1) the approximate
standard
√
error of θ̂kk based on (8.66) is 1/ T , so that θkk = 0 is rejected if | T θ̂kk | > 1.96.
This way one can look at the PACF and test for which lags the partial autocorrelation
coefﬁcient differs from zero. For a genuine AR(p) model the partial autocorrelations
will be close to zero after the p-th lag.
For moving average models it can be shown that the partial autocorrelations do not
have a cut off point but tail off to zero, just like the autocorrelations in an autoregressive
model. In summary, an AR(p) process is described by:
1. an ACF that is inﬁnite in extent (it tails off).
2. a PACF that is (close to) zero for lags larger than p.
For an MA(q) process we have that:
1. an ACF that is (close to) zero for lags larger than q.
2. a PACF that is inﬁnite in extent (it tails off).
In the absence of any of these two situations, a combined ARMA model may provide
a parsimonious representation of the data.
8.7.3 Diagnostic Checking

As a last step in the model-building cycle some checks on the model adequacy are
required. Possibilities are doing a residual analysis and overﬁtting the speciﬁed model.
For example, if an ARMA(p, q) model is chosen (on the basis of the sample ACF and
PACF), we could also estimate an ARMA(p + 1, q) and an ARMA(p, q + 1) model
and test the signiﬁcance of the additional parameters.
A residual analysis is usually based on the fact that the residuals of an adequate
model should be approximately white noise. A plot of the residuals can be a useful tool
in checking for outliers. Moreover, the estimated residual autocorrelations are usually
examined. Recall that for a white noise series the autocorrelations are zero. Therefore
the signiﬁcance of the residual autocorrelations is often checked by comparing with

CHOOSING A MODEL

285

√
approximate two standard error bounds ±2/ T . To check the overall acceptability of
the residual autocorrelations, the Ljung–Box (1978) portmanteau test statistic,
QK = T (T + 2)

K

k=1

1
r2
T −k k

(8.67)

is often used. Here, the rk are the estimated autocorrelation coefﬁcients of the residuals
ε̂t and K is a number chosen by the researcher. Values of Q for different K may be
computed in a residual analysis. For an ARMA(p, q) process (for Yt ) the statistic QK is
approximately Chi-squared distributed with K − p − q − 1 degrees of freedom (under
the null hypothesis that the ARMA(p, q) is correctly speciﬁed). If a model is rejected at
this stage, the model-building cycle has to be repeated. Note that this test only makes
sense if K > p + q + 1.
8.7.4 Criteria for Model Selection

Because economic theory does not provide any guidance to the appropriate choice of
model, some additional criteria can be used to choose from alternative models that
are acceptable from a statistical point of view. As a more general model will always
provide a better ﬁt (within the sample) than a restricted version of it, all such criteria
provide a trade-off between goodness-of-ﬁt and the number of parameters used to
obtain that ﬁt. For example, if an MA(2) model would provide the same ﬁt as an
AR(10) model, we would prefer the ﬁrst as it is more parsimonious. As discussed in
Chapter 3, a well known criterion is Akaike’s Information Criterion (AIC ) (Akaike,
1973). In the current context it is given by
AIC = log σ̂ 2 + 2

p+q +1
,
T

(8.68)

where σ̂ 2 is the estimated variance of εt . An alternative is Schwarz’s Bayesian Information Criterion (SC, BIC or SBC ), proposed by Schwarz (1978), which is given by
BIC = log σ̂ 2 +

p+q +1
log T .
T

(8.69)

Both criteria are likelihood-based and represent a different trade-off between ‘ﬁt’, as
measured by the loglikelihood value, and ‘parsimony’, as measured by the number
of free parameters, p + q + 1 (assuming that a constant is included in the model).
Usually, the model with the smallest AIC or BIC value is preferred, although one can
choose to deviate from this if the differences in criterion values are small for a subset
of the models.
While the two criteria differ in their trade-off between ﬁt and parsimony, the BIC
criterion can be preferred because it has the property that it will almost surely select
the true model, if T → ∞, provided that the true model is in the class of ARMA(p,
q) models for relatively small values of p and q. The AIC criterion tends to result
asymptotically in overparametrized models (see Hannan, 1980).

UNIVARIATE TIME SERIES MODELS

286

8.7.5 Illustration: Modelling the Price/Earnings Ratio

In Subsection 8.4.4, we saw that the null hypothesis of a unit root in the log of
the Standard and Poor’s Price to Earnings ratio could not be rejected. Therefore, we
shall in this section try to model the ﬁrst differenced series, the relative change in the
price/earnings ratio. The sample autocorrelation and partial autocorrelation function are
presented in Figures 8.7 and 8.8, respectively. At the 5% level, the autocorrelations
only have a signiﬁcant peak at lag 4, while the partial autocorrelations are statistically
signiﬁcant at lags 2 and 4.
From an inspection of the sample ACF and PACF, there is no obvious model that
comes to mind. Because it can be argued that both ACF and PACF are zero after
lag 4, we initially consider the estimation of an AR(4) and an MA(4) model. These
Sample ACF

0.6
0.4
0.2
0
−0.2
−0.4
−0.6

1

2

3

4

5

6

7

8
k

9

10 11

12 13 14

15

Figure 8.7 Sample autocorrelation function of log(P/E)

Sample PACF

0.6
0.4
0.2
0
−0.2
−0.4
−0.6
1

2

3

4

5

6

7

8
k

9

10 11

12 13 14

Figure 8.8 Sample partial autocorrelation function of log(P/E)

15

CHOOSING A MODEL

287

two models are estimated by least squares, including an intercept term. For the AR(4)
model we obtain
Yt = 0.009 + 0.060 Yt−1 − 0.203 Yt−2 − 0.023 Yt−3
(0.012) (0.088)
(0.089)
(0.089)
−0.212 Yt−4 + ε̂t ,
(0.089)

σ̂ = 0.179

Q6 = 0.279 (p = 0.598),
AIC = −3.437,

Q12 = 3.975 (p = 0.783)
BIC = −3.246,

while estimating the MA(4) model produces
Yt = 0.008 + 0.047 ε̂t−1 − 0.189 ε̂t−2 − 0.041 ε̂t−3 − 0.150 ε̂t−4 +ε̂t , σ̂ = 0.178
(0.010) (0.088)
(0.089)
(0.089)
(0.089)
Q6 = 1.610 (p = 0.204),
AIC = −3.452,

Q12 = 5.285 (p = 0.625)
BIC = −3.266.

For neither of the speciﬁcations, we can reject the null hypothesis that the residuals
correspond to a white noise process. The Ljung–Box statistics for the ﬁrst K = 6
and 12 residual autocorrelations do not reject. With the same number of parameters,
the moving average speciﬁcation provides a slightly better ﬁt than the autoregressive
one, which is conﬁrmed by the values of the AIC and BIC criteria. Note that both
speciﬁcations contain at least two insigniﬁcant lags.
It is interesting to see whether a more parsimonious model could provide almost
the same ﬁt (but with fewer parameters). Because the ﬁrst and third order (partial)
autocorrelation coefﬁcients of Yt are very small, we consider fourth order AR and
MA speciﬁcations but with the ﬁrst and third lag excluded. This leads to the following
autoregressive model:
Yt = 0.009 − 0.202 Yt−2 − 0.217
(0.011) (0.088)
(0.088)
Q6 = 0.714 (p = 0.870),
AIC = −3.448,

Yt−4 + ε̂t ,

σ̂ = 0.178

Q12 = 4.240 (p = 0.894)
BIC = −3.337,

while the moving average model is given by
Yt = 0.008 −0.183 ε̂t−2 − 0.161 ε̂t−4 +ε̂t ,
(0.010) (0.088)
(0.088)
Q6 = 2.260 (p = 0.520),
AIC = −3.464,

σ̂ = 0.177

Q12 = 5.683 (p = 0.771)
BIC = −3.353.

UNIVARIATE TIME SERIES MODELS

288

On the basis of the AIC and BIC criteria, both speciﬁcations are preferred to their
more general counterparts that include the ﬁrst and third lag. In the autoregressive
model, both coefﬁcients are signiﬁcantly different from zero, while the coefﬁcient for
the fourth lag in the moving average model is insigniﬁcant (at the 5% level). Let us
consider two further simpliﬁcations that only include the second lag. This produces
the following results:
Yt = 0.009 − 0.168 Yt−2 + ε̂t ,
(0.014) (0.088)
Q6 = 8.039 (p = 0.090),
AIC = −3.422,

σ̂ = 0.181

Q12 = 12.344 (p = 0.263)
BIC = −3.347

and
Yt = 0.008 − 0.254ε̂t−2 + ε̂t ,
(0.012) (0.086)
Q6 = 5.244 (p = 0.263),
AIC = −3.450,

σ̂ = 0.178

Q12 = 9.429 (p = 0.491)
BIC = −3.373.

From these last two models, the moving average model provides a better ﬁt. None of the
considered models is rejected at the 5% level by the Ljung–Box tests, although for the
AR(2) model a rejection is found at the 10% level. (This may not be surprising, given
the signiﬁcance of the partial autocorrelation function at lag 4.) The AIC criterion
prefers the restricted MA(4) model (with lags 2 and 4 included), whereas the BIC
criterion prefers the more parsimonious MA(2) model. Because the differences in model
ﬁt are small, the MA(2) speciﬁcation appears to provide the most appropriate model
for the annual change in the log price/earnings ratio.

8.8

Predicting with ARMA Models

A main goal of building a time series model is predicting the future path of economic
variables. It can be noted that ARMA models usually perform quite well in this respect
and often outperform more complicated structural models. Of course, ARMA models
do not provide any economic insight in one’s predictions and are unable to forecast
under alternative economic scenarios. In this section, we discuss the optimal predictor,
which is simply the conditional expectation of a future value given the available information, and how it can be derived in ARMA models. Furthermore, we pay attention
to forecast accuracy.
8.8.1 The Optimal Predictor

Suppose we are at time T and are interested in predicting YT +h , the value of Yt h
periods ahead. A predictor for YT +h will be based on an information set, denoted IT ,
that contains the information that is available and potentially used at the time of making
the forecast. Ideally it contains all the information that is observed and known at time

PREDICTING WITH ARMA MODELS

289

T . In univariate time series modelling we will usually assume that the information set
at any point t in time contains the value of Yt and all its lags. Thus we have
IT = {Y−∞ , . . . , YT −1 , YT }.

(8.70)

In general, the predictor ŶT +h|T (the predictor for YT +h as constructed at time T ) is a
function of (variables in) the information set IT . Our criterion for choosing a predictor
from the many possible ones is to minimize the expected quadratic prediction error
E{(YT +h − ŶT +h|T )2 |IT },

(8.71)

where E{.|IT } denotes the conditional expectation given the information set IT . It is
not very hard to show that the best predictor for YT +h given the information set at
time T is the conditional expectation of YT +h given the information IT . We denote
this optimal predictor as
YT +h|T ≡ E{YT +h |IT }.
(8.72)
Because the optimal predictor is a conditional expectation, it satisﬁes the usual
properties of expectation operators. Most importantly, the conditional expectation of
a sum is the sum of conditional expectations. Further, it holds that the conditional
expectation of YT +h given an information set I T , where I T is a subset of IT , is
at best as good as YT +h|T based on IT . In line with our intuition, it holds that the
more information one uses to determine the predictor (the larger IT is), the better
the predictor is. For example, E{YT +h |YT , YT −1 , YT −2 , . . .} will usually be a better
predictor than E{YT +h |YT } or E{YT +h } (an empty information set).
To simplify things we shall, in the sequel, assume that the parameters in the ARMA
model for Yt are known. In practice one would simply replace the unknown parameters
by their consistent estimates. Now, how do we determine these conditional expectations when Yt follows an ARMA process? To simplify the notation, we shall consider
forecasting of yT +h , noting that YT +h|T = µ + yT +h|T . As a ﬁrst example consider an
AR(1) process where it holds by assumption that
yT +1 = θyT + εT +1 .
Consequently,
yT +1|T = E{yT +1 |yT , yT −1 , . . .} = θyT + E{εT +1 |yT , yT −1 , . . .} = θyT ,

(8.73)

where the latter equality follows from the fact that the white noise process is unpredictable. To predict two periods ahead (h = 2), we write
yT +2 = θyT +1 + εT +2 ,
from which it follows that
E{yT +2 |yT , yT −1 , . . .} = θ E{yT +1 |yT , yT −1 , . . .} = θ 2 yT .

(8.74)

UNIVARIATE TIME SERIES MODELS

290

In general, we obtain yT +h|T = θ h yT . Thus, the last observed value yT contains all
the information to determine the predictor for any future value. When h is large the
predictor of yT +h converges to 0 (the unconditional expectation of yt ), provided (of
course) that |θ | < 1. With a nonzero mean, the best predictor for YT +h is directly
obtained as µ + yT +h|T = µ + θ h (YT − µ). Note that this differs from θ h YT .
As a second example, consider an MA(1) process, where
yt = εt + αεt−1 .
Then we have
E{yT +1 |yT , yT −1 , . . .} = αE{εT |yT , yT −1 , . . .} = αεT ,
where, implicitly, we assumed that εT is observed (contained in IT ). This is an innocent
assumption provided the MA process is invertible. In that case we can write
εT =

∞


(−α)j yT −j ,

j =0

and determine the one-period ahead predictor as
yT +1|T

∞

=α
(−α)j yT −j .

(8.75)

j =0

Predicting two periods ahead gives
yT +2|T = E{εT +2 |yT , yT −1 , . . .} + αE{εT +1 |yT , yT −1 , . . .} = 0,

(8.76)

which shows that the MA(1) model is uninformative for forecasting two periods ahead:
the best predictor is simply the (unconditional) expected value of yt , normalized at 0.
This also follows from the autocorrelation function of the process, because the ACF
is zero after one lag. That is, the ‘memory’ of the process is only one period.
For the general ARMA(p, q) model,
yt = θ1 yt−1 + · · · + θp yt−p + εt + α1 εt−1 + · · · + αq εt−q
we can derive the following recursive formula to determine the optimal predictors:
yT +h|T = θ1 yT +h−1|T + · · · + θp yT +h−p|T + εT +h|T + α1 εT +h−1|T + · · · + αq εT +h−q|T ,
(8.77)
where εT +K|T is the optimal predictor for εT +K at time T , and
yT +k|T = yT +k
εT +k|T = 0
εT +k|T = εT +k

if k ≤ 0
if k > 0
if k ≤ 0,

PREDICTING WITH ARMA MODELS

291

where the latter innovation can be solved from the autoregressive representation of
the model. For this we have used the fact that the process is stationary and invertible,
in which case the information set {yT , yT −1 , . . .} is equivalent to {εT , εT −1 , . . .}. That
is, if all εt s are known from −∞ to T , then all yt s are known from −∞ to T and
vice versa.
To illustrate this, consider an ARMA(1, 1) model, where
yt = θyt−1 + εt + αεt−1 ,
such that
yT +1|T = θyT |T + εT +1|T + αεT |T = θyT + αεT .
Using that (assuming invertibility)
yt − θyt−1 = (1 + αL)εt
can be rewritten as
εt = (1 + αL)−1 (yt − θyt−1 ) =

∞


(−α)j Lj (yt − θyt−1 ),

j =0

we can write for the one-period ahead predictor
yT +1|T = θyT + α

∞


(−α)j (yT −j − θyT −j −1 ).

(8.78)

j =0

Predicting two periods ahead gives
yT +2|T = θyT +1|T + εT +2|T + αεT +1|T = θyT +1|T .

(8.79)

Note that this does not equal θ 2 yT .
8.8.2 Prediction Accuracy

In addition to the prediction itself, it is important (sometimes even more important)
to know how accurate this prediction is. To judge forecasting precision, we deﬁne
the prediction error as YT +h − YT +h|T = yT +h − yT +h|T and the expected quadratic
prediction error as
Ch ≡ E{(yT +h − yT +h|T )2 } = V {yT +h |IT },

(8.80)

where the latter step follows from the fact that yT +h|T = E{yT +h |IT }. Determining Ch ,
corresponding to the variance of the h-period ahead prediction error, is relatively easy
with the moving average representation.

UNIVARIATE TIME SERIES MODELS

292

To start with the simplest case, consider an MA(1) model. Then we have
C1 = V {yT +1 |yT , yT −1 , . . .} = V {εT +1 + αεT |εT , εT −1 , . . .} = V {εT +1 } = σ 2 .
Alternatively, we explicitly solve for the predictor, which is yT +1|T = αεT and determine the variance of yT +1 − yT +1|T = εT +1 , which gives the same result. For the
two-period ahead predictor we have
C2 = V {yT +2 |yT , yT −1 , . . .} = V {εT +2 + αεT +1 |εT , εT −1 , . . .} = (1 + α 2 )σ 2 .
As expected, the accuracy of the prediction decreases if we predict further into the
future. It will not, however, increase any further if h is increased beyond 2. This
becomes clear if we compare the expected quadratic prediction error with that of a
simple unconditional predictor,
ŷT +h|T = E{yT +h } = 0
(empty information set). For this predictor we have
Ch = E{(yT +h − 0)2 } = V {yT +h } = (1 + α 2 )σ 2 .
Consequently, this gives an upper bound on the inaccuracy of the predictors. The
MA(1) model thus gives more efﬁcient predictors only if one predicts one period
ahead. More general ARMA models, however, will yield efﬁciency gains also in further
ahead predictors.
Suppose the general model is ARMA(p, q), which we write as an MA(∞) model,
with αj coefﬁcients to be determined,
yt =

∞


αj εt−j

with α0 ≡ 1.

j =0

The h-period ahead predictor (in terms of εt s) is given by
yT +h|T = E{yT +h |yT , yT −1 , . . .} =

∞


αj E{εT +h−j |εT , εT −1 , . . .} =

yT +h − yT +h|T =

αj εT +h−j ,

j =h

j =0

such that

∞


h−1


αj εT +h−j .

j =0

Consequently, we have
E{(yT +h − yT +h|T ) } = σ
2

2

h−1

j =0

αj2 .

(8.81)

ILLUSTRATION: THE EXPECTATIONS THEORY OF THE TERM STRUCTURE

293

This shows how the variances of the forecast errors can easily be determined from
the coefﬁcients in the moving average representation of the model. Recall that for the
computation of the predictor, the autoregressive representation was most convenient.
As an illustration, consider the AR(1) model where αj = θ j . The expected quadratic
prediction errors are given by
C1 = σ 2 ,

C2 = σ 2 (1 + θ 2 ),

C3 = σ 2 (1 + θ 2 + θ 4 ),

etc. For h going to inﬁnity, we have C∞ = σ 2 (1 + θ 2 + θ 4 + · · ·) = σ 2 /(1 − θ 2 ),
which is the unconditional variance of yt and therefore the expected quadratic prediction error of a constant predictor ŷT +h|T = E{yT +h } = 0. Consequently, the informational value contained in an AR(1) process slowly decays over time. In the long run
the predictor equals the unconditional predictor, being the mean of the yt series (as is
the case in all stationary time series models). Note that for a random walk, with θ = 1,
the forecast error variance increases linearly with the forecast horizon.
In practical cases, the parameters in ARMA models will be unknown and we replace
them by their estimated values. This introduces additional uncertainty in the predictors.
Usually, however, this uncertainty is ignored. The motivation is that the additional
variance that arises because of estimation error disappears asymptotically when the
sample size T goes to inﬁnity. In practice, the increase in the forecast error variance
if one would take it into account is usually fairly small.

8.9

Illustration: The Expectations Theory of the Term
Structure

Quite often, building a time series model is not a goal in itself, but a necessary ingredient in an economic analysis. To illustrate this, we shall in this section pay attention
to the term structure of interest rates. The term structure has attracted considerable
attention in both the macro-economics and the ﬁnance literature (see, for example,
Pagan, Hall and Martin, 1996) and the expectations hypothesis plays a central role in
many of these studies.
To introduce the problem, we consider an n-period discount bond, which is simply
a claim to one dollar paid to you n periods from today. The (market) price at time t
(today) of this discount bond is denoted as pnt . The implied interest rate rnt can then
be solved from
1
pnt =
.
(8.82)
(1 + rnt )n
The yield curve describes rnt as a function of its maturity n, and may vary from one
period t to the other. This depicts the term structure of interest rates. Models for the
term structure try to simultaneously model how the different interest rates are linked
and how the yield curve moves over time.
The pure expectations hypothesis, in a linearized form, can be written as
n−1

rnt =

1
E{r1,t+h |It },
n h=0

(8.83)

UNIVARIATE TIME SERIES MODELS

294

where It denotes the information set containing all information available at time t.
This says that the long term interest rate is the average of the expected short term rates
over the same interval. The left-hand side of this can be interpreted as the certain yield
of an n-period investment, while the right-hand side corresponds to the expected 9 yield
from investing in one-period bonds over an n-period horizon. Thus, expected returns
on bonds of different maturities are assumed to be equal.
The expectations hypothesis, in a more general form, allows for risk premia by
assuming that expected returns on different bonds can differ by constants, which can
depend on maturity but not on time. This extends (8.83) to
n−1

1
rnt =
E{r1,t+h |It } +
n h=0

n,

(8.84)

where n denotes a risk or term premium that varies with maturity n. Instead of testing
the expectations hypothesis in this form, which is the subject of many studies (see
Campbell and Shiller, 1991), we shall look at a simple implementation of (8.84). Given
that the term premia are constant, we can complete the model by making assumptions
about the relevant information set It and the time series process of the one-period
interest rate.
Let us assume, for simplicity, that
It = {r1t , r1,t−1 , r1,t−2 , . . . .},

such that the relevant information set contains the current and lagged short interest
rates only. If r1t can be described by an AR(1) process,
r1t − µ = θ (r1,t−1 − µ) + εt ,
with 0 < θ ≤ 1, the optimal s-period ahead predictor (see (8.74)) is given by
E{r1,t+h |It } = µ + θ h (r1t − µ).
Substituting this into (8.84) results in
n−1

rn,t =

1
[µ + θ h (r1t − µ)] +
n h=0

= µ + ξn (r1t − µ) +

n

n,

(8.85)

where, for 0 < θ < 1,
n−1

ξn =

1 h
1 1 − θn
θ =
< ξn−1 < 1,
n h=0
n 1−θ

(8.86)

while for θ = 1 we have ξn = 1 for each maturity n.
9

We impose rational expectations, which means that economic agents have expectations that correspond
to mathematical expectations, conditional upon some information set.

ILLUSTRATION: THE EXPECTATIONS THEORY OF THE TERM STRUCTURE

295

The rather simple model of the term structure in (8.85) implies that long rates depend
linearly on short rates and that short rate changes have less impact on longer rates than
shorter rates since ξn is decreasing with n if 0 < θ < 1. Note, for example, that
V {rnt } = ξn2 V {r1t },

(8.87)

which, with 0 < θ < 1, implies that short rates are more volatile than long rates. The
result in (8.85) also implies that there is just one factor which drives interest rates at
any maturity and thus one factor which shifts the term structure.
If all risk premia are zero ( n = 0) an inverted yield curve (with short rates
exceeding long rates) occurs if the short rate is above its mean µ, which – when
the distribution of εt is symmetric around zero (e.g. normal) – happens in 50% of the
cases. The reason is that when the short rate is below its average, it is expected to
increase to its average again, which increases the long rates. In practice, we see inverted
yield curves in less than 50% of the periods. For the United States,10 for example, we
displayed the one-month and the 5-year bond yields from January 1970 to February
1991 in Figure 8.9 (T = 254). Usually, the long rate is above the short rate, but there
are a few periods of inversion where this is not the case, for example from June 1973
to March 1974.
Clearly the time series properties of the short-term interest rate are important for
the cross-sectional relationships between the interest rates at different maturities. If the
short rate follows an AR(1) process we obtain the fairly simple expression in (8.85),
for which we can note that the values of ξn are very sensitive to the precise value
18
16
14
12
10
8
6
4
2
70

72

74

76

78

80

One-month rate

Figure 8.9
10

82

84

86

88

90

5-year rate

One-month and 5-year interest rates (in %), 1970:1–1991:2

The data used in this section are taken from the McCulloch and Kwon data set (see McCulloch and
Kwon, 1993). They are available in the ﬁles IRATES.

UNIVARIATE TIME SERIES MODELS

296

of θ , particularly for large maturities, if θ is close to unity. For more general time
series processes, we obtain similar expressions but the result will not just involve the
current short rate r1t . Because the optimal predictor for an AR(2) model, for example,
depends upon the two last observations, an AR(2) process for the short rate would give
an expression similar to (8.85) that involves r1t and r1,t−1 .
A debatable issue is that of stationarity. In many cases, the presence of a unit root in
the short term interest rate cannot be rejected statistically, but this does not necessarily
mean that we have to accept the unit root hypothesis. Economically, it seems hard to
defend nonstationarity of interest rates, although their persistence is known to be high.
That is, even with stationarity it takes a very long time for the series to go back to its
mean. Different authors make different judgements on this question and you will ﬁnd
empirical studies on the term structure of interest rates that choose either way. Let us
ﬁrst estimate an AR(1) model for the one-month interest rate. Estimation by OLS gives
(standard errors in parentheses):
r1t = 0.350 + 0.951 r1,t−1 + et , σ̂ = 0.820.

(8.88)

(0.152) (0.020)
The implied estimate for µ is 0.350/(1 − 0.951), which corresponds to approximately
7.2%, while the sample average is 7.3%. We can determine the Dickey–Fuller test
statistic from this regression as (0.951 − 1)/0.020 = −2.49, which means that we cannot reject the null hypothesis of a unit root at the 5% level.11 Because the AR(1) model
may be too restrictive, we also performed a number of augmented Dickey–Fuller tests
with 1, 3 and 6 additional lags included. The resulting test statistics were: −2.63, −2.29
and −1.88, respectively. Only the ﬁrst test implies a rejection at the 10% level. Thus,
we ﬁnd that a unit root in the short term interest rate cannot be rejected statistically.
Despite this, we will not impose it a priori in the sequel.
The short term interest rate is fairly well described by the ﬁrst order autoregressive
process in (8.88). Estimating AR(2) or ARMA(1, 1) speciﬁcations, for example, does
not result in a signiﬁcant improvement. The estimated autocorrelation function of the
residuals of the AR(1) model is given in Figure 8.10. It shows that only one of the
residual autocorrelation coefﬁcients is statistically signiﬁcant, at lag 8, which provides
only weak evidence against the hypothesis that the error term in (8.88) is a white noise
process. Moreover, none of the Ljung–Box tests rejects.
A way to test the expectations hypothesis is to regress a long interest rate on the
short rate, that is
rnt = β1 + β2 r1t + ut .
(8.89)
If (8.85) is taken to be literally true, the error term in this regression should be negligibly small (that is, the R 2 should be rather close to unity) and the true value of
β2 should equal ξn . The results of these regressions for maturities n = 3, 12 and 60
are given in Table 8.5. Given the high sensitivity of ξn with respect to θ , which was
not signiﬁcantly different from one, the estimated values for ξn do not, a priori, seem
in conﬂict with the time series model for the short rate. It must be mentioned, however, that the R 2 of the regression with the 5-year bond yield is fairly low. This
11

From Table 8.1, the appropriate critical value is −2.88.

AUTOREGRESSIVE CONDITIONAL HETEROSKEDASTICITY

297

Residual ACF

0.6
0.4
0.2
0
−0.2
−0.4
−0.6

1

2

3

4

5

6

7

8
k

9

10 11

12 13 14

15

Figure 8.10 Residual autocorrelation function, AR(1) model, 1970:1–1991:2
Table 8.5

value of ξn with θ = 0.95
value of ξn with θ = 1
OLS estimate of ξn
(standard error)
R 2 of regression

The term structure of interest rates
Quarterly n = 3

Annual n = 12

5-year n = 60

0.951
1
1.009
(0.009)
0.982

0.766
1
0.947
(0.017)
0.929

0.318
1
0.739
(0.028)
0.735

implies that other factors affect the long term yield in addition to the short rate.
One explanation is time variation in the risk premium n . Alternatively, the presence
of measurement errors in the interest rates may reduce their cross-sectional correlations.
At a more general level, the above example illustrates the delicate dependence of
long-run forecasts on the imposition of a unit root. While the estimated value of 0.95
is not signiﬁcantly different from one, imposing the unit root hypothesis would imply
that interest rates follow a random walk and that the forecast for any future period
is the last observation, in this case 5.68%. Using θ = 0.95, the optimal forecast 10
periods ahead is 6.3%, while the forecast for a 5-year horizon is virtually identical to
the unconditional mean of the series, 7.2%.

8.10

Autoregressive Conditional Heteroskedasticity

In ﬁnancial time series one often observes what is referred to as volatility clustering.
In this case big shocks (residuals) tend to be followed by big shocks in either direction,
and small shocks tend to follow small shocks. For example, stock markets are typically
characterized by periods of high volatility and more ‘relaxed’ periods of low volatility.
This is particularly true at high frequencies, for example with daily or weekly returns,

UNIVARIATE TIME SERIES MODELS

298

but less clear at lower frequencies. One way to model such patterns is to allow the
variance of εt to depend upon its history.
8.10.1 ARCH and GARCH Models

The seminal paper in this area is Engle (1982), which proposes the concept of autoregressive conditional heteroskedasticity (ARCH). It says that the variance of the error
term at time t depends upon the squared error terms from previous periods. The most
simple form is
2
σt2 ≡ E{εt2 |It−1 } =  + αεt−1
,
(8.90)
where It−1 denotes the information set, typically including εt−1 and its entire history.
This speciﬁcation is called an ARCH(1) process. To ensure that σt2 ≥ 0 irrespective of
2
εt−1
we need to impose that  ≥ 0 and α ≥ 0. The ARCH(1) model says that when a
big shock happens in period t − 1 it is more likely that εt has a large (absolute) value
2
as well. That is, when εt−1
is large, the variance of the next innovation εt is also large.
The speciﬁcation in (8.90) does not imply that the process for εt is nonstationary. It
2
just says that the squared values εt2 and εt−1
are correlated. The unconditional variance
of εt is given by
2
σ 2 = E{εt2 } =  + αE{εt−1
}
and has a stationary solution
σ2 =


1−α

(8.91)

provided that 0 ≤ α < 1. Note that the unconditional variance does not depend upon t.
The ARCH(1) model is easily extended to an ARCH(p) process, which we can
write as
2
2
2
2
σt2 =  + α1 εt−1
+ α2 εt−2
+ · · · + αp εt−p
=  + α(L)εt−1
,

(8.92)

where α(L) is a lag polynomial of order p − 1. To ensure that the conditional variance
is non-negative  and the coefﬁcients in α(L) must be non-negative. To ensure that
the process is stationary it is also required that α(1) < 1. The effect of a shock j
periods ago on current volatility is determined by the coefﬁcient αj . In an ARCH(p)
model, old shocks of more than p periods ago have no effect on current volatility.
The presence of ARCH errors in a regression or autoregressive model does not invalidate OLS estimation. It does imply, however, that more efﬁcient (nonlinear) estimators
exist than OLS. More importantly, it may be of interest to predict future variances,
for example, because they correspond to the riskiness of an investment. Consequently,
it is relevant to test for the presence of ARCH effects and, if needed, to estimate
the model allowing for it. Testing for p-th order autoregressive heteroskedasticity can
be done along the lines of the Breusch–Pagan test for heteroskedasticity discussed
in Chapter 4. It sufﬁces to run an auxiliary regression of squared OLS residuals et2
2
2
upon lagged squares et−1
, . . . , et−p
and a constant and compute T times the R 2 . Under
the null hypothesis of homoskedasticity (α1 = · · · = αp = 0) the resulting test statistic asymptotically follows a Chi-squared distribution with p degrees of freedom. In

AUTOREGRESSIVE CONDITIONAL HETEROSKEDASTICITY

299

other words, testing homoskedasticity against the alternative that the errors follow an
ARCH(p) process is very simple.
ARCH models have been generalized in many different ways. A useful variant,
proposed by Bollerslev (1986), is the generalized ARCH or GARCH model. In its
general form, a GARCH(p, q) model can be written as
σt2

=+

p


2
αj εt−j

j =1

+

q


2
βj σt−j

(8.93)

j =1

or
2
2
σt2 =  + α(L)εt−1
+ β(L)σt−1
,

(8.94)

where α(L) and β(L) are lag polynomials. In practice a GARCH(1,1) speciﬁcation
often performs very well. It can be written as
2
2
σt2 =  + αεt−1
+ βσt−1
,

(8.95)

which has only three unknown parameters to estimate. Non-negativity of σt2 requires
that , α and β are non-negative. If we deﬁne the surprise in the squared innovations
as vt ≡ εt2 − σt2 , the GARCH(1, 1) process can be rewritten as
2
+ vt − βvt−1 ,
εt2 =  + (α + β)εt−1

which shows that the squared errors follow an ARMA(1, 1) process. While the error vt is
uncorrelated over time (because it is a surprise term), it does exhibit heteroskedasticity.
The root of the autoregressive part is α + β, so that stationarity requires that α + β < 1.
Values of α + β close to unity imply that the persistence in volatility is high.12 Noting
2
2
that,13 under stationarity, E{εt−1
} = E{σt−1
} = σ 2 , the unconditional variance of εt
can be written as
σ 2 =  + ασ 2 + βσ 2
or
σ2 =


.
1−α−β

(8.96)

We can recursively substitute lags of (8.95) into itself to obtain
2
2
2
σt2 =  (1 + β + β 2 + · · ·) + α(εt−1
+ βεt−2
+ β 2 εt−3
+ · · ·)

=

∞


2
β j −1 εt−j
,
+α
1−β
j =1

(8.97)

which shows that the GARCH(1,1) speciﬁcation is equivalent to an inﬁnite order ARCH
model with geometrically declining coefﬁcients. It implies that the effect of a shock
The integrated GARCH(1,1) or IGARCH(1,1) process arises when α + β = 1 and volatility shocks have
a permanent effect (see Engle and Bollerslev, 1986).
13
The equality that follows only holds if εt does not exhibit autocorrelation.
12

UNIVARIATE TIME SERIES MODELS

300

on current volatility decreases over time. Consequently, a GARCH speciﬁcation may
provide a parsimonious alternative to a higher order ARCH process. Equation (8.97)
can also be rewritten as
σt2 − σ 2 = α

∞


2
β j −1 (εt−j
− σ 2 ),

(8.98)

j =1

which is convenient for forecasting.
Many alternative speciﬁcations to model conditional volatility are proposed in the
literature, corresponding to a variety of different acronyms (see Bollerslev, Chou and
Kroner, 1992; Bera and Higgins, 1993; Bollerslev, Engle and Nelson, 1994; or Diebold
and Lopez, 1995, for a review). An important restriction of the ARCH and GARCH
speciﬁcations above is their symmetry: only the absolute values of the innovations
matter, not their sign. That is, a big negative shock has the same impact on future
volatility as a big positive shock of the same magnitude. An interesting extension
is towards asymmetric volatility models, in which good news and bad news have a
different impact on future volatility. Note that the distinction between good and bad
news is more sensible for stock markets than for exchange rates, where agents typically
are on both sides of the market. That is, good news for one agent may be bad news
for another.
An asymmetric model should allow for the possibility that an unexpected drop
in price (‘bad news’) has a larger impact on future volatility than an unexpected
increase in price (‘good news’) of similar magnitude. A fruitful approach to capture
such asymmetries is provided by Nelson’s (1990) exponential GARCH or EGARCH
model, given by
2
log σt2 =  + β log σt−1
+γ

εt−1
|ε |
+ α t−1 ,
σt−1
σt−1

(8.99)

where α, β and γ are constant parameters. Because the level εt−1 /σt−1 is included,
the EGARCH model is asymmetric as long as γ = 0. When γ < 0, positive shocks
generate less volatility than negative shocks (‘bad news’). It is possible to extend the
EGARCH model by including additional lags. Note that we can rewrite (8.99) as
εt−1
σt−1
ε
2
+ (γ − α) t−1
=  + β log σt−1
σt−1

2
log σt2 =  + β log σt−1
+ (γ + α)

if εt−1 > 0
if εt−1 < 0.

The logarithmic transformation guarantees that variances will never become negative.
Typically, one would expect that γ + α > 0 while γ < 0.
Engle and Ng (1993) characterize a range of alternative models for conditional
volatility by a so-called news impact curve, which describes the impact of the last
return shock (news) on current volatility (keeping all information dated t − 2 or before
constant and ﬁxing all lagged conditional variances at the unconditional variance σ 2 ).
Compared to GARCH(1,1) the EGARCH model has an asymmetric news impact curve
(with a larger impact for negative shocks). Moreover, because the effect upon σt2 is

AUTOREGRESSIVE CONDITIONAL HETEROSKEDASTICITY

301

exponential, rather than quadratic, the news impact curve of the EGARCH model
typically has larger slopes (see Engle and Ng, 1993).
Financial theory tells us that certain sources of risk are priced by the market. That
is, assets with more ‘risk’ may provide higher average returns to compensate it. If σt2
is an appropriate measure of risk, the conditional variance may enter the conditional
mean function of yt . One variant of the ARCH-in-mean or ARCH-M model of Engle,
Lilien and Roberts (1987) speciﬁes that
yt = xt θ + δσt2 + εt ,
where εt is described by an ARCH(p) process (with conditional variance σt2 ).
Campbell, Lo and MacKinlay (1997, Section 12.2) provide additional discussion on the
links between ARCH-M models and asset pricing models, like the CAPM discussed
in Section 2.7.
8.10.2 Estimation and Prediction

There are different approaches to the estimation of conditional volatility models. Let
us assume that εt is the error term of a model like14 yt = xt θ + εt , where xt may
include lagged values of yt . As a special case xt is just a constant. Furthermore, let
the conditional variance of εt be described by an ARCH(p) process. Now, if we make
assumptions about the (conditional) distribution of εt we can estimate the model by
maximum likelihood. To see how, let
εt = σ t v t

with

vt ∼ NID(0, 1).

This implies that conditional upon the information in It−1 the innovation εt is normal
with mean zero and variance σt2 . It does not imply, however, that the unconditional
distribution of εt is normal, because σt becomes a random variable if we do not
condition upon It−1 . Typically, the unconditional distribution has fatter tails than a
normal one. From this, we can write down the conditional distribution of yt as
1
1
exp − (εt2 /σt2 )
f (yt |xt , It−1 ) = 
2
2
2πσt
2
2
+ · · · + αp εt−p
and εt = yt − xt θ . From this, the loglikeliwhere σt2 =  + α1 εt−1
hood function can be determined as the sum over all t of the log of the above expression,
substituting the appropriate expressions for σt2 and εt . The result can be maximized in
the
pusual way with respect to θ, α1 , . . . , αp and  . Imposing the stationarity condition
( j =1 αj < 1) and the non-negativity condition (αj ≥ 0 for all j ) may be difﬁcult in
practice, so that large values for p are not recommended.
If vt does not have a standard normal distribution, the above maximum likelihood
procedure may provide consistent estimators for the model parameters, even though
the likelihood function is incorrectly speciﬁed. The reason is that under some fairly
weak assumptions, the ﬁrst order conditions of the maximum likelihood procedure
14

To avoid confusion with the GARCH parameters, the regression coefﬁcients are referred to as θ .

UNIVARIATE TIME SERIES MODELS

302

are also valid when vt is not normally distributed. This is referred to as quasimaximum likelihood estimation (see Section 6.4). Some adjustments have to be
made, however, for the computation of the standard errors (see Hamilton, 1994, p. 663
for details).
A computationally simpler approach would be feasible GLS (see Chapter 4). In this
case, θ is ﬁrst estimated consistently by applying OLS. Second, a regression is done
2
2
of the squared OLS residuals et2 upon et−1
, . . . , et−p
and a constant. This is the same
regression that was used for the heteroskedasticity test described above. The ﬁtted
values from this regression are estimates for σt2 and can be used to transform the
model and compute a weighted least squares (EGLS) estimator for θ . This approach
only works well if the ﬁtted values for σt2 are all strictly positive. Moreover, it does
not provide asymptotically efﬁcient estimators for the ARCH parameters.
In ﬁnancial markets, GARCH models are frequently used to forecast volatility of
returns, which is an important input for investment, option pricing, risk management
and ﬁnancial market regulation (see Poon and Granger, 2003, for a review). Forecasting
the conditional variance from an ARCH(p) model is straightforward. To see this,
rewrite the model ‘in deviations from means’ as
2
2
σt2 − σ 2 = α1 (εt−1
− σ 2 ) + · · · + αp (εt−p
− α2)

with σ 2 =  /(1 − α1 · · · − αp ). Assuming for notational convenience that the model
parameters are known, the one-period ahead forecast follows as
2
2
2
σt+1|t
≡ E{εt+1
|It } = σ 2 + α1 (εt2 − σ 2 ) + · · · + αp (εt−p+1
− σ 2 ).

This is analogous to predicting from an AR(p) model for yt as discussed in Section 8.8.
Forecasting the conditional volatility more than one period ahead can be done using
the recursive formula
2
2
2
2
σt+h|t
≡ E{εt+h
|It } = σ 2 + α1 (σt+h−1|t
− σ 2 ) + · · · + αp (σt+h−p|t
− σ 2 ),
2
2
where σt+j
|t = εt+j if j ≤ 0. The h-period ahead forecast converges to the unconditional variance σ 2 if h becomes large (assuming that α1 + · · · + αp < 1).
For a GARCH model prediction and estimation can take place along the same lines
if we use (8.97), (8.98) or a higher order generalization. For example, the one-period
ahead forecast for a GARCH(1, 1) model is given by
2
σt+1|t
= σ 2 + α(εt2 − σ 2 ) + β(σt2 − σ 2 )

where σt2 = σ 2 + α
ten as

∞

j =1

2
β j −1 (εt−j
− σ 2 ). The h-period ahead forecast can be writ-

2
2
σt+h|t
= σ 2 + (α + β)[σt+h−1|t
− σ 2]

= σ 2 + (α + β)h−1 [α(εt2 − σ 2 ) + β(σt2 − σ 2 )],
which shows that the volatility forecasts converge to the unconditional variance at
a rate α + β. For EGARCH models estimation can also be done by maximum like-

AUTOREGRESSIVE CONDITIONAL HETEROSKEDASTICITY

303

lihood, although simple closed-form expressions for multi-period forecasts are not
available. Empirically the likelihood function for an EGARCH model is more difﬁcult
to maximize and problems of nonconvergence occasionally occur.
8.10.3 Illustration: Volatility in Daily Exchange Rates

To illustrate some of the volatility models discussed above, we consider a series of
daily exchange rates between the US$ and the Deutsche Mark (DM) from 1 January
1980 to 21 May 1987. Excluding days for which no prices are quoted (New Year’s day
etc.), this results in a total of T = 1867 observations. As (log) exchange rates are to a
rough approximation described by a random walk, we consider a model where yt is the
change in the log exchange rate and the conditional mean only includes an intercept.
The series for yt is plotted in Figure 8.11, which shows the existence of periods with
low volatility and periods with high volatility.
The OLS residuals et of regressing yt upon a constant correspond, of course, with
yt minus its sample average. On the basis of these residuals we can perform tests
for ARCH effects, by regressing et2 upon a constant and p of its lags. A test of
homoskedasticity against ARCH(1) errors produces a test statistic (computed as T
times the R 2 of the auxiliary regression) of 21.77, which is highly signiﬁcant for a
Chi-squared distribution with one degree of freedom. Similarly, we can test against
ARCH(6) errors with a statistic of 83.46, which also results in a clear rejection of the
homoskedasticity assumption.
The following three models are estimated: an ARCH(6), a GARCH(1,1) and a
standard exponential GARCH model15 (EGARCH(1,1)) and we present the results
0.06

0.04

0.02

0.00

−0.02

−0.04
500

Figure 8.11
15

1000

5000

Daily change in log exchange rate US$/DM, 2 January 1980–21 May 1987

To improve numerical optimization, the series is multiplied by 100. The results in Table 8.6 were obtained
by RATS 6.0. Standard software to estimate these models is also available in, for example, EViews and
MicroFit. Depending on the optimization routine, starting values and convergence criteria used by these
programs, the estimation results may be slightly different.

UNIVARIATE TIME SERIES MODELS

304

Table 8.6

GARCH estimates for change in log exchange rate US$/DM
ARCH(6)

GARCH(1,1)

EGARCH

0.228
(0.023)
0.092
(0.026)
0.081
(0.025)
0.123
(0.028)
0.138
(0.033)
0.122
(0.028)
0.101
(0.027)
–

0.016
(0.005)
0.110
(0.016)
–

−0.185
(0.023)
0.215
(0.027)

constant
2
εt−1
2
εt−2
2
εt−3
2
εt−4
2
εt−5
2
εt−6
2
σt−1

|εt−1 |/σt−1

–
–
–
–
0.868
(0.018)

2
log(σt−1
)

0.968
(0.009)
−0.017
(0.013)

εt−1 /σt−1

1.6

1.4

1.2

1.0

0.8

0.6

0.4
1770

1780

1790

ARCH6

Figure 8.12
1987

1800

1810

1820

EGARCH11

1830

1840

1850

1860

GARCH11

Conditional standard deviations implied by different models, 1 January–21 May

WHAT ABOUT MULTIVARIATE MODELS?

305

in Table 8.6. All speciﬁcations are estimated by maximum likelihood assuming that
the conditional distribution of the errors is normal. The results for the ARCH(6) speciﬁcation show that all 6 lags have a signiﬁcant and positive effect. Moreover, the
coefﬁcients do not appear to decrease to zero very quickly. The more parsimonious
GARCH(1,1) model also indicates that the effect of lagged shocks dies out only very
slowly. The estimated value of α + β is 0.978, so that the estimated process is close to
being nonstationary. This is a typical ﬁnding in empirical applications. For the exponential GARCH model we do not ﬁnd evidence of asymmetry as the γ coefﬁcient has
a t-ratio of only −1.35. As indicated above, this is not an unusual ﬁnding for exchange
rates. The large coefﬁcient for log σt2 also reﬂects the high degree of persistence in
exchange rate volatility.
To compare the alternative volatility models, Figure 8.12 plots the estimated standard
deviations σ̂t as implied by the parameter estimates. To minimize the impact of
initial conditions and to appreciate the differences across models we only present
results for the ﬁve months of 1987. The graph shows that the volatility implied
by the ARCH(6) speciﬁcation is less smooth than those for the GARCH(1,1) and
EGARCH(1,1) models. Apparently, six lags are insufﬁcient to capture the persistence
of volatility.

8.11

What about Multivariate Models?

This chapter has concentrated on a more or less pure statistical approach of ﬁtting an
adequate time series model (from the class of ARMA models) to an observed time
series. This is what we referred to as univariate time series modelling. In real life,
it is obvious that many economic variables are related to each other. This, however,
does not imply that a pure time series analysis is wrong. Building structural models in
which variables are linked to each other (often based on economic theory) is a different
branch. It gives insight into the interrelationships between variables and how a certain
policy (shock) affects the economy (not just what its ﬁnal effect is). Of course, these
advantages do require a ‘correct’ representation of the underlying economy. In the
time series approach, one is more concerned with predicting future values, including
future uncertainty (variances). To this end, (in univariate time series analysis) only
the history of the variable under concern is taken into account. As said before, from
the predictive point of view, a pure time series approach often outperforms a more
structural approach.
To illustrate the relationships, assume that the following regression model describes
the relationship between two (demeaned) variables yt and xt ,
yt = βxt + εt ,
where εt is a white noise error term. If xt can be described by some ARMA model,
then yt is the sum of an ARMA process and a white noise process and will therefore
also follow an ARMA process. For example, if xt can be described by a ﬁrst order
moving average model
xt = ut + αut−1 ,

UNIVARIATE TIME SERIES MODELS

306

where ut is a white noise error independent of εt , then we can write
yt = βut + αβut−1 + εt .
From this, we can easily derive that the autocovariances of yt are V {yt } = σε2 + β 2 (1 +
α 2 )σu2 , cov{yt , yt−1 } = β 2 ασu2 and cov{yt , yt−k } = 0 for k = 2, 3, . . . . Consequently,
yt follows a ﬁrst order moving average process, with parameters that can be solved
from the above covariances. Thus, the fact that two variables are related does not imply
that a pure times series approach is invalid.
In the next chapter, we shall extend the univariate time series approach to a multivariate setting. This allows us to consider the time series properties of a number of
series simultaneously, along with their short-run and long-run dependencies.

Exercises
Exercise 8.1 (ARMA Models and Unit Roots)

A researcher uses a sample of 200 quarterly observations on Yt , the number (in 1000s)
of unemployed persons, to model the time series behaviour of the series and to generate predictions. First, he computes the sample autocorrelation function, with the
following results:
k
ρ̂k
a.

1
2
3
4
0.83 0.71 0.60 0.45

5
6
7
0.44 0.35 0.29

8
9
10
0.20 0.11 −0.01

What do we mean by the sample autocorrelation function? Does the above pattern
indicate that an autoregressive or moving average representation is more appropriate? Why?

Next, the sample partial autocorrelation function is determined. It is given by
k
θ̂kk

1
2
3
4
0.83 0.16 −0.09 0.05

5
6
7
8
0.04 −0.05 0.01 0.10

9
−0.03

10
−0.01

b. What do we mean by the sample partial autocorrelation function? Why is the ﬁrst
partial autocorrelation equal to the ﬁrst autocorrelation coefﬁcient (0.83)?
c. Does the above pattern indicate that an autoregressive or moving average representation is more appropriate? Why?
The researcher decides to estimate, as a ﬁrst attempt, a ﬁrst order autoregressive model
given by
Yt = δ + θ Yt−1 + εt .
(8.100)
The estimated value for θ1 is 0.83 with a standard error of 0.07.
d.

Which estimation method is appropriate for estimating the AR(1) model? Explain
why it is consistent.

EXERCISES

307

e. The researcher wants to test for a unit root. What is meant by ‘a unit root’? What
are the implications of the presence of a unit root? Why are we interested in it?
(Give statistical or economic reasons.)
f. Formulate the hypothesis of a unit root and perform a unit root test based on the
above regression.
g. Perform a test for the null hypothesis that θ = 0.90.
Next, the researcher extends the model to an AR(2), with the following results (standard
errors in parentheses):
Yt = 50.0 + 0.74Yt−1 + 0.16Yt−2 + ε̂t .
(5.67) (0.07)
(0.07)

(8.101)

h. Would you prefer the AR(2) model to the AR(1) model? How would you check
whether an ARMA(2, 1) model may be more appropriate?
i. What do the above results tell you about the validity of the unit root test of f?
j. How would you test for the presence of a unit root in the AR(2) model?
k. From the above estimates, compute an estimate for the average number of unemployed E{Yt }.
l. Suppose the last two quarterly unemployment levels for 1996-III and 1996-IV
were 550 and 600, respectively. Compute predictions for 1997-I and 1997-II.
m. Can you say anything sensible about the predicted value for the quarter 2023-I?
(And its accuracy?)
Exercise 8.2 (Modelling Daily Returns – Empirical)

In the ﬁles SP500 daily returns on the S&P 500 index are available from January 1981
to April 1991 (T = 2783). Returns are computed as ﬁrst differences of the log of the
S&P 500 US stock price index.
a.

Plot the series and determine the sample autocorrelation and partial autocorrelation function.
b. Estimate an AR(1) up to AR(7) model and test the individual and joint signiﬁcance
of the AR coefﬁcients. Why would a signiﬁcance level of 1% or less be more
appropriate than the usual 5%?
c. Perform Ljung–Box tests on residual autocorrelation in these seven models for
K = 6 (when appropriate), 12 and 18.
d. Compare AIC and BIC values. Use them, along with the results of the statistical
tests, to choose a preferred speciﬁcation.
For the next questions use your preferred speciﬁcation.
e. Save the residuals of your model and test against p-th order autoregressive heteroskedasticity (choose several alternative values for p).
f. Re-estimate your model allowing for ARCH(p) errors (where p is chosen on the
basis of the above tests). Compare the estimates with those of the test regressions.

308

UNIVARIATE TIME SERIES MODELS

g.

Re-estimate your model allowing for GARCH(1,1) errors. Is there any indication
of nonstationarity?
h. Re-estimate your model allowing for EGARCH errors. (Be sure to check that the
program has converged.) Is there any evidence for asymmetry?
Exercise 8.3 (Modelling Quarterly Income – Empirical)

In the ﬁles INCOME, information is available on quarterly disposable income in the
United Kingdom for the quarters 1971:I to 1985:II, measured in million pounds and
current prices (T = 58).
a.

b.

c.

d.
e.
f.
g.
h.

Produce a graph of the natural logarithm of quarterly income. Estimate a standard
Dickey–Fuller regression, with an intercept term, for log income and compute the
Dickey–Fuller test statistic for a unit root. What do you conclude? Repeat the test
while including a linear time trend.
Perform augmented Dickey–Fuller tests including 1 up to 6 lags, with and without
including a linear trend. What do you conclude about the presence of a unit root
in log income?
Transform the series into ﬁrst differences and produce a graph. Perform augmented
Dickey–Fuller tests on the change in log income including 1 up to 6 lags. What
do you conclude? Motivate why you did or did not include a time trend.
Determine the sample ACF and PACF for the change in log income. Is there an
obvious model suggested by these graphs?
Estimate an AR(4) and an MA(4) model for the change in log income.
Test for serial correlation in the residuals of these two models. Can you reject the
null hypothesis of white noise errors?
Find a parsimonious model that adequately describes the process generating the
change in log income. Motivate the steps that you take.
Use the model to forecast quarterly disposable income in 1985:III.

9

Multivariate Time
Series Models

In the previous chapter we considered models for the stochastic process of a single
economic time series. One reason why it may be more interesting to consider several
series simultaneously is that it may improve forecasts. For example, the history of a
second variable, Xt say, may help forecasting future values of Yt . It is also possible
that particular values of Xt are associated with particular movements in the Yt variable.
For example, oil price shocks may be helpful in explaining gasoline consumption. In
addition to the forecasting issue, this also allows us to consider ‘what if’ questions.
For example, what is the expected future development of gasoline consumption if oil
prices are decreasing by 10% over the next couple of years?
In this chapter we consider multivariate time series models. In Section 9.1, we shall
consider explaining one variable from its own past including current or lagged values of a second variable. This way, the dynamic effects of a change in Xt upon Yt
can be modelled and estimated. To apply standard estimation or testing procedures
in a dynamic time series model, it is typically required that the various variables are
stationary, since the majority of econometric theory is built upon the assumption of
stationarity. For example, regressing a nonstationary variable Yt upon a nonstationary
variable Xt may lead to a so-called spurious regression, in which estimators and test
statistics are misleading. The use of nonstationary variables not necessarily results in
invalid estimators. An important exception arises when two or more I (1) variables are
cointegrated, that is, if there exists a particular linear combination of these nonstationary variables which is stationary. In such cases a long-run relationship between these
variables exists. Often, economic theory suggests the existence of such long-run or
equilibrium relationships, for example, purchasing power parity or the quantity theory
of money. The existence of a long-run relationship also has its implications for the
short-run behaviour of the I (1) variables, because there has to be some mechanism that
drives the variables to their long-run equilibrium relationship. This mechanism is modelled by an error-correction mechanism, in which the ‘equilibrium error’ also drives

MULTIVARIATE TIME SERIES MODELS

310

the short-run dynamics of the series. Section 9.2 introduces the concept of cointegration when only two variables are involved, and relates it to error-correction models.
In Section 9.3 an empirical illustration is provided on purchasing power parity, which
can be characterized as corresponding to a long-run cointegrating relationship.
Another starting point of multivariate time series analysis is the multivariate generalization of the ARMA processes of Chapter 8. This is the topic of Section 9.4, where
particular emphasis is placed on vector autoregressive models (VARs). The existence
of cointegrating relationships between the variables in the VAR has important implications on the way it can be estimated and represented. Section 9.5 discusses how
hypotheses regarding the number of cointegrating relationships can be tested, and how
an error-correction model representing the data can be estimated. Finally, Section 9.6
concludes with an empirical illustration.
There exists a fairly large number of recent textbooks on time series analysis
that discuss cointegration, vector autoregressions and error-correction models. For
economists attractive choices are Mills (1990), Harris (1995), Franses (1998), Patterson
(2000) and Enders (2004). More technical detail is provided in, for example, Banerjee,
Dolado, Galbraith and Hendry (1993), Hamilton (1994), Johansen (1995), Maddala and
Kim (1998), Boswijk (1999) and Gourieroux and Jasiak (2001). Most of these texts
also discuss topics that are not covered in this chapter, including structural VARs,
Granger causality, seasonality and structural breaks.

9.1

Dynamic Models with Stationary Variables

Considering an economic time series in isolation and applying techniques from the
previous chapter to model it may provide good forecasts in many cases. It does not,
however, allow us to determine what the effects are of, for example, a change in a
policy variable. To do so, it is possible to include additional variables in the model.
Let us consider two (stationary) variables1 Yt and Xt , and assume that it holds that
Yt = δ + θ Yt−1 + φ0 Xt + φ1 Xt−1 + εt .

(9.1)

As an illustration, we can think of Yt as ‘company sales’ and Xt as ‘advertising’, both
in month t. If we assume that εt is a white noise process, independent of Xt , Xt−1 , . . .
and Yt−1 , Yt−2 , . . . , the above relation is sometimes referred to as an autoregressive distributed lag model.2 To estimate it consistently, we can simply use ordinary
least squares.
The interesting element in (9.1) is that it describes the dynamic effects of a change
in Xt upon current and future values of Yt . Taking partial derivatives, we can derive
that the immediate response is given by
∂Yt /∂Xt = φ0 .
1

(9.2)

In line with the previous chapter we use capital letters to denote the original series and small letters for
deviations from the mean.
2
More details can be found in, for example, Davidson and MacKinnon (1993, Section 19.4) or Johnston
and Dinardo (1997, Chapter 8).

DYNAMIC MODELS WITH STATIONARY VARIABLES

311

Sometimes this is referred to as the impact multiplier. An increase in X with one unit
has an immediate impact on Y of φ0 units. The effect after one period is
∂Yt+1 /∂Xt = θ ∂Yt /∂Xt + φ1 = θ φ0 + φ1 ,

(9.3)

∂Yt+2 /∂Xt = θ ∂Yt+1 /∂Xt = θ (θ φ0 + φ1 )

(9.4)

and after two periods

and so on. This shows that after the ﬁrst period, the effect is decreasing if |θ | < 1.
Imposing this so-called stability condition allows us to determine the long-run effect of
a unit change in Xt . It is given by the long-run multiplier (or equilibrium multiplier)
φ0 + (θ φ0 + φ1 ) + θ (θ φ0 + φ1 ) + · · ·
= φ0 + (1 + θ + θ 2 +· · ·)(θ φ0 + φ1 ) =

φ0 + φ1
.
1−θ

(9.5)

This says that if advertising Xt increases with one unit, the expected cumulative
increase in sales is given by (φ0 + φ1 )/(1 − θ ). If the increase in Xt is permanent,
the long-run multiplier also has the interpretation of the expected long-run permanent
increase in Yt . From (9.1) the long-run equilibrium relation between Y and X can be
seen to be (imposing E{Yt } = E{Yt−1 })
E{Yt } = δ + θ E{Yt } + φ0 E{Xt } + φ1 E{Xt }
or
E{Yt } =

δ
φ + φ1
+ 0
E{Xt },
1−θ
1−θ

(9.6)
(9.7)

which presents an alternative derivation of the long-run multiplier. We shall write (9.7)
shortly as E{Yt } = α + βE{Xt }, with obvious deﬁnitions of α and β.
There is an alternative way to formulate the autoregressive distributed lag model
from (9.1). Subtracting Yt−1 from both sides of (9.1) and some rewriting gives
Yt = δ − (1 − θ )Yt−1 + φ0 Xt + (φ0 + φ1 )Xt−1 + εt
or

Yt = φ0 Xt − (1 − θ )[Yt−1 − α − βXt−1 ] + εt .

(9.8)

This formulation is an example of an error-correction model. It says that the change in
Yt is due to the current change in Xt plus an error-correction term. If Yt−1 is above the
equilibrium value that corresponds to Xt−1 , that is, if the ‘equilibrium error’ in square
brackets is positive, an additional negative adjustment in Yt is generated. The speed
of adjustment is determined by 1 − θ , which is the adjustment parameter. Assuming
stability ensures that 1 − θ > 0.
It is also possible to consistently estimate the error-correction model by least squares.
Because the residual sum of squares that is minimized with (9.8) is the same as that
of (9.1), the resulting estimates are numerically identical.3
3

The model in (9.8) can be estimated by nonlinear least squares or by OLS after reparametrization and
solving for the original parameters from the resulting estimates. The results are the same.

MULTIVARIATE TIME SERIES MODELS

312

Both the autoregressive distributed lag model in (9.1) and the error-correction model
in (9.8) assume that the values of Xt can be treated as given, that is, as being uncorrelated with the equations’ error terms. Essentially this says that (9.1) is appropriately
describing the expected value of Yt given its own history and conditional upon current
and lagged values of Xt . If Xt is simultaneously determined with Yt and E{Xt εt } = 0,
OLS in either (9.1) or (9.8) would be inconsistent. The typical solution in this context
is to consider a bivariate model for both Yt and Xt (see Section 9.5 below).
Special cases of the model in (9.1) can be derived from alternative models that have
some economic interpretation. For example, let Yt∗ denote the optimal or desired level
of Yt and assume that
Yt∗ = α + βXt + ηt ,
(9.9)
for some unknown coefﬁcients α and β, and where ηt is an error term independent of
Xt , Xt−1 , . . . . The actual value Yt differs from Yt∗ because adjustment to its optimal
level corresponding to Xt is not immediate. Suppose that the adjustment is only partial
in the sense that
Yt − Yt−1 = (1 − θ )(Yt∗ − Yt−1 )
(9.10)
where 0 < θ < 1. Substituting (9.9) we obtain
Yt = Yt−1 + (1 − θ )α + (1 − θ )βXt − (1 − θ )Yt−1 + (1 − θ )ηt
= δ + θ Yt−1 + φ0 Xt + εt ,

(9.11)

where δ = (1 − θ )α, φ0 = (1 − θ )β and εt = (1 − θ )ηt . This is a special case of (9.1)
as it does not include Xt−1 . The model given by (9.9) and (9.10) is referred to as a
partial adjustment model.
The autoregressive distributed lag model in (9.1) can be easily generalized. Restricting attention to two variables only, we can write a general form as
θ (L)Yt = δ + φ(L)Xt + εt ,
where

(9.12)

θ (L) = 1 − θ1 L − · · · − θp Lp
φ(L) = φ0 + φ1 L + · · · + φq Lq

are two lag polynomials. Note that the constant in φ(L) is not restricted to be one.
Assuming that θ (L) is invertible (see Subsection 8.2.2), we can write
Yt = θ −1 (1)δ + θ −1 (L)φ(L)Xt + θ −1 (L)εt .

(9.13)

The coefﬁcients in the lag polynomial θ −1 (L)φ(L) describe the dynamic effects of Xt
upon current and future values of Yt . The long-run effect of Xt is obtained as
θ −1 (1)φ(1) =

φ0 + φ1 + · · · + φq
1 − θ1 − · · · − θp

,

(9.14)

MODELS WITH NONSTATIONARY VARIABLES

313

which generalizes the result above. Recall from Subsection 8.2.2 that invertibility of
θ (L) requires that θ1 + θ2 + · · · + θp < 1, which guarantees that the denominator in
(9.14) is nonzero. A special case arises if θ (L) = 0, so that the model in (9.13) does
not contain any lags of Yt . This is referred to as a distributed lag model.
As long as it can be assumed that the error term εt is a white noise process, or – more
generally – is stationary and independent of Xt , Xt−1 , . . . and Yt−1 , Yt−2 , . . . , the distributed lag models can be estimated consistently by ordinary least squares. Problems
may arise, however, if along with Yt and Xt the implied εt is also nonstationary. This
is discussed in the next section.

9.2

Models with Nonstationary Variables

9.2.1 Spurious Regressions

The assumption that the Yt and Xt variables are stationary is crucial for the properties of
standard estimation and testing procedures. To show consistency of the OLS estimator,
for example, we typically use the result that when the sample size increases, sample
(co)variances converge to population (co)variances. Unfortunately, when the series
are nonstationary the latter (co)variances are ill-deﬁned because the series are not
ﬂuctuating around a constant mean.
As an illustration, consider two variables, Yt and Xt , generated by two independent
random walks,
Yt = Yt−1 + ε1t ,
ε1t ∼ IID(0, σ12 )
(9.15)
Xt = Xt−1 + ε2t ,

ε2t ∼ IID(0, σ22 )

(9.16)

where ε1t and ε2t are mutually independent. There is nothing in this data generating
mechanism that leads to a relationship between Yt and Xt . A researcher, unfamiliar
with these processes, may want to estimate a regression model explaining Yt from Xt
and a constant,4
Yt = α + βXt + εt .
(9.17)
The results from this regression are likely to be characterized by a fairly high R 2
statistic, highly autocorrelated residuals and a signiﬁcant value for β. This phenomenon
is the well-known problem of nonsense or spurious regressions (see Granger and
Newbold, 1974). In this case, two independent nonstationary series are spuriously
related due to the fact that they are both trended. As argued by Granger and Newbold,
in these situations, characterized by a high R 2 and a low Durbin–Watson (dw ) statistic,
the usual t- and F-tests on the regression parameters may be very misleading. The
reason for this is that the distributions of the conventional test statistics are very
different from those derived under the assumption of stationarity. In particular, as
shown by Phillips (1986), the OLS estimator does not converge in probability as the
sample size increases, the t- and F-test statistics do not have well-deﬁned asymptotic
4

To ensure consistent notation throughout this chapter, the constant term is denoted by α and the slope
coefﬁcient by β. It will be clear from what follows that the role of the constant is often fundamentally
different from the slope coefﬁcients when variables are nonstationary.

MULTIVARIATE TIME SERIES MODELS

314

Table 9.1 Spurious regression: OLS involving two independent random walks
Dependent variable: Y
Variable

Estimate

Standard error

t-ratio

constant
X

3.9097
−0.4435

0.2462
0.0473

15.881
−9.370

s = 3.2698 R 2 = 0.3072 R̄ 2 = 0.3037 F = 87.7987
dw = 0.1331

distributions, and the dw statistic converges to zero. The reason is that with Yt and Xt
being I (1) variables, the error term εt will also be a nonstationary I (1) variable.
To illustrate the spurious regression result, we generated two series of 200 observations5 according to (9.15) and (9.16) with normal error terms, starting with Y0 = X0 = 0
and setting σ12 = σ22 = 1. The results of a standard OLS regression of Yt upon Xt and
a constant are presented in Table 9.1. While the parameter estimates in this table
would be completely different from one simulation to the next, the t-ratios, R 2 and
dw statistic show a very typical pattern: using the usual signiﬁcance levels both the
constant term and Xt are highly signiﬁcant, the R 2 of 31% seems reasonable, while
the Durbin–Watson statistic is extremely low. (Remember from Chapter 4 that values
close to 2 correspond to the null hypothesis of no autocorrelation.) Estimation results
like this should not be taken seriously. Because both Yt and Xt contain a stochastic trend, the OLS estimator tends to ﬁnd a signiﬁcant correlation between the two
series, even if they are completely unrelated. Statistically, the problem is that εt is
nonstationary.
If lagged values of both the dependent and independent variables are included in
the regression, as in (9.1), no spurious regression problem arises because there exist
parameter values (viz. θ = 1 and φ0 = φ1 = 0) such that the error term εt is I (0),
even if Yt and/or Xt are I (1). In this case the OLS estimator is consistent for all
parameters. Thus, including lagged values in the regression is sufﬁcient to solve many
of the problems associated with spurious regression (see Hamilton, 1994, p. 562).
9.2.2 Cointegration

An important exception to the results in the previous subsection arises when the two
nonstationary series have the same stochastic trend in common. Consider two series,
integrated of order one, Yt and Xt , and suppose that a linear relationship exists between
them. This is reﬂected in the proposition that there exists some value β such that
Yt − βXt is I (0), although Yt and Xt are both I (1). In such a case it is said that Yt
and Xt are cointegrated, and that they share a common trend. Although the relevant
asymptotic theory is nonstandard, it can be shown that one can consistently estimate β
from an OLS regression of Yt on Xt as in (9.17). In fact, in this case, the OLS estimator
b is said to be super consistent for β, because it converges to β at a much
√ faster rate
than with conventional asymptotics. In the standard
case,
we
have
that
T (b − β) is
√
asymptotically normal and we say that b is T -consistent for β. In the cointegration
5

These simulated series are available in SPURIOUS.

MODELS WITH NONSTATIONARY VARIABLES

315

√
case, T (b − β) is degenerate, which means that b √
converges to β at such a fast rate
that the difference b − β, multiplied by an increasing T factor, still converges to zero.
Instead, the appropriate asymptotic distribution is the one of T (b − β). Consequently,
conventional inference procedures do not apply.
The intuition behind the super consistency result is quite straightforward. Suppose
the estimated regression model is
Yt = a + bXt + et .

(9.18)

For the true value of β, Yt − βXt is I (0). Clearly, for b = β, the OLS residual et will
be nonstationary and hence will have a very large variance in any ﬁnite sample. For
b = β, however, the estimated variance of et will be much smaller. Since ordinary
least squares chooses a and b to minimize the sample variance of et , it is extremely
good in ﬁnding an estimate close to β.
If Yt and Xt are both I (1) and there exists a β such that Zt = Yt − βXt is I (0),
Yt and Xt are cointegrated, with β being called the cointegrating parameter, or, more
generally, (1, −β) being called the cointegrating vector. When this occurs, a special
constraint operates on the long-run components of Yt and Xt . Since both Yt and Xt
are I (1), they will be dominated by ‘long wave’ components, but Zt , being I (0), will
not be: Yt and βXt must therefore have long-run components that virtually cancel out
to produce Zt .
This idea is related to the concept of a long-run equilibrium. Suppose that such an
equilibrium is deﬁned by the relationship
Yt = α + βXt .

(9.19)

Then zt = Zt − α is the ‘equilibrium error’, which measures the extent to which the
value of Yt deviates from its ‘equilibrium value’ α + βXt . If zt is I (0), the equilibrium
error is stationary and ﬂuctuating around zero. Consequently, the system will, on average, be in equilibrium. However, if Yt and Xt are not cointegrated and, consequently,
zt is I (1), the equilibrium error can wander widely and zero-crossings would be very
rare. Under such circumstances, it does not make sense to refer to Yt = α + βXt as
a long-run equilibrium. Consequently, the presence of a cointegrating vector can be
interpreted as the presence of a long-run equilibrium relationship.
From the discussion above, it is obvious that it is important to distinguish cases
where there is a cointegrating relationship between Yt and Xt and spurious regression
cases. Suppose we know from previous results that Yt and Xt are integrated of order
one, and suppose we estimate the ‘cointegrating regression’
Yt = α + βXt + εt .

(9.20)

If Yt and Xt are cointegrated the error term in (9.20) is I (0). If not, εt will be I (1).
Hence, one can test for the presence of a cointegrating relationship by testing for a unit
root in the OLS residuals et from (9.20). It seems that this can be done by using the
Dickey–Fuller tests of the previous section. For example, one can run the regression,
et = γ0 + γ1 et−1 + ut

(9.21)

MULTIVARIATE TIME SERIES MODELS

316

Table 9.2 Asymptotic critical values residual unit root
tests for cointegration (with constant term) (Davidson
and MacKinnon, 1993)
Signiﬁcance level

Number of variables
(incl. Yt )

1%

5%

10%

2
3
4
5

−3.90
−4.29
−4.64
−4.96

−3.34
−3.74
−4.10
−4.42

−3.04
−3.45
−3.81
−4.13

and test whether γ1 = 0 (a unit root). There is, however, an additional complication in
testing for unit roots in OLS residuals rather than in observed time series. Because the
OLS estimator ‘chooses’ the residuals in the cointegrating regression (9.20) to have
as small a sample variance as possible, even if the variables are not cointegrated, the
OLS estimator will make the residuals ‘look’ as stationary as possible. Thus using
standard DF or ADF tests we may reject the null hypothesis of nonstationarity too
often. As a result, the appropriate critical values are more negative than those for the
standard Dickey–Fuller tests and are presented in Table 9.2. If et is not appropriately
described by a ﬁrst order autoregressive process, one should add lagged values of et to
(9.21), leading to the augmented Dickey–Fuller (ADF) tests, with the same asymptotic
critical values. This test can be extended to test for cointegration between three or more
variables. If more than a single Xt variable is included in the cointegrating regression,
the critical values shift further to the left. This is reﬂected in the additional rows in
Table 9.2.
An alternative test for cointegration is based on the usual Durbin–Watson statistic
from (9.20). Note that the presence of a unit root in εt asymptotically corresponds
with a zero value for the dw statistic. Thus under the null hypothesis of a unit root,
the appropriate test is whether dw is signiﬁcantly larger than zero. Unfortunately,
critical values for this test, commonly referred to as the Cointegrating Regression
Durbin–Watson test or CRDW test (see Sargan and Bhargava, 1983), depend upon
the process that generated the data. Nevertheless, the value of the Durbin–Watson
statistic is often suggestive for the presence or absence of a cointegrating relationship.
When the data are generated by a random walk, 5% critical values are given in Table 9.3
for a number of different sample sizes. Note that when T goes to inﬁnity, and Yt and
Xt are not cointegrated, the dw statistic converges to zero (in probability).
The cointegration tests discussed here test the presence of a unit root in regression
residuals. This implies that the null hypothesis of a unit root corresponds to no cointegration. So, if we cannot reject the presence of a unit root in the OLS residuals,
this implies that we cannot reject that Yt and Xt are not cointegrated. Sometimes, it
may be more appropriate to test the null hypothesis that two or more variables are
cointegrated against the alternative that they are not. Recently, several authors have
suggested tests for the null of cointegration; see Maddala and Kim (1998, Section 4.5)
for a review.
If Yt and Xt are cointegrated, OLS applied to (9.20) produces a super consistent
estimator of the cointegrating vector, even if short-run dynamics are incorrectly omitted.
The reason for this is that the nonstationarity asymptotically dominates all forms of

MODELS WITH NONSTATIONARY VARIABLES

317

Table 9.3 5% Critical values CRDW tests for cointegration (Banerjee et al., 1993)
Number of variables
(incl. Yt )
2
3
4
5

Number of observations
50

100

200

0.72
0.89
1.05
1.19

0.38
0.48
0.58
0.68

0.20
0.25
0.30
0.35

misspeciﬁcation in the stationary part of (9.20). Thus, incomplete short-run dynamics,
autocorrelation in εt , omitted (stationary) variables, endogeneity of Xt are all problems
in the stationary part of the regression which can be neglected (that is, are of lower
order) when looking at the asymptotic distribution of the super consistent estimator b. In
general, however, the OLS estimator for the cointegrating parameter has a non-normal
distribution, and inferences based on its t-statistic tend to be misleading.
Another problem with the OLS estimator is that, despite the super consistency property, Monte Carlo studies indicate that in small samples the bias in the estimated
cointegrating relation may be substantial (see Banerjee et al., 1993, Section 7.4). Typically these biases are small if the R 2 of the cointegrating regression is close to unity.
A large number of alternative estimators have been proposed in the literature (see
Hargreaves, 1994, for a review). A simple alternative is the so-called dynamic OLS
estimator, suggested by Stock and Watson (1993), based on extending the cointegrating
regression by adding leads and lags of Xt . Under appropriate conditions, the resulting
estimator for β has an approximate normal distribution and standard t-statistics (based
on HAC standard errors) are valid. A more complicated alternative is the so-called
fully modiﬁed OLS estimator, suggested by Phillips and Hansen (1990); see Patterson
(2000, Chapter 9) for discussion.
Asymptotically, one can interchange the role of Yt and Xt and estimate
Xt = α ∗ + β ∗ Yt + u∗t ,

(9.22)

to get super consistent estimates of α ∗ = −α/β and β ∗ = 1/β. It is important to note
that this would not occur if Yt and Xt were stationary, in which case the distinction
between endogenous and exogenous variables is crucial. For example, if (Yt , Xt ) is
i.i.d. bivariate normal with expectations zero, variances σy2 , σx2 and covariance σxy , the
conditional expectation of Yt given Xt equals σxy /σx2 Xt = βXt and the conditional
expectation of Xt given Yt is σxy /σy2 Yt = β ∗ Yt (see Appendix B). Note that β ∗ =
1/β, unless Yt and Xt are perfectly correlated (σxy = σx σy ). As perfect correlation
also implies that the R 2 equals unity, this also suggests that the R 2 obtained from
a cointegrating regression should be quite high (as it converges to one if the sample
size increases).
Although the existence of a long-run relationship between two variables is of interest,
it may be even more relevant to analyse the short-run properties of the two series. This
can be done using the result that the presence of a cointegrating relationship implies that
there exists an error-correction model that describes the short-run dynamics consistently
with the long-run relationship.

MULTIVARIATE TIME SERIES MODELS

318

9.2.3 Cointegration and Error-correction Mechanisms

The Granger representation theorem (Granger, 1983; Engle and Granger, 1987) states
that if a set of variables are cointegrated, then there exists a valid error-correction
representation of the data. Thus, if Yt and Xt are both I (1) and have a cointegrating
vector (1, −β) , there exists an error-correction representation, with Zt = Yt − βXt , of
the form
θ (L)Yt = δ + φ(L)Xt−1 − γ Zt−1 + α(L)εt ,
(9.23)
where εt is white noise6 and where θ (L), φ(L) and α(L) are polynomials in the lag
operator L (with θ0 ≡ 1). Let us consider a special case of (9.23),
Yt = δ + φ1 Xt−1 − γ (Yt−1 − βXt−1 ) + εt ,

(9.24)

where the error term has no moving average part and the systematic dynamics are kept
as simple as possible. Intuitively, it is clear why the Granger representation theorem
should hold. If Yt and Xt are both I (1) but have a long-run relationship, there must be
some force which pulls the equilibrium error back towards zero. The error-correction
model does exactly this: it describes how Yt and Xt behave in the short-run consistent
with a long-run cointegrating relationship. If the cointegrating parameter β is known,
all terms in (9.24) are I (0) and no inferential problems arise: we can estimate it by
OLS in the usual way.
When Yt = Xt−1 = 0 we obtain the ‘no change’ steady state equilibrium
Yt − βXt =

δ
,
γ

(9.25)

which corresponds with (9.19) if α = δ/γ . In this case the error-correction model can
be written as
Yt = φ1 Xt−1 − γ (Yt−1 − α − βXt−1 ) + εt ,
(9.26)
where the constant is only present in the long-run relationship. If, however, the errorcorrection model (9.24) contains a constant that equals δ = αγ + λ, with λ = 0, this
implies deterministic trends in both Yt and Xt and the long-run equilibrium corresponds
to a steady state growth path with Yt = Xt−1 = λ/(1 − φ1 ). Recall from Chapter 8
that a nonzero intercept in a univariate ARMA model with a unit root also implies that
the series has a deterministic trend.
In some cases it makes sense to assume that the cointegrating vector is known a priori
(e.g. when the only sensible equilibrium is Yt = Xt ). In that case, inferences in (9.23) or
(9.24) can be made in a standard way. If β is unknown, the cointegrating vector can be
estimated (super)
√ consistently from the cointegrating regression (9.20). Consequently,
with standard T asymptotics, one can ignore the fact that β is estimated and apply
conventional theory to the estimation of the parameters in (9.23).
Note that the precise lag structure in (9.23) is not speciﬁed by the theorem, so we
probably need to do some speciﬁcation analysis in this direction. Moreover, the theory
6

The white noise term εt is assumed to be independent of both Yt−1 , Yt−2 , . . . and Xt−1 , Xt−2 , . . . .

ILLUSTRATION: LONG-RUN PURCHASING POWER PARITY (PART 2)

319

is symmetric in its treatment of Yt and Xt , so that there should also exist an errorcorrection representation with Xt as the left-hand side variable. Because at least one
of the variables has to adjust to deviations from the long-run equilibrium, at least one of
the adjustment parameters γ in the two error-correction equations has to be nonzero.
If Xt does not adjust to the equilibrium error (has a zero adjustment parameter), it
is weakly exogenous for β (as deﬁned by Engle, Hendry and Richard, 1983). This
means that we can include Xt in the right-hand side of (9.24) without affecting the
error-correction term −γ (Yt−1 − βXt−1 ). That is, we can condition upon Xt in the
error-correction model for Yt (see Section 9.5 below).
The representation theorem also holds conversely, i.e. if Yt and Xt are both I (1)
and have an error-correction representation, then they are necessarily cointegrated. It is
important to realize that the concept of cointegration can be applied to (nonstationary)
integrated time series only. If Yt and Xt are I (0) the generating process can always be
written in an error-correction form (see Section 9.1).

9.3

Illustration: Long-run Purchasing Power Parity
(Part 2)

In the previous chapter, we introduced the topic of purchasing power parity (PPP),
which requires the exchange rate between two currencies to equal the ratio of the two
countries’ price levels. In logarithms, absolute PPP can be written as
st = pt − pt∗ ,

(9.27)

where st is the log of the spot exchange rate, pt the log of domestic prices and pt∗
the log of foreign prices. Few proponents of PPP would argue for a strict adherence
to PPP. Rather, PPP is usually seen as determining the exchange rate in the long
run, while a variety of other factors, such as trading restrictions, productivity and
preference changes, may inﬂuence the exchange rate in conditions of disequilibrium.
Consequently, (9.27) is viewed as an equilibrium or cointegrating relationship.
Using monthly observations for France and Italy from January 1981 until June 1996,
as before, we are thus looking for a cointegrating relationship between pt , pt∗ and st .
In Section 8.5 we already concluded that nonstationarity of the real exchange rate
rs t ≡ st − pt + pt∗ could not be rejected. This implies that (1, −1, 1) is rejected as a
cointegrating vector. In this section we test whether another cointegrating relationship
exists, initially using only two variables: st , the log exchange rate and ratio t ≡ pt − pt∗ ,
the log of the price ratio. Intuitively, such relationship would imply that a change
in relative prices corresponds to a less than (or more than) proportionate change in
the exchange rate, while imposing symmetry. The corresponding cointegrating regression is
st = α + βratio t + εt ,
(9.28)
where β = 1 corresponds to (9.27). Note that pt and pt∗ are not based on prices but
price indices. Therefore, one may expect that the constant in (9.28) is different from
zero. Consequently, we can only test for relative PPP instead of absolute PPP.

MULTIVARIATE TIME SERIES MODELS

320

The evidence in Section 8.5 suggested that st was I (1). For the log price ratio, ratio t ,
the results of the (augmented) Dickey–Fuller tests are given in Table 9.4. Clearly, we
cannot reject the null hypothesis of a unit root in ratio t , a conclusion which is in line
with what can be inferred from the graph in Figure 8.5.
We are now ready to estimate the cointegrating regression and test for cointegration
between st and pt − pt∗ . First, we estimate (9.28) by ordinary least squares. This gives
the results in Table 9.5. The test for the existence of a cointegrating relationship is
a test for stationarity of the residuals in this regression. We can test for a unit root
in the residuals by means of the CRDW test, based on the Durbin–Watson statistic.
Clearly, the value of 0.055 is not signiﬁcant at any reasonable level of signiﬁcance
and consequently, we cannot reject the null hypothesis of a unit root in the residuals.
Instead of the CRDW test we can also apply the augmented Dickey–Fuller tests, the
results of which are given in Table 9.6. The appropriate 5% critical value is −3.37
(see Table 9.2). Again, the null hypothesis of a unit root cannot be rejected and,
consequently, there is no evidence in the data that the spot exchange rate and the
price ratio are cointegrated. This conclusion corresponds with that in, e.g. Corbae and
Ouliaris (1988), who conclude that there is no long-run tendency for exchange rates
and relative prices to settle down on an equilibrium track.
A potential explanation for this rejection is that the restriction imposed, viz. that
pt and pt∗ enter (9.28) with coefﬁcient β and −β, respectively, is invalid, due to,
Table 9.4 Unit root tests for log price ratio
Italy vs. France
Statistic

Without trend

With trend

DF
ADF(1)
ADF(2)
ADF(3)
ADF(4)
ADF(5)
ADF(6)

−1.563
−0.993
−1.003
−1.058
−1.014
−1.294
−2.015

−2.692
−2.960
−2.678
−3.130
−2.562
−2.493
−3.096

Table 9.5

OLS results

Dependent variable: st (log exchange rate)
Variable
constant
ratio t = pt − pt∗

Estimate

Standard error

t-ratio

5.4872
0.9822

0.00678
0.05133

809.706
19.136

s = 0.0860 R 2 = 0.6656 R̄ 2 = 0.6638 F = 366.191
dw = 0.0555 T = 186
Table 9.6 ADF (cointegration) tests of residuals
DF
ADF(1)
ADF(2)
ADF(3)

−1.904
−1.850
−1.896
−1.952

ADF(4)
ADF(5)
ADF(6)

−1.910
−1.946
−2.249

VECTOR AUTOREGRESSIVE MODELS

321

Table 9.7 OLS results
Dependent variable: st (log exchange rate)
Variable

Estimate

Standard error

t-ratio

constant
pt
pt∗

12.5092
3.0964
−4.6291

0.5170
0.1508
0.2710

24.194
19.372
−17.085

s = 0.0609 R 2 = 0.8335 R̄ 2 = 0.8316 F = 357.902
dw = 0.1525 T = 186
Table 9.8 ADF (cointegration) tests of residuals
DF
ADF(1)
ADF(2)
ADF(3)

−2.806
−3.159
−2.964
−2.872

ADF(4)
ADF(5)
ADF(6)

−2.863
−2.923
−2.918

for example, transportation costs or measurement error. We can estimate (9.28) with
unconstrained coefﬁcients, so that we can test the existence of a more general cointegrating relationship between the three variables, st , pt and pt∗ . However, when we
consider more than two-dimensional systems, the number of cointegrating relationships may be more than one. For example, there may be two different cointegrating
relationships between three I (1) variables, which makes the analysis somewhat more
complicated than in the two-dimensional case. Section 9.5 will pay attention to the
more general case.
When there exists only one cointegrating vector, we can estimate the cointegrating
relationship, as before, by regressing one variable upon the other variables. This does
require, however, that the cointegrating vector involves the left-hand side variable of
this regression, because its coefﬁcient is implicitly normalized to minus one. In our
example, we regress st upon pt and pt∗ to obtain the results reported in Table 9.7. The
ADF tests on the residuals produce the results in Table 9.8, where the appropriate 5%
critical value is −3.77 (see Table 9.2). Again, we have to conclude that we cannot
reject the null hypothesis that there is no cointegrating relationship between the log
exchange rate and the log price indices of France and Italy. It does not seem to be the
case that some (weak) form of purchasing power parity holds for these two countries.
Of course, it could be the case that our sample period is just not long enough to
ﬁnd sufﬁcient evidence for a cointegrating relationship. This seems to be in line with
what people ﬁnd in the literature. With longer samples, up to a century or more, the
evidence is more in favour of some long-run tendency to PPP (see the survey in Froot
and Rogoff, 1994).

9.4

Vector Autoregressive Models

The autoregressive moving average models of the previous chapter can be readily
extended to the multivariate case, in which the stochastic process that generates the time
series of a vector of variables is modelled. The most common approach is to consider

MULTIVARIATE TIME SERIES MODELS

322

a vector autoregressive (VAR) model. A VAR describes the dynamic evolution of a
number of variables from their common history. If we consider two variables, Yt and
Xt , say, the VAR consists of two equations. A ﬁrst order VAR would be given by
Yt = δ1 + θ11 Yt−1 + θ12 Xt−1 + ε1t

(9.29)

Xt = δ2 + θ21 Yt−1 + θ22 Xt−1 + ε2t ,

(9.30)

where ε1t and ε2t are two white noise processes (independent of the history of Y and
X) that may be correlated. If, for example, θ12 = 0 it means that the history of X helps
explaining Y . The system (9.29)–(9.30) can be written as

  
    
δ1
θ11 θ12
Yt−1
ε1t
Yt
=
+
+
(9.31)
Xt
δ2
θ21 θ22
Xt−1
ε2t
or, with appropriate deﬁnitions, as
Yt = δ +



1 Yt−1

+ εt ,

(9.32)

where Yt = (Yt , Xt ) and εt = (ε1t , ε2t ) . This extends the ﬁrst order autoregressive
model from Chapter 8 to the more dimensional case. In general, a VAR(p) model for
a k-dimensional vector Yt is given by
Yt = δ +



1 Yt−1

+··· +



p Yt−p

+ εt ,

(9.33)

where each j is a k × k matrix and εt is a k-dimensional vector of white noise terms
with covariance matrix . As in the univariate case, we can use the lag operator to
deﬁne a matrix lag polynomial
(L) = Ik −

1L

− ··· −

p
pL ,

where Ik is the k-dimensional identity matrix, so that we can write the VAR as
(L)Yt = δ + εt .
The matrix lag polynomial is a k × k matrix where each element corresponds to a
p-th order polynomial in L. Extensions to vectorial ARMA (VARMA) models can be
obtained by premultiplying εt with a (matrix) lag polynomial.
The VAR model implies univariate ARMA models for each of its components. The
advantages of considering the components simultaneously include that the model may
be more parsimonious and includes fewer lags, and that more accurate forecasting is
possible, because the information set is extended to also include the history of the other
variables. From a different perspective, Sims (1980) has advocated the use of VAR
models instead of structural simultaneous equations models because the distinction
between endogenous and exogenous variables does not have to be made a priori, and
‘arbitrary’ constraints to ensure identiﬁcation are not required (see, e.g. Canova, 1995,
for a discussion). Like a reduced form a VAR is always identiﬁed.

VECTOR AUTOREGRESSIVE MODELS

323

The expected value of Yt can be determined if we impose stationarity. This gives
 + ··· +

E{Yt } = δ +
or



1 E{Yt }

µ = E{Yt } = (I −

1

−··· −

p E{Yt }

−1
p) δ

=

(1)−1 δ,

which shows that stationarity will require that the k × k matrix (1) is invertible.7
For the moment we shall assume that this is the case. As before, we can subtract the
mean and consider yt = Yt − µ, for which we have that
yt =

t−1
1y

+ ··· +

t−p
py

+ εt .

(9.34)

We can use the VAR model for forecasting in a straightforward way. For forecasting
from the end of the sample period (period T ), the relevant information set now includes
the vectors yT , yT −1 , . . . , and we obtain for the optimal one-period ahead forecast
yT +1|T = E{
yT +1 |
yT , yT −1 , . . .} =

T
1y

+··· +

T −p+1 .
py

(9.35)

The one-period ahead forecast error variance is simply V {
yT +1 |
yT , yT −1 , . . .} = .
Forecasts more than one period ahead can be obtained recursively. For example,
yT +2|T =
=

T +1|T
1y

+··· +

T −p+2
py

1(

+ ··· +

T −p+1 )
Py

T
1y

+ ··· +

T −p+2 .
py

(9.36)

To estimate a vector autoregressive model we can simply use ordinary least squares
equation by equation,8 which is consistent because the white noise terms are assumed
to be independent of the history of yt . From the residuals of each of the k equations,
e1t , . . . , ekt , we can estimate the (i, j )-element in  as9
σ̂ij =
so that
ˆ =


T

1
e e ,
T − p t=p+1 it j t

(9.37)

T

1
e e ,
T − p t=p+1 t t

(9.38)

where et = (e1t , . . . , ekt ) .
Recall from Chapter 8 that in the AR(p) case stationarity requires that θ (1) = 0, so that θ (1)−1 exists.
Because the explanatory variables are the same for each equation, a system estimator, like SUR (see
Greene, 2003, Section 14.2), provides the same estimates as OLS applied to each equation separately. If
different restrictions are imposed upon the equations, SUR estimation will be more efﬁcient than OLS,
though OLS remains consistent.
9
Assuming that observations are available from t = 1, . . . , T , the number of useful observations is T − p.
Note that a degrees of freedom correction can be applied, as in the linear regression model (see Chapter 2).
7
8

324

MULTIVARIATE TIME SERIES MODELS

Determining the lag length p in an empirical application is not always easy and
univariate autocorrelation or partial autocorrelation functions will not help; see Canova
(1995) for a discussion. A reasonable strategy is to estimate a VAR model for different
values of p and then select on the basis of the Akaike or Schwarz information criteria,
as discussed in Chapters 3 and 8, or on the basis of statistical signiﬁcance.
If (1) is invertible, it means that we can write the vector autoregressive model as a
vector moving average (VMA) model by premultiplying with (L)−1 . This is similar
to deriving the moving average representation of a univariate autoregressive model.
This gives
Yt = (1)−1 δ + (L)−1 εt = µ + (L)−1 εt ,
(9.39)
which describes each element in Yt as a weighted sum of all current and past shocks
in the system. Writing (L)−1 = Ik + A1 L + A2 L2 + · · ·, we have that
Yt = µ + εt + A1 εt−1 + A2 εt−2 + · · ·

(9.40)

If the white noise vector εt increases by a vector d, the effect upon Yt+s (s > 0) is
given by As d. Thus the matrix
∂ Y
As = t+s
(9.41)
∂εt
has the interpretation that its (i, j )-element measures the effect of a one-unit increase
in εj t upon Yi,t+s . If only the ﬁrst element ε1t of εt changes, the effects are given by
the ﬁrst column of As . The dynamic effects upon the j -th variable of such a one-unit
increase are given by the elements in the ﬁrst column and j -th row of Ik , A1 , A2 , . . . .
A plot of these elements as a function of s is called the impulse-response function. It
measures the response of Yj,t+s to an impulse in Y1t , keeping constant all other variables
dated t and before. Although it may be hard to derive expressions for the elements in
(L)−1 , the impulse-responses can be determined fairly easily by simulation methods
(see Hamilton, 1994).
If (1) is not invertible it cannot be the case that all variables in Yt are stationary
I (0) series. At least one stochastic trend must be present. In the extreme case where
we have k independent stochastic trends, all k variables are integrated of order one,
while no cointegrating relationships exist. In this case (1) is equal to a null matrix.
The intermediate cases are more interesting: the rank of the matrix (1) equals the
number of linear combinations of variables in Yt that are I (0), that is determines the
number of cointegrating vectors. This is the topic of the next section.

9.5

Cointegration: the Multivariate Case

When more than two variables are involved, cointegration analysis is somewhat more
complex, because the cointegrating vector generalizes to a cointegrating space, the
dimension of which is not known a priori. That is, when we have a set of k I (1)
variables, there may exist up to k − 1 independent linear relationships that are I (0),
while any linear combination of these relationships is – by construction – also I (0).
This implies that individual cointegrating vectors are no longer statistically identiﬁed;

COINTEGRATION: THE MULTIVARIATE CASE

325

only the space spanned by these vectors is. Ideally, vectors in the cointegrating space
can be found that have an economic interpretation and can be interpreted as representing
long-run equilibria.
9.5.1 Cointegration in a VAR

If the variables of interest are stacked in the k-dimensional vector Yt , the elements of
which are assumed to be I (1), there may be different vectors β such that Zt = β  Yt is
I (0). That is, there may be more than one cointegrating vector β. It is clearly possible
for several equilibrium relations to govern the long-run behaviour of the k variables.
In general, there can be r ≤ k − 1 linearly independent cointegrating vectors,10 which
are gathered together into the k × r cointegrating matrix11 β. By construction, the
rank of the matrix12 β is r, which will be called the cointegrating rank of Yt . This
 t = β  Yt is I (0), while each
means that each element in the r-dimensional vector Z

element in the k-dimensional vector Yt is I (1).
The Granger representation theorem (Engle and Granger, 1987) directly extends
to this more general case and claims that if Yt is cointegrated, there exists a valid
error-correction representation of the data. While there are different ways to derive
and describe such a representation, we shall here start from the vector autoregressive
model for Yt introduced in the previous section:
Yt = δ +
or



1 Yt−1



p Yt−p

+··· +

+ εt

(9.42)

(L)Yt = δ + εt .

(9.43)

For the case with p = 3 we can write this as

or

Yt = δ + (

1

+

2

− Ik )Yt−1 −

=δ+(

1

+

2

+

3



2 Yt−1

− Ik )Yt−1 −

Yt = δ + 1 Yt−1 + 2 Yt−2 + (



+ εt



−

3 (Yt−1

+

2

+

3 Yt−3

2 Yt−1

1

+



3

+ Yt−2 ) + εt ,

− Ik )Yt−1 + εt ,

where 1 = − 2 − 3 and 2 = − 3 . Similarly, we can write for general values of
p that13
Yt = δ + 1 Yt−1 + · · · + p−1 Yt−p+1 + Yt−1 + εt ,
(9.44)
The existence of k cointegrating relationships between the k elements in Yt would imply that there exist
k independent linear combinations that are I (0), such that, necessarily, all individual elements in Yt must
be I (0). Clearly, this is in conﬂict with the deﬁnition of cointegration as a property of I (1) variables, and
it follows that r ≤ k − 1.
11
We follow the convention in the cointegration literature to denote the cointegrating matrix by a Greek
lower case β.
12
See Appendix A for the deﬁnition of the rank of a matrix.
13
It is possible to rewrite the VAR such that any of the lags appear in levels on the right-hand side, with
the same coefﬁcients in the ‘long-run matrix’ . For comparison with the univariate case, we choose for
inclusion of the ﬁrst lag.
10

MULTIVARIATE TIME SERIES MODELS

326

where the ‘long-run matrix’
 ≡ − (1) = −(Ik −

1

− ··· −

p)

(9.45)

determines the long-run dynamic properties of Yt .14 This equation is a direct generalization of the regressions used in the augmented Dickey–Fuller test. Because Yt
and εt are stationary (by assumption), it must be the case that Yt−1 in (9.44) is
also stationary. This could reﬂect three different situations. First, if all elements in
Yt are integrated of order one and no cointegrating relationships exist, it must be the
case that  = 0 and (9.44) presents a (stationary) VAR model for Yt . Second, if
all elements in Yt are stationary I (0) variables, the matrix  = − (1) must be of
full rank and invertible so that we can write a vector moving average representation
Yt = −1 (L)(δ + εt ). Third, if  is of rank r (0 < r < k) the elements in Yt−1 are
linear combinations that are stationary. If the variables in Yt are I (1), these linear
combinations must correspond to cointegrating vectors. This intermediate case is the
most interesting one. If  has a reduced rank of r ≤ k − 1, this means that there are
r independent linear combinations of the k elements in Yt that are stationary, that is:
there exist r cointegrating relationships. Note that the existence of k cointegrating relationships is impossible: if k independent linear combinations produce stationary series,
all k variables themselves must be stationary.
If  has reduced rank it can be written as the product of a k × r matrix γ and an
r × k matrix β  that both have rank r.15 That is,  = γβ  . Substituting this produces
the model in error-correction form
Yt = δ + 1 Yt−1 + · · · + p−1 Yt−p+1 + γβ  Yt−1 + εt .

(9.46)

The linear combinations β  Yt−1 present the r cointegrating relationships. The coefﬁcients in γ measure how the elements in Yt are adjusted to the r ‘equilibrium errors’
 t−1 = β  Yt−1 . Thus, (9.46) is a generalization of (9.24) and is referred to as a vector
Z
error-correction model (VECM).
If we take expectations in the error-correction model we can derive that
 t−1 }.
(I − 1 − · · · − p−1 )E{Yt } = δ + γ E{Z

(9.47)

There is no deterministic trend in any of the variables if E{Yt } = 0. Under the
assumption that the matrix (I − 1 − · · · − p−1 ) is nonsingular, this requires that δ +
 t−1 } corresponds to the vector
 t−1 } = 0 (compare Subsection 9.2.3), where E{Z
γ E{Z
of intercepts in the cointegrating relations. If we impose this restriction, intercepts
appear in the cointegrating relationships only, and we can rewrite the error-correction
 t−1 − E{Z
 t−1 } and have no intercepts, i.e.
model to include zt = Z
Yt = 1 Yt−1 + · · · + p−1 Yt−p+1 + γ (−α + β  Yt−1 ) + εt ,
In the univariate case, the long-run properties are determined by θ (1), where θ (L) is the AR polynomial
(see Chapter 8).
15
This means that the r columns in γ are linearly independent, and that the r rows in β  are independent
(see Appendix A).
14

COINTEGRATION: THE MULTIVARIATE CASE

327

 t−1 } =
where α is an r-dimensional vector of constants, satisfying E{β  Yt−1 } = E{Z
α. As a result, all terms in this expression have mean zero and no deterministic
trends exist.
If we add one common constant to the vector error-correction model, we obtain
Yt = λ + 1 Yt−1 + · · · + p−1 Yt−p+1 + γ (−α + β  Yt−1 ) + εt ,
where λ is a k-dimensional vector with identical elements λ1 . Now the long-run equilibrium corresponds to a steady state growth path with growth rates for all variables
given by
E{Yt } = (I − 1 − · · · − p−1 )−1 λ.
The deterministic trends in each Yj t are assumed to cancel out in the long run, so
that no deterministic trend is included in the error-correction term. We can go as far
as allowing for k − r separate deterministic trends that cancel out in the cointegrating
relationships, in which case we are back at speciﬁcation (9.46) without restrictions
on δ. In this case, δ is capturing r intercept terms in the long-run relationships and
k − r different deterministic trends in the variables in Yt . If there would be more than
k − r separate deterministic trends, they cannot cancel out in β  Yt−1 and we should
include a deterministic trend in the cointegrating equations. See Harris (1995, p. 96)
for additional discussion and some alternatives.

9.5.2 Example: Cointegration in a Bivariate VAR

As an example, consider the case where k = 2. In this case the number of cointegrating
vectors may be zero or one (r = 0, 1). Let us consider a ﬁrst order (nonstationary) VAR
for Yt = (Yt , Xt ) . That is,


Yt
Xt




=

θ11

θ12

θ21

θ22



Yt−1




+

Xt−1

ε1t


,

ε2t

where, for simplicity, we did not include intercept terms. The matrix  is given by

 = − (1) =

θ11 − 1

θ12

θ21

θ22 − 1


.

This matrix is a zero matrix if θ11 = θ22 = 1 and θ12 = θ21 = 0. This corresponds to
the case where Yt and Xt are two random walks. The matrix  has reduced rank if
(θ11 − 1)(θ22 − 1) − θ21 θ12 = 0.
If this is the case,



β  = θ11 − 1 θ12

(9.48)

MULTIVARIATE TIME SERIES MODELS

328

is a cointegrating vector (where we chose an arbitrary normalization) and we can write




1
θ11 − 1 θ12 .
 = γβ  =
θ21 /(θ11 − 1)
Using this, we can write the model in error-correction form. First, write
 
  

  
θ12
Yt−1
θ11 − 1
ε1t
Yt−1
Yt
=
+
+
.
Xt
Xt−1
θ21
θ22 − 1
Xt−1
ε2t
Next, we rewrite this as

 
 

Yt
1
ε1t
=
.
((θ11 − 1)Yt−1 + θ12 Xt−1 ) +
θ21 /(θ11 − 1)
Xt
ε2t

(9.49)

The error-correction form is thus quite simple, as it excludes any dynamics. Note that
both Yt and Xt adjust to the equilibrium error, because θ21 = 0 is excluded. (Also note
that θ21 = 0 would imply θ11 = θ22 = 1 and no cointegration.)
The fact that the linear combination Zt = (θ11 − 1)Yt + θ12 Xt is I (0) also follows
from this result. Note that we can write


 
1
ε1t
Zt = ( θ11 − 1 θ12 )
Zt−1 + ( θ11 − 1 θ12 )
ε2t
θ21 /(θ11 − 1)
or (using (9.48)):
Zt = Zt−1 + (θ11 − 1 + θ22 − 1)Zt−1 + vt = (θ11 + θ22 − 1)Zt−1 + vt ,
where vt = (θ11 − 1)ε1t + θ12 ε2t is a white noise error term. Consequently, Zt
is described by a stationary AR(1) process unless θ11 = 1 and θ22 = 1, which
was excluded.
9.5.3 Testing for Cointegration
If it is known that there exists at most one cointegrating vector, a simple approach
to test for the existence of cointegration is the Engle–Granger approach described in
Section 9.2.2. It requires running a regression of Y1t (being the ﬁrst element of Yt ) on
the other k − 1 variables Y2t , . . . , Ykt and testing for a unit root in the residuals. This can
be done using the ADF tests on the OLS residuals applying the critical values from
Table 9.2. If the unit root hypothesis is rejected, the hypothesis of no-cointegration
is also rejected. In this case, the static regression gives consistent estimates of the
cointegrating vector, while in a second stage, the error-correction model can be estimated using the estimated cointegrating vector from the ﬁrst stage.
There are some problems with this Engle–Granger approach. First, the results of the
tests are sensitive to the left-hand side variable of the regression, that is, to the normalization applied to the cointegrating vector. Second, if the cointegrating vector happens

COINTEGRATION: THE MULTIVARIATE CASE

329

not to involve Y1t but only Y2t , . . . , Ykt , the test is not appropriate and the cointegrating
vector will not be consistently estimated by a regression of Y1t upon Y2t , . . . , Ykt . Third,
the residual-based test tends to lack power because it does not exploit all the available information about the dynamic interactions of the variables. Fourth, it is possible
that more than one cointegrating relationship exists between the variables Y1t , . . . , Ykt .
If, for example, two distinct cointegrating relationships exist, OLS typically estimates
a linear combination of them (see Hamilton, 1994, p. 590). Fortunately, as the null
hypothesis for the cointegration tests is that there is no cointegration, the tests are still
appropriate for their purpose.
An alternative approach that does not suffer from these drawbacks was proposed by
Johansen (1988), who developed a maximum likelihood estimation procedure, which
also allows one to test for the number of cointegrating relations. The details of the
Johansen procedure are very complex and we shall only focus on a few aspects.
Further details can be found in Johansen and Juselius (1990) and Johansen (1991),
or in textbooks like Banerjee et al. (1993, Chapter 8); Hamilton (1994, Chapter 20);
Johansen (1995, Chapter 11) and Stewart and Gill (1998, Sections 9.4 and 9.5). The
starting point of the Johansen procedure is the VAR representation of Yt given in (9.44)
and reproduced here:
Yt = δ + 1 Yt−1 + · · · + p−1 Yt−p+1 + Yt−1 + εt ,

(9.50)

where εt is NID(0, ). Note that the use of maximum likelihood requires us to impose
a particular distribution for the white noise terms. Assuming that Yt is a vector of I (1)
variables, while r linear combinations of Yt are stationary, we can write
 = γβ  ,

(9.51)

where, as before, γ and β are of dimension k × r. Again, β denotes the matrix
of cointegrating vectors, while γ represents the matrix of weights with which each
cointegrating vector enters each of the Yt equations. The approach of Johansen is
based on the estimation of the system (9.50) by maximum likelihood, while imposing
the restriction in (9.51) for a given value of r.
The ﬁrst step in the Johansen approach involves testing hypotheses about the rank
of the long-run matrix , or – equivalently – the number of columns in β. For a
given r, it can be shown (see, e.g. Hamilton, 1994, Section 20.2) that the ML estimate for β equals the matrix containing the r eigenvectors corresponding to the r
largest (estimated) eigenvalues of a k × k matrix that can be estimated fairly easily using an OLS package. Let us denote the (theoretical) eigenvalues of this matrix
in decreasing order as λ1 ≥ λ2 ≥ · · · ≥ λk . If there are r cointegrating relationships
(and  has rank r) it must be the case that log(1 − λj ) = 0 for the smallest k − r
eigenvalues, that is, for j = r + 1, r + 2, . . . , k. We can use the (estimated) eigenvalues, say λ̂1 > λ̂2 > · · · > λ̂k , to test hypotheses about the rank of . For example,
the hypothesis H0 : r ≤ r0 versus the alternative H1 : r0 < r ≤ k, can be tested using
the statistic
k

λtrace (r0 ) = −T
log(1 − λ̂j ).
(9.52)
j =r0 +1

MULTIVARIATE TIME SERIES MODELS

330

This test is the so-called trace test. It checks whether the smallest k − r0 eigenvalues
are signiﬁcantly different from zero. Furthermore, we can test H0 : r ≤ r0 versus the
more restrictive alternative H1 : r = r0 + 1 using
λmax (r0 ) = −T log(1 − λ̂r0 +1 ).

(9.53)

This alternative test is called the maximum eigenvalue test, as it is based on the
estimated (r0 + 1)th largest eigenvalue.
The two tests described here are actually likelihood ratio tests (see Chapter 6), but
do not have the usual Chi-squared distributions. Instead, the appropriate distributions
are multivariate extensions of the Dickey–Fuller distributions. As with the unit root
tests, the percentiles of the distributions depend on the fact whether a constant (and
a time trend) are included. Critical values for two cases are presented in Table 9.9.
Case 1 assumes that there are no deterministic trends and includes r intercepts in
the cointegrating relationships. Case 2 is based on the inclusion of k unrestricted
intercepts in the VAR, which implies k − r separate deterministic trends and r intercepts
in the cointegration vectors. The critical values depend upon k − r0 , the number of
nonstationary components under the null hypothesis. Note that when k − r0 = 1 the
two test statistics are identical and thus have the same distribution.
There are many studies that show that the small sample properties of the test statistics
in (9.52) and (9.53) differ substantially from the asymptotic properties. As a result,
the tests are biased towards ﬁnding cointegration too often when asymptotic critical
values are used (see Chueng and Lai, 1993). A small sample correction, which is now
commonly used, was suggested by Ahn and Reinsel (1990) and Reimers (1991), and
implies that the test statistics are multiplied by a factor (T − pk)/T , where p denotes
the number of lags in the VAR model. A more accurate correction factor is derived in
Johansen (2002).
It is important to realize that the parameters γ and β are not uniquely identiﬁed in
the sense that different combinations of γ and β can produce the same matrix  = γβ  .
Table 9.9 Critical values Johansen’s LR tests for cointegration (Pesaran,
Shin and Smith, 2000)
λtrace -statistic
H0 : r ≤ r0 vs H1 : r > r0
k − r0

5%

10%

λmax -statistic
H0 : r ≤ r0 vs H1 : r = r0 + 1
5%

10%

Case 1: restricted intercepts in VAR (in cointegrating relations only)
1
2
3
4
5

9.16
20.18
34.87
53.48
75.98

7.53
17.88
31.93
49.95
71.81

9.16
15.87
22.04
28.27
34.40

7.53
13.81
19.86
25.80
31.73

8.07
14.88
21.12
27.42
33.64

6.50
12.98
19.02
24.99
31.02

Case 2: unrestricted intercepts in VAR
1
2
3
4
5

8.07
17.86
31.54
48.88
70.49

6.50
15.75
28.78
45.70
66.23

COINTEGRATION: THE MULTIVARIATE CASE

331

This is because γβ  = γ P P −1 β  for any invertible r × r matrix P . In other words,
what the data can determine is the space spanned by the columns of β, the cointegration
space, and the space spanned by γ . Consequently, the cointegrating vectors in β have
to be normalized in some way to obtain unique cointegrating relationships. Often, it
is hoped that these relationships are so-called ‘structural’ cointegrating relationships
that have a sensible economic interpretation. In general, it may not be possible to
statistically identify these structural cointegrating relationships from the estimated β
matrix; see Davidson (2000, Section 16.6) for a discussion.
9.5.4 Illustration: Long-run Purchasing Power Parity (Part 3)

In this section, we reconsider the above example concerning long-run purchasing
power parity. We shall analyse the existence of one or more cointegrating relationships between the three variables st , pt and pt∗ , using Johansen’s technique described
in the previous section. This is a standard option available in, for example, MicroFit.
The ﬁrst step in this procedure is the determination of p, the maximum order of the
lags in the autoregressive representation given in (9.42). It appears that, in general,
too few lags in the model lead to rejection of the null hypotheses too easily, while too
many lags in the model decrease the power of the tests. This indicates that there is
some optimal lag length. In addition to p, we have to decide upon whether to include a
time trend in (9.42) or not. In the absence of a time trend, an intercept is automatically
included in the cointegrating relationship(s). Let us, more or less arbitrarily, consider
the case with p = 3 excluding a time trend. The ﬁrst step in Johansen’s procedure yields
the results16 in Table 9.10. These results present the estimated eigenvalues λ̂1 , . . . , λ̂k
(k = 3) in descending order. Recall that each nonzero eigenvalue corresponds to a
cointegrating vector. A range of test statistics based on these estimated eigenvalues is
given as well. These results indicate that:
1. The null hypothesis of no cointegration (r = 0) has to be rejected at a 5% level,
when tested against the hypothesis of one cointegrating vector (r = 1), because
65.5 exceeds the critical value of 22.04.
2. The null hypothesis of zero or one cointegrating vector (r ≤ 1) also has to be
rejected against the alternative of two cointegrating relationships (r = 2).
3. The null hypothesis of two or fewer cointegrating vectors cannot be rejected against
the alternative of r = 3. Recall that r = 3 corresponds to stationarity of each of the
three series, which was also rejected by the univariate unit root tests.
Table 9.10 Maximum eigenvalue tests for cointegration
Null hypothesis

Alternative

λmax -statistic

5% critical value

H0 : r = 0
H0 : r ≤ 1
H0 : r ≤ 2

H1 : r = 1
H1 : r = 2
H1 : r = 3

65.509
22.032
6.371

22.04
15.87
9.16

lag length p = 3 intercepts included T = 183
Estimated eigenvalues: 0.3009, 0.1134, 0.0342
16

The results in this subsection are obtained by MicroFit 4.0, Oxford University Press.

MULTIVARIATE TIME SERIES MODELS

332

From these results we have to choose the number of cointegrating vectors. Given our
previous results it is somewhat surprising that Johansen’s tests seem to indicate the
presence of two cointegrating relationships. In the ﬁrst Engle–Granger steps, we could
not reject no-cointegration in any of the cases we considered. A possible explanation
for this ﬁnding may be that the number of lags in the VAR is too small. Similar to
what we found before with the univariate unit root tests on pt and pt∗ , the inclusion of
too few lags may lead to the wrong conclusion that the series are stationary, or – in
this case – are cointegrated.17 Table 9.11 shows what happens if we repeat the above
procedure with a lag length of p = 12, motivated by the fact that we have monthly data.
What is quite clear from these results is that the evidence in favour of one or two
cointegrating vectors is much weaker than before. The ﬁrst test that considers the null
hypothesis of no cointegration (r = 0) versus the alternative of one cointegrating relationship (r = 1) does not lead to rejection of the null. The second test though, implies
a marginal rejection of the hypothesis of the existence of zero or one cointegrating
vector. Suppose we continue our analysis despite our reservations, while we decide
that the number of cointegrating vectors is equal to one (r = 1). The next part of the
results consists of the estimated cointegrating vector β, presented in Table 9.12. The
normalized cointegrating vector is given in the third column and corresponds to
st = 6.347pt − 14.755pt∗ ,

(9.54)

which does not seem to correspond to an economically interpretable long-run relationship.
As the conclusion that there exists one cointegrating relationship between our three
variables is most probably incorrect, we do not pursue this example any further. To
appropriately test for long-run purchasing power parity via the Johansen procedure,
we will probably need longer time series. Alternatively, some authors use several
Table 9.11

Maximum eigenvalue tests for cointegration

Null hypothesis

Alternative

λmax -statistic

5% critical value

H0 : r = 0
H0 : r ≤ 1
H0 : r ≤ 2

H1 : r = 1
H1 : r = 2
H1 : r = 3

19.521
16.437
6.180

22.04
15.87
9.16

lag length p = 12 intercepts included T = 174
Estimated eigenvalues: 0.1061, 0.0901, 0.0349
Table 9.12

Johansen estimation results

Estimated cointegrating vector
Variable
st
pt
pt∗

Normalized
−0.092
0.583
−1.354

−1.000
6.347
−14.755

Based on VAR with p = 12
17

Note, for example, that the ‘cointegrating’ vector (0, 0, 1) corresponds to stationarity of the last element.

ILLUSTRATION: MONEY DEMAND AND INFLATION

333

sets of countries simultaneously and apply panel data cointegration techniques (see
Chapter 10). Another problem may lie in measuring the two price indices in an accurate
way, comparable across the two countries.

9.6

Illustration: Money Demand and Inﬂation

One of the advantages of cointegration in multivariate time series models is that it
may help improving forecasts. The reason is that forecasts from a cointegrated system
are tied together by virtue of the existence of one or more long-run relationships.
Typically, this advantage is realized when forecasting over medium or long horizons (compare Engle and Yoo, 1987). Hoffman and Rasche (1996) and Lin and Tsay
(1996) empirically examine the forecast performance in a cointegrated system. In this
section, based on the Hoffman and Rasche study, we consider an empirical example
concerning a ﬁve-dimensional vector process. The empirical work is based on quarterly data for the United States from 1954:1 to 1994:4 (T = 164) for the following
variables:18
mt : log of real M1 money balances
inﬂ t : quarterly inﬂation rate (in % per year)
cpr t : commercial paper rate
yt : log real GDP (in billions of 1987 dollars)
tbr t : treasury bill rate
The commercial paper rate and the treasury bill rate are considered as risky and riskfree returns on a quarterly horizon, respectively. The series for M1 and GDP are
seasonally adjusted. Although one may dispute the presence of a unit root in some of
these series, we shall follow Hoffman and Rasche (1996) and assume that these ﬁve
variables are all well described by an I (1) process.
A priori one could think of three possible cointegrating relationships governing the
long-run behaviour of these variables. First, we can specify an equation for money
demand as
mt = α1 + β14 yt + β15 tbr t + ε1t ,
where β14 denotes the income elasticity and β15 the interest rate elasticity. It can be
expected that β14 is close to unity, corresponding to a unitary income elasticity, and
that β15 < 0. Second, if real interest rates are stationary we can expect that
inﬂ t = α2 + β25 tbr t + ε2t
corresponds to a cointegrating relationship with β25 = 1. This is referred to as the
Fisher relation, where we are using actual inﬂation as a proxy for expected inﬂation.19
Third, it can be expected that the risk premium, as measured by the difference between
18
19

The data are available in the ﬁles MONEY.
The real interest rate is deﬁned as the nominal interest rate minus the expected inﬂation rate.

MULTIVARIATE TIME SERIES MODELS

334

Table 9.13 Univariate cointegrating regressions by OLS (standard errors
in parentheses), intercept estimates not reported
Money demand
mt
inﬂ t
cpr t
yt
tbr t
R2
dw
ADF (6)

−1
0
0
0.423
(0.016)
−0.031
(0.002)
0.815
0.199
−3.164

Fisher equation
0
−1
0
0
0.558
(0.053)
0.409
0.784
−1.888

Risk premium
0
0
−1
0
1.038
(0.010)
0.984
0.705
−3.975

the commercial paper rate and the treasury bill rate, is stationary, so that a third
cointegrating relationship is given by
cpr t = α3 + β35 tbr t + ε3t
with β35 = 1.
Before proceeding to the vector process of these ﬁve variables, let us consider the
OLS estimates of the above three regressions. These are presented in Table 9.13. To
ease comparison with later results the layout stresses that the left-hand side variables
are included in the cointegrating vector (if it exists) with a coefﬁcient of −1. Note
that the OLS standard errors are inappropriate if the variables in the regression are
integrated. Except for the risk premium equation, the R 2 s are not close to unity, which
is an informal requirement for a cointegrating regression. The Durbin–Watson statistics
are small and if the critical values from Table 9.3 are appropriate, we would reject the
null hypothesis of no cointegration at the 5% level for the last two equations but not for
the money demand equation. Recall that the critical values in Table 9.3 are based on
the assumption that all series are random walks, which may be correct for interest rate
series but may be incorrect for money supply and GDP. Alternatively, we can test for
a unit root in the residuals of these regressions by the augmented Dickey–Fuller tests.
The results are not very sensitive to the number of lags that is included and the test
statistics for 6 lags are reported in Table 9.13. The 5% asymptotic critical value from
Table 9.2 is given by −3.77 for the regression involving three variables and −3.37 for
the regressions with two variables. Only for the risk premium equation we can thus
reject the null hypothesis of no cointegration.
The empirical evidence for the existence of the suggested cointegrating relationships
between the ﬁve variables is somewhat mixed. Only for the risk premium equation we
ﬁnd an R 2 close to unity, a sufﬁciently high Durbin–Watson statistic and a signiﬁcant
rejection of the ADF test for a unit root in the residuals. For the two other regressions
there is little reason to reject the null hypothesis of no cointegration. Potentially this is
caused by the lack of power of the tests that we employ, and it is possible that a multivariate vector analysis provides stronger evidence for the existence of cointegrating
relationships between these ﬁve variables. Some additional information is provided
if we plot the residuals from these three regressions. If the regressions correspond

ILLUSTRATION: MONEY DEMAND AND INFLATION

335

0.15
0.10
0.05
0.00
−0.05
−0.10
−0.15
55

60

65

70

75

80

85

90

Figure 9.1 Residuals of money demand regression

8
6
4
2
0
−2
−4
55

60

65

Figure 9.2

70

75

80

85

90

Residuals of Fisher regression

to cointegration these residuals can be interpreted as long-run equilibrium errors and
should be stationary and ﬂuctuating around zero. For the three regressions, the residuals are displayed in Figures 9.1, 9.2 and 9.3, respectively. Although a visual inspection
of these graphs is ambiguous, the residuals of the money demand and risk premium
regressions could be argued to be stationary on the basis of these graphs. For the Fisher
equation, the current sample period provides less evidence of mean reversion.
The ﬁrst step in the Johansen approach involves testing for the cointegrating rank
r. To compute these tests we need to choose the maximum lag length p in the vector
autoregressive model. Choosing p too small will invalidate the tests and choosing

MULTIVARIATE TIME SERIES MODELS

336

3

2

1

0

−1
55

60

65

Figure 9.3

70

75

80

85

90

Residuals of risk premium regression

p too large may result in a loss of power. In Table 9.14 we present the results20
of the cointegrating rank tests for p = 5 and p = 6. The results show that there is
some sensitivity with respect to the choice of the maximum lag length in the vector
autoregressions, although qualitatively the conclusion changes only marginally. At the
5% level all tests reject the null hypotheses of none or one cointegrating relationship.
The tests for the null hypothesis that r = 2 only reject at the 5% level, albeit marginally,
if we choose p = 6 and use the trace test statistic. As before, we need to choose the
cointegrating rank r from these results. The most obvious choice is r = 2, although
one could consider r = 3 as well (see Hoffman and Rasche, 1996).
Table 9.14 Trace and maximum eigenvalue tests for cointegration
Test statistic
Null hypothesis

Alternative

p=5

p=6

5% critical value

127.801
72.302
35.169
16.110

75.98
53.48
34.87
20.18

55.499
37.133
19.059
11.860

34.40
28.27
22.04
15.87

λtrace -statistic
H0 :
H0 :
H0 :
H0 :

r
r
r
r

=0
≤1
≤2
≤3

H1 :
H1 :
H1 :
H1 :

r
r
r
r

≥1
≥2
≥3
≥4

H0 :
H0 :
H0 :
H0 :

r
r
r
r

=0
≤1
≤2
≤3

H1 :
H1 :
H1 :
H1 :

r
r
r
r

=1
=2
=3
=4

108.723
59.189
29.201
13.785
λmax -statistic
49.534
29.988
15.416
9.637

intercepts included T = 164
20

The results reported in this table are obtained from MicroFit 4.0; critical values taken from Table 9.9.

ILLUSTRATION: MONEY DEMAND AND INFLATION

337

If we restrict the rank of the long-run matrix to be equal to two we can estimate
the cointegrating vectors and the error-correction model by maximum likelihood, following the Johansen procedure. Recall that statistically the cointegrating vectors are
not individually deﬁned, only the space spanned by these vectors is. To identify individual cointegrating relationships we thus need to normalize the cointegrating vectors
somehow. When r = 2 we need to impose two normalization constraints on each
cointegrating vector. Note that in the cointegrating regressions in Table 9.13 a number
of constraints are imposed a priori, including a −1 for the right-hand side variables
and zero restrictions on some of the other variables’ coefﬁcients. In the current case
we need to impose two restrictions and, assuming that the money demand and risk
premium relationships are the most likely candidates, we shall impose that mt and
cpr t have coefﬁcients of −1, 0 and 0, −1, respectively. Economically, we expect that
inﬂ t does not enter in any of the cointegrating vectors. With these two restrictions,
the cointegrating vectors are estimated by maximum likelihood, jointly with the coefﬁcients in the vector error-correction model. The results for the cointegrating vectors
are presented in Table 9.15.
The cointegrating vector for the risk premium equation corresponds closely to our
a priori expectations, with the coefﬁcients for inﬂ t , yt and tbr t being insigniﬁcantly
different from zero, zero and one, respectively. For the vector corresponding to the
money demand equation inﬂ t appears to enter the equation signiﬁcantly. Recall that
mt corresponds to real money demand, which should normally not depend upon the
inﬂation rate. The coefﬁcient estimate of −0.023 implies that, ceteris paribus, nominal money demand (mt + inﬂ t ) increases somewhat less than proportionally with the
inﬂation rate.
It is possible to test our a priori cointegrating vectors by using likelihood ratio
tests. These tests require that the model is re-estimated imposing some additional
restrictions on the cointegrating vectors. This way we can test the following hypotheses:21

Table 9.15 ML estimates of cointegrating vectors (after normalization) based on VAR with
p = 6 (standard errors in parentheses), intercept
estimates not reported

mt
inﬂ t
cpr t
yt
tbr t

Money demand

Risk premium

−1
−0.023
(0.006)
0
0.425
(0.033)
−0.028
(0.005)

0
0.041
(0.031)
−1
−0.037
(0.173)
1.017
(0.026)

loglikelihood value: 808.2770

21

The tests here are actually overidentifying restrictions tests (see Chapter 5). We interpret them as regular
hypotheses tests taking the a priori restrictions in Table 9.15 as given.

MULTIVARIATE TIME SERIES MODELS

338

H0a : β12 = 0,
H0b :

β14 = 1;

β22 = β24 = 0,

β25 = 1; and

H0c : β12 = β22 = β24 = 0,

β14 = β25 = 1,

where β12 denotes the coefﬁcient for inﬂ t in the money demand equation and β22 and
β24 are the coefﬁcients for inﬂation and GDP in the risk premium equation, respectively. The loglikelihood values for the complete model, estimated imposing H0a , H0b
and H0c , respectively, are given by 782.3459, 783.7761 and 782.3196. The likelihood
ratio test statistics, deﬁned as twice the difference in loglikelihood values, for the three
null hypotheses are thus given by 51.86, 49.00 and 51.91. The asymptotic distributions
under the null hypotheses of the test statistics are the usual Chi-squared distributions, with degrees of freedom given by the number of restrictions that is tested (see
Chapter 6). Compared with the Chi-squared critical values with 3, 2 or 5 degrees of
freedom, each of the hypotheses is clearly rejected.
As a last step we consider the vector error-correction model for this system. This
corresponds to a VAR of order p − 1 = 5 for the ﬁrst-differenced series, with the
inclusion of two error-correction terms in each equation, one for each cointegrating
vector. Note that the number of parameters estimated in this vector error-correction
model is well above 100, so we shall concentrate on a limited part of the results only.
The two error-correction terms are given by
ecm1t = −mt − 0.023inﬂ t + 0.425yt − 0.028tbr t + 3.362;
ecm2t = −cpr t + 0.041inﬂ t − 0.037yt + 1.017tbr t + 0.687.
The adjustment coefﬁcients in the 5 × 2 matrix γ , with their associated standard errors,
are reported in Table 9.16. The long-run money demand equation contributes signiﬁcantly to the short-run movements of both money demand and income. The short-run
behaviour of money demand, inﬂation and the commercial paper rate appears to be
signiﬁcantly affected by the long-run risk premium relationship. There is no statistical
Table 9.16 Estimated matrix of adjustment
coefﬁcients (standard errors in parentheses),
∗
indicates signiﬁcance at the 5% level
Error-correction term
Equation
mt
inﬂ t
cpr t
yt
tbr t

ecm1t−1

ecm2t−1

∗

0.0090∗
(0.0024)
−1.1618∗
(0.5287)
0.6626∗
(0.2618)
−0.0013
(0.0028)
0.3195
(0.2365)

0.0276
(0.0104)
1.4629
(2.3210)
−2.1364
(1.1494)
0.0687∗
(0.0121)
−1.2876
(1.0380)

EXERCISES

339

evidence that the treasury bill rate adjusts to any deviation from long-run equilibria,
so that it could be treated as weakly exogenous.

9.7

Concluding Remarks

The literature on cointegration and related issues is of a recent date and still expanding.
In this chapter we have been fairly brief on some topics, while other topics have
been left out completely. Fortunately, there exists a substantial number of specialized
textbooks on the topic that provide a more extensive coverage. Examples of relatively
non-technical textbooks are Mills (1990), Harris (1995), Franses (1998), Patterson
(2001), and Enders (2004). More technical discussion is available in Lütkepohl (1991),
Cuthbertson, Hall and Taylor (1992), Banerjee et al. (1993), Hamilton (1994), Johansen
(1995), and Boswijk (1999).

Exercises
Exercise 9.1 (Cointegration Theory)

a. Assume that the two series yt and xt are I (1) and assume that both yt − β1 xt and
yt − β2 xt are I (0). Show that this implies that β1 = β2 , showing that there can be
only one unique cointegrating parameter.
b. Explain intuitively why the Durbin-Watson statistic in a regression of the I (1)
variables yt upon xt is informative about the question of cointegration between yt
and xt .
c. Explain what is meant by ‘super consistency’.
d. Consider three I (1) variables yt , xt and zt . Assume that yt and xt are cointegrated, and that xt and zt are cointegrated. Does this imply that yt and zt are also
cointegrated? Why (not)?
Exercise 9.2 (Cointegration)

Consider the following very simple relationship between aggregate savings St and
aggregate income Yt .
St = α + βYt + εt ,

t = 1, . . . , T .

(9.55)

For some country this relationship is estimated by OLS over the years 1946–1995
(T = 50). The results are given in Table 9.17.
Table 9.17 Aggregate savings explained from aggregate
income; OLS results
Variable

Coefﬁcient

Standard error

t-ratio

constant
income

38.90
0.098

4.570
0.009

8.51
10.77

T = 50 s = 22.57 R 2 = 0.93 dw = 0.70

340

MULTIVARIATE TIME SERIES MODELS

Assume, for the moment, that the series St and Yt are stationary. (Hint: if needed
consult Chapter 4 for the ﬁrst set of questions.)
a. How would you interpret the coefﬁcient estimate of 0.098 for the income variable?
b. Explain why the results indicate that there may be a problem of positive autocorrelation. Can you give arguments why, in economic models, positive autocorrelation
is more likely than negative autocorrelation?
c. What are the effects of autocorrelation on the properties of the OLS estimator?
Think about unbiasedness, consistency and the BLUE property.
d. Describe two different approaches to handle the autocorrelation problem in the
above case. Which one would you prefer?
From now on, assume that St and Yt are nonstationary I (1) series.
e.
f.
g.
h.
i.
j.
k.
l.
m.
n.

Are there indications that the relationship between the two variables is ‘spurious’?
Explain what we mean by ‘spurious regressions’.
Are there indications that there is a cointegrating relationship between St and Yt ?
Explain what we mean by a ‘cointegrating relationship’.
Describe two different tests that can be used to test the null hypothesis that St
and Yt are not cointegrated.
How do you interpret the coefﬁcient estimate of 0.098 under the hypothesis that
St and Yt are cointegrated?
Are there reasons to correct for autocorrelation in the error term when we estimate
a cointegrating regression?
Explain intuitively why the estimator for a cointegrating parameter is super
consistent.
Assuming that St and Yt are cointegrated, describe what we mean by an errorcorrection mechanism. Give an example. What do we learn from it?
How can we consistently estimate an error-correction model?

Exercise 9.3 (Cointegration – Empirical)
In the ﬁles INCOME we ﬁnd quarterly data on UK nominal consumption and income,
for 1971:1 to 1985:2 (T = 58). Part of these data was used in Exercise 8.3.

a.
b.
c.
d.
e.

f.

Test for a unit root in the consumption series using several augmented
Dickey–Fuller tests.
Perform a regression by OLS explaining consumption from income. Test for
cointegration using two different tests.
Perform a regression by OLS explaining income from consumption. Test for
cointegration.
Compare the estimation results and R 2 s from the last two regressions.
Determine the error-correction term from one of the two regressions and estimate an error-correction model for the change in consumption. Test whether the
adjustment coefﬁcient is zero.
Repeat the last question for the change in income. What do you conclude?

10

Models Based on
Panel Data

A panel data set contains repeated observations over the same units (individuals,
households, ﬁrms), collected over a number of periods. Although panel data are typically collected at the micro-economic level, it has become more and more practice to
pool individual time series of a number of countries or industries and analyse them
simultaneously. The availability of repeated observations on the same units allows
economists to specify and estimate more complicated and more realistic models than
a single cross-section or a single time series would do. The disadvantages are more
of a practical nature: because we repeatedly observe the same units, it is usually
no longer appropriate to assume that different observations are independent. This
may complicate the analysis, particularly in nonlinear and dynamic models. Furthermore, panel data sets very often suffer from missing observations. Even if these
observations are missing in a random way (see below), the standard analysis has to
be adjusted.
This chapter provides an introduction to the analysis of panel data. A simple linear
panel data model is presented in Section 10.1 and some advantages compared to crosssectional or time series data are discussed in the context of this model. Section 10.2
pays attention to the so-called ﬁxed effects and random effects models, and discusses
issues relating to the choice between these two basic models. An empirical illustration is provided in Section 10.3. The introduction of a lagged dependent variable
in the linear model complicates consistent estimation, and, as will be discussed in
Section 10.4, instrumental variables procedures or GMM provide interesting alternatives. Section 10.5 provides an empirical example on the estimation of short- and
long-run wage elasticities of labour demand. Increasingly, panel data approaches are
used in a macro-economic context to investigate the dynamic properties of economic
variables. Section 10.6 discusses the recent literature on unit root and cointegration
tests in heterogeneous panels. In micro-economic applications, the model of interest
often involves limited dependent variables and panel data extensions of logit, probit

MODELS BASED ON PANEL DATA

342

and tobit models are discussed in Section 10.7. Finally, we discuss the problems associated with incomplete panels and selection bias in Section 10.8. Extensive discussions
of the econometrics of panel data can be found in Baltagi (2001), Wooldridge (2002),
Hsiao (2003) and Arellano (2003).

10.1

Advantages of Panel Data

An important advantage of panel data compared to time series or cross-sectional data
sets is that it allows identiﬁcation of certain parameters or questions, without the need
to make restrictive assumptions. For example, panel data make it possible to analyse
changes on an individual level. Consider a situation in which the average consumption
level rises with 2% from one year to another. Panel data can identify whether this rise
is the result of, for example, an increase of 2% for all individuals or an increase of
4% for approximately one half of the individuals and no change for the other half (or
any other combination). That is, panel data are not only suitable to model or explain
why individual units behave differently but also to model why a given unit behaves
differently at different time periods (for example, because of a different past).
We shall, in the sequel, index all variables by an i for the individual1 (i = 1, . . . , N )
and a t for the time period (t = 1, . . . , T ). In very general terms, we could specify a
linear model as
yit = xit βit + εit ,
where βit measures the partial effects of xit in period t for unit i. Of course, this model
is much too general to be useful, and we need to put more structure on the coefﬁcients
βit . The standard assumption, used in many empirical cases, is that βit is constant for
all i and t, except – possibly – the intercept term. This could be written as
yit = αi + xit β + εit ,

(10.1)

where xit is a K-dimensional vector of explanatory variables, not including a constant.2
This means that the effects of a change in x are the same for all units and all periods,
but that the average level for unit i may be different from that for unit j . The αi
thus capture the effects of those variables that are peculiar to the i-th individual and
that are constant over time. In the standard case, εit is assumed to be independent and
identically distributed over individuals and time, with mean zero and variance σε2 . If
we treat the αi as N ﬁxed unknown parameters, the model in (10.1) is referred to as
the standard ﬁxed effects model.
An alternative approach assumes that the intercepts of the individuals are different
but that they can be treated as drawings from a distribution with mean µ and variance σα2 . The essential assumption here is that these drawings are independent of the
explanatory variables in xit (see below). This leads to the random effects model, where
the individual effects αi are treated as random. The error term in this model consists
1

While we refer to the cross-sectional units as individuals, they could also refer to other units like ﬁrms,
countries, industries, households or assets.
2
The elements in β are indexed as β1 to βK , where the ﬁrst element – unlike the previous chapters – does
not refer to the intercept.

ADVANTAGES OF PANEL DATA

343

of two components: a time-invariant component3 αi and a remainder component εit
that is uncorrelated over time.4 It can be written as
yit = µ + xit β + αi + εit ,

(10.2)

where µ denotes the intercept term.
The possibility of treating the αi s as ﬁxed parameters has some great advantages,
but also some disadvantages. Most panel data models are estimated under either the
ﬁxed effects or the random effects assumption and we shall discuss this extensively
in Section 10.2. First, the next two subsections discuss some potential advantages of
panel data in more detail.
10.1.1 Efﬁciency of Parameter Estimators

Because panel data sets are typically larger than cross-sectional or time series data
sets, and explanatory variables vary over two dimensions (individuals and time) rather
than one, estimators based on panel data are quite often more accurate than from other
sources. Even with identical sample sizes, the use of a panel data set will often yield
more efﬁcient estimators than a series of independent cross-sections (where different
units are sampled in each period). To illustrate this, consider the following special case
of the random effects model in (10.2) where we only include time dummies, i.e.
yit = µt + αi + εit ,

(10.3)

where each µt is an unknown parameter corresponding to the population mean in
period t. Suppose we are not interested in the mean µt in a particular period, but in
the change of µt from one period to another. In general the variance of the efﬁcient
estimator for µt − µs (s = t), µ̂t − µ̂s , is given by
V {µ̂t − µ̂s } = V {µ̂t } + V {µ̂s } − 2 cov{µ̂t , µ̂s }

(10.4)


with µ̂t = 1/N N
i=1 yit (t = 1, . . . , T ). Typically, if a panel data set is used the
covariance between µ̂t and µ̂s will be positive, in particular – if the random effects
assumptions of model (10.2) hold – equal σα2 /N . However, if two independent crosssectional data sets are used different periods will contain different individuals so µ̂t
and µ̂s will have zero covariance. In other words, if one is interested in changes from
one period to another, a panel will yield more efﬁcient estimators than a series of
cross-sections.
Note, however, that the reverse is also true, in the sense that repeated cross-sections
will be more informative than a panel when one is interested in a sum or average
of µt over several periods. At a more intuitive level, panel data may provide better
information because the same individuals are repeatedly observed. On the other hand,
having the same individuals rather than different ones may imply less variation in
the explanatory variables and thus relatively inefﬁcient estimators. A comprehensive
3
4

In the random effects model, the αi s are redeﬁned to have a zero mean.
The model is sometimes referred to as a (one-way) error components model.

MODELS BASED ON PANEL DATA

344

analysis on the choice between a pure panel, a pure cross-section and a combination
of these two data sources, is provided in Nijman and Verbeek (1990). Their results
indicate that when exogenous variables are included in the model and one is interested
in the parameters which measure the effects of these variables, a panel data set will
typically yield more efﬁcient estimators than a series of cross-sections with the same
number of observations.
10.1.2 Identiﬁcation of Parameters

A second advantage of the availability of panel data is that it reduces identiﬁcation
problems. Although this advantage may come under different headings, in many cases
it involves identiﬁcation in the presence of endogenous regressors or measurement
error, robustness to omitted variables and the identiﬁcation of individual dynamics.
Let us start with an illustration of the last of these. There are two alternative explanations for the often observed phenomenon that individuals who have experienced
an event in the past are more likely to experience that event in the future. The ﬁrst
explanation is that the fact that an individual has experienced the event changes his preferences, constraints, etc., in such a way that he is more likely to experience that event
in the future. The second explanation says that individuals may differ in unobserved
characteristics which inﬂuence the probability of experiencing the event (but are not
inﬂuenced by the experience of the event). Heckman (1978a) terms the former explanation true state dependence and the latter spurious state dependence. A well-known
example concerns the ‘event’ of being unemployed. The availability of panel data
will ease the problem of distinguishing between true and spurious state dependence,
because individual histories are observed and can be included in the model.
Omitted variable bias arises if a variable that is correlated with the included variables is excluded from the model. A classical example is the estimation of production
functions (Mundlak, 1961). In many cases, especially in the case of small ﬁrms, it
is desirable to include management quality as an input in the production function.
In general, however, management quality is unobservable. Suppose that a production
function of the Cobb–Douglas type is given by
yit = µ + xit β + mi βK+1 + εit

(10.5)

where yit denotes log output, xit is a K-dimensional vector of log inputs, both for
ﬁrm i at time t, and mi denotes the management quality for ﬁrm i (which is assumed
to be constant over time). The unobserved variable mi is expected to be negatively
correlated with the other inputs in xit , since a high quality management will probably
result in a more efﬁcient use of inputs. Therefore, unless βK+1 = 0, deletion of mi
from (10.5) will lead to biased estimates of the other parameters in the model. If
panel data are available this problem can be resolved by introducing a ﬁrm speciﬁc
effect αi = µ + mi βK+1 and considering this as a ﬁxed unknown parameter. Note that
without additional information it is not possible to identify βK+1 ; a restriction that
identiﬁes βK+1 is the imposition of constant returns to scale.5
In a similar way, a ﬁxed time effect can be included in the model to capture the
effect of all (observed and unobserved) variables that do not vary over the individual
5

Constant returns to scale implies that βK+1 = 1 − (β1 + · · · + βK ).

THE STATIC LINEAR MODEL

345

units. This illustrates the proposition that panel data can reduce the effects of omitted
variable bias, or – in other words – estimators from a panel data set may be more
robust to an incomplete model speciﬁcation.
Finally, in many cases panel data will provide ‘internal’ instruments for regressors
that are endogenous or subject to measurement error. That is, transformations of the
original variables can often be argued to be uncorrelated with the model’s error term
and correlated with the explanatory variables themselves and no external instruments
are needed. For example, if xit is correlated with αi , it can be argued that xit − x̄i ,
where x̄i is the time-average for individual i, is uncorrelated with αi and provides a
valid instrument for xit . More generally, estimating the model under the ﬁxed effects
assumption eliminates αi from the error term and, consequently, eliminates all endogeneity problems relating to it. This will be illustrated in the next section. An extensive
discussion of the beneﬁts and limitations of panel data is provided in Hsiao (1985).

10.2

The Static Linear Model

In this section we discuss the static linear model in a panel data setting. We start with
two basic models, the ﬁxed effects and the random effects model, and subsequently
discuss the choice between the two, as well as alternative procedures that could be
considered to be somewhere between a ﬁxed effects and a random effects treatment.
10.2.1 The Fixed Effects Model

The ﬁxed effects model is simply a linear regression model in which the intercept
terms vary over the individual units i, i.e.
yit = αi + xit β + εit ,

εit ∼ IID(0, σε2 ),

(10.6)

where it is usually assumed that all xit are independent of all εit . We can write this in
the usual regression framework by including a dummy variable for each unit i in the
model. That is,
N

yit =
αj dij + xit β + εit ,
(10.7)
j =1

where dij = 1 if i = j and 0 elsewhere. We thus have a set of N dummy variables in the
model. The parameters α1 , . . . , αN and β can be estimated by ordinary least squares in
(10.7). The implied estimator for β is referred to as the least squares dummy variable
(LSDV) estimator. It may, however, be numerically unattractive to have a regression
model with so many regressors. Fortunately one can compute the estimator for β in a
simpler way. It can be shown that exactly the same estimator for β is obtained if the
regression is performed in deviations from individual means. Essentially, this implies
that we eliminate the individual effects αi ﬁrst by transforming the data. To see this,
ﬁrst note that
ȳi = αi + x̄i β + ε̄i ,
(10.8)

MODELS BASED ON PANEL DATA

346

where ȳi = T −1


t

yit and similarly for the other variables. Consequently, we can write
yit − ȳi = (xit − x̄i ) β + (εit − ε̄i ).

(10.9)

This is a regression model in deviations from individual means and does not include the
individual effects αi . The transformation that produces observations in deviation from
individual means, as in (10.9), is called the within transformation. The OLS estimator
for β obtained from this transformed model is often called the within estimator or
ﬁxed effects estimator, and it is exactly identical to the LSDV estimator described
above. It is given by


β̂FE

N 
T

=
(xit − x̄i )(xit − x̄i )
i=1 t=1

−1

N 
T


(xit − x̄i )(yit − ȳi ).

(10.10)

i=1 t=1

If it is assumed that all xit are independent of all εit (compare assumption (A2) from
Chapter 2), the ﬁxed effects estimator can be shown to be unbiased for β. If, in addition,
normality of εit is imposed, β̂FE also has a normal distribution. For consistency,6 it is
required that
(10.11)
E{(xit − x̄i )εit } = 0
(compare assumption (A7) in Chapters 2 and 5). Sufﬁcient for this is that xit is uncorrelated with εit and that x̄i has no correlation with the error term. These conditions are
in turn implied by
(10.12)
E{xit εis } = 0 for all s, t,
in which case we call xit strictly exogenous. A strictly exogenous variable is not
allowed to depend upon current, future and past values of the error term. In some
applications this may be restrictive. Clearly, it excludes the inclusion of lagged dependent variables in xit , but any xit variable which depends upon the history of yit would
also violate the condition. For example, if we are explaining labour supply of an individual, we may want to include years of experience in the model, while quite clearly
experience depends upon the person’s labour history.
With explanatory variables independent of all errors, the N intercepts are estimated
unbiasedly as
α̂i = ȳi − x̄i β̂FE , i = 1, . . . , N.
Under assumption (10.11) these estimators are consistent for the ﬁxed effects αi provided T goes to inﬁnity. The reason why α̂i is inconsistent for ﬁxed T is clear: when
T is ﬁxed the individual averages ȳi and x̄i do not converge to anything if the number
of individuals increases.
6

Unless stated otherwise, we consider in this chapter consistency for the number of individuals N going to
inﬁnity. This corresponds with the common situation that we have panels with large N and small T .

THE STATIC LINEAR MODEL

347

The covariance matrix for the ﬁxed effects estimator β̂FE , assuming that εit is i.i.d.
across individuals and time with variance σε2 , is given by

V {β̂FE } =

σε2

N 
T


−1


(xit − x̄i )(xit − x̄i )

.

(10.13)

i=1 t=1

Unless T is large, using the standard OLS estimate for the covariance matrix based
upon the within regression in (10.9) will underestimate the true variance. The reason
is that in this transformed regression the error covariance matrix is singular (as the T
transformed errors of each individual add up to zero) and the variance of εit − ε̄i is
(T − 1)/T σε2 rather than σε2 . A consistent estimator for σε2 is obtained as the within
residual sum of squares divided by N (T − 1). That is,
σ̂ε2 =
=

N

T

N

T


1
(y − α̂i − xit β̂FE )2
N (T − 1) i=1 t=1 it

1
(y − ȳi − (xit − x̄i ) β̂FE )2 .
N (T − 1) i=1 t=1 it

(10.14)

It is possible to apply the usual degrees of freedom correction in which case K is subtracted from the denominator. Note that using the standard OLS covariance matrix in
model (10.7) with N individual dummies is reliable, because the degrees of freedom
correction involves N additional unknown parameters corresponding to the individual intercept terms. Under weak regularity conditions, the ﬁxed effects estimator is
asymptotically normal, so that the usual inference procedures can be used (like t and
Wald tests).
Essentially, the ﬁxed effects model concentrates on differences ‘within’ individuals.
That is, it is explaining to what extent yit differs from ȳi and does not explain why ȳi
is different from ȳj . The parametric assumptions about β on the other hand, impose
that a change in x has the same (ceteris paribus) effect, whether it is a change from
one period to the other or a change from one individual to the other. When interpreting
the results, however, from a ﬁxed effects regression, it may be important to realize that
the parameters are identiﬁed only through the within dimension of the data.
10.2.2 The Random Effects Model

It is commonly assumed in regression analysis that all factors that affect the dependent
variable, but that have not been included as regressors, can be appropriately summarized
by a random error term. In our case, this leads to the assumption that the αi are random
factors, independently and identically distributed over individuals. Thus we write the
random effects model as
yit = µ + xit β + αi + εit ,

εit ∼ IID(0, σε2 );

αi ∼ IID(0, σα2 ),

(10.15)

where αi + εit is treated as an error term consisting of two components: an individual
speciﬁc component, which does not vary over time, and a remainder component, which

MODELS BASED ON PANEL DATA

348

is assumed to be uncorrelated over time. That is, all correlation of the error terms
over time is attributed to the individual effects αi . It is assumed that αi and εit are
mutually independent and independent of xjs (for all j and s). This implies that the OLS
estimator for µ and β from (10.15) is unbiased and consistent. The error components
structure implies that the composite error term αi + εit exhibits a particular form of
autocorrelation (unless σα2 = 0). Consequently, routinely computed standard errors for
the OLS estimator are incorrect and a more efﬁcient (GLS) estimator can be obtained
by exploiting the structure of the error covariance matrix.
To derive the GLS estimator,7 ﬁrst note that for individual i all error terms can be
stacked as αi ιT + εi , where ιT = (1, 1, . . . , 1) of dimension T and εi = (εi1 , . . . , εiT ) .
The covariance matrix of this vector is (see Hsiao, 2003, Section 3.3)
V {αi ιT + εi } =  = σα2 ιT ιT + σε2 IT ,

(10.16)

where IT is the T -dimensional identity matrix. This can be used to derive the generalized least squares (GLS) estimator for the parameters in (10.15). For each individual,
we can transform the data by premultiplying the vectors yi = (yi1 , . . . , yiT ) etc. by
−1 , which is given by
−1



=

σε−2




σα2

ι ι ,
IT − 2
σε + T σα2 T T

which can also be written as
−1



=

σε−2



1
IT − ιT ιT
T

where
ψ=




1 
+ ψ ιT ιT ,
T

σε2
.
σε2 + T σα2

Noting that IT − (1/T )ιT ιT transforms the data in deviations from individual means
and (1/T )ιT ιT takes individual means, the GLS estimator for β can be written as

β̂GLS =

N 
T


(xit − x̄i )(xit − x̄i ) + ψT

i=1 t=1



×



N 
T

i=1 t=1

N


−1
(x̄i − x̄)(x̄i − x̄)

i=1

(xit − x̄i )(yit − ȳi ) + ψT



N



(x̄i − x̄)(ȳi − ȳ) ,

(10.17)

i=1


where x̄ = (1/(NT )) i,t xit denotes the overall average of xit . It is easy to see that
for ψ = 0 the ﬁxed effects estimator arises. Because ψ → 0 if T → ∞, it follows that
the ﬁxed and random effects estimators are equivalent for large T . If ψ = 1, the GLS
7

It may be instructive to re-read the general introduction to GLS estimation in Section 4.2.

THE STATIC LINEAR MODEL

349

estimator is just the OLS estimator (and  is diagonal). From the general formula for
the GLS estimator it can be derived that
β̂GLS = β̂B + (Ik − )β̂FE ,
where


β̂B =

N


−1


(x̄i − x̄)(x̄i − x̄)

i=1

N

(x̄i − x̄)(ȳi − ȳ)
i=1

is the so-called between estimator for β. It is the OLS estimator in the model for
individual means
(10.18)
ȳi = µ + x̄i β + αi + ε̄i , i = 1, . . . , N.
The matrix  is a weighting matrix and is proportional to the inverse of the covariance
matrix of β̂B (see Hsiao, 2003, Section 3.4, for details). That is, the GLS estimator is a
matrix-weighted average of the between estimator and the within estimator, where the
weight depends upon the relative variances of the two estimators. (The more accurate
one gets the higher the weight.)
The between estimator effectively discards the time series information in the data
set. The GLS estimator, under the current assumptions, is the optimal combination
of the within estimator and the between estimator, and is therefore more efﬁcient
than either of these two estimators. The OLS estimator (with ψ = 1) is also a linear
combination of the two estimators, but not the efﬁcient one. Thus, GLS will be more
efﬁcient than OLS, as usual. If the explanatory variables are independent of all εit and
all αi , the GLS estimator is unbiased. It is a consistent estimator for N or T or both
tending to inﬁnity if in addition to (10.11) it also holds that E{x̄i εit } = 0 and most
importantly that
E{x̄i αi } = 0.
(10.19)
Note that these conditions are also required for the between estimator to be consistent.
An easy way to compute the GLS estimator is obtained by noting that it can be
determined as the OLS estimator in a transformed model (compare Chapter 4), given by
(yit − ϑ ȳi ) = µ(1 − ϑ) + (xit − x̄i ) β + uit ,

(10.20)

where ϑ = 1 − ψ 1/2 . The error term in this transformed regression is i.i.d. over individuals and time. Note again that ψ = 0 corresponds to the within estimator (ϑ = 1).
In general, a ﬁxed proportion ϑ of the individual means is subtracted from the data to
obtain this transformed model (0 ≤ ϑ ≤ 1).
Of course, the variance components σα2 and σε2 are unknown in practice. In this
case we can use the feasible GLS estimator (EGLS), where the unknown variances are
consistently estimated in a ﬁrst step. The estimator for σε2 is easily obtained from the
within residuals, as given in (10.14). For the between regression the error variance is
σα2 + (1/T )σε2 , which we can estimate consistently by
σ̂B2 =

N
1 
(ȳ − µ̂B − x̄i β̂B )2 ,
N i=1 i

(10.21)

MODELS BASED ON PANEL DATA

350

where µ̂B is the between estimator for µ. From this, a consistent estimator for σα2
follows as
1
(10.22)
σ̂α2 = σ̂B2 − σ̂ε2 .
T
Again, it is possible to adjust this estimator by applying a degrees of freedom correction,
implying that the number of regressors K + 1 is subtracted in the denominator of
(10.21) (see Hsiao, 2003, Section 3.3). The resulting EGLS estimator is referred to as
the random effects estimator for β (and µ), denoted below as β̂RE . It is also known
as the Balestra–Nerlove estimator.
Under weak regularity conditions, the random effects estimator is asymptotically
normal. Its covariance matrix is given by

V {β̂RE } =

σε2

N 
T




(xit − x̄i )(xit − x̄i ) + ψT

i=1 t=1

N


−1


(x̄i − x̄)(x̄i − x̄)

, (10.23)

i=1

which shows that the random effects estimator is more efﬁcient than the ﬁxed effects
estimator as long as ψ > 0. The gain in efﬁciency is due to the use of the between
variation in the data (x̄i − x̄). The covariance matrix in (10.23) is routinely estimated
by the OLS expressions in the transformed model (10.20).
In summary, we have seen a range of estimators for the parameter vector β. The
basic two are:
1. The between estimator, exploiting the between dimension of the data (differences
between individuals), determined as the OLS estimator in a regression of individual averages of y on individual averages of x (and a constant). Consistency, for
N → ∞, requires that E{x̄i αi } = 0 and E{x̄i ε̄i } = 0. Typically this means that the
explanatory variables are strictly exogenous and uncorrelated with the individual
speciﬁc effect αi .
2. The ﬁxed effects (within) estimator, exploiting the within dimension of the data
(differences within individuals), determined as the OLS estimator in a regression in
deviations from individual means. It is consistent for β for T → ∞ or N → ∞,
provided that E{(xit − x̄i )εit } = 0. Again this requires the x-variables to be strictly
exogenous, but it does not impose any restrictions upon the relationship between
αi and xit .
The other two estimators are:
3. The OLS estimator, exploiting both dimensions (within and between) but not efﬁciently. Determined (of course) as OLS in the original model given in (10.15). Consistency for T → ∞ or N → ∞ requires that E{xit (εit + αi )} = 0. This requires
the explanatory variables to be uncorrelated with αi but does not impose that they
are strictly exogenous. It sufﬁces that xit and εit are contemporaneously uncorrelated.
4. The random effects (EGLS) estimator, combining the information from the
between and within dimensions in an efﬁcient way. It is consistent for T → ∞
or N → ∞ under the combined conditions of 1 and 2. It can be determined as a

THE STATIC LINEAR MODEL

351

weighted average of the between and within estimator or as the OLS estimator in a
regression where the variables are transformed as yit − ϑ̂ ȳi , where ϑ̂ is an estimate
for ϑ = 1 − ψ 1/2 with ψ = σε2 /(σε2 + T σα2 ).
10.2.3 Fixed Effects or Random Effects?

Whether to treat the individual effects αi as ﬁxed or random is not an easy question
to answer. It can make a surprising amount of difference in the estimates of the β
parameters in cases where T is small and N is large. When only a few observations
are available for each individual it is very important to make the most efﬁcient use of
the data. The most common view is that the discussion should not be about the ‘true
nature’ of the effects αi . The appropriate interpretation is that the ﬁxed effects approach
is conditional upon the values for αi . That is, it essentially considers the distribution
of yit given αi , where the αi s can be estimated. This makes sense intuitively if the
individuals in the sample are ‘one of a kind’, and cannot be viewed as a random draw
from some underlying population. This interpretation is probably most appropriate
when i denotes countries, (large) companies or industries, and predictions we want
to make are for a particular country, company or industry. Inferences are thus with
respect to the effects that are in the sample.
In contrast, the random effects approach is not conditional upon the individual αi s,
but ‘integrates them out’. In this case, we are usually not interested in the particular
value of some person’s αi ; we just focus on arbitrary individuals that have certain
characteristics. The random effects approach allows one to make inference with respect
to the population characteristics. One way to formalize this is noting that the random
effects model states that
E{yit |xit } = xit β,
(10.24)
while the ﬁxed effects model estimates
E{yit |xit , αi } = xit β + αi .

(10.25)

Note that the β coefﬁcients in these two conditional expectations are the same only if
E{αi |xit } = 0. To summarize this, a ﬁrst reason why one may prefer the ﬁxed effects
estimator is that some interest lies in αi , which makes sense if the number of units
is relatively small and of a speciﬁc nature. That is, identiﬁcation of individual units
is important.
However, even if we are interested in the larger population of individual units, and
a random effects framework seems appropriate, the ﬁxed effects estimator may be
preferred. The reason for this is that it may be the case that αi and xit are correlated,
in which case the random effects approach, ignoring this correlation, leads to inconsistent estimators. We saw an example of this above, where αi included management
quality and was argued to be correlated with the other inputs included in the production function. The problem of correlation between the individual effects αi and the
explanatory variables in xit can be handled by using the ﬁxed effects approach, which
essentially eliminates the αi from the model, and thus eliminates any problems that
they may cause.
Hausman (1978) has suggested a test for the null hypothesis that xit and αi are
uncorrelated. The general idea of a Hausman test is that two estimators are compared:

MODELS BASED ON PANEL DATA

352

one which is consistent under both the null and alternative hypothesis and one which
is consistent (and typically efﬁcient) under the null hypothesis only. A signiﬁcant
difference between the two estimators indicates that the null hypothesis is unlikely to
hold. In the present case, assume that E{εit xis } = 0 for all s, t, so that the ﬁxed effects
estimator β̂FE is consistent for β irrespective of the question whether xit and αi are
uncorrelated, while the random effects estimator β̂RE is consistent and efﬁcient only
if xit and αi are not correlated. Let us consider the difference vector β̂FE − β̂RE . To
evaluate the signiﬁcance of this difference, we need its covariance matrix. In general
this would require us to estimate the covariance between β̂FE and β̂RE , but because
the latter estimator is efﬁcient under the null hypothesis, it can be shown that (under
the null)
V {β̂FE − β̂RE } = V {β̂FE } − V {β̂RE }.
(10.26)
Consequently, we can compute the Hausman test statistic as
ξH = (β̂FE − β̂RE ) [V̂ {β̂FE } − V̂ {β̂RE }]−1 (β̂FE − β̂RE ),

(10.27)

where the V̂ s denote estimates of the true covariance matrices. Under the null hypothesis, which implicitly says that plim(β̂FE − β̂RE ) = 0, the statistic ξH has an asymptotic
Chi-squared distribution with K degrees of freedom, where K is the number of elements in β.
The Hausman test thus tests whether the ﬁxed effects and random effects estimator
are signiﬁcantly different. Computationally, this is relatively easy because the covariance matrix satisﬁes (10.26). An important reason why the two estimators would be
different is the existence of correlation between xit and αi , although other sorts of
misspeciﬁcation may also read to rejection (we shall see an example of this below).
A practical problem when computing (10.27) is that the covariance matrix in square
brackets may not be positive deﬁnite in ﬁnite samples, such that its inverse cannot be
computed. As an alternative, it is possible to test for a subset of the elements in β.
10.2.4 Goodness-of-ﬁt

The computation of goodness-of-ﬁt measures in panel data applications is somewhat
uncommon. One reason is the fact that one may attach different importance to explaining the within and between variation in the data. Another reason is that the usual R 2
or adjusted R 2 criteria are only appropriate if the model is estimated by OLS.
Our starting point here is the deﬁnition of the R 2 in terms of the squared correlation
coefﬁcient between actual and ﬁtted values, as presented in Section 2.4. This deﬁnition
has the advantage that it produces values within the [0, 1] interval, irrespective of the
estimator that is used to generate the ﬁtted values. Recall that it corresponds to the
standard deﬁnition of the R 2 (in terms of sums of squares) if the model is estimated
by OLS (provided that an intercept term is included). In the current context, the total
variation in yit can be written as the sum of the within variation and the between
variation, that is
1 
1 
1 
(yit − ȳ)2 =
(yit − ȳi )2 +
(ȳ − ȳ)2 ,
NT i,t
NT i,t
N i i

(10.28)

THE STATIC LINEAR MODEL

353

where ȳ denotes the overall sample average. Now, we can deﬁne alternative versions
of an R 2 measure, depending upon the dimension of the data that we are interested in.
For example, the ﬁxed effects estimator is chosen to explain the within variation as
well as possible, and thus maximizes the ‘within R 2 ’ given by
2
(β̂FE ) = corr2 {ŷitFE − ŷiFE , yit − ȳi },
Rwithin

(10.29)

where ŷitFE − ŷiFE = (xit − x̄i ) β̂FE and corr2 denotes the squared correlation coefﬁcient.
The between estimator, being an OLS estimator in the model in terms of individual
means, maximizes the ‘between R 2 ’, which we deﬁne as
2
(β̂B ) = corr2 {ŷiB , ȳi },
Rbetween

(10.30)

where ŷiB = x̄i β̂B . The OLS estimator maximizes the overall goodness-of-ﬁt and thus
the overall R 2 , which is deﬁned as
2
(β̂) = corr2 {ŷit , yit },
Roverall

(10.31)

for an arbitrary
with ŷit = xit b. It is possible to deﬁne within, between and overall R 2 s
estimator 
β̂ for β by using as ﬁtted values ŷit = xit β̂, ŷi = (1/T ) t ŷit and ŷ =
(1/(N T )) i,t ŷit , where the intercept terms are omitted (and irrelevant).8 For the
ﬁxed effects estimator this ignores the variation captured by the α̂i s. If we would take
into account the variation explained by the N estimated intercepts α̂i , the ﬁxed effects
model perfectly ﬁts the between variation. This is somewhat unsatisfactory though, as
it is hard to argue that the ﬁxed effects α̂i explain the variation between individuals,
they just capture it. Put differently, if we ask ourselves: why does individual i consume
on average more than another individual, the answer provided by α̂i is simply: because
it is individual i. Given this argument, and because the α̂i s are often not computed, it
seems appropriate to ignore this part of the model.
Taking the deﬁnition in terms of the squared correlation coefﬁcients, the three measures above can be computed for any of the estimators that we considered. If we take
the random effects estimator, which is (asymptotically) the most efﬁcient estimator if
the assumptions of the random effects model are valid, the within, between and overall
R 2 s are necessarily smaller than for the ﬁxed effects, between and OLS estimator,
respectively. This, again, stresses that goodness-of-ﬁt measures are not adequate to
choose between alternative estimators. They provide, however, possible criteria for
choosing between alternative (potentially non-nested) speciﬁcations of the model.
10.2.5 Alternative Instrumental Variables Estimators

The ﬁxed effects estimator eliminates anything that is time-invariant from the model.
This may be a high price to pay for allowing the x-variables to be correlated with the
individual speciﬁc heterogeneity αi . For example, we may be interested in the effect
of time-invariant variables (like gender) on a person’s wage. Actually, there is no need
to restrict attention to the ﬁxed and random effects assumptions only, as it is possible
8

These deﬁnitions correspond to the R 2 measures as computed in Stata 7.0.

MODELS BASED ON PANEL DATA

354

to derive instrumental variables estimators that can be considered to be in between a
ﬁxed and random effects approach.
To see this, let us ﬁrst of all note that we can write the ﬁxed effects estimator as

β̂FE =

=

N 
T


−1


(xit − x̄i )(xit − x̄i )

i=1 t=1
N 
T

i=1 t=1

−1
(xit −

x̄i )xit

N 
T


(xit − x̄i )(yit − ȳi )

i=1 t=1
N 
T


(xit − x̄i )yit .

(10.32)

i=1 t=1

Writing the estimator like this shows that it has the interpretation of an instrumental
variables estimator9 for β in the model
yit = µ + xit β + αi + εit ,
where each explanatory variable is instrumented by its value in deviation from
the individual speciﬁc mean. That is, xit is instrumented by xit − x̄i . Note that
E{(xit − x̄i )αi } = 0 by construction (if we take expectations over i and t), so that
the IV estimator is consistent provided E{(xit − x̄i )εit } = 0, which is implied by
the strict exogeneity of xit . Clearly, if a particular element in xit is known to be
uncorrelated with αi there is no need to instrument it; that is, this variable can be
used as its own instrument. This route may also allow us to estimate the effect of
time-invariant variables.
To describe the general approach, let us consider a linear model with four groups of
explanatory variables (Hausman and Taylor, 1981)


yit = µ + x1,it
β1 + x2,it
β2 + w1i γ1 + w2i γ2 + αi + εit ,

(10.33)

where the x-variables are time-varying and the w -variables are time-invariant. The
variables with index 1 are assumed to be uncorrelated with both αi and all εis . The
variables x2,it and w2i are correlated with αi but not with any εis . Under these assumptions, the ﬁxed effects estimator would be consistent for β1 and β2 , but would not
identify the coefﬁcients for the time-invariant variables. Moreover, it is inefﬁcient
because x1,it is needlessly instrumented. Hausman and Taylor (1981) suggest to estimate (10.33) by instrumental variables using the following variables as instruments:
x1,it , w1i and x2,it − x̄2i , x̄1i . That is, the exogenous variables serve as their own instruments, x2,it is instrumented by its deviation from individual means (as in the ﬁxed
effects approach) and w2i is instrumented by the individual average of x1,it . Obviously,
identiﬁcation requires that the number of variables in x1,it is at least as large as that
in w2i . The resulting estimator, the Hausman–Taylor estimator, allows us to estimate the effect of time-invariant variables, even though the time-varying regressors
are correlated with αi . The trick here is to use the time averages of those time-varying
regressors that are uncorrelated with αi as instruments for the time-invariant regressors.
Clearly, this requires that sufﬁcient time-varying variables are included that have no
correlation with αi . Of course, it is a straightforward extension to include additional
9

It may be instructive to re-read Section 5.3 for a general discussion of instrumental variables estimation.

THE STATIC LINEAR MODEL

355

instruments in the procedure that are not based on variables included in the model.
This is what one is forced to do in the cross-sectional case, where no transformations
are available that can be argued to produce valid instruments. The strong advantage of
the Hausman–Taylor approach is that one does not have to use external instruments.
With sufﬁcient assumptions instruments can be derived within the model. Despite this
important advantage, the Hausman–Taylor estimator plays a surprisingly minor role in
empirical work. A notable exception is Chowdhury and Nickell (1985).
Hausman and Taylor also show that the instrument set is equivalent to using
x1,it − x̄1i , x2,it − x̄2i and x1,it , w1i . This follows directly from the fact that taking
different linear combinations of the original instruments does not affect the estimator.
Hausman and Taylor also show how the nondiagonal covariance matrix of the error
term in (10.33) can be exploited to improve the efﬁciency of the estimator. Nowadays,
this would typically be handled in a GMM framework, as we shall see in the next
section (see Arellano and Bover, 1995).
Two subsequent papers try to improve upon the efﬁciency of the Hausman–Taylor
instrumental variables estimator by proposing a larger set of instruments. Amemiya
and MaCurdy (1986) suggest the use of the time-invariant instruments x1,i1 − x̄1i up to
x1,iT − x̄1i . This requires that E{(x1,it − x̄1i )αi } = 0 for each t. This assumption makes
sense if the correlation between αi and x1,it is due to a time-invariant component in
x1,it , such that E{x1,it αi } for a given t does not depend upon t. Breusch, Mizon and
Schmidt (1989) nicely summarize this literature and suggest as additional instruments
the use of the time-invariant variables x2,i1 − x̄2i up to x2,iT − x̄2i .
10.2.6 Robust Inference
Both the random effects and the ﬁxed effects models assume that the presence of
αi captures all correlation between the unobservables in different time periods. That
is, εit is assumed to be uncorrelated over individuals and time. Provided that the xit
variables are strictly exogenous, the presence of autocorrelation in εit does not result
in inconsistency of the standard estimators. It does, however, invalidate the standard
errors and resulting tests, just as we saw in Chapter 4. Moreover, it implies that the
estimators are no longer efﬁcient. For example, if the true covariance matrix  does
not satisfy (10.16), the random effects estimator no longer corresponds to the feasible
GLS estimator for β. As we know, the presence of heteroskedasticity in εit or – for
the random effects model – in αi , has similar consequences.
One way to avoid misleading inferences, without the need to impose alternative
assumptions on the structure of the covariance matrix , is the use of the OLS, random
effects or ﬁxed effects estimators for β, while adjusting its standard errors for general
forms of heteroskedasticity and autocorrelation. Consider the following model10

yit = xit β + uit ,

(10.34)

without the assumption that uit has an error components structure. Consistency of the
(pooled) OLS estimator
−1 N T
 N T



xit xit
xit yit
(10.35)
b=
i=1 t=1
10

i=1 t=1

For notational convenience, the constant is assumed to be included in xit .

MODELS BASED ON PANEL DATA

356

for β requires that

E{xit uit } = 0.

(10.36)

Assuming that error terms of different individuals are uncorrelated (E{uit ujs } = 0 for
all i = j ), the OLS covariance matrix can be estimated by a variant of the Newey–West
estimator from Chapter 4, given by

V̂ {b} =

N 
T

i=1 t=1

−1
xit xit

N 
T 
T

i=1 t=1 s=1


ûit ûis xit xis

N 
T


−1
xit xit

,

(10.37)

i=1 t=1

where ûit denotes the OLS residual. This estimator allows for general forms of heteroskedasticity as well as autocorrelation (within a given individual). If the (conditional) covariances of ui1 , . . . , uiT are not related to the explanatory variables, the
middle matrix in (10.37) can be replaced by
N 
T 
T

i=1 t=1 s=1




N
1 
û û x x  ,
N i=1 it is it is

(10.38)


where (1/N ) N
i=1 ûit ûis is a consistent estimator for ts = E{uit uis }. This allows
the variance of uit to be different over time, with arbitrary covariances. In a similar
fashion, it is also possible to construct a robust estimator for the covariance matrix of
the random effects estimator β̂RE using the transformed model in (10.20). Note that
the random effects estimator is not the appropriate EGLS estimator under these more
general conditions.
If uit has a time-invariant component αi that might be correlated with the explanatory
variables, the ﬁxed effects estimator would be more appropriate than OLS and a similar
correction for heteroskedasticity and autocorrelation (in εit ) can be employed (Arellano,
1987). The resulting expression is similar to (10.37) but replaces each xit with its
within transformation xit − x̄i and the OLS residual by the within residual. These
robust covariance matrix estimates are consistent for ﬁxed T and N → ∞ under weak
regularity conditions. If, conversely, N is ﬁxed and T → ∞, consistency requires the
use of Bartlett weights in (10.37) as discussed in Subsection 4.10.2; see Arellano (2003,
Section 2.3) for more details.
If one is willing to make speciﬁc assumptions about the form of heteroskedasticity or
autocorrelation, it is possible to improve upon the efﬁciency of the OLS, random effects
or ﬁxed effects estimators, by exploiting the structure of the error covariance matrix
using a feasible GLS or maximum likelihood approach. An overview of a number
of such estimators, which are typically computationally unattractive, is provided in
Baltagi (2001, Chapter 5). Kmenta (1986) suggests a relatively simple feasible GLS
estimator that allows for ﬁrst order autocorrelation in uit combined with individual
speciﬁc heteroskedasticity, but does not allow for a time-invariant component in uit
(see Baltagi, 2001, Section 10.4). Kiefer (1980) proposes a GLS estimator for the ﬁxed
effects model that allows for arbitrary covariances between εit and εis ; see Arellano
(2003, Section 2.3) or Hsiao (2003, Section 3.8) for more details. Wooldridge (2002,
Subsection 10.4.3) describes a feasible GLS estimator where the covariance matrix 
is estimated unrestrictedly from the pooled OLS residuals. Consistency of this estimator

THE STATIC LINEAR MODEL

357

basically requires the same conditions as required by the random effects estimator, but
it does not impose the error components structure. When N is sufﬁciently large relative
to T , this feasible GLS estimator may provide an attractive alternative to the random
effects approach.
10.2.7 Testing for Heteroskedasticity and Autocorrelation

Most of the tests that can be used for heteroskedasticity or autocorrelation in the random
effects model are computationally burdensome. For the ﬁxed effects model, which is
essentially estimated by OLS, things are relatively less complex. Fortunately, as the
ﬁxed effects estimator can be applied even if we make the random effects assumption
that αi is i.i.d. and independent of the explanatory variables, the tests for the ﬁxed
effects model can also be used in the random effects case.
A fairly simple test for autocorrelation in the ﬁxed effects model is based upon the
Durbin–Watson test discussed in Chapter 4. The alternative hypothesis is that
εit = ρεi,t−1 + vit ,

(10.39)

where vit is i.i.d. across individuals and time. This allows for autocorrelation over time
with the restriction that each individual has the same autocorrelation coefﬁcient ρ. The
null hypothesis under test is H0 : ρ = 0 against the one-sided alternative ρ < 0 or ρ > 0.
Let ε̂it denote the residuals from the within regression (10.9) or – equivalently – from
(10.7). Then Bhargava, Franzini and Narendranathan (1983) suggest the following
generalization of the Durbin–Watson statistic
N T
dwp =

2
t=2 (ε̂it − ε̂i,t−1 )
.
N T 2
i=1
t=1 ε̂it

i=1

(10.40)

Using similar derivations as Durbin and Watson, the authors are able to derive lower
and upper bounds on the true critical values that depend upon N, T , and K only. Unlike
the true time series case, the inconclusive region for the panel data Durbin–Watson
test is very small, particularly when the number of individuals in the panel is large. In
Table 10.1 we present some selected lower and upper bounds for the true 5% critical
values that can be used to test against the alternative of positive autocorrelation. The
numbers in the table conﬁrm that the inconclusive regions are small and also indicate
that the variation with K, N or T is limited. In a model with three explanatory variables
estimated over 6 time periods, we reject H0 : ρ = 0 at the 5% level if dwp is smaller
than 1.859 for N = 100 and 1.957 for N = 1000, both against the one-sided alternative
Table 10.1

5% lower and upper bounds panel Durbin–Watson test (Bhargava et al., 1983)
N = 100

T =6
T = 10

K
K
K
K

=3
=9
=3
=9

N = 500

N = 1000

dL

dU

dL

dU

dL

dU

1.859
1.839
1.891
1.878

1.880
1.902
1.904
1.916

1.939
1.935
1.952
1.949

1.943
1.947
1.954
1.957

1.957
1.954
1.967
1.965

1.959
1.961
1.968
1.970

MODELS BASED ON PANEL DATA

358

of ρ > 0. For panels with very large N , Bhargava et al. suggest simply to test if the
computed statistic dwp is less than two, when testing against positive autocorrelation.
Because the ﬁxed effects estimator is also consistent in the random effects model, it is
also possible to use this panel data Durbin–Watson test in the latter model.
To test for heteroskedasticity in εit , we can again use the ﬁxed effects residuals ε̂it .
The auxiliary regression of the test regresses the squared within residuals ε̂it2 upon a
constant and the J variables zit that we think may affect heteroskedasticity. This is a
variant of the Breusch–Pagan test11 for heteroskedasticity discussed in Chapter 4. Its
alternative hypothesis is that
V {εit } = σ 2 h(zit α),

(10.41)

where h is an unknown continuously differentiable function with h(0) = 1, so that
the null hypothesis that is tested is given by H0 : α = 0. Under the null hypothesis,
the test statistic, computed as N (T − 1) times the R 2 of the auxiliary regression, will
have an asymptotic Chi-squared distribution, with J degrees of freedom. An alternative
test can be computed from the residuals of the between regression, and is based upon
N times the R 2 of an auxiliary regression of the between residuals upon z̄i or, more
generally, upon zi1 , . . . , ziT . Under the null hypothesis of homoskedastic errors, the test
statistic has an asymptotic Chi-squared distribution, with degrees of freedom equal to
the number of variables included in the auxiliary regression (excluding the intercept).
The alternative hypothesis of the latter test is less well-deﬁned.

10.3

Illustration: Explaining Individual Wages

In this section we shall apply a number of the above estimators when estimating an
individual wage equation. The data12 are taken from the Youth Sample of the National
Longitudinal Survey held in the USA, and comprise a sample of 545 full-time working
males who have completed their schooling by 1980 and then followed over the period
1980–1987. The males in the sample are young, with an age in 1980 ranging from
17 to 23, and entered the labour market fairly recently, with an average of 3 years
of experience in the beginning of the sample period. The data and speciﬁcations we
choose are similar to those in Vella and Verbeek (1998). Log wages are explained from
years of schooling, years of experience and its square, dummy variables for being a
union member, working in the public sector and being married and two racial dummies.
The estimation results13 for the between estimator, based upon individual averages,
and the within estimator, based upon deviations from individual means, are given in the
ﬁrst two columns of Table 10.2. First of all, it should be noted that the ﬁxed effects or
within estimator eliminates any time-invariant variables from the model. In this case,
it means that the effects of schooling and race are wiped out. The differences between
11

In a panel data context, the term Breusch–Pagan test is usually associated with a Lagrange multiplier test
in the random effects model for the null hypothesis that there are no individual speciﬁc effects (σα2 = 0);
see Baltagi (2001, Subsection 4.2.1) or Wooldridge (2002, Subsection 10.4.4). In applications, this test
almost always rejects the null hypothesis.
12
The data used in this section are available as MALES.
13
The estimation results in this section are obtained by Stata 7.0.

ILLUSTRATION: EXPLAINING INDIVIDUAL WAGES

359

Table 10.2 Estimation results wage equation, males 1980–1987 (standard errors in
parentheses)
Dependent variable: log(wage)
Variable

Between

Fixed effects

OLS

Random effects

constant

0.490
(0.221)
0.095
(0.011)
−0.050
(0.050)
0.0051
(0.0032)
0.274
(0.047)
0.145
(0.041)
−0.139
(0.049)
0.005
(0.043)
−0.056
(0.109)

–

0.035
(0.039)

−0.034
(0.065)
0.099
(0.005)
0.089
(0.010)
−0.0028
(0.0007)
0.180
(0.017)
0.108
(0.016)
−0.144
(0.024)
0.016
(0.021)
0.004
(0.037)

−0.104
(0.111)
0.101
(0.009)
0.112
(0.008)
−0.0041
(0.0006)
0.106
(0.018)
0.063
(0.017)
−0.144
(0.048)
0.020
(0.043)
0.030
(0.036)

0.1782
0.0006
0.0642

0.1679
0.2027
0.1866

0.1776
0.1835
0.1808

schooling
experience
experience 2
union member
married
black
hispanic
public sector
within R 2
between R 2
overall R 2

0.0470
0.2196
0.1371

–
0.116
(0.008)
−0.0043
(0.0006)
0.081
(0.019)
0.045
(0.018)
–
–

the two sets of estimates seem substantial, and we shall come back to this below. In the
next column the OLS results are presented applied to the random effects model, where
the standard errors are not adjusted for the error components structure. The last column
presents the random effects EGLS estimator. As discussed in Subsection 10.2.2, the
variances of the error components αi and εit can be estimated from the within and
between residuals. In particular, we have σ̂B2 = 0.1209 and σ̂ε2 = 0.1234. From this,
we can consistently estimate σα2 as σ̂α2 = 0.1209 − 0.1234/8 = 0.1055. Consequently,
the factor ψ is estimated as
ψ̂ =

0.1234
= 0.1276,
0.1234 + 8 × 0.1055

leading to ϑ̂ = 1 − ψ̂ 1/2 = 0.6428. This means that the EGLS estimator can be obtained
from a transformed regression where 0.64 times the individual mean is subtracted
from the original data. Recall that OLS imposes ϑ = 0 while the ﬁxed effects estimator employs ϑ = 1. Note that both the OLS and the random effects estimates are in
between the between and ﬁxed effects estimates.
If the assumptions of the random effects model are satisﬁed, all four estimators in
Table 10.2 are consistent, the random effects estimator being the most efﬁcient one. If,
however, the individual effects αi are correlated with one or more of the explanatory
variables, the ﬁxed effects estimator is the only one that is consistent. This hypothesis
can be tested by comparing the between and within estimators, or the within and

MODELS BASED ON PANEL DATA

360

random effects estimators, which leads to tests that are equivalent. The simplest one
to perform is the Hausman test discussed in Subsection 10.2.3, based upon the latter
comparison. The test statistic takes a value of 31.75 and reﬂects the differences in the
coefﬁcients on experience, experience squared and the union, married and public sector
dummies. Under the null hypothesis, the statistic follows a Chi-squared distribution
with 5 degrees of freedom, so that we have to reject the null at any reasonable level
of signiﬁcance.
Marital status is a variable that is likely to be correlated with the unobserved heterogeneity in αi . Typically one would not expect an important causal effect of being
married upon one’s wage, so that the marital dummy is typically capturing other (unobservable) differences between married and unmarried workers. This is conﬁrmed by
the results in the table. If we eliminate the individual effects from the model and consider the ﬁxed effects estimator the effect of being married reduces to 4.5%, while for
the between estimator, for example, it is almost 15%. Note that the effect of being
married in the ﬁxed effects approach is identiﬁed only through people that change
marital status over the sample period. Similar remarks can be made for the effect of
union status upon a person’s wage. Recall, however, that all estimators assume that
the explanatory variables are uncorrelated with the idiosyncratic error term εit . If such
correlations would exist, even the ﬁxed effects estimator would be inconsistent. Vella
and Verbeek (1998) concentrate on the impact of endogenous union status on wages
for this group of workers and consider alternative, more complicated, estimators.
The goodness-of-ﬁt measures conﬁrm that the ﬁxed effects estimator results in the
largest within R 2 and thus explains the within variation as well as possible. The OLS
estimator maximizes the usual (overall) R 2 , while the random effects estimator results
in reasonable R 2 s in all dimensions. Recall that the standard errors of the OLS estimator
are misleading as they do not take into account the correlation across different error
terms. The correct standard errors for the OLS estimator should be larger than those
for the efﬁcient EGLS estimator that exploits these correlations.

10.4

Dynamic Linear Models

Among the major advantages of panel data is the ability to model individual dynamics.
Many economic models suggest that current behaviour depends upon past behaviour
(persistence, habit formation, partial adjustment, etc.), so in many cases we would like
to estimate a dynamic model on an individual level. The ability to do so is unique for
panel data.
10.4.1 An Autoregressive Panel Data Model

Consider the linear dynamic model with exogenous variables and a lagged dependent
variable, that is
yit = xit β + γ yi,t−1 + αi + εit ,
where it is assumed that εit is IID(0, σε2 ). In the static model, we have seen arguments
of consistency (robustness) and efﬁciency for choosing between a ﬁxed or random
effects treatment of the αi . In a dynamic model the situation is substantially different,

DYNAMIC LINEAR MODELS

361

because yi,t−1 will depend upon αi , irrespective of the way we treat αi . To illustrate
the problems that this causes, we ﬁrst consider the case where there are no exogenous
variables included and the model reads
yit = γ yi,t−1 + αi + εit ,

|γ | < 1.

(10.42)

Assume that we have observations on yit for periods t = 0, 1, . . . , T .
The ﬁxed effects estimator for γ is given by
N T
γ̂FE =

t=1 (yit − ȳi )(yi,t−1 − ȳi,−1 )
,
N T
2
i=1
t=1 (yi,t−1 − ȳi,−1 )

i=1

(10.43)



where ȳi = (1/T ) Tt=1 yit and ȳi,−1 = (1/T ) Tt=1 yi,t−1 . To analyse the properties of
γ̂FE we can substitute (10.42) into (10.43) to obtain
γ̂FE = γ +

N T

t=1 (εit − ε̄i )(yi,t−1 − ȳi,−1 )
.
N T
(1/(NT )) i=1 t=1 (yi,t−1 − ȳi,−1 )2

(1/(N T ))

i=1

(10.44)

This estimator, however, is biased and inconsistent for N → ∞ and ﬁxed T , as the
last term in the right-hand side of (10.44) does not have expectation zero and does not
converge to zero if N goes to inﬁnity. In particular, it can be shown that (see Nickell,
1981; or Hsiao, 2003, Section 4.2)
N
T
σ 2 (T − 1) − T γ + γ T
1 
(εit − ε̄i )(yi,t−1 − ȳi,−1 ) = − ε2 ·
= 0.
T
(1 − γ )2
N→∞ N T i=1 t=1
(10.45)
Thus, for ﬁxed T we have an inconsistent estimator. Note that this inconsistency is not
caused by anything we assumed about the αi s, as these are eliminated in estimation.
The problem is that the within transformed lagged dependent variable is correlated
with the within transformed error. If T → ∞, (10.45) converges to 0 so that the ﬁxed
effects estimator is consistent for γ if both T → ∞ and N → ∞.
One could think that the asymptotic bias for ﬁxed T is quite small and therefore not
a real problem. This is certainly not the case, as for ﬁnite T the bias can hardly be
ignored. For example, if the true value of γ equals 0.5, it can easily be computed that
(for N → ∞)
plim γ̂FE = −0.25 if T = 2

plim

plim γ̂FE = −0.04 if T = 3
plim γ̂FE = 0.33 if T = 10,
so even for moderate values of T the bias is substantial. Fortunately, there are relatively
easy ways to avoid these biases.
To solve the inconsistency problem, we ﬁrst of all start with a different transformation
to eliminate the individual effects αi , in particular we take ﬁrst differences. This gives
yit − yi,t−1 = γ (yi,t−1 − yi,t−2 ) + (εit − εi,t−1 ),

t = 2, . . . , T .

(10.46)

MODELS BASED ON PANEL DATA

362

If we estimate this by OLS we do not get a consistent estimator for γ because yi,t−1
and εi,t−1 are, by deﬁnition, correlated, even if T → ∞. However, this transformed
speciﬁcation suggests an instrumental variables approach. For example, yi,t−2 is correlated with yi,t−1 − yi,t−2 but not with εi,t−1 , unless εit exhibits autocorrelation (which
we excluded by assumption). This suggests an instrumental variables estimator14 for
γ as
N T
yi,t−2 (yit − yi,t−1 )
γ̂IV = N i=1T t=2
.
(10.47)
i=1
t=2 yi,t−2 (yi,t−1 − yi,t−2 )
A necessary condition for consistency of this estimator is that
N

plim

T


1
(ε − εi,t−1 )yi,t−2 = 0
N (T − 1) i=1 t=2 it

(10.48)

for either T , or N , or both going to inﬁnity. The estimator in (10.47) is one of the
estimators proposed by Anderson and Hsiao (1981). They also proposed an alternative,
where yi,t−2 − yi,t−3 is used as an instrument. This gives
(2)
γ̂IV

N T

(yi,t−2 − yi,t−3 )(yit − yi,t−1 )
= N i=1T t=3
,
i=1
t=3 (yi,t−2 − yi,t−3 )(yi,t−1 − yi,t−2 )

(10.49)

which is consistent (under regularity conditions) if
N

plim

T


1
(ε − εi,t−1 )(yi,t−2 − yi,t−3 ) = 0.
N (T − 2) i=1 t=3 it

(10.50)

Consistency of both of these estimators is guaranteed by the assumption that εit has
no autocorrelation.
Note that the second instrumental variables estimator requires an additional lag
to construct the instrument, such that the effective number of observations used in
estimation is reduced (one sample period is ‘lost’). The question which estimator one
should choose is not really an issue. A method of moments approach can unify the
estimators and eliminate the disadvantages of reduced sample sizes. A ﬁrst step in this
approach is to note that
N

T


1
plim
(ε − εi,t−1 )yi,t−2 = E{(εit − εi,t−1 )yi,t−2 } = 0
N (T − 1) i=1 t=2 it

(10.51)

is a moment condition (compare Chapter 5). Similarly,
N

T


1
(ε − εi,t−1 )(yi,t−2 − yi,t−3 )
plim
N (T − 2) i=1 t=3 it
= E{(εit − εi,t−1 )(yi,t−2 − yi,t−3 )} = 0
14

See Section 5.3 for a general introduction to instrumental variables estimation.

(10.52)

DYNAMIC LINEAR MODELS

363

is a moment condition. Both IV estimators thus impose one moment condition in
estimation. It is well known that imposing more moment conditions increases the
efﬁciency of the estimators (provided the additional conditions are valid, of course).
Arellano and Bond (1991) suggest that the list of instruments can be extended by
exploiting additional moment conditions and letting their number vary with t. To do
this, they keep T ﬁxed. For example, when T = 4, we have
E{(εi2 − εi1 )yi0 } = 0
as the moment condition for t = 2. For t = 3, we have
E{(εi3 − εi2 )yi1 } = 0,
but it also holds that
E{(εi3 − εi2 )yi0 } = 0.
For period t = 4, we have three moment conditions and three valid instruments
E{(εi4 − εi3 )yi0 } = 0
E{(εi4 − εi3 )yi1 } = 0
E{(εi4 − εi3 )yi2 } = 0.
All these moment conditions can be exploited in a GMM framework. To introduce
the GMM estimator deﬁne for general sample size T ,

εi = 

εi2 − εi1
εi,T




···
− εi,T −1

(10.53)

as the vector of transformed error terms, and



Zi = 



[yi0 ]

0

0
..
.

[yi0 , yi1 ]

0

···

...

0
0

..

.

0

0








(10.54)

[yi0 , . . . , yi,T −2 ]

as the matrix of instruments. Each row in the matrix Zi contains the instruments that
are valid for a given period. Consequently, the set of all moment conditions can be
written concisely as
E{Zi εi } = 0.
(10.55)
Note that these are 1 + 2 + 3 + · · · + T − 1 conditions. To derive the GMM estimator,
write this as
E{Zi (yi − γ yi,−1 )} = 0.
(10.56)

MODELS BASED ON PANEL DATA

364

Because the number of moment conditions will typically exceed the number of unknown
coefﬁcients, we estimate γ by minimizing a quadratic expression in terms of the corresponding sample moments (compare Chapter 5), that is





N
N
1  
1  
min
Z (yi − γ yi,−1 ) WN
Z (yi − γ yi,−1 ) ,
γ
N i=1 i
N i=1 i

(10.57)

where WN is a symmetric positive deﬁnite weighting matrix.15 Differentiating this with
respect to γ and solving for γ gives

γ̂GMM =

N




yi,−1
Zi

i=1


×

N



WN



yi,−1
Zi WN

i=1

N


−1
Zi yi,−1

i=1



N



Zi yi .

(10.58)

i=1

The properties of this estimator depend upon the choice for WN , although it is consistent
as long as WN is positive deﬁnite, for example, for WN = I , the identity matrix.
The optimal weighting matrix is the one that gives the most efﬁcient estimator, i.e.
that gives the smallest asymptotic covariance matrix for γ̂GMM . From the general theory
of GMM in Chapter 5, we know that the optimal weighting matrix is (asymptotically)
proportional to the inverse of the covariance matrix of the sample moments. In this
case, this means that the optimal weighting matrix should satisfy
plim WN = V {Zi εi }−1 = E{Zi εi εi Zi }−1 .

N→∞

(10.59)

In the standard case where no restrictions are imposed upon the covariance matrix of
εi , this can be estimated using a ﬁrst-step consistent estimator of γ and replacing the
expectation operator by a sample average. This gives

opt
ŴN

=

N
1  
Z ε̂ ε̂ Z
N i=1 i i i i

−1
,

(10.60)

where ε̂i is a residual vector from a ﬁrst-step consistent estimator, for example using
WN = I .
The general GMM approach does not impose that εit is i.i.d. over individuals and
time, and the optimal weighting matrix is thus estimated without imposing these restrictions. Note, however, that the absence of autocorrelation was needed to guarantee the
validity of the moment conditions. Instead of estimating the optimal weighting matrix
unrestrictedly, it is also possible (and potentially advisable in small samples) to impose
15

The sufﬁx N reﬂects that WN can depend upon the sample size N and does not reﬂect the dimension of
the matrix.

DYNAMIC LINEAR MODELS

365

the absence of autocorrelation in εit , combined with a homoskedasticity assumption.
Noting that under these restrictions

E{εi εi }

=

σε2 G

=

σε2

2


 −1


 0

..
.

−1
2

..

.

..

..

.

.

···

0


0 

,
−1 


−1

0


(10.61)

2

the optimal weighting matrix can be determined as

opt
WN

=

N
1  
Z GZi
N i=1 i

−1
.

(10.62)

Note that this matrix does not involve unknown parameters, so that the optimal GMM
estimator can be computed in one step if the original errors εit are assumed to be
homoskedastic and exhibit no autocorrelation.
Under weak regularity conditions, the GMM estimator for γ is asymptotically normal
for N → ∞ and ﬁxed T , with its covariance matrix given by


N
1  
plim 
yi,−1 Zi
N i=1
N→∞



N
1  
Z ε ε Z
N i=1 i i i i

−1 

−1
N
1  
 .
Z y
N i=1 i i,−1

(10.63)
This follows from the more general expressions in Section 5.6. With i.i.d. errors the
middle term reduces to

−1
N
1  
2 opt
2
σε W N = σ ε
Z GZi
.
N i=1 i
Alvarez and Arellano (2003) show that, in general, the GMM estimator is also consistent when both N and T tend to inﬁnity despite the fact that the number of moment
conditions tends to inﬁnity with the sample size. For large T , however, the GMM
estimator will be close to the ﬁxed effects estimator, which provides a more attractive
alternative.
10.4.2 Dynamic Models with Exogenous Variables

If the model also contains exogenous variables, we have
yit = xit β + γ yi,t−1 + αi + εit ,

(10.64)

which can also be estimated by the generalized instrumental variables or GMM
approach. Depending upon the assumptions made about xit , different sets of additional

MODELS BASED ON PANEL DATA

366

instruments can be constructed. If the xit are strictly exogenous in the sense that they
are uncorrelated with any of the εis error terms, we also have that
E{xis εit } = 0 for each s, t,

(10.65)

so that xi1 , . . . , xiT can be added to the instruments list for the ﬁrst-differenced equation
in each period. This would make the number of rows in Zi quite large. Instead, almost
the same level of information may be retained when the ﬁrst-differenced xit s are used
as their own instruments.16 In this case, we impose the moment conditions
E{xit εit } = 0 for each t
and the instrument matrix can be written as


[yi0 , xi2
]
0
···


0
[yi0 , yi1 , xi3 ]

Zi = 
..
..

.
.

0

···

0

(10.66)

0
0
0




.




[yi0 , . . . , yi,T −2 , xiT
]

If the xit variables are not strictly exogenous but predetermined, in which case current and lagged xit s are uncorrelated with current error terms, we only have that
E{xit εis } = 0 for s ≥ t. In this case, only xi,t−1 , . . . , xi1 are valid instruments for
the ﬁrst-differenced equation in period t. Thus, the moment conditions that can be
imposed are
E{xi,t−j εit } = 0 for j = 1, . . . , t − 1 (for each t).

(10.67)

In practice, a combination of strictly exogenous and predetermined x-variables may
occur rather than one of these two extreme cases. The matrix Zi should then be adjusted
accordingly. Baltagi (2001, Chapter 8) provides additional discussion and examples.
Arellano and Bover (1995) provide a framework to integrate the above approach
with the instrumental variables estimators of Hausman and Taylor (1981) and others
discussed in Subsection 10.2.5. Most importantly, they discuss how information in
levels can also be exploited in estimation. That is, in addition to the moment conditions
presented above, it is also possible to exploit the presence of valid instruments for the
levels equation (10.64) or its average over time (the between regression). This is of
particular importance when the γ coefﬁcient is close to unity; see also Blundell and
Bond (1998) and Arellano (2003, Section 6.6).

10.5

Illustration: Wage Elasticities of Labour Demand

In this section we consider a model that explains labour demand of ﬁrms from wages,
output, lagged labour demand and some other variables. Our goal is to derive estimates
16

We give up potential efﬁciency gains if some xit variables help ‘explaining’ the lagged endogenous
variables.

ILLUSTRATION: WAGE ELASTICITIES OF LABOUR DEMAND

367

for the short- and long-run wage elasticities of labour demand in Belgium. The data and
models are taken from Konings and Roodhooft (1997), who use a panel of 2800 large
Belgian ﬁrms over the period 1986–1994. On the basis of a wage-bargaining model
between unions and ﬁrms, the authors derive a static demand equation for labour as
log Lit = β1 + β2 log wit + β3 log Kit + β4 log Yit + β5 log wjt + uit ,
where Lit denotes desired employment of ﬁrm i in period t (labour demand), wit
denotes the unit cost of labour, while Kit and Yit denote capital stock and output level
respectively. The last variable wjt denotes the industry average of the real wage, which
is included because it reﬂects an outside option that is important for unions that bargain
about wages. This relationship is interpreted as a long-run result, as it ignores costs
of adjustment.
For the short run, Konings and Roodhooft experiment with alternative dynamic speciﬁcations. The simplest one assumes that
log Lit = β1 + β2 log wit + β3 log Kit + β4 log Yit + β5 log wjt + γ log Li,t−1 + uit .
In estimation, Yit is approximated by the value added. The dynamic model that we
estimate is then given by
log Lit = β1 + β2 log wit + β3 log Kit + β4 log Yit
+ β5 log wjt + γ log Li,t−1 + αi + εit ,
where it is assumed that the error term consists of two components. The component αi
denotes unobserved ﬁrm speciﬁc time-invariant heterogeneity. First differencing this
equation, as in the previous section, eliminates αi but does not result in an equation that
can be estimated consistently by OLS. First, there is correlation between  log Li,t−1
and εit (as above). Second, it is by no means obvious that the wage costs are given
exogenously. In particular, trade unions are assumed to bargain with employers over
wages and employment, in which case wages are determined simultaneously with
employment. Thus, we expect that
E{ log wit εit } = 0.
Therefore,  log wit is also instrumented in estimation. Valid instruments are given by
log wi,t−2 , log wi,t−3 , . . . , similar to the instruments for  log Li,t−1 . The number of
available instruments thus increases with t.
In Table 10.3 we present the estimation results of the static and dynamic model
discussed above. This is a subset of the results in Konings and Roodhooft (1997), who
also consider models with additional lags of the other variables. The ﬁrst column gives
the estimates for the static (long-run) labour demand function. The wage is treated as
endogenous and instrumented as indicated above. In the second column, lagged labour
demand is included, which is also instrumented in the manner described above. Both
speciﬁcations also include regional and time dummies. To test the model against a
non-speciﬁed alternative we can use the overidentifying restrictions tests, as discussed
in Chapter 5. The test statistics of 29.7 and 51.66 have to be compared with the critical

MODELS BASED ON PANEL DATA

368

Table 10.3

Estimation results labour demand equation (Konings and Roodhooft, 1997)

Dependent variable: log Lit
Variables

Static model

log Li,t−1
log Yit
log wit
log wjt
log Kit

–
0.021
−1.78
0.16
0.08

overidentifying restr. test
number of observations

29.7
(df = 15)
10599

(0.009)
(0.60)
(0.07)
(0.011)
(p = 0.013)

Dynamic model
0.60
0.008
−0.66
0.054
0.078
51.66
(df = 29)

(0.045)
(0.005)
(0.19)
(0.033)
(0.006)
(p = 0.006)

10599

values from a Chi-squared distribution with 15 and 29 degrees of freedom, respectively.
With p-values of 0.013 and 0.006 the overidentifying restrictions are, at the 1% level,
on the boundary of being rejected for both speciﬁcations. The signiﬁcance of the lagged
dependent variable (standard errors are given in parentheses) suggests that the dynamic
speciﬁcation should be preferred.
The estimated short-run wage elasticity from the last column is −0.66, while the
long-run elasticity is −0.66/(1 − 0.60) = −1.6, which is close to the estimate of −1.78
from the static long-run model. Both of these estimates are quite high. For example,
they suggest that in the long run a 1% wage increase results in a 1.6% decrease in
labour demand. These estimates are much higher than was initially believed based on
macro-economic time series data. Apparently, the possibility to correct for observed and
unobserved ﬁrm heterogeneity has a substantial impact on the estimates. A potential
problem of the results in Table 10.3 lies in the way the data are constructed. First,
the panel is unbalanced (see Section 10.7 below), while the model ignores changes
in labour demand due to ﬁrms that enter or leave the sample (for example, because
of ﬁnancial distress). In addition, employment is measured as the mean number of
employees in a given year, while wages (unit labour costs) are computed as total
labour costs divided by the number of employees. Clearly, this ignores the problem
of a reduction in average labour time per worker, which may have taken place in this
decade. For example, if a ﬁrm replaces one full-time worker by two part-time workers,
employment increases while labour costs decrease. In reality, however, no real changes
have taken place. See Konings and Roodhooft (1997) for additional discussion.

10.6

Nonstationarity, Unit Roots and Cointegration

The recent literature exhibits an increasing integration of techniques and ideas from
time-series analysis, such as unit roots and cointegration, into the area of panel data
modelling. The underlying reason for this development is that researchers have increasingly realized that cross-sectional information is a useful additional source of information that should be exploited. To analyse the effect of a certain policy measure, for
example adopting a road tax or a pollution tax, it may be more fruitful to compare with
other countries than to try to extract information about these effects from the country’s own history. Pooling data from different countries may also help to overcome

NONSTATIONARITY, UNIT ROOTS AND COINTEGRATION

369

the problem that sample sizes of time series are fairly small, so that tests regarding
long-run properties are not very powerful.
A number of recent articles discuss issues relating to unit roots, spurious regressions
and cointegration in panel data. Most of this literature focuses upon the case in which
the number of time periods T is fairly large, while the number of cross-sectional units
N is small or moderate. As a consequence, it is quite important to deal with potential
nonstationarity of the data series, while the presence of a unit root or cointegration may
be of speciﬁc economic interest. For example, a wide range of applications exist concerning purchasing power parity, including Oh (1996), focusing on (non)stationarity of
real exchange rates for a set of countries, or on testing for cointegration between nominal exchange rates and prices (compare Sections 8.5 and 9.3 and Subsection 9.5.4).
For ease of discussion, we shall refer below to the cross-sectional units as countries,
although they may also correspond to ﬁrms, industries or regions.
A crucial issue in analysing the time series on a number of countries simultaneously
is that of heterogeneity. Because it is possible to estimate a separate regression for
each country, it is natural to think of the possibility that model parameters are different
across countries, a case commonly referred to as ‘heterogeneous panels’. Robertson
and Symons (1992) and Pesaran and Smith (1995) stress the importance of parameter
heterogeneity in dynamic panel data models and analyse the potentially severe biases
that may arise from handling it in an inappropriate manner. Such biases are particularly
misleading in a nonstationary world as the relationships of the individual series may
be completely destroyed.
As long as we consider each time series individually, and the series are of sufﬁcient
length, there is nothing wrong with applying the time series techniques from Chapters 8
and 9. However, if we pool different series, we have to be aware of the possibility
that their processes not all have the same characteristics or are described by the same
parameters. For example, it is conceivable that yit is stationary for country 1 but
integrated of order one for country 2. Even when all variables are integrated of order
one in each country, heterogeneity in cointegration properties may lead to problems.
For example, if for each country i the variables yit and xit are cointegrated with
parameter βi , it holds that yit − βi xit is I (0) for each i, but in general there does not
exist a common cointegrating parameter β that makes yit − βxit stationaryfor all i.
Similarly, there is no guarantee that the cross-sectional averages ȳt = (1/N ) i yit and
x̄t are cointegrated, even if all underlying individual series are cointegrated.
In Subsections 10.6.1 and 10.6.2, we pay attention to panel data unit root tests and
cointegration tests, respectively. Basically, the tests are directed at testing the joint null
hypothesis of a unit root (or the absence of cointegration) for each of the countries
involved. In comparison to the single time-series case, panel data tests raise a number
of additional issues, including cross-sectional dependence, heterogeneity in dynamics
and error-term properties, and the type of asymptotics that is employed. While most
asymptotic analysis is done with both N and T tending to inﬁnity, there are various
ways that this can be done.
10.6.1 Panel Data Unit Root Tests

To introduce panel data unit root tests, consider the autoregressive model
yit = αi + γi yi,t−1 + εit ,

(10.68)

370

MODELS BASED ON PANEL DATA

which we can rewrite as
yit = αi + πi yi,t−1 + εit ,

(10.69)

where πi = γi − 1. The null hypothesis that all series have a unit root then becomes
H0 : πi = 0 for all i. A ﬁrst choice for the alternative hypothesis is that all series are
stationary with the same mean-reversion parameter, that is, H1 : πi = π < 0 for each
country i, and is used in the approaches of Levin and Lin (1992),17 Quah (1994)
and Harris and Tzavalis (1999). A more general alternative allows the mean-reversion
parameters to be potentially different across countries and states that H1 : πi < 0 for at
least one country i. This alternative is used by Maddala and Wu (1999), Choi (2001),
Im, Pesaran and Shin (2003)18 and others. As in the time-series case discussed in
Chapter 8, the properties of the test statistics (and their computation) depend crucially
upon the deterministic regressors included in the test equation. For example, in (10.69)
we have included a dummy for each country, corresponding to the ﬁxed effect. Alternative tests are available in cases where the equation includes a common intercept, or
in cases where a deterministic trend is added to the ﬁxed effect.
For all tests, the null hypothesis is that the time series of all individual countries
have a unit root. This implies that the null hypothesis can be rejected (in sufﬁciently
large samples) if any one of the N coefﬁcients πi is less than zero. Rejection of the
null hypothesis therefore does not indicate that all series are stationary. As Smith
and Fuertes (2003) note, if the hypothesis of interest is that all series are stationary
(for example, real exchange rates under purchasing power parity), it would be more
appropriate to use a panel version of the KPSS test, as discussed in Section 8.4, where
stationarity is the null hypothesis rather than the alternative. However, a test like this
may reject if just one series is nonstationary, which may not be interesting either.
Because of these issues, Maddala, Wu and Liu (2000) argue that for purchasing power
parity panel data unit root tests are the wrong answer to the low power of unit root
tests in single time series.
In addition to the choice of deterministic regressors in the test equations, panel data
unit root tests offer three additional technical issues in comparison with the single
time-series case. First, one has to make assumptions on the cross-sectional dependence
between εit s, noting that virtually all the existing nonstationary panel data literature
assume cross-sectional independence. Second, we need to be speciﬁc on the properties of εit and how they are allowed to vary across the different units. This includes
serial correlation and the possibility of heteroskedasticity across units. Third, asymptotic properties of estimators and tests depend crucially upon the way in which N ,
the number of cross-sectional units, and T , the number of time periods, tend to inﬁnity (see Phillips and Moon, 1999). Some tests assume that either T or N is ﬁxed
and assume that the other dimension tends to inﬁnity. Many tests are based on a
sequential limit, where ﬁrst T tends to inﬁnity for ﬁxed N , and subsequently N
tends to inﬁnity. Alternatively, some tests assume that both N and T tend to inﬁnity
along a speciﬁc path (e.g. T /N being ﬁxed). While the type of asymptotics that is
applied may seem a theoretical issue, remember that we are using asymptotic theory to approximate the properties of estimators and tests in the ﬁnite sample that we
17
18

A revised version of the Levin and Lin (1992) paper is available in Levin, Lin and Chu (2002).
A ﬁrst version of this paper dates back to 1995.

NONSTATIONARITY, UNIT ROOTS AND COINTEGRATION

371

happen to have. Although it is hard to make general statements on this matter, some
asymptotic approximations are simply better than others. Many papers in this area
therefore also contain a Monte Carlo study to analyse the ﬁnite sample behaviour of
the proposed tests under controlled circumstances. A common ﬁnding for many of
the tests below is that they tend to be oversized. That is, when the null hypothesis is true, the tests tend to reject more frequently than their nominal size (say, 5%)
suggests. Further, many tests do not perform very well when the error terms are crosssectionally correlated, or in the presence of cross-country cointegration. For example,
when real exchange rates are I (1) and cointegrated across countries, the null hypothesis tends to be rejected too often (see Banerjee, Marcellino and Osbat, 2001, for an
illustration).
While it is beyond the scope of this text to discuss alternative panel data unit
root tests in great technical detail, a brief discussion of some tests is warranted.
More details can be found in Banerjee (1999), Baltagi (2001, Chapter 12) or Enders
(2004, Section 4.11). Levin and Lin (1992) and Harris and Tzavalis (1999) base
their tests upon the OLS estimator for π, assuming that εit is i.i.d. across countries
and time. Depending upon the deterministic regressors included, the OLS estimator may be biased, even asymptotically. When ﬁxed effects are included, the estimator corresponds to the ﬁxed effects estimator for π based on (10.69), which is
biased for ﬁxed T (see Section 10.4). With appropriate correction and standardization factors, test statistics can be derived that are asymptotically normal for N → ∞
and ﬁxed T (Harris and Tzavalis) or both N, T → ∞ (Levin and Lin); see Baltagi (2001, Section 12.2). While the test statistics can be modiﬁed to allow for serial
correlation in εit , they do not allow cross-sectional dependence. This assumption is
rather strong, and as stressed by O’Connell (1998) in a panel study on purchasing power parity, allowing for cross-sectional dependence may substantially affect
inferences about the presence of a unit root. Because individual observations in a
panel typically have no natural ordering, modelling cross-sectional dependence is
not obvious.
The above two sets of tests are restrictive because they assume that πi is the
same across all countries, also under the alternative hypothesis. The test proposed
by Im, Pesaran and Shin (2003) allows πi to be different across individual units.
It is based on averaging the augmented Dickey–Fuller (ADF) test statistics (see
Section 8.4) over the cross-sectional units, while allowing for different orders of
serial correlation. In fact, the alternative hypothesis states that πi < 0 for at least
one i and thus allows that πi = 0 for a subset of the countries. Im, Pesaran and
Shin (2003) also propose a test based on the N Lagrange multiplier statistics for
πi = 0, averaged over all countries. The idea underlying these tests is quite simple: if
you have N independent test statistics, their average will be asymptotically normally
distributed for N → ∞. Consequently, the tests are based on comparison of appropriately scaled cross-sectional averages with critical values from a standard normal
distribution.
An alternative approach to combine information from individual unit root tests is
employed by Maddala and Wu (1999) and Choi (2001), who propose panel data unit
root tests based on combining the p-values of the N cross-sectional tests. Let pi denote
the p-value of the (augmented) Dickey–Fuller test for unit i. Under the null hypothesis,
pi will have a uniform distribution over the interval [0, 1], small values corresponding

MODELS BASED ON PANEL DATA

372

to rejection. The combined test statistic is given by
P = −2

N


log pi .

(10.70)

i=1

For ﬁxed N , this test statistic will have a Chi-squared distribution with 2N degrees of
freedom as T → ∞, so that large values of P lead us to reject the null hypothesis.
While this test (sometimes referred to as the Fisher test) is attractive because it allows
the use of different ADF tests and different time-series lengths per unit, a disadvantage is that it requires individual p-values that have to be derived by Monte Carlo
simulations.
While the latter tests may seem attractive and easy to use, a word of caution is
appropriate. Before one can apply the individual ADF tests underlying the Maddala
and Wu (1999) and Im, Pesaran and Shin (2003) approaches, one has to determine the
number of lags and determine whether a trend should be included. It is not obvious
how this should be done. For a single time series, a common approach is to perform the
ADF test for a range of alternative lag values. For example, in Table 8.2 we presented
26 different (augmented) Dickey–Fuller test statistics for the log price index. If we
were to combine the ADF tests for N different countries, in whatever way, this creates
a wide range of possible combinations. Smith and Fuertes (2003) warn for pre-test
biases in this context.
10.6.2 Panel Data Cointegration Tests

A wide range of alternative tests is available to test for cointegration in a dynamic
panel data setting, and research in this area is evolving rapidly. A substantial number
of these tests are based on testing for a unit root in the residuals of a panel cointegrating regression. The drawbacks and complexities associated with the panel unit root
tests are also relevant in the cointegration case. Several additional issues are of potential importance when testing for cointegration: heterogeneity in the parameters of the
cointegrating relationships, heterogeneity in the number of cointegrating relationships
across countries and the possibility of cointegration between the series from different
countries. A ﬁnal issue is that of estimating the cointegrating vectors, for which several
alternative estimators are available, with different small and large sample properties
(depending upon the type of asymptotics that is chosen).
When the cointegrating relationship is unknown, which is almost always the case,
most cointegration tests start with estimating the cointegrating regression. Let us focus
on the bivariate case and write the panel regression as
yit = αi + βi xit + εit ,

(10.71)

where both yit and xit are integrated of order one. Cointegration implies that εit is stationary for each i. Homogeneous cointegration, in addition, requires that βi = β. If the
cointegrating parameter is heterogeneous, and homogeneity is imposed, one estimates
yit = αi + βxit + [(βi − β)xit + εit ],

(10.72)

MODELS WITH LIMITED DEPENDENT VARIABLES

373

and in general the composite error term is integrated of order one, even if εit is stationary. However, the problem of spurious regressions may be less relevant in this
situation. This is because a pooled estimator will also average over i, so that the noise
in the equation will be attenuated. In many circumstances, when N → ∞, the ﬁxed
effects estimator for β is actually consistent for the long-run average relation parameter, as well as asymptotically normal, despite the absence of cointegration (see Phillips
and Moon, 1999). However, the meaning of this long-run relationship, in the absence
of cointegration, is open to some interpretation (see Hsiao, 2003, Section 10.2 for some
discussion). With heterogeneous cointegration, the long-run average estimated from the
pooled regression may differ substantially from the average of the cointegration parameters averaged over countries (see Pesaran and Smith, 1995). Consequently, if there is
heterogeneous cointegration, it is much better to estimate the individual cointegrating
regressions rather than using a pooled estimator. Obviously, this requires T → ∞.
To test for cointegration, the panel data unit root tests from the previous section can
be applied to the residuals from these regressions, provided that the critical values are
appropriately adjusted (see Pedroni, 1999, or Kao, 1999). Recall that these tests assume
cross-sectional independence. Some tests assume homogeneity of the cointegrating
parameter and use a pooled OLS or dynamic OLS estimator (see Subsection 9.2.2).
Additional discussion on these tests can be found in Banerjee (1999), Baltagi (2001,
Chapter 4) or Smith and Fuertes (2003).
With more than two variables, an additional complication may arise because more
than one cointegrating relationship may exist for one or more of the countries. Further,
even with one cointegrating vector per country, the results will be sensitive to the
normalization constraint (left-hand side variable) that is chosen. Finally, the existence
of between-country cointegration may seriously distort the results of within-country
cointegration tests.

10.7

Models with Limited Dependent Variables

Panel data are relatively often used in micro-economic problems where the models of
interest involve nonlinearities. Discrete or limited dependent variables are an important
phenomenon in this area, and their combination with panel data usually complicates
estimation. The reason is that with panel data it can usually not be argued that different observations on the same unit are independent. Correlations between different error
terms typically complicate the likelihood functions of such models and therefore complicate their estimation. In this section we discuss the estimation of panel data logit,
probit and tobit models. More details on panel data models with limited dependent
variables can be found in Maddala (1987) or Hsiao (2003, Chapters 7–8).
10.7.1 Binary Choice Models

As in the cross-sectional case, the binary choice model is usually formulated in terms
of an underlying latent model. Typically, we write19
yit∗ = xit β + αi + εit ,
19

To simplify the notation we shall assume that xit includes a constant, whenever appropriate.

(10.73)

MODELS BASED ON PANEL DATA

374

where we observe yit = 1 if yit∗ > 0 and yit = 0 otherwise. For example, yit may indicate whether person i is working in period t or not. Let us assume that the idiosyncratic
error term εit has a symmetric distribution with distribution function F (.), i.i.d. across
individuals and time and independent of all xis . Even in this case the presence of αi
complicates estimation, both when we treat them as ﬁxed unknown parameters and
when we treat them as random error terms.
If we treat αi as ﬁxed unknown parameters, we are essentially including N dummy
variables in the model. The loglikelihood function is thus given by (compare (7.12))

yit log F (αi + xit β)
log L(β, α1 , . . . , αN ) =
i,t

+



(1 − yit ) log[1 − F (αi + xit β)].

(10.74)

i,t

Maximizing this with respect to β and αi (i = 1, . . . , N ) results in consistent estimators
provided that the number of time periods T goes to inﬁnity. For ﬁxed T and N → ∞,
the estimators are inconsistent. The reason is that for ﬁxed T , the number of parameters
grows with sample size N and we have what is known as an ‘incidental parameter’
problem. Clearly, we can only estimate αi consistently if the number of observations for
individual i grows, which requires that T tends to inﬁnity. In general, the inconsistency
of α̂i for ﬁxed T will carry over to the estimator for β.
The incidental parameter problem, where the number of parameters increases with
the number of observations, arises in any ﬁxed effects model, including the linear
model; see Lancaster (2000) for a recent discussion. For the linear case, however, it
was possible to eliminate the αi s, such that β could be estimated consistently, even
though all the αi parameters could not. For most nonlinear models, however, the
inconsistency of α̂i leads to inconsistency of the other parameter estimators as well.
Also note that from a practical point of view, the estimation of more than N parameters
may not be very attractive if N is fairly large.
Although it is possible to transform the latent model such that the individual effects
αi are eliminated, this does not help in this context because there is no mapping from,
∗
for example, yit∗ − yi,t−1
, to observables like yit − yi,t−1 . An alternative strategy is the
use of conditional maximum likelihood (see Andersen, 1970; or Chamberlain, 1980).
In this case, we consider the likelihood function conditional upon a set of statistics ti
that are sufﬁcient for αi . This means that conditional upon ti an individual’s likelihood
contribution no longer depends upon αi , but still depends upon the other parameters
β. In the panel data binary choice model, the existence of a sufﬁcient statistic depends
upon the functional form of F , that is, depends upon the distribution of εit .
At the general level let us write the joint density or probability mass function of
yi1 , . . . , yiT as f (yi1 , . . . , yiT |αi , β), which depends upon the parameters β and αi .
If a sufﬁcient statistic ti exists, this means that there exists an observable variable ti
such that f (yi1 , . . . , yiT |ti , αi , β) = f (yi1 , . . . , yiT |ti , β) and so does not depend upon
αi . Consequently, we can maximize the conditional likelihood function, based upon
f (yi1 , . . . , yiT |ti , β) to get a consistent estimator for β. Moreover, we can use all the
distributional results from Chapter 6 if we replace the loglikelihood by the conditional
loglikelihood function. For the linear model with normal errors, a sufﬁcient statistic
for αi is ȳi . That is, the conditional distribution of yit given ȳi does not depend upon

MODELS WITH LIMITED DEPENDENT VARIABLES

375

αi , and maximizing the conditional likelihood function can be shown to reproduce the
ﬁxed effects estimator for β. Unfortunately, this result does not automatically extend
to nonlinear models. For the probit model, for example, it has been shown that no
sufﬁcient statistic for αi exists. This means that we cannot estimate a ﬁxed effects
probit model consistently for ﬁxed T .
10.7.2 The Fixed Effects Logit Model

For the ﬁxed effects logit model, the situation is different. In this model ti = ȳi is
a sufﬁcient statistic for αi and consistent estimation is possible by conditional maximum likelihood. It should be noted that the conditional distribution of yi1 , . . . , yiT
is degenerate if ti = 0 or ti = 1. Consequently, such individuals do not contribute to
the conditional likelihood and should be discarded in estimation. Put differently, their
behaviour would be completely captured by their individual effect αi . This means
that only individuals that change status at least once are relevant for estimating β. To
illustrate the ﬁxed effects logit model, we consider the case with T = 2.
Conditional upon ti = 1/2, the two possible outcomes are (0, 1) and (1, 0). The
conditional probability of the ﬁrst outcome is
P {(0, 1)|ti = 1/2, αi , β} =

P {(0, 1)|αi , β}
.
P {(0, 1)|αi , β} + P {(1, 0)|αi , β}

(10.75)

Using that
P {(0, 1)|αi , β} = P {yi1 = 0|αi , β}P {yi2 = 1|αi , β}
with20
P {yi2 = 1|αi , β} =


β}
exp{αi + xi2
,

1 + exp{αi + xi2
β}

it follows that the conditional probability is given by
P {(0, 1)|ti = 1/2, αi , β} =

exp{(xi2 − xi1 ) β}
,
1 + exp{(xi2 − xi1 ) β}

(10.76)

which indeed does not depend upon αi . Similarly,
P {(1, 0)|ti = 1/2, αi , β} =

1
.
1 + exp{(xi2 − xi1 ) β}

(10.77)

This means that we can estimate the ﬁxed effect logit model for T = 2 using a standard
logit with xi2 − xi1 as explanatory variables and the change in yit as the endogenous
event (1 for a positive change, 0 for a negative one). Note that in this ﬁxed effects
binary choice model it is even more clear than in the linear case that the model is only
identiﬁed through the ‘within dimension’ of the data; individuals who do not change
status are simply discarded in estimation as they provide no information whatsoever
about β. For the case with larger T it is a bit more cumbersome to derive all the
20

See (7.6) in Chapter 7 for the logistic distribution function.

MODELS BASED ON PANEL DATA

376

necessary conditional probabilities, but in principle it is a straightforward extension of
the above case (see Chamberlain, 1980, or Maddala, 1987). Chamberlain (1980) also
discusses how the conditional maximum likelihood approach can be extended to the
multinomial logit model.
If it can be assumed that the αi are independent of the explanatory variables in xit ,
a random effects treatment seems more appropriate. This is most easily achieved in
the context of a probit model.
10.7.3 The Random Effects Probit Model

Let us start with the latent variable speciﬁcation
yit∗ = xit β + uit ,
with

(10.78)

yit = 1 if yit∗ > 0
yit = 0 if yit∗ ≤ 0,

(10.79)

where uit is an error term with mean zero and unit variance, independent of
(xi1 , . . . , xiT ). To estimate β by maximum likelihood we will have to complement
this with an assumption about the joint distribution of ui1 , . . . , uiT . The likelihood
contribution of individual i is the (joint) probability of observing the T outcomes
yi1 , . . . , yiT . This joint probability is determined from the joint distribution of the latent
∗
∗
variables yi1
, . . . , yiT
by integrating over the appropriate intervals. In general, this will
thus imply T integrals, which in estimation are typically to be computed numerically.
When T = 4 or more, this makes maximum likelihood estimation infeasible. It
is possible to circumvent this ‘curse of dimensionality’ by using simulation-based
estimators, as discussed in, for example, Keane (1993), Weeks (1995) and Hajivassiliou
and McFadden (1998). Their discussion is beyond the scope of this text.
Clearly, if it can beassumed that all uit are independent, we have that f (yi1 , . . . ,
yiT |xi1 , . . . , xiT , β) = t f (yit |xit , β), which involves T one-dimensional integrals
only (as in the cross-sectional case). If we make an error components assumption,
and assume that uit = αi + εit , where εit is independent over time (and individuals),
we can write the joint probability as
 ∞
f (yi1 , . . . , yiT |xi1 , . . . , xiT , β) =
f (yi1 , . . . , yiT |xi1 , . . . , xiT , αi , β)f (αi )dαi

=

−∞
∞
−∞




t


f (yit |xit , αi , β) f (αi )dαi ,

(10.80)

which requires numerical integration over one dimension. This is a feasible speciﬁcation that allows the error terms to be correlated across different periods, albeit in a
restrictive way. The crucial step in (10.80) is that conditional upon αi the errors from
different periods are independent.
In principle arbitrary assumptions can be made about the distributions of αi and
εit . For example, one could assume that εit is i.i.d. normal while αi has a logistic

MODELS WITH LIMITED DEPENDENT VARIABLES

377

distribution. However, this may lead to distributions for αi + εit that are nonstandard.
For example, the sum of two logistically distributed variables in general does not have
a logistic distribution. This implies that individual probabilities, like f (yit |xit , β), are
hard to compute, and do not correspond to a cross-sectional probit or logit model.
Therefore, it is more common to start from the joint distribution of ui1 , . . . , uiT . The
multivariate logistic distribution has the disadvantage that all correlations are restricted
to be 1/2 (see Maddala, 1987), so that it is not very attractive in practice. Consequently,
the most common approach is to start from a multivariate normal distribution, which
leads to the random effects probit model.
Let us assume that the joint distribution of ui1 , . . . , uiT is normal with zero means
and variances equal to 1 and cov{uit , uis } = σα2 , s = t. This corresponds to assuming
that αi is NID(0, σα2 ) and εit is NID(0, 1 − σα2 ). Recall that as in the cross-sectional
case we need a normalization on the errors’ variances. The normalization chosen here
implies that the error variance in a given period is unity, such that the estimated β
coefﬁcients are directly comparable to estimates obtained from estimating the model
from one wave of the panel using cross-sectional probit maximum likelihood. For the
random effects probit model, the expressions in the likelihood function are given by

xit β + αi
f (yit |xit , αi , β) =  
1 − σα2


xit β + αi
=1− 
1 − σα2


if yit = 1
if yit = 0,

(10.81)

where  denotes the cumulative density function of the standard normal distribution.
The density of αi is given by


1
1 α2
exp − i2 .
f (αi ) = 
2 σα
2πσα2

(10.82)

The integral in (10.80) has to be computed numerically, which can be done using
the algorithm described in Butler and Mofﬁtt (1982). Several software packages (for
example, LIMDEP and Stata) have standard routines for estimating the random effects
probit model.
It can be shown (Robinson, 1982) that ignoring the correlations across periods and
estimating the β coefﬁcients using standard probit maximum likelihood on the pooled
data is consistent, though inefﬁcient. Moreover, routinely computed standard errors are
incorrect. Nevertheless, these values can be used as initial estimates in an iterative
maximum likelihood procedure based on (10.80).
10.7.4 Tobit Models

The random effects tobit model is very similar to the random effects probit model, the
only difference is in the observation rule. Consequently, we can be fairly brief here.
Let us start with
yit∗ = xit β + αi + εit ,
(10.83)

MODELS BASED ON PANEL DATA

378

while

yit = yit∗

if yit∗ > 0

yit = 0

if yit∗ ≤ 0.

(10.84)

We make the usual random effects assumption that αi and εit are i.i.d. normally
distributed, independent of xi1 , . . . , xiT , with zero means and variances σα2 and σε2 ,
respectively. Using f as generic notation for a density or probability mass function,
the likelihood function can be written as in (10.80),

f (yi1 , . . . , yiT |xi1 , . . . , xiT , β) =

∞



−∞

t

f (yit |xit , αi , β)f (αi )dαi ,

where f (αi ) is given by (10.82) and f (yit |xit , αi , β) is given by


1 (yit − xit β − αi )2
f (yit |xit , αi , β) = 
exp −
2
σε2
2πσε2

 
x β + αi
= 1 −  it
σε
1

if yit > 0
if yit = 0.

(10.85)

Note that the latter two expressions are similar to the likelihood contributions in the
cross-sectional case, as discussed in Chapter 7. The only difference is the inclusion of
αi in the conditional mean.
In a completely similar fashion, other forms of censoring can be considered, to obtain,
for example, the random effects ordered probit model. In all cases, the integration over
αi has to be done numerically.
10.7.5 Dynamics and the Problem of Initial Conditions

The possibility of including a lagged dependent variable in the above models is of
economic interest. For example, suppose we are explaining whether or not an individual
is unemployed over a number of consecutive months. It is typically the case that
individuals who have a longer history of being unemployed are less likely to leave the
state of unemployment. As discussed in the introductory section of this chapter there
are two explanations for this: an individual with a longer unemployment history may
be discouraged in looking for a job or may (for whatever reason) be less attractive
for an employer to hire. This is referred to as state dependence: the longer you are
in a certain state the less likely you are to leave it. Alternatively, it is possible that
unobserved heterogeneity is present such that individuals with certain unobserved
characteristics are less likely to leave unemployment. The fact that we observe a
spurious state dependence in the data is simply due to a selection mechanism: the
long-term unemployed have certain unobservable (time-invariant) characteristics that
make it less likely for them to ﬁnd a job anyhow. In the binary choice models discussed
above, the individual effects αi capture the unobserved heterogeneity. If we include a
lagged dependent variable, we can distinguish between the above two explanations.

MODELS WITH LIMITED DEPENDENT VARIABLES

379

Let us consider the random effect probit model, although similar results hold for the
random effects tobit case. Suppose the latent variable speciﬁcation is changed into
yit∗ = xit β + γ yi,t−1 + αi + εit ,

(10.86)

with yit = 1 if yit∗ > 0 and 0 otherwise. In this model γ > 0 indicates positive state
dependence: the ceteris paribus probability that yit = 1 is larger if yi,t−1 is also one.
Let us consider maximum likelihood estimation of this dynamic random effects probit
model, making the same distributional assumptions as before. In general terms, the
likelihood contribution of individual i is given by21
f (yi1 , . . . , yiT |xi1 , . . . , xiT , β)
 ∞
f (yi1 , . . . , yiT |xi1 , . . . , xiT , αi , β)f (αi )dαi
=

=
where

−∞
∞

T


−∞

t=2


f (yit |yi,t−1 , xit , αi , β) f (yi1 |xi1 , αi , β)f (αi )dαi , (10.87)


xit β + γ yi,t−1 + αi
f (yit |yi,t−1 , xit , αi , β) = 

1 − σα2


xit β + γ yi,t−1 + αi

=1−
1 − σα2


if yit = 1
if yit = 0.

This is completely analogous to the static case and yi,t−1 is simply included as an
additional explanatory variable. However, the term f (yi1 |xi1 , αi , β) in the likelihood
function may cause problems. It gives the probability of observing yi1 = 1 or 0 without
knowing the previous state but conditional upon the unobserved heterogeneity term αi .
If the initial value is exogenous in the sense that its distribution does not depend upon
αi , we can put the term f (yi1 |xi1 , αi , β) = f (yi1 |xi1 , β) outside the integral. In this
case, we can simply consider the likelihood function conditional upon yi1 and ignore
the term f (yi1 |xi1 , β) in estimation. The only consequence may be a loss of efﬁciency
if f (yi1 |xi1 , β) provides information about β. This approach would be appropriate if
the initial state is necessarily the same for all individuals or if it is randomly assigned
to individuals. An example of the ﬁrst situation is given in Nijman and Verbeek (1992),
who model nonresponse with respect to consumption and the initial period refers to
the month before the panel and no nonresponse was necessarily observed.
However, it may be hard to argue in many applications that the initial value yi1 is
exogenous and does not depend upon a person’s unobserved heterogeneity. In that case
we would need an expression for f (yi1 |xi1 , αi , β) and this is problematic. If the process
that we are estimating has been going on for a number of periods before the current
sample period, f (yi1 |xi1 , αi , β) is a complicated function that depends upon person i’s
unobserved history. This means that it is typically impossible to derive an expression for
21


For notational convenience, the time index is deﬁned such that the ﬁrst observation is (yi1 , xi1
).

MODELS BASED ON PANEL DATA

380

the marginal probability f (yi1 |xi1 , αi , β) that is consistent with the rest of the model.
Heckman (1981) suggests an approximate solution to this initial conditions problem
that appears to work reasonably well in practice. It requires an approximation for the
marginal probability of the initial state by a probit function using as much pre-sample
information as available, without imposing restrictions between its coefﬁcients and
the structural β and γ parameters. Vella and Verbeek (1999a) provide an illustration
of this approach in a dynamic random effects tobit model. The impact of the initial
conditions diminishes if the number of sample periods T increases, so one may decide
to ignore the problem when T is fairly large; see Hsiao (2003, Subsection 7.5.2) for
more discussion.
10.7.6 Semi-parametric Alternatives

The binary choice and censored regression models discussed above suffer from two
important drawbacks. First, the distribution of εit conditional upon xit (and αi ) needs
to be speciﬁed, and second, with the exception of the ﬁxed effects logit model, there
is no simple way to estimate the models treating αi as ﬁxed unknown parameters.
Several semi-parametric approaches have been suggested for these models that do not
require strong distributional assumptions on εit and somehow allow αi to be eliminated
before estimation.
In the binary choice model, it is possible to obtain semi-parametric estimators for
β that are consistent up to a scaling factor whether αi is treated as ﬁxed or random.
For example, Manski (1987) suggests a √
maximum score estimator (compare Subsection 7.1.8), while Lee (1999) provides a N -consistent estimator for the static binary
choice model; see Hsiao (2003, Section 7.4) for more discussion. Honoré and Kyriazidou (2000) propose a semi-parametric estimator for discrete choice models with a
lagged dependent variable.
A tobit model as well as a truncated regression model with ﬁxed effects can be
estimated consistently using the generalized method of moments exploiting the moment
conditions given by Honoré (1992) or Honoré (1993) for the dynamic model. The
essential trick of these estimators is that a ﬁrst-difference transformation, for appropriate
subsets of the observations, no longer involves the incidental parameters αi ; see Hsiao
(2003, Sections 8.4 and 8.6) for more discussion.

10.8

Incomplete Panels and Selection Bias

Because of a variety of reasons, empirical panel data sets are often incomplete. For
example, after a few waves of the panel people may refuse cooperation, households
may not be located again or have split up, ﬁrms may have ﬁnished business or have
merged with another ﬁrm, or investment funds may be closed down. On the other
hand, ﬁrms may enter business at a later stage, refreshment samples may have been
drawn to compensate attrition, or the panel may be collected as a rotating panel. In a
rotating panel, each period a ﬁxed proportion of the units is replaced. A consequence
of all these events is that the resulting panel data set is no longer rectangular. If the
total number of individuals equals N and the number of time periods is T then the
total number of observations is substantially smaller than NT .

INCOMPLETE PANELS AND SELECTION BIAS

381

A ﬁrst consequence of working with an incomplete panel is a computational one.
Most of the expressions for the estimators given above are no longer appropriate if
observations are missing. A simple ‘solution’ is to discard any individual from the panel
that has incomplete information and to work with the completely observed units only.
In this approach estimation uses the balanced sub-panel only. This is computationally
attractive, but potentially highly inefﬁcient: a substantial amount of information may
be ‘thrown away’. This loss in efﬁciency can be prevented by using all observations
including those on individuals that are not observed in all T periods. This way, one
uses the unbalanced panel. In principle this is straightforward, but computationally it
requires some adjustments to the formulae in the previous sections. We shall discuss
some of these adjustments in Subsection 10.8.1. Fortunately, most software that can
handle panel data also allows for unbalanced data.
Another potential and even more serious consequence of using incomplete panel data
is the danger of selection bias. If individuals are incompletely observed because of an
endogenous reason, the use of either the balanced sub-panel or the unbalanced panel
may lead to biased estimators and misleading tests. To elaborate upon this, suppose
that the model of interest is given by
yit = xit β + αi + εit .

(10.88)

Furthermore, deﬁne the indicator variable rit (‘response’) as rit = 1 if (xit , yit ) is
observed and 0 otherwise. The observations on (xit , yit ) are missing at random if rit is
independent of αi and εit . This means that conditioning upon the outcome of the selection process does not affect the conditional distribution of yit given xit . If we want to
concentrate upon the balanced sub-panel, the conditioning is upon ri1 = · · · = riT = 1
and we require that rit is independent of αi and εi1 , . . . , εiT . In these cases, the usual
consistency properties of the estimators are not affected if we restrict attention to the
available or complete observations only. If selection depends upon the equations’ error
terms the OLS, random effects and ﬁxed effects estimators may suffer from selection
bias (compare Chapter 7). Subsection 10.8.2 provides additional details on this issue,
including some simple tests. In cases with selection bias, alternative estimators have to
be used, which are typically computationally unattractive. This is discussed in Subsection 10.8.3. Additional details and discussion on incomplete panels and selection bias
can be found in Verbeek and Nijman (1992, 1996).
10.8.1 Estimation with Randomly Missing Data
The expressions for the ﬁxed and random effects estimators are easily extended to
the unbalanced case. The ﬁxed effects estimator, as before, can be determined as
the OLS estimator in the linear model where each individual has its own intercept
term. Alternatively, the resulting estimator for β can be obtained directly by applying
OLS to the within transformed model, where now all variables are in deviation from
the mean over the available observations. Individuals that are observed only once
provide no information on β and should be discarded in estimation. Deﬁning ‘available
means’ as22
T
T
t=1 rit yit
t=1 rit xit
ȳi = 
;
x̄
=
,
T
i
T
t=1 rit
t=1 rit
22

We assume that

T

t=1 rit

≥ 1, i.e. each individual is observed at least once.

MODELS BASED ON PANEL DATA

382

the ﬁxed effects estimator can be concisely written as

β̂FE =

T
N 


−1


rit (xit − x̄i )(xit − x̄i )

i=1 t=1

T
N 


rit (xit − x̄i )(yit − ȳi ).

(10.89)

i=1 t=1

That is, all sums are simply over the available observations only.
In a similar way, the random effects estimator can be generalized. The random
effects estimator for the unbalanced case can be obtained from

β̂GLS =

N 
T


rit (xit − x̄i )(xit − x̄i ) +

i=1 t=1



×
T

t=1 rit

−1
ψi Ti (x̄i − x̄)(x̄i − x̄)

i=1

N 
T


rit (xit − x̄i )(yit − ȳi ) +

i=1 t=1

where Ti =

N


N



ψi Ti (x̄i − x̄)(ȳi − ȳ) ,

(10.90)

i=1

denotes the number of periods individual i is observed and
ψi =

σε2
.
σε2 + Ti σα2

Alternatively, it is obtained from OLS applied to the following transformed model
(yit − ϑi ȳi ) = µ(1 − ϑi ) + (xit − ϑi x̄i ) β + uit ,

(10.91)

1/2

where ϑi = 1 − ψi . Note that the transformation applied here is individual speciﬁc
as it depends upon the number of observations for individual i.
Essentially, the more general formulae for the ﬁxed effects and random effects
estimators are characterized by the fact that all summations and means are over the
available observations only and that Ti replaces T . Completely analogous adjustments
apply to the expressions for the covariance matrices of the two estimators given in
(10.13) and (10.23). Consistent estimators for the unknown variances σα2 and σε2 are
given by
σ̂ε2 = N

i=1

1

N 
T


Ti − N

i=1 t=1

and
σ̂α2


2
rit yit − ȳi − (xit − x̄i ) β̂FE


N 
1 
1 2

2
=
(ȳi − x̄i β̂B ) − σ̂ε ,
N i=1
Ti

(10.92)

(10.93)

respectively, where β̂B is the between estimator for β (computed as the OLS estimator
in (10.18), where the means now reﬂect ‘available means’). Because the efﬁciency of
the estimators for σα2 and σε2 asymptotically has no impact on the efﬁciency of the
random effects estimator, it is possible to use computationally simpler estimators for
σα2 and σε2 that are consistent. For example, one could use the standard estimators

INCOMPLETE PANELS AND SELECTION BIAS

383

computed from the residuals obtained from estimating with the balanced sub-panel
only, and then use (10.90) or (10.91) to compute the random effects estimator.
10.8.2 Selection Bias and Some Simple Tests

In addition to the usual conditions for consistency of the random effects and ﬁxed
effects estimator, based on either the balanced sub-panel or the unbalanced panel, it
was assumed above that the response indicator variable rit was independent of all unobservables in the model. This assumption may be unrealistic. For example, explaining
the performance of mutual funds may suffer from the fact that funds with a bad performance are less likely to survive (Ter Horst, Nijman and Verbeek, 2001), analysing
the effect of an income policy experiment may suffer from biases if people that beneﬁt
less from the experiment are more likely to drop out of the panel (Hausman and Wise,
1979), or estimating the impact of the unemployment rate on individual wages may
be disturbed by the possibility that people with relatively high wages are more likely
to leave the labour market in case of increasing unemployment (Keane, Mofﬁtt and
Runkle, 1988).
If rit depends upon αi or εit , selection bias may arise in the standard estimators
(see Chapter 7). This means that the distribution of y given x and conditional upon
selection (into the sample) is different from the distribution of y given x (which is
what we are interested in). For consistency of the ﬁxed effects estimator it is now
required that
E{(xit − x̄i )εit |ri1 , . . . , riT } = 0.
(10.94)
This means that the ﬁxed effects estimator is inconsistent if the fact whether an individual is in the sample or not tells us something about the expected value of the error
term that is related with xit . Clearly, if (10.11) holds and rit is independent of αi and
all εis (for given xis ), the above condition is satisﬁed. Note that sample selection may
depend upon αi without affecting consistency of the ﬁxed effects estimator for β. In
fact, εit may even depend upon rit as long as their relationship is time-invariant (see
Verbeek and Nijman, 1992, 1996 for additional details).
In addition to (10.94), the conditions for consistency of the random effects estimator
are now given by E{x̄i εit |ri1 , . . . , riT } = 0 and
E{x̄i αi |ri1 , . . . , riT } = 0.

(10.95)

This does not allow the expected value of either error component to depend on the
selection indicators. If individuals with certain values for their unobserved heterogeneity αi are less likely to be observed in some wave of the panel, this will typically
bias the random effects estimator. Similarly, if individuals with certain shocks εit are
more likely to drop out, the random effects estimator is typically inconsistent. Note
that because the ﬁxed effects estimator allows selection to depend upon αi and upon
εit in a time-invariant way, it is more robust against selection bias than the random
effects estimator. Another important observation made by Verbeek and Nijman (1992)
is that estimators from the unbalanced panel do not necessarily suffer less from selection bias than those from the balanced sub-panel. In general, the selection biases in the
estimators from the unbalanced and balanced samples need not be the same, and their
relative magnitude is not known a priori.

384

MODELS BASED ON PANEL DATA

Verbeek and Nijman (1992) suggest a number of simple tests for selection bias based
upon the above observations. First, as the conditions for consistency state that the error
terms should – in one sense or another – not depend upon the selection indicators,
one can test this by simply including some function of ri1 , . . . , riT in the model and
checking its signiﬁcance. Clearly, the null hypothesis says that whether an individual
was observed in any of the periods 1 to T should not give us any information about
his unobservables in the model. Obviously, adding rit to the model in (10.88) leads to
multicollinearity as rit = 1 for all observations in the sample.
 Instead, one could add
functions of ri1 , . . . , riT , like ri,t−1 , ci = Tt=1 rit or Ti = Tt=1 rit , indicating whether
unit i was observed in the previous period, whether he was observed over all periods,
and the total number of periods i is observed, respectively. Note that in the balanced
sub-panel all variables are identical for all individuals and thus incorporated in the
intercept term. Verbeek and Nijman (1992) suggest that the inclusion of ci and Ti may
provide a reasonable procedure to check for the presence of selection bias. Note that
this requires that the model is estimated under the random effects assumption, as the
within transformation would wipe out both ci and Ti . Of course, if the tests do not
reject, there is no reason to accept the null hypothesis of no selection bias, because the
power of the tests may be low.
Another group of tests is based upon the idea that the four different estimators,
random effects and ﬁxed effects, using either the balanced sub-panel or unbalanced
panel, usually all suffer differently from selection bias. A comparison of these estimators may therefore give an indication for the likelihood of selection bias. Although
any pair of estimators can be compared (see Verbeek and Nijman, 1992, or Baltagi,
2001, Section 11.5), it is known that ﬁxed effects and random effects estimators may
be different for other reasons than selection bias (see Subsection 10.2.3). Therefore, it
is most natural to compare either the ﬁxed effects or the random effects estimator using
the balanced sub-panel, with its counterpart using the unbalanced panel. If different
samples, selected on the basis of ri1 , . . . , riT , lead to signiﬁcantly different estimators,
it must be the case that the selection process tells us something about the unobservables
in the model. That is, it indicates the presence of selection bias. As the estimators using
the unbalanced panel are efﬁcient within a particular class of estimators, we can use
the result of Hausman again and derive a test statistic based upon the random effects
estimator as (compare (10.27))
B
U 
B
U
B
U
ξRE = (β̂RE
− β̂RE
) [V̂ {β̂RE
} − V̂ {β̂RE
}]−1 (β̂RE
− β̂RE
),

(10.96)

where the V̂ denote estimates of the covariance matrices and the superscripts B and U
refer to the balanced and unbalanced sample, respectively. Similarly, a test based on the
two ﬁxed effects estimators can be derived. Under the null hypothesis, the test statistic
follows a Chi-squared distribution with K degrees of freedom. Note that the implicit
U
B
null hypothesis for the test is that plim(β̂RE
− β̂RE
) = 0. If this is approximately true
and the two estimators suffer similarly from selection bias, the test has no power.23
Again, it is possible to test for a subset of the elements in β.
23

The test suggested here is not a real Hausman test because none of the estimators is consistent under
the alternative hypothesis. This does not invalidate the test as such but may result in limited power in
certain directions.

EXERCISES

385

10.8.3 Estimation with Nonrandomly Missing Data
As in the cross-sectional case (see Section 7.6), selection bias introduces an identiﬁcation problem. As a result, it is not possible to consistently estimate the model
parameters in the presence of selection bias, unless additional assumptions are imposed.
As an illustration, let us assume that the selection indicator rit can be explained by a
random effects probit model, that is

rit∗ = zit γ + ξi + ηit ,

(10.97)

where rit = 1 if rit∗ > 0 and 0 otherwise, and zit is a (well-motivated) vector of exogenous variables that includes xit . The model of interest is given by
yit = xit β + αi + εit .

(10.98)

Let us assume that the error components in the two equations have a joint normal
distribution. This is a generalization of the cross-sectional sample selection model
considered in Subsection 7.5.1. The effect of sample selection in (10.98) is reﬂected
in the expected values of the unobservables, conditional upon the exogenous variables
and the selection indicators, that is

and

E{αi |zi1 , . . . , ziT , ri1 , . . . , riT }

(10.99)

E{εit |zi1 , . . . , ziT , ri1 , . . . , riT }.

(10.100)

It can be shown (Verbeek and Nijman, 1992) that (10.100) is time-invariant if
cov{εit , ηit } = 0 or if zit γ is time-invariant. This is required for consistency of the ﬁxed
effects estimator. Further, (10.99) is zero if cov{αi , ξi } = 0, while (10.100) is zero if
cov{εit , ηit } = 0 so that the random effect estimator is consistent if the unobservables
in the primary equation and the selection equation are uncorrelated.
Estimation in the more general case is relatively complicated. Hausman and Wise
(1979) consider a case where the panel has two periods and attrition only takes place
in the second period. In the more general case, using maximum likelihood to estimate
the two equations simultaneously requires numerical integration over two dimensions
(to integrate out the two individual effects). Nijman and Verbeek (1992) and Vella
and Verbeek (1999) present alternative estimators based upon the two-step estimation
method for the cross-sectional sample selection model. Essentially, the idea is that
the terms in (10.99) and (10.100), apart from a constant, can be determined from the
probit model in (10.97), so that estimates of these terms can be included in the primary
equation. Wooldridge (1995) presents some alternative estimators based on somewhat
different assumptions.

Exercises
Exercise 10.1 (Linear Model)
Consider the following simple panel data model

yit = xit β + αi∗ + εit ,

i = 1, . . . , N,

t = 1, . . . , T ,

(10.101)

MODELS BASED ON PANEL DATA

386

where β is one-dimensional, and where it is assumed that
αi∗ = x̄i λ + αi ,

with αi ∼ NID(0, σα2 ),

εit ∼ NID(0, σε2 ),


mutually independent and independent of all xit s, where x̄i = (1/T ) Tt=1 xit .
The parameter β in (10.101) can be estimated by the ﬁxed effects (or within) estimator given by
N T
t=1 (xit − x̄i )(yit − ȳi )
β̂FE = i=1
.
N T
2
i=1
t=1 (xit − x̄i )
As an alternative, the correlation between the error term αi∗ + εit and xit can be handled
by an instrumental variables approach.
a.

Give an expression for the IV estimator β̂IV for β in (10.101) using xit − x̄i as an
instrument for xit . Show that β̂IV and β̂FE are identical.

Another way to eliminate the individual effects αi∗ from the model is obtained by
taking ﬁrst differences. This results in
yit − yi,t−1 = (xit − xi,t−1 )β + (εit − εi,t−1 ),

t = 2, . . . , T .
(10.102)
b. Denote the OLS estimator based on (10.102) by β̂FD . Show that β̂FD is identical
to β̂IV and β̂FE if T = 2. This identity does no longer hold for T > 2. Which of
the two estimators would you prefer in that case? Explain. (Note: for additional
discussion see Verbeek, 1995.)
c. Consider the between estimator β̂B for β in (10.101). Give an expression for β̂B
and show that it is unbiased for β + λ.
d. Finally, suppose we substitute the expression for αi∗ into (10.101), giving
yit = xit β + x̄i λ + αi + εit ,

i = 1, . . . , N,

i = 1, . . . , N,

t = 1, . . . , T .

(10.103)

The vector (β, λ) can be estimated by GLS (random effects) based on (10.103).
It can be shown that the implied estimator for β is identical to β̂FE . Does this
imply that there is no real distinction between the ﬁxed effects and random effects
approaches? (Note: for additional discussion see Hsiao, 2003, Section 3.4.2a.)
Exercise 10.2 (Hausman–Taylor Model)

Consider the following linear panel data model




yit = x1,it
β1 + x2,it
β2 + w1,i
γ1 + w2,i
γ2 + αi + εit ,

(10.104)

where wk,i are time-invariant and xk,it are time-varying explanatory variables. The variables with index 1 (x1,it and w1,i ) are strictly exogenous in the sense that E{x1,it αi } =
0, E{x1,is εit } = 0 for all s, t, E{w1,i αi } = 0 and E{w1,i εit } = 0. It is also assumed that

EXERCISES

387

E{w2,i εit } = 0 and that the usual regularity conditions (for consistency and asymptotic
normality) are met.
a. Under which additional assumptions would OLS applied to (10.104) provide a
consistent estimator for β = (β1 , β2 ) and γ = (γ1 , γ2 ) ?
b. Consider the ﬁxed effects (within) estimator. Under which additional assumption(s)
would it provide a consistent estimator for β?
c. Consider the OLS estimator for β based upon a regression in ﬁrst differences.
Under which additional assumption(s) will this provide a consistent estimator
for β?
d. Discuss one or more alternative consistent estimators for β and γ if it can be
assumed that E{x2,is εit } = 0 (for all s, t), and E{w2,i εit } = 0. What are the restrictions, in this case, on the number of variables in each of the categories?
e. Discuss estimation of β if x2,it equals yi,t−1 .
f. Discuss estimation of β if x2,it includes yi,t−1 .
g. Would it be possible to estimate both β and γ consistently if x2,it includes yi,t−1 ?
If so, how? If not, why not? (Make additional assumptions, if necessary.)
Exercise 10.3 (Dynamic and Binary Choice Models)

Consider the following dynamic wage equation
wit = xit β + γ wi,t−1 + αi + εit ,

(10.105)

where wit denotes an individual’s log hourly wage rate and xit is a vector of personal
and job characteristics (age, schooling, gender, industry, etc.).
a. Explain in words why OLS applied to (10.105) is inconsistent.
b. Also explain why the ﬁxed effects estimator applied to (10.105) is inconsistent for
N → ∞ and ﬁxed T , but consistent for N → ∞ and T → ∞. (Assume that εit
is i.i.d.)
c. Explain why the results from a and b also imply that the random effects (GLS)
estimator in (10.105) is inconsistent for ﬁxed T .
d. Describe a simple consistent (for N → ∞) estimator for β, γ , assuming that αi
and εit are i.i.d. and independent of all xit s.
e. Describe a more efﬁcient estimator for β, γ under the same assumptions.
In addition to the wage equation, assume there is a binary choice model explaining
whether an individual is working or not. Let rit = 1 if individual i was working in
period t and zero otherwise. Then the model can be described as
rit∗ = zit δ + ξi + ηit
rit = 1 if rit∗ > 0
= 0 otherwise,

(10.106)

388

MODELS BASED ON PANEL DATA

where zit is a vector of personal characteristics. Assume that ξi ∼ NID(0, σξ2 ) and
ηit ∼ NID(0, 1 − σξ2 ), mutually independent and independent of all zit s. The model in
(10.106) can be estimated by maximum likelihood.
f.
g.
h.

Give an expression for the probability that rit = 1 given zit and ξi .
Use the expression from f to obtain a computationally tractable expression for the
likelihood contribution of individual i.
Explain why it is not possible to treat the ξi s as ﬁxed unknown parameters and
estimate δ consistently (for ﬁxed T ) from this ﬁxed effects probit.

From now on, assume that the appropriate wage equation is static and given by (10.105)
with γ = 0.
What are the consequences for the random effects estimator in (10.105) if ηit and
εit are correlated? Why?
j. What are the consequences for the ﬁxed effects estimator in (10.105) if ξi and αi
are correlated (while ηit and εit are not)? Why?
i.

A

Vectors and Matrices

At occasional places in this text use is made of results from linear algebra. This
Appendix is meant to review the concepts that are used. More details can be found
in textbooks on linear algebra or, for example, in Davidson and MacKinnon (1993,
Appendix A), Davidson (2000, Appendix A), or Greene (2003, Appendix A). Some
of the more complex topics are used in a limited number of places in the text. For
example, eigenvalues and the rank of a matrix only play a role in Chapter 9, while the
rules of differentiation are only needed in Chapters 2 and 5.

A.1 Terminology
In this book a vector is always a column of numbers, denoted



a1
 a2 
 
a =  ..  .
 . 
an
The transpose of a vector, denoted a  = (a1 , a2 , . . . , an ) is a row of numbers, sometimes called a row vector. A matrix is a rectangular array of numbers. Of dimension
n × k, it can be written as


a11

a
 21
A=


an1

a12
a22
an2

···
..

a1k

.

···

ank




.



VECTORS AND MATRICES

390

The ﬁrst index of the element aij refers to i-th row, the second index to the j -th
column. Denoting the vector in the j -th column of this matrix by aj , it is seen that A
consists of k vectors a1 to ak , which we can denote as

A = a1

a2


. . . ak .

The symbol  denotes the transpose of a matrix or vector, obtained as


a11

a
 12
A =




a21
a22

· · · an1
..

.



an2 

.. 
.
. 

· · · ank

a1k

The columns of A are the rows of A and vice versa. A matrix is square if n = k. A
square matrix A is symmetric if A = A . A square matrix A is called a diagonal matrix
if aij = 0 for all i = j . Note that a diagonal matrix is symmetric by construction. The
identity matrix I is a diagonal matrix with all diagonal elements equal to one.

A.2

Matrix Manipulations

If two matrices or vectors have the same dimensions they can be added or subtracted. Let A and B be two matrices of dimension n × k with typical elements aij
and bij , respectively. Then A + B has typical element aij + bij , while A − B has typical
element aij − bij . It easily follows that A + B = B + A and (A + B) = A + B  .
A matrix A of dimension n × k and a matrix B of dimension k × m can be multiplied to produce a matrix of dimension n × m. Let us consider the special case of
k = 1 ﬁrst. Then A = a  is a row vector and B = b is a column vector. Then we deﬁne
 
b1
b 
 2

AB = a  b = (a1 , a2 , . . . , an ) 
 ..  = a1 b1 + a2 b2 + · · · + an bn .
 . 
bn
We call a  b the inner product of the vectors a and b. Note that a  b = b a. Two vectors
are called orthogonal if a  b = 0. For any vector a, except the null vector, we have
that a  a > 0. The outer product of a vector a is aa  , which is of dimension n × n.
Another special case arises for m = 1, in which case A is an n × k matrix and B = b
is a vector of dimension k. Then c = Ab is also a vector, but of dimension n. It has
typical elements
ci = ai1 b1 + ai2 b2 + · · · + aik bk ,
which is the inner product between the vector obtained from the i-th row of A and the
vector b.

PROPERTIES OF MATRICES AND VECTORS

391

When m > 1, B is a matrix and C = AB is a matrix of dimension n × m with
typical elements
cij = ai1 b1j + ai2 b2j + · · · + aik bkj ,
being the inner products between the vectors obtained from the i-th row of A and the
j -th column of B. Note that this can only make sense if the number of columns in A
equals the number of rows in B.
As an example, consider

A=

1 2 3
4 5 0

,

1

2

0

5


B = 3




4

and
AB =

7 25
19 28

.

It is important to note that AB = BA. Even if AB exists, BA may not be deﬁned
because the dimensions of B and A do not match. If A is of dimension n × k and B
is of dimension k × n, then AB exists and has dimension n × n, while BA exists with
dimension k × k. In the above example, we have


9 12 3





BA =  19 26 9  .
20 25 0
For the transpose of a product of two matrices, it holds that
(AB) = B  A .
From this (and (A ) = A) it follows that both A A and AA exist and are symmetric.
Finally, multiplying a scalar and a matrix is the same as multiplying each element in
the matrix by this scalar. That is, for a scalar c we have that cA has typical element
caij .

A.3 Properties of Matrices and Vectors
If we consider a number of vectors a1 to ak we can take a linear combination of
these vectors. With scalar weights c1 , . . . , ck this produces the vector c1 a1 + c2 a2 +
· · · + ck ak , which we can shortly write as Ac, where, as before, A = [a1 · · · ak ] and
c = (c1 , . . . , ck ) .
A set of vectors is linearly dependent if any of the vectors can be written as a
linear combination of the others. That is, if there exist values for c1 , . . . , ck , not all

VECTORS AND MATRICES

392

zero, such that c1 a1 + c2 a2 + · · · + ck ak = 0 (the null vector). Equivalently, a set of
vectors is linearly independent if the only solution to
c1 a1 + c2 a2 + · · · + ck ak = 0
is
c1 = c2 = · · · = ck = 0.
That is, if the only solution to Ac = 0 is c = 0.
If we consider all possible vectors that can be obtained as linear combinations of
the vectors a1 , . . . , ak , these vectors form a vector space. If the vectors a1 , . . . , ak are
linearly dependent, we can reduce the number of vectors without changing this vector
space. The minimal number of vectors needed to span a vector space is called the
dimension of that space. This way, we can deﬁne the column space of a matrix as the
space spanned by its columns, and the column rank of a matrix as the dimension of
its column space. Clearly, the column rank can never exceed the number of columns.
A matrix is of full column rank if the column rank equals the number of columns.
The row rank of a matrix is the dimension of the space spanned by the rows of the
matrix. In general, it holds that the row rank and the column rank of a matrix are
equal, so we can unambiguously deﬁne the rank of a matrix. Note that this does not
imply that a matrix that is of full column rank is automatically of full row rank (this
only holds if the matrix is square).
A useful result in regression analysis is that for any A
rank (A) = rank (A A) = rank (AA ).

A.4

Inverse Matrices

A matrix B, if it exists, is the inverse of a matrix A if AB = I and BA = I . A necessary
requirement for this is that A is a square matrix and has full rank, in which case A is
also called invertible or nonsingular. In this case, we can deﬁne B = A−1 , and
AA−1 = I

and A−1 A = I.

Note that the deﬁnition implies that A = B −1 . Thus we have that (A−1 )−1 = A. If A−1
does not exist, we say that A is singular. Analytically, the inverse of a diagonal matrix
and a 2 × 2 matrix are easily obtained. For example,


−1  −1
a11
0
0
a11 0
0




−1
0 
 0 a22 0  = 
 0 a22

−1
0
0 a33
0
0 a33
and
a11

a12

a21

a22

−1

=

1
a11 a22 − a12 a21

a22

−a12

−a21

a11

.

IDEMPOTENT MATRICES

393

If a11 a22 − a12 a21 = 0 the 2 × 2 matrix A is singular: its columns are linearly dependent, and so are its rows. We call a11 a22 − a12 a21 the determinant of this 2 × 2 matrix
(see below).
Suppose we are asked to solve Ac = d for given A and d, where A is of dimension
n × n and both c and d are n-dimensional vectors. This is a system of n linear equations
with n unknowns. If A−1 exists, we can write
A−1 Ac = c = A−1 d
to obtain the solution. If A is not invertible, the system of linear equations has linear
dependencies. There are two possibilities. Either more than one vector c satisﬁes Ac =
d, so no unique solution exists; or the equations are inconsistent, so there is no solution
to the system. If d is the null vector, only the ﬁrst possibility remains.
It is straightforward to derive that
(A−1 ) = (A )−1
and
(AB )−1 = B −1 A−1
(assuming that both inverse matrices exist).

A.5 Idempotent Matrices
A special class of matrices is that of symmetric and idempotent matrices. A matrix P
is symmetric if P  = P and idempotent if PP = P . A symmetric idempotent matrix
P has the interpretation of a projection matrix. This means that the projection vector
Px is in the column space of P , while the residual vector x − P x is orthogonal to any
vector in the column space of P .
A projection matrix which projects upon the column space of a matrix A can
be constructed as P = A(A A)−1 A . Clearly, this matrix is symmetric and idempotent. Projecting twice upon the same space should leave the result unaffected so we
should have PPx = P x, which follows directly. The residual from the projection
is x − Px = (I − A(A A)−1 A )x, so that M = I − A(A A)−1 A is also a projection
matrix with MP = PM = 0 and MM = M = M  . Thus the vectors Mx and Px are
orthogonal.
An interesting projecting matrix (used in Chapter 10) is Q = I − (1/n)ιι , where
ι is an n-dimensional vector of ones (so that ιι is a matrix of ones). The diagonal
elements in this matrix are 1 − 1/n and all off-diagonal elements are −1/n. Now Qx
is a vector containing x in deviation from its mean. A vector of means is produced by
the transformation matrix P = (1/n)ιι . Note that PP = P and QP = 0.
The only nonsingular projection matrix is the identity matrix. All other projection
matrices are singular, each having rank equal to the dimension of the space upon which
they project.

VECTORS AND MATRICES

394

A.6

Eigenvalues and Eigenvectors

Let A be a symmetric n × n matrix. Consider the following problem of ﬁnding combinations of a vector c (other than the null vector) and a scalar λ that satisfy
Ac = λc.
In general, there are n solutions λ1 , . . . , λn , called the eigenvalues (characteristic roots)
of A, corresponding with n vectors c1 , . . . , cn , called the eigenvectors (characteristic
vectors). If c1 is a solution then so is kc1 for any constant k, so the eigenvectors are
deﬁned up to a constant. The eigenvectors of a symmetric matrix are orthogonal, that
is ci cj = 0 for all i = j .
If an eigenvalue is zero, the corresponding vector c satisﬁes Ac = 0, which implies
that A is not of full rank and thus singular. Thus a singular matrix has at least one
zero eigenvalue. In general, the rank of a symmetric matrix corresponds to the number
of nonzero eigenvalues.
A symmetric matrix is called positive deﬁnite if all its eigenvalues are positive. It is
called positive semi-deﬁnite if all its eigenvalues are non-negative. A positive deﬁnite
matrix is invertible. If A is positive deﬁnite, it holds for any vector x (not the null
vector) that
x  Ax > 0.
The reason is that any vector x can be written as a linear combination of the eigenvectors as x = d1 c1 + · · · + dn cn for scalars d1 , . . . , dn , and we can write
x  Ax = (d1 c1 + · · · + dn cn ) A(d1 c1 + · · · + dn cn )
= λ1 d12 c1 c1 + · · · + λn dn2 cn cn > 0.
Similarly, for a positive semi-deﬁnite matrix A we have that for any vector x
x  Ax ≥ 0.
The determinant of a symmetric matrix equals the product of its n eigenvalues. The
determinant of a positive deﬁnite matrix is positive. A symmetric matrix is singular if
the determinant is zero (i.e. if one of the eigenvalues is zero).

A.7

Differentiation

Let x be an n-dimensional column vector. If c is also an n-dimensional column vector,
c x is a scalar. Let us consider c x as a function of the vector x. Then, we can consider
the vector of derivatives of c x with respect to each of the elements in x, that is
∂c x
= c.
∂x

SOME LEAST SQUARES MANIPULATIONS

395

This is a column vector of n derivatives, the typical element being ci . More generally,
we have for a vectorial function Ax (where A is a matrix) that
∂Ax
= A .
∂x
The element in column i, row j of this matrix is the derivative of the j -th element in
the function Ax with respect to xi .
Further,
∂x  Ax
= 2Ax
∂x
for a symmetric matrix A. If A is not symmetric, we have
∂x  Ax
= (A + A )x.
∂x
All these results follow from collecting the results from an element-by-element differentiation.

A.8 Some Least Squares Manipulations
Let xi = (xi1 , xi2 , . . . , xiK ) with xi1 ≡ 1 and β = (β1 , β2 , . . . , βK ) . Then
xi β = β1 + β2 xi2 + · · · + βK xiK .
The matrix

N
i=1

xi xi

xi1



x 
 i2 
 .  (x , x , . . . , x )
=
iK
 .  i1 i2
i=1  . 
N





=




xiK
N
i=1

2
xi1

..
.
..
.
N
i=1

xi1 xiK

N
i=1

xi2 xi1

N
i=1

2
xi2

···

N
i=1

xiK xi1

.

..
.

···

N
i=1

..










2
xiK

is a K × K symmetric matrix containing sums of squares and cross-products. The
vector

 N
i=1 xi1 yi

 N
N


 i=1 xi2 yi 
xi yi = 

.
..


i=1


N
i=1 xiK yi

VECTORS AND MATRICES

396

has length K, so that the system
N

xi xi b =

i=1

N

xi yi
i=1


is a system of K equations with K unknowns (in b). If N
i=1 xi xi is invertible a unique
N

solution exists. Invertibility requires that i=1 xi xi is of full rank. If it is not full rank,
a nonzero K-dimensional vector c exists such that xi c = 0 for each i and a linear

dependence exists between the columns/rows of the matrix N
i=1 xi xi .
With matrix notation, the N × K matrix X is deﬁned as

x11
 ..
X= .

x12
..
.

···
..
.


x1K
.. 
. 

xN1

xN2

···

xNK



and y = (y1 , y2 , . . . , yN ) . From this it is easily veriﬁed that
X X =

N

xi xi

i=1

and
X y =

N

xi yi .
i=1

The matrix X X is not invertible if the matrix X is not of full rank. That is, if a linear
dependence exists between the columns of X (‘regressors’).

B

Statistical and
Distribution Theory

This Appendix brieﬂy reviews some statistical and distribution theory that is used in
this text. More details can be found in, for example, Davidson and MacKinnon (1993,
Appendix B) or Greene (2003, Appendix B).

B.1 Discrete Random Variables
A random variable is a variable that can take different outcomes depending upon ‘the
state of nature’. For example, the outcome of throwing once with a dice is random,
with possible outcomes 1, 2, 3, 4, 5, and 6. Let us denote an arbitrary random variable
by Y . If Y denotes the outcome of the dice experiment (and the dice is fair and thrown
randomly), the probability of each outcome is 1/6. We can denote this as
P {Y = y} = 1/6

for y = 1, 2, . . . , 6.

The function which links possible outcomes (in this case y = 1, 2, . . . , 6) to the corresponding probabilities is the probability mass function or, more generally, the
probability distribution function. We can denote it by
f (y) = P {Y = y}.
Note that f (y) is not a function of the random variable Y , but of all its possible outcomes.
The function f (y) has the property that if we sum it over all possible outcomes the
result is one. That is

f (yj ) = 1.
j

STATISTICAL AND DISTRIBUTION THEORY

398

The expected value of a discrete random variable is a weighted average of all possible
outcomes, where the weights correspond to the probability of that particular outcome.
We denote

E{Y } =
yj f (yj ).
j

Note that E{Y } does not necessarily correspond to one of the possible outcomes. In
the dice experiment, for example, the expected value is 3.5.
A distribution is degenerate if it is concentrated in one point only, that is if
P {Y = y} = 1 for one particular value of y and zero for all other values.

B.2

Continuous Random Variables

A continuous random variable can take an inﬁnite number of different outcomes, for
example, any value in the interval [0, 1]. In this case each individual outcome has a
probability of zero. Instead of a probability mass function, we deﬁne the probability
density function f (y) ≥ 0 as

P {a ≤ Y ≤ b} =

b
a

f (y)dy.

In a graph, P {a ≤ Y ≤ b} is the area under the function f (y) between the points a
and b. Taking the integral of f (y) over all possible outcomes gives
 ∞
f (y)dy = 1.
−∞

If Y takes values within a certain range only, it is implicitly assumed that f (y) = 0
anywhere outside this range.
We can also deﬁne the cumulative density function (cdf) as
 y
F (y) = P {Y ≤ y} =
f (t)dt,
−∞

such that f (y) = F  (y) (the derivative). The cumulative density function has the property that 0 ≤ F (y) ≤ 1, and is monotonically increasing, i.e.
F (y) ≥ F (x)

if y > x.

It easily follows that P {a ≤ Y ≤ b} = F (b) − F (a).
The expected value or mean of a continuous random variable, often denoted as µ,
is deﬁned as
 ∞
µ = E{Y } =
yf (y)dy.
−∞

Other measures of location are the median, which is the value m for which we have
P {Y ≤ m} ≥ 1/2 and P {Y ≥ m} ≤ 1/2.

EXPECTATIONS AND MOMENTS

399

So 50% of the observations is below the median and 50% above. The mode is simply
the value for which f (y) takes its maximum. It is not often used in econometric
applications.
A distribution is symmetric around its mean if f (µ − y) = f (µ + y). In this case
the mean and the median of the distribution are identical.

B.3 Expectations and Moments
If Y and X are random variables and a and b are constants, then it holds that
E{aY + bX} = aE{Y } + bE{X},
showing that the expectation is a linear operator. Similar results do not necessarily hold
if we consider a nonlinear transformation of a random variable. For a nonlinear function
g, it does not hold in general that E{g(Y )} = g(E{Y }). If g is concave (g  (Y ) < 0),
Jensen’s inequality says that
E{g(Y )} ≤ g(E{Y }).
For example, E{log Y } ≤ log E{Y }. The implication of this is that we cannot determine
the expected value of a function of Y from the expected value of Y only. Of course,
it holds by deﬁnition that

E{g(Y )} =

∞
−∞

g(y)f (y)dy.

The variance of a random variable, often denoted by σ 2 , is a measure of the dispersion of the distribution. It is deﬁned as
σ 2 = V {Y } = E{(Y − µ)2 }
and equals the expected quadratic deviation from the mean. It is sometimes called the
second central moment. A useful result is that
E{(Y − µ)2 } = E{Y 2 } − 2E{Y }µ + µ2 = E{Y 2 } − µ2 ,
where E{Y 2 } is the second moment. If Y has a discrete distribution, its variance is
determined as

V {Y } =
(yj − µ)2 f (yj ),
j

where j indexes the different outcomes. For a continuous distribution we have

V {Y } =

∞
−∞

(y − µ)2 f (y)dy.

STATISTICAL AND DISTRIBUTION THEORY

400

Using these deﬁnitions it is easily veriﬁed that
V {aY + b} = a 2 V {Y },
where a and b are arbitrary constants. Often we will also use the standard deviation of
a random variable, denoted σ , deﬁned as the square root of the variance. The standard
deviation is expressed in the same units as Y .
In most cases the distribution of a random variable is not completely described by
its mean and variance, and we can deﬁne the k-th central moment as
E{(Y − µ)k },

k = 1, 2, 3, . . .

In particular, the third central moment is a measure of the asymmetry of the distribution
around its mean, while the fourth central moment measures the peakedness of the
distribution. Typically, skewness is deﬁned as S ≡ E{(Y − µ)3 }/σ 3 , while kurtosis is
deﬁned as K ≡ E{(Y − µ)4 }/σ 4 . Kurtosis of a normal distribution is 3, so that K − 3
is referred to as excess kurtosis. A distribution with positive excess kurtosis is called
leptokurtic.

B.4

Multivariate Distributions

The joint density function of two random variables Y and X, denoted f (y, x), is
deﬁned by
 b1  b2
P {a1 < Y < b1 , a2 < X < b2 } =
f (y, x)dydx.
a1

a2

If Y and X are independent, it holds that f (y, x) = f (y)f (x), such that
P {a1 < Y < b1 , a2 < X < b2 } = P {a1 < Y < b1 }P {a2 < X < b2 }.
In general, the marginal distribution of Y is characterized by the density function

f (y) =

∞
−∞

f (y, x)dx.

This implies that the expected value of Y is given by

E{Y } =

∞
−∞


yf (y)dy =

∞
−∞



∞
−∞

yf (y, x)dxdy.

The covariance between Y and X is a measure of linear dependence between the
two variables. It is deﬁned as
σxy = cov{Y, X} = E{(Y − µy )(X − µx )},

CONDITIONAL DISTRIBUTIONS

401

where µy = E{Y } and µx = E{X}. The correlation coefﬁcient is given by the covariance standardized by the two standard deviations, that is
σxy
cov{Y, X}
ρyx = √
=
.
σx σy
V {Y }V {X}
The correlation coefﬁcient is always between −1 and 1 and is not affected by the
scaling of the variables. If cov{Y, X} = 0, Y and X are said to be uncorrelated. When
a, b, c, d are constants, it holds that
cov{aY + b, cX + d} = ac cov{Y, X}.
Further,
cov{aY + bX, X} = a cov{Y, X} + b cov{X, X} = a cov{Y, X} + bV {X}.
It also follows that two variables Y and X are perfectly correlated (ρyx = 1) if Y = aX
for some nonzero value of a. If Y and X are correlated, the variance of a linear function
of Y and X depends upon their covariance. In particular,
V {aY + bX} = a 2 V {Y } + b2 V {X} + 2ab cov{Y, X}.
If we consider a K-dimensional vector of random variables, Y = (Y1 , . . . , YK ) , we
can deﬁne its expectation vector as



E{Y1 }
 . 
E{Y } =  .. 
E{YK }
and its variance-covariance matrix (or simply covariance matrix) as


V {Y } = 

V {Y1 }
..
.

···
..
.

cov{YK , Y1 } · · ·


cov{Y1 , YK }
..

.
.
V {YK }

Note that this matrix is symmetric. If we consider one or more linear combinations of
the elements in Y , say R Y , where R is of dimension J × K, it holds that
V {R Y } = RV {Y }R  .

B.5 Conditional Distributions
A conditional distribution describes the distribution of a variable, say Y , given the
outcome of another variable X. For example, if we throw with two dice, X could
denote the outcome of the ﬁrst dice and Y could denote the total of the two dice. Then

402

STATISTICAL AND DISTRIBUTION THEORY

we could be interested in the distribution of Y conditional upon the outcome of the ﬁrst
dice. For example, what is the probability of throwing 7 in total if the ﬁrst dice had
an outcome of 3? Or an outcome of 3 or less? The conditional distribution is implied
by the joint distribution of the two variables. We deﬁne
f (y|X = x) = f (y|x) =

f (y, x)
.
f (x)

If Y and X are independent, it immediately follows that f (y|x) = f (y). From the
above deﬁnition it follows that
f (y, x) = f (y|x)f (x),
which says that the joint distribution of two variables can be decomposed in the product
of a conditional distribution and a marginal distribution. Similarly, we can write
f (y, x) = f (x|y)f (y).
The conditional expectation of Y given X = x is the expected value of Y from the
conditional distribution. That is,

E{Y |X = x} = E{Y |x} = yf (y|x)dy.
The conditional expectation is a function of x, unless Y and X are independent.
Similarly, we can deﬁne the conditional variance as

V {Y |x} = (y − E{Y |x})2 f (y|x)dy,
which can be written as
V {Y |x} = E{Y 2 |x} − (E{Y |x})2 .
It holds that
V {Y } = Ex {V {Y |X}} + Vx {E{Y |X}},
where Ex and Vx denote the expected value and variance, respectively, based upon
the marginal distribution of X. The terms V {Y |X} and E{Y |X} are functions of the
random variable X and therefore random variables themselves.
Let us consider the relationship between two random variables Y and X, where
E{Y } = 0. Then it follows that Y and X are uncorrelated if
E{Y X} = cov{Y, X} = 0.
If Y is conditional mean independent of X it means that
E{Y |X} = E{Y } = 0.

THE NORMAL DISTRIBUTION

403

This is stronger than zero correlation because E{Y |X} = 0 implies that E{Y g(X)} = 0
for any function g. If Y and X are independent this is again stronger and it implies that
E{g1 (Y )g2 (X)} = E{g1 (Y )}E{g2 (X)}
for arbitrary functions g1 and g2 . It is easily veriﬁed that this implies conditional mean
independence and zero correlation. Note that E{Y |X} = 0 does not necessarily imply
that E{X|Y } = 0.

B.6 The Normal Distribution
In econometrics the normal distribution plays a central role. The density function for
a normal distribution with mean µ and variance σ 2 is given by
f (y) = √

1
2πσ 2

exp −

1 (y − µ)2
,
2
σ2

which we write as Y ∼ N(µ, σ 2 ). It is easily veriﬁed that the normal distribution is
symmetric. A standard normal distribution is obtained for µ = 0 and σ = 1. Note that
the standardized variable (Y − µ)/σ is N(0, 1) if Y ∼ N(µ, σ 2 ). The density of a
standard normal distribution, typically denoted by φ, is given by
1
1
φ(y) = √ exp − y 2 .
2
2π
A useful property of a normal distribution is that a linear function of a normal variable
is also normal. That is, if Y ∼ N(µ, σ 2 ) then
aY + b ∼ N(aµ + b, a 2 σ 2 ).
The cumulative density function of the normal distribution does not have a closed form
expression. We have
P {Y ≤ y} = P

Y −µ
y −µ
≤
σ
σ

y−µ
=
σ


=

(y−µ)/σ
−∞

φ(t)dt,

where  denotes the cdf of the standard normal distribution. Note that (y) = 1 −
(−y) due to the symmetry.
The symmetry also implies that the third central moment of a normal distribution
is zero. It can be shown that the fourth central moment of a normal distribution is
given by
E{(Y − µ)4 } = 3σ 4 .
Note that this implies that E{Y 4 } = 4σ 4 . Typically these properties of the third and
fourth central moments are exploited in tests against non-normality.

STATISTICAL AND DISTRIBUTION THEORY

404

If (Y, X) have a bivariate normal distribution with mean vector µ = (µy , µx )
and covariance matrix

σy2 σyx
,
=
σyx σx2
denoted (Y, X) ∼ N(µ, ), the joint density function is given by
f (y, x) = f (y|x)f (x),
where both the conditional density of Y given X and the marginal density of X are
normal. The conditional density function is given by


1 (y − µy|x )2
1
,
exp −
f (y|x) = 
2
2
σy|x
2πσ 2
y|x

where µy|x is the conditional expectation of Y given X, given by
µy|x = µy + (σyx /σx2 )(x − µx ),
2
and σy|x
is the conditional variance of Y given X,
2
2
= σy2 − σyx2 /σx2 = σy2 (1 − ρyx
),
σy|x

with ρyx denoting the correlation coefﬁcient between Y and X. These results have some
important implications. First, if two (or more) variables have a joint normal distribution,
all marginal distributions and conditional distributions are also normal. Second, the
conditional expectation of one variable given the other(s) is a linear function (with an
intercept term). Third, if ρyx = 0 it follows that f (y|x) = f (y) so that
f (y, x) = f (y)f (x),
and Y and X are independent. Thus, if Y and X have a joint normal distribution
with zero correlation then they are automatically independent. Recall that in general
independence is a stronger requirement than uncorrelatedness.
Another important result is that a linear function of normal variables is also normal,
that is, if (Y, X) ∼ N(µ, ) then
aY + bX ∼ N(aµy + bµx , a 2 σy2 + b2 σx2 + 2abσyx ).
These results can be generalized to a general K-variate normal distribution. If the
K-dimensional vector Y has a normal distribution with mean vector µ and covariance
matrix , that is
Y ∼ N(µ, ),
it holds that the distribution of R Y , where R is a J × K matrix, is a J -variate normal
distribution, given by
R Y ∼ N(Rµ, RR  ).

RELATED DISTRIBUTIONS

405

In models with limited dependent variables we often encounter forms of truncation.
If Y has density f (y), the distribution of Y truncated from below at a given point c
(Y ≥ c) is given by
f (y|Y ≥ c) =

f (y)
P {Y ≥ c}

if y ≥ c

and 0 otherwise.

If Y is a standard normal variable, the truncated distribution of Y ≥ c has mean
E{Y |Y ≥ c} = λ1 (c),
where
λ1 (c) =
and variance

φ(c)
,
1 − (c)

V {Y |Y ≥ c} = 1 − λ1 (c)[λ1 (c) − c].

If the distribution is truncated from above (Y ≤ c) it holds that
E{Y |Y ≤ c} = λ2 (c),
with
λ2 (c) =

−φ(c)
.
(c)

If Y has a normal density with mean µ and variance σ 2 the truncated distribution
Y ≥ c has mean
E{Y |Y ≥ c} = µ + σ λ1 (c∗ ) ≥ µ
where c∗ = (c − µ)/σ , and, similarly,
E{Y |Y ≤ c} = µ + σ λ2 (c∗ ) ≤ µ.
When (Y, X) have a bivariate normal distribution, as above, we obtain that
E{Y |X ≥ c} = µy + (σyx /σx2 )[E{X|X ≥ c} − µx ]
= µy + (σyx /σx )λ1 (c∗ ).
More details can be found in Maddala (1983, Appendix).

B.7 Related Distributions
Besides the normal distribution several other distributions are important. First, we
deﬁne the Chi-squared distribution as follows. If Y1 , . . . , YJ is a set of independent
standard normal variables, it holds that
ξ=

J

j =1

Yj2

STATISTICAL AND DISTRIBUTION THEORY

406

has a Chi-squared distribution with J degrees of freedom. We denote ξ ∼ χJ2 . More
generally, if Y1 , . . . , YJ is a set of independent normal variables with mean µ and
variance σ 2 , if follows that
J

(Yj − µ)2
ξ=
σ2
j =1
is Chi-squared with J degrees of freedom. Most generally, if Y = (Y1 , . . . , YJ ) is a
vector of random variables that has a joint normal distribution with mean vector µ and
(nonsingular) covariance matrix , it follows that
ξ = (Y − µ)  −1 (Y − µ) ∼ χJ2 .
If ξ has a Chi-squared distribution with J degrees of freedom it holds that E{ξ } = J
and V {ξ } = 2J .
Next, we consider the t distribution (or Student distribution). If X has a standard
normal distribution, X ∼ N(0, 1), and ξ ∼ χJ2 and if X and ξ are independent, the ratio
t=√

X
ξ /J

has a t distribution with J degrees of freedom. Like the standard normal distribution,
the t distribution is symmetric around zero, but it has fatter tails, particularly for small
J . If J approaches inﬁnity, the t distribution approaches the normal distribution.
If ξ1 ∼ χJ21 and ξ2 ∼ χJ22 and if ξ1 and ξ2 are independent, it follows that the ratio
f =

ξ1 /J1
ξ2 /J2

has an F distribution with J1 and J2 degrees of freedom in numerator and denominator, respectively. It easily follows that the inverse ratio
ξ2 /J2
ξ1 /J1
also has an F distribution, but with J2 and J1 degrees of freedom, respectively. The
F distribution is thus the distribution of the ratio of two independent Chi-squared
distributed variables, divided by their respective degrees of freedom. When J1 = 1, ξ1
is a squared normal variable, say ξ1 = X2 , and it follows that
t =
2

X

2


ξ2 /J2

=

ξ1
= f ∼ FJ12 .
ξ2 /J2

Thus, with one degree of freedom in the numerator, the F distribution is just the square
of a t distribution. If J2 is large, the distribution of
J1 f =

ξ1
ξ2 /J2

RELATED DISTRIBUTIONS

407

is well approximated by a Chi-squared distribution with J1 degrees of freedom. For
large J2 the denominator is thus negligible.
Finally, we consider the lognormal distribution. If log Y has a normal distribution
with mean µ and variance σ 2 then Y > 0 has a so-called lognormal distribution. The
lognormal density is often used to describe the population distribution of (labour)
income or the distribution of asset returns (see Campbell, Lo and MacKinlay, 1997).
While E{log Y } = µ, it holds that
1
E{Y } = exp µ + σ 2
2
(compare Jensen’s inequality above).

Bibliography
Ahn, S. K. and Reinsel, C. G. (1990), Estimation for Partially Nonstationary Multivariate
Autoregressive Models, Journal of the American Statistical Association, 85, 813–823.
Akaike, H. (1973), Information Theory and an Extension of the Maximum Likelihood Principle.
In: B. N. Petrov and F. Cszaki, eds., Second International Symposium on Information Theory,
Akademiai Kiado, Budapest, 267–281.
Alvarez, J. and Arellano, M. (2003), The Time Series and Cross-Section Asymptotics of
Dynamic Panel Data Estimators, Econometrica, 71, 1121–1159.
Amemiya, T. (1981), Qualitative Response Models: A Survey, Journal of Economic Literature,
19, 1483–1536.
Amemiya, T. (1984), Tobit Models: A Survey, Journal of Econometrics, 24, 3–61.
Amemiya, T. and MaCurdy, T. (1986), Instrumental-Variable Estimation of an ErrorComponents Model, Econometrica, 54, 869–881.
Andersen, E. B. (1970), Asymptotic Properties of Conditional Maximum Likelihood Estimation,
Journal of the Royal Statistical Society, Series B, 32, 283–301.
Anderson, T. W. and Hsiao, C. (1981), Estimation of Dynamic Models with Error Components,
Journal of the American Statistical Association, 76, 598–606.
Andrews, D. W. K. and Schafgans, M. A. (1998), Semiparametric Estimation of the Intercept
of a Sample Selection Model, Review of Economic Studies, 63, 497–517.
Anglin, P. M. and Gençay, R. (1996), Semiparametric Estimation of a Hedonic Price Function,
Journal of Applied Econometrics, 11, 633–648.
Angrist, J. D. and Imbens, G. W. (1995), Two-Stage Least Squares Estimation of Average
Causal Effects in Models with Variable Treatment Intensity, Journal of the American Statistical Association, 90, 431–442.
Angrist, J. D., Imbens, G. W. and Rubin, D. B. (1996), Identiﬁcation and Causal Effects Using
Instrumental Variables, Journal of the American Statistical Association, 91, 444–455.
Angrist, J. D. and Krueger, A. B. (1991), Does Compulsory School Attendance Affect Schooling
and Earnings?, Quarterly Journal of Economics, 106, 979–1014.
Arellano, M. (1987), Computing Robust Standard Errors for Within-Groups Estimators, Oxford
Bulletin of Economics and Statistics, 49, 431–434.
Arellano, M. (2003), Panel Data Econometrics, Oxford University Press, Oxford.
Arellano, M. and Bond, S. (1991), Some Tests of Speciﬁcation for Panel Data: Monte Carlo
Evidence and an Application to Employment Equations, Review of Economic Studies, 58,
277–294.

410

BIBLIOGRAPHY

Arellano, M. and Bover, O. (1995), Another Look at the Instrumental Variable Estimation of
Error-Components Models, Journal of Econometrics, 68, 29–51.
Atkinson, A. B., Gomulka, J. and Stern, N. H. (1990), Spending on Alcohol: Evidence from the
Family Expenditure Survey 1970–1983, Economic Journal, 100, 808–827.
Baltagi, B. H. (2001), Econometric Analysis of Panel Data, 2nd edition, John Wiley and Sons,
New York.
Banerjee, A. (1999), Panel Data Unit Roots and Cointegration: An Overview, Oxford Bulletin
of Economics and Statistics, 61, 607–629.
Banerjee, A., Dolado, J., Galbraith, J. W. and Hendry, D. F. (1993), Co-Integration, ErrorCorrection, and the Econometric Analysis of Non-Stationary Data, Oxford University Press.
Banerjee, A., Marcellino, M. and Osbar, C. (2001), Testing for PPP: Should We Use Panel
Methods? working paper, European University Institute, Firenze.
Banks, J., Blundell, R. and Lewbel, A. (1997), Quadratic Engel Curves and Consumer Demand,
Review of Economics and Statistics, 74, 527–539.
Banz, R. (1981), The Relation between Returns and Market Value of Common Stocks, Journal
of Financial Economics, 9, 3–18.
Bårdsen, G. (2001), Review of PCGets 1 for Windows, Econometrics Journal, 4, 311–318.
Bera, A. K. and Higgins, M. L. (1993), ARCH Models: Properties, Estimation and Testing,
Journal of Economic Surveys, 7, 305–366.
Bera, A. K., Jarque, C. M. and Lee, L. F. (1984), Testing the Normality Assumption in Limited
Dependent Variable Models, International Economic Review, 25, 563–578.
Berndt, E. R. (1991), The Practice of Econometrics, Classic and Contemporary, Addison-Wesley,
Reading.
Berndt, E. R., Hall, B. H., Hall, R. E. and Hausman, J. A. (1974), Estimation and Inference in
Nonlinear Structural Models, Annals of Economic and Social Measurement, 3, 653–665.
Bhargava, A., Franzini, L. and Narendranathan, W. (1983), Serial Correlation and the Fixed
Effects Model, Review of Economic Studies, 49, 533–549.
Blundell, R. and Bond, S. (1998), Initial Conditions and Moment Restrictions in Dynamic Panel
Data Models, Journal of Econometrics, 87, 115–143.
Bollerslev, T. (1986), Generalized Autoregressive Conditional Heteroskedasticity, Journal of
Econometrics, 31, 307–327.
Bollerslev, T., Chou, R. Y. and Kroner, K. F. (1992), ARCH Modeling in Finance. A Review
of the Theory and Empirical Evidence, Journal of Econometrics, 52, 5–59.
Bollerslev, T., Engle, R. F. and Nelson, D. B. (1994), ARCH Models. In: R. F. Engle and
D. L. McFadden, eds., Handbook of Econometrics, Volume IV, Elsevier Science, Amsterdam,
2961–3038.
Boswijk, H. P. (1999), Asymptotic Theory for Integrated Processes, Oxford University Press,
Oxford.
Bound, J., Jaeger, D. A. and Baker, R. M. (1995), Problems with Instrumental Variables Estimation when the Correlation between the Instrument and the Endogenous Variable is Weak,
Journal of the American Statistical Association, 90, 443–450.
Box, G. E. P. and Jenkins, G. M. (1976), Time Series Analysis: Forecasting and Control, revised
edition, Holden-Day.
Breusch, T. (1978), Testing for Autocorrelation in Dynamic Linear Models, Australian Economic
Papers, 17, 334–355.
Breusch, T. and Pagan, A. (1980), A Simple Test for Heteroskedasticity and Random Coefﬁcient
Variation, Econometrica, 47, 1287–1294.
Breusch, T., Mizon, G. and Schmidt, P. (1989), Efﬁcient Estimation Using Panel Data, Econometrica, 57, 695–700.

BIBLIOGRAPHY

411

Butler, J. S. and Mofﬁtt, R. (1982), A Computationally Efﬁcient Quadrature Procedure for the
One-Factor Multinomial Probit Model, Econometrica, 50, 761–764.
Cameron, A. C. and Trivedi, P. K. (1986), Econometric Models Based on Count Data: Comparisons and Applications of Some Estimators and Tests, Journal of Applied Econometrics,
1, 29–53.
Cameron, A. C. and Trivedi, P. K. (1998), Regression Analysis of Count Data, Cambridge University Press, Cambridge.
Cameron, A. C. and Windmeijer, F. A. G. (1996), r-squared Measures for Count Data Regression Models with Applications to Health Care Utilization, Journal of Business and Economic
Statistics, 14, 209–220.
Cameron, A. C. and Windmeijer, F. A. G. (1997), An R-squared Measure of Goodness of
Fit for Some Common Nonlinear Regression Models, Journal of Econometrics, 77,
329–342.
Campbell, J. Y. and Shiller, R. J. (1991), Yield Spreads and Interest Rate Movements: A Bird’s
Eye View, Review of Economic Studies, 58, 495–514.
Campbell, J. Y. and Shiller, R. J. (1998), Valuation Ratios and the Long-Run Stock Market
Outlook, Journal of Portfolio Management, 24, 11–26.
Campbell, J. Y., Lo, A. W. and MacKinlay, A. C. (1997), The Econometrics of Financial Markets, Princeton University Press, Princeton.
Canova, F. (1995), The Economics of VAR Models. In: K. D. Hoover, ed., Macroeconometrics:
Developments, Tensions and Prospects, Kluwer Academic Publishers, Boston, 57–97.
Card, D. (1995), Using Geographical Variation in College Proximity to Estimate the Return to
Schooling. In: L. N. Christoﬁdes, E. K. Grant and R. Swidinsky, eds., Aspects of Labour Market Behaviour: Essays in Honour of John Vanderkamp, University of Toronto Press, Toronto,
201–222.
Card, D. (1999), The Causal Effect of Education on Earnings. In: O. Ashenfelter and D. Card,
eds., Handbook of Labor Economics, Volume IIIA, Elsevier Science, Amsterdam.
Carroll, J. D. and Green, P. E. (1995), Psychometric Methods in Marketing Research: Part 1
Conjoint Analysis, Journal of Marketing Research, 32, 385–391.
Chamberlain, G. (1980), Analysis of Covariance with Qualitative Data, Review of Economic
Studies, 47, 225–238.
Charemza, W. W. and Deadman, D. F. (1999), New Directions in Econometric Practice. General to Speciﬁc Modelling, Cointegration and Vector Autoregression, 2nd edition, Edward
Elgar, Aldershot.
Cheung, Y.-W. and Lai, K. S. (1993), Finite Sample Sizes of Johansen’s Likelihood Ratio Tests
for Cointegration, Oxford Bulletin of Economics and Statistics, 55, 313–332.
Choi, I. (2001), Unit Root Tests for Panel Data, Journal of International Money and Finance,
20, 249–272.
Chow, G. (1960), Tests of Equality Between Sets of Coefﬁcients in Two Linear Regressions,
Econometrica, 28, 591–605.
Chowdhury, G. and Nickell, S. (1985), Hourly Earnings in the United States: Another Look at
Unionization, Schooling, Sickness and Unemployment using PSID Data, Journal of Labor
Economics, 3, 38–69.
Cincera, M. (1997), Patents, R&D, and Technological Spillovers at the Firm Level: Some Evidence from Econometric Count Models for Panel Data, Journal of Applied Econometrics, 12,
265–280.
Cochrane, J. H. (1996), A Cross-Sectional Test of an Investment-Based Asset Pricing Model,
Journal of Political Economy, 104, 572–621.
Cochrane, J. H. (2001), Asset Pricing, Princeton University Press, Princeton.

412

BIBLIOGRAPHY

Cochrane, D. and Orcutt, G. (1949), Application of Least Squares Regression to Relationships
Containing Autocorrelated Error Terms, Journal of the American Statistical Association, 44,
32–61.
Corbae D. and Ouliaris, S. (1988), Cointegration and Tests of Purchasing Power Parity, Review
of Economics and Statistics, 70, 508–511.
Cuthbertson K., Hall, S. G. and Taylor, M. P. (1992), Applied Econometric Techniques, Philip
Allan, Hemel Hempstead.
Davidson, J. (2000), Econometric Theory, Blackwell Publishers, Oxford.
Davidson, R. and MacKinnon, J. G. (1981), Several Tests for Model Speciﬁcation in the Presence of Alternative Hypotheses, Econometrica, 49, 781–793.
Davidson, R. and MacKinnon, J. G. (1993), Estimation and Inference in Econometrics, Oxford
University Press, Oxford.
Deaton, A. and Muellbauer, J. (1980), Economics and Consumer Behavior, Cambridge University Press, Cambridge.
Diamond, P. A. and Hausman, J. A. (1994), Contingent Valuation: Is Some Number Better than
No Number?, Journal of Economics Perspectives, 8, 45–64.
Dickey, D. A. and Fuller, W. A. (1979), Distribution of the Estimators for Autoregressive Time
Series with a Unit Root, Journal of the American Statistical Association, 74, 427–431.
Diebold, F. X. (1998), Forecasting, South-Western College Publishing, Cincinnati, OH.
Diebold, F. X. and Lopez, J. A. (1995), Modeling Volatility Dynamics. In: K. D. Hoover, ed.,
Macroeconometrics: Developments, Tensions and Prospects, Kluwer Academic Publishers,
Boston, 427–466.
Durbin, J. and Watson, G. (1950), Testing for Serial Correlation in Least Squares Regression – I,
Biometrika, 37, 409–428.
Eicker, F. (1967), Limit Theorems for Regressions with Unequal and Dependent Errors. In:
L. LeCam and J. Neyman, eds., Proceedings of the Fifth Berkeley Symposium on Mathematical
Statistics and Probability, University of California Press, Berkeley, 59–82.
Elton, E. J. and Gruber, M. J. (2003), Modern Portfolio Theory and Investment Analysis, 6th
edition, John Wiley and Sons, New York.
Enders, W. (2004), Applied Econometric Time Series, 2nd edition, John Wiley and Sons, New
York.
Engle, R. F. (1982), Autoregressive Conditional Heteroskedasticity with Estimates of the Variance of United Kingdom Inﬂation, Econometrica, 50, 987–1007.
Engle, R. F. (1984), Wald, Likelihood Ratio and Lagrange Multiplier Tests in Econometrics.
In: Z. Griliches and M. D. Intriligator, eds., Handbook of Econometrics, Volume II, Elsevier
Science, Amsterdam, 775–826.
Engle, R. F. and Bollerslev, T. (1986), Modelling the Persistence of Conditional Variances,
Econometric Reviews, 5, 1–50.
Engle, R. F. and Granger, C. W. J. (1987), Cointegration and Error Correction: Representation,
Estimation and Testing, Econometrica, 55, 251–276.
Engle, R. F. and Ng, V. K. (1993), Measuring and Testing the Impact of News on Volatility,
Journal of Finance, 48, 1749–1778.
Engle, R. F. and Yoo, B. S. (1987), Forecasting and Testing in Co-Integrated Systems, Journal
of Econometrics, 35, 143–159.
Engle, R. F., Hendry, D. F. and Richard, J.-F. (1983), Exogeneity, Econometrica, 51, 277–304.
Engle, R. F., Lilien, D. M. and Robins, R. P. (1987), Estimating Time Varying Risk Premia in
the Term Structure: The ARCH-M Model, Econometrica, 55, 591–407.
Fama, E. F. (1991), Efﬁcient Capital Markets II, Journal of Finance, 46, 1575–1617.
Fama, E. F. and French, K. R. (1988), Permanent and Temporary Components of Stock Prices,
Journal of Political Economy, 81, 246–273.
Franses, P. H. B. F. (1998), Time Series Models for Business and Economic Forecasting, Cambridge University Press, Cambridge.

BIBLIOGRAPHY

413

Franses, P. H. B. F. and Paap, R. (2001), Quantitative Models in Marketing Research, Cambridge University Press, Cambridge.
Froot, K. A. and Rogoff, K. (1996), Perspectives on PPP and Long-run Exchange Rates. In:
S. Grossman and K. Rogoff, eds., Handbook of International Economics, Volume III, Elsevier
Science, Amsterdam.
Fuller, W. A. (1976), Introduction to Statistical Time-Series, John Wiley and Sons, New York.
Glewwe, P. (1997), A Test of the Normality Assumption in the Ordered Probit Model, Econometric Reviews, 16, 1–19.
Godfrey, L. (1978), Testing against General Autoregressive and Moving Average Error Models
when the Regressors Include Lagged Dependent Variables, Econometrica, 46, 1293–1302.
Godfrey, L. (1988), Misspeciﬁcation Tests in Econometrics. The Lagrange Multiplier Principle
and Other Approaches, Cambridge University Press, Cambridge.
Goldfeld, S. and Quandt, R. (1965), Some Tests for Homoskedasticity, Journal of the American
Statistical Association, 60, 539–547.
Gourieroux, C. and Jasiak, J. (2001), Financial Econometrics, Princeton University Press,
Princeton.
Gouriéroux, C., Monfort, A. and Trognon, A. (1984), Pseudo-maximum Likelihood Methods:
Theory, Econometrica, 42, 681–700.
Gouriéroux, C., Monfort, A., Renault, E. and Trognon, A. (1987), Generalized Residuals, Journal of Econometrics, 34, 5–32.
Granger, C. W. J. (1983), Co-Integrated Variables and Error-Correcting Models, Unpublished
Discussion Paper 83-13, University of California, San Diego.
Granger, C. W. J. and Newbold, P. (1974), Spurious Regressions in Econometrics, Journal of
Econometrics, 35, 143–159.
Granger, C. W. J. and Swanson, N. R. (1997), An Introduction to Stochastic Unit-Root Processes, Journal of Econometrics, 80, 35–62.
Greene, W. H. (2003), Econometric Analysis, 5th edition, Prentice Hall.
Gregory, A. W. and Veall, M. R. (1985), On Formulating Wald Tests of Nonlinear Restrictions,
Econometrica, 53, 1465–1468.
Griliches, Z. (1977), Estimating the Returns to Schooling: Some Econometric Problems, Econometrica, 45, 1–22.
Gronau, R. (1974), Wage Comparisons: A Selectivity Bias, Journal of Political Economy, 82,
1119–1143.
Hajivassiliou, V. A. and McFadden, D. F. (1998), The Method of Simulated Scores for the
Estimation of LDV Models, Econometrica, 66, 863–896.
Hall, A. (1993), Some Aspects of Generalized Method of Moments Estimation. In:
G. S. Maddala, C. R. Rao and H. D. Vinod, eds., Handbook of Statistics, Volume XI, Elsevier
Science, Amsterdam, 393–417.
Hamilton, J. D. (1994), Time Series Analysis, Princeton University Press, Princeton.
Hanemann, W. M. (1994), Valuing the Environment through Contingent Valuation, Journal of
Economic Perspectives, 8, 19–44.
Hannan, E. J. (1980), The Estimation of the Order of an ARMA Process, Annals of Statistics,
8, 1071–1081.
Hansen, L. P. (1982), Large Sample Properties of Generalized Method of Moments Estimators,
Econometrica, 50, 1029–1054.
Hansen, L. P., Heaton, J. and Yaron, A. (1996), Finite Sample Properties of Some Alternative
GMM Estimators, Journal of Business and Economic Statistics, 14, 262–280.
Hansen, L. P. and Singleton, K. (1982), Generalized Instrumental Variables Estimation of Nonlinear Rational Expectations Models, Econometrica, 50, 1269–1286.
Hargreaves, C. P. (1994), A Review of Methods of Estimating Cointegrating Relationships. In:
C. P. Hargreaves, ed., Nonstationary Time Series Analysis and Cointegration, Oxford University Press, Oxford.

414

BIBLIOGRAPHY

Harris, R. D. F. and Tzavalis, E. (1999), Inference for Unit Roots in Dynamic Panels Where
the Time Dimension is Fixed, Journal of Econometrics, 91, 201–226.
Harris, R. I. D. (1995), Using Cointegration Analysis in Econometric Modelling, Prentice
Hall–Harvester Wheatsheaf, London.
Hasza, D. P. and Fuller, W. A. (1979), Estimation for Autoregressive Processes with Unit Roots,
Annals of Statistics, 7, 1106–1120.
Hausman, J. A. (1978), Speciﬁcation Tests in Econometrics, Econometrica, 46, 1251–1271.
Hausman, J. A., Hall, B. H. and Griliches, Z. (1984), Econometric Models Based on Count Data
with an Application to the Patents-R&D Relationship, Econometrica, 52, 909–938.
Hausman, J. A. and McFadden, D. F. (1984), A Speciﬁcation Test for the Multinomial Logit
Model, Econometrica, 52, 1219–1240.
Hausman, J. A. and Taylor, W. E. (1981), Panel Data and Unobservable Individual Effects,
Econometrica, 49, 1377–1398.
Hausman, J. A. and Wise, D. A. (1979), Attrition Bias in Experimental and Panel Data: The
Gary Income Maintenance Experiment, Econometrica, 47, 455–473.
Hayashi, F. (2000), Econometrics, Princeton University Press, Princeton, NJ.
Heckman, J. J. (1978a), Simple Statistical Models for Discrete Panel Data Developed and
Applied to Test the Hypothesis of True State Dependence against the Hypothesis of Spurious
State Dependence, Annales de l’INSEE, 30/31, 227–269.
Heckman, J. J. (1978b), Dummy Endogenous Variables in a Simultaneous Equations System,
Econometrica, 46, 931–960.
Heckman, J. J. (1979), Sample Selection Bias as a Speciﬁcation Error, Econometrica, 47,
153–161.
Heckman, J. J. (1981), The Incidental Parameters Problem and the Problem of Initial Conditions in Estimating a Discrete Time–Discrete Data Stochastic Process. In: C. F. Manski and
D. F. McFadden, eds., Structural Analysis of Discrete Data with Econometric Applications,
MIT Press, Cambridge, 179–195.
Heckman, J. J. (1990), Varieties of Selection Bias, American Economic Review, 80, 313–318.
Heckman, J. J. (1997), Instrumental Variables: A Study of Implicit Behavioral Assumptions
Used in Making Program Evaluations, Journal of Human Resources, 32, 441–462.
Heckman, J. J. (2001), Micro Data, Heterogeneity, and the Evaluation of Public Policy: Nobel
Lecture, Journal of Political Economy, 109, 673–748.
Heckman, J. J., Tobias, J. L. and Vytlacil, E. (2003), Simple Estimators for Treatment Parameters in a Latent-Variable Framework, Review of Economics and Statistics, 85, 748–755.
Henriksson, R. D. and Merton, R. C. (1981), On Market Timing and Investment Performance.
II. Statistical Procedures for Evaluating Forecasting Skills, Journal of Business, 54, 513–533.
Hildreth, C. and Lu, J. (1960), Demand Relations with Autocorrelated Disturbances, Technical
Bulletin No. 276, Michigan State University.
Hoffman, D. L. and Rasche, R. H. (1996), Assessing Forecast Performance in a Cointegrated
System, Journal of Applied Econometrics, 11, 495–517.
Honoré, B. E. (1992), Trimmed LAD and Least Squares Estimation of Truncated and Censored
Regression Models with Fixed Effects, Econometrica, 60, 533–565.
Honoré, B. E. (1993), Orthogonality Conditions for Tobit Models with Fixed Effects and Lagged
Dependent Variables, Journal of Econometrics, 59, 35–61.
Honoré, B. E. and Kyriazidou, E. (2000), Panel Data Discrete Choice Models with Lagged
Dependent Variables, Econometrica, 68, 839–874.
Hoover, K. D. and Perez, S. J. (1999), Data Mining Reconsidered: Encompassing and the
General-to-Speciﬁc Approach to Speciﬁcation Search, Econometrics Journal, 2, 167–191.
Horowitz, J. L. (1992), A Smoothed Maximum Score Estimator for the Binary Response Model,
Econometrica, 60, 505–531.

BIBLIOGRAPHY

415

Horowitz, J. L. (1998), Semiparametric Methods in Econometrics, Springer-Verlag, New York.
Hsiao, C. (1985), Beneﬁts and Limitations of Panel Data, Econometric Reviews, 4, 121–174.
Hsiao, C. (2003), Analysis of Panel Data, 2nd edition, Cambridge University Press, Cambridge.
Im, K., Pesaran, M. H. and Shin, Y. (2003), Testing for Unit Roots in Heterogeneous Panels,
Journal of Econometrics, 115, 29–52.
Imbens, G. W. and Angrist, J. D. (1994), Identiﬁcation and Estimation of Local Average Treatment Effects, Econometrica, 62, 467–476.
Jarque, C. M. and Bera, A. K. (1980), Efﬁcient Tests for Normality, Homoskedasticity and Serial
Independence of Regression Residuals, Economics Letters, 6, 255–259.
Jenkins, S. P. (2003), Survival Analysis, unpublished manuscript, ISER, University of
Essex, Colchester. Downloadable from http://www.iser.essex.ac.uk/teaching/stephenj/ec968/
index.php.
Johansen, S. (1988), Statistical Analysis of Cointegration Vectors, Journal of Economic Dynamics and Control, 12, 231–254.
Johansen, S. (1991), Estimation and Hypothesis Testing of Cointegrating Vectors in Gaussian
Vector Autoregressive Models, Econometrica, 59, 1551–1580.
Johansen, S. (1995), Likelihood-Based Inference in Cointegrated Vector Autoregressive Models,
Oxford University Press, Oxford.
Johansen, S. (2002), A Small Sample Correction for the Test of Cointegrating Rank in the Vector
Autoregressive Model, Econometrica, 70, 1929–1961.
Johansen, S. and Juselius, K. (1990), Maximum Likelihood Estimation and Inference on Cointegration – with Applications to the Demand for Money, Oxford Bulletin of Economics and
Statistics, 52, 169–210.
Johnston, J. and Dinardo, J. (1997), Econometric Methods, 4th edition, McGraw-Hill, New York.
Kao, C. (1999), Spurious Regression and Residual-Based Tests for Cointegration in Panel Data,
Journal of Econometrics, 90, 1–44.
Keane, M. P. (1993), Simulation Estimation for Panel Data Models with Limited Dependent
Variables. In: G. S. Maddala, C. R. Rao and H. D. Vinod, eds., Handbook of Statistics, Volume
XI, Elsevier Science, Amsterdam, 545–571.
Keane, M. P., Mofﬁtt, R. and Runkle, D. (1988), Real Wages over the Business Cycle: Estimating the Impact of Heterogeneity with Micro Data, Journal of Political Economy, 96,
1232–1266.
Kiefer, N. (1980), Estimation of Fixed Effects Models for Time Series of Cross-Sections with
Arbitrary Intertemporal Covariance, Journal of Econometrics, 14, 195–202.
Kiefer, N. (1988), Economic Duration Data and Hazard Functions, Journal of Economic Literature, 26, 646–679.
Kmenta, J. (1986), Elements of Econometrics, MacMillan, New York.
Konings, J. and Roodhooft, F. (1997), How Elastic is the Demand for Labour in Belgian Enterprises? Results from Firm Level Panel Data, 1986–1994, De Economist, 145, 229–241.
Krolzig, H.-M. and Hendry, D. F. (2001), Computer Automation of General-to-Speciﬁc Model
Selection Procedures, Journal of Economic Dynamics and Control, 25, 831–866.
Kwiatkowski, D., Phillips, P. C. B., Schmidt, P. and Shin, Y. (1992), Testing the Null Hypothesis of Stationarity Against the Alternative of a Unit Root: How Sure Are We That Economic
Time Series Have a Unit Root? Journal of Econometrics, 54, 159–178.
Lafontaine, F. and White, K. J. (1986), Obtaining Any Wald Statistic You Want, Economics
Letters, 21, 35–40.
Lancaster, T. (2000), The Incidental Parameter Problem Since 1948, Journal of Econometrics,
95, 391–413.
Layard, R. and Nickell, S. J. (1986), Unemployment in Britain, Economica (Supplement: Unemployment), 53, S121–S169.

416

BIBLIOGRAPHY

Leamer, E. (1978), Speciﬁcation Searches, John Wiley and Sons, New York.
Lee, M. J. (1996), Methods of Moments and Semiparametric Econometrics for Limited Dependent Variable Models, Springer-Verlag, New York.
Lee, M. J. (1999), A Root-N Consistent Semiparametric Estimator for Related Effects Binary
Response Panel Data, Econometrica, 67, 427–433.
Lee, L. F. and Maddala, G. S. (1985), The Common Structure of Tests for Selectivity Bias,
Serial Correlation, Heteroskedasticity and Non-Normality in the Tobit Model, International
Economic Review, 26, 1–20.
Leung, S. F. and Yu, S. (1996), On the Choice Between Sample Selection and Two-Part Models,
Journal of Econometrics, 72, 197–229.
Levin, A. and Lin, C.-F. (1993), Unit Root Tests in Panel Data: New Results, Discussion Paper,
Department of Economics, University of San Diego.
Levin, A., Lin, C.-F. and Chu, S.-S. J. (2002), Unit Root Tests in Panel Data: Asymptotic and
Finite-Sample Properties, Journal of Econometrics, 108, 1–24.
Lin, J.-L. and Tsay, R. S. (1996), Co-Integration Constraint and Forecasting: An Empirical
Examination, Journal of Applied Econometrics, 11, 519–538.
Little, R. J. A. and Rubin, D. B. (1987), Statistical Analysis with Missing Data, John Wiley and
Sons, New York.
Ljung, G. M. and Box, G. E. P. (1978), On a Measure of Lack of Fit in Time Series Models,
Biometrika, 65, 297–303.
Lo, A. and MacKinlay, C. (1990), Data-Snooping Biases in Tests of Financial Asset Pricing
Models, Review of Financial Studies, 3, 431–468.
Louviere, J. J. (1988), Conjoint Analysis Modeling of Stated Preferences. A Review of Theory,
Methods, Recent Developments and External Validity, Journal of Transport Economics and
Policy, 22, 93–119.
Lovell, M. C. (1983), Data Mining, Review of Economics and Statistics, 65, 1–12.
Lütkepohl, H. (1991), Introduction to Multiple Time Series Analysis, Springer-Verlag, Berlin.
MacKinnon, J. G. (1991), Critical Values for Cointegration Tests. In: R. F. Engle and
C. W. J. Granger, eds., Long-Run Economic Relationships: Readings in Cointegration, Oxford
University Press, 267–276.
MacKinnon, J. G., White, H. and Davidson, R. (1983), Test for Model Speciﬁcation in the Presence of Alternative Hypotheses: Some Further Results, Journal of Econometrics, 21, 53–70.
Maddala, G. S. (1983), Limited-Dependent and Qualitative Variables in Econometrics, Cambridge University Press, Cambridge.
Maddala, G. S. (1987), Limited Dependent Variable Models Using Panel Data, The Journal of
Human Resources, 22, 307–338.
Maddala, G. S. (2001), Introduction to Econometrics, 3rd edition, John Wiley and Sons, Chichester.
Maddala, G. S. and Kim, I.-M. (1998), Unit Roots, Cointegration and Structural Change, Cambridge University Press, New York.
Maddala, G. S. and Wu, S. (1999), A Comparative Study of Unit Root Tests with Panel Data
and a New Simple Test, Oxford Bulletin of Economics and Statistics, 61, 631–652.
Maddala, G. S., Wu, S. and Liu, P. C. (2000), Do Panel Data Rescue Purchasing Power Parity
(PPP) Theory? In: J. Krishnakumar and E. Ronchetti, eds., Panel Data Econometrics: Future
Directions, North Holland, Amsterdam, 35–51.
Manski, C. F. (1975), Maximum Score Estimation of the Stochastic Utility Model of Choice,
Journal of Econometrics, 3, 205–228.
Manski, C. F. (1985), Semiparametric Analysis of Discrete Response, Journal of Econometrics,
27, 313–333.
Manski, C. F. (1987), Semiparametric Analysis of Random Effects Linear Models from Binary
Panel Data, Econometrica, 55, 357–362.

BIBLIOGRAPHY

417

Manski, C. F. (1989), Anatomy of the Selection Problem, The Journal of Human Resources, 24,
243–260.
Manski, C. F. (1994), The Selection Problem. In: C. A. Sims, ed., Advances in Econometrics,
Sixth World Congress, Volume I, Cambridge University Press, Cambridge, 143–170.
Martins, M. (2001), Parametric and Semiparametric Estimation of Sample Selection Models: An
Empirical Application to the Female Labour Force in Portugal, Journal of Applied Econometrics, 16, 23–40.
Marquering, W. and Verbeek, M. (1999), An Empirical Analysis of Intertemporal Asset Pricing
Models with Transactions Costs and Habit Persistence, Journal of Empirical Finance, 6,
243–265.
McCall, B. P. (1995), The Impact of Unemployment Insurance Beneﬁt Levels on Recipiency,
Journal of Business and Economic Statistics, 13, 189–198.
McCulloch, J. H. and Kwon, H. C. (1993), U.S. Term Structure Data, 1947–1991, Ohio State
working paper 93–6, Ohio State University, Columbus, OH.
McFadden, D. F. (1974), Conditional Logit Analysis of Qualitative Choice Behavior. In:
P. Zaremba, ed., Frontiers in Econometrics, Academic Press, New York, 105–142.
Mehra, R. and Prescott, E. (1985), The Equity Premium: A Puzzle, Journal of Monetary Economics, 15, 145–161.
Mills, T. C. (1990), Time Series Techniques for Economists, Cambridge University Press, Cambridge.
Mizon, G. E. (1984), The Encompassing Approach in Econometrics. In: K. F. Wallis and
D. F. Hendry, eds., Quantitative Economics and Econometric Analysis, Basil Blackwell,
Oxford, 135–172.
Mizon, G. E. and Richard, J. F. (1986), The Encompassing Principle and its Application to
Testing Non-Nested Hypotheses, Econometrica, 54, 657–678.
Mundlak, Y. (1961), Empirical Production Function Free of Management Bias, Journal of Farm
Economics, 43, 44–56.
Nelson, C. R. and Plosser, C. I. (1982), Trends and Random Walks in Macro-economic Time
Series: Some Evidence and Implications, Journal of Monetary Economics, 10, 139–162.
Nelson, D. (1990), Conditional Heteroskedasticity in Asset Returns: A New Approach, Econometrica, 59, 347–370.
Newey, W. K. (1985), Maximum Likelihood Speciﬁcation Testing and Conditional Moment
Tests, Econometrica, 53, 1047–1070.
Newey, W. K. and West, K. (1987), A Simple Positive Semi-Deﬁnite, Heteroskedasticity and
Autocorrelation Consistent Covariance Matrix, Econometrica, 55, 703–708.
Newey, W. K., Powell, J. L. and Walker, J. R. (1990), Semiparametric Estimation of Selection
Models: Some Empirical Results, American Economic Review, 80, 324–328.
Nickell, S. (1981), Biases in Dynamic Models with Fixed Effects, Econometrica, 49, 1417–1426.
Nijman, Th. E. and Verbeek, M. (1990), Estimation of Time Dependent Parameters in Linear
Models Using Cross Sections, Panels or Both, Journal of Econometrics, 46, 333–346.
Nijman, Th. E. and Verbeek, M. (1992), Nonresponse in Panel Data: The Impact on Estimates
of a Life Cycle Consumption Function, Journal of Applied Econometrics, 7, 243–257.
O’Connell, P. G. J. (1998), The Overvaluation of Purchasing Power Parity, Journal of International Economics, 44, 1–19.
Oh, K. (1996), Purchasing Power Parity and Unit Root Tests Using Panel Data, Journal of
International Money and Finance, 15, 405–418.
Ongena, S. and Smith, D. C. (2001), The Duration of Bank Relationships, Journal of Financial
Economics, 61, 449–475.
Owen, P. D. (2003), General-to-Speciﬁc Modelling Using PcGets, Journal of Economic Surveys,
17, 609–628.
Pagan, A. and Ullah, A. (1999), Nonparametric Econometrics, Cambridge University Press,
Cambridge.

418

BIBLIOGRAPHY

Pagan, A. and Vella, F. (1989), Diagnostic Tests for Models Based on Individual Data: A Survey,
Journal of Applied Econometrics, 4, S29–S59.
Pagan, A., Hall, A. D. and Martin, V. (1996), Modeling the Term Structure. In: G. S. Maddala
and C. R. Rao, eds., Handbook of Statistics, Volume XIV, Elsevier Science, Amsterdam,
91–118.
Patterson, K. D. (2000), An Introduction to Applied Econometrics: A Time Series Approach,
MacMillan Press, London.
Pedroni, P. (1999), Critical Values for Cointegration Tests in Heterogeneous Panels with Multiple
Regressors, Oxford Bulletin of Economics and Statistics, 61, 653–678.
Pesaran, M. H. and Smith, R. P. (1995), Estimation of Long-Run Relationships from Dynamic
Heterogeneous Panels, Journal of Econometrics, 68, 79–113.
Pesaran, M. H., Shin, Y. and Smith, R. J. (2000), Structural Analysis of Vector Error Correction
Models with Exogenous I (1) Variables, Journal of Econometrics, 97, 293–343.
Phillips, P. C. B. (1986), Understanding Spurious Regressions in Econometrics, Journal of
Econometrics, 33, 311–340.
Phillips, P. C. B. and Hansen, B. E. (1990), Statistical Inference in Instrumental Variables
Regression with I (1) Processes, Review of Economic Studies, 57, 99–125.
Phillips, P. C. B. and Moon, H. R. (1999), Linear Regression Limit Theory for Nonstationary
Panel Data, Econometrica, 67, 1057–1111.
Phillips, P. C. B. and Park, J. Y. (1988), On the Formulation of Wald Tests of Nonlinear Restrictions, Econometrica, 56, 1065–1083.
Phillips, P. C. B. and Perron, P. (1988), Testing for a Unit Root in Time Series Regression,
Biometrika, 75, 335–346.
Poon, S.-H. and Granger, C. W. J. (2003), Forecasting Volatility in Financial Markets: A
Review, Journal of Economic Literature, 41, 478–539.
Portney, P. R. (1994), The Contingent Valuation Debate: Why Should Economists Care?, Journal of Economic Perspectives, 8, 3–18.
Prais, S. and Winsten, C. (1954), Trend Estimation and Serial Correlation, Cowles Commission
Discussion Paper 383, Chicago.
Quah, D. (1994), Exploiting Cross-Section Variation for Unit Root Inference in Dynamic Data,
Economics Letters, 44, 9–19.
Ramsey, J. B. (1969), Tests for Speciﬁcation Errors in Classical Linear Least Squares Regression
Analysis, Journal of the Royal Statistical Society B, 32, 350–371.
Reimers, H.-E. (1992), Comparison of Tests for Multivariate Cointegration, Statistical Papers,
33, 335–359.
Robertson, D. and Symons, J. (1992), Some Strange Properties of Panel Data Estimators, Journal
of Applied Econometrics, 7, 175–189.
Robinson, P. M. (1982), On the Asymptotic Properties of Estimators of Models Containing
Limited Dependent Variables, Econometrica, 50, 27–41.
Rosen, S. (1974), Hedonic Prices and Implicit Markets: Product Differentiation in Perfect Competition, Journal of Political Economy, 82, 34–55.
Rubin, D. B. (1976), Inference and Missing Data, Biometrika, 63, 581–592.
Ruud, P. A. (1984), Test of Speciﬁcation in Econometrics, Econometric Reviews, 3, 211–242.
Said, S. E. and Dickey, D. A. (1984), Testing for Unit Roots in Autoregressive Moving Average
Models of Unknown Order, Biometrika, 71, 599–607.
Sargan, J. D. and Bhargava, A. S. (1983), Testing Residuals from Least Squares Regression for
Being Generated by the Gaussian Random Walk, Econometrica, 51, 213–248.
Savin, N. E. and White, K. J. (1977), The Durbin–Watson Test for Serial Correlation with
Extreme Sample Sizes or Many Regressors, Econometrica, 45, 1989–1996.

BIBLIOGRAPHY

419

Schwarz, G. (1978), Estimating the Dimension of a Model, Annals of Statistics, 6, 461–464.
Sims, C. A. (1980), Macroeconomics and Reality, Econometrica, 48, 1–48.
Smith, R. P. and Fuertes, A.-M. (2003), Panel Time Series, working paper, Dept. of Economics,
Birkbeck College, London.
Staiger, D. and Stock, J. H. (1997), Instrumental Variables Regression with Weak Instruments,
Econometrica, 65, 557–586.
Stewart, J. and Gill, L. (1998), Econometrics, 2nd edition, Prentice Hall, London.
Stock, J. H. and Watson, M. W. (1993), A Simple Estimator of Cointegrating Vectors in Higher
Order Integrated Systems, Econometrica, 61, 783–820.
Stock, J. H. and Watson, M. W. (2003), Introduction to Econometrics, Addison-Wesley, Boston.
Stock, J. H. and Wright, J. H. (2003), GMM with Weak Identiﬁcation, Econometrica, 68,
1055–1096.
Sullivan, R., Timmermann, A. and White, H. (2001), Dangers of Data-Driven Inference: The
Case of Calender Effects in Stock Returns, Journal of Econometrics, 105, 249–286.
Tauchen, G. E. (1985), Diagnostic Testing and Evaluation of Maximum Likelihood Models,
Journal of Econometrics, 30, 415–443.
Taylor, M. P. (1995), The Economics of Exchange Rates, Journal of Economic Literature, 33,
12–47.
Ter Horst, J. R., Nijman, Th. E. and Verbeek, M. (2001), Eliminating Look-Ahead Bias in
Evaluating Persistence in Mutual Fund Performance, Journal of Empirical Finance, 8,
345–373.
Tobin, J. (1958), Estimation of Relationships for Limited Dependent Variables, Econometrica,
26, 24–36.
Vella, F. (1998), Estimating Models with Sample Selection Bias: A Survey, Journal of Human
Resources, 33, 127–169.
Vella, F. and Verbeek, M. (1998), Whose Wages Do Unions Raise? A Dynamic Model of Unionism and Wage Rate Determination for Young Men, Journal of Applied Econometrics, 13,
163–183.
Vella, F. and Verbeek, M. (1999a), Two-Step Estimation of Panel Data Models with Censored
Endogenous Variables and Selection Bias, Journal of Econometrics, 90, 239–263.
Vella, F. and Verbeek, M. (1999b), Estimating and Interpreting Models with Endogenous Treatment Effects, Journal of Business and Economic Statistics, 17, 473–478.
Verbeek, M. (1995), Alternative Transformations to Eliminate Fixed Effects, Econometric
Reviews, 14, 205–211.
Verbeek, M. and Nijman, Th. E. (1992), Testing for Selectivity Bias in Panel Data Models,
International Economic Review, 33, 681–703.
Verbeek, M. and Nijman, Th. E. (1996), Incomplete Panels and Selection Bias. In: L. Mátyás,
and P. Sevestre, eds., The Econometrics of Panel Data. A Handbook of the Theory with Applications, 2nd revised edition, Kluwer Academic Publishers, Dordrecht, 449–490.
Wallis, K. F. (1979), Topics in Applied Econometrics, 2nd edition, Basil Blackwell, Oxford.
Weeks, M. (1995), Circumventing the Curse of Dimensionality in Applied Work Using Computer
Intensive Methods, Economic Journal, 105, 520–530.
White, H. (1980), A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct
Test for Heteroskedasticity, Econometrica, 48, 817–838.
White, H. (1982), Maximum Likelihood Estimation of Misspeciﬁed Models, Econometrica, 50,
1–25.
White, H. (1990), A Consistent Model Selection Procedure Based on m-Testing. In:
C. W. J. Granger, ed., Modelling Economic Series: Readings in Econometric Methodology,
Clarendon Press, Oxford, 369–383.

420

BIBLIOGRAPHY

Winkelmann, R. (2003), Econometric Analysis of Count Data, 4th edition, Springer-Verlag,
Heidelberg.
Wooldridge, J. M. (1995), Selection Corrections for Panel Data Models under Conditional Mean
Independence Assumptions, Journal of Econometrics, 68, 115–132.
Wooldridge, J. M. (2002), Econometric Analysis of Cross-Section and Panel Data, MIT Press,
Cambridge, MA.
Wooldridge, J. M. (2003), Introductory Econometrics: A Modern Approach, 2nd edition, Thomson South-Western, Mason, OH.

Index
2SLS estimator, 145
ability bias, 138
abstention, 233
ACF, see autocorrelation function
ADF test, see augmented Dickey-Fuller test
adjusted R squared, 22, 58
Akaike Information Criterion (AIC), 58, 285
Almost Ideal Demand System, 223
alternative hypothesis, 24
Anderson-Hsiao estimator, 362
AR (p) process, 261
ARCH, see autoregressive conditional
heteroskedasticity
ARMA models, 261
choosing lag length, 281
diagnostic checking, 284
estimation, 279
forecast accuracy, 291
forecasting, 290
formulation, 261
model selection, 285
parsimony, 285
stationarity, 267
artiﬁcial nesting model, 60
asymptotic distribution, 34
asymptotic efﬁciency, 167
asymptotic theory, 32, 36
ATE, 241
average treatment effect, 241
augmented Dickey–Fuller test, 271
on regression residuals, 316
autocorrelation, 16, 97, 259
Breusch–Godfrey test, 101
Cochrane–Orcutt estimator, 100
Cochrane–Orcutt transformation, 100
Durbin–Watson test, 102
ﬁrst order autocorrelation, 98, 259

fourth order autocorrelation, 107
HAC standard errors, 111, 125
in panel data models, 356
iterative Cochrane Orcutt estimator, 101
lagged dependent variable, 126
Lagrange multiplier test, 101, 181
moving average, 107, 117
Newey–West standard errors, 111, 125,
356
overlapping samples, 116
Prais–Winsten estimator, 100
testing for ﬁrst order autocorrelation, 101
autocorrelation function (ACF), 259, 281
estimation, 282
autocovariance, 257, 259
autoregressive conditional heteroskedasticity,
297
ARCH(1) process, 298
ARCH(p) process, 298
ARCH-in-mean, 301
EGARCH process, 300
estimation of ARCH models, 301
exponential GARCH process, 300
forecasting, 302
GARCH model, 299
integrated GARCH, 299
news impact curve, 300
testing for ARCH(p) errors, 298
autoregressive distributed lag model, 310
autoregressive panel data model, 360
autoregressive process, 98, 256
moving average representation, 258, 264
auxiliary regression, 63
backshift operator, see lag operator
balanced sub-panel, 381

422

Balestra–Nerlove estimator, 350
Bartlett weights, 110
baseline hazard function, 246
best linear approximation, 9, 63
best linear unbiased estimator, 18
between estimator, 349
BIC, 58, 285
binary choice model, 190
estimation, 193
generalized residual, 193
goodness-of-ﬁt, 194
hit rate, 196
linear probability model, 190
logit model, 190
maximum score estimator, 202
probit model, 190
semi-parametric estimation, 202
single index model, 201
speciﬁcation tests, 199
with panel data, 373
binomial distribution, 169
BLUE, 18, 82
Box–Cox transformation, 61
Breusch–Godfrey test, 101
Breusch–Pagan test, 91, 180
in panel data models, 358
cancelling roots, 265
Capital Asset Pricing Model (CAPM), 38
causal effect, 76, 137, 240, 360
causal interpretation, 132
censored regression model, see tobit model
censoring, 219, 247
ceteris paribus condition, 15, 29, 52, 132
characteristic equation, 264
characteristic roots, 264
Chebycheff inequality, 32
Chi-squared distribution, 405
Chow test, 64
CM test, 184
Cochrane–Orcutt estimator, 100
Cochrane–Orcutt transformation, 100
coefﬁcient of relative risk aversion, 150
cointegrating matrix, 325
cointegrating parameter, 315
cointegrating rank, 325
Cointegrating Regression Durbin–Watson
test, 316
cointegrating space, 324
cointegrating vector, 315
cointegration, 314
cointegrating matrix, 325
cointegrating rank, 325
cointegrating regression, 315
cointegrating space, 324
cointegrating vector, 315

INDEX

CRDW test, 316
dynamic OLS, 317
error-correction representation, 318
fully modiﬁed OLS, 317
Granger representation theorem, 318, 325
in a VAR, 325
in panel data, 372
Johansen procedure, 329
long-run matrix, 326
long-run-equilibrium, 315
maximum eigenvalue test, 330
multivariate case, 324
structural cointegrating relationships, 331
super consistency, 314
testing for cointegration, 316, 328
testing in a VAR, 328
trace test, 330
common roots, 265
common trend, 314
conditional density function, 402
conditional expectation, 15, 402
of a normal distribution, 404
conditional heteroskedasticity, 123, 298
conditional maximum likelihood
ARMA models, 280
panel data models, 374
conditional mean independence, 122, 402
conditional moment test, 184
conditional normal distribution, 404
conditional variance
of a normal distribution, 404
conﬁdence interval, 25
consistency, 33
rate of convergence, 35, 202, 314
consumption-based asset pricing model, 154
contingent valuation, 205
control variable, 52
convergence in distribution, 35
convergence in probability, 33
correlation coefﬁcient, 401
correlogram, see autocorrelation function
count data model, 211
goodness-of-ﬁt, 216
likelihood function, 212
NegBin I model, 213
NegBin II model, 213
overdispersion, 213
Poisson regression model, 211
covariance, 400
covariance matrix, 401
covered interest rate parity, 112
Cramer–Rao lower bound, 168
CRDW test, 316
critical value, 24
cumulative density function (cdf), 398
curse of dimensionality, 376

INDEX

data mining, 56
data snooping, 56
degrees of freedom correction, 18
demand function, 222
deterministic regressors, 14
deterministic trend, 269, 318, 326
DF test, see Dickey–Fuller test
dichotomous models, 190
Dickey–Fuller test, 269
difference stationarity, 270
distributed lag model, 310
distribution
conditional distribution, 402
continuous distribution, 398
degenerate distribution, 398
discrete distribution, 398
leptokurtic, 400
median, 398
mode, 399
skewness, 400
symmetric distribution, 399
disturbance term, see error term
double-bounded procedure, 205
drift, see deterministic trend
dummy endogenous variable, 241
dummy variable, 12, 42
dummy variable trap, 73
duration dependence, 247
duration models, 244
duration dependence, 247
estimation, 248
ﬂow sampling, 247
hazard function, 245
left-censoring, 247
proportional hazard models, 246
right-censoring, 247
stock sampling, 247
survivor function, 245
Durbin–Watson test, 102
in cointegrating regression, 316
in spurious regression, 314
in panel data models, 357
Durbin–Wu–Hausman test, 135
dynamic forecast, 110
efﬁcient market hypothesis, 113, 122
EGLS, 82, 86
elasticity, 53, 222
encompassing, 60
encompassing F -test, 60
endogenous regressors, 132
Engel curve, 222, 233, 252
equidispersion, 212
equilibrium, 311, 315
steady state equilibrium, 318
steady state growth path, 318

423

equilibrium multiplier, 311
equity premium puzzle, 156
error components model, see random effects
model
error term, 14
i.i.d., 122
n.i.d., 19, 164
normally distributed, 19, 20
error-correction model, 311, 318
cointegrated variables, 318
obtained from VAR, 316
speed of adjustment, 311
stability, 311
with stationary variables, 311
estimate, 15
estimator, 15
best linear unbiased, 18
consistent, 32
linear estimator, 18
unbiased, 17
exact identiﬁcation, 144
exact multicollinearity, 42, 73
excess kurtosis, 185, 201, 226, 400
exclusion restriction, 135, 230, 239, 244
exogeneity, 15, 129
predeterminedness, 366
strict exogeneity, 346, 366
weak exogeneity, 319
expectations hypothesis, 294
expected value, 398
F distribution, 406
F -test, 27, 85
encompassing F -test, 60
non-nested F -test, 60
fat tails, 185, 201, 400
feasible generalized least squares (FGLS), 82,
86
Fisher relation, 333
ﬁxed effects estimator, 346
as IV estimator, 354
covariance matrix, 347
in dynamic model, 361
with unbalanced data, 382
ﬁxed effects logit model, 375
ﬁxed effects model, 342, 345
ﬁxed effects vs. random effects, 351
ﬂow sampling, 247
forward discount, 114
forward rate, 112
fractional integration, 274
functional form, 62
interaction terms, 52, 64
RESET test, 63
structural break, 63
testing, 63

424

GARCH, see autoregressive conditional
heteroskedasticity
Gauss–Markov assumptions, 16
Gauss–Markov theorem, 18
general-to-speciﬁc modelling, 57
generalized instrumental variables estimator
(GIVE), 145
generalized least squares, 82
generalized method of moments, 148
asymptotic distribution, 151
conditional moment condition, 149
GMM estimator, 151
iterated GMM estimator, 156
one-step GMM estimator, 156
optimal GMM estimator, 151
optimal weighting matrix, 151
overidentifying restrictions test, 152
two-step GMM estimator, 151
generalized residual, 193, 200, 225, 243
GIV estimator, 145
GLS estimator, 82
with autocorrelation, 99
with heteroskedasticity, 84
in random effects model, 348
GMM estimator, 151
Goldfeld-Quandt test, 91
goodness-of-ﬁt, 20
in binary choice models, 194
in count data models, 216
in linear models, 20
in model selection, 58
in panel data models, 352
likelihood ratio index, 195, 216
Granger representation theorem, 318, 325
Hausman test, 135, 351, 384
endogeneity test, 135
ﬁxed effects versus random effects, 351
sample selection bias in panel data, 384
Hausman–Taylor estimator, 354
hazard function, 245
baseline hazard, 246
log-logistic hazard, 247
Weibull hazard, 247
Heckman’s lambda, 229, 243
hedonic price, 65
heteroskedasticity, 70, 82, 123
Breusch–Pagan test, 91, 180
conditional heteroskedasticity, 123
GARCH models, 299
Goldfeld–Quandt test, 91
in binary choice models, 202
in panel data models, 355
in tobit models, 226
LM test in linear model, 91, 180
multiplicative heteroskedasticity, 89

INDEX

testing for, 90
weighted least squares, 84
White test, 92
heteroskedasticity-and-autocorrelationconsistent standard errors, 111,
125
in panel data models, 356
heteroskedasticity-consistent standard errors,
88
hit rate, 196
homoskedasticity, 16
human capital earnings function, 137
hypothesis testing, 23
actual size, 37
alternative hypothesis, 24
critical values, 24
null hypothesis, 23
one-sided test, 24
p-value, 31
power, 31
signiﬁcance level, 24
size, 31
type I error, 31
type II error, 31
i.i.d., 122
identiﬁcation, 131
exact identiﬁcation, 144
normalization constraints, 204
overidentiﬁcation, 144
overidentifying restrictions test, 147, 152
sample selection problem, 238
underidentiﬁcation, 144
identity, 129
idiosyncratic risk, 41
ignorable selection rule, 237
IID, 122
impact multiplier, 311
impulse-response function, 324
incidental parameter problem, 374
incomplete panel data, 380
inconclusive region, 103
independence, 400, 403
independence of irrelevant alternatives, 209
independent logit model, see multinomial
logit model
inferior good, 222
information matrix, 167
block diagonality, 175, 179
information matrix test, 184
information set, 288
initial conditions problem, 380
instrument, see instrumental variable
instrumental variable, 133
instrumental variables estimator, 133
asymptotic distribution, 134

INDEX

dummy endogenous variables, 244
generalized instrumental variables
estimator, 142, 145
Hausman test for endogeneity, 135
in dynamic panel data model, 362
in static panel data model, 353
two-stage least squares, 145
weak instruments, 147
weighting matrix, 144
integration, 267
interaction term, 52, 64, 244
intertemporal marginal rate of substitution,
155
iterated EGLS, 82
iterated GMM, 156
IV estimator, 133

425

Keynesian consumption function, 129
KPSS test, 271
kurtosis, 185, 400

limited dependent variables, 189
in panel data, 373
linear probability model, 190
linear regression model, 14
LM test, see Lagrange multiplier test
log Weibull distribution, 208
logistic distribution, 191, 247
logit model, 190
ﬁxed effects logit model, 375
generalized residual, 193
goodness-of-ﬁt, 194
hit rate, 196
likelihood function, 193
LM test, 200
nested logit model, 210
ordered logit model, 203
with heteroskedasticity, 202
loglikelihood function, 162
loglinear model, 53, 61, 71
log-logistic hazard, 247
lognormal distribution, 54, 407
long-run equilibrium, 315
long-run multiplier, 311
LR test, see likelihood ratio test
LSE methodology, 57
luxury good, 222

lag operator, 262
lag polynomial, 262
characteristic roots, 264
invertibility, 263
matrix lag polynomial, 322
lagged dependent variable, 122, 126
Lagrange multiplier test, 173
autocorrelation, 101, 181
heteroskedasticity, 91
heteroskedasticity in binary choice models,
200
heteroskedasticity in linear model, 179
heteroskedasticity in tobit model, 226
in tobit model, 225
normality in probit model, 201
omitted variables in binary choice model,
200
omitted variables in linear model, 178
outer product gradient (OPG) version, 176
latent model, 192
latent variable, 192, 203, 218
law of large numbers, 34
least squares dummy variable (LSDV)
estimator, 345
left-censoring, 247
leptokurtic, 400
likelihood function, 162
likelihood ratio index, 195, 216
likelihood ratio test, 173

M test, 184
MA process, 257
MA(q) process, 261
Maddala–Wu test, 372
marginal distribution, 400
marginal propensity to consume, 129
market portfolio, 38
market risk, 38, 41
matrix lag polynomial, 322
maximum likelihood, 162
ARIMA models, 280
asymptotic distribution, 167
BHHH estimator, 168
binary choice models, 193
cointegrating VAR, 329
conditional moment test, 184
consistency, 167
constrained ML estimator, 172
count data models, 211
duration models, 247
efﬁciency, 167
ﬁrst order condition, 163, 183
GARCH models, 301
Hessian matrix, 168
information matrix, 167
information matrix test, 184
likelihood contributions, 166
likelihood function, 162
likelihood ratio test, 172

J -test, 61
Jarque–Bera test, 185
Jensen’s inequality, 399

INDEX

426

maximum likelihood (continued )
loglikelihood function, 162
maximum likelihood estimator, 163
outer product of the gradients, 176
panel data binary choice models, 374
panel data tobit models, 378
pseudo-maximum likelihood, 183
quasi-maximum likelihood, 183
score vector, 166
second order condition, 163
tobit models, 220
maximum score estimator, 202, 380
mean reversion, 268
mean variance efﬁcient, 38
measurement error, 127, 136, 138
median, 398
missing at random, 381
misspeciﬁcation, 55, 62, 108, 178
mode, 399
model test, 28
model selection, 56, 285
moment conditions, 132, 142, 362
money demand equation, 333
Monte Carlo study, 36
moving average process, 107, 257
multi-response models, 202
multicollinearity, 9, 42
dummy variable trap, 73
exact multicollinearity, 42, 73
multinomial logit model, 209
independence of irrelevant alternatives, 209
odds ratio, 209
probability ratio, 209
multinomial probit model, 210
multiplier, 311
necessary good, 222
NegBin I model, 213
NegBin II model, 213
nested logit model, 210
nested models, 55
Newey–West standard errors, 111, 125, 356
news impact curve, 300
NID, 19, 164
noise-to-signal ratio, 128
non-nested F -test, 60
non-nested models, 59
nonlinear least squares, 62
in ARMA models, 280
nonlinear model, 62
nonresponse, 237
nonsense regression, 313
normal distribution, 19, 403
bivariate, 404
normal equations, 9
normality test, 185

in linear model, 185
in ordered probit model, 207
in probit model, 201
in tobit model, 226
Jarque–Bera test, 185
null hypothesis, 23
OLS, 8
estimator, 15
matrix notation, 12, 395
normal equations, 9
residual, 9
OLS estimator, 15
asymptotic properties, 32, 36, 122
small sample properties, 16, 36
omitted variable bias, 55, 344
omitted variables, 178
in panel data, 344
in probit model, 200
in tobit model, 226
one-sided test, 24
optimal predictor, 288
ordered logit model, 203
ordered probit model, 203
ordered response model, 203
ordinary least squares, see OLS
orthogonality, 55
overdispersion, 213
overﬁtting, 284
overidentiﬁcation, 144
overidentifying restrictions test, 147, 152
overlapping samples, 116, 125
p-value, 31
PACF, see partial autocorrelation function
panel data, 341
panel data cointegration tests, 372
panel data unit root tests, 369
parsimony, 58, 285
partial adjustment model, 312
partial autocorrelation coefﬁcient, 283
partial autocorrelation function, 284
patents, 215
PE test, 61
Peason family of distributions, 201
Phillips–Perron test, 273
plim, 33
Poisson distribution, 211
Poisson regression model, 211
population, 14
power, 31
PPP, see purchasing power parity
Prais–Winsten estimator, 100
predetermined regressors, 366
prediction, 44

INDEX

with ARMA models, 288
with GARCH models, 302
with VAR models, 323
prediction accuracy, 291
prediction error, 45, 291
prediction interval, 45
predictor, 45
unbiased, 45
pricing error, 156
probability density function (pdf), 398
joint density function, 400
probability limit, 33
probability mass function, 397
probit model, 191
generalized residual, 193
goodness-of-ﬁt, 194
hit rate, 196
likelihood function, 193
LM tests, 200
multinomial probit model, 210
normality test, 201
ordered probit model, 203
random effects ordered probit model, 378
random effects probit model, 376
with heteroskedasticity, 202
projection matrix, 13, 393
proportional hazards model, 246
pseudo-maximum likelihood, see
quasi-maximum likelihood
purchasing power parity, 276, 319, 331
pure expectations hypothesis, 293
quasi-maximum likelihood, 183
in count data models, 212
in GARCH models, 302
R squared, 21
adjusted R squared, 22, 58
between R squared, 353
likelihood ratio index, 195, 216
McFadden R squared, 195
overall R squared, 353
pseudo R squared, 195
uncentred R squared, 21
within R squared, 353
random effects estimator, 350
covariance matrix, 350
with unbalanced data, 382
random effects model, 342, 347
random effects ordered probit model, 378
random effects probit model, 376
random effects tobit model, 377
random sample, 14, 221, 228
random utility framework, 208
random variable, 14, 397

427

random walk, 117, 266
random walk with drift, 270
rate of convergence, 35, 202, 314
R&D expenditures, 215
real exchange rate, 278
reduced form, 130, 135, 148
reference group, 73
regression line, 10
replacement rate, 197
reservation wage, 229
RESET test, 63
residual, 9, 18
residual analysis, 284
residual sum of squares, 10
returns to schooling, 137
robust inference, 88, 111, 124, 355
right-censoring, 247
risk premium, 38, 112, 154, 333
Said–Dickey test, 273
sample, 14
sample selection model, 228
Heckman’s lambda, 229
likelihood function, 230
semi-parametric estimation, 239
two-step estimation, 231
sample selection problem, 76, 228, 237, 384
duration models, 247
identiﬁcation problem, 238
in panel data, 381, 383
sampling frame, 237
selection rule, 237
self-selection, 237
single index, 239
tests in panel data, 384
tobit II model, 228
treatment effects, 240
sampling frame, 237
sampling process, 14, 247
Sargan test, 147, 152
Schwarz Information Criterion, 58, 285
score test, see Lagrange multiplier test
score vector, 166
selection bias, see sample selection problem
selection rule, 237
self-selection, 237, 244
semi-parametric estimation
ﬁxed effects tobit model, 380
binary choice model, 202
panel data binary choice model, 380
sample selection model, 239
serial correlation, see autocorrelation
shadow prices of restrictions, 174
signiﬁcance level, 24
signiﬁcant, 25
simple linear regression, 10

428

simplicity, 58
simulation, 36, 376
simultaneity, 129
simultaneous equations model, 129
complete simultaneous equations model,
135
identiﬁcation, 131
reduced form, 130
structural form, 129
single index, 201, 239
single index model, 201
size, 31
skewness, 185, 201, 226, 400
small ﬁrm effect, 156
spurious regression, 313
in panel data, 373
standard deviation, 400
standard error, 19
state dependence, 344, 378
stationarity, 258, 266
covariance stationarity, 258
difference stationarity, 270
strict stationarity, 258
trend stationarity, 270
unit roots, 267
weak stationarity, 258
statistical model, 14
stochastic discount factor, 155
stochastic process, 256
stock sampling, 247
structural break, 63
structural form, 129
structural model, 129
Student distribution, see t distribution
subsample, 74, 88, 381
sufﬁcient statistic, 374
super consistency, 314
survivor function, 245
switching regression model, 242
symmetric distribution, 399
systematic risk, 38
t distribution, 406
t-ratio, 23
t-test, 23, 26, 84
t-value, 23
take-up rate, 197
term structure of interest rates, 293
test, see hypothesis testing
test statistic, 24
tobit model, 219
corner solution, 218, 223
extensions, 227
ﬁxed effects tobit model, 380
generalized residual, 225
heteroskedasticity, 226

INDEX

likelihood function, 220
normality test, 227
random effects tobit model, 377
second order generalized residual, 225
speciﬁcation tests, 225
standard tobit model (type I), 219
tobit II model, see sample selection model
tobit III model, 233
truncated regression model, 221
unobserved heterogeneity, 218, 378
total expenditure elasticity, 222
transformation matrix, 81, 348
treatment effect, 241
trend stationarity, 270
truncated normal distribution, 207, 219, 229,
405
truncated regression model, 221
truncation, 207, 221, 405
two-sided test, 24
TT, 241
type I error, 31, 56
type I extreme value distribution, 208
type II error, 31
unbalanced panel data, 381
unbiased, 17
uncentred R squared, 21
uncorrelatedness, 401
uncovered interest rate parity, 112
underidentiﬁcation, 144
unemployment beneﬁts, 197
unit root, 264
in panel data, 369
seasonal unit root, 274
stochastic unit root, 273
test for unit root in AR(1) model, 269
test for unit root in general ARIMA
model, 273
test for unit root in higher order AR
process, 272
unobserved heterogeneity, 378
unordered models, 208
VAR model, see vector autoregressive model
variance, 399
VARMA model, 322
vector autoregressive model (VAR), 322
determining lag length, 324
estimation, 323
forecasting, 323
impulse-response function, 324
stationarity, 326
vector error-correction model (VECM), 326
vector moving average model (VMA), 324
volatility clustering, 297

INDEX

Wald test, 30, 85, 172
weak exogeneity, 319
weak instruments, 147
weak stationarity, 258
Weibull hazard, 247
weighted least squares, 84
weighting matrix, 89, 144, 349
optimal weighting matrix (GMM), 151
optimal weighting matrix (IV), 144

429

white noise process, 256
White standard errors, 88
willingness to pay, 205
within estimator, 346
within transformation, 346

yield curve, 293
Yule–Walker equations, 283

Source Exif Data:

File Type                       : PDF
File Type Extension             : pdf
MIME Type                       : application/pdf
PDF Version                     : 1.6
Linearized                      : Yes
Create Date                     : 2005:08:31 17:59:22+06:00
Modify Date                     : 2013:04:14 01:52:49+01:00
Subject                         : 0470857749.pdf
Has XFA                         : No
XMP Toolkit                     : Adobe XMP Core 5.2-c001 63.139439, 2010/09/27-13:37:26
Metadata Date                   : 2013:04:14 01:52:49+01:00
Format                          : application/pdf
Description                     : 0470857749.pdf
Document ID                     : uuid:d853a72c-a8b2-4589-8f81-a3b086102302
Instance ID                     : uuid:373d7e1a-bdb3-4c59-a962-64806c2be83a
Page Layout                     : SinglePage
Page Mode                       : UseOutlines
Page Count                      : 448

EXIF Metadata provided by EXIF.tools

A Guide To Modern Econometrics

Navigation menu

Versions of this User Manual:

Views

Navigation