P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM

CB495-FMA CB495/Train September 18, 2002 10:54 Char Count= 0

Discrete Choice Methods

with Simulation

Kenneth E. Train
University of California, Berkeley
and
National Economic Research Associates, Inc.

iii
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-FMA CB495/Train September 18, 2002 10:54 Char Count= 0

published by the press syndicate of the university of cambridge

The Pitt Building, Trumpington Street, Cambridge, United Kingdom

cambridge university press

The Edinburgh Building, Cambridge CB2 2RU, UK
40 West 20th Street, New York, NY 10011-4211, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
Ruiz de Alarcón 13, 28014 Madrid, Spain
Dock House, The Waterfront, Cape Town 8001, South Africa
http://www.cambridge.org


C Kenneth E. Train 2003

This book is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.

First published 2003

Printed in the United Kingdom at the University Press, Cambridge

Typeface Times Roman 11/13 pt. System LATEX 2ε [tb]

A catalog record for this book is available from the British Library.

Library of Congress Cataloging in Publication Data

Train, Kenneth.
Discrete choice methods with simulation / Kenneth E. Train.
p. cm.
Includes bibliographical references and index.
ISBN 0-521-81696-3 – ISBN 0-521-01715-7 (pb.)
1. Decision making – Simulation methods. 2. Consumers’ preferences –
Simulation methods. I. Title.
HD30.23 .T725 2003
003 .56 – dc21 2002071479

ISBN 0 521 81696 3 hardback

ISBN 0 521 01715 7 paperback

iv
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-FMA CB495/Train September 18, 2002 10:54 Char Count= 0

Contents

1 Introduction page 1
1.1 Motivation 1
1.2 Choice Probabilities and Integration 3
1.3 Outline of Book 7
1.4 Topics Not Covered 8
1.5 A Couple of Notes 11
Part I Behavioral Models
2 Properties of Discrete Choice Models 15
2.1 Overview 15
2.2 The Choice Set 15
2.3 Derivation of Choice Probabilities 18
2.4 Specific Models 21
2.5 Identification of Choice Models 23
2.6 Aggregation 33
2.7 Forecasting 36
2.8 Recalibration of Constants 37
3 Logit 38
3.1 Choice Probabilities 38
3.2 The Scale Parameter 44
3.3 Power and Limitations of Logit 46
3.4 Nonlinear Representative Utility 56
3.5 Consumer Surplus 59
3.6 Derivatives and Elasticities 61
3.7 Estimation 64
3.8 Goodness of Fit and Hypothesis Testing 71
3.9 Case Study: Forecasting for a New
Transit System 75
3.10 Derivation of Logit Probabilities 78

v
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-FMA CB495/Train September 18, 2002 10:54 Char Count= 0

vi Contents

4 GEV 80
4.1 Introduction 80
4.2 Nested Logit 81
4.3 Three-Level Nested Logit 90
4.4 Overlapping Nests 93
4.5 Heteroskedastic Logit 96
4.6 The GEV Family 97
5 Probit 101
5.1 Choice Probabilities 101
5.2 Identification 104
5.3 Taste Variation 110
5.4 Substitution Patterns and Failure of IIA 112
5.5 Panel Data 114
5.6 Simulation of the Choice Probabilities 118
6 Mixed Logit 138
6.1 Choice Probabilities 138
6.2 Random Coefficients 141
6.3 Error Components 143
6.4 Substitution Patterns 145
6.5 Approximation to Any Random Utility Model 145
6.6 Simulation 148
6.7 Panel Data 149
6.8 Case Study 151
7 Variations on a Theme 155
7.1 Introduction 155
7.2 Stated-Preference and Revealed-Preference Data 156
7.3 Ranked Data 160
7.4 Ordered Responses 163
7.5 Contingent Valuation 168
7.6 Mixed Models 170
7.7 Dynamic Optimization 173
Part II Estimation
8 Numerical Maximization 189
8.1 Motivation 189
8.2 Notation 189
8.3 Algorithms 191
8.4 Convergence Criterion 202
8.5 Local versus Global Maximum 203
8.6 Variance of the Estimates 204
8.7 Information Identity 205
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-FMA CB495/Train September 18, 2002 10:54 Char Count= 0

Contents vii

9 Drawing from Densities 208

9.1 Introduction 208
9.2 Random Draws 208
9.3 Variance Reduction 217
10 Simulation-Assisted Estimation 240
10.1 Motivation 240
10.2 Definition of Estimators 241
10.3 The Central Limit Theorem 248
10.4 Properties of Traditional Estimators 250
10.5 Properties of Simulation-Based Estimators 253
10.6 Numerical Solution 260
11 Individual-Level Parameters 262
11.1 Introduction 262
11.2 Derivation of Conditional Distribution 265
11.3 Implications of Estimation of θ 267
11.4 Monte Carlo Illustration 270
11.5 Average Conditional Distribution 272
11.6 Case Study: Choice of Energy Supplier 273
11.7 Discussion 283
12 Bayesian Procedures 285
12.1 Introduction 285
12.2 Overview of Bayesian Concepts 287
12.3 Simulation of the Posterior Mean 294
12.4 Drawing from the Posterior 296
12.5 Posteriors for the Mean and Variance
of a Normal Distribution 297
12.6 Hierarchical Bayes for Mixed Logit 302
12.7 Case Study: Choice of Energy Supplier 308
12.8 Bayesian Procedures for Probit Models 316

Bibliography 319
Index 331
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

1 Introduction

1.1 Motivation
When I wrote my first book, Qualitative Choice Analysis, in the mid-
1980s, the field had reached a critical juncture. The breakthrough con-
cepts that defined the field had been made. The basic models – mainly
logit and nested logit – had been introduced, and the statistical and eco-
nomic properties of these models had been derived. Applications had
proven successful in many different areas, including transportation, en-
ergy, housing, and marketing – to name only a few.
The field is at a similar juncture today for a new generation of proce-
dures. The first-generation models contained important limitations that
inhibited their applicability and realism. These limitations were well
recognized at the time, but ways to overcome them had not yet been
discovered. Over the past twenty years, tremendous progress has been
made, leading to what can only be called a sea change in the approach
and methods of choice analysis. The early models have now been sup-
plemented by a variety of more powerful and more flexible methods.
The new concepts have arisen gradually, with researchers building on
the work of others. However, in a sense, the change has been more like
a quantum leap than a gradual progression. The way that researchers
think about, specify, and estimate their models has changed. Importantly,
a kind of consensus, or understanding, seems to have emerged about the
new methodology. Among researchers working in the field, a definite
sense of purpose and progress prevails.
My purpose in writing this new book is to bring these ideas together,
in a form that exemplifies the unity of approach that I feel has emerged,
and in a format that makes the methods accessible to a wide audience.
The advances have mostly centered on simulation. Essentially, simu-
lation is the researcher’s response to the inability of computers to per-
form integration. Stated more precisely, simulation provides a numerical

1
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

2 Introduction

approximation to integrals, with different methods offering different

properties and being applicable to different kinds of integrands.
Simulation allows estimation of otherwise intractable models. Prac-
tically any model can be estimated by some form of simulation.
The researcher is therefore freed from previous constraints on model
specification – constraints that reflected mathematical convenience
rather than the economic reality of the situation. This new flexibility
is a tremendous boon to research. It allows more realistic representation
of the hugely varied choice situations that arise in the world. It enables
the researcher to obtain more information from any given dataset and, in
many cases, allows previously unapproachable issues to be addressed.
This flexibility places a new burden on the researcher. First, the meth-
ods themselves are more complicated than earlier ones and utilize many
concepts and procedures that are not covered in standard econometrics
courses. Understanding the various techniques – their advantages and
limitations, and the relations among them – is important when choosing
the appropriate method in any particular application and for developing
new methods when none of the existing models seems right. The purpose
of this book is to assist readers along this path.
Second, to implement a new method or a variant on an old method,
the researcher needs to be able to program the procedure into computer
software. This means that the researcher will often need to know how
maximum likelihood and other estimation methods work from a compu-
tational perspective, how to code specific models, and how to take exist-
ing code and change it to represent variations in behavior. Some models,
such as mixed logit and pure probit (in addition, of course, to standard
logit), are available in commercially available statistical packages. In
fact, code for these and other models, as well as manuals and sample
data, are available (free) at my website http://elsa.berkeley.edu/∼train.
Whenever appropriate, researchers should use available codes rather
than writing their own. However, the true value of the new approach to
choice modeling is the ability to create tailor-made models. The com-
putational and programming steps that are needed to implement a new
model are usually not difficult. An important goal of the book is to
teach these skills as an integral part of the exposition of the models
themselves. I personally find programming to be extremely valuable
pedagogically. The process of coding a model helps me to understand
how exactly the model operates, the reasons and implications of its
structure, what features constitute the essential elements that cannot be
changed while maintaining the basic approach, and what features are
arbitrary and can easily be changed. I imagine other people learn this
way too.
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

Introduction 3

1.2 Choice Probabilities and Integration

To focus ideas, I will now establish the conceptual basis for discrete
choice models and show where integration comes into play. An agent
(i.e., person, firm, decision maker) faces a choice, or a series of choices
over time, among a set of options. For example, a customer chooses
which of several competing products to buy; a firm decides which
technology to use in production; a student chooses which answer to
give on a multiple-choice test; a survey respondent chooses an integer
between 1 and 5 on a Likert-scale question; a worker chooses whether
to continue working each year or retire. Denote the outcome of the de-
cision(s) in any given situation as y, indicating the chosen option or
sequence of options. We assume for the purposes of this book that the
outcome variable is discrete in that it takes a countable number of values.
Many of the concepts that we describe are easily transferable to situa-
tions where the outcome variable is continuous. However, notation and
terminology are different with continuous outcome variables than with
discrete ones. Also, discrete choices generally reveal less information
about the choice process than continuous-outcome choices, so that the
econometrics of discrete choice is usually more challenging.
Our goal is to understand the behavioral process that leads to the
agent’s choice. We take a causal perspective. There are factors that col-
lectively determine, or cause, the agent’s choice. Some of these factors
are observed by the researcher and some are not. The observed factors
are labeled x, and the unobserved factors ε. The factors relate to the
agent’s choice through a function y = h(x, ε). This function is called
the behavioral process. It is deterministic in the sense that given x and
ε, the choice of the agent is fully determined.
Since ε is not observed, the agent’s choice is not deterministic and
cannot be predicted exactly. Instead, the probability of any particular
outcome is derived. The unobserved terms are considered random with
density f (ε). The probability that the agent chooses a particular outcome
from the set of all possible outcomes is simply the probability that the
unobserved factors are such that the behavioral process results in that
outcome: P(y | x) = Prob(ε s.t. h(x, ε) = y).
We can express this probability in a more usable form. Define an
indicator function I [h(x, ε) = y] that takes the value of 1 when the
statement in brackets is true and 0 when the statement is false. That
is, I [·] = 1 if the value of ε, combined with x, induces the agent to
choose outcome y, and I [·] = 0 if the value of ε, combined with x,
induces the agent to choose some other outcome. Then the probability
that the agent chooses outcome y is simply the expected value of this
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

4 Introduction

indicator function, where the expectation is over all possible values of

the unobserved factors:

P(y | x) = Prob(I [h(x, ε) = y] = 1)


(1.1) = I [h(x, ε) = y] f (ε) dε.

Stated in this form, the probability is an integral – specifically an integral

of an indicator for the outcome of the behavioral process over all possible
values of the unobserved factors.
To calculate this probability, the integral must be evaluated. There are
three possibilities.

1.2.1. Complete Closed-Form Expression

For certain specifications of h and f , the integral can be ex-
pressed in closed form. In these cases, the choice probability can be
calculated exactly from the closed-form formula. For example, consider
a binary logit model of whether or not a person takes a given action, such
as buying a new product. The behavioral model is specified as follows.
The person would obtain some net benefit, or utility, from taking the
action. This utility, which can be either positive or negative, consists of
a part that is observed by the researcher, β  x, where x is a vector of
variables and β is a vector of parameters, and a part that is not observed,
ε: U = β  x + ε. The person takes the action only if the utility is positive,
that is, only if doing so provides a net benefit. The probability that the per-
son takes
 the action, given what the researcher can observe, is therefore

P = I [β x + ε > 0] f (ε) dε, where f is the density of ε. Assume that
ε is distributed logistically, such that its density is f (ε) = e−ε /(1 + e−ε )2
with cumulative distribution F(ε) = 1/(1 + e−ε ). Then the probability
of the person taking the action is

P = I [β  x + ε > 0] f (ε) dε

= I [ε > −β  x] f (ε) dε
 ∞
= f (ε) dε
ε=−β  x
1
= 1 − F(−β  x) = 1 −
1 + eβ  x

eβ x
= .
1 + eβ  x
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

Introduction 5

For any x, the probability can be calculated exactly as P = exp(β  x)/

(1 + exp(β  x)).
Other models also have closed-form expressions for the probabilities.
Multinomial logit (in Chapter 3), nested logit (Chapter 4), and ordered
logit (Chapter 7) are prominent examples. The methods that I described
in my first book and that served as the basis for the first wave of interest in
discrete choice analysis relied almost exclusively on models with closed-
form expressions for the choice probabilities. In general, however, the
integral for probabilities cannot be expressed in closed form. More to
the point, restrictions must be placed on the behavioral model h and
the distribution of random terms f in order for the integral to take
a closed form. These restrictions can make the models unrealistic for
many situations.

1.2.2. Complete Simulation

Rather than solve the integral analytically, it can be approxi-
mated through simulation. Simulation is applicable in one form or an-
other to practically any specification of h and f . Simulation relies on the
fact that integration
 over a density is a form of averaging. Consider the
integral t = t(ε) f (ε) dε, where t(ε) is a statistic based on ε which has
¯
density f (ε). This integral is the expected value of t over all possible
values of ε. This average can be approximated in an intuitively straight-
forward way. Take numerous draws of ε from its distribution f , calculate
t(ε) for each draw, and average the results. This simulated average is an
unbiased estimate of the true average. It approaches the true average as
more and more draws are used in the simulation.
This concept of simulating an average is the basis for all simulation
methods, at least all of those that we consider in this book. As given in
equation (1.1), the probability of a particular outcome is an average of
the indicator I (·) over all possible values of ε. The probability, when
expressed in this form, can be simulated directly as follows:
1. Take a draw of ε from f (ε). Label this draw ε 1 , where the
superscript denotes that it is the first draw.
2. Determine whether h(x, ε 1 ) = y with this value of ε. If so, create
I 1 = 1; otherwise set I 1 = 0.
3. Repeat steps 1 and 2 many times, for a total of R draws. The
indicator for each draw is labeled I r for r = 1, . . . , R.
4. Calculate the average of the I r ’s. This average is the simulated
probability: P̌(y | x) = R1 rR=1 I r . It is the proportion of times
that the draws of the unobserved factors, when combined with
the observed variables x, result in outcome y.
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

6 Introduction

As we will see in the chapters to follow, this simulator, while easy to

understand, has some unfortunate properties. Choice probabilities can
often be expressed as averages of other statistics, rather than the average
of an indicator function. The simulators based on these other statistics
are calculated analogously, by taking draws from the density, calculating
the statistic, and averaging the results. Probit (in Chapter 5) is the most
prominent example of a model estimated by complete simulation. Vari-
ous methods of simulating the probit probabilities have been developed
based on averages of various statistics over various (related) densities.

1.2.3. Partial Simulation, Partial Closed Form

So far we have provided two polar extremes: either solve the
integral analytically or simulate it. In many situations, it is possible to
do some of both.
Suppose the random terms can be decomposed into two parts labeled
ε1 and ε2 . Let the joint density of ε1 and ε2 be f (ε) = f (ε1 , ε2 ). The
joint density can be expressed as the product of a marginal and a condi-
tional density: f (ε1 , ε2 ) = f (ε2 | ε1 ) · f (ε1 ). With this decomposition,
the probability in equation (1.1) can be expressed as

P(y | x) = I [h(x, ε) = y] f (ε) dε
  
= I [h(x, ε1 , ε2 ) = y] f (ε2 | ε1 ) dε2 f (ε1 ) dε1 .
ε1 ε2

Now suppose that a closedform exists for the integral in large brackets.
Label this formula g(ε1 ) ≡ ε2 I [h(x, ε1 , ε2 ) = y] f (ε2 | ε1 ) dε2 , which
 on the value of ε1 . The probability then becomes
is conditional
P(y | x) = ε1 g(ε1 ) f (ε1 ) dε1 . If a closed-form solution does not ex-
ist for this integral, then it is approximated through simulation. Note
that it is simply the average of g over the marginal density of ε1 . The
probability is simulated by taking draws from f (ε1 ), calculating g(ε1 )
for each draw, and averaging the results.
This procedure is called convenient error partitioning (Train, 1995).
The integral over ε2 given ε1 is calculated exactly, while the integral over
ε1 is simulated. There are clear advantages to this approach over com-
plete simulation. Analytic integrals are both more accurate and easier to
calculate than simulated integrals. It is useful, therefore, when possible,
to decompose the random terms so that some of them can be integrated
analytically, even if the rest must be simulated. Mixed logit (in Chap-
ter 6) is a prominent example of a model that uses this decomposition
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

Introduction 7

effectively. Other examples include Gourieroux and Monfort’s (1993)

binary probit model on panel data and Bhat’s (1999) analysis of ordered
responses.

1.3 Outline of Book

Discrete choice analysis consists of two interrelated tasks: specification
of the behavioral model and estimation of the parameters of that model.
Simulation plays a part in both tasks. Simulation allows the researcher to
approximate the choice probabilities that arise in the behavioral model.
As we have stated, the ability to use simulation frees the researcher
to specify models without the constraint that the resulting probabilities
must have a closed form. Simulation also enters the estimation task.
The properties of an estimator, such as maximum likelihood, can change
when simulated probabilities are used instead of the actual probabilities.
Understanding these changes, and mitigating any ill effects, is important
for a researcher. In some cases, such as with Bayesian procedures, the
estimator itself is an integral over a density (as opposed to the choice
probability being an integral). Simulation allows these estimators to be
implemented even when the integral that defines the estimator does not
take a closed form.
The book is organized around these two tasks. Part I describes be-
havioral models that have been proposed to describe the choice process.
The chapters in this section move from the simplest model, logit, to
progressively more general and consequently more complex models. A
chapter is devoted to each of the following: logit, the family of gener-
alized extreme value models (whose most prominent member is nested
logit), probit, and mixed logit. This part of the book ends with a chapter
titled “Variations on a Theme,” which covers a variety of models that
build upon the concepts in the previous chapters. The point of this chap-
ter is more than simply to introduce various new models. The chapter
illustrates the underlying concept of the book, namely, that researchers
need not rely on the few common specifications that have been pro-
grammed into software but can design models that reflect the unique
setting, data, and goals of their project, writing their own software and
using simulation as needed.
Part II describes estimation of the behavioral models. Numerical max-
imization is covered first, since most estimation procedures involve
maximization of some function, such as the log-likelihood function.
We then describe procedures for taking draws from various kinds of
densities, which are the basis for simulation. This chapter also describes
different kinds of draws, including antithetic variants and quasi-random
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

8 Introduction

sequences, that can provide greater simulation accuracy than indepen-

dent random draws. We then turn to simulation-assisted estimation, look-
ing first at classical procedures, including maximum simulated likeli-
hood, method of simulated moments, and method of simulated scores.
Finally, we examine Bayesian estimation procedures, which use simula-
tion to approximate moments of the posterior distribution. The Bayesian
estimator can be interpreted from either a Bayesian or classical perspec-
tive and has the advantage of avoiding some of the numerical difficulties
associated with classical estimators. The power that simulation provides
when coupled with Bayesian procedures makes this chapter a fitting
finale for the book.

1.4 Topics Not Covered

I feel it is useful to say a few words about what the book does not cover.
There are several topics that could logically be included but are not.
One is the branch of empirical industrial organization that involves esti-
mation of discrete choice models of consumer demand on market-level
data. Customer-level demand is specified by a discrete choice model,
such as logit or mixed logit. This formula for customer-level demand is
aggregated over consumers to obtain market-level demand functions that
relate prices to shares. Market equilibrium prices are determined as the
interaction of these demand functions with supply, based on marginal
costs and the game that the firms are assumed to play. Berry (1994)
and Berry et al. (1995) developed methods for estimating the demand
parameters when the customer-level model takes a flexible form such as
mixed logit. The procedure has been implemented in numerous markets
for differentiated goods, such as ready-to-eat cereals (Nevo, 2001).
I have decided not to cover these procedures, despite their importance
because doing so would involve introducing the literature on market-
level models, which we are not otherwise considering in this book. For
market demand, price is typically endogenous, determined by the in-
teraction of demand and supply. The methods cited previously were
developed to deal with this endogeneity, which is probably the central
issue with market-level demand models. This issue does not automati-
cally arise in customer-level models. Prices are not endogenous in the
traditional sense, since the demand of the customer does not usually
affect market price. Covering the topic is therefore not necessary for our
analysis of customers’ choices.
It is important to note, however, that various forms of endogeneity
can indeed arise in customer-level models, even if the traditional type of
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

Introduction 9

endogeneity does not. For example, suppose a desirable attribute of prod-

ucts is omitted from the analysis, perhaps because no measure of it exists.
Price can be expected to be higher for products that have high levels of
this attribute. Price therefore becomes correlated with the unobserved
components of demand, even at the customer level: the unobserved part
of demand is high (due to a high level of the omitted attribute) when
the price is high. Estimation without regard to this correlation is incon-
sistent. The procedures cited above can be applied to customer-level
models to correct for this type of endogeneity, even though they were
originally developed for market-level data. For researchers who are con-
cerned about the possibility of endogeneity in customer-level models,
Petrin and Train (2002) provide a useful discussion and application of
the methods.
A second area that this book does not cover is discrete–continuous
models. These models arise when a regression equation for a continuous
variable is related in any of several ways to a discrete choice. The most
prominent situations are the following.

1. The continuous variable depends on a discrete explanatory

variable that is determined endogenously with the dependent
variable. For example, consider an analysis of the impact of job-
training programs on wages. A regression equation is specified
with wages as the dependent variable and a dummy variable for
whether the person participated in a job-training program. The
coefficient of the participation dummy indicates the impact of
the program on wages. The situation is complicated, however, by
the fact that participation is voluntary: people choose whether to
participate in job-training programs. The decision to participate
is at least partially determined by factors that also affect the per-
son’s wage, such as the innate drive, or “go-for-it” attitude, of the
person. Estimation of the regression by ordinary least squares is
biased in this situation, since the program-participation dummy
is correlated with the errors in the wage equation.
2. A regression equation is estimated on a sample of observations
that are selected on the basis of a discrete choice that is de-
termined endogenously with the dependent variable. For exam-
ple, a researcher might want to estimate the effect of weather
on peak energy load (that is, consumption during the highest-
demand hour of the day). Data on energy loads by time of day are
available only for households that have chosen time-of-use rates.
However, the households’ choice of rate plan can be expected
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

10 Introduction

to be related to their energy consumption, with customers who

have high peak loads tending not to choose time-of-use rates,
since those rates charge high prices in the peak. Estimation of
the regression equation on this self-selected sample is biased
unless the endogeneity of the sample is allowed for.
3. The continuous dependent variable is truncated. For example,
consumption of goods by households is necessarily positive.
Stated statistically, consumption is truncated below at zero, and
for many goods (such as opera tickets) observed consumption
is at this truncation point for a large share of the population.
Estimation of the regression without regard to the truncation
can cause bias.

The initial concepts regarding appropriate treatment of discrete–

continuous models were developed by Heckman (1978, 1979) and Dubin
and McFadden (1984). These early concepts are covered in my earlier
book (Train, 1986, Chapter 5). Since then, the field has expanded tremen-
dously. An adequate discussion of the issues and procedures would take
a book in itself. Moreover, the field has not reached (at least in my view)
the same type of juncture that discrete choice modeling has reached.
Many fundamental concepts are still being hotly debated, and poten-
tially valuable new procedures have been introduced so recently that
there has not been an opportunity for researchers to test them in a vari-
ety of settings. The field is still expanding more than it is coalescing.
There are several ongoing directions of research in this area. The
early procedures were highly dependent on distributional assumptions
that are hard to verify. Researchers have been developing semi- and
nonparametric procedures that are hopefully more robust. The special
1986 issue of the Journal of Econometrics provides a set of important
articles on the topic. Papers by Lewbel and Linton (2002) and Levy
(2001) describe more recent developments. Another important devel-
opment concerns the representation of behavior in these settings. The
relation between the discrete and continuous variables has been gen-
eralized beyond the fairly simple representation that the early methods
assumed. For example, in the context of job training, it is likely that
the impact of the training differs over people and that people choose to
participate in the training program on the basis of the impact it will have
on them. Stated in econometric terms: the coefficient of the participation
dummy in the wage equation varies over people and affects the value of
the dummy. The dummy is correlated with its own coefficient, as well
as with the unobserved variables that enter the error of the regression.
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

Introduction 11

A recent discussion of approaches to this issue is provided by Carneiro

et al. (2001).

1.5 A Couple of Notes

Throughout the book, I refer to the researcher as “she” and the decision
maker as “he.” This usage, as well as being comparatively gender-neutral
(or at least symmetrically noninclusive), allows both people to be re-
ferred to in the same paragraph without confusion.
Many colleagues have provided valuable comments and sugges-
tions on earlier drafts of the book. I am very grateful for this help. I
thank Greg Allenby, Moshe Ben-Akiva, Chandra Bhat, Denis Bolduc,
David Brownstone, Siddhartha Chib, Jon Eisen-Hecht, Florian Heiss,
David Hensher, Joe Herriges, Rich Johnson, Frank Koppelman, Jordan
Louviere, Aviv Nevo, Juan de Dios Ortúzar, Ken Small, Joan Walker,
Cliff Winston, Joachim Winter, and the students in my graduate econo-
metrics course.
I welcome readers to contact me if you feel I have not covered material
that you consider important, or if I have confused rather than clarified
any of the material that I do cover. Hopefully, another edition of this
book will someday materialize.
P1: GEM/IKJ P2: GEM/IKJ QC: GEM/ABE T1: GEM
CB495-01Drv CB495/Train KEY BOARDED September 18, 2002 11:2 Char Count= 0

12