Journal of Business Research 51 (2001) 115 ± 125

Random utility models in marketing research: a survey

George Baltasa,*, Peter Doyleb
a
Athens University of Economics and Business, Athens, Greece
b
Warwick Business School, University of Warwick, Coventry CV4 7AL, England, UK
Accepted 1 February 2000

Abstract

Random utility (RU) models are well-established methods for describing discrete choice behavior. Recently, there has been a strong
upsurge in interest driven by advances in data gathering and estimation technology. This review paper describes the principles and issues, and
develops a taxonomy of three major families of models. The paper summarizes and classifies the different approaches. The advantages and
limitations of the various alternatives are outlined. Practical issues in implementing the models are also discussed. D 2000 Elsevier Science
Inc. All rights reserved.

Keywords: Random utility models; Marketing research; Discrete choice behavior

1. Introduction 2. Overview

With an ever increasing importance of market intelli- Consider an individual agent choosing a single option
gence, the need to understand advanced methods of market among a finite set of alternatives, for example, a consumer
research has never been greater. Random utility (RU) deciding which brand to buy. This is the realm of behavior
models have been developed to describe choice among that is considered in RU modeling. In RU models, prefer-
mutually exclusive discrete alternatives and received con- ences for such discrete alternatives are determined by the
siderable academic and industry attention. This paper sur- realization of latent indices of attractiveness, called product
veys RU models, discusses key issues and develops a utilities. Utility maximization is the objective of the decision
structurally meaningful synthesis of the different formula- process and leads to observed choice in the sense that the
tions. It is intended to help managers and researchers keep consumer chooses the alternative for which utility is max-
informed of a fast-changing and important area which may imal. Individual preferences depend on characteristics of the
not directly fall within their own specific professional or alternatives and the tastes of the consumer. An RU model
research interests. The paper is organized as follows. The defines a mapping from observed characteristics into pre-
subsequent section illustrates the underlying principles of ferences. The analyst however cannot observe all the factors
RU models and explains how probabilistic choice flows affecting preferences and the latter are treated as random
from utility maximization. The third section discusses the variables. By its abstraction from various idiosyncratic
behavior of RU models and issues that arise from restrictive factors, the model uses stochastic assumptions to describe
stochastic assumptions. The fourth section introduces a unmeasured variation in preferences. An operational way to
taxonomy and describes the different models. The fifth allow for maximization of latent preferences is to consider a
section is concerned with implementation and experimental utility function that is decomposable into two additively
data. The sixth section discusses some practical issues and separable parts, (1) a deterministic component specified as a
the last section concludes. function of measured attributes of the alternatives and/or the
individual, and (2) a stochastic component representing
unobserved attributes affecting choice, interindividual dif-
ferences in utilities depending upon the heterogeneity in
* Corresponding author. tastes, measurement errors, and functional misspecification
E-mail address: [email protected] (G. Baltas). (Manski, 1977). In the next sections, we shall consider

0148-2963/00/$ ± see front matter D 2000 Elsevier Science Inc. All rights reserved.
PII: S 0 1 4 8 - 2 9 6 3 ( 9 9 ) 0 0 0 5 8 - 2
116 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125

different models based on alternative hypotheses about the expression in Eq. (4) can be derived in a great number of
``unknown.'' Proceeding further in the same vein, let ways (McFadden, 1973; Train, 1986; Anderson et al.,
1992). Having laid out the necessary background, we turn to
Uij ˆ Vij ‡ eij …1†
the stochastic assumptions of the models.
be the utility of alternative j for consumer i, where Vij is the
deterministic component and eij the random component.
Typically, the deterministic component Vij has been assumed 3. Stochastic assumptions of RU models
to have an additively separable linear form Vij = xijTB, where
xij and B are the vectors of exogenous variables and Suppose we observe members of a population of con-
parameters, respectively. In the hypothetical case that V sumers, each member i of which has a utility function Uij =
contains perfect information about the determinants of Vij + eij for each product j of a set C = [1,2,3, . . . ,M]. Vij is
utility, the consumer would simply choose the product with the non-stochastic function mapping attributes into utility
the highest Vij. The stochastic terms eij shaping the true and and eij accounts for factors not included in Vij.
latent utility in Eq. (1), introduce uncertainty regarding the The simple MNL model accounts for unobserved deter-
choice and therefore, choice probabilities are invoked to minants of choice by IID random terms. That is they are
describe choice behavior. The probabilistic description of assumed to have the same distribution, with the same mean
choice has been introduced not to reflect behavior that is and variance and also to be uncorrelated across and within
probabilistic. Rather, it is the lack of information that leads individuals. An interesting property is the effect of increas-
the analyst to treat utility as a random variable and ing unexplained stochastic variation on the identified coef-
consequently to describe choice in a probabilistic fashion. ficients. Since the variance (assumed the same for all ej) is
In fact, the properties of RU models can be attributed to the related with the parameter m, it is obvious in Eq. (4) that the
specific assumptions that each model implies about the variance discounts the value of the estimated parameters in
stochastic terms. Under the utility maximization rule, a the non-stochastic function V. Since the variables in V are
consumer facing a set of available products C = {1,2,3, . . . , exogenous, the estimated coefficients absorb the variance
M} will choose a product j with probability P( j) = P(Uj > effect. Intuitively, high variance implies limited ability of
Uk) for all k 2 C, k 6ˆ j, or as it follows from Eq. (1): the observed variables to explain choices and therefore leads
to smaller values of the coefficients. Since m is a transfor-
P… j† ˆ P…ek < Vj ÿ Vk ‡ ej † for all k 2 C; k 6ˆ j: …2†
mation of the variation of the random disturbances, it can be
The probability that j is chosen is then obtained by making seen as an index of unobserved variation in preferences that
assumptions about the form of the distribution of the cannot be explained by the variables in the non-stochastic
random variables and integrating Eq. (2) over a continuum function V. In the MNL, the price coefficient reflects the
of all possible values for ej. From Eq. (2), we can write the response of choice probabilities to prices and its magnitude
selection probability for, say, the first alternative as is related to the variance parameter m. As unobserved
Z Z Z variation decreases, the value of the identified price para-
1 V1 ÿV2 ‡e1 V1 ÿV3 ‡e1
meter in V increases and vice versa. Therefore, the identified
P…1† ˆ ...
ÿ1 ÿ1 ÿ1 price coefficient can be regarded as an index of average
Z V1 ÿVM ‡e1 substitutability among alternatives that is related to stochas-
 f …e1 ; e2 ; e3 ; . . . eM †deM . . . de3 de2 de1 : tic variation. Although the MNL accommodates varying
ÿ1
…3† rates of symmetric substitution, the assumption of IID
random components remains restrictive and imposes the
In words, Eq. (3) states that the choice probability of independence of irrelevant alternatives (IIA) property (Ben-
alternative 1 is the probability of over all possible values of Akiva and Lerman, 1985). Under this structural restriction,
e1, all the other random terms being less than V1 ÿ Vj + e1, the odds of the consumer choosing j over k remain the same
8j 2 C. The popular multinomial logit (MNL) model is regardless of the composition of the choice set. An analo-
derived by assuming that the random terms are indepen- gous and possibly more important drawback is that the
dently identically distributed (IID) according to the double model cannot postulate any pattern of differential substitut-
exponential distribution with mode zero and variance m2p2/ ability between products. An improvement in an alternati-
6, where m is a positive scale parameter. The choice ve's systematic utility will have a proportionally equal
probability in Eq. (3) then takes the compact form, impact on the selection probabilities of all other alternatives.
  Thus, an implication of the IIA property is that the cross-
V
exp mj elasticity of the probability of brand j with respect to a
P… j† ˆ P  : …4† change in Vk is the same for all j with j 6ˆ k.
exp Vmk The assumption of independent preferences is restrictive.
k2C
In reality, alternatives may not be equally dissimilar. Differ-
The analytic form of the MNL probabilities has greatly ential similarities among products due to shared character-
contributed to the popularity of the MNL model. The istics lead to correlated utilities. When these conditions
 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125 117

apply, the random utility components are not independent. 1997a). Similarly, state dependence refers to effects of prior
An important implication is that the rates of substitution are experience in a clear causal sense while inertia describes
not uniform since cross-alternative substitutability is related short-run persistency. In conclusion, utilities may be corre-
to perceived similarity (McFadden, 1981). lated over time within people because they are evaluated by
Similarly, the assumption of identically distributed uti- the same tastes (heterogeneity), because they are determined
lities may not be pragmatic. Unmeasured factors may have by past experiences (state dependence), or because they
unequal variances across alternatives. For instance, the depend on past utilities (inertia). We shall consider inter-
suitability on consumption occasions may vary more for personal heterogeneity and intrapersonal dynamics as two
brand A than for brand B. Similarly, interpersonal differ- related forms of taste variation, i.e. variation over people
ences in preferences may not be uniform for all brands, that and variation over time.
is some brands have utilities that vary more over people. To summarize: (a) unobserved product attributes may
Where these conditions apply, the non-stochastic function V lead to non-IID (over products) random utilities, (b) un-
does not explain equally well choices of different alterna- observed taste heterogeneity may lead to non-IID (over
tives (Baltas and Doyle, 1998). products) random utilities across individuals and also to
When an RU model is used as a model of population non-IID (over time) random utilities within individuals, (c)
behavior, another issue may arise because of parameter intrapersonal dynamics may lead to non-IID (over time)
heterogeneity across consumers. From a formal point of random utilities within individuals.
view, the stochastic utility of alternative j for individual i These distinctions are instructive. For example, a homo-
at time t, eijt, embodies both interindividual and intraindi- geneous probit model accounts for unobserved product
vidual unmeasured variation in preferences. The interindi- attributes but does not deal explicitly with interindividual
vidual variation may include unobserved factors, such as differences. Analogously, as we discuss, subsequently, a
intrinsic brand preference and deviations from average random-coefficients MNL model account for interindividual
population sensitivity to marketing variables, which are differences but does not deal explicitly with unobserved
persistent over time for each individual. The intraindivi- product attributes. Although both are free from the IIA
dual variation may include unobserved factors, such as property when used as models of population behavior, their
different states of mind, different consumption occasions structural flexibility has different origins. For the first, it
and dynamic taste formation, which vary over time for arises out of an approximation of individual decision that
each individual. accounts for unobserved product attributes. For the second,
To appreciate these ideas consider that in the specifica- it flows from a description of population behavior that
tion Uijt = xijt B + eijt, the parameter vector describes mean accounts for aggregation over heterogeneous utility func-
population tastes, Ebt = B, where bi may vary over indivi- tions. We now turn to describe and integrate alternative
duals. The unobserved random component can be expressed formulations that lie within the class of RU models.
as eijt = xijt (B ÿ bi) + xijt where xijt is a random term that is
assumed IID over alternatives and time. Therefore, the
unobserved component eijt is correlated over j because of 4. A taxonomy of RU models
the presence of the individual specific taste deviation (B ÿ
bi) (e.g. Train, 1998). We propose a characterization of the models along the
Unobserved heterogeneity may create correlation not following lines. As Fig. 1 illustrates, the major branches
only over alternatives but also over time. Since the realiza- represent the three major issues: (a) unobserved heteroge-
tion of the random variable Uij over successive occasions t neity in alternatives, (b) taste variation, and (c) heteroge-
depends on the invariable component (B ÿ bi), the utilities neous choice sets.
attached to a product are serially correlated. Unobserved The first branch describes alternative treatments of un-
heterogeneity would appear as serial correlation in the observed product attributes. First, we distinguish between
random terms (Keane, 1997b). models that assume IIA and non-IIA unobserved product
Heterogeneity however is not the only reason for serial heterogeneity. The first class includes only the MNL. In
correlation. Another possible explanation is the presence of theory, other models (e.g. a restricted probit with IID terms)
intrapersonal dynamics. State dependence and short-run can be cast as members of the same class. In practice, only
preference inertia can also lead to serially correlated utilities. the MNL has been used. The second class allows utilities
As the name suggests, state dependence is the influence of that are non-IID over products and account for unobserved
past outcomes (states) on present choices. On the other product heterogeneity that violates the IIA property. Con-
hand, preference inertia or habit-persistence (Heckman, sequently, the models can postulate more flexible patterns of
1981a; Roy et al., 1996) is the dependence of current substitution. McFadden's (1979) generalized extreme value
utilities on past utilities. A main behavioral distinction (GEV) model, Hausman and Wise's (1978) multinomial
between heterogeneity and inertia is that heterogeneity probit (MNP) model, and the heteroscedastic extreme value
refers to time-invariant patterns while inertia describes (HEV) model introduced by Daganzo (1979) can be re-
behavior that tends to be similar in the short-run (Keane, garded as members of this class.
118 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125

Fig. 1. A taxonomy of random utility models.

The second branch deals with variation in tastes. The first important for applied work. The transparent and elegant
class accounts for taste variation over people, that is, probability formula allows a clear interpretation in terms
heterogeneous preferences and responses to product char- of the relative utilities. In addition, in practice, estimation
acteristics and marketing variables. Several different ap- and forecasting are easy. The share of the new product
proaches have been used to account for problems of will equal the proportional decreases in the share of all
aggregation over heterogeneous utility functions. One may existing products and the response of demand to policy
include individual background variables in the utility func- variables is given by compact own and cross elasticities.
tion, assign consumers to segments, or allow parameters to The MNL can also be estimated on a subset of alter-
vary over people either by fixed or random effects (FE) natives, when the IIA is sustained. However, as is often
models (e.g. Chintagunta et al., 1991). the case, the unmatched tractability of the MNL comes at
The second class considers taste variation over time. As a cost in flexibility.
mentioned earlier, in the presence of choice dynamics, In early marketing studies, the MNL model has been
individual tastes change over time in response to previous used to study a new product strategy (Hauser and Urban,
experiences. However, the determination of dynamic ef- 1977; Silk and Urban, 1978), choice of business schools
fects in RU models requires explicit control of heteroge- (Punj and Staelin, 1978), and store choice (Gensch and
neity (see, e.g. Roy et al., 1996). Distinguishing the effect Recker, 1979). Guadagni and Little (1983) used scanner
of heterogeneity from the impact of the past on current panel data to calibrate an MNL model. Since then,
tastes and consequently, choices is a central issue in scanner data have become the dominant source of ob-
empirical work. servations in applied RU analyses. The MNL model has
To this point, the marketing literature has followed or been used to study several other issues such as variety
extended ideas with econometric origins. The third family seeking (Lattin, 1987), reference-dependent behavior
however can be viewed as a departure from this tradi- (Winer, 1986; Lattin and Bucklin, 1989; Hardie et al.,
tion. This stream of research attempts to include choice 1993), promotion effects (Allenby and Rossi, 1991;
set effects in RU models, a concept that has been given Bronnenberg and Wathieu, 1996), advertising effects
much attention in marketing (see Shocker et al., 1991 for (Tellis 1988, Deigthon et al., 1994), brand equity (Swait
a review). The inclusion of heterogeneous choice sets et al., 1993) and price effects (Krishnamurthi and Raj,
lead to models that describe the idiosyncratic composi- 1988, 1991).
tion of the group of alternatives evaluated prior to The models included in the non-IIA class describe
choice. As will be seen below, this heterogeneity in the unobserved product heterogeneity that violates the IIA
availability of alternatives is created by a multistage property. In practical terms, these models relax the stochas-
process where individuals reduce their options before tic structure of the MNL model by accommodating prefer-
reaching a final choice. ences that are correlated, non-identical, or both across
The following sections discuss the structure and indivi- alternatives (Fig. 1).
dual models of the taxonomy. The standard GEV or nested logit (NMNL) model (Ben-
Akiva, 1973; McFadden, 1979) partitions the set of avail-
4.1. Unobserved product heterogeneity able alternatives into subsets of relatively homogeneous
alternatives and postulates that substitution is greater within
The first criterion in classifying RU models is how they than between subsets. This choice structure appears like a
deal with unobserved attributes of the choice alternatives. tree with branches defining choices among group of ele-
The most simple way is to assume the IIA property. As mental alternatives and twigs representing choices of alter-
noted above, this is the assumption of the MNL model. natives within branches. The corresponding transition
The MNL has played a historical role and it is still probabilities are MNL. The choice probabilities can be
 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125 119

expressed as products of respective transition probabilities variables across alternative travel modes. For example, the
or can be written compactly, utility of the train may have more variation than the utility
of the car since factors like comfort are less predictable by
" #…lq ÿ1†
P the commuter. In this manner, uncertainty comes not only
exp…Vjq =lq † exp…Vkq =lq † from the analyst's observational imperfections but also from
k2q
P… j; q† ˆ  lm …5† the decision maker's imperfect information. Baltas and
P
n P Doyle (1998) used the HEV to describe brand choice in a
exp…Vkm =lm †
mˆ1 k2m
frequently bought category. In this study, the model de-
scribes unobserved inter and intraindividual fluctuation in
where Vjq is the non-stochastic utility for alternative j in preferences that may be unequal across brands. For exam-
subset q and lq is an index of the inverse of utility correlation ple, preferences over consumers and suitability over con-
or alternative dissimilarity within subsets. The model allows sumption occasions may not vary equally for all competing
alternatives in the same subset to have correlated utilities as brands. The origin of the HEV is attributed to Daganzo
they share unobserved attributes. In addition, the NMNL (1979) who developed an analogous model.
allows the variance to differ across subsets. Although the The MNP model (Fig. 1) introduced by Bock and Jones
NMNL is often used to describe hierarchical choice, it does (1968) assumes a multivariate normal distribution and
not require a multistage decision process. The nested allows for the most general configuration of substitution
structure simply illustrates the correlation among the random across products via arbitrary covariance matrices (Hausman
utility components. The NMNL is popular in marketing (e.g. and Wise, 1978; Daganzo, 1979; Johnson and Hensher,
Dubin, 1986; Buckley, 1988; Kannan and Wright, 1991; 1982). The random utilities are allowed to be both correlated
Kamakura et al., 1996; Baltas et al., 1997). Among all the and unequal across alternatives, i.e. non-IID. Note that if we
non-IIA models, the NMNL is the most empirically tractable. set all covariance terms to zero, we get a model similar to
Estimation, forecasting, and evaluation of market response HEV with independent random utilities. Likewise, if we
(elasticities) are straightforward. The current approach further force all variance terms to be equal, we get a model
however is with limitations. While the IIA property does similar to MNL with both independent and identical random
not hold between groups of choices, it is still imposed within utilities. Similarly, if we restrict the covariance terms to be
each group. This can be regarded as a restriction of the model. equal within subsets of alternatives but zero between sub-
Also, the analyst has to specify a priori the partitioning of the sets, we get a structure similar to the NMNL model.
choice set in relatively homogeneous groups of alternatives. Unfortunately, despite intellectual appeal, the flexibility of
This however allows specific hypotheses to be tested and the MNP model comes at the cost of estimation problems.
prevents ad hoc rationalization of empirical results. More precisely, there is no closed-form expression for the
The MNL probabilities assume a common scale para- choice probabilities, which, in principle, require the calcula-
meter m for all random components ei implying equal tion of an (Mÿ1)-dimensional integral. Thus, the primary
variances across all choice alternatives. The HEV model difficulty in using the MNP has been the lack of practical,
(see Fig. 1) employs the double exponential distribution for accurate methods for approximating the choice probabilities
the random terms but allows them not to be identically when the number of alternatives is large (McFadden, 1981).
distributed. More specifically, the assumption of common However, recent advances in estimation technology and, in
variance is relaxed by allowing alternative-specific scale particular, in simulation-assisted inference (e.g. McFadden,
parameters mj. The HEV nests the MNL, i.e. the MNL is a 1989; Hajivassiliou et al., 1996) deal with the tractability
constrained form of the HEV (Bhat, 1995). Estimation of problem of the MNP. In practice, the applied researcher still
the HEV model yields estimates for the parameter vector of encounters problems such as parameter estimability and
the explanatory variables and also for the scale parameters computational resources in large problems. Future develop-
of the stochastic terms. The HEV accommodates differential ments may resolve these issues. Marketing applications
competitive effects by controlling the impact of changes in have employed simulated maximum likelihood methods
the deterministic components on choice probabilities via the (Kamakura and Srivastava, 1984, 1986; Papatla and Krish-
scale parameters. However, this model has received the least namurthi, 1992), McFadden's (1989) method of simulated
attention not only in marketing but also in the economic moments (Chintagunta, 1992a,b; Chintagunta and Honore,
literature. Further, alternative but not necessarily conflicting 1996; Keane 1997a) and Bayesian approaches (McCulloch
interpretations have been given. Allenby and Ginter (1995) and Rossi, 1994). A discussion of the MNP model in
used it in their research on consideration sets. Inspired by connection with recent developments particularly in simula-
the work of Hausman and Ruud (1987), this study assumes tion-based inference is given in Weeks (1997).
that for brands under consideration by the consumer, sys-
tematic factors have a greater role in determining utilities 4.2. Taste variation
and therefore, the latter would exhibit less stochastic varia-
tion. In transportation research, Bhat (1995) proposed the In the preceding section, we have been concerned with
HEV to account for different variances of unobserved models that are able to describe several patterns of non-
120 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125

IIA unobserved product heterogeneity. This section dis- In another related work, consumers are assigned to
cusses models that deal explicitly with two related forms segments by a deterministic (Currim, 1981; Gensch, 1985)
of taste variation Ð interpersonal heterogeneity and intra- or probabilistic method (Kamakura and Russell, 1989,
personal dynamics. Bucklin and Gupta, 1992; Kamakura and Russell, 1993;
We begin by considering the treatment of heterogeneity Gupta and Chintagunta, 1994; Kamakura et al., 1996; Mela
in RU models. From the modeler's standpoint, it makes et al., 1997). Here, the assumption is that taste parameters
sense to distinguish between two forms of heterogeneity: are constant within segments, which contain homogeneous
heterogeneous tastes for observed attributes and heteroge- consumers. The latent-class (LC) model (Kamakura and
neous intrinsic tastes for choice alternatives. Russell 1989) assumes that each consumer has a probability
More formally, let the systematic utility V of alternative j of belonging to several latent classes or segments. Thus, the
for consumer i be a function of structural covariates xij and underlying distribution of tastes is assumed to be discrete.
an alternative-specific dummy whose coefficient is an Alternatively, models allow parameters to vary across all
alternative-specific intercept aj, which can be interpreted households either by estimating parameters for each house-
as intrinsic brand preference, hold, i.e. fixed effects (FE) model or by assuming that these
parameters are distributed according to a probability dis-
Vij ˆ aj ‡ xij B ‡ eij : …6†
tribution and estimate the parameters of this distribution, i.e.
As alluded to earlier, people may have different tastes for RE model.
observed characteristics of the alternatives and different There are three main approaches to the FE model. The
sensitivity to marketing variables. Similarly, consumers may first available approach involves, at least in theory, estima-
have heterogeneous preferences for unmeasured properties of tion of household-specific terms, for instance a constant for
the alternatives which are approximated by nominal variables each household for each brand. Unfortunately, it quickly
such as alternative-specific dummies (McFadden, 1980). becomes impractical as it requires sufficiently long purchase
This implies that the structural parameters of the model B strings and estimation of N(Mÿ1) alternative-specific inter-
and also the alternative-specific intercepts aj vary over cepts where N, number of households and M, number of
people. Particularly, in working with observations on re- choice alternatives.
peated choices of the sampled individuals (e.g. panel data), An alternative treatment of heterogeneity in intrinsic
this interindividual variation in tastes should not be treated alternative preferences has been developed by Chamberlain
as a random event. Without adjusting for interindividual (1980) and demonstrated in marketing data by Jones and
differences, the model may yield biased estimates of aggre- Landwehr (1988). Chamberlain proposed a conditional
gate market response (Chamberlain, 1980; Hsiao, 1986). As maximum likelihood method that yields consistent esti-
depicted in Fig. 1, the literature reflects various approaches mates of B by conditioning on sufficient statistics of the
to account for taste heterogeneity. intrinsic alternative preferences aij. Note that the condi-
Starting from the most simple strategy, characteristics of tional method takes out of the likelihood function the
the individuals can enter the utility function in cases where a parameters aij and does not estimate them. Its usefulness
systematic association between individual descriptors and lies in the consistent estimation of B. Household-level
taste variation is assumed. For example, income may be probabilities are not inferred (Chintagunta et al., 1991).
correlated with the importance attached to price. Despite its As the number of alternatives increases, the conditional ML
simplicity, this approach can yield immediate insights into becomes quite impractical.
segmentation and targeting as it associates specific indivi- A third approach to the FE specification is Bayesian
dual characteristics with product and attribute preferences. (Rossi and Allenby, 1993). In this technique, the individual
Specific hypotheses can also be tested. For example, a purchase string updates pooled parameter estimates to form
related explanation for the different tastes of consumers individual-level counterparts. The prior information ensures,
with different buying power was introduced by Muellbauer in essence, that all probabilities are positive and resolves the
(1975), discussed by Deaton and Muellbauer (1980, p. 262) problem of unidentified parameters when some alternatives
and applied in purchase data by Allenby and Rossi (1991). are never bought by the household.
Chakraborty et al. (1992) proposed a screening method to The RE approach can take two primary forms. The first is
identify interactions between demographic variables and a parametric form in which tastes are distributed according
attribute coefficients. to some predetermined continuous distributions such as
Other studies use purchase history as an explanatory normal, lognormal, or gamma (Steckel and Vanhonacker,
variable, given as a function of observed past behavior (e.g. 1988; Gonul and Srinivasan, 1993; Allenby and Lenk,
Guadagni and Little, 1983; Krishnamurthi and Raj, 1988) to 1995; Kim et al., 1995). Using a continuous distribution,
account for cross-sectional heterogeneity (see also Fader we allow tastes (parameters) to vary across the population
and Lattin, 1993). It is straightforward to employ these two according to the specific distribution function and estimate
approaches in a standard MNL model. The best way to its mean and variance. Attention should be given to the
characterize them is as corrections for the homogeneous empirical implications of the distributional assumptions. An
model, not as explicit treatments of heterogeneity. incorrect probability distribution would result in biased
 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125 121

estimates (Heckman and Singer, 1984). Similarly, the dis- From a managerial standpoint, distinguishing heteroge-
tribution that is invoked to characterize taste heterogeneity neity from state dependence is important because the degree
should agree with theory about consumer tastes. For exam- of the latter determines the long-term results of short-term
ple, Gonul and Srinivasan (1993), McCulloch and Rossi policies. State dependence implies that short-term activities
(1994), and Allenby and Lenk (1995) use a normal distribu- have long-term potential since there exists a clear causal
tion for the price parameter which forces some households link between current and future choice behavior (see, e.g.
to have positive price coefficient. Since economic theory Roy et al., 1996).
dictates a negative price coefficient, an appropriate distribu- The explicit treatment of unobserved heterogeneity
tion with restricted range should be employed (Kim et al., makes FE and RE models (e.g. Chintagunta et al., 1991)
1995; Brownstone and Train 1996; Train, 1998). For ex- particularly well-suited for determining dynamic choice in
ample, the lognormal distribution for the coefficient of the panel data. In practice, state dependence is often accom-
negative of price can do this job. modated in RU models by making the systematic utility
The second approach to RE modeling is semiparametric. component a function of past choices. This is usually done
It approximates the underlying distribution of tastes with a by introducing a dummy for lagged purchase (Jones and
discrete probability distribution. The location and probabil- Landwehr, 1988) or a variable summarizing the entire
ity masses associated with the support points are estimated choice history (see Keane, 1997a). Another possibility is
empirically (Chintagunta et al., 1991; Chintagunta, 1992a; to let attributes of previously selected alternatives to influ-
Jain et al., 1994; Chintagunta and Honore, 1996). ence current attribute tastes (Erdem, 1996).
Note that the main difference between the semipara- In addition, current utilities may depend not only on past
metric-RE model and the LC model is one of interpretation. choices but also on their past values (Heckman, 1981a; Roy
Each support point of the former can be regarded as an LC et al., 1996). Economists have called this phenomenon,
and similarly each LC can be interpreted as a support point. habit persistence. Stated intuitively, the current utility of a
The difference is that the LC model postulates a finite set of brand depends not only on prior buying behavior but also on
segments of homogeneous members while the RE model, prior propensities to buy. This short-run inertia in prefer-
with a discrete distribution, postulates a continuous distri- ences may lead to serially correlated disturbances even in
bution of tastes that is approximated by a finite set of the presence of lagged choice variables which account for
support points. true state dependence. Therefore, the empirical detection of
A strong pattern in the above studies is that the average habit-persistence according to Heckman (1981a) requires
impact of policy variables on choice is greater than the control for state dependence.
level identified by models ignoring heterogeneity (see, e.g. As alluded to earlier, serial correlation can arise out of
Chintagunta et al., 1991; Chintagunta, 1992a; Gonul and unobserved heterogeneity, state dependence or short-run
Srinivasan, 1993; Chintagunta and Honore, 1996). inertia in preferences (Roy et al., 1996; Keane, 1997a,b). A
Loosely speaking, controlling for interindividual variation first attempt at the joint determination of heterogeneity, state
increases the explanatory power of non-stochastic factors dependence, and habit has been made by Roy et al. (1996).
included in the model. Dealing with heterogeneity may This study also provides an instructive discussion of the
also afford insights for segmentation and fine-tuning of issues involved in dynamic models. Keane (1997a) develops
marketing efforts. another quite comprehensive framework that admits struc-
Having dealt with issues of interpersonal heterogeneity, tural state dependence and autoregressive error terms while
we now turn to intrapersonal dynamics. As mentioned controlling for very rich heterogeneity structures.
above, we shall consider state dependence and inertia as
two forms of dynamic choice. 4.3. Heterogeneous choice sets
An important empirical problem is the confounding of
heterogeneity and state dependence. Time-invariant taste This research stream suggests that consumers may not
differences over people (i.e. heterogeneity) and effects of consider all the available alternatives before making a
the past experiences on current utilities (i.e. state depen- choice. Thus, choice set may be idiosyncratic for each
dence) lead to choice behavior that exhibits persistence over consumer reflecting a multistage choice process. The eco-
time. Loosely speaking, current behavior conveys informa- nomic explanation for this restriction on alternative evalua-
tion for past choices of the same individual because of two tion is that consumers continue to search for information as
underlying factors: common tastes and dynamic links. long as the returns from that search exceed the respective
Without controlling for heterogeneity, past choices may costs. From a psychological viewpoint, this reduction is
appear to be a determinant of current behavior solely viewed as a way to cope with complexity by filtering the
because they are a proxy for temporally persistent hetero- alternatives using simple criteria before the detailed evalua-
geneous preferences (e.g. Heckman, 1981a,b; Hsiao, 1986). tion of the reduced set.
Similarly, if state dependence does exist and is neglected, The more specific question as to whether choice set
the degree of time-invariant heterogeneity will be overstated effects should be incorporated in RU models does not have
(Keane, 1997a). a universally accepted answer. For example, Horowitz and
122 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125

Louviere (1995) argue that choice sets reflect preferences, to the researcher. Therefore, there is less scope for
which are normally described by the utility function. unobserved attributes to influence utilities. Loosely speak-
The relevant literature reflects various approaches to ing, the stochastic component here deals more with
choice set effects. Gensch (1987) hypothesizes an attribute unobserved heterogeneity in decision makers and less with
processing stage that determines the options forming the unobserved heterogeneity in alternatives. Unexplained var-
relevant consideration set. Hauser and Wernerfelt (1990) iation at the individual level can be viewed as noise that
proposed a model where set formation is a trade-off arises out of confusion, indifference, and inattention and
between costs and benefits for including and excluding therefore can be treated as a pure chance event that is
alternatives. Silk and Urban (1978), Roberts and Lattin randomly distributed (Hensher, 1984). The emphasis on
(1991), and Horowitz and Louviere (1995) specified sets taste heterogeneity can also be attributed to the character
on the basis of direct consumer reports. Siddarth et al. of the typical experimental task where each respondent is
(1995) define the choice set as the set of options which offered a series of choices. Hence, the RU model should
has been chosen at least once during a time interval. account for correlation in unobserved utility over repeated
Bronnenberg and Vanhonacker (1996) develop a model choices by each heterogeneous individual (Revelt and
that incorporates the probability of each alternative be- Train, 1997). This argument however should not be
longing to a consumer's choice set. Ben-Akiva and over-interpreted. Certain design factors (e.g. brand name)
Boccara (1995) uses information both from consumer may convey information about various objectives or per-
reports and observed choice to infer probabilistic choice ceived characteristics and induce unobserved heterogeneity
set generation. Allenby and Ginter (1995) include choice in alternatives.
set effects in a single-stage model and examine the Finally, a related approach introduced by Beggs et al.
influence of merchandising variables. Andrews and Srini- (1981) and Chapman and Staelin (1982) is to use an ordered
vasan (1995) enumerate all possible choice sets and then logit model on ranked individual data which allows, in
model the probability that each of them is the choice set principle, more information to be gathered from respondents
for a particular consumer. The current trend is towards but involves other problems (Ben-Akiva et al., 1992). For
probabilistic formulations which view choice sets as forceful statements of ECA along with a detailed discussion
latent, unobservable constructs that cannot be imputed of several practical issues, the interested reader is referred to
with certainty on the basis of observational data. Shocker Louviere (1994).
et al. (1991) and Ben-Akiva and Boccara (1995) note that
this implies a more realistic approximation of individual
choice behavior. 6. Some practical issues

The literature has given little attention to some prac-

5. RU models and experimental data tical, yet important issues encountered in applied work.
One such problem is the definition of the choice alter-
RU models have been applied primarily to observational native. In most marketing applications, the elemental
data such as scanner and survey data. This is the typical alternative has been the brand. However, a brand normally
application for any econometric model, with observations covers several variants, e.g. different flavors, formulas, etc.
from natural ``experiments'' provided by the market. De- Assuming away most of these details may yield poor
spite the appeal of real-world data, analysts might want to policy implications. Aggregation of alternatives may also
collect experimental choice data for specific projects where introduce an aggregation bias that increases with the
observational data are less revealing, for example, new heterogeneity of elemental alternatives (Parsons and Kealy,
product design (McFadden, 1986; Louviere, 1992). 1992). Recently, Fader and Hardie (1996) raised this issue
In experimental choice analysis (ECA), respondents are and proposed a way of using detailed attribute information
asked to choose among attribute bundles (e.g. Louviere and in RU models. The issue of alternative aggregation is
Hensher, 1983; Louviere and Woodworth, 1983; Hensher, directly related to the definition of the alternative set. In a
1984; Louviere, 1988; Louviere and Batsell, 1991; Chakra- subcategory, for example decaffeinated instant coffee, there
borty et al., 1992; Elrod et al., 1992). Then, an RU model is is less attribute variation than in the coffee product class.
invoked to estimate the mapping from the experimental In some cases, the only measurable attributes that vary
conditions (attributes of the alternatives and decision ma- across alternatives may be merchandising variables and
kers) to the results (choices). In this respect, the nature of brand name.
ECA approximates real market behavior better than tradi- When the number of alternatives gets too large, the
tional conjoint methods where there is no corresponding alternative set may not be manageable. A solution intro-
market behavior to the usual rating and ranking tasks (see duced by McFadden (1978) is to randomly sample from the
McFadden, 1986; Louviere, 1994). full alternative set and treat the consumer choice as having
Notice that in ECA, all attributes characterizing the come from the reduced set. This procedure exploits the IIA
hypothetical multiattribute alternatives are known a priori property and leads to consistent inference with the MNL
 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125 123

(Parsons and Kealy, 1992) but it is not certain how it works Acknowledgments
with other models. Another strategy to reduce the computa-
tional burden of a large number of alternatives is to sample Because of space limitations some contributions have not
from the full alternative set with unequal selection prob- been given the attention that they deserve. The authors
abilities which reflect the importance of alternatives (Ben- thank two anonymous JBR reviewers and the editors for
Akiva and Lerman, 1985, pp. 261 ± 269). Alternatively, the helpful comments.
researcher can cluster items sharing the same attribute and
estimate a NMNL model (Kannan and Wright, 1991) or
assume heterogeneous choice sets and employ one of the References
procedures discussed in the previous section.
A related empirical issue is how to deal with a large Allenby GM, Ginter JL. The effects of in-store displays and feature adver-
number of explanatory variables in discrete choice mod- tising on consideration sets. Int J Res Mark 1995;12:67 ± 79.
Allenby GM, Lenk PJ. Reassessing brand loyalty, price sensitivity, and
els Ð a problem often encountered in working with indivi-
merchandising effects on consumer brand choice. J Bus Econ Stat
dual descriptors. As alluded to earlier, Chakraborty et al. 1995;13:281 ± 9.
proposed a method for empirically selecting significant Allenby GM, Rossi P. Quality perceptions and asymmetric switching be-
variables in choice models. tween brands. Mark Sci (Summer) 1991;10:185 ± 204.
Another important empirical issue noted by Krishna- Anderson SP, de Palma A, Thisse J-F. Discrete Choice Theory of Product
Differentiation. Cambridge, MA: MIT Press, 1992.
murthi and Raj (1991) and analyzed empirically by Kim
Andrews RL, Srinivasan TC. Studying consideration effects in empirical
and Rossi (1994) relates to the exclusion of infrequent choice models using scanner panel data. J Mark Res (February)
purchasers because of estimation or other requirements. 1995;32:30 ± 41.
Such sample selection rules may induce a heavy-user bias Baltas G. An integrated model of category demand and brand choice. J
(Kim and Rossi, 1994). Most research on scanner data also Mark Res Soc 1998;40:295 ± 306.
Baltas G, Doyle P. A flexible model for consumer choice in packaged goods
tends to drop small brands for lack of observations, or to
categories. J Mark Res Soc 1998;40:141 ± 53.
simplify the analysis. While such sample selection practices Baltas G, Doyle P, Dyson P. A model of consumer choice for national vs.
could be justified for illustrating a new model or approach, store brands. J Oper Res Soc 1997;48:988 ± 95.
they are less acceptable in applied, business settings where a Beggs S, Cardell S, Hausman J. Assessing the potential demand for electric
complete picture of the market is required. cars. J Econometrics 1981;16:1 ± 19.
Ben-Akiva M. Structure of Passenger Travel Demand Models. PhD dis-
Finally, a general concern relates to overall model practi-
sertation. Department of Civil Engineering, MIT, 1973.
cality. As our discussion illustrates, recent developments Ben-Akiva M, Boccara B. Discete choice models with latent choice sets. Int
have increased model complexity and made estimation, J Res Mark 1995;12:9 ± 24.
interpretation, and forecasting less straightforward. Some Ben-Akiva M, Lerman S. Discrete Choice Analysis: Theory and Applica-
specifications are still rather impractical. The issue can be tion to Predict Travel Demand. Cambridge, MA: MIT Press, 1985.
Ben-Akiva M, Morikawa T, Shiroishi F. Analysis of the reliability of pre-
viewed as the common dilemma between simplicity and
ference ranking data. J Bus Res 1992;24:149 ± 64.
flexibility. There is no universal answer to this question as it Bhat CR. A heteroscedastic extreme value model of intercity travel mode
depends on one's rate of exchange between the two criteria. choice. Transp Res B 1995;29:471 ± 83.
Bock RD, Jones LV. The Measurement and Prediction of Judgement and
Choice. San Francisco, CA: Holden-Day, 1968.
Bronnenberg BJ, Vanhonacker WR. Limited choice sets, local price re-
7. Concluding remarks
sponse, and implied measures of price competition. J Mark Res
(May) 1996;33:163 ± 73.
In this paper, we have been concerned with RU models Bronnenberg BJ, Wathieu L. Asymmetric promotion effects and brand
for discrete choice. The theoretical framework laid out in the positioning. Mark Sci (Fall) 1996;15:379 ± 94.
outset of our discussion allowed several issues and models Brownstone D, Train K. Forecasting New Product Penetration with Flexible
Substitution Patterns. Working Paper, Department of Economics, Uni-
to be presented and interpreted. It is evident that RU models
versity of California, Berkeley, 1996.
have proved useful methods for assessing the effects of Buckley PG. Nested multinomial logit analysis for a hierarchical choice
marketing mix, guiding new product development, and model. J Bus Res (September) 1988;17:133 ± 54.
forecasting. Part of their current popularity stems from Bucklin RE, Gupta S. Brand choice, purchase incidence and segmentation:
recent advances in the collection of behavioral data through an integrated modeling approach. J Mark Res 1992;29:201 ± 15.
Chakraborty G, Woodworth G, Gaeth GJ, Ettenson R. Screening for inter-
retail checkouts. Besides packaged goods, RU models can
actions between design factors and demographics in choice-based con-
also be used to analyze discrete choice behavior in services, joint. J Bus Res (March) 1992;24:115 ± 33.
distribution, education, and other business settings invol- Chamberlain G. Analysis of covariance with qualitative data. Rev Econ
ving optimizing decisions. Finally, the models discussed in Stud 1980;47:225 ± 38.
this paper can be incorporated in more general demand Chapman R, Staelin R. Exploiting rank ordered choice set data within the
stochastic utility model. J Mark Res 1982;18:288 ± 301.
systems that attempt the independent (e.g. Gupta, 1988) or
Chiang J. A simultaneous approach to whether, what and how much to buy
joint (e.g. Chiang, 1991; Chintagunta, 1993; Baltas 1998) questions. Mark Sci 1991;10:297 ± 315.
determination of discrete and continuous decisions such as Chintagunta PK. Heterogeneity in nested logit models: an estimation ap-
which brand and how many units to buy. proach and empirical results. Int J Res Mark 1992;9:161 ± 75.
124 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125

Chintagunta PK. Estimating a multinomial probit model of brand choice conditions in estimating a discrete time ± discrete data stochastic pro-
using the method of simulated moments. Mark Sci (Fall) 1992;11: cess. In: Manski CF, McFadden D, editors. Structural Analysis of Dis-
386 ± 407. crete Data with Econometric Applications. Cambridge, MA: MIT Press,
Chintagunta PK. Investigating purchase incidence, brand choice and 1981. pp. 179 ± 95.
purchase quantity decisions of households. Mark Sci (Spring) 1993; Heckman JJ, Singer BA. Method for minimizing the impact of distribu-
12:184 ± 208. tional assumptions in econometric models for duration data. Econome-
Chintagunta PK, Honore BE. Investigating the effects of marketing vari- trica 1984;52:271 ± 320.
ables and unobserved heterogeneity in a multinomial probit model. Int J Hensher DA. Achieving representativeness of the observable component of
Res Mark 1996;13:1 ± 15. the indirect utility function in logit choice models: an empirical revela-
Chintagunta PK, Jain DC, Vilcassim NJ. Investigating heterogeneity in tion. J Bus 1984;57:265 ± 80.
brand preferences in logit models for panel data. J Mark Res (Novem- Horowitz JL, Louviere JJ. What is the role of consideration sets in choice
ber) 1991;28:417 ± 28. modeling. Int J Res Mark 1995;12:39 ± 54.
Currim IS. Using segmentation approaches for better prediction and under- Hsiao C. Analysis of Panel Data. Cambridge: Cambridge Univ. Press, 1986.
standing from consumer choice models. J Mark Res (August) Jain CJ, Vilcassim NF, Chintagunta PK. A random-coefficients logit brand-
1981;18:301 ± 9. choice model applied to panel data. J Bus Econ Stat 1994;12:317 ± 28.
Daganzo C. Multinomial Probit: The Theory and Applications to Demand Johnson EJ, Hensher D. Application of multinomial probit to a two-period
Forecasting. New York: Academic Press, 1979. panel data set. Transp Res A 1982;16:457 ± 64.
Deaton A, Muellbauer J. Economics and Consumer Behavior. Cambridge: Jones MJ, Landwehr JT. Removing heterogeneity bias from logit model
Cambridge Univ. Press, 1980. estimation. Mark Sci (Winter) 1988;7:41 ± 59.
Deigthon J, Henderson CM, Nelsin SA. The effects of advertising on Kamakura WA, Kim B-D, Lee J. Modeling preference and structural het-
brand switching and repeat purchasing. J Mark Res (February) erogeneity in consumer choice. Mark Sci 1996;15:152 ± 72.
1994;34:28 ± 43. Kamakura WA, Russell GJ. A probabilistic choice model for market seg-
Dubin JA. A nested logit model of space and water heat system choice. mentation and elasticity structure. J Mark Res ( November) 1989;26:
Mark Sci 1986;5:112 ± 24. 379 ± 90.
Elrod TS, Louviere J, Kavey KS. An empirical comparison of ratings-based Kamakura WA, Russell GJ. Measuring brand value with scanner data. Int J
and choice-based conjoint models. J Mark Res 1992;29:368 ± 77. Res Mark 1993;10:9 ± 22.
Erdem T. A dynamic analysis of market structure based on panel data. Mark Kamakura WA, Srivastava RK. Predicting choice shares under conditions
Sci 1996;15:359 ± 78. of brand interdependence. J Mark Res (November) 1984;21:420 ± 34.
Fader PS, Hardie BGS. Modeling consumer choice among SKUs. J Mark Kamakura WA, Srivastava RK. An ideal-point probabilistic choice model
Res 1996;33:442 ± 52. for heterogeneous preferences. Mark Sci (Summer) 1986;5:199 ± 218.
Fader PS, Lattin JM. Accounting for heterogeneity and nonstationarity in a Kannan PK, Wright GP. Modeling and testing structured markets: a nested
cross-sectional model of consumer purchase behavior. Mark Sci (Sum- logit approach. Mark Sci 1991;10:58 ± 82.
mer) 1993;12:304 ± 17. Keane MP. Modeling heterogeneity and state dependence in consumer
Gensch DH. Empirically testing a disaggregate choice model for segments. choice behavior. J Bus Econ Stat 1997;15:310 ± 11.
J Mark Res (November) 1985;22:462 ± 7. Keane MP. Current issues in discrete choice modeling. Mark Lett
Gensch DH. A two-stage disaggregate attribute choice model. Mark Sci 1997;8:307 ± 22.
(Summer) 1987;6:223 ± 39. Kim B-D, Blattberg RC, Rossi PE. Modeling the distribution of price
Gensch DH, Recker WW. The multinomial, multiattribute logit choice sensitivity and implications for optimal retail pricing. J Bus Econ Stat
model. J Mark Res (February) 1979;16:124 ± 32. 1995;13:291 ± 303.
Gonul F, Srinivasan K. Modeling multiple sources of heterogeneity in Kim B-D, Rossi PE. Purchase frequency, sample selection and price sensi-
multinomial logit models: methodological and managerial issues. Mark tivity: the heavy user bias. Mark Lett 1994;5:57 ± 68.
Sci (Summer) 1993;12:213 ± 29. Krishnamurthi L, Raj SP. A model of brand choice and purchase quantity
Guadagni PM, Little JD. A logit model of brand choice calibrated on price sensitivities. Mark Sci (Winter) 1988;7:1 ± 20.
scanner data. Mark Sci 1983;2:203 ± 38. Krishnamurthi L, Raj SP. An empirical analusis of the relationship between
Gupta S. Impact of sales promotions on when, what, and how much to buy. brand loyalty and consumer price elasticity. Mark Sci 1991;10:172 ± 83.
J Mark Res (November) 1988;25:342 ± 55. Lattin J, Bucklin RE. Reference effects of price and promotion on brand
Gupta S, Chintagunta PK. On using demographic variables to determine choice behavior. J Mark Res (August) 1989;26:299 ± 310.
segment membership in logit mixture models. J Mark Res (February) Lattin JM. A model of balanced choice behavior. Mark Sci 1987;6:48 ± 65.
1994;31:128 ± 36. Louviere J. Conjoint analysis. In: Bagozzi RP, editor. Advanced Methods of
Hajivassiliou V, McFadden D, Ruud P. Simulation of multivariate normal Marketing Research. Cambridge, MA: Blackwell, 1994. pp. 223 ± 59.
rectangle probabilities and their derivatives. J Econometrics 1996;72: Louviere J, Hensher DA. Using discrete choice models with experimental
85 ± 134. design data to forecast consumer demand for a unique cultural event.
Hardie BGS, Johnson EJ, Fader PS. Modeling loss aversion and reference J Consum Res (December) 1983;10:348 ± 61.
dependence effects on brand choice. Mark Sci (Fall) 1993;12:378 ± 94. Louviere J, Woodworth G. Design and analysis of simulated consumer
Hauser JR, Urban GL. A normative methodology for modeling consumer choice or allocation experiments: an approach based on aggregate data.
response to innovation. Oper Res 1977;25:579 ± 619. J Mark Res 1983;20:350 ± 67.
Hauser JR, Wernerfelt B. An evaluation cost model of consideration sets. J Louviere JJ. Conjoint modeling of stated preferences: a review of theory,
Consum Res (March) 1990;16:393 ± 408. methods, recent developments and external validity. J Transp Econ
Hausman J, Ruud P. Specifying and testing econometric models for rank- Policy 1988;10:93 ± 119.
order data. J Econometrics 1987;34:83 ± 104. Louviere JJ. Experimental choice analysis: introduction and overview.
Hausman J, Wise DA. A conditional probit model for qualitative choice: J Bus Res 1992;24:89 ± 95.
discrete decisions recognising interdependence and heterogeneous pre- Louviere JJ, Batsell RR. Experimental analysis of choice. Mark Lett
ferences. Econometrica 1978;46:403 ± 26. 1991;2:199 ± 214.
Heckman JJ. Statistical models for discrete panel data. In: Manski CF, Manski CF. The structure of random utility models. Theory Decis
McFadden D, editors. Structural Analysis of Discrete Data with Econo- 1977;8:229 ± 54.
metric Applications. Cambridge, MA: MIT Press, 1981. pp. 114 ± 78. McCulloch R, Rossi PE. An exact likelihood analysis of the multinomial
Heckman JJ. The incidental parameters problem and the problem of initial probit model. J Econometrics 1994;64:207 ± 40.
 G. Baltas, P. Doyle / Journal of Business Research 51 (2001) 115±125 125

McFadden D. Conditional logit analysis of qualitative choice behavior. In: Choices of Appliance Efficiency Level, working paper, Department of
Zaremka P, editor. Frontiers in Econometrics. New York: Academic Economics, University of California, Berkeley, 1997.
Press, 1973. pp. 105 ± 42. Roberts JH, Lattin JM. Development and testing of a model of considera-
McFadden D. Modelling the choice of residential location. In: Karquist A, tion set composition. J Mark Res (November) 1991;28:429 ± 40.
et al., editors. Spatial Interaction Theory and Planning Models. Amster- Rossi PE, Allenby GM. A Bayesian approach to estimating households
dam: North-Holland, 1978. pp. 75 ± 96. parameters. J Mark Res (May) 1993;30:171 ± 82.
McFadden D. Modeling the choice of residential location. Transp Res Rec Roy R, Chintagunta PK, Haldar S. A framework for investigating habits,
1979;673:72 ± 8. ``The Hand of the Past,'' and heterogeneity in dynamic brand choice.
McFadden D. Econometric models for probabilistic choice among products. Mark Sci 1996;15:280 ± 99.
J Bus 1980;53:513 ± 29. Shocker AD, Ben-Akiva M, Boccara B, Negungedi P. Consideration set
McFadden D. Econometric models of probabilistic choice. In: Manski CF, influences on consumer decision making and choice: issues, models,
McFadden D, editors. Structural Analysis of Discrete Data with and suggestions. Mark Lett 1991;2:181 ± 97.
Econometric Applications. Cambridge, MA: MIT Press, 1981. pp. Siddarth S, Bucklin RE, Morrison DG. Making the cut: analyzing choice set
198 ± 272. restriction in scanner panel data. J Mark Res (August) 1995;32:255 ± 66.
McFadden D. The choice theory approach to market research. Mark Sci Silk AJ, Urban GL. Pre-test-market evaluation of new packaged goods: a
1986;5:275 ± 97. model and measurement methodology. J Mark Res 1978;15:171 ± 91.
McFadden D. A method of simulated moments for the estimation of dis- Steckel JH, Vanhonacker WR. A heterogeneous conditional logit model of
crete response models without numerical integration. Econometrica choice. J Bus Econ Stat 1988;6:391 ± 8.
1989;57:995 ± 1026. Swait J, Erdem T, Louviere J, Dubelaar C. The equalization price: a
Mela CF, Gupta S, Lehmann DR. The long-term impact of promo- measure of consumer-perceived brand equity. Int J Res Mark
tion and advertising on consumer brand choice. J Mark Res 1993;10:23 ± 45.
(May) 1997;34:248 ± 61. Tellis GJ. Advertising exposure, loyalty and brand purchase: a two-stage
Muellbauer J. The cost of living and taste and quality changes. J Econ model of choice. J Mark Res 1988;25:134 ± 44.
Theory 1975;10:269 ± 83. Train K. Qualitative Choice Analysis. Cambridge, MA: MIT Press, 1986.
Papatla P, Krishnamurthi L. A probit model of choice dynamics. Mark Sci Train KE. Recreation demand models with taste differences over people.
(Spring) 1992;11:189 ± 206. Land Econ 1998;74:230 ± 9.
Parsons GR, Kealy MJ. Randomly drawn opportunity sets in a random Weeks M. The multinomial probit model revisited: a discussion of para-
utility model of lake recreation. Land Econ 1992;68:93 ± 106. meter estimability, identification and specification testing. J Econ Surv
Punj GN, Staelin R. The choice process for graduate business schools. 1997;11:297 ± 320.
J Mark Res 1978;15:588 ± 98. Winer RS. A reference price model of brand choice for frequently pur-
Revelt D, Train K. Mixed Logit with Repeated Choices: Households' chased products. J Consum Res (September) 1986;13:250 ± 6.