256
Journal of Marketing Research
Vol. XLII (August 2005), 256–265
*Anand V. Bodapati is Assistant Professor of Marketing (e-mail:
bodapati@ucla.edu), and Aimee Drolet is Associate Professor of Market-
ing (e-mail: adrolet@anderson.ucla.edu), Anderson School of Manage-
ment, University of California, Los Angeles. The authors acknowledge the
many thoughtful suggestions of the JMR reviewers. In addition, the authors
thank Don Morrison, Itamar Simonson, and Pat Suppes for comments on
previous drafts. The authors are especially grateful to Randy Bucklin for
his invaluable feedback and guidance at various stages of this research.
ANAND V. BODAPATI and AIMEE DROLET*
In the usual multinomial choice model, consumers choose to use
“actual value” information; that is, utilities are continuous functions of
product attributes (e.g., choices depend on actual magnitudes of price
differences). The authors propose an alternative model in which con-
sumers use only “ordered value” information; that is, utilities are functions
only of the relative orderings of the attributes’ values across alternatives
(e.g., choices depend only on the price ordering and not on actual
prices). The ordered value model is attractive because it fits well with
psychological evidence that consumers often favor decision mechanisms
that are cognitively less demanding. Using a supermarket shopper panel
data set, the authors evaluate four models in which (1) all consumers use
actual values; (2) all consumers use ordered values; (3) some con-
sumers use actual values all the time, and some consumers use ordered
values all the time; and (4) all consumers use both actual values and
ordered values but with different propensities. In the analysis, the
ordered value model finds stronger support than the actual value model:
Model 1 outperforms Model 2; in the two hybrid choice models (Models 3
and 4), ordered value processing is more prevalent than actual value
processing. These results suggest that consumers in some product cate-
gories engage more heavily in ordered value processing than in actual
value processing.
A Hybrid Choice Model That Uses Actual
and Ordered Attribute Value Information
The traditional assumption derived from both marketing
and economics is that consumers’ predefined preferences or
utility functions determine both their attitudinal and behav-
ioral responses to changes in products’ actual attribute val-
ues (for a recent review, see Gustafsson, Herrmann, and
Huber 2000). Techniques that are designed to predict con-
sumer choices, notably conjoint analysis, are founded on
this assumption. However, contrary to this assumption,
growing research suggests that consumers frequently do not
have predefined preferences but instead construct their pref-
erences as needed to make choices (for a review, see
Bettman, Luce, and Payne 1998).
To construct their preferences, consumers use a variety of
choice rules. The use of different rules is contingent on var-
ious factors of the decision problem (e.g., framing, context)
and of the decision maker (e.g., expertise) (Payne, Bettman,
and Johnson 1992, 1993; Simonson and Tversky 1992).
These factors can systematically and often dramatically
influence choice outcomes, even though, normatively
speaking, they should not. For example, the framing of
choices (e.g., in terms of “lives saved” versus “lives lost”;
Tversky and Kahneman 1991) can cause large preference
reversals, even though the consequences of choices are for-
mally identical.
These rules all transform information inputs into final
goal states of knowledge, but they can vary in the type of
attribute value information they use (Payne, Bettman, and
Johnson 1993). Some rules use actual attribute values (e.g.,
$.10, 10X magnification power) as inputs. For example, the
weighted-additive rule considers the actual values of all
alternatives on all the relevant attributes. Other rules make
use of ordered values, or values that are defined by the rela-
tive positions of considered options along attributes (e.g.,
higher magnification, most expensive). For example, the
lexicographic rule chooses the alternative with the relatively
best value on the most important attribute.
Rules using actual values as inputs are in accord with the
idea that “fundamental” values underlie consumer choices.
Such rules require that consumers assess the true or psycho-

A Hybrid Choice Model 257
logical significance of a product’s actual values or actual
value differences between a product and other considered
products. However, for consumers, these tasks are difficult,
even for simple products and even when full information is
available (Ariely, Loewenstein, and Prelec 2003; Frederick
and Fischhoff 1998). In contrast, consumers can make rela-
tive valuations easily. Accordingly, much research shows
that consumers rely heavily on rules that use ordered values
as information inputs. Such rules convert actual values to
more psychologically meaningful, better-defined ordered
values (e.g., relative attribute importance) (Drolet, Simon-
son, and Tversky 2000).
If choices are based on ordered rather than actual values,
adjustments in considered products’ actual attribute value
profiles will presumably not have an appreciable effect on
choice shares as long as products’ relative positions along
attributes are preserved. If preserved, consumers should be
less responsive to actual value changes than traditional
models, such as the multinomial logit (MNL) model, pre-
dict. Because these models assume that consumers assess
the psychological value of actual value differences, even
minor adjustments in products’ actual values should affect
preferences. In contrast, if choices are based on ordered val-
ues and the relative positions of products on attributes are
not preserved, even slightly perceptible actual value
changes can have a dramatic effect on choice shares. This
(disproportionately large) response is at odds with what tra-
ditional models predict.
To illustrate, consider three situations in which a consumer
chooses between two products, A and B. In Situation 1, the
price of Product A is $7.55, and the price of Product B is
$7.54; the price difference between A and B is $.01. In Situ-
ation 2, Product A is $7.54, and Product B is $7.55; the dif-
ference between A and B is –$.01. In Situation 3, Product A
is $7.52, and Product B is $7.55; the price difference between
A and B is –$.03. The MNL model predicts that the increase
in Product A’s choice share when moving from Situation 1 to
Situation 2 will be approximately equal to the increase in
Product A’s choice share when moving from Situation 2 to
Situation 3 because the corresponding shifts in actual value
difference in price are equal. However, intuition and experi-
mental research (e.g., Drolet and Bodapati 2003) suggest that
Product A’s choice share will increase more when moving
from Situation 1 to Situation 2 than from Situation 2 to Situ-
ation 3. Although the two moves are quantitatively equiva-
lent, they differ qualitatively. The move from Situation 1 to
Situation 2 alters which product (A or B) is priced lower. The
move from Situation 2 to Situation 3 does not. Consumers’
heavy use of decision rules based on ordered values suggests
that consumers would pay greater attention to the ordering of
the prices than to their actual values and that there would be
a greater increase in Product A’s choice share if the ordering
of the two products’ prices is altered.
Nevertheless, consumers do not base all of their choices
solely on ordered values. If they did, we would predict that
the change in the choice share of Product A when moving
from Situation 2 to Situation 3 would be zero. The move
from Situation 2 to Situation 3 preserves the ordering of
prices and alters only the magnitude of the price differ-
ence. Similar to the MNL model’s prediction that con-
sumers respond only to actual value differences, this pre-
diction appears equally incorrect in light of empirical
evidence that even order-preserving changes in price have
an effect on choice shares. Consumers typically do not
consider only the ordering of prices, because even order-
preserving changes alter choice if the changes are large
enough.
In short, it appears that models that predict choices on
the basis of only one kind of attribute value processing
(actual or ordered) predict choices less effectively than do
models that represent decision making as involving the
use of both actual value processing and ordered value pro-
cessing. In this article, we develop an econometric model
that allows consumers to choose either actual or ordered
value information. We estimate the hybrid model using
real consumer purchase data from an Information
Resources Inc. (IRI) scanner panel data set. The results
indicate that the hybrid model predicts choices better than
do models that allow consumers to use only one type of
attribute value information, actual or ordered. This article
concludes with a discussion of the implications of this
research.
MODEL
Our proposed model assumes that on any one choice
occasion, a consumer may operate in one of two decision
modes, actual value mode (AVM) or ordered value mode
(OVM). We assume that a consumer who operates in OVM
evaluates each product only according to its position along
attributes relative to those of other considered products. In
contrast, a consumer who operates in AVM evaluates the
numerical differences between the actual attribute values
of a product and the actual attribute values of other prod-
ucts. The probability of choosing a product i in AVM or
OVM is given by p(choice = i|AVM) and p(choice =
i|OVM), respectively; we derive algebraic expressions sub-
sequently. Because we do not ordinarily know in which of
the two modes the consumer operates on any given occa-
sion, we posit a probability pAVM for the consumer operat-
ing in AVM. The overall probability of the consumer’s
choosing product i is a weighted average of p(choice =
i|AVM) and p(choice = i|OVM); the weight is governed by
pAVM.
Choice Probabilities for AVM
For a consumer operating in AVM, the utility for product
i from a set of I products and described on a space of A
attributes as (xi1, xi2, …, xiA) is as follows:
where ei is an error term. The consumer chooses the option
I, which maximizes utility UiAVM, and the probability that
the consumer chooses i is as follows:
If we assume that the stochastic components e1, e2, …, ei
are drawn from the extreme value distribution, the choice
probability is given by the following:
( ) ( , )2
1
p pu u j i b x e b x ei j a ia i a ja j
a
A
> ∀ ≠ = + > +
=
∑∑
=
∀ ≠






a
A
j i
1
, .
( ) ,1
1
U b x ei a ia
a
A
iAVM
= +
=
∑