Consumer Decision Making in Online

Shopping Environments: The Effects of

Interactive Decision Aids
Gerald Häubl • Valerie Trifts
Faculty of Business, University of Alberta, Edmonton, Alberta, Canada T6G 2R6
[email protected] • [email protected]

consumer’s own utility function (attribute importance

Abstract weights and minimum acceptable attribute levels), the RA
Despite the explosive growth of electronic commerce and the
generates a personalized list of recommended alternatives.
rapidly increasing number of consumers who use interactive
The second decision aid, a comparison matrix (CM), is de-
media (such as the World Wide Web) for prepurchase infor-
signed to help consumers make in-depth comparisons
mation search and online shopping, very little is known
among selected alternatives. The CM allows consumers to
about how consumers make purchase decisions in such set-
organize attribute information about multiple products in an
tings. A unique characteristic of online shopping environ-
alternatives ⳯ attributes matrix and to have alternatives
ments is that they allow vendors to create retail interfaces
sorted by any attribute.
with highly interactive features. One desirable form of inter-
Based on theoretical and empirical work in marketing,
activity from a consumer perspective is the implementation
judgment and decision making, psychology, and decision
of sophisticated tools to assist shoppers in their purchase de-
support systems, we develop a set of hypotheses pertaining
cisions by customizing the electronic shopping environment
to the effects of these two decision aids on various aspects of
to their individual preferences. The availability of such
consumer decision making. In particular, we focus on how
tools, which we refer to as interactive decision aids for consum-
use of the RA and CM affects consumers’ search for product
ers, may lead to a transformation of the way in which shop-
pers search for product information and make purchase de- information, the size and quality of their consideration sets,
cisions. The primary objective of this paper is to investigate and the quality of their purchase decisions in an online shop-
the nature of the effects that interactive decision aids may ping environment.
have on consumer decision making in online shopping A controlled experiment using a simulated online store
environments. was conducted to test the hypotheses. The results indicate
While making purchase decisions, consumers are often that both interactive decision aids have a substantial impact
unable to evaluate all available alternatives in great depth on consumer decision making. As predicted, use of the RA
and, thus, tend to use two-stage processes to reach their de- reduces consumers’ search effort for product information,
cisions. At the first stage, consumers typically screen a large decreases the size but increases the quality of their consid-
set of available products and identify a subset of the most eration sets, and improves the quality of their purchase de-
promising alternatives. Subsequently, they evaluate the lat- cisions. Use of the CM also leads to a decrease in the size but
ter in more depth, perform relative comparisons across prod- an increase in the quality of consumers’ consideration sets,
ucts on important attributes, and make a purchase decision. and has a favorable effect on some indicators of decision
Given the different tasks to be performed in such a two-stage quality.
process, interactive tools that provide support to consumers In sum, our findings suggest that interactive tools de-
in the following respects are particularly valuable: (1) the signed to assist consumers in the initial screening of available
initial screening of available products to determine which alternatives and to facilitate in-depth comparisons among se-
ones are worth considering further, and (2) the in-depth com- lected alternatives in an online shopping environment may
parison of selected products before making the actual pur- have strong favorable effects on both the quality and the ef-
chase decision. This paper examines the effects of two deci- ficiency of purchase decisions—shoppers can make much
sion aids, each designed to assist consumers in performing better decisions while expending substantially less effort. This
one of the above tasks, on purchase decision making in an suggests that interactive decision aids have the potential to
online store. drastically transform the way in which consumers search for
The first interactive tool, a recommendation agent (RA), al- product information and make purchase decisions.
lows consumers to more efficiently screen the (potentially (Decision Making; Online Shopping; Electronic Commerce; De-
very large) set of alternatives available in an online shopping cision Aids; Recommendation Agents; Consumer Behavior; Infor-
environment. Based on self-explicated information about a mation Search; Consideration Sets; Information Processing)

Marketing Science 䉷 2000 INFORMS 0732-2399/00/1901/0004/$05.00

Vol. 19, No. 1, Winter 2000, pp. 4–21 1526-548X electronic ISSN
 CONSUMER DECISION MAKING IN
ONLINE SHOPPING ENVIRONMENTS

Introduction (Alba et al. 1997, Winer et al. 1997), theorizing about

The popularity of interactive media such as the World the nature of these differences has been sparse. We
Wide Web (WWW) has been growing at a very rapid propose that consumer behavior in an online shopping
pace (see, e.g., GVU 1999). From a marketing perspec- environment is determined largely by the degree and
tive, this has manifested itself primarily in two ways: type of machine interactivity that is implemented in
(1) a drastic increase in the number of companies that such a setting. Specifically, we hypothesize that the
seek to use the WWW to communicate with (potential) way in which consumers search for product informa-
customers, and (2) the rapid adoption of the WWW by tion and make purchase decisions is a function of the
broad consumer segments for a variety of purposes, particular interactive tools available in an online shop-
including prepurchase information search and online ping environment. We refer to such tools as interactive
shopping (Alba et al. 1997). The combination of these decision aids for consumers.
two developments provides a basis for substantial In this paper, we identify two types of interactive
growth in the commercial use of interactive media. decision aids that, in light of established notions about
The focus of this paper is on one specific type of purchase decision processes, seem particularly valu-
commercial use of interactive media: shopping in online able to consumers. Based on theoretical and empirical
environments. We conceptualize this behavior as a work in marketing, judgment and decision making,
shopping activity performed by a consumer via a psychology, and decision support systems, we develop
computer-based interface, where the consumer’s com- a set of hypotheses pertaining to the effects of each of
puter is connected to, and can interact with, a retailer’s these tools on consumers’ search for product infor-
digital storefront (implemented on some computer) mation, the size and quality of their consideration sets,
through a network (e.g., the WWW). A consumer can and the quality of their purchase decisions in an online
engage in online shopping in any location, but our con- store. The results of a controlled experiment indicate
ceptualization is based on the assumptions that the that each of the interactive decision aids has a substan-
products of interest are not physically present at the tial impact on consumer decision making, thus provid-
time and that no face-to-face assistance is available to ing a demonstration of how the availability of such
the shopper. tools may transform the way in which individuals
A unique characteristic of online shopping environ- search for information and make purchase decisions in
ments is that they allow for the implementation of very online environments.
high degrees of interactivity. The latter is a multidi- The paper is organized as follows. First, we briefly
mensional construct, the key facets of which include discuss the relevant literature on human decision mak-
reciprocity in the exchange of information, availability ing and decision aids. Next, we provide an overview
of information on demand, response contingency, cus- of interactive decision aids for online shopping and
tomization of content, and real-time feedback (Alba et motivate the choice of the two particular tools inves-
al. 1997, Ariely 2000, Zack 1993). In the context of tigated in our study. We then develop a set of hypoth-
computer-mediated communication, a distinction has eses pertaining to how we expect each of these decision
been made between person interactivity and machine aids to affect different aspects of consumer decision
interactivity. While the former describes the ability to making in online shopping environments. This is fol-
communicate with other individuals, the latter refers lowed by a description of the method used to test these
to the ability to interactively access information in an hypotheses. We then report the results of our empirical
online database (Hoffman and Novak 1996, p. 53). study. The paper concludes with a general discussion
Given our conceptualization of shopping in online en- of the findings.
vironments, the concept of machine interactivity is of
particular interest. Human Decision Making and
While it has been hypothesized that consumers’ Decision Aids
shopping behavior in online stores may be fundamen- Humans adapt their decision making strategies to spe-
tally different from that in traditional retail settings cific situations and environments (see, e.g., Payne

Marketing Science/Vol. 19, No. 1, Winter 2000 5

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

1982). They can be described as “cognitive misers” is subject to delay and ambiguity, decision makers may
who strive to reduce the amount of cognitive effort be inclined to focus more on reducing cognitive effort
associated with decision making (Shugan 1980). The than on improving decision accuracy (Einhorn and
notion that individuals are typically willing to settle Hogarth 1978, Kleinmuntz and Schkade 1993). Thus,
for imperfect accuracy of their decisions in return for decision aids may lead individuals to merely reduce
a reduction in effort is well supported (Bettman et al. effort without improving the quality of their decisions.
1990, Johnson and Payne 1985) and consistent with the In fact, there is empirical evidence that the use of de-
idea of bounded rationality (Simon 1955). Because of cision aids does not necessarily enhance decision mak-
this trade-off between effort and accuracy, decision ing performance (cf. Benbasat and Nault 1990), and
makers frequently choose options that are satisfactory that the latter may even be reduced as a result (Todd
but would be suboptimal if decision costs were zero. and Benbasat 1992, p. 373). Given this mixed evidence,
This is particularly common when alternatives are nu- it cannot be assumed that a consumer’s use of inter-
merous and/or difficult to compare, i.e., when the active decision aids in an online shopping context will
complexity of the decision environment is high (Payne lead to increased decision quality. Rather, this repre-
et al. 1993). sents an open question, which is addressed in this
One form of coping with highly complex decision paper.
environments is to use decision support systems. The In the following section, we first provide a general
latter are computer-based technologies designed to as- overview of interactive decision aids available to con-
sist an individual (or a group) in making a decision or sumers for the purpose of online shopping. Based on
choosing a course of action in a nonroutine situation established notions about purchase decision making
that requires judgment (Kasper 1996). Decision sup-
and on characteristic features of online shopping en-
port systems contain one or more tools, or decision aids,
vironments, two decision aids are selected for inclu-
that perform distinct information processing tasks or
sion in our empirical study. These two tools are then
functions (e.g., search a database or sort objects by
discussed in detail.
some criterion). The motivating principle underlying
decision aids is that resource-intensive, but standar-
dizable, information processing tasks are performed
by a computer-based system, thus freeing up some of Interactive Decision Aids for Online
the human decision maker’s processing capacity. De- Shopping
termining an adequate “division of labor” between hu-
Overview of Tools
man and computer is crucial. Human decision makers
are typically good at selecting variables that are rele- The technology available for implementing machine
vant in the decision process, but weak at integrating interactivity in online shopping environments has the
and retaining large amounts of information. Effective potential to provide consumers with unparalleled op-
decision aids should be designed to capitalize on the portunities to locate and compare product offerings
strengths and compensate for the inherent weaknesses (Alba et al. 1997, p. 38). Such capabilities are particu-
of their users (Hoch and Schkade 1996). larly valuable given that online stores cannot offer
A standard assumption in past research on decision physical contact with products, do not allow face-to-
support systems, most of which has focused on man- face interaction with a salesperson, and may offer a
agerial decisions (e.g., Pearson and Shim 1994), is that very large number of alternatives because of their vir-
decision makers who are provided with decision aids tually infinite “shelfspace,” i.e., the lack of physical
that have adequate information processing capabilities constraints with respect to product display.
will use these tools to analyze problems in greater Interactive decision aids that may be of use to con-
depth and, as a result, make better decisions (cf. Hoch sumers who wish to shop online include a wide variety
and Schkade 1996). However, behavioral decision the- of software tools, ranging from general-purpose search
ory suggests that because feedback on effort expendi- engines (e.g., www.infoseek.com, www.lycos.com) to
ture tends to be immediate while feedback on accuracy sophisticated agent-mediated electronic commerce

6 Marketing Science/Vol. 19, No. 1, Winter 2000

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

systems (e.g., compare.net, www.jango.com). A com- Recommendation Agent: A Tool for Screening
mon classification of interactive shopping agents is Alternatives
based on whether a tool is designed to help a consumer We conceptualize a recommendation agent (RA) as an
determine (1) what to buy or (2) whom to buy from. interactive decision aid that assists consumers in the
These two tasks may be referred to as product brokering initial screening of the alternatives that are available in
and merchant brokering, respectively (see Guttman et al. an online store. Based on information provided by the
1998). For the purpose of this paper, we confine our shopper regarding his/her own preference, an RA
attention to the former. “recommends” a set of products that are likely to be
Among tools for product brokering, a distinction can attractive to that individual. Elementary forms of this
be made between decision aids that operate within a type of decision aid are currently implemented on a
particular merchant’s online store (e.g., www.person number of online retail sites (e.g., www.macys.com,
alogic.com) and ones that operate across merchants www.netmarket.com). A real-world tool that corre-
(e.g., www.shopper.com). The primary focus of this sponds very closely to our conceptualization of RA is
paper is on the former. The decision aids we investi- the consumer decision guide developed by Persona-
gate are implemented within an online store (see be- Logic娃 (www.personalogic.com).
low). However, this research also pertains to those The RA used in the present study generates a per-
cross-merchant decision aids that allow shoppers di- sonalized list of recommended alternatives, in which
rect access to a common product database (e.g., alternatives are described by their brand and model
www.jango.com), provided that these tools do not dis- name.1 This recommendation is based on three types
criminate between products on the basis of which ven- of parameters provided by the consumer. First, a con-
dor they are associated with.
sumer’s self-explicated attribute importance weights
A well-known phenomenon regarding decision
are used to compute a summary score for each alter-
making in complex environments is that individuals
native as the sum over all products of (standardized)
are often unable to evaluate all available alternatives
attribute level scale value and corresponding impor-
in great depth prior to making a choice (Beach 1993).
tance weight.2 This score determines the order of al-
Instead, they tend to use two-stage processes to reach
ternatives in the RA’s output. Thus, the RA is effec-
their decisions, where the depth of information pro-
tively an automated implementation of a weighted
cessing varies by stage (Payne 1982, Payne et al. 1988).
additive evaluation rule (Payne et al. 1993). Second, the
In the context of purchase decision making, a typical
RA allows consumers to specify minimum acceptable
two-stage process may unfold as follows. First, the
consumer screens a large set of relevant products, attribute levels, and only alternatives that meet all such
without examining any of them in great depth, and specifications are included in the personalized list. This
identifies a subset that includes the most promising corresponds to an automated implementation of a con-
alternatives. Subsequently, s/he evaluates the latter in junctive decision rule (see Wright 1975). Finally, the
more depth, performs comparisons across products on RA allows shoppers to impose a quota cut-off
important attributes, and makes a purchase decision. (Feinberg and Huber 1996), i.e., to limit the number of
Given the different tasks to be performed in the course products to be included in the list.
of such two-stage purchase decision processes, inter- Comparison Matrix: A Tool for Organizing Product
active tools that provide support to consumers in the Information
following two respects seem particularly valuable: (1) The second decision aid we examine, a comparison ma-
the initial screening of available products to determine trix (CM), is conceptualized as an interactive tool that
which ones are worth considering further, and (2) the
assists consumers in making in-depth comparisons
in-depth comparison of selected products before mak-
ing the actual purchase decision. We focus on two de- 1
From there, detailed information about a product may be requested
cision aids, each designed to assist consumers in per- by clicking on its model name.
forming one of these key tasks. The two interactive 2
These scores are imperfect, approximate indicators of an alterna-
tools are discussed in turn. tive’s (unknown) true utility to a consumer.

Marketing Science/Vol. 19, No. 1, Winter 2000 7

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

among those alternatives that appear most promising Consideration set is conceptualized as the set of alter-
based on the initial screening. The CM allows shoppers natives that a consumer considers seriously for pur-
to organize attribute information about multiple prod- chase (Hauser and Wernerfelt 1990).3 We use both the
ucts. Very basic forms of this type of decision aid, usu- size and the quality of this set as dependent variables.
ally referred to as a shopping cart or basket, are im- The former is simply the number of products consid-
plemented on many online retail sites (e.g., ered seriously, which can be viewed as an indicator of
www.amazon.com, www.shopping.com). Most of a shopper’s relative product uncertainty when making
these tools do not currently allow for side-by-side com- a purchase decision. Consideration set quality is con-
parisons of products in terms of their attributes. How- ceptualized as the share of considered products that
ever, one real-world comparison aid that does closely are “non-dominated,” i.e., not objectively inferior to
match the above definition of CM is available on any alternative (see the Method section for details).
CompareNet’s site (compare.net). Decision quality is measured using both objective and
The CM used in the present study is implemented subjective indicators. This concept can be defined by
as an interactive display format in which product in- basic principles of coherence, such as not selecting
formation is presented in an alternatives (rows) ⳯ at- dominated alternatives (Payne et al. 1993). Thus, one
tributes (columns) matrix. It is designed to enable indicator of objective decision quality is whether or not
shoppers to compare products more efficiently and ac- a consumer purchases a nondominated alternative.4 Our
curately. While viewing detailed information about an second measure of objective decision quality is
alternative in the online shopping environment, a con- whether or not a shopper, after making a purchase de-
sumer can choose to have the product added to his/ cision, changes his/her mind and switches to another
her personal CM. (Once included, alternatives may alternative when given an opportunity to do so. Switch-
also be deleted from the CM.) The display format is ing indicates poor initial decision quality (see the
interactive in that a shopper can have all products in Method section for details). Finally, subjective decision
the CM sorted by any attribute. Use of this decision aid quality is conceptualized as the consumer’s degree of
should result in a shift in emphasis from memory- confidence in a purchase decision.
based to stimulus-based purchase decisions in the
sense that retaining specific attribute information 3
An alternative conceptualization of consideration set is to view it as
about relevant alternatives in memory becomes less a dynamic construct that evolves over time as products are being
important (see Alba et al. 1997). added to and dropped from the set (e.g., Nedungadi 1990). Within
this dynamic framework, our conceptualization of consideration set
corresponds to the final consideration set, i.e., the set of alternatives
considered at the time the actual purchase decision is made.
Hypotheses 4
Measuring the quality of purchase decisions and consideration sets
is a very ambitious endeavor. In this context, quality is conceptual-
Dependent Variables ized as the degree of match or fit between heterogeneous consumer
We expect that use of the RA and the CM will have an preferences and differentiated products. Because an individual’s
impact on three general aspects of consumer decision preferences are not subject to direct observation, it is impossible to
making in an online shopping environment: (1) accurately measure decision quality in uncontrolled real-world set-
tings. The measurement approach used in the present study is based
amount of information search, (2) consideration sets,
on the idea of an objective standard for quality and requires a com-
and (3) decision quality. bination of objectively dominated and nondominated alternatives. The
Amount of search for product information is conceptu- sets of available products were constructed in such a way that, ir-
alized as the number of alternatives for which detailed respective of an individual’s utility function, the purchase of partic-
information is acquired (Moorthy et al. 1997). In our ular alternatives indicates (with certainty) that s/he made a poor
decision. Choices of dominated alternatives in our controlled study
study, this coincides with the number of pages con-
are the equivalent of real-world purchase decisions that are subop-
taining attribute information about a particular prod- timal given an individual’s utility function at the time of purchase
uct that are viewed. This is an indicator of the effort and the set of available products, irrespective of whether or not any
an individual expends to screen available alternatives. of the alternatives are objectively dominated.

8 Marketing Science/Vol. 19, No. 1, Winter 2000

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

We present a set of hypotheses about how these six product information. As a result, the marginal benefits
aspects of consumer decision making are affected by of including additional products in the consideration
use of each of the two interactive decision aids, RA and set diminishes much more rapidly than in a situation
CM. All hypotheses are stated in terms of the expected where the consumer has no prior information about
difference, everything else being equal,5 between a sce- the relative utility of alternatives. Therefore, we expect
nario in which one of these tools is used in an online that individuals who use the RA will have smaller con-
shopping encounter and a case where it is not. The base sideration sets than those who do not.
case in which neither tool is available to a shopper cor-
Hypothesis H2. Use of the recommendation agent leads
responds to a typical, “bare-bones” online store.
to a reduction in the number of alternatives considered se-
Effects of Using the Recommendation Agent riously for purchase.

Amount of Search for Product Information. The Consideration Set Quality. Because the RA uses
amount of search for product information is deter- self-explicated attribute importance weights and min-
mined by consumers’ uncertainty about the absolute imum acceptable attribute levels to produce a person-
utility associated with an alternative and about the alized list of recommended alternatives, the products
relative utility of alternatives in a set (Moorthy et al. with the highest subjective utility will tend to appear
1997, Ratchford and Srinivasan 1993). In an online towards the top of a shopper’s list. Therefore, consum-
shopping environment, the amount of information ers should be less likely to consider inferior alterna-
search is also dependent upon the consumer’s ability tives for purchase. In addition, consideration sets are
to screen information effectively (Alba et al. 1997, more likely to be composed of products with similar
Bakos 1997). Because the RA automatically sorts avail- utility values than products with dissimilar ones
able products based on criteria provided by the shop- (Lehmann and Pan 1994). Therefore, we expect that the
per, the latter is better able to determine the relative share of alternatives included in the consideration set
utility of alternatives, and this should, in turn, lead to that are nondominated will be greater when the RA is
a reduction in the amount of search (Moorthy et al. used than when it is not.
1997). Thus, we hypothesize that individuals who have
Hypothesis H3. Use of the recommendation agent leads
this tool to assist them in their shopping task will view
to a larger share of nondominated alternatives in the set of
attribute information about fewer products than those
alternatives considered seriously for purchase.
who do not.
Decision Quality. The RA enables shoppers to
Hypothesis H1. Use of the recommendation agent leads
screen products using complex decision rules with
to a reduction in the number of alternatives for which de-
very low effort. Research on decision support systems
tailed product information is viewed.
indicates that decision aids designed to screen large
Consideration Set Size. Models of consideration numbers of alternatives may reduce decision makers’
set size are typically based on the notion of a trade-off cognitive effort (Todd and Benbasat 1994) and improve
between the marginal benefits and costs of considering decision quality by enabling individuals to make com-
an additional alternative (e.g., Hauser and Wernerfelt plex decisions with high accuracy (Singh and Ginzberg
1990, Roberts and Lattin 1991). These models assume 1996). By applying decision rules in an automated
that a product’s utility is unknown to the consumer fashion, such tools can reduce the amount of superflu-
prior to evaluation. However, because the RA screens ous information to be processed and, thus, augment
and ranks alternatives based on consumer-specified human information processing capabilities. In addi-
criteria, it provides information about the relative util- tion, the ability to screen alternatives in an efficient
ity of available products prior to processing specific manner enhances the “quality” of the information that
is processed, which, combined with reduced informa-
5
For brevity, we suppress the statement “ceteris paribus” in all our tion quantity, should have a positive impact on deci-
hypotheses. sion quality (Keller and Staelin 1987, 1989). Finally,

Marketing Science/Vol. 19, No. 1, Winter 2000 9

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

Widing and Talarzyk’s (1993) findings suggest that accurately eliminate unwanted products from their
electronic decision formats based on weighted average consideration set. Decision aids that help organize in-
scores for alternatives lead to less switching after initial formation have been found to reduce the number of
choice. Thus, we hypothesize that consumers’ use of alternatives considered by decision makers (Goslar et
the RA will have the following effects on the three in- al. 1986). The CM improves consumers’ ability to both
dicators of decision quality. determine their personal efficient frontiers and identify
dominated alternatives (Winer et al. 1997). As a result,
Hypothesis H4. Use of the recommendation agent leads
use of the CM reduces relative product uncertainty
to an increased probability of a nondominated alternative
and, thus, the marginal benefit of including an addi-
being selected for purchase.
tional product in the consideration set (Hauser and
Hypothesis H5. Use of the recommendation agent leads Wernerfelt 1990, Roberts and Lattin 1991). Thus, we
to a reduced probability of switching to another alternative hypothesize that consumers who use the CM will se-
(after making the initial purchase decision). riously consider fewer alternatives for purchase than
Hypothesis H6. Use of the recommendation agent leads those who do not.
to a higher degree of confidence in purchase decisions. Hypothesis H8. Use of the comparison matrix leads to
Effects of Using the Comparison Matrix (CM) a reduction in the number of alternatives considered seri-
ously for purchase.
Amount of Search for Product Information.
Through its capability for organizing information, the Consideration Set Quality. The CM’s alternatives
CM allows consumers to more efficiently compare and ⳯ attributes format facilitates side-by-side compari-
determine the relative attractiveness of alternatives. sons of products in terms of their attributes. This dis-
When searching for detailed product information, play format, in conjunction with the CM’s capability
shoppers who have access to the CM will anticipate for sorting all selected alternatives by any attribute,
being able to subsequently use this tool to make ac- reduces the demand on memory and improves con-
curate side-by-side comparisons of products and, sumers’ ability to identify suboptimal alternatives (see
therefore, tend to initially acquire information about a Payne et al. 1993, Winer et al. 1997). Because the CM
larger number of alternatives. If a product appears at- allows for efficient discrimination between products
tractive at first glance, the consumer can add it to the with respect to their subjective overall utility, it will
CM, evaluate it in direct comparison with other alter- render consumers less likely to either eliminate excel-
natives, and then decide whether or not to retain it in lent alternatives from or retain inferior alternatives in
the matrix. Because the CM facilitates stimulus-based, their consideration set. Therefore, we expect that con-
as opposed to memory-based, comparisons (see Alba sideration set quality will be higher for shoppers who
et al. 1997), it reduces the combined marginal cost of use the CM than for those who do not.
acquiring and processing attribute information about Hypothesis H9. Use of the comparison matrix leads to
an alternative. Therefore, we expect that individuals a larger share of nondominated alternatives in the set of al-
who use this tool will view information about more ternatives considered seriously for purchase.
products than those who do not.
Decision Quality. The findings of numerous stud-
Hypothesis H7. Use of the comparison matrix leads to
ies suggest that the way in which information is dis-
an increase in the number of alternatives for which detailed
played influences decision processes by affecting the
product information is viewed.
ease of carrying out different processing operations
Consideration Set Size. While availability of the (see Kleinmuntz and Schkade 1993). Because decision
CM should increase the amount of search (see H7), we makers generally try to conserve cognitive effort, they
expect that once consumers actually take advantage tend to use processing strategies that are facilitated by
of this decision aid’s comparison-facilitating capabili- a given display format (e.g., Russo 1977). The CM en-
ties, they will be able to more quickly and more hances consumers’ ability to compare products in

10 Marketing Science/Vol. 19, No. 1, Winter 2000

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

terms of their attributes (Alba et al. 1997). As a result, alternatives were constructed (9 models for each of 6
use of this tool should lead to a shift in emphasis from brands). Six actual brand names were used in each
memory-based to stimulus-based choice. The latter has product category. All model names were fictitious but
been found to result in a reduced probability of an in- representative of the respective category. Each alter-
ferior alternative being chosen (Muthukrishnan 1995, native was described on seven attributes in addition to
Exp. 2). In addition, information display formats that brand and model name. Five of these attributes were
reduce task difficulty have been found to lower the varied systematically across alternatives, while two at-
frequency of preference reversals (Johnson et al. 1988). tributes were held constant. The tent attributes that
The latter may be viewed as indicators of suboptimal were varied are (number of levels in parentheses): pole
choice. Thus, we hypothesize that use of the CM will material (3), warranty (3), weight (12), durability rating
have the following effects on the three indicators of (12), and price (12). Fly fabric and vestibule were held
decision quality. constant across alternatives. For stereos, the varied at-
Hypothesis H10. Use of the comparison matrix results tributes are: CD player type (3), tuner presets (3), out-
in an increased probability of a nondominated alternative put power (12), sound quality rating (12), and price
being selected for purchase. (12). Cassette decks and remote control were held
constant.6
Hypothesis H11. Use of the comparison matrix leads to
The measurement of consideration set quality and of
a reduced probability of switching to another alternative (af-
two aspects of decision quality requires alternatives
ter making the initial purchase decision).
that are known to be nondominated. For each product
Hypothesis H12. Use of the comparison matrix leads to category, six nondominated alternatives—one for each
a higher degree of confidence in purchase decisions. brand—were constructed. That is, 6 of the 54 products
were mutually nondominated. They did, however,
dominate all remaining models.7 Having one nondom-
Method inated alternative for each brand guaranteed that, ir-
A controlled experiment was conducted to test the
respective of an individual’s relative preference for
above hypotheses about the effects of the RA and the
brand names, one of the nondominated products was
CM on the six dependent measures of interest. The
the single most preferred alternative. The six nondom-
main task consisted of shopping for and making a pur-
inated alternatives were constructed by first assigning
chase of a product in each of two categories—back-
to them the best level of the two three-level attributes.
packing tents and compact stereo systems—in an on-
Next, for the three attributes with 12 levels, all six were
line store. In this section, we discuss (1) the
assigned the best level of one, the second best of an-
experimental design of the study, (2) the modeling ap-
proach, (3) the sample and incentive, and (4) the ex- other, and the third best of the remaining attribute. All
perimental procedure. possible combinations of first, second, and third best
were used. The two best levels of the 12-level attributes
Experimental Design were reserved for the nondominated alternatives. The
A 24 full-factorial experimental design was used. The remaining 48 products were constructed by means of
manipulated factors are: RA (yes, no), CM (yes, no),
product category (backpacking tent, compact stereo 6
The exact descriptions of all 54 models of backpacking tents are
system), and product category order (tent first, stereo provided in Appendix A. The corresponding information for com-
first). While product category is a within-subjects fac- pact stereo systems is available from the authors upon request.
7
tor, RA, CM, and order were manipulated between An alternative is dominated if there is at least one other alternative
that is superior on at least one attribute while not being inferior on
subjects. Respondents were randomly assigned to one
any attribute. That is, a dominated alternative is known to be within
of the eight conditions of the 23 between-subjects the efficient frontier of any consumer. By contrast, an alternative is
subdesign. nondominated if no other alternative is superior on an attribute with-
For each of the two product categories, a total of 54 out, at the same time, being inferior on at least one other attribute.

Marketing Science/Vol. 19, No. 1, Winter 2000 11

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

an iterative algorithm that approximated a target ma- a hierarchical sampling pattern or are otherwise clus-
trix of plausible across-attribute correlations, while ad- tered by design (Liang and Zeger 1986). Because prod-
hering to the required pattern of (non-)dominance uct category is a within-subject factor in our experi-
among alternatives. ment, the responses for each dependent variable are
clustered (by respondent) rather than independent.
Modeling Approach Thus, the ability to account for systematic relationships
We use Generalized Estimating Equations (GEE) models among multiple observations for an individual is es-
(Diggle et al. 1995, Liang and Zeger 1986) to test our sential to our study. For each dependent variable, a
hypotheses. This modeling technology generalizes “working correlation” between the two responses of
classical linear models in two ways, both of which are an individual is estimated as a free parameter (Zeger
essential to our study. We briefly discuss each of these and Liang 1986). This is required for an adequate mod-
extensions in turn. eling of the data, although the working correlations are
First, GEE models can accommodate a variety of re- not of substantive interest here.
sponse distributions in addition to the common nor- Two of our dependent measures—amount of search
mal (Gaussian) distribution. Given the different re- and consideration set size—are based on count data
sponse types used in the present study, this capability with no effective upper limit and are properly treated
is required for the proper modeling of our dependent as following a Poisson distribution. In our GEE mod-
variables. Within the GEE framework, the relationship els, this is implemented by specifying g(l) ⳱ log(l) as
between a dependent variable and a set of predictors the link function and V(l) ⳱ l as the variance function,
is expressed as where l ⳱ E(Y | x). Two other dependent variables—
p choice of a nondominated alternative and switching—
g(E(Y|x)) ⳱ b0 Ⳮ 兺 bixi, (1) are binary. In addition, consideration set quality is
i⳱1 measured as a fraction based on a set of binary re-
sponses. A binomial distribution is adequate for these
where Y is the dependent variable and x ⳱ (x1, . . . ,
three response variables. This is handled by using g(l)
xp) are the values of a set of predictor variables X1, . . . ,
⳱ log(l/(1 ⳮ l)) and V(l) ⳱ l(1 ⳮ l) as the GEE
Xp. The intercept b0 and a coefficient bi for each pre-
model’s link and variance functions, respectively. Fi-
dictor are estimated. The link function g, which may be
nally, confidence in purchase decisions is measured on
any monotonic differentiable function, allows nonlin-
a nine-point rating scale and can thus be treated as
ear relationships between predictor and outcome vari-
standard Gaussian (i.e., using a GEE model with iden-
ables (McCullagh and Nelder 1989).
tity link function and constant variance).
In addition, GEE models relax the assumption that
responses have independent distributions with con- Sample and Incentive
stant variance. In particular, the variance of the depen- Eighty undergraduate psychology students partici-
dent variable can be specified as a function of the mean pated (for partial course credit) in a pilot study aimed
response E(Y | x) via a variance function V such that at testing the validity of the experimental manipula-
tions. All manipulations were successful, and no tech-
var(Y) ⳱ V(E(Y|x))␾, (2) nical problems were encountered with repsect to the
electronic shopping environment. For the main study,
where ␾ is a scale parameter. The link and variance
249 undergraduate business students completed the
functions allow a wide range of non-normal response
online shopping task for both product categories.8 In
distributions, including binomial, Poisson, and gamma
addition to partial course credit, subjects in the main
(Zeger and Liang 1986).
study participated in a lottery designed to increase the
The second generalization of classical linear models
reflected in GEE models is the relaxation of the as- 8
The number of respondents per cell in the 23 between-subjects sub-
sumption of independence among observations, which design ranged from 29 to 33. All subjects had prior experience using
allows a more adequate modeling of data that follow the WWW, as did all participants in the pilot study.

12 Marketing Science/Vol. 19, No. 1, Winter 2000

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

validity of the findings by making the shopping task No-CM conditions made their purchase from one of
more consequential. Prior to entering the electronic the individual product pages. In all conditions, re-
shopping environment, subjects were informed that spondents were asked to confirm the product of their
two randomly selected winners were to receive one of choice before the purchase was finalized.
the two products they “purchased” during their shop- After subjects made their first purchase, a measure
ping trip, plus the difference between $500 and the of confidence in their purchase decision (“How confi-
price of that product in cash. One tent and one stereo dent are you that the product you just purchased is
system were dispensed. Because the alternatives used really the best choice for you?”) was obtained using a
in the study were constructed (see above), the two win- nine-point rating scale. Next, they were presented with
ners received the real-world model that best matched the list of alternatives and asked to report their consid-
their chosen alternative (i.e., same brand name and eration set (“Please indicate which of the products you
similar attribute levels). Each prize’s cash component considered seriously before making your purchase de-
was based on the chosen model’s price used in the cision.”). Subjects could then switch from the pur-
shopping environment. chased alternative to each of six (five) nondominated
Procedure alternatives.10 This switching task was presented as a
Data were collected in a university computer lab in series of pairwise comparisons in which complete de-
sessions of 15 to 20 subjects. Upon arrival, participants scriptions of both products were displayed side by
were assigned to a personal computer and informed side.11 Subjects were encouraged to switch whenever
that they would be pilot-testing a new online store by they saw an alternative they preferred over their initial
shopping for a product in each of two categories. The choice. They were informed that the lottery winners
experimenter then held a 10-minute practice session would receive whatever product they had “in their
aimed at demonstrating the features of the shopping basket” after the switching task. The same procedure
environment. was repeated for the second product category. After
Before shopping for the first product, subjects were that, subjects completed a questionnaire containing
asked to rate their level of product category knowledge manipulation checks. Finally, the administrator de-
and interest (using nine-point rating scales). They then briefed the participants and concluded the session.
read a description of the task and of the lottery incen-
tive. Subjects in the No-RA conditions were taken to a Results
hierarchically structured website with all six brands Manipulation Checks
listed at the top level and all models for a brand listed To verify that the experimental manipulations were
at the lower level. Subjects accessed detailed informa- successful, subjects responded to manipulation-check
tion about a product by first clicking on a brand name questions after completion of their second shopping
and then on a model name. In the conditions in which
trip. First, they expressed (using a nine-point rating
the RA was available,9 subjects started by providing
scale) how difficult it was for them to locate the prod-
attribute importance weights using a 100-point con-
ucts that best matched their personal preferences. This
stant sum scale, minimum-acceptable attribute levels,
was used to check the RA manipulation. The mean rat-
and the maximum number of alternatives to be in-
ings obtained from the RA and No-RA conditions are
cluded in their personalized recommendation list.
2.95 and 4.67, respectively. This difference in means is
From that list, they were able to request detailed in-
highly significant (p ⬍ 0.001, x2 ⳱ 0.178) and in the
formation about a particular product. Subjects who
used the CM were able to add the attribute information 10
The nondominated products used for this purpose were identical
displayed on a product’s page to the CM, from where to the ones used during the shopping task. The number of switching
they eventually made their purchase. Subjects in the opportunities depended upon whether a subject had initially chosen
a dominated (6) or nondominated alternative (5).
9 11
In the RA conditions, all subjects actually used this decision aid. The This switching task is similar to a method used by Widing and
same is true for the CM conditions. Talarzyk (1993).

Marketing Science/Vol. 19, No. 1, Winter 2000 13

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

Table 1 Model Results: Coefficient Estimates

Dependent Variables

Amount Consideration Consideration Purchase of Confidence in

Predictors of Search Set Size Set Quality Nondominated Alternative Switching Purchase Decision

Intercept 2.158a 1.054 0.469 1.686 ⳮ0.441 6.554

(0.032) (0.027) (0.098) (0.182) (0.109) (0.086)
*** *** *** *** *** ***
Recommendation ⳮ0.574 ⳮ0.094 2.362 2.086 ⳮ1.728 0.314
Agent (RA) (0.063) (0.054) (0.199) (0.359) (0.218) (0.173)
*** * *** *** *** *
Comparison 0.022 ⳮ0.543 0.970 0.497 ⳮ0.452 ⳮ0.068
Matrix (CM) (0.063) (0.054) (0.195) (0.358) (0.218) (0.173)
*** *** *
RA ⳯ CM ⳮ0.191 0.282 0.455 1.114 ⳮ0.094 ⳮ0.519
Interaction (0.126) (0.109) (0.391) (0.717) (0.436) (0.346)
**
Product 0.109 0.085 ⳮ0.487 ⳮ0.569 0.013 0.849
Category (0.035) (0.042) (0.106) (0.192) (0.185) (0.116)
** * *** xf.** ****
Order 0.284 ⳮ0.076 0.383 0.212 ⳮ0.177 ⳮ0.291
Position (0.035) (0.043) (0.107) (0.192) (0.185) (0.116)
*** *** *

a
Cell Format: Level of Significance:
Coefficient Estimate * denotes significance at 0.05 level
(Standard error) ** denotes significance at 0.01 level
具t value典 *** denotes significance at 0.001 level
Level of Significance

intended direction. The CM manipulation was checked these models: main effects for RA, CM, product cate-
by asking subjects to rate (on a nine-point rating scale) gory, and order position, plus an RA ⳯ CM interaction
how difficult it was for them to compare different effect. The RA and CM main effects are of primary
products. The mean ratings obtained from the CM and substantive interest. Product category, order position,
No-CM conditions are 2.80 and 4.44, respectively. This and the RA ⳯ CM interaction are included for an ad-
difference in means is highly significant (p ⬍ 0.001, x2 equate representation of the data and for exploratory
⳱ 0.176) and in the intended direction. We conclude purposes. The two levels of RA and CM were coded
that both of our manipulations were successful. ⳮ0.5 (not used) and 0.5 (used). Product category was
coded ⳮ0.5 (tent) and 0.5 (stereo), and order position
Hypothesis Tests was coded ⳮ0.5 (first product category) and 0.5 (sec-
To test the hypotheses regarding the effects of the RA ond product category).13 The RA ⳯ CM interaction
and CM, a GEE model was estimated for each of the
six dependent variables.12 In addition to an intercept,
the following predictor variables were included in
brary function developed by V. J. Carey, Department of Biostatistics,
12
All GEE models were estimated using the S-Plus statistical analysis Harvard University.
13
and programming environment (MathSoft, Inc. 1998) and a GEE li- In models that include interaction terms in the form of products of

14 Marketing Science/Vol. 19, No. 1, Winter 2000

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

was treated as the product of the RA and CM predictor Figure 1 Effects of RA and CM on Amount of Search Within Online
variables. Store
Table 1 provides an overview of the model results.
Each column contains the GEE coefficient estimates
with respect to one of the six dependent variables.
Standard errors are in parentheses. The level of statis-
tical significance of a coefficient is indicated by aster-
isks.14 The tests of our hypotheses are based on the
coefficients for the RA and CM main effects. Figures 1
through 6 contain the cell means, percentages, or ratios
of the dependent variables as a function of whether or
not the RA and CM were used. We discuss the effects
of the RA and the CM in turn.
The effect of the RA on the amount of search for
product information is highly significant (p ⬍ 0.001)
and in the expected direction. As shown in Figure 1,
subjects viewed detailed product information for sub-
stantially fewer alternatives when the RA was used Figure 2 Effects of RA and CM on Consideration Set Size
(6.58 on average) than when it was not (11.78). This
provides strong support for H1. As expected, use of
the RA also led to smaller consideration sets. The av-
erage number of alternatives considered seriously for
purchase was 2.78 in the RA conditions and slightly
above 3 in the No-RA conditions (See Figure 2). This
effect is significant at p ⬍ 0.05 and supports H2. While
reducing consideration set size, use of the RA resulted
in a drastic increase in consideration set quality. The
average share of the alternatives considered seriously
for purchase that were nondominated was 0.85 when
the RA was used and 0.42 when it was not (see Figure
3). This effect is highly significant (p ⬍ 0.001) and pro-
vides strong support for H3.
Use of the RA had the following effects on the three

Figure 3 Effects of RA and CM on Consideration Set Quality

categorical predictor variables, the coefficient estimate and statistical
test for a predictor that is included in an interaction term are not
invariant to the coding of other predictors included in the same in-
teraction term (see Irwin and McClelland 2000). We use standard-
ized mean-centered coding for all main effects. As a result of the
mean-centering, all main and interaction effects (i.e., coefficients) are
with respect to the average of the two levels of other factors. As a
result of standardization, all main effects are expressed in terms of
the difference between the two levels of a factor.
14
Because we are testing directional hypotheses regarding the effects
of the RA and CM, the level of significance of these effects is based
on one-tailed t values. For all other effects, two-tailed t values are
used.

Marketing Science/Vol. 19, No. 1, Winter 2000 15

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

Figure 4 Effects of RA and CM on Purchase of Nondominated Alter- measures of decision quality.15 The share of subjects
native who purchased a nondominated alternative was about
93% when the RA was used, and about 65% when it
was not (see Figure 4). This highly significant result (p
⬍ 0.001) supports H4. The proportion of subjects who
switched to another alternative when given an oppor-
tunity to do so during the post-purchase switching
task was only slightly above 20% in the RA conditions
but about 60% among those subjects who did not use
the RA (see Figure 5). This effect is also highly signifi-
cant (p ⬍ 0.001) and provides strong support for H5.
Finally, use of the RA resulted in a significant (p ⬍
0.05) improvement in consumers’ confidence in their
choice (means of 6.71 vs. 6.41) as predicted by H6 (see
Figure 6). This effect of the RA on the subjective mea-
sure of decision quality is noticeably weaker than its
effect on the two objective measures. In sum, use of the
Figure 5 Effects of RA and CM on Switching During Post-Purchase
Switching Task RA reduces search effort for product information, de-
creases the size but increases the quality of consider-
ation sets, and improves the quality of purchase
decisions.
We now turn to the effects of the CM. As predicted,
use of the CM led to an increase in the average number
of alternatives for which subjects viewed detailed
product information. However, this effect is not sig-
nificant (p ⬎ 0.1), and thus H7 is not supported. By
contrast, the CM’s impact on consideration set size is
very substantial. The average number of alternatives
considered seriously for purchase was 2.19 when re-
spondents used the CM and 3.77 when they did not
(see Figure 2). This effect is highly significant (p ⬍
0.001) and provides strong support for H8. In addition,
the average share of the alternatives considered for
Figure 6 Effects of RA and CM on Confidence in Purchase Decision purchase that were nondominated was about 0.68
when the CM was used and about 0.57 otherwise (see

15
The pairwise Spearman correlation coefficients among these three
measures are q ⳱ ⳮ0.44 (p ⬍ 0.001) between purchase of nondom-
inated alternative and switching, q ⳱ 0.09 (p ⬍ 0.05) between pur-
chase of nondominated alternative and confidence, and q ⳱ ⳮ0.12
(p ⬍ 0.01) between switching and confidence. All three correlations
are in the expected direction. The fact that the pairwise correlations
between each of the two objective measures and the subjective one
are of small magnitude shows that the two types of indicators are
not merely redundant with each other and that obtaining objective
measures of decision quality provides valuable insight above and
beyond what subjective indicators may reveal.

16 Marketing Science/Vol. 19, No. 1, Winter 2000

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

Figure 3). This favorable effect of the CM on consid- and tested in the form of interaction terms between
eration set quality is also highly significant (p ⬍ 0.001), RA, CM, or the RA ⳯ CM interaction on one hand and
thus supporting H9. a potential moderating variable on the other. Each of
Consistent with H10, use of the CM led to an in- these terms was included in a separate model that also
crease in the share of subjects who purchased a non- contained main effects for RA, CM, product category,
dominated alternative. This effect is only marginally and order position, as well as an RA ⳯ CM interaction
significant (p ⬍ 0.1). As predicted by H11, the propor- effect. None of these moderating effects were statisti-
tion of subjects who switched to another alternative cally significant (at p ⳱ 0.05) with respect to any of the
when given an opportunity to do so during the post- six dependent variables. This suggests that the gener-
purchase switching task was lower when the CM was alizability of the substantive findings across product
used (38%) than when it was not (44%) (see Figure 5). categories is satisfactory.
This positive effect of the CM on decision quality is
significant at p ⬍ 0.05. Finally, we do not find any sup-
port for H12—the CM did not reliably affect subjects’ Discussion
confidence in their choice. In sum, use of the CM leads A characteristic feature of electronic shopping environ-
to a decrease in the size but an increase in the quality ments is the lack of physical constraints with respect
of consideration sets, and it tends to have a favorable to product display. The virtually infinite “shelfspace”
effect on objective decision quality. available in online stores allows vendors to offer an
Other Results extremely large number of alternatives within a prod-
In addition to the tests of our hypotheses, several other uct category. From a consumer perspective, having ac-
results are of interest. First, interaction effects between cess to a very large number of products is highly de-
RA and CM were included in the models discussed sirable. At the same time, however, consumers have
above. For one of the six dependent variables, consid- limited cognitive resources and may simply be unable
eration set size, this interaction is statistically signifi- to process the potentially vast amounts of information
cant (p ⬍ 0.01). Follow-up tests suggest that use of the about these alternatives. A potential solution to this
RA leads to a reduction in consideration set size when dilemma is to provide consumers with sophisticated
the CM is not available (p ⬍ 0.01) but has no such effect interactive decision aids designed to help them effec-
when the CM is available (p ⬎ 0.7). tively manage and capitalize on the enormous
For an adequate representation of the data, main ef- amounts of product information that may be available
fects for product category and order position were in- in electronic shopping environments.
cluded in the models. While we have no substantive The objective of the present study was to examine
interest in these two factors, we note that each affects the effects of such interactive decision aids on various
some of the dependent variables. Product category ap- aspects of consumer decision making in an online
pears to have an impact on five of the outcomes. How- shopping context. In particular, we focused on two
ever, inclusion of subjects’ knowledge and interest tools that represent obvious choices given the well-
with respect to the category as predictors renders the established notion that consumers often reach pur-
effects of product category on all but two dependent chase decisions via a two-stage process. The RA assists
variables (consideration set quality, purchase of a non- consumers in the initial screening of alternatives, and
dominated alternative) statistically insignificant. The the CM facilitates in-depth comparisons of selected al-
effects of order position may reflect increasing famil- ternatives that are considered seriously for purchase.
iarity with the shopping environment over time. The results of our study indicate that use of these tools
To examine whether the effects of the RA and CM has a substantial impact on the amount of search for
were moderated by product category, order position, product information, the size and quality of shoppers’
knowledge, or interest, a number of additional GEE consideration sets, and the quality of their purchase
models were estimated for each of the dependent vari- decisions.
ables. The moderating relationships were expressed In a nutshell, the two interactive decision aids allow

Marketing Science/Vol. 19, No. 1, Winter 2000 17

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

consumers to make much better decisions while expend- substantial imperfections. For example, they may ne-
ing substantially less effort. Given the well-established glect relevant attributes, overlook some attractive al-
notion that a trade-off between effort and accuracy is ternatives entirely, or even be systematically biased in
inherent to human decision making in traditional en- favor of a subset of products (e.g., those of a certain
vironments (Payne et al. 1993), it is interesting that brand). Therefore, we do not conclude that recommen-
tools like the RA and the CM can simultaneously in- dation agents will always and unconditionally lead to
crease decision quality and reduce effort. The findings desirable outcomes for consumers. Rather, the effects
of our study show how drastically interactive decision of the RA found in the present study should be viewed
aids implemented in online shopping environments as a demonstration of the potential effects of typical,
may transform the way in which consumers search for well-functioning recommendation tools.
product information and make purchase decisions. While the present study provides valuable insights
The present study examines the effects of interactive into the effects of interactive decision aids on consumer
decision aids on consumer decision making in a par- decision making in online shopping environments,
ticular electronic shopping setting. How well do our further research will be needed to obtain a deeper un-
empirical results generalize to other types of online re- derstanding of these effects. In particular, an exami-
tail environments? While both the RA and the CM nation of potential moderators would be valuable. Fac-
were operationalized as decision aids available within tors that might moderate the effects reported here
an online retailer’s site, the relevance of our findings include the number of available alternatives, the
is not limited to interactive tools that are exclusive to amount of risk associated with a purchase, and con-
an individual merchant. In particular, the results apply sumers’ confidence in the integrity of the interactive
to all within- and cross-store decision aids that allow decision aids.
online shoppers direct access to a common database of In conclusion, the findings of the present study sug-
products and that do not discriminate between prod- gest that interactive decision aids designed to assist
ucts on the basis of which vendor they are associated consumers in the initial screening of available products
with. To the extent that such decision aids are imple- and to facilitate in-depth comparisons among selected
mented in the context of multi-retailer online malls, alternatives may have highly desirable properties in
cross-merchant comparison schemes, or groups of terms of consumer decision making. Such tools allow
stores that allow for unrestricted cross-store compari- shoppers to more easily detect products that are over-
sons, our results are of relevance. priced or otherwise dominated by competing alterna-
Three important boundary conditions with respect tives, thus increasing market efficiency. More gener-
to the present study’s findings should be made explicit. ally, the availability of interactive decision aids in
First, the focus of this research is on consumers’ goal- online shopping environments should enhance the
directed shopping behavior in an online environment, ability of individuals to identify products that match
rather than on exploratory navigation behavior. Thus, their personal preferences and, therefore, lead to sub-
no conclusions about the latter may be drawn based stantial positive welfare effects for consumers.16
on the results reported here. Second, our findings do
16
not pertain to situations in which a consumer has not This research was supported by a Nova Corporation Faculty Fel-
lowship awarded to the first author by the University of Alberta, a
yet selected an electronic store, online mall, or other
contribution from Macromedia Inc., as well as grants from Telus
entity with common product offerings. Hierarchical Communications Inc. and the University of Alberta through its So-
decision processes, such as first selecting one of a set cial Science Research Program. The authors thank Terry Elrod for
of competing online merchants and subsequently se- his valuable assistance, V. J. Carey for making available his S-Plus
function for fitting GEE models, Barry Ard for his programming
lecting a product from that merchant’s offerings,
assistance, and Adam Finn, Julie Irwin, as well as the Editor, the
should be examined in future research. Finally, the RA Area Editor, and two anonymous Marketing Science reviewers for
used in the present study is a high-quality decision aid. their insightful comments. Correspondence should be addressed to
Real-world recommendation systems may suffer from the first author.

18 Marketing Science/Vol. 19, No. 1, Winter 2000

Appendix A Product Descriptions: Backpacking Tents

Pole Material Weight Durability Rating Warranty Price

Brand Model (aluminum type) Fly Fabric (kilograms) (0 to 100 points) Vestibule (years) ($)

Eureka Adventurer regular-strength 2.3 oz. nylon 3.3 84 yes 2 329.99

Eureka Challenger regular-strength 2.3 oz. nylon 3.5 86 yes 2 324.99
Eureka Drifter high-strength 2.3 oz. nylon 3.3 89 yes 4 359.99
Eureka Huntsman regular-strength 2.3 oz. nylon 3.7 88 yes 3 324.99
Eureka Mountaineer regular-strength 2.3 oz. nylon 3.3 86 yes 3 324.99
Eureka Naturalist ultrahigh-strength 2.3 oz. nylon 3.3 88 yes 4 369.99
Eureka Outfitter high-strength 2.3 oz. nylon 3.6 89 yes 2 329.99
Eureka Traveler ultrahigh-strength 2.3 oz. nylon 3.3 92 yes 4 314.99
Eureka Wanderer high-strength 2.3 oz. nylon 3.5 86 yes 2 329.99
Kelty Glacier Lake ultrahigh-strength 2.3 oz. nylon 3.4 84 yes 2 324.99
Kelty Lakeside ultrahigh-strength 2.3 oz. nylon 4.0 89 yes 3 324.99
Kelty Mountain Springs high-strength 2.3 oz. nylon 4.1 91 yes 3 324.99
Kelty Oasis ultrahigh-strength 2.3 oz nylon 3.5 91 yes 4 369.99
Kelty Raging Tide ultrahigh-strength 2.3 oz. nylon 4.2 91 yes 3 324.99
Kelty River Rapid ultrahigh-strength 2.3 oz. nylon 3.9 85 yes 3 324.99
Kelty Seabreeze ultrahigh-strength 2.3 oz. nylon 3.1 91 yes 4 319.99
Kelty Swift Current ultrahigh-strength 2.3 oz. nylon 3.5 89 yes 4 354.99
Kelty Waterfall ultrahigh-strength 2.3 oz. nylon 3.5 82 yes 3 334.99
Outbound Galaxy ultrahigh-strength 2.3 oz. nylon 3.3 91 yes 2 359.99
Outbound Lunar Eclipse ultrahigh-strength 2.3 oz. nylon 3.9 83 yes 4 324.99
Outbound Moonscape ultrahigh-strength 2.3 oz. nylon 3.1 92 yes 4 324.99
Outbound Neptune ultrahigh-strength 2.3 oz. nylon 3.6 86 yes 3 329.99
Outbound North Star ultrahigh-strength 2.3 oz. nylon 3.3 86 yes 3 334.99
Outbound Skyline ultrahigh-strength 2.3 oz. nylon 3.3 84 yes 3 329.99
Outbound Stargazer ultrahigh-strength 2.3 oz. nylon 3.3 90 yes 3 354.99
Outbound Sunlight ultrahigh-strength 2.3 oz. nylon 3.3 89 yes 3 359.99
Outbound Westwind high-strength 2.3 oz. nylon 4.1 89 yes 2 324.99
REI Bear Paw regular-strength 2.3 oz. nylon 3.6 84 yes 2 324.99
REI Coyote high-strength 2.3 oz. nylon 4.2 87 yes 4 324.99
REI Eagle high-strength 2.3 oz. nylon 3.3 82 yes 4 334.99
REI Grizzly ultrahigh-strength 2.3 oz. nylon 4.2 86 yes 4 324.99
REI Mountain Lion ultrahigh-strength 2.3 oz. nylon 3.2 91 yes 4 314.99
REI Night Owl ultrahigh-strength 2.3 oz. nylon 3.3 91 yes 4 369.99
REI Raven ultrahigh-strength 2.3 oz. nylon 3.8 89 yes 4 344.99
REI Red Fox ultrahigh-strength 2.3 oz. nylon 3.3 91 yes 3 359.99
REI Timberwolf ultrahigh-strength 2.3 oz. nylon 3.6 91 yes 4 349.99
Sierra Designs Backtrail regular-strength 2.3 oz. nylon 3.3 84 yes 3 324.99
Sierra Designs Badlands regular-strength 2.3 oz. nylon 3.8 88 yes 3 324.99
Sierra Designs Big Country high-strength 2.3 oz. nylon 3.4 91 yes 4 359.99
Sierra Designs Forest Mist high-strength 2.3 oz. nylon 3.5 82 yes 2 329.99
Sierra Designs Landscape high-strength 2.3 oz. nylon 3.8 89 yes 2 329.99
Sierra Designs Mountain Range ultrahigh-strength 2.3 oz. nylon 3.3 93 yes 4 319.99
Sierra Designs Rocky Ridge ultrahigh-strength 2.3 oz. nylon 3.4 90 yes 4 369.99
Sierra Designs South Ridge regular-strength 2.3 oz. nylon 3.6 83 yes 3 324.99
Sierra Designs Summit regular-strength 2.3 oz. nylon 3.3 82 yes 2 324.99
Quest Daydream high-strength 2.3 oz. nylon 3.9 91 yes 2 324.99
Quest Freestyle regular-strength 2.3 oz. nylon 4.0 88 yes 4 329.99
Quest Great Escape high-strength 2.3 oz. nylon 3.3 88 yes 4 369.99
Quest Journey ultrahigh-strength 2.3 oz. nylon 3.2 93 yes 4 324.99
Quest Peacemaker ultrahigh-strength 2.3 oz. nylon 4.2 91 yes 4 324.99
Quest Serenity high-strength 2.3 oz. nylon 3.4 89 yes 4 349.99
Quest Solitude regular-strength 2.3 oz. nylon 3.3 84 yes 3 324.99
Quest Spirit ultrahigh-strength 2.3 oz. nylon 4.0 84 yes 4 324.99
Quest Tranquility ultrahigh-strength 2.3 oz. nylon 3.8 91 yes 4 359.99

Note: Boldface indicates nondominated alternatives.

Marketing Science/Vol. 19, No. 1, Winter 2000 19

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

References Kleinmuntz, Don N., David A. Schkade. 1993. Information displays

Alba, Joseph, John Lynch, Barton Weitz, Chris Janiszewski, Richard and decision processes. Psych. Sci. 4 221–227.
Lutz, Alan Sawyer, Stacey Wood. 1997. Interactive home shop- Lehmann, Donald R., Yigang Pan. 1994. Context effects, new brand
ping: consumer, retailer, and manufacturer incentives to par- entry, and consideration sets. J. Marketing Res. 31 (3) 364–374.
ticipate in electronic marketplaces. J. Marketing 61 (July) 38–53. Liang, Kung-Yee, Scott L. Zeger. 1986. Longitudinal data analysis
Ariely, Dan. 2000. Controlling the information flow: effects on con- using generalized linear models. Biometrika 73 (1) 13–22.
sumers’ decision making and preferences. J. Consumer Res. MathSoft, Inc. 1998. S-Plus 4.5. MathSoft, Seattle, WA.
Forthcoming. McCullagh, P., J. A. Nelder. 1989. Generalized Linear Models. Second
Bakos, J. Yannis. 1997. Reducing buyer search costs: implications for Edition, Chapman and Hall, London, UK.
electronic marketplaces. Management Sci. 43 (12) 1676–1692. Moorthy, Sridhar, Brian T. Ratchford, Debabrata Talukdar. 1997.
Beach, Lee R. 1993. Broadening the definition of decision making: Consumer information search revisited: theory and empirical
the role of prechoice screening of options. Psych. Sci. 4 (4) 215– analysis. J. Consumer Res. 23 (March) 263–277.
220. Muthukrishnan, A. V. 1995. Decision ambiguity and incumbent
Benbasat, Izak, Barrie R. Nault. 1990. An evaluation of empirical brand advantage. J. Consumer Res. 22 (1) 98–109.
research in managerial support systems. Decision Support Sys- Nedungadi, Prakash. 1990. Recall and consumer consideration sets:
tems 6 (2) 203–226. influencing choice without altering brand evaluations. J. Con-
Bettman, James R., Eric J. Johnson, John W. Payne. 1990. A compo- sumer Res. 17 (December) 263–276.
nential analysis of cognitive effort in choice. Organ. Behavior Payne, John W. 1982. Contingent decision behavior. Psych. Bull. 92
Human Decision Processes 45 111–139. 382–402.
Diggle, Peter J., Kung-Yee Liang, Scott L. Zeger. 1995. Analysis of ——, James R. Bettman, Eric J. Johnson. 1988. Adaptive strategy se-
Longitudinal Data. Clarendon Press, Oxford, UK. lection in decision making. J. Experiment. Psych.: Learning, Mem-
Einhorn, H., R. Hogarth. 1978. Confidence in judgment: persistence ory, and Cognition 14 534–552.
of the illusion of validity. Psych. Rev. 85 395–416. ——, ——, ——. 1993. The Adaptive Decision Maker. Cambridge Uni-
Feinberg, Fred M., Joel Huber. 1996. A theory of cutoff formation versity Press, Cambridge, UK.
under imperfect information. Management Sci. 42 (1) 65–84. Pearson, J. Michael, J. P. Shim. 1994. An empirical investigation into
Goslar, Martin D., Gary I. Green, Terry H. Hughes. 1986. Decision decision support systems capabilities: a proposed taxonomy.
support systems: an empirical assessment for decision making. Inform. Management 27 45–57.
Decision Sci. 17 79–91. Ratchford, Brian T., Narasimhan Srinivasan. 1993. An empirical in-
Guttman, Robert H., Alexandros G. Moukas, Pattie Maes. 1998. vestigation of returns to search. Marketing Sci. 12 (Winter) 73–
Agent-mediated electronic commerce: a survey. Knowledge 87.
Engrg. Rev. 13 (2) 147–160. Roberts, John H., James M. Lattin. 1991. Development and testing of
GVU. 1999. GVU’s WWW User Survey. Georgia Institute of Technol-
a model of consideration set composition. J. Marketing Res. 28
ogy: Graphic, Visualization, and Usability Center [http://
(November) 429–40.
www.gvu.gatech.edu/user surveys].
Russo, J. Edward. 1977. The value of unit price information. J. Mar-
Hauser, John R., Birger Wernerfelt. 1990. An evaluation cost model
keting Res. 14 (May) 193–201.
of consideration sets. J. Consumer Res. 16 (March) 393–408.
Shugan, Steven M. 1980. The cost of thinking. J. Consumer Res. 7
Hoch, Stephen J., David A. Schkade. 1996. A psychological approach
(September) 99–111.
to decision support systems. Management Sci. 42 (1) 51–64.
Simon, Herbert A. 1955. A behavioral model of rational choice.
Hoffman, Donna L., Thomas P. Novak. 1996. Marketing in hyper-
Quart. J. Econom. 69 99–118.
media computer-mediated environments: conceptual founda-
Singh, Danièle T., Michael J. Ginzberg. 1996. An empirical investi-
tions. J. Marketing 60 (July) 50–68.
Irwin, Julie R., Gary H. McClelland. 2000. Heuristics for moderated gation of the impact of process monitoring on computer-
regression models. J. Marketing Res. forthcoming. mediated decision making performance. Organ. Behavior Hu-
Johnson, Eric J., John W. Payne. 1985. Effort and accuracy in choice. man Decision Process. 67 (2) 156–169.
Management Sci. 31 394–414. Todd, Peter, Izak Benbasat. 1992. The use of information in decision
——, ——, James R. Bettman. 1988. Information displays and pref- making: an experimental investigation of the impact of
erence reversals. Organ. Behavior Human Decision Process. 42 1– computer-based decision aids. MIS Quart. 16 (September) 373–
21. 393.
Kasper, George M. 1996. A theory of decision support system design ——, ——. 1994. The influence of decision aids on choice strategies:
for user calibration. Inform. Systems Res. 7 (2) 215–232. an experimental analysis of the role of cognitive effort. Organ.
Keller, Kevin L., Richard Staelin. 1987. Effects of quality and quantity Behavior Human Decision Process. 60 36–74.
of information on decision effectiveness. J. Consumer Res. 14 Widing, Robert E. II, W. Wayne Talarzyk. 1993. Electronic informa-
200–213. tion systems for consumers: an evaluation of computer-assisted
——, ——. 1989. Assessing biases in measuring decision effective- formats in multiple decision environments. J. Marketing Res. 30
ness and information overload. J. Consumer Res. 15 504–508. (May) 125–141.

20 Marketing Science/Vol. 19, No. 1, Winter 2000

 HÄUBL AND TRIFTS
Consumer Decision Making in Online Shopping Environments

Winer, Russell S., John Deighton, Sunil Gupta, Eric J. Johnson, Zeger, Scott L., Kung-Yee Liang. 1986. Longitudinal data analysis for
Barbara Mellers, Vicki G. Morwitz, Thomas O’Guinn, Arvind discrete and continuous outcomes. Biometrics 42 121–130.
Rangaswamy, Alan G. Sawyer. 1997. Choice in computer- Zack, Michael H. 1993. Interactivity and communication mode
mediated environments. Marketing Lett. 8 (3) 287–296. choice in ongoing management groups. Inform. Systems Res. 4
Wright, Peter. 1975. Consumer choice strategies: simplifying vs. op- (3) 207–239.
timizing. J. Marketing Res. 12 (February) 60–67.

This paper was received May 13, 1998, and has been with the authors 9 months for 3 revisions; Processed by William Boulding.

Marketing Science/Vol. 19, No. 1, Winter 2000 21