Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

INFORMS

Brand Effects on Choice and Choice Set Formation under Uncertainty


Author(s): Joffre Swait and Tülin Erdem
Source: Marketing Science, Vol. 26, No. 5 (Sep. - Oct., 2007), pp. 679-697
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/40057089
Accessed: 17-10-2015 15:44 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Marketing Science.

http://www.jstor.org

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Marketing Science infjUHl
Vol.26, No. 5, September-October 2007,pp. 679-697
issn 0732-23991eissn 1526-548X107 1260510679 doi 10.1287/mksc.l060.0260
©2007 INFORMS

Brand Effects on Choice and Choice Set


Formation Under Uncertainty

JoffreSwait
Advanis Inc. and Universityof Alberta,Suite 1600,Sun Life Place, 1012399th Street,
Edmonton,Alberta,Canada,joffre_swait@advanis.ca
Tulin Erdem
SternSchool of Business,New YorkUniversity,New York,New York10012,terdem@stern.nyu.edu

paper examines the effects of brand credibility,a central concept in information economics-based
approachesto brand effects and brand equity, on consumer choice and choice set formation.We investi-
gate the mechanismsthroughwhich credibilityeffectsmaterialize,namely,throughperceivedquality,perceived
risk, and informationcosts saved. The credibilityof a brand as a signal is defined as the believabilityof the
product position informationcontained in a brand, which depends on consumer perceptionsof the willing-
ness and ability of firms to deliver what they have promised. The choice set is defined as the collection of
brands that have a nonzero probabilityof being chosen among those actually available for choice in a given
context.
Furthermore,we study the impact of brand credibilityon the varianceof the stochasticcomponentof utility.
Not only do choice model parameterscapturethe impact of systematicutility differenceson choice probabili-
ties, but also the magnitude of this systematicimpact is moderatedby the relative importanceof the stochas-
tic utility component in preference.We term this moderationphenomenon preference discrimination, which we
conceptualize as the decision makers' capacity to effectively discriminatebetween products' utilities in choice
situations.
We estimate a discretechoice model of brand choice set formationand preferencediscriminationon exper-
imental data in two categories- juice and personal computers- and find strong evidence for brand credibility
effectsand differentialmechanismsthroughwhich brand credibility'simpact materializeson brand choice con-
ditional on choice set, choice set formation,and preferencediscrimination.
Keywords:informationeconomics;perceived quality;perceivedrisk;brandpreference;branding;brandchoice;
choice models; personalcomputers;juice
History:This paper was received October10, 2003, and was with the authors24 months for 3 revisions;
processedby Joel Huber.

1. Introduction consumer choice set formation and preference dis-


As "a name, term, sign, symbol or design, or a com- crimination,the lattera concepton which we will sub-
bination of them which is intended to identify the sequently elaborate. Brandcredibility(the credibility
goods and services of one seller or a group of sellers of a brand as a signal) is defined as the believabil-
and to differentiatethem from those of competitors" ity of the product position informationcontained in
(Kotler 1997, p. 443), the brandplays multiple roles a brand. This information depends on consumers'
in consumerchoice. These roles may include brands' perceptions of the willingness and ability of firms
effects on consumerpreferences;on brand and quan- to deliver what they have promised. This perceived
tity choice; and on consideration, to name a few. willingness and ability to deliver constitute the two
Additionally, these effects may materialize through components of brand credibility:trustworthinessand
various mechanismssuch as psychological(e.g., asso- expertise. When asymmetric information character-
ciative network memory), sociological (such as brand izes a market, economic agents (i.e., consumers and
communities),and economicprocesses (brandsas sig- firms) may use signals (i.e., manipulableattributesor
nals under uncertainty)(Keller2002). activities) to convey informationabout their charac-
In this paper, we investigate the impact of brand teristics (Spence 1974). To be effective, such signals
credibility, a central concept in economics-based must be credible (Tirole1990).Previous literaturehas
approaches to brand effects and brand equity, on studied the credibilityof a brand as a signal of qual-
679

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem:BrandEffectson ChoiceandChoiceSet FormationUnderUncertainty
680 Science26(5),pp.679-697,©2007INFORMS
Marketing

ity or product positioning (Wernerfelt1988, Rao and are robust (for example, one needs to link brand
Ruekkert1994, Erdemand Swait 1998). choice to formation of subsets from the universal
We define the choiceset as the brands that have set in a unified framework, control for unobserved
a nonzero probabilityof being chosen among those heterogeneity,and incorporateprice in the model to
actuallyavailablefor choice in a given context, at the explore the effect of brand credibility,none of which
time of choice. We should note that in the literature steps were taken in Erdemand Swait 2004);(b) it is of
some researchersused the terms consideration set interest in itself to researchthe mechanisms through
and choice set interchangeably (for a review, see which brand credibilityoperatesin brand choice con-
Roberts and Nedungadi 1995), while others have ditional on choice set versus the choice set forma-
drawn a distinction. Following Shockeret al. (1991), tion stage; (c) it is necessary to determine whether
we make such a distinction:althoughthese are closely brand credibilityalso affects choice set formation,in
relatedconstructs,consideration sets referto long-term, addition to consideration(Erdemand Swait 2004).We
dynamic sets that vary within and across usage and tackle all three issues in this paper. Furthermore,in
purchaseoccasions, whereas choicesets are conceptu- this paper we also investigate the impact of brand
alized as the set of alternatives considered immedi- credibility(and the mechanismsthrough which these
on discrimination.We
ately prior to choice, and are thus more instantaneous effects materialize) preference
in nature.(Of course, the choice set could itself be the conduct our analysis on experimentaldata collected
end result of a dynamicprocess.)This paper proposes through one study involving both the juice and per-
and shows that brand credibility not only affects sonal computer(PC)productcategories,and a second
choice set formation and conditional brand choice, study involving only PCs.1
but it also argues and tests for the differentialmech- We review relevant literatureand define the con-
anisms through which credibility operates at each ceptual frameworkguiding our testing in §2; we dis-
cuss our modeling approachand data in §3 and our
stage. resultsin §4. We conclude the paper with a discussion
Besidesthe impactof brandcredibilityon choice set
of managerialimplicationsand future research.
formation,we also explore the effect of brand credi-
bility on what we term preferencediscrimination.It
is well known that in discrete choice models, many 2. Related Work and Conceptual
of which possess a characteristicallysigmoidal shape Framework
(e.g., multinomiallogit, nested logit, and probit), the 2.1. Choice Set Formationand Brand Choice
more the slope coefficient grows in magnitude, the
Conditional on Choice Set
steeperthe sigmoidal shape becomes. In fact, at one of
its limits the sigmoidal shape will arbitrarilyclosely Previous empiricalwork on how consumersmay nar-
emulate a step function (see, e.g., Ben-Akiva and row attention to a subset of brands out of a bigger set
Lerman1985, pp. 70-72). A behavioralinterpretation
has focused on modeling choices as the outcome of
a two-stage process of considerationset or choice set
of this phenomenon is that decision makers'capacity
formation and conditional brand choice (e.g., Swait
to discriminatebetween alternativesmay be higher in
and Ben-Akiva 1987a, b; Roberts and Lattin 1991;
certain contexts. This discriminationcan be concep-
Andrews and Srinivasan1995;Chiang et al. 1999).
tualized as preferencediscrimination,that is, a given
consumer's choice probabilitieswill be consistently model However, there has been no published empirical
of consideration or choice set formation in
low or high for certain brands since the consumer the context of brand choice that
can discriminatebetter among such alternatives.We the explicitly captures
explanations offered for brands' effects on con-
conjecturethat brand credibilitywill affectpreference sideration or choice set formation.This is not to say
discriminationthus defined. that the wider literaturehas not considered appro-
While previous research has explored the impact theoreticalbases for specifying consideration
priate
of credibilityon product utility in adopting a struc- or choice set formation models. For example, Meyer
tural equation modeling framework (Erdem and
(1979) formulated a theory of destination choice set
Swait 1998) and separately on brand consideration formation,
and choice (Erdem and Swait 2004), we investigate constraints; incorporatinginformationavailabilityand
in the transportationdemand area, Swait
brand credibility'seffect on choice set formationand and Ben-Akiva(1987a,b) modeled the impact of dif-
brand choice, conditional on choice set in an inte- ferent constraintson choice.
grated framework. For example, Erdem and Swait While empiricalwork tried to link choice or consid-
(2004) estimated simple binary logistic models on erationset formationstages to the brandchoice stage,
self-reportedconsideration data from several prod-
uct categories and found evidence for brand credi- 1We should note that Erdemand Swait
(1998,2004)have not used
bility effects. These results call for further research any experimentaldata and conducted their analysis on consumer
since (a) one needs to test whether such effects self-reportson several items.

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set Formation Under Uncertainty
MarketingScience26(5),pp. 679-697,©2007 INFORMS 681

the main approachin the literatureto conceptualizing Brand credibility, in turn, may (a) increase con-
the formationof subsets from universal sets of alter- sumer expected attribute(e.g., quality) levels (Aaker
natives has been the cost-benefit approach (Hauser 1991),(b) decreasethe varianceof consumerattribute
and Wernerfelt 1990). This approach employs the beliefs, i.e., consumerperceived risk3(Srinivasanand
expected utility maximizationframeworkto advance Ratchford1991), and (c) decrease information costs
the notion that consumers weigh the cost of brand (Shugan1980)by crediblysignaling productpositions
evaluation for membership in this subset against when there is consumer uncertainty about brands
the benefits of adding or dropping the brand. This (Erdemand Swait 1998, Montgomeryand Wernerfelt
approach based on expected utility maximization 1992,Wernerfelt1988).
implies consumer uncertainty about brands, since
consumersare unsureabout the utility they would get 2.3. Brand Credibility, Choice Set Formation, and
from each availablebrand and, hence, need to make Conditional Brand Choice
expected utility-expectedcost calculationsin forming Such a signaling frameworkof brand effects on con-
their considerationor choice sets. sumer brand utility and choice also implies that
when there is consumer uncertainty about brands
2.2. Brand Credibility and information is costly to obtain or process, the
The importanceof credibility under uncertaintyhas credibility of a brand may be an important fac-
been established in several contexts (e.g., Xie and tor underlying the formation of choice sets (as well
Shugan 2001, Godes and Mayzlin 2004). There is as consideration sets; see Erdem and Swait 2004).
also a growing literatureon the importanceof brand The higher perceived value and lower perceived risk
credibility under consumer uncertainty.When con- associated with a higher-credibilitybrand are antic-
sumers are uncertainabout brands and the marketis ipated to increase expected benefits (Hauser and
characterizedby asymmetric information (i.e., firms Wernerfelt1990).Additionally,the lower information
know more than consumers do about their prod- costs associated with credible brands are likely to
ucts),brandscan serve as signals of productpositions decrease expected costs, while the credibility of a
(Wernerfelt1988). As a signal of product position- brand decreases perceived risk because it increases
ing, the most important characteristicof a brand is consumers' confidence in a firm's product claims.
its credibility.A firm can use various marketingmix Credibilityalso decreasesinformationcosts since con-
elements in addition to the brand to signal prod- sumers may use credible brands as a source of
uct quality:for example, charginga high price, offer- knowledge to economize on information gathering
ing a certain warranty, or distributing via certain and processing costs (e.g., reading ConsumerReports
channels. Each of these actions may or may not be or doing online searchesfor product reviews).
credible depending on market conditions, including In the context of brand choice, the cost-benefit
competitiveand consumerbehavior.However, credi- approach implies consumer uncertainty about the
ble signals that set brands apart from the individual attributes(i.e.,quality)of brands.Althoughnot explic-
marketingmix elements is that the former embody itly included in models of considerationor choice set
the cumulativeeffect of past marketingmix strategies formationand brandchoice,models of Bayesianlearn-
and activities, as well as consumer interactionswith ing (Robertsand Urban 1988,Erdemand Keane 1996)
the firm.Thishistoricalnotion that credibilityis based are consistent with a cost-benefitapproachbased on
on the sum of past behaviors has been referred to expected utility maximization to model brand eval-
as reputationin the informationeconomics literature uations. Erdem and Keane (1996) model consumers'
(see Herbig and Milewicz 1995).
Credibilityis broadly defined as the believabilityof investments, ceteris paribus (Erdem and Swait 1998). Consistency
an entity's intentionsat a particulartime. It is posited refers to the degree of harmony and convergence among the mar-
to have two main components:trustworthinessand keting mix elements and the stability of marketing mix strategies
expertise. Thus, brand credibility is defined as the and attribute levels over time. Brand investments, on the other
hand, are resources that firms spend on brands to (a) assure con-
believabilityof the product informationcontained in sumers that brand promises will be kept and (b) demonstrate
a brand,which requiresthat consumersperceive that
longer-term commitment to brands (Klein and Leffler 1981). Fur-
the brand have the ability (i.e., expertise) and will- thermore, it has also been shown that the clarity (i.e., lack of
ingness (i.e., trustworthiness)to continuously deliver ambiguity) of the product information contained in a brand is an
what has been promised (if and when brands do antecedent to brand credibility (Erdem and Swait 1998).
3Consumer
not deliver what is promised, their brand equity will uncertainty about product attributes generates con-
sumer perceived risk (which can be conceptualized as the variance
erode).2 of consumer attribute beliefs) because "any action of a consumer
will produce consequences which he cannot anticipate with any-
2The
credibility of a brand has been shown to be higher for brands thing approximating certainty, and some of which at least are likely
with higher marketing mix consistency over time and higher brand to be unpleasant" (Robertson et al. 1984, p. 184).

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand ChoiceSet Formation Under Uncertainty
682 MarketingScience26(5),pp. 679-697,©2007 INFORMS

learning about imperfectlyknown product attributes Figure1 TheImpactof EffectSizeonBinaryLogitChoiceProbabilities


throughadvertisingand experience,which alterscon-
sumers' mean attribute evaluations (expected qual-
ity) and their variance (perceived risk). Their model
is consistent with Howard and Sheth's (1969) the-
ory of formation of subsets from universal sets of
alternatives, which posits that consumers will try
different brands to learn about them and then buy
from among only a small subset of availablebrands.
Mehta et al. (2003) estimated such an explicit model
on scanner panel data; in their model, experience
provides informationabout brands and reduces con-
sumer uncertainty.
We postulate that the more credible the brand,
the greater the likelihood of a consumer including
it in his choice set due to its possible impact on
perceived quality, perceived risk, and information
costs. While this notion follows directlyfrom the cost- This effect-size parameterr is directly related to the
benefit approach (Hauser and Wernerfelt1990), the relativeimportanceof the stochasticutility difference
magnitudeof credibility'simpacton choice set forma- (e1 - e2) compared to the systematic utility differ-
tion versus its impact on the conditionalbrand choice ence (V\ - V2); precisely, t is inversely related to
the variance of this stochastic difference. Figure 1
stage, and the differentialmechanismsthroughwhich
brand credibility may operate at these two stages, maps out the choice probability of Brand 1 as a
have not been tested explicitly. function of the systematic utility difference and the
effect-size parameter r. As r -» 0 (i.e., the variance
of (f^ - s2) approaches infinity), the probability of
2.4. Brand Credibility and Preference
Brand1 approaches1/2 everywhere,which is to say,
Discrimination
choice is random. On the other hand, as t -> oo
Choicemodels are used to predictbehavioras a func-
tion of attribute-basedutility differences. To elabo- (i.e., the variance of (s1 - e2) - 0), the probabilityof
Brand1 being chosen approachesa step functionthat
rate and set the stage for subsequent explanations, transitions abruptly at Vl - V2 close to zero, which
we remind the reader that in these latent variable is to say, in the limit choice is deterministic with
choice models the total utility of a product is gener-
respectto the systematicutility differencebetween the
ally decomposed into a systematiccomponent V and brands:if Vl-V2< 0, Brand2 certainlywill be chosen,
a stochasticcomponent e (also commonly termed the if Vt - V2> 0 then Brand 1 certainly will be chosen,
error,in analogy to linear models). Substantivetheo- and if the differenceis zero, choice will be random.
retical developments in the area of choice modeling This effect is discussed in Ben-Akiva and Lerman
have concentratedalmost exclusively on the system-
(1985, pp. 70-72) and Louviere et al. (2000, p. 236),
atic component,but the realizationthat the stochastic and is known to hold for multinomiallogit and pro-
component of utility also influences a model's abil- bit, nested logit, and other choice model forms.
ity to predict choice behavior is gradually growing In effect,not only do choice model parameterscap-
among researchers (e.g., Swait and Louviere 1993, ture the impact of systematic utility differences on
Louviereet al. 2000). choice probabilities,but the magnitudeof this system-
Choice models exhibit an interestingpropertywith atic impact is moderated by the relative importance
respect to the size of these systematic utility differ- of the stochasticutility.We term this moderationphe-
ences: in different contexts and for different deci- nomenon preference discrimination.As Figure 1 clearly
sion makers,the same systematicutility differencecan illustrates, for the same systematic utility difference,
result in more-extremechoice probabilities,depend- a larger effect size will translate into more-extreme
ing on the relative sizes (or scales) of the stochas- probabilities(i.e., closer to zero or one). As an effect
tic utility components.To illustrate,consider a binary size grows (i.e., stochasticutility decreasesin impor-
logit probabilitymodel between Brands1 and 2 that is tance with respect to total product utility), systematic
a functionof the systematicutility differencebetween preferencesbecome more and more able to discrim-
the brands, V2- V2,as well as an effect size parameter inate between alternatives, with the choice process
r > 0, thus tending toward the step-functionlimiting condition;
this is nothing more than the model's attribution
P^Hl + expt-r^-V,,))}-1, T>0. (1) of perfect discrimination between the products or

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set FormationUnder Uncertainty
MarketingScience26(5),pp. 679-697,©2007 INFORMS 683

brands to the decision maker on the basis of known size can be either positive or negative, depending
attributes. on the role that credibilityplays in choice heuristics.
A recent stream of literaturein the discrete choice Since the choice heuristics consumers apply under
area has studied the impact of errorheteroscedastic- uncertainty(and, hence, the impact of credibilityon
ity on choice model parameterinferences(see, among choice) may vary across different decision contexts,
others, Swait and Louviere 1993, Allenby and Ginter we test the net impact of credibility on impact size
1995,Dellaertet al. 1999,Louviereet al. 2000).In that in two specific decision contexts, characterizedby
literature,heteroscedasticityplays the same role that the variability(fixed versus variable)and size of the
effect size has played above. However, the effect-size choice set (small versus large).
phenomenon we have discussed above is more gen-
eral because the heteroscedasticityexplanation can
3. Empirical Tests
only be used in the context of random utility models, We have conducted two separate studies to test
whereas the effect-sizeexplanationcan apply for any
the concepts discussed above. In the first, more-
probabilisticchoice model (PCM).
Erdem and Swait (1998) found that credibility comprehensive study, we investigate the impacts of
affects product evaluation by causing utility (i.e., brand credibilityon choice set formation,preference
discrimination,and conditional choice in two prod-
product attractiveness) to increase with increasing uct categories (juices and PCs) for which we expect
quality, decreasing risk perceptions and increasing differentialbrand impacts. This first study involves
informationcosts saved. The effect-sizephenomenon
we are postulating to be a function of brand credibil- an experimentalchoice task, designed with fixed-size
sets using five brands. Having found strong support
ity is over and above the wholesale differencesErdem for the hypothesized impacts of brand credibility,it
and Swait (1998) found in their structuralequation
models. Referring to Figure 1, Erdem and Swait's was decided to conduct a second study on the PC
results would suggest that credibilitycauses vertical product category only, but substantiallyincrease the
shifts in the sigmoidal shape; preferencediscrimina- number of brands and make the size of choice sets
tion or effect size, on the other hand, refers to the vary considerably.The purposes of the second study
are twofold: (a) to lend face validity to the struc-
steepness or sharpnessof the sigmoidal shape around turalchoice set formationmodel by elicitingchoice set
its point of balance(here,where utility differencesare
zero). membershipinformation,in addition to choice data;
This steepeningor flatteningof the sigmoidal shape and (b) to investigate the robustnessof the results of
is meaningful only to the analyst, of course, and the first study when the complexityof the choice con-
not to the consumer. Theoretically,the consumer is text is increasedby using varying set sizes and nine
assumed to be aware of a product's full utility attri- differentbrands (almostdouble the numberof brands
bution (systematic and stochastic). What the effect used in the first study). The second study results are
size is describingis simply the relativerelianceof the found to strongly support the first study results.
consumer on known (to the analyst) sources versus 3.1. Study 1
unknown (again, to the analyst) sources of utility
when making a choice. The steepening of the sig- 3.1.1. Model Description. We have formulated a
moidal shape indicates a sharper (i.e., steeper) sen- PCM that structurallyincludes choice set formation
sitivity with respect to all the variables included and also allows for capturing preferencediscrimina-
in the systematic utility specification- thus, the con- tion effects due to brand credibility.Our specification
sumer will be relying more strongly on the sys- is an elaborationon Swait's (1984)independentavail-
tematic component of utility than on the stochastic ability multinomial logit (IA-MNL)model (see also
component. This greater consumer reliance on the Swait and Ben-Akiva1987a,Andrews and Srinivasan
systematic component would be reflected in more- 1995, Ben-Akiva and Boccara 1995). The two-stage
discriminating(i.e., more-extreme)behaviorin regard model we use in our researchadds two features to
to the choice alternatives on the part of consumers. the IA-MNLmodel:
Our expectation is that brand credibility may affect 1. Preference discrimination, parametrized as a
the degree of this discriminatingbehavior.For exam- functionof brand credibility,which allows us to char-
ple, if high (low) credibilityserves as a quick heuristic acterize how consumers continue to rely on brand
for consumers to discriminateamong brands, to the credibilityas a decision aid during the second stage of
analystsuch relianceby consumerson brand credibil- choice. We accomplish this by defining and estimat-
ity would reflect itself in a conditional choice model ing an effect-sizefunction r, which is parametrizedin
throughincreased(decreased)parametermagnitudes, terms of brand credibility,among other factors.
that is, increased (decreased) effect size. Thus, it is 2. Random parameters, to capture individual-
possible that the net impact of credibilityon impact specific heterogeneity in product evaluations. As

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem:BrandEffectson ChoiceandChoiceSet FormationUnderUncertainty
684 Science26(5),pp.679-697,©2007INFORMS
Marketing

shown by Heckman (1981), among others, if unob- given by (2d) and (2e):5
served heterogeneity is unaccounted for, biased
parameterestimates can result. We thereforeinclude Qn(C,Wn\8)
person-specificrandomeffects. 1g))0Wc[l " AknWkn
I*)1)
The derivation of the basic probabilistic choice _ (UjecAjniWjn
model we employ is availablefrom the authorsupon i-(n^[i-4(\is)])
request. CeAM. (2d)
We now describe in detail the econometricspecifi-
cation we use, beginning with Pin,the unconditional Ain(Win\8)= [l + exp(-8Win)]-\ VfeM, (2e)
probabilitythat individual n chooses alternativei where Ainis the probabilityalternative,/ is included
in the choice set, Winis a vector of product- and
Pin= f E Pin\c(Xn,Zn\P,0) person-specificcharacteristics,and 8 is a conformable
parametervector. The normalizationconstant in the
'QH(C,Wn\8)<Kfi\fi,Zfi)dp, V/gM, (2a) denominatorof (2d) excludes the possibility of a null
choice set, a logical impossibilitywhen choice is actu-
where C is a choice set in AM,which is the set of ally observed, as it will be in the data we collected.
all possible choice sets derived from the universal (or Thus, Model (2a-2e) details a conditional MNL
available) set M of brands eligible for choice at the choice model with random coefficients (mean j3
time of decision, Pin|Cis the conditional probability and variance-covariancematrix2^) and parametrized
of choosing i from set C, and Qn(C)is the likelihood effect-size function, plus an independent availability
of C being the choice set from which i is chosen. The choice set formation model in which the individual
conditionalchoice probabilityis defined by the multi- probabilitiesof brand inclusion in the choice set are
nomial logit (MNL)model with effect-sizefunctionr:4 given by a logistic CDF,which is a functionof covari-
ates. Despite the fact that this model is more complex
|/3,0)
P,.n|C(Xn,Zn than standardchoice models, particularlydue to the
exp[Tin(Zin\e)-Vin(Xin\l}n)] V.CCM explicit formulationof a two-stage decision process,it
~ ' is not our intent to claim that decision makersactually
E;ecexp[r/n(Z;.J0) Vjn(Xjn\pn)]'
employ such a two-stage approachto choice. Instead,
=0, V^C, (2b) we put this model forwardas an improved paramor-
where Vinis the deterministicutility, defined subse- phic representationof observed choice behavior.
The three parametrized functions in the model
quently as a function of product attributesand per- above are detailed below:
sonal characteristics,as well as person-specifictaste
weights /?„, which are distributed according to the
multivariatedensity function </>(/3
| j8,2^) (see Expres- Vin= ain+ p1ln(Pi)+ p2Cin ieM, (3a)
sion (2a)), where j8 is the population mean and 2^ rin= exp(0a + 01Cin+ 02C?), i € M, (3b)
is the variance-covariancematrix. The vector Zin of
Ain= [l + exp(-(8n^82Pi^83Cin))]-\ ieM, (3c)
product-and person-specificcharacteristicsis used to
parametrizethe effect-sizefunction where
rin(Zin\d)= exp(6Zin), V/eM, (2c) Cin= brand credibilityconstructfor brand i,
person n;
where 6 is a conformableparametervector;the expo- =
p, price of brand i; and
nentiationoperatoris used to constraineffect-size rin a, (5, 8, 6 = parametervectors to be estimated.
to be nonnegative.The probabilityof a choice set C is
5As
pointed out by Swait (1984),the IA-MNLmodel (and, conse-
4
Independenceof irrelevantalternatives(IIA)is a well-knownchar- quently,our extensionof it) has the characteristicthat the inclusion
acteristicof the basicMNLmodel, as well as this variantof it. How- of any one alternativein the considerationset is probabilistically
ever, even though the conditionalchoice model of our proposed independentof the inclusionof any otheralternativefrom M. This
model system displays IIA, the overall choice probabilitymodel assumptionis made principallyfor tractability.Withoutthis simpli-
(Expression(2a)) does not share this property.Intuitively,because fication,the probabilityof each considerationset must be estimated
there is explicit considerationof choice set formation,and choice separately.For a set with / alternatives,there are (2J- 1) sub-
sets are essentiallyan expressionof specificmarketstructures,the sets whose probabilitiesmust be estimated, which is in practice
overallchoice model does not have the IIA property.Furthermore, intractablefor even small / (say, / > 5). This assumptionof inde-
we estimate a random effects specificationand allow for a con- pendentavailabilityprecludesthe possibilitythatbrandsmay share
tinuous distributionin tastes, which implies that the brand choice unobservedattributesand characteristicsthat cause brand groups
model we estimateis not subjectto IIA at the aggregatelevel (that to jointly have greateror smaller probabilitiesof joint occurrence
is, IIA holds only for each individualseparately). in the choice set than would be predictedby independence.

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set Formation Under Uncertainty
Science26(5),pp.679-697,©2007INFORMS
Marketing 685

Price is included in systematicutility Function(3a) credibilityas a common antecedent;and (b) that the
in logarithmicform to capturea decreasingmarginal impacts of credibilityon utility are entirely mediated
impact on utility; price is also included in the avail- by these three constructs.Accordingly,we also esti-
ability Function(3c) to captureprice thresholdeffects mate variant models of (3a-3c) that substitute these
in choice set formation, which might otherwise be end constructsin the place of brand credibility:
confoundedwith brand credibility.In the latter func-
tion, price is included linearly because the logistic Vm= otin+ ft ln(p,)+ riPQln+ y2PR,n+ y3ICSin,
functionlends itself to capturinga thresholdeffect for
ieM, (4a)
variablesincluded in a linear fashion in its argument.
It will also be noted that (3b) includes both linear and rin= exp(0fl+ i^PQfc+ k2PQ*,+ K3PRin + k4PR^
quadratic credibility terms. We have argued before + k5ICSi.b
+ ic6ICS?i)/ ieM, (4b)
that high or low levels of brand credibility might
lead to increasedpreferencediscrimination.However, Ain= [1 + exp(-(8I1 + 8lVi+ ^PQfo + tt2PR,w
it is possible that the marginal impact of credibility
+ 7r3lCSin))]-\ ieM. (4c)
on preferencediscriminationwill decreaseat extreme
credibilitylevels due to a thresholdeffect:at very high Vectors y, k, and tt are additional parametersto be
or very low credibility levels, the marginal impact estimated,along with or instead of previouslydefined
of credibilityon preferencediscriminationis likely to
parameters.Working our way through the implicit
diminish because some residual level of uncertainty
hypotheses in (4a-4c), it is our expectation that in
will persist.This same kind of argumentcan be made the choice set formationstage all three end constructs
in terms of the impact of credibility on choice set are likely to affect brand inclusion in Model (4c):
formation;Expression(3c) allows this kind of thresh- (a) employment of a quality thresholdshould lead to
old effect of choice set formation directly through a significantand positive t^; (b) the desire to exclude
the logistic CDF functionalform (see subsequentdis-
cussion of results). However, in the case of (3b), the high PR brands should lead to a significantly neg-
ative tt2 effect; and (c) economies with respect to
possibility of a threshold effect must be allowed for decision-makingcosts should be indicated by a sta-
explicitly through the quadraticterm since the expo- tistically and substantivelypositive tt3. In the utility
nentiation operator is strictly monotonic in its argu- Function (4a), which is of course estimated condi-
ment. These three functions should be understood tional on brandinclusionin the choice set, (a) it seems
simply as flexible forms that will be determined by to us that relative quality differencesamong brands
the data; there is no a priori theory to guide their in the choice set should play a role, leading to a dis-
specifications,though we've tried above to justify to cerniblepositive effect for yx; (b) with respect to risk
some extent the specific forms shown. and informationcosts, it is not so clear from a priori
It will be noted in (3a-3c) that brand credibility considerationsthat these variablesshould continueto
enters the model in three ways: First,it affects choice
set formation via the brand inclusion probability play a role in determining product attractivenessor
Model (3c)- increasingbrand credibility is expected utility, once their impact has been controlled for in
the choice set formation stage. Whether or not they
to increasea brand's likelihood of being in the latent do must be determined empirically.However, this is
choice set. Second, conditional on the brand being not to say that systematicutility is wholly unaffected
in the choice set, brand credibility is expected to
affect utility by shifting utility upward, as in Expres- by PR and ICS:we expect these and PQ to display
sion (3a), in which /32 is expected to be positive. significantimpacts on the effect-size Function (4b) if
preferencediscriminationis operativewithin the con-
Third,it is hypothesized that the effect-sizefunction r ditional brand evaluation stage of choice.
will reflectgreaterpreferencediscrimination(perhaps
with diminishing marginal impact) at the extremes 3.1.2. Data Collection and Calculationof Indices.
of credibility since that construct may be used by Data were collected through paper-and-pencilsur-
consumers to define credibilitythresholds,below or veys, the subjectsof which were undergraduatestu-
above which preferencesare better discriminatedby dents at two majorNorth Americanuniversitieswho
known attributeor variableimpacts (i.e., attributesor receivedcourse creditfor participation.The final sam-
variablesincluded in the systematicutility). ple sizes were 391 respondents for juice and 366
One of our goals is to understand the mecha- for PCs.
nisms whereby brand credibilityaffects the different Constructvalue estimates for brand credibility,PQ,
stages of choice behavior. Previous research (Erdem PR, and ICSwere obtained in a straightforwardfash-
and Swait 1998, 2004) strongly suggests that (a) per- ion. Based on the structuralequation model reported
ceived quality (PQ),perceivedrisk (PR),and informa- in Erdem and Swait (1998),as well as on subsequent
tion costs saved (ICS)are constructsthat share brand studies that have used the scales there developed

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand ChoiceSet FormationUnder Uncertainty
686 MarketingScience26(5),pp. 679-697,©2007 INFORMS

Table1 andRelatedConstruct
BrandCredibility Measures
ItemsandReliability
Construct Item a
Cronbach's

Credibility Thisbrandremindsmeof someonewho'scompetent andknowswhathe/sheis doing.(+) 0.85


Thisbrandhastheabilityto deliverwhatit promises.(+)
Thisbranddeliverswhatit promises.(+)
Thisbrand's productclaimsarebelievable. (+)
Thisbrandhasa nameyoucantrust.(+)
Thisbranddoesn'tpretend to be somethingit isn't.(+)
PQ Thequalityof thisbrandis veryhigh.(+) 0.77
Intermsof overallquality,I'dratethisbrandas a. . . (+)
PR I'dhaveto tryit severaltimesto figureoutwhatthisbrandis like.(+) 0.64
I neverknowhowgoodthisbrandwillbe beforeI buyit. (+)
ICS I needlotsmoreinformation aboutthisbrandbeforeI'dbuyit. (-) 0.75
I knowwhatI'mgoingto getfromthisbrand,whichsavestimeshoppingaround.(+)
I knowI cancountonthisbrandbeingthereinthefuture.(+)
ThisbrandgivesmewhatI want,whichsavesmetimeandefforttryingto do better.(+)
Notes,(a) Allscalesformedas simpleaveragesof componentitems,reversescoredfornegativeitems,(b) Allitemsmeasured
on
nine-point scaleexceptforsecondPQitem,whichwas measured
agree-disagree 9 = high
scalewith1 = lowquality,
on nine-point
(c) (+/-) signsindicate
quality, a priorisignexpectation.

(Erdem and Swait 2004), we employed the first six expected interrelationshipsbetween them, as speci-
items presented in Table 1 to measure brand credi- fied by Erdemand Swait (1998),we estimated simul-
bility. The two components of the brand credibility taneous equation models involving only observables
construct,namely expertise(firsttwo items in Table1) (i.e., the calculatedindices) by productclass, and con-
and trustworthiness(next four items in Table 1), are firmed that the credibility index, decomposed into
both included in credibility.These items were mea- trustworthinessand expertise,is an antecedentof the
sured on nine-point agree-disagreescales; credibility other three construct indices. These results, not pre-
for a person and brand combinationwas calculated sented here, are availablefrom the authors.
as the simple average of all six items. There is the In addition to rating each of five brands6in terms
possibility,however, that such measures might sim- of the items in Table 1 (as well as providing cer-
tain other brand-levelinformation),respondentsalso
ply reflect individual differences (i.e., an individual
rates all brands better or worse than other individu- completed a simple pricing choice experimentinvolv-
als); thus, the final constructestimate was defined for ing 17 choice sets or scenarios for both product cat-
brand i, person n, as the mean-centeredvalue Cin= egories that they rated on the aforementioneditems.
An orthogonalmain effects design from the 45"3 facto-
Qn ~ Cn, where rial was used to construct16 choice sets. This design
Cin= E XqnU yields one choice set with all brands priced at their
6
lowest levels, so a 17th set with all brands priced at
;=i
their highest levels was added to capturea pure cat-
is the simple average of the six items x^n,and Cn= egory demand effect; we also included a "none of
these" option in each set to allow respondentsto opt
£, Cin/Kis the person-specificaverage of Cinover the out of the category entirely if they didn't like the
K brands.
Table 1 also shows measurement scales for three brands,the prices, or both. Subjectscould thus choose
one of the five brands at the prices offered, or select
other constructs:PQ, PR, and ICS. Recall that Erdem
"none of these."
and Swait (1998) showed that brand credibility is
We placed anotherquestion between the two main
an antecedent to these three constructs, which are tasks of interestthat asked respondents(a) to evaluate
the mediatingmechanismswhereby credibilityaffects the degree of confidence they felt in assessing a new
productevaluation and utility. product in each of 21 product categories (including
We defined person- and brand-specificPQ, PR, and frozen orange juice concentrateand PCs) before trial,
ICSmeasuresby averagingthe respectiveitems given after one trial and after one year of use, by means
in Table 1, then centering each constructwithin per- of seven-point agree-disagreescales; and (b) to asso-
son, analogously to the procedure used for credi- ciate statements with each of the same 21 categories
bility. These measures permit examination of how
credibilityaffectschoice set formationand preference 6 The brands used in
Study 1 were, for juices, Dole, Minute Maid,
discrimination.To ascertainthat these indices of credi- Sunkist, Tropicana, Welch's; for PCs, Apple, Compaq, Dell, Gate-
bility,quality,risk, and informationcosts maintainthe way, IBM.

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set Formation Under Uncertainty
MarketingScience26(5),pp. 679-697,©2007 INFORMS 687

Table2 ResultsforJuices(Study1)
Estimation
t-statistics)
(Asymptotic
(b) MNL (d) IAL (e) IALPQ,
(a) MNL and
credibility (c) MNL &
credibility PR,ICSand
brandhet. brandhet. PQ,PR,ICS brandhet. brandhet.
function(Vin)
Utility
Brand1 3.713(12.4) 3.087(29.1) 3.174(29.8) 2.378(6.1) 3.936(4.4)
Brand2 4.523(11.4) 3.651(34.5) 3.643(34.3) 2.845(7.3) 3.613(4.6)
Brand3 2.957(13.3) 2.837(26.3) 2.918(26.6) 2.254(6.4) 3.661(4.5)
Brand4 4.325(11.6) 3.554(33.4) 3.514(33.2) 2.534(6.3) 4.210(4.8)
Brand5 3.058(12.9) 2.884(26.2) 2.991(27.3) 2.285(5.4) 2.810(4.8)
In(Price) -7.429 (-7.9) -5.931 (-39.7) -6.129 (-38.8) -6.206 (-10.1) -10.523 (-5.0)
Credibility 0 0.893 (38.3) 1.802 (8.1)
PQ 0.213 (10.3) 2.286 (4.9)
PR -0.338 (-11.8)
ICS 0.459 (13.7)
crb2rand 2.352(2.0) 3.76E-10H) 0 3.76E-10(- ) 0.810(1.1)
Effect-size
function(In%)
Brand1 0.072 (0.8) -0.593 (-3.6)
Brand2 0.12 (1.3) -0.454 (-3.2)
Brand3 0.25 (2.3) -0.254 (-1.4)
Brand4 0.011 (0.1) -0.415 (-3.1)
Brand5 0 0
Credibility 0.366 (5.3)
Credibility^ -0.051 (-1.7)
PQ 0.295 (4.0)
PQ**2 0.037 (1.5)
PR 0.004 (0.1)
PR**2 -0.352 (-5.7)
ICS 0.241 (2.4)
ICS"2 0.258 (3.9)
Brandinclusion
functions
Brand1 4.267 (9.5) 4.867 (14.2)
Brand2 4.638(10.6) 5.723(16.3)
Brand3 3.838 (8.9) 4.597 (14.2)
Brand4 4.965 (10.6) 5.179 (15.6)
Brand5 4.024(8.4) 5.116(14.3)
None -0.458 (-1.4) -1.148 (-7.4)
Price -2.339 (-8.1) -3.421 (-16.7)
Credibility -0.132 (-1.1) 0
PQ -0.043 (-1.2)
PR -0.382 (-11.0)
ICS 0.376 (9.2)
LL(convergence) -4,036.67 -3,538.66 -3,425.49 -3,417.74 -3,235.02
Akaike
Rho-squared 0.2164 0.3129 0.3344 0.3336 0.3677
Numberof parameters 7 7 9 21 28
Notes,(a)AllSMLestimatesbasedon R= 150 Haltonreplicates,
(b)Allconstructindicesaremean-centered
foreachindividual
respondent,
(c) Number 6,632choicesfrom391 respondents,
of observations: fromIIDnormalvariates.
(d)Stochasticbrandeffectestimated

that describe their familiaritywith it, potential risks of potentialuncertaintyand sensitivity to uncertainty.
involved in a purchase, the type of benefits offered, We wanted to use two product categoriesthat would
their level of involvement, hedonistic aspects of pur- vary widely with respect to level of involvement,
chasing in the category,and so forth (these measure- degree of potential PR, and information costs. One
ment items are detailed in Erdemet al. 2002,Table2). factor that affects potential information costs and
This informationwas considered useful as a means PR is the imperfect observability of attributes, a
of supporting interpretationof results about the two concept describing the extent to which consumers
productclasses. can evaluate perfectly the product attributesjust by
The two product categories investigated in our search, through just one or perhaps a few consump-
empirical research (frozen orange juice concentrate tion experiences,or through a long consumptionhis-
and PCs) were chosen to cover a wide enough range tory (Nelson 1970, 1974). Indeed, certain attributes

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set FormationUnder Uncertainty
688 MarketingScience26(5),pp. 679-697,©2007 INFORMS

may never be perfectly observable (i.e., credence (c) to show that one has a better and more accu-
attributes;Darby and Karni 1974). The intervening rate understandingof the mechanismsthroughwhich
task mentioned above clearly showed that these sub- brand credibilityoperates in models that incorporate
jects, as a group, viewed orange juice concentrateas both choice set and conditional brand choice stage.
a search or short-termexperience good and the PC We shall concentrateour presentationon the last two
as a much-longer-termexperience or credence good. models for each product class. In the paragraphsthat
The subjectsalso viewed PCs to be a more complex follow, we will discuss the results for the two product
and higher-involvement category. Considering also classes in parallel.
that PCs are several ordersof magnitudemore expen- Model (1) is a simple MNL model that has only
sive than orange juice, potential information costs brand constants, price and brand heterogeneity.The
and PR are expected to be higher in PCs. Studying latter is specified as independent and identicallydis-
the impacts of brand credibility on choice in such tributed (IID) normal variates across brands. (The
widely differingproduct classes permits us to exam- assumption of identical distributionacross brands is
ine whether credibilitymay have differentimpacts on made for parsimony, based on empirical findings.)
choice across differinglevels of potential information From Tables 2 and 3 it is apparentthat there is sig-
costs and PR. nificant brand-level heterogeneity,since the variance
One would expect, for example, the effect of cred- of the normally distributedstochasticbrand effect is
ibility on choice set formation to be larger in PCs statisticallydifferentfrom zero at the 95%confidence
than in juice, given that potential uncertainty and level, in both product classes. However, Model (2)
risk are higher for PCs; the data collected through adds the credibilityconstructto the MNL utility func-
the intervening task described above confirmed this tion in Model (1); the log likelihood increasessignif-
perception on the part of respondents. In regard to icantly, and most interestingly,the stochastic brand
the differentialmechanismsthroughwhich credibility effect (the varianceof taste distributionthat had been
set to be the same for all brands) goes to zero (in
may operate in PC versus juice, in low involvement
categories, for example, consumers may have high juice) or becomes nonsignificant (in PCs). For these
sensitivity to informationcosts (they are not willing data, at least, it seems that the inclusion of the brand-
to bear informationcosts, even though the potential level credibilityconstructnot only accountsfor all the
costs may be low), hence ICS due to credibilitymay variabilitycapturedby the unobservedheterogeneity
be a more importantfactor in the choice set genera- in tastes, but also considerably improves goodness-
tion stage for juice than for PCs (in which consumers of-fit.8The credibilityimpact on the estimated utility
are willing to bear some information costs). There function, confirming prior results from Erdem and
is ample empirical evidence, for example, that con- Swait (1998),is strong and positive.
sumers are very sensitive to informationcosts in low Model (3) is the final baseline model: we substi-
involvement product categories. For example, Inter- tute the component credibility constructs (PQ, PR,
net shopping agents provide a great deal of infor- and ICS) into the utility function in place of their
mation about different retailers even in the case of common antecedent.All three constructsare found to
ratherhomogeneous products (e.g., books and CDs), have strong and statisticallysignificantimpactsin the
but consumersare willing to pay price premiums for expected directions. To contrast Models (2) and (3),
credibleretailerssince they do not want to invest in note that in both product categories the latter model
has a much better log likelihood value than the for-
comparingdifferentonline retailers(see, e.g., Iyer and
mer, at an additionalcost of only one parameter.This
Pazgal 2003).
suggests that the simple use of the brand credibil-
3.1.3. Model Estimation Results. Tables 2 and 3 ity construct to explore how brand effects are actu-
present parameterestimation results7for five choice ally operationalizedfor consumers is less insightful
models for juice and PCs, respectively.The first three than using its succeedingconstructs,PQ, PR,and ICS.
models for each product category are intended (a) to However, Model (3) also highlights the reduced form
provide baseline comparisons,(b) to make a specific nature of the baseline models: although all three suc-
point about brand-levelstochasticheterogeneity,and ceeding constructsare statisticallysignificantin both
7 Parameter estimation was done 8 It should be noted that, while the literature on brand choice mod-
by simulated maximum likelihood
(SML), with 150 quasirandom Halton replications used throughout. els estimated on scanner panel data shows significant unobserved
We refer the reader to Keane (1993) for an overview of SML esti- heterogeneity in tastes, those models typically do not incorporate
mation methods. Note that the log likelihood function considered behavioral or attitudinal observables that vary across consumers,
each choice replication of the individual to be independent, some- thus leading to unobserved heterogeneity. A complementary expla-
thing necessary by the very structure of the choice set formation nation for this result is that our subjects are quite homogeneous in
models, which permit the choice set to vary from one replication age and other sociodemographic characteristics, which would not
to the next. be the case in a scanner panel.

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set Formation Under Uncertainty
MarketingScience26(5),pp. 679-697,©2007 INFORMS 689

product classes, these models in essence say nothing reduction;ICSseems not to have played a statistically
more than that the utility of a product rises in qual- significant role in screening PC brands or products.
ity, drops in risk, and rises in saved decision costs. Note that the nonsignificanceof brand credibilityin
Model (3) simply confirms the structural equation choice set formation in Model (4) may well have to
models of Erdem and Swait (1998),without any fur- do with the opposite valences of PR and ICS in juice
ther explanation of how brand credibility works its (or PR and PQ in PCs) in Model (5), a consistenteffect
impacts.Since the purpose of our paper is to explore in both product classes.
the impact of brand credibilityon choice set forma- Turning our attention to the utility function in
tion and preferencediscrimination,Models (l)-(3) of Model (5), it will be noted that only PQ is included
Tables2 and 3 cannot help us in this regard. in this model component.In both productclasses, the
Instead,we must look to Model (4) in these tables, inclusion of PR and ICS in the utility function, over
which implementsthe IALmodel specified in Expres- and above their inclusionin other model components,
sions (2a-2e) and (3a-3c). This model includes brand did not result in statisticallyimproved models;hence,
credibility in the choice set formation availability we decided to present only these simpler versions
functions, in the effect-size function, and in the util- with PQ. It is possible that this result is due to mul-
ity function. Note that, consistent with models dis- ticollinearitybetween these three constructs;Study 2,
cussed earlier, no statistically significant stochastic because of a different experimental design, will be
brand heterogeneity is found in either product cat- useful in clarifying this issue. Quality, as expected,
egory. Somewhat unexpectedly, brand credibility is is found in the present data to have a strong pos-
found to have a nonsignificantimpact in choice set itive impact on brand comparisons within the indi-
formation;this result could be explained in juices by vidual's choice set. Note also the maintainedsupport
the fact that because this is a low-involvement prod- for stochasticbrand homogeneity with the inclusion
uct categoryabout which these respondentsmay feel of perceived quality in the utility functions.Again, it
they'renot well informed, brand credibilitydoes not is notable that the explanatorypower of Model (5) is
serve as the basis for an early brand triage;however, significantlysuperiorto that of Model (4), suggesting
the result is also found in PCs, so we must look for the usefulness of decomposing the impact of brand
an explanation elsewhere (we will elaborate on this credibilityinto that of its successor constructs.
later in this section). Note, however, that in the util- Withrespectto preferencediscrimination,the effect-
ity function of Model (4), in both products credibil- size functions t of both product categories show
ity is found to have the expected strong and posi- themselves to increase significantly with increasing
tive effect on utility; in addition, the effect-size func- PQ, decreasing PR (for juices, but not PCs), and
tion r shows that preferencesare better discriminated increasing ICS. Both PQ and ICS display a strong
among brandsthat have high credibility,and that this increasing effect through the tested range of these
effect has a somewhat marginallydiminishingimpact constructsin the juice category,without any amelio-
as credibilitygrows. rationof the impact at higher constructlevels; PR has
Model (5) substitutes the successors of credibility, the same type of impact for juices, but in the oppo-
PQ, PR,and ICS,into the two-stage choice model sys- site direction,as expected. These opposite effects may
tem (Expressions(2a)-(2e) and (4a)-(4c). This model well account for the overall diminishing marginal
allows us to explore the mechanismswhereby brand impact of brand credibilityon preferencediscrimina-
credibility differences are translated into impacts in tion, as found in the effect-sizefunction of Model (4).
the different choice stages. In the brand choice set For PCs, preference discriminationis also found to
inclusion functions, Model (5) shows that PQ has a be strongly affected by credibility (Model 4), with a
statisticallynegligible impact on brand inclusion in diminishingmarginalimpact.Model (5) indicatesthat
the choice set for the juice category,but an impor- PQ and ICS are the likely means whereby preference
tant role in the PC class; in both product categories, discriminationis actually accomplished in PCs; PR
increasing PR decreases the probability of brand seems to have played its importantrole in the choice
inclusion,althoughthe impact is more pronouncedin set formationprocess.
juice than in PCs;finally,increasingICS increasesthe It is helpful to visualize the differentimpacts cap-
likelihoodof inclusionof the brandin the individual's tured in Model (5) through some simple graphs. In
juice choice set, but the correspondingimpact in PCs Figure 2, we present the estimated brand inclusion
is not statisticallysignificant at the 95%level. Thus, probabilitiesfor Brand 1, both for juice and for PCs.
in the juice category,quality seems not to have been (Brand 1 was arbitrarilyselected for this illustration
employed by these respondentsin their screening of among the five presented, and it representsa differ-
alternatives,but PR and informationprocessingcosts ent brand in each productcategory.)The probabilities
were used as screeningcriteria.In the PC class, how- of inclusion in the latent choice set are presentedas a
ever, quality was a screening criterion, as was risk function of price and credibility(set at low, medium,

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand ChoiceSet FormationUnder Uncertainty
690 MarketingScience26(5),pp. 679-697,©2007 INFORMS

Figure2 as a Function
ChoiceSet Formation of PriceandBrandquite a bit largerin the PC category than in the juice
Credibility category.
While detailed comparisonof the two chartsin Fig-
ure 2 leads to the conclusion of the greater relative
overall importanceof brand credibilityfor choice set
formation in the PC category compared to the juice
category,the simultaneous use of the coefficientsof
Model (5) will suggest how these impacts are cre-
ated via PQ, PR, and ICS. In PCs (see Table 3), for
example, we have alreadynoted that brand inclusion
probabilitiesare mainly affected by PQ and PR, and
only marginallyby ICS.Furthermore,between quality
and risk, quality is by far the greaterdeterminantof
brand inclusion. This suggests that the role of brand
credibility in this first stage of choice in this high-
involvement,high-priceproductcategoryis to permit
a quality-basedscreeningof alternativesto occur;PR,
on the other hand, has only something like one-third
the impact of PQ in this stage. In juices, by contrast,
brand inclusion is strongly and about equally deter-
mined by PR and ICS, while PQ has no impact to
speak of. This suggests that for choice set formation
in the juice category,brand credibilityis the basis for
economizingon informationcosts, as well as risk min-
imization;quality will play a role subsequently dur-
ing the comparativeevaluation of brands within the
choice set.
Figure3 is shown to describehow choice probabil-
ities, conditionalon choice set formation,are affected
by credibility.Note that the conditional choice prob-
abilities that we have estimated are not only a func-
and high levels). For a given level of credibility,PQ, tion of brand utility, but also of effect size, which
PR,and ICSare estimatedaccordingto the simultane- in turn are affected by PQ, PR, and ICS. The prefer-
ous equationmodels (a differentone for each product ence discriminationeffect we hypothesized and illus-
category) alluded to before. In both product cate- trated in Figure 1 is quite evident in Figure 3, where
gories, inclusion of Brand1 in the subject'schoice set we show for both product categories the conditional
is predictedto become less likely as price is increased choice probabilityfor Brand 1 (again arbitrarilycho-
and as credibilitydecreases.While a cursoryglance at sen) as a function of price and credibility (varying
Figure2 might suggest that the overallimpactof cred- from low to high). (We once more use the simultane-
ibility on brand inclusion in the choice set is greater ous equation model to cascade the effects of credibil-
in the juice category than in the PC category,closer ity via the PQ, PR, and ICS constructs.)These choice
examinationwill find that to be a misapprehension. probabilitiesare calculated using the estimated util-
Specifically,for these two brands calculation of the ity and effect size functions for Brand 1, assuming
odds of inclusion in the choice set for medium- and a binary choice against an aggregate good with zero
high-credibilityperceptions of the brands, compared utility. The impact of credibilityis clearly discernible
to low-credibilityperception of the brands, varies as in both product categories, but is quite dramaticin
follows: (a) for juices, medium-credibilityperception the juice category.The greatersensitivity of Brandl's
makes it 1.6 to 2.7 times more likely that the brand is conditional choice probabilityin juices, compared to
included over the price range shown than is the case PCs, is in large part due to the effect-size function
for a low perceptionof credibility,while the odds vary in the former product category: in the juice cate-
from 2 to 6.4 times for a high-credibilityperception gory, credibilityhas strong discernibleimpactsvia all
of Brand 1 comparedto a low-credibilityperception; three credibilitysuccessor constructs,whereas in the
(b) in PCs, medium-credibilityodds vary from 6.5 PC category,credibility'simpact is mostly restricted
to 18.5 times greaterlikelihood,while high-credibility to informationcosts. This may reflect the possibility
odds vary from 9 to 220. Thus, the overall impact of that once the choice set is formed, brand credibil-
credibilityon choice set formationis estimated to be ity actually plays a more determinantrole in juices

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set Formation Under Uncertainty
MarketingScience26(5),pp. 679-697,©2007 INFORMS 691

Figure3 BrandCredibility,
PreferenceDiscrimination
and confirm these results in the PC category,where one
Conditional
ChoiceProbability
expects the more substantive brand impacts to be
present; it was also conducted to test their robust-
ness to certainchallengingconditions:choice contexts
with more brands than used in Study 1, and with
choice set sizes varying from scenario to scenario.In
addition, Study 2 also has the objectiveof collecting
informationto support analyses of choice set forma-
tion versus choice behavior in a mannerindependent
of the IAL model, thus permitting a check on the
validity of the latent choice set formation model as
a means to exploring the impact of brand on differ-
ent stages of choice. As will be subsequently seen,
this design change proved quite useful in elucidating
certaineffects that remainedsomewhat ambiguousin
Study 1.
3.2.1. Data Collection. Using a survey instrument
very similar to that of Study 1 (with certain notable
differencesto be discussed), 287 undergraduatebusi-
ness students at two North American universities
were used as sources of data for Study 2.
The main differencesbetween the two instruments
are as follows: (a) Study 2 involved only PCs,whereas
Study 1 also had juices, (b) The experimentalchoice
task for Study 2 used eight national brands plus
a local computer store generic brand (or a total of
nine brands), while Study 1 had only five brand,
(c) WhereasStudy 1 had fixed choice sets of six alter-
natives (always showing the same five brands plus
the "None"alternative),Study 2 used choice sets with
as few as two and as many as nine brands, in addi-
tion to the "None" alternative.9(d) WhereasStudy 1
than in PCs, aiding the consumermore in making the elicited only the single final choice from respondents,
final decision among the reduced set of brands. In in Study 2 they were requested to provide the single
PCs, however, brand credibility seems to have been final choice alternativeas well as all other alternatives
strongerin determiningthe choice set itself, and then seriously consideredfor choice.
its role in the second-stage evaluative and compara- Subjectsin Study 2 also provided responses to the
tive process is basicallylimited to the impact of qual- same items as shown in Table 1. Constructestimates
ity differences. were calculatedin an identical fashion to calculations
Finally,it is worth noting the relative importance in Study 1.
of the subconstructsof brand credibility,trustworthi- 3.2.2. Modeling Results. Only three models were
ness (Tr), and expertise (Ex) in the two categories estimatedfrom the data of Study 2, and these are pre-
we analyze. An analysis of the simultaneous equa- sented in Table4. The firsttwo models are straightfor-
tion system involving the indices for Tr,Ex, PQ, PR, ward:the firstmodel is a binarylogit model estimated
and ICS(model availableupon request)indicatesthat on the elicited considerationdata, and confirms that
the majorimpact of credibilityon the latterthree con- the PQ, PR, and ICS constructs have a statistically
structsis via expertisein the juice category(this is not
important impact on consideration,in the expected
to say that trustworthinessdoes not play some role), directions;the second model is a simple conditional
but is balanced between trustworthinessand exper- MNL model with effect-size function, based on the
tise in the PC category.
9 The
3.2. Study 2 experimental design for Study 2 involved 52 different runs.
These were initially generated randomly, then an exchange algo-
As indicatedbefore,Study 1 focused on exploringthe rithm was used to improve design characteristics. The design was
validity of certainbrand impacts on differentcompo- blocked into four groups of 13 runs. The brands used in Study 2
nents of the choice process across two widely vary- were Apple, Compaq, Dell, Gateway, HP, IBM, Sony, Toshiba, plus
ing product categories. Study 2 was conducted to the generic moniker "local computer store."

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set FormationUnder Uncertainty
692 MarketingScience26(5),pp. 679-697,©2007 INFORMS

Table3 ResultsforPCs(Study1)
Estimation
f-statistics)
(Asymptotic
(b) MNL (d) IAL (e) IALPQ,PR,
(a) MNL and
credibility (c) MNL and
credibility ICSand
brandhet. brandhet. PQ,PR,ICS brandhet. brandhet.

function(Vin)
Utility
Brand1 50.063(7.7) 38.603(14.1) 36.078(37.6) 40.325(11.0) 26.779(3.0)
Brand2 49.92 (7.7) 39.136(14.1) 37.034(38.3) 40.748(11.2) 21.299(6.0)
Brand3 52.457(7.7) 40.256(14.1) 37.68 (38.6) 40.765(11.1) 20.983(6.1)
Brand4 50.077(7.7) 39.105(14.1) 36.901(38.1) 40.395(11.1) 21.498(6.0)
Brand5 51.603(7.7) 39.698(14.1) 37.188(38.3) 40.826(11.5) 21.081(6.2)
In(Price) -7.224 (-7.5) -5.554 (-13.5) -5.208 (-37.2) -5.589 (-11.0) -2.898 (-6.2)
Credibility 0 0.798 (14.2) 0.99 (7.0)
PQ 0.486 (15.4) 0.142 (2.3)
PR -0.061 (-2.6)
ICS 0.407 (11.3)
(rb2rand 4.065 (2.4) 0.625 (1.5) 0 0 0.007 (0.1)
Effect-size
function(Inrln)
Brand1 -0.295 (-2.7) -1.9 (-1.8)
Brand2 0.246 (2.1) -0.051 (-0.2)
Brand3 0.06 (0.8) 0.319 (2.7)
Brand4 0.035 (0.3) -0.283 (-0.9)
Brand5 0 0
Credibility 0.236 (4.3)
Credibility*^ -0.02 (-1.5)
PQ 0.103 (2.0)
PQ"2 0.013 (1.0)
PR -0.017 (-0.9)
PR**2 0.003 (1.0)
ICS 0.416 (4.3)
ICS"2 -0.085 (-3.5)
Brandinclusion
functions
Brand1 3.279(6.4) 6.417(6.8)
Brand2 2.656 (6.6) 7.805 (7.7)
Brand3 4.748 (9.4) 9.883 (7.7)
Brand4 3.284 (6.4) 7.837 (7.4)
Brand5 3.664 (8.5) 8.833 (7.7)
None ooc 0.170 (0.2)
Price -2.623 (-6.4) -0.007 (-10.2)
Credibility -0.101 (-1.5) 0
PQ 0.637 (6.3)
PR -0.188 (-2.9)
ICS 0.212 (1.6)
LL(convergence) -3,771.08 -3,294.8 -3,089.92 -3,152.50 -3,034.50
Akaike
Rho-squared 0.2575 0.3509 0.391 0.3764 0.3982
Numberof parameters 7 8 9 20 28
Notes,(a)AllSMLestimatesbasedon R= 150 Haltonreplicates, foreachindividual
(b)Allconstructindicesaremeancentered respondent,
intheavailability
(c) Infinity function is alwayspresent.Thisconstraint
thatalternative
indicates (d)Stochastic
wasimposedduringestimation,
brandeffectestimatedfromIIDnormalvariates.(e) Number 6,194choicesfrom366 respondents.
of observations:

stated choice sets, again confirming that these con- of brands present from one scenario to another,
structs(and, hence, brand credibility)have a substan- whereas in Study 1 the number of brands and the set
tive impact on utility,conditionalon stated choice set. of brands shown were held constantacross scenarios.
The only statisticallysignificanteffect among the con- (The conditional MNL model does not include size
structsof interestin the effect size is PR (see §4). effects in the utility function due to our expectation
The binary considerationmodel in Table 4 differs that product utility should not be a function of num-
from the models in Tables 2 and 3 from Study 1: ber of brands.)Of interesthere is that the likelihoodof
it includes several choice set size effects that are of the brand being considered or included in the choice
interest.The reason they can be identified in Study 2 set decreaseswith number of brands, suggesting that
is due to the explicit variationin brands and number context complexity leads to choice set size reduction,

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set Formation Under Uncertainty
MarketingScience26(5),pp. 679-697,©2007 INFORMS 693

ceterisparibus.In addition,it is found that the impact in the IAL frameworkof Model (3), and see how well
of PQ as a screening mechanism for brand inclusion this combinationpredictsthe unconditionaldata used
in the stated choice set increaseswith the number of to estimate Model (3). The rationale for doing this
brands, whereas the impacts of PR and ICS are not is that the coefficientsof Model (1) are analogous to
sensitive to this context variable. Turningour atten- those of the brand inclusion function of Model (3),
tion to the MNL model conditional on stated choice, while those of Model (2) correspondto the utility and
we find that number of brands affects the PR con- effect-size functions of Model (3). From Table 4, it
struct negatively: as the number of brands increases, can be seen that the log likelihood at convergenceof
the effect size function decreases in PR. This sug- Model (3) is -4,174.75, which is the global optimal
gests that as risk perceptionsincreasein the presence value achieved by the maximum likelihood estima-
of more brands, preferencediscriminationdecreases. tor; the application of the coefficients of Models (1)
Hence, this model suggests that it is possible that and (2) in the IAL model is necessarily worse, at
as the number of brands presented increases,respon- -4,314.76. The Pearson chi-squaredstatistic, a mea-
dents behave as if nonsystematicfactorsplay a greater sure of asymptoticnormalizedresidualsbehaviorthat
role in defining product utility. converges to unity in a well-specified model, is 1.12
The third model in Table 4 is analogous to the for Model (3) and a somewhat higher 1.27 for the
specificationof Model (5) in Table 3: it is a full IAL application of Models (1) and (2) coefficients.These
model with person-brandheterogeneity.However, it results show that the structuralmodel using latent
differsfrom Model (5) in Table3 due to the inclusion choice sets has better performanceon the common
of a number of brand effects in the brand inclusion data, suggesting that the IAL model better captures
and effect-size functions. This choice set formation the underlyingbehaviorthan do the two partialmod-
and choice specificationon Study 2's data basically els. Overall, the differences in performanceare not
confirms the effects previously observed in Study 1: overwhelming,but they are substantivelysignificant.
brandcredibilityhas a discernible,substantiveimpact In addition, it will be noted that particularparam-
on all three components of the model (choice set for- eters are sometimes markedly differentbetween the
mation,effectsize, and choice),in expected directions. model systems:e.g., in the brandinclusionfunctionof
The number-of-brandscontext variableshows that as Model (3), the impactof ICSis estimatedto be negligi-
this increases, ceteris paribus, (a) any brand is less ble, whereas in Model (1) there is a statisticallysignif-
likely to be considered;(b) the "None" alternativeis icant effect estimated for this construct:the structural
morelikely to be considered(anotherresponseto con- choice set formation model results suggest that ICS
text complexity);(c) the PR constructplays an increas- plays no role in brand inclusion,but the more limited
ingly stronger role as a screening mechanism, while brand considerationmodel suggests the contrary.We
the impact of PQ and ICS is not sensitive to these believe that the former results are more credibledue
increasesin contextualcomplexity;and (d) preference to explicit modeling of the choice set formationin the
discriminationis less negatively affected, suggesting IAL model.
that this constructis correlatedwith a lower reliance There are some noteworthy comparisons that can
on systematic utility sources when more brands are be made between the two PC IAL models of Stud-
being presented. ies 1 and 2. In making these comparisons,it is impor-
It is informativeto examine whether using the full- tant to keep in mind that the essential design differ-
informationstructuralchoice set formationModel (3) ences between the studies is that the latterhas signifi-
of Table4 is superior to the use of the separate con- cantly more brands and the numberof brands shown
siderationand conditional choice Models (1) and (2) in a scenariois variable.In the brand inclusion func-
of the same table. It is not possible to directly com- tion, Model (3) of Table 4 indicates that PQ and PR
pare log likelihoods or performnested statisticaltests have significantroles in determiningchoice set mem-
because the data structures utilized for the models bership; this is essentially the same result shown in
differ:Model (1) uses every alternativepresented,but Model (5) of Table 3, where ICS is not statistically
in a yes or no format;Model (2) is restrictedto the significantat the 95%confidence level. In the utility
use of the one choice set composed only of those function of Model (5) in Table 3, PQ is statistically
alternativesindicated to have been seriously consid- significantbut the other constructsare not (as men-
ered; and Model (3) uses all alternatives presented, tioned before, this result may have been caused by
in a choose-one-from-manyformat, but with explicit multicollinearitybetween the three constructs,lead-
structural(and latent) consideration of every possi- ing us to present the best model with a single con-
ble choice set that can be formed from the alterna- struct); all three constructs are cleanly identified in
tives presented in the scenario.However, one means Study 2, showing that PQ and ICS continue to play
of comparingthese models is to apply the estimated significantroles in brand comparisons,conditionalon
coefficientsfrom the nonstructuralModels (1) and (2) choice set. This improvement in the estimability of

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand ChoiceSet FormationUnder Uncertainty
694 MarketingScience26(5),pp. 679-697,©2007 INFORMS

Table4 Estimation
ResultsforPCs(Study2)

(Asymptotic
f-statistics)
logit
(a) Binary (b) MNLchoicemodelconditional (c) IALwithbrand
consideration
model set
on statedconsideration heterogeneity
function(Vin)
Utility
Brand1 17.351 (5.26) 14.195 (2.39)
Brand2 17.959(5.35) 14.715(2.39)
Brand3 18.459(5.27) 14.809(2.40)
Brand4 17.354(5.28) 14.513(2.39)
Brand5 17.68 (5.29) 14.091 (2.39)
Brand6 17.868(5.25) 14.473(2.39)
Brand7 18.316(5.25) 15.398(2.39)
Brand8 18.447(5.27) 15.320(2.39)
Brand9 17.877(5.63) 17.257(2.41)
In(Price) -2.493 (-5.30) -2.043 (-2.39)
PQ 0.396 (4.24) 0.647 (2.36)
PR -0.132 (-1.82) -0.096 (-1.55)
ICS 0.126 (0.91) 0.640 (2.30)
Var(brand constants) 0 (0)
Effectsizefunction(In%)
Brand1 0.114(0.37) 0.754(4.02)
Brand2 0.097(0.55) 0.664(5.71)
Brand3 0.158(0.87) 0.615(5.38)
Brand4 0.216(1.03) 0.752(4.90)
Brand5 0.199(1.13) 0.703(6.14)
Brand6 0.161 (0.85) 0.668 (5.74)
Brand7 -0.045 (-0.24) 0.374(3.10)
Brand8 -0.161 (-0.83) 0.384(3.25)
Brand9 0 0
PQ 0.131 (1.87) -0.01 (-0.13)
PQA2 -0.016 (-1.65) 0.033 (5.26)
PR 0.187 (3.36) 0.087 (1.81)
PRA2 0.006 (1.06) 0.005 (1.81)
ICS -0.046 (-0.46) -0.158 (-1.74)
ICSA2 0.015 (1.35) 0.030 (3.08)
Size -0.021 (-1.15) 0.114 (2.24)
Size* PQ 0.005 (0.61) -0.018 (-2.24)
Size*PR -0.021 (-3.89) -0.005 (-0.94)
Size*ICS 0.004 (0.44) -0.004 (-0.40)
Brandinclusion functions
Brand1 1.041 (9.00) 0.470 (0.96)
Brand2 2.204 (22.4) 1.039 (3.36)
Brand3 2.909 (28.06) 2.281 (7.42)
Brand4 1.597 (15.56) 0.503 (1.46)
Brand5 2.003(20.73) 1.762(5.51)
Brand6 1.946 (18.80) 1.402 (4.78)
Brand7 2.167(20.36) 1.112(3.87)
Brand8 1.696(16.36) 0.779(2.54)
Brand9 1.307(12.64) -0.451 (-1.73)
None 0 -2.258 (-3.61)
Price/1,000 -1.401 (-19.02) -1.051 (-6.2)
PQ 0.470 (7.15) 0.388 (3.21)
PR -0.038 (-0.69) 0.111 (1.09)
ICS 0.254 (2.57) 0.016 (0.09)
Size[Brands 1-9] -0.138 (-16.3) -0.141 (-4.46)
Size[None] 0 0.567 (3.56)
Size[1-9]*PQ 0.032 (3.80) 0.007 (0.43)
Size[1-9]*PR -0.01 (-1.37) -0.039 (-2.79)
Size[1-9]*ICS 0.013 (1.03) 0.016 (0.68)
LL(convergence) -10,053.6 -3,019.63 -4,174.75
Rho-squared Akaike 0.3071 0.3116 0.3788
Number of parameters 17 31 51
Notes,(a)AllSMLestimatesbasedonR= 150Halton (b)Loglikelihoods
replicates, arenotcomparableacrossmodelssincedata
structures
areuniqueto eachmodel,(c) Allconstructindicesaremeancenteredforeachindividual (d) Stochastic
respondent,
brandeffectestimated
fromIIDnormalvariates.(e) Number of observations:
24,700choicesfrom287 respondents.

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set Formation Under Uncertainty
MarketingScience26(5),pp. 679-697,©2007 INFORMS 695

these parametersis due to the explicit variation of greater weight on quality differences among lower-
presence or absence of brands by the experimental quality brands than among higher-quality brands.
design used in Study 2. Finally,in terms of the effect- This could well be because in a changing, dynamic
size functions, the two studies indicate that PQ and environment that inhibits the development of rule-
ICS affect preferencediscriminationin the expected based choice set formation,and in which high-quality
directions(and that the impact of PQ decreasessome- brands may not be included, preferencediscrimina-
what with number of brands presented), while the tion (i.e., sharper tastes) will be discernible at low
impact of PR is not statisticallysignificantat the 95% quality levels, as opposed to high quality levels.
level. Our empirical results revealed statistically non-
Overall, Study 2 has confirmed all the important significant credibility construct effects at the choice
results of Study 1, while extending the latter's find- set formation stage. However, we believe this find-
ings by showing that they are robust to significant ing to be an outcome of the offsettingeffects of brand
changes in decision context complexity (number of credibilityon choice sets through PQ and PR in PCs,
brands and changing availableset composition). and ICS and PR in juice. Models containingonly the
end constructs(PQ, PR, and ICS) show that, indeed,
4. Discussion, Conclusions, and PQ and PR in PCs, and PR and ICS in juice have
Future Research statisticallysignificanteffects on choice set formation.
We have proposed that brand credibility will affect (Higher PQ, lower PR, and higher ICS increase the
the formation of choice sets, consumer preferences, probabilityof a brand being included in a choice set,
ceteris paribus.)
and preference discrimination.We have also inves-
Noteworthy also are the differential mechanisms
tigated the mechanisms through which credibility
effects materialize,namely through PQ, PR, and ICS. through which credibility operates at the choice set
formationstage in juice versus PCs. ICSis (not) signif-
We have found strong evidence for brand credi-
icant in juice (PCs),whereas PQ is (not) significantin
bility effects and differential mechanisms through PCs (juice).This is an interestingresult, which can be
which brand credibility's impact materializes on
choice set formation, brand utility, and preference explained by the differentialinvolvement and levels
discrimination. of uncertaintyand sensitivity to such uncertaintyin
In both the juice and PC categorieswe found that these two categories:in juice, consumersare not will-
PQ continuedto play a significantrole in defining the ing to accrueinformationcosts (sensitivityto informa-
tion costs is high) even though potential information
utility of brands,even aftercredibilityeffects are cap- costs are likely low, whereas in PCs, consumers are
tured at the choice set formation stage. In addition,
Study 2 (because of its different and more-flexible likely to be willing to bear higher informationcosts.
study design) results suggest that in the PC category, Furthermore,in the juice category, PQ differences
ICS has just as important a role in defining utility among brands may not be large enough for PQ to
as quality.Risk perceptions,however, have negligible determinethe likelihood of a brand to be included in
the choice set, whereas in the PC category,consumers
impact at this evaluative stage, conditionalon choice
set formation.Comparingthese findings with those of may perceive larger differences across brands. Our
a simple MNL model with no explicit choice set for- results also indicate that credibility'simpact on con-
mation process modeled, which suggest that all three sumer choice processes through PR, PQ, and ICS are
constructs play a role in determining product util- mainly due to the expertisedimension of brand cred-
ity (see Models (3) in Tables2 and 3), highlights the ibility in juice, whereasboth subcomponentsof brand
need for an explicit, structuralchoice set formation credibility (expertise and trustworthiness)are affect-
model to understandthe mechanismswhereby credi- ing consumerchoice processes in the PC category.
bility affectsbrand choice. Theseresultsappearreasonable,given thatrespond-
In both categories,we found that credibilityaffects ents rated PCs to be a more complex and higher-
preferencediscrimination.We found that choice set involvement category with more imperfectlyobserv-
size moderatesthe effect of brand credibilityon pref- able attributesthan juice. The respondentsalso indi-
erencediscrimination:higherbrandcredibilityis asso- cated that they are likely (a) to know less about PCs
ciated with higher preference discrimination when than about juices, but (b) to be more willing to gather
choice set sizes are constant or the universal brand informationto make a decision about PCs than about
set remainsunchanged across choice instances (as in juices.
Study 1). When choice set size varies greatly (as in Our overall results suggest that management of
Study 2), lower brand credibility(throughlower per- credibilityis a key issue in brand management.This
ceived quality) is associated with higher preference is, of course,not a new insight:others had previously
discrimination,suggesting that decision makersplace pointed out the importanceof credibility(e.g., Aaker

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand ChoiceSet FormationUnder Uncertainty
696 MarketingScience26(5),pp. 679-697,© 2007INFORMS

1991). However, our results also show the mecha- characteristicsin the brandinclusionprobabilities,our
nisms through which credibilityoperates in different results will be unaffectedby the omitted correlations,
contexts. For example, to encourage a brand to be but if the unobserved characteristicsare also corre-
included in the choice set in a low PR-low involve- lated with the brand characteristicsthat parametrize
ment category such as juice would require focusing the inclusion probabilities,our estimated effects in
on encouraging the brand to be used as a choice those probabilities could be either under- or over-
heuristic, principally by decreasing consumer infor- stated. In addition, our modeling assumes that the
mation costs, whereas in a high involvement-high PR MNL model (and consequently,IIA) holds whatever
category such as PCs, marketingefforts would need the structureof the choice set. In reality,correlations
to be geared mainly toward increasingmean quality between the errorterms of differentalternativesin a
expectations,and to some extent decreasingthe vari- choice would lead to more-complexmodels of choice
ance of such expectations. conditional on choice set. Future researchcan prof-
In regardto the relativeimportanceof trustworthi- itably investigatewhether less-restrictiveassumptions
ness versus expertise in juice versus PC, the results in the conditionalchoice model are warranted.
obtained from the simultaneous equation system Anotherlimitationimposed on both our studies has
involving the indices for Tr, Ex, PQ, PR, and ICS, to do with the measurementof credibilityand its use
along with the results from Tables 2 and 3, sug- in conjunctionwith the choice experiments. In our
gest certaindirectionsfor marketingeffortsthat differ method, we measured credibilityof each brand prior
by product category.In juice, for instance, consumer to observing choices made in designed choice sets.
communicationsshould emphasize quality control to During the choice experiment,however, respondents
increasebrand credibility.Accordingto the results of may have been exposed to brand and price combina-
Table2, this should (a) enhance the use of the brand tions that were, to them, incompatible- e.g., a high
as a choice set formationheuristic,leading to greater credibilitybrand at a very low price, which might be
likelihood of brand inclusion in the choice set; and interpretedby respondents as inconsistent with the
(b) improve qualityperceptionsvis-a-vis otherbrands perceived credibilityof the brand, hence leading to a
in the choice set, leading to greater odds of being revision of their credibilityperceptions.To the extent
chosen. that such revisionsmight occurwithin the contextof a
On the other hand, in a high-involvement, high- 20-minutepaper-and-pencilsurvey,it seems to us that
price category such as PCs, these results suggest that the majorimpactwould be on the effect-sizefunction,
advertisingand other marketingactions should build which capturesthe relativerole of unobservedcharac-
up perceptionsof both expertiseand trustworthiness. teristicson utility.Furtherresearchon the occurrence
Thus, the message to consumers has to be broader and magnitude of such an effect is needed.
and deeper than in the case of juices: to be sure, the Thereare several other avenues for futureresearch.
manufacturer'sexpertiseneeds to be transmitted,but Our analysis can be extended to include dynamics to
it is just as necessaryto communicatewith consumers explore issues such as choice set formationover time
about such topics as the firm's longevity in the mar- and the dynamics of credibility formation. A larger
ket, the firm'slikelihood of futureexistence,customer number of product categories can also be studied
satisfactionwith prior purchases, customer support to draw more extensive empiricalgeneralizations,as
afterpurchase,etc. well as to explore factors that may moderate brand
We wish to note here an important assumption
credibility effects. Finally, our study is one of the
in our model development that may impose limita- first attemptsto link behavioralprocessesto the mea-
tions on our conclusions. Specifically,the choice set surement of the stochasticcomponent of utility. Fur-
formation stage in the statistical model (Expression ther study of such behavioralmechanismswould be
(2d)) assumes that the inclusion or exclusion of an fruitful.
alternativefrom the choice set is completely uninfor-
mative about the inclusion or exclusion of another
alternative.This assumption was made to maintain References
model tractabilityand make estimation of the statis- Aaker, D. A. 1991. Managing Brand Equity. The Free Press, New
tical model feasible. However, in reality,brands may York.
share unobserved characteristicsthat cause inclusion Allenby, G. M, J. L. Ginter. 1995. The effects of in-store displays
and feature advertising on consideration sets. Internat. J. Res.
probabilitiesto decrease or increase together. In our
models, inclusion probabilitiesare parametrizedby a Marketing12(1) 67-80.
number of observed brand characteristics,but poten- Andrews, R., T. C. Srinivasan. 1995. Studying consideration effects
in empirical choice models using scanner panel data. /. Mar-
tially unobserved characteristicsare omitted due to keting Res. 32 30-41.
the independence assumption. To the extent that the Ben-Akiva, M., B. Boccara. 1995. Discrete choice models with latent
unobserved characteristicsare orthogonal to brand choice sets. Internat.J. Res. Marketing12 9-24.

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions
Swait and Erdem: Brand Effectson Choiceand Choice Set Formation Under Uncertainty
MarketingScience26(5),pp. 679-697,© 2007INFORMS 697

Ben-Akiva, M., S. Lerman. 1985. Discrete Choice Analysis: Theory Meyer, R. 1979. Theory of destination choice-set formation
and Applicationto PredictTravelDemand.MIT Press, Cambridge, under informational constraints. TransportationRes. Record750
MA. 6-12.
Boulding, W., A. Kirmani. 1993. A consumer-side experimental Montgomery, C, B. Wernerfelt. 1992. Risk reduction and umbrella
examination of signaling theory: Do consumers perceive war- branding. /. Bus. 65 31-50.
ranties as signals of quality? /. ConsumerRes. 20 111-123. Nelson, P. 1970. Information and consumer behavior. /. Political
Chiang, J., S. Chib, C. Narasimhan. 1999. Markov chain Monte Econom. 78 311-329.
Carlo and model of consideration set and parameter hetero- Nelson, P. 1974. Advertising as information. /. Political Econom. 81
geneity. /. Econometrics89 223-248. 729-754.
Dallaert, B., J. Brazell, J. Louviere. 1999. The effect of attribute Rao, A., R. W. Ruekkert. 1994. Brand alliances as signals of product
variation on consumer choice consistency. MarketingLett. 10(2)
quality. Sloan ManagementRev. 35(Fall) 87-97.
139-147.
Roberts, J., J. Lattin. 1991. Development and testing of a model of
Darby, M., E. Karni. 1974. Free competition and the optimal amount consideration set composition. /. MarketingRes. 28 429-440.
of fraud. /. Law Econom.16 67-88.
Roberts, J., P. Nedungadi. 1995. Studying consideration in con-
Erdem, T, M. P. Keane. 1996. Decision making under uncertainty: sumer decision process: Progress and challenges. Internat. J.
Capturing dynamic brand choice processes in turbulent con- Res. Marketing 12 3-7.
sumer goods markets. Marketing Sci. 15(1) 1-20.
Roberts, J., G. Urban. 1988. Modeling multiattribute utility, risk
Erdem, T, J. Swait. 1998. Brand equity as a signaling phenomenon. and belief dynamics for new consumer durable brand choice.
/. ConsumerPsych. 7(2) 131-157. ManagementSci. 34 167-185.
Erdem, T, J. Swait. 2004. Brand credibility and its role in brand Robertson, T. S., J. Zelinski, S. Ward. 1984. ConsumerBehavior.Scott,
choice and consideration. /. ConsumerRes. 31(June) 191-199. Foresman, Chicago, IL.
Erdem, T, J. Swait, J. Louviere. 2002. The impact of brand credibil- Shocker, A., M. Ben-Akiva, B. Boccara, P. Nedungadi. 1991. Con-
ity on consumer price sensitivity. Internat. J. Res. Marketing 19 sideration set influences on consumer decision making and
1-19. choice: Issues, models and suggestions. Marketing Lett. 2(3)
Godes, D., D. Mayzlin. 2004. Using online conversations to study 181-197.
word-of-mouth communication. Marketing Sci. 23(4) 545-560.
Shugan, S. 1980. The cost of thinking. /. ConsumerRes. 7 99-111.
Hauser, J., B. Wernerfelt. 1990. An evaluation cost model of consid- Spence, M. 1974. Market Signaling: Informational Transfer in Hir-
eration sets. /. ConsumerRes. 16 393-408.
ing and Related Screening Processes. Harvard University Press,
Heckman, J. J. 1981. Statistical models for discrete panel data. Cambridge, MA.
C. Manski, D. McFadden, eds. The StructuralAnalysis of Discrete Srinivasan, N., B. T. Ratchford. 1991. An empirical test of a model
Data. MIT Press, Cambridge, MA, 114-178. of external search in automobiles. /. ConsumerRes. 18 233-242.
Herbig, P., J. Milewicz. 1995. The relationship of reputation and Swait, J. 1984. Probabilistic choice set formation in transportation
credibility to brand success. /. ConsumerMarketing14 5-10. Demand models. Unpublished doctoral thesis, department of
Howard, J. A., J. N. Sheth. 1969. The Theoryof Buyer Behavior.Wiley, Civil Engineering, MIT, Cambridge, MA.
New York.
Swait, J. 2001. Choice set generation within the generalized extreme
Iyer, G., A. Pazgal. 2003. Internet shopping agents: Virtual value family of discrete choice models. TransportationRes.
co-location and competition. MarketingSci. 22(1) 85-106. Part B 35 643-666.
Keane, M. 1993. Simulation estimation for panel data models with Swait, J., M. Ben-Akiva. 1987a. Incorporating random constraints in
limited dependent Varaibles. G. S. Maddala, C. R. Rao, H. D. discrete choice models of choice set generation. Transportation
Vinod, eds. Handbook of Statistics. North Holland Publisher, Res. B 21B 91-102.
New York, 545-571.
Swait, J., M. Ben-Akiva. 1987b. Empirical test of a constrained
Keller, K. L. 2002. Brandingand Brand Equity. MSI, Cambridge, MA. choice discrete model: Mode choice in Sao Paulo. Transportation
Klein, B., K. B. Leffler. 1981. The role of market forces in assuring Res. B 21B 103-115.
contractual performance. /. Political Econom.89(4) 615-639. Swait, J., J. Louviere. 1993. The role of the scale parameter in the
Kotler, P. 1997. MarketingManagement,7th ed. Prentice Hall, Upper estimation and use of multinomial logit models. /. Marketing
Saddle River, NJ. Res. 30 305-314.
Louviere, J., D. Hensher, J. Swait. 2000. Stated ChoiceMethods:Anal- Tirole, J. 1990. The Theoryof IndustrialOrganization.MIT Press, Cam-
ysis and Applications in Marketing, Transportationand Environ- bridge, MA.
mental Valuation.Cambridge University Press, Cambridge, UK. Wernerfelt, B. 1988. Umbrella branding as a signal of new product
Manski, C. 1977. Structure of random utility models. TheoryDecision quality: An example of signalling by posting a bond. RAND J.
8 229-254. Econom.19 458-466.
Mehta, N., S. Rajiv, K. Srinivasan. 2003. Price uncertainty and con- Xie, J., S. M. Shugan. 2001. Electronic tickets, smart cards, and
sumer search: A structural model of consideration set forma- online prepayments: When and how to advance sell. Marketing
tion. MarketingSci. 22(1) 58-84. Sci. 20(3) 219-243.

This content downloaded from 139.86.7.217 on Sat, 17 Oct 2015 15:44:18 UTC
All use subject to JSTOR Terms and Conditions

You might also like