See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/239984438

Exploiting Rank Ordered Choice Set Data Within the Stochastic Utility Model

Article  in  Journal of Marketing Research · August 1982

DOI: 10.2307/3151563

CITATIONS READS
397 666

2 authors, including:

Richard Staelin
Duke University
143 PUBLICATIONS   17,825 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The Role of Paid, Earned and Owned Media in Building Entertainment Brands: Reminding Informing and Enhancing Enjoyment View project

A “Meta-Analysis” of Multibrand, Multioutlet Channel Systems View project

All content following this page was uploaded by Richard Staelin on 20 March 2015.

The user has requested enhancement of the downloaded file.

 Exploiting Rank Ordered Choice Set Data within the Stochastic Utility Model
Author(s): Randall G. Chapman and Richard Staelin
Source: Journal of Marketing Research, Vol. 19, No. 3 (Aug., 1982), pp. 288-301
Published by: American Marketing Association
Stable URL: http://www.jstor.org/stable/3151563
Accessed: 20-03-2015 15:33 UTC

REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/3151563?seq=1&cid=pdf-reference#references_tab_contents

You may need to log in to JSTOR to access the linked references.

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.

American Marketing Association is collaborating with JSTOR to digitize, preserve and extend access to Journal of Marketing
Research.

http://www.jstor.org

This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
I I

RANDALL
G. CHAPMANand RICHARD
STAELIN*

Theauthorsreporton a procedurefor exploitingthe informationcontentof rank

ordered choice sets to estimateefficientlythe parametersof the multinomiallogit
modelformulationof the stochasticutilitymodelof choice behavior.The!availability
of rankorderedchoiceset data leads to an "explosion"or decompositionprocedure
for exploitingsuchextra information.This"explosion"process involvesthe decom-
positionof a rankedchoice set into a series of unrankedand statisticallyindepen-
dent choice sets. In relationto explosion strategies, several heuristicsand an an-
alytical procedurefor determiningthe "optimal"explosion depth are discussedin
detail. The resultsof a Monte Carlo study of the small sample propertiesof the
conditionallogit estimationprocedure(the maximumlikelihoodestimationprocedure
used to develop parameterestimatesof the multinomiallogit model formulationof
the stochasticutilitymodel)are reportedand interpreted.A college choice empirical
applicationillustratesthe proceduresdeveloped.

Exploiting Rank Ordered Choice Set Data

Within the Stochastic Utility Model

Accordingto a popularchoice behaviormodelingpar- associatedwith estimatingthe parametersof such choice

adigm, the consumer decision maker evaluates each
availablealternativein terms of its componentparts,
assessingthe relativeimportanceof the componentsand
ultimatelychoosing the alternativewith the greatest
weighted aggregate score. Marketing scientists have
positeda varietyof preferenceandchoice modelsof how
consumerdecisionmakerscombinethe componentscores
to producea weighted aggregate score. We are con-
cerned with one general class of consumer choice
models, the stochasticutility model. Withinthe choice
modelingliterature,this model is commonlycited as the
Luce (1959) choice model; in actualempiricalapplica-
tion, it is typically parameterizedas the multinomial
logit model (cf. Gensch and Recker 1979). Econome-
tricians,most notablyMcFadden(cf. 1974, 1890), have
madesignificantstridesin refiningstatisticaltechniques
*RandallG. Chapmanis VisitingAssistantProfessorof Marketing,
GraduateSchoolof Business,Universityof Chicago.RichardStaelin
is T. Austin Finch Professor of Business Administration,Fuqua
School of Business,Duke University.
The helpfulcommentsandadviceof CharlesF. Manski,J. Edward
Russo, and an anonymousreviewer are most gratefullyacknowl-
edged. Responsibilityfor errorsof omission and commissionrests
with the authors.
288
models. Some illustrativeapplicationsof this choice be-
havior modeling methodologyinclude the selection of
a college (Chapman1979; Kohn, Manski, and Mundel
1976; Punjand Staelin 1978), a mode of transportation
(cf. Domencichand McFadden1975), a grocery store
(GenschandRecker1979), a shoppingcenter(Chapman
1980), a home (Li 1977), an occupation(Boskin 1974),
and an electricutility fuel (Joskowand Mishkin 1977).
Although the assumptionsassociated with the sto-
chastic utility model do not restrictits applicabilityto
situationsin which each consumer decision maker is
observedmakingjust a single choice from amonga set
of known alternatives,most marketingapplicationsof
the stochasticutility model are limited to those cases.
In such situations,the analysisis cross-sectionalandthe
econometricproblemis to assess the relativeimportance
of the quantifiableattributesthatdescribethe objectsof
choice, suchestimationbeingbasedon (theresearcher's)
knowledgeof each consumer's actual choice and the
compositionof each consumer'schoice set.
We develop a procedureto enhancethe estimationof
the parametersof the stochasticutility model. This en-
hancementis achievedby exploitingthe additionalin-
formationcontained in preference rank orderings of
choice set alternatives,dataoften availableto marketing
researchers.In the next section, several key aspects of

Journal of Marketing Research

Vol. XIX (August 1982), 288-301

This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
STOCHASTIC
UTILITY
MODEL
the stochasticutilitymodelof choice behaviorarebriefly
reviewed. The principleof decomposingrank ordered
choice sets into a series of statisticallyindependentun-
rankedchoice sets is explained. Strategies and tech-
niquesfor coping with "noisy" and possibly unreliable
rankorderinformationare then considered.The results
of a Mont6Carlostudydesignedto investigatethe small
samplepropertiesof the parameterestimatesof the con-
ditionallogit estimationprocedure(the maximumlike-
lihood estimationprocedureused to develop parameter
estimatesof the multinomiallogit model formulationof
the stochasticutilitymodel)arereportedandinterpreted.
Particularattentionis focused on the incrementalcon-
tributionof using the additionalinformationcontained
in the preferencerankorderingratherthanjust employ-
ing knowledgeof the chosen choice set alternative.We
concludewith an empiricalapplicationand a discussion
of the implicationsof these results for marketingre-
searcherswho use rankorderedchoice set data to esti-
mate the parametersof the stochasticutility model to
drawinferencesaboutchoice behavior.
THESTOCHASTIC UTILITY MODEL
The generalnatureof the choice behaviorbeing mod-
eled with the stochasticutilitymodel, and the natureof
the data(assumed)availableto the empiricalmarketing
researcher,may be describedas follows. Eachconsumer
decision maker i (i = 1,2,...,/) has a choice set Ci
consisting of Ji alternatives (1 < Ji < oo). The choice set
alternativesare assumedto be characterizedby N quan-
tifiable attributes.Each decision makeris observedto
choose an alternativefrom his or her choice set. The
decision makersare assumedto be utility maximizers
(i.e., rational)whose choices representtheir most pre-
ferredalternativesat the time of choice. Also, because
the data used typically are cross-sectional,the sample
of decision makers is assumed to have homogeneous
tastesandpreferencesin termsof the relativeimportance
of the attributescharacterizingthe alternatives.
Let xj denote a vector of relevantattributesof alter-
nativej, di denotea vectorof individualdecisionmaker
demographicattributes,and yijdenote a vectorof inter-
active variablesrelatingdecision makeri to alternative
j. It is assumedthat a utility function U exists which
measuresthe unobserveddesirabilityor attractivenessof
an alternativewith attributevectorxj to a decisionmaker
with demographicvector di and associated decision
maker/alternative vector yij.
(1) U, = U(xi,d,,y,)
Measurementerroris typical in the modeling process
becausethe x, d, andy vectorsgenerallydo not capture
all of the factorsinfluencingthe choice process, the cor-
rectfunctionalformfor the model may not be specified,
and individualsoften exhibit aspects of idiosyncratic
behavior.Thus, Uijis assumedto be composedof two
parts-a deterministic component,V(xi,di,yY),
representing
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
289
the systematic component of the model, and an er-
ror term e, which capturesthe measurementerrorsin
the modelng process. If one assumes these two com-
ponents are independentand additive, the model in
equation1maybewrittenintheform
(2) Ui.= Vy+ Eij
where V = V(xj,di,yi). The presenceof the stochastic
errorterm in equation2 leads to this model being de-
scribed as a stochastic utility model.
Supposethat individuali is observedto choose alter-
nativej* fromCi. If rationalchoice behavioris assumed,
revealed pre erence implies that Ui,*2 U.i (for j =
1,2,...,Ji). Because the utility function is partly sto-
chastic, the probabilityof this event occurringmay be
writtenas
(3) P,* = Prob(Ui, > U, j = 1,2,.. . ,J)
= Prob(Ei - ei* <
V,i - Vi, j = 1,2,...,i)
where Pi* is the probabilitythat decision maker i
chooses alternativej*. Furtherdevelopmentand simpli-
fication of equation3 requirethat a joint distribution
functionbe specified for the errorterms. In principle,
any joint distributionfunction could be used and an
expressionfor the choice probabilitiescould be devel-
oped. Unfortunately,the choice of most distributions,
includingthe usual normaldistributionassumptionfor
errortermsin statisticalmodels, necessitatesthe calcu-
lation of a formidableseries of numericalintegrations
to determineexplicitly the choice probabilities.How-
ever, if the stochasticerrortermsareassumedto be iden-
tically and independentlydistributed(IID) accordingto
the doubleexponentialdistribution,such that
Prob(Eij t) = exp [-exp (-t)],
(4)
one can show thatthe choice probabilitieshave the fol-
lowing form (cf. McFadden1974).
(5) PU*= exp(Vi,) forj* = 1,2,...,.
W ')*?- J, ,V
I
exp(V,)
j=l
The value of the double exponentialdistributionas-
sumptionis that a tractableclosed-formexpressionre-
sults for the choice probabilities.This particularpara-
metricformof the stochasticutilitymodelis often called
the multinomiallogit model because it is the multiple
choice generalizationof the binarylogit model.
To operationalizethe choice probabilityexpressionin
equation5, the functionalformof the deterministiccom-
ponentof the stochasticutilitymodel mustbe specified.
A linear-in-parameters specificationassumptionwould
lead to
N
(6) j
n=n=l ln
where Zijn is the measured value of attribute n for alter-
290
native j to decision maker i and Zij = Zijn(xj,di,,y), and
0, is the relative importance of attributen to the sample
of decisionmakers.The 0 values in equation6 are the
parametersof the stochasticutility model that must be
estimatedfrom the availablesamplechoice set data.
One particularfeatureof the stochasticutility model
with doubleexponentiallydistributederrortermsshould
be noted.Becausethe varianceof the doubleexponential
distributionis a knownfixed constant,not dependenton
the estimated0 values, the magnitudeof the estimated
Ovaluesdirectlyinfluencesthe "percentof varianceex-
plained" by the systematic component of the utility
function. As shown in the Appendix, the quantity
N
o2 = 02 is a measureof the "size" of the systematic
n=i
n=l
component. As (2 -? oo,the systematic component dom-
inatesthe errorcomponentand the choice probabilities
approachunityfor the alternativesin the choice sets with
greatestutility, with all other choice probabilitiesap-
proaching zero. Conversely, as -)2-) 0, the error com-
ponentin the stochasticutilitymodel dominatesthe sys-
tematiccomponentand each choice set alternativehas
approximatelyequal probabilityof being chosen. Other
detailsof these scale effects are given in the Appendix
and in subsequentdiscussion.
Values of the parametersof the multinomiallogit
model may be estimatedby maximizingthe likelihood
functionassociatedwith the probabilisticchoice model
in equation5. Standardsoftwarepackages(cf. Manski
1974) are availableto calculatethe maximumlikelihood
estimates.Because maximumlikelihood estimatesare,
in general,consistentand asymptoticallynormallydis-
tributed,approximatelarge sample confidence bounds
on parameterestimates may be constructedand hy-
pothesesmay be tested in standardways.
EXTENDING THE MODEL FOR
RANK ORDERED DATA
The estimationof the parametersof the stochastic
utility model requiresthe availabilityof the following
data from a representativesample of decision makers
from the populationof interest:(1) the alternativesin
each decision maker's choice set; (2) the actual alter-
nativechosen (i.e., preferred)by each decision maker;
and(3) the numericalvalueof each quantifiableattribute
associated with the choice set alternatives (i.e., the Zij
values).The modeloperateson the principleof revealed
preference:the alternativeactuallychosen by a decision
makeris assumedto be preferredto all otheralternatives
in the decisionmaker'schoice set. The basic parameter
estimation methodology can be extended if the re-
searcherhas available (or could conveniently gather
along with the otherrequireddata) a completerankor-
deringof all of the alternativesin the decision makers'
choice sets. To exploitthe informationcontentof a pref-
erencerankorderingof choice set alternatives,one must
relaterankingbehaviorto choice behavior.The theoret-
ical justificationfor relatingrankingbehaviorto choice
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
JOURNALOF MARKETING AUGUST1982
RESEARCH,
behavior,withinthe class of choice modelsof whichthe
stochasticutility model is a member,is providedby a
proof reportedby Luce and Suppes (1965, p. 354-6).
Althoughthe Luce and Suppes proof is for constant
utilitymodels (i.e., choice models in which each alter-
native has a fixed utility value and the probabilityof
choosingone alternativeover anotheris a functionof the
distancebetween their utilities), both the constantand
stochastic(random)utility models have an econometric
specificationsimilarto that of equation5 (cf. Luce and
Suppes 1965, p. 332-9). Consequently,the Luce and
Suppesproofrelatingrankingbehaviorto choice behav-
ior can be extendedto the lattermodel.
The Luce andSuppesRankingChoiceTheoremstates
thatfor any rankorderedpreferenceset which has been
derivedfrom a constantutility model,
(7) Pr(a,b,c,...)= Pr(alC)-Pr(b,c,...)
where Pr(a,b,c,...) is the probabilityof observingthe
rankorderof alternativea being preferredto alternative
b being preferredto alternativec, and so on, and Pr(al
C) is the probabilityof alternativea being chosen from
the set of alternatives C = {a,b,c,...}. This Ranking
Choice Theoremenables the probabilityof a ranking
event, Pr(a,b,c,...), to be decomposedinto the product
of two probabilities-the probabilityof a choice event,
Pr(aIC), and the probabilityof a subrankingevent,
Pr(b,c,...). By successively applying this Ranking Choice
Theoremto the subrankingevents, one can derive a
probabilityexpressionfor the rankingevent which is the
productof the probabilitiesof J - 1 choice events, i.e.,
(8) Pr(a,b,c, ...) = Pr(alC)'Pr(blC
- {a})
*Pr(clC- {a,b}) ...
whereC - {a} is the set of alternativesexcludingalter-
native a. Equation8 is equivalentto saying that the
probabilityof the joint rankingevent of J alternativesis
composed of J - 1 statistically independentchoice
events.
If one appliesthe RankingChoiceTheoremto the sto-
chasticutilitymodel, assumingthatthe alternativeindex
j is now interpretedas a serial preferenceindex, it fol-
lows that:
J,
(9) Prob (Ui, 2 Ui2 > ... UJ) = [I Prob (U, > Uij,
j*=l
for j= *,...,J).
The left side of equation9 is the joint probabilitythat
alternative1 is preferredto alternative2 which is pre-
ferredto alternative3, and so on to alternativeJi - 1
which is preferredto alternativeJ, for decision maker
i. The rightside of equation9 may be interpretedas the
statisticaldefinitionof the independenceof the events
(UilI
= = 2,3,...Ji),
Uij, 1,2,...Ji), (Ui2 Uij, j
..., (UiJ, Uij,).
The statistical independencecondition implied by
equation9 leadsto the notionof an "explosion" process
 UTILITY
STOCHASTIC MODEL
for exploitingthe informationcontentin preferencerank
orderedchoice sets. More formally:
Rankorder "explosionrule" definition:Given rankor-
deredchoice set data, the originalrank orderedobser-
vationthatdecisionmakeri prefersthe alternativesin the
order1,2,...,Ji-i.e., that U,1> U,a ... > U--can
be "exploded"(decomposed)intoJi - 1 statisticallyin-
dependentchoice observationsof the form (Uil > Uij,j
= 1,2,...,Ji), (U 2 i, j = 2,3,...,Ji), ..., (Ui,i-l
' U,i). (Notethatthe trivialcase (Ui, - Uj,) is ignored.)
To illustratethe use of the explosionrule, considerthe
automobilechoice processin which a decision makeris
observedto rankorderhis or her choice set in the fol-
lowing manner:Pinto is preferredto Chevettepreferred
to Omnipreferredto Pacer.Invokingthe explosionrule,
one can formthreechoice sets {Pintopreferredto Chev-
ette, Omni,andPacer},{Chevettepreferredto Omniand
Pacer}, and {Omni preferredto Pacer}. Because these
threeexploded (decomposed)rank orderedchoice sets
are statisticallyindependent,they can be thoughtof as
equivalentlybeing obtainedfrom independentdecision
makers.Thus, the explosion process appliedto prefer-
ence rankorderedchoice set data results in additional
choice observationsfor analysisat the estimationstage.
As the accuracyof the parameterestimatesof the sto-
chasticutility model dependson, among other factors,
the numberof independentchoice sets availablefor anal-
ysis, the explosion(decomposition)processexploitsthe
informationcontent in a preferencerank orderingby
generatingmultiplestatisticallyindependentchoice ob-
servationsfromeachdecisionmaker'srankorderedpref-
erences. These additionalchoice observationswill, in
general,lead to "better"estimatesof the parametersof
the stochasticutility model, "better" being interpreted
in the sense of reducedsamplingvariance.The explo-
sion processis also importantin termsof samplingcost.
Obtainingadditionalchoice sets from independentde-
cision makersis generallymore expensive than asking
a smallerset of decision makersto supplythe complete
preferencerankorderingof all elementsin theirchoice
sets (in additionto the chosen alternative).Hence, the
rankorderexplosionprocesscan be viewed in eitherof
two complementaryways: (1) for a given numberof
rankorderedchoice sets, the explosionprocessleads to
incremental independentchoice observationsand, hence,
to decreasedsamplingvarianceof the parameteresti-
mates, or (2) to reduce samplingcosts, rank ordered
choice sets can be obtainedfrom independentdecision
makers(for a relativelysmall incrementalcost overjust
collectingdataon the chosen alternative)and, by use of
the explosionprocess, sufficientindependentunranked
choice sets can be generatedfrom the rank ordered
choice set to yield a desireddegree of precisionin the
resultingparameterestimates.
Two other implicationsof the explosion rule should
be noted.First,the rulestatesonly thatthe Ji - 1 choice
eventsderivedfrom a decisionmaker'spreferencerank
orderingare statisticallyindependentand, thus, can be
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
291
thoughtof as equivalentlybeing obtainedfrom Ji - 1
independentdecision makers.(No statementapparently
is madeaboutthe errortermdistributionassociatedwith
the exploded choice sets.) Because the elicitation of
preferencerank orderedalternativesmay involve the
decision maker in forecastinghypotheticalchoice be-
havior(whichmay not be identicalto actualchoice be-
havior),some "noise" may be introducedinto the data
collectionprocess. This is equivalentto observingerror
terms with larger variancesfor lower ranked alterna-
tives. Sucha phenomenonmightoccurbecausethe con-
sumerdecision makersgive less thoughtto "choices"
associated with lower ranked alternatives.However,
such a premiseis not consistentwith the underlyingas-
sumptionsthatlead to the developmentof the explosion
rule, as the derivationof equation5 is based on the as-
sumptionthatthe errortermsare IID. Consequently,if
one suspects that the lower ranked alternativesare
"noisier" and have greater error variance than the
higherrankedalternatives,one must explicitly test for
such a conditionpriorto use of the explosion rule.
A secondimplicationof note is the Luce and Suppes
RankingChoiceTheoremassumptionthatdecisionmak-
ers ranktheirchoice set alternativesfromtop to bottom.
Thoughthis behavioralmodel of the rankingprocess is
certainlyplausible,choice set alternativesmightbe rank
orderedin otherways. Interestingly,if one assumesthat
the decision makersreversethis procedure(by succes-
sively deletinginferiorchoice set alternativesfromcon-
sideration),the choice probabilitiesgeneratedfrom such
a bottom-to-topprocedurewill be equal to the top-to-
bottom choice probabilitiesonly if the alternativesin
each choice set are equallylikely to be chosen, a rather
restrictivesituation(Luce and Suppes 1965, p. 356-8).
Therefore,it mattersin a theoreticalsense whetherde-
cision makersranktheirchoice sets fromtop to bottom,
becausethe explosionprocessholds only if this is true.
Consequently,if the informationcontentof rankordered
choice sets is to be exploited, the marketingresearcher
musttry to ensurethatthe decisionmakers'behaviorin
rankingchoice set alternativesapproximatesthe top-to-
bottombehavioralmodel. Any violation of the top-to-
bottomrankingmodel will lead to "noisy" choice sets
beinggeneratedby the explosionprocess. (Two obvious
unresolvedresearchquestionsflow fromthis discussion:
"How do consumerdecision makersactuallygo about
the process of rankorderinga given choice set?" and
"How well are alternativerankingparadigmsapproxi-
matedby the top-to-bottombehavioralrankingassump-
tion?" The latter question is of particularnote in the
contextof generatinga rankingof choice set alternatives
from ratingsdata suppliedby consumerdecision mak-
ers.)
EXPLOSION STRATEGIES
The ideal environmentfor gathering rank ordered
choice set data would be a laboratorysetting. The de-
cision makerscould be presentedwith a finite set of al-
ternativesand asked to choose the most preferredele-
292
ment. This most preferredelement would be assigned
rankone and then removedfrom each choice set. Each
decision makerwould then be asked to select the best
alternativefrom amongthe remainingones. These sec-
ond choice alternativeswould be assigned rank two.
Thisprocesswouldbe iterateduntilthe choice sets were
completelyrankedfrom top to bottom.
Complicationsarise, however, in gatheringsuch rank
orderedpreferencedata with a survey researchdesign.
One usually cannotcontrolhow the respondentunder-
takesa rankorderingtask. Some respondentsmight in-
dicateonly the alternativeactuallychosen, whereasoth-
ers might provide rank order informationfor the first
coupleof alternativesbut not for the rest of the elements
in the choice set. Therefore,the questionarises of how
to cope with partial and/or "noisy" preferencerank
orderedchoice sets.
Partiallyrankorderedchoice sets can be treatedin a
straightforward manner.Denote ri as the depthof avail-
able rankorderinformationfor decision makeri on the
alternatives in choice set Ci (where 1 - ri - Ji - 1).
Then, the maximumnumberof explodedchoice obser-
vationsthatcould be made availablefor analysiswould
be equalto I= min(ri,Ji- 1). Thepracticalsignificance
of partiallycompleterankorderinformationis that the
unrankedalternativescannotbe exploded,butthe ranked
alternatives can still be decomposed into unranked
choice events in the usual fashion.
The conceptualproblem associated with using rank
ordereddatawherethe errortermsmightnot be IID and/
or the rankingmighthave been performedin some man-
ner otherthantop down is more complicated.One so-
lutionis to ignore the availablerankorderinformation
and analyzethe choice set data accordingto the basic
stochasticutility model framework.A second, more
constructiveapproachmight be to ask respondentsto
provideratingsof the choice set alternatives.Suchrating
data could then be used to derive choice set rankings.
This approachwould allow respondentsto provideiden-
tical ratingsof alternativesthey feel less sure about in
the preferencerank ordering(as well as allowing the
respondentnot to rate some alternativesbecauseof lack
of relevantinformationor unfamiliarity).A third ap-
proachmightbe to elicit confidencejudgmentsdirectly
fromrespondentson the degreeof confidenceassociated
with the ranking(or rating)of each choice set alterna-
tive. A fourthapproachis to explode only partiallythe
availablerankorderedchoice sets. DefineE as the depth
of the explosion(decomposition)processas specifiedby
the marketingresearcher.Then I(E), the numberof in-
dependentchoice observationsavailableafterexploding
the rankorderedchoice sets to a depthof E, is defined
as
(10) I(E) = E min(E,r,Ji1 - 1).
i=l
If the marketingresearcherdecides to use all of the
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
JOURNALOF MARKETING AUGUST1982
RESEARCH,
availablerankorderinformation(andtreatall of the rank
orderedalternativesas being equallyreliable),the max-
imumnumberof explodedchoice observationswouldbe
I
(11)
i=1
Givena distributionof ri andJi values acrossthe sam-
ple of decision makers, I(E) will be a monotonically
nondecreasingfunctionof E, andI(E) will typicallyin-
crease at a decreasingrate. The definition of I(E) in
equation 10 implies that I(E) - I(E + 1), where the
equalityholds when E is greaterthan or equal to the
maximum of ri or Ji - 1 for all i.
In statisticalterms,the importantstrategicquestionis
how E shouldbe chosen. Largervalues of E will gen-
eratemore independentchoice observations,leading to
reducedsamplingvarianceof the parameterestimatesof
the stochasticutility model, at least in principle.How-
ever, largervalues of E result in explosion of observa-
tions that are fartherdown in the rank orderssupplied
by the sampleof decision makers.Becausethese order-
ings are more likely to be characterizedas randomor
"noisy" (or at least less systematicwith respectto the
rulesused by the decisionmakersto choose the first few
alternativesin the preferencerankordering),the effect
of addingsuch "noisy" orderingswould be to bias the
parameterestimatestowardzero. Thus, the tradeoffin-
volves reducingsamplingvarianceof the parameteres-
timates versus the possibility of less reliability (and
greater"noise") for the incrementalchoice observations
generatedby explosions. Hence, an importantfactor in
the choice of E is the responsivenessof samplingvari-
ance to increasesin the numberof choice sets available
for analysis.
An heuristicapproachto choosingE involves plotting
valuesof a likelihoodratioindex versusdifferentvalues
of E. McFadden(1974) has proposedthe followinglike-
lihoodratioindex as an overallgoodness-of-fitmeasure
(analogousto the multiplecorrelationcoefficient in lin-
ear statisticalmodels).
L(O=
(12)
2 - L( = 0)
L(O= 0)
whereL is the log likelihoodfunction.To the extentthat
the maximumlikelihoodparameterestimatesexplainthe
choiceprocesscompletely,R2will approachone in value
becauseL(0 = 0) will approachzero in value. If 0, the
vectorof MLEparameterestimates,is essentiallyequal
to 0 (implyingrandomchoice among the choice set al-
ternatives), L(O = 0) will approach L(0 = 0) in value,
implyingthat R2 will approachzero in value. Hence,
R2 will vary between zero and one dependingon the
"explanatorypower" of 0. Now, becauseR2 does not
dependon the numberof availablechoice observations,
the calculatedR2 values should remain approximately
constantas E increases.If R2startsto decreasesubstan-
tiallyat some value of E, this would implythat "noisy"
STOCHASTIC
UTILITY
MODEL
exploded choice observationshad been added to the
availablepool of choice sets, and the explosionprocess
should be terminatedjust prior to the "elbow in the
curve."
A secondconsiderationin choosingE is relatedto the
researcher's prior beliefs and knowledge about the
choice process being studied. Such beliefs and knowl-
edge mightlead the researcherto not considervalues of
E beyondsome practicalupperbound.
Anotherfactorto consideris the distributionsof ri and
Ji values which serve to restrictthe possible range of
explosion depths. For example, if the mean depth of
available rank order information is small-say
r = 2.5-I(5) probablywill not be much larger than
I(4). Thus, the researchermight decide not to explode
beyonda depthof 4 because few additionalchoice ob-
servationswill be generated and the extra exploded
choice observationsbeyondE = 4 might be of a priori
questionablequality.
The strengthof these heuristicapproachesis thatthey
arerelativelysimpleto use. However,an obviousques-
tion is whetheran analyticalapproachexists for deter-
mining the extent to which rank orderedchoice sets
shouldbe exploded(decomposed).One formalapproach
involves groupingchoice observationsby depth of ex-
plosion and sequentialhypothesis testing. Define the
first subgroupof choice observationsto consist of the
I(E) choice sets generatedby an explosionto a depthof
E. The secondsubgroupthenconsistsof the incremental
I(E + 1) - I(E) choice sets generatedby exploding to
a depthof E + 1. If I(E + 1) - I(E) is large enough
to allowreasonableparameterestimatesto be developed,
the hypothesis that 0(E) = 0E+ can be tested by a sta-
tisticaltestingproceduresuggestedby WatsonandWes-
tin (1975). This test affords an assessmentof whether
two datasubgroupsshouldbe pooled for estimationpur-
poses. To test the hypothesis that 0(1)= 0(2), the appro-
priatetest statisticis
-2{L( = @P))- [L(O = ()) + L(O = o(2)]}
where O(P)is the MLE of 0 obtained by pooling the data
subgroups, and 0(1)and 0(2) are the MLEs for the separate
datasubgroups.This test statisticwill be asymptotically
distributedchi squarewith N degreesof freedom(Wald
1943). In applicationof the Watsonand Westinpooling
test, the failureto rejectthe null hypothesisimplies that
the two datasubgroupscome fromthe same underlying
choiceprocess,with the sameerrortermstructure.Con-
sequently,these two data subgroupscan be pooled for
the purposesof estimation.Thus, this is an exact test of
whetherthe assumptionsunderlyingthe explosion pro-
cess hold with real data: if the null hypothesis that
0(E)= (E+1)cannot be rejected, the available choice set
dataare consistentwith the requisiteassumptionsof the
constantutilitymodel. If this null hypothesisis rejected,
no informationis providedas to which assumption(or
assumptions)is violatedby the data. This groupingand
sequentialtesting procedurecan be iteratedfor succes-
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
293
sive values of E until either the hypothesis that the
subgroupparametervectors are equal is rejectedor the
quantityI(E + 1) - I(E) yields too few explodedchoice
observationsto providemeaningfulparameterestimates.
The explosionprocess would terminatewhen either of
these conditionsis encountered.
SMALLSAMPLE PROPERTIES OF THE
CONDITIONAL LOGITESTIMATION PROCEDURE
The known propertiesof maximumlikelihood esti-
matorshave been derived only under asymptoticcon-
ditions.To assess the value of generatingadditionalsta-
tistically independentobservationsby exploding rank
orderedchoice sets, one mustinvestigatethe small sam-
ple propertiesof the multinomiallogit model parameter
estimates.Of particularinterestto marketingresearchers
is the situationdescribedby McFadden(1974) as "con-
ditionallogit estimation,"in which thereis one choice
set per decision makerand no replications.As well as
examininghow the precisionof the estimatesis affected
by the number of choice sets available for analysis
(among other factors), we investigatedseveral related
issues such as unbiasednessof the estimatesand com-
putationalcosts.
Design of the Monte'Carlo Experiments
The generalprocedureused to investigatethe small
samplepropertiesof the conditionallogit estimateswas
Monte Carlo experimentationon artificiallygenerated
data. Choice sets with controllablecharacteristicswere
generatedandthe choice processwas simulatedwith the
probabilisticchoice model in equation5. The simulated
rank orderedchoice sets were then exploded and the
conditionallogit estimationprocedurewas appliedto the
resultingunrankedchoice observations.Becausethe true
modelparameters wereknown,we were ableto examine
the extentto which the estimationtechniquerecaptured
the trueparametervector-i.e., the extent to which O
approximates0.
The experimentalfactors hypothesizedto affect the
ability of the conditionallogit estimationprocedureto
recapturethe true parametervalues included (1) the
numberof parametersto be estimated(N), (2) the num-
ber of choice sets availablefor analysisafterexplosion
(NOBS),(3) the averagenumberof alternativesin the
choice sets afterexplosion (SIZE), (4) the relative size
of the deterministicportionof the choice model in com-
parisonwith the error componentas discussed in the
Appendix(SCALE),and (5) the collinearityamong the
variablesin the model (COLL).SIZEandNOBS are in-
verselyrelatedbecauseboth dependon the depthof the
explosionchosen by the analyst.1
'Becauseeachsuccessiveexplosiongeneratesadditionalchoice sets
with one less alternative(becausethe most preferredalternativeis
removedfromconsiderationin formingan explodedchoice set), SIZE
decreasesas E increases.Furthermore,as NOBS increaseswith in-
creasesin E, the indirecteffect is thatSIZEandNOBSare inversely
294
The scale of the model parameters is defined as
N 1/2
(13) SCALE =E o2
n=l
Because the size of the error component was known and
the variance of the Z vectors could be determined, we
were able to select SCALE values to represent different
levels of "explained variance" ranging from about 10%
(SCALE = 0.45) to about 85% (SCALE = 3.50).2
The collinearity among the individual components of
Z is capturedby an index value designed to measure the
overall average correlation among the variables describ-
ing the choice set alternatives. The collinearity index
used in this study is
N
E2 --N
(14) COLL=
N(N - 1)
where X, is the nt eigenvalue associated with the gen-
erated data (where the usual convention of ordering the
eigenvalues from largest to smallest is followed). COLL
is bounded between zero and one. If the variables are
orthogonal (i.e., completely uncorrelated), Xh = 2 =
...= = 1 and COLL = 0; if the data are completely
correlated and have rank 1, Xh= N and the other eigen-
values will be equal to zero and COLL = 1.
The experimental design was factorial with one rep-
lication per cell. The factorial design set four factors at
each of the following levels.
N = 2, 4, 7, and 10,
I = 40, 100, 200, and 400,
E = 1, 2, 3, 5, and 10,
SCALE= 0.45, 0.85, 1.375, and 3.50.
Hence, a total of 320 (= 4 x 4 x 5 x 4) experiments
were conducted. Within this Monte Carlo study, NOBS
is defined as
I
(15) NOBS = E min(E,Ji - 1).
i=I
Within each cell in the experimental design, SIZE and
related.This follows becauseboth SIZEandNOBS are functionsof
E. Note thatSIZEandNOBS are only correlatedwithin the context
of an explosion,becausewhenE increasesso does NOBS,but at the
expenseof slightlydecreasingthe averagechoice set size (SIZE).Ex
ante (beforeexplosion)measuresof SIZEandNOBSareuncorrelated
withinthe factorialexperimentaldesign used in this study.
2Thesevalues of SCALEwere chosen by noting that if the com-
ponentsof Z are standardizedto have unit variance,it follows that
Var(U)= SCALE2+ Var(e).BecauseVar e = 1.645 for the double
exponentialdistribution,the usual partitioningof total variationinto
explainedvariationand unexplainedvariationimpliesthat:
"proportion of _ SCALE2
explainedvariation" SCALE2+ 1.645
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
JOURNALOF MARKETING AUGUST1982
RESEARCH,
COLL were chosen stochastically, but in such a way as
to reflect the usual kinds of conditions an empirical mar-
keting researcher might encounter in real choice behav-
ior process data. With regard to choice set size prior to
explosion, each choice set was drawn from a normal
distribution (whose mean was determined by a draw
from a uniform distribution with a range of 2 to 10 and
a standarddeviation drawn from a uniform distribution
with a range of 1 to 2). The actual size of any choice
set (prior to explosion) was truncated so that 2 - Ji - 10
for each choice set i. (The lower limit on choice set size
is the absolute minimum number of alternatives that can
constitute a choice set; the upper limit of 10 was chosen
to reflect the empirical reality of bounded choice set
sizes.) Collinearity was induced into the variables char-
acterizing the choice set alternatives by a transformation
procedure described by Chapman (1981).
The artificial choice set data in each cell of the ex-
perimental design were generated in the following man-
ner. A total of I choice sets were generated, each choice
set containing Ji alternatives (where the Ji values were
determined as described above). Each of the N attributes
of each alternative were generated by drawing indepen-
dently distributed normal random variables with mean
zero and variance one. Collinearity was induced into
these data. The transformed collinear data were stan-
dardized so that each attribute had mean zero and vari-
ance one. COLL was then calculated.
The next series of steps involved simulating the rank
order choice process. A true parameter vector was gen-
erated by drawing N independent values from a uniform
distribution with a range of zero to one. Positive and
negative signs were attached to each of the parameters
(the signs being determined with a probability of 0.5)
and the resulting true parameter values were rescaled so
that relation 13 was satisfied. Next, the probabilistic
choice model in equation 5 was used to assign the choice
probabilities to each alternative. The "chosen" alter-
native was assigned rank one.3 The choice probabilities
of the "nonchosen" alternatives were then rescaled to
sum to 1.0, and the "choice" process was repeated to
determine the rank two alternative. This procedure was
iterated to rank all of the alternatives in each simulated
choice set. Finally, choice sets were exploded to a depth
of E, resulting in NOBS choice sets. These exploded
choice sets were input to a conditional logit estimation
program (Manski 1974) to obtain the parameter esti-
mates.
Several additional points should be noted. First, this
type of experimental design involves an implicit as-
sumption that each simulated decision maker employs
the same underlying choice model. Thus, there is no
heterogeneity problem to confound the estimation pro-
cess. Second, reliability is not an issue in these experi-
3Byuse of a randomnumbergenerator,an alternativewas "cho-
sen" with probabilityequal to the choice probabilityassignedto the
alternative.
 UTILITY
STOCHASTIC MODEL
ments because the same underlying probabilistic choice
process was employed to rank all of the choice set al-
ternatives. Third, only the non-extreme values of
COLL - 0.75 were considered, reflecting the usual pres-
ence of modest degrees of interrelationships among the
attributesof the choice set alternatives. Because of these
points, the results reported in the rest of this section
probably represent an upper bound on the degree to
which the conditional logit model estimates recapturethe
"true" underlying choice model. These results are "bet-
ter" than might be obtained with real choice data be-
cause potential problem sources such as respondent het-
erogeneity, measurement and specification errors, and
unreliable rank order information are not present.
In each cell of the experimental design, several per-
formance measures were calculated. These measures
were designed to represent the extent to which the con-
ditional logit estimation procedure recaptured the true
parameter vector values associated with the underlying
multinomial logit model. The performance measures in-
cluded the following.
1. RMSE(root mean squarederror),where
I N \1/2
2
RMSE = I E(O - On)
2. MAE(meanabsoluteerror),where
1N
MAE = -
In - Onl
Nn=1
3. BIAS(bias), where
N - 0.6863 log SCALE+ 1.6146 log N
1
BIAS =-- (On- n)
n=l
4. TIME(CPU executiontime in seconds, includingonly
the time used for estimationand excludingdatagenera-
tion time requirements)
In performancemeasures 1-3, Onis the known true value
of the nh parameter in the multinomial logit model and
n,is the maximum likelihood estimate obtained from the
conditional logit estimation procedure.
Results of the Monte Carlo Experiments
The results of these experiments were analyzed by re-
gressing variants of the performance measures on theo-
retically plausible causal experimental factors.
Sampling variance. One component of the accuracy
of the conditional logit estimation procedure is sampling
variance. To reflect relative sampling variation (and to
avoid the problem associated with, ceteris paribus, large
parameterestimates having associated large standarder-
rors), the performance measure chosen was the relative
root squared variation (RELRSV), defined as
RMSE
(16) RELRSV =
|9|
where:
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
295
(17)
161 1 NI
= Nn=l le .
Thus, RELRSV is a measure of relative variation anal-
ogous to a coefficient of variation.
Theoretically, RELRSV = f(NOBS, SIZE, COLL,
SCALE, N) with prior beliefs that, ceteris paribus,
RELRSVshould decrease with increases in NOBS, SIZE,
and SCALE and decrease with decreases in COLL and
N. The theoretical arguments associated with the signs
on NOBS, SIZE, and N all reflect "degrees of freedom"
concerns. The expected positive sign on COLL is due
to collinearity being posited to lead to less precise esti-
mates of the model's parameters. SCALE reflects the
relative magnitude of the systematic component in the
model and, thus, should be negatively related to
RELRSV-a priori, a greater magnitude for the "ex-
plained variance" component of the model should be
associated with an improvement in the precision of the
estimates (a decrease in RELRSV). An interactive func-
tional relationship was postulated and the model was
estimated after appropriate logarithmic transformation.
The OLS results were
(18)
log RELRSV= 0.1812 - 0.5753 logNOBS
(0.4270) (0.0487)
1
- 0.2164 log(SIZE-1) + 0.1849
(0.1260) (0.1138) 1-COLL
(0.0670) (0.0919)
R2 = 0.65 No. of obs. = 320
where the standarderrors of the coefficient estimates are
in parentheses and log is the naturallogarithm. The form
of the regression model in equation 18 reflects the the-
oretical relationship between COLL and RELRSV. As
COLLapproaches one in value, RELRSVincreases with-
out bound regardless of the values of the other factors.
Also, note that SIZE - 1 is used to describe the choice
set sizes because the conditional logit estimation pro-
cedure is based on the number of within-choice-set com-
parisons available for analysis (i.e., the chosen alter-
native versus the other J - 1 inferior alternatives).
Given the interactive functional form, the coefficients
on NOBS, SIZE - 1, SCALE, and N can be interpreted
as elasticities.
The estimated coefficients in equation 18 all have the
theoretically posited signs. The main conclusion to be
drawn from these results is that the explosion process
leads to important sampling variance efficiencies. The
coefficient estimate of -0.5753 for log NOBS implies
that RELRSV decreases approximately with the square
root of the number of choice sets available for analysis
after explosion, ceteris paribus. A comparison of the
coefficients on log NOBS and log (SIZE - 1) is useful
296
for assessing the marginal tradeoffs between the number
of available choice observations after explosion and the
average choice set size after explosion. For example, a
researcherwho has 100 choice sets each with five ranked
alternativescould use E = 1 (and NOBS = 100 and SIZE
= 5) or the data could be exploded to E = 4 (where
NOBS = 400 and SIZE correspondingly declines to 3.5).
Using equation 18, we can easily show that RELRSVfor
the former strategy (E = 1) is about twice that for the
latter strategy (E = 4). Further, to obtain the same level
of RELRSVas that when E = 4 by retaining E = 1 and
increasing the average choice set sizes, we would have
to increase SIZE from five to about 100. (In this ex-
perimental design, there was much more variation in
NOBS than in SIZE; NOBS had a range of 40 to about
2600 whereas SIZE ranged from about 2.6 to 9.2. If
SIZE had been varied much more, say to 100, it might
have exerted a stronger influence on RELRSV. Still, as
most empirical applications of the multinomial logit
model would probably involve mean choice set sizes of
fewer than 10 alternatives, this experimental framework
seems realistic. Note also that, in many applications, the
size of the choice sets is determined by the consumer
decision makers, and is not at the control of the mar-
keting researcher. The number of sample decision mak-
ers, however, is often at the choice modeler's control.)
Collinearity among the attributes seems to have only
a negligible impact on sampling variance. This result is
very encouraging because real choice data should be ex-
pected to exhibit patterns of collinearity.
As was expected, the precision of the conditional logit
estimates is directly related to SCALE, a surrogate for
the degree of systematic behavior exhibited by the sam-
ple of decision makers. The coefficient estimate on log
SCALE implies that a doubling of SCALE would lead to
about a 68% decrease in RELRSV.
A doubling of N, the number of model parameters,
leads to about a 160% increase in RELRSV, ceteris par-
ibus. Because theoretical considerations guide the de-
terminationof the number of parametersin the stochastic
utility model, these results are useful to a choice modeler
in assessing the appropriate sample size to obtain a de-
sired level of precision in the resulting parameter esti-
mates. For example, to obtain equivalent precision for
parameterestimates in a four-parametermodel estimated
with 100 choice sets, the choice modeler would require
about 260 choice sets for a more theoretically complex
model with eight parameters.
The model reported in equation 18 was also estimated
with relative absolute variation as the dependent vari-
able. The OLS results were virtually identical to those
reported in equation 18. Apparently, the relationship
between the external factors (i.e., NOBS, SIZE, and so
on) and the sampling variance of the conditional logit
estimates does not depend on whether accuracy is mea-
sured in terms of a quadratic or a linear loss function.
Although the small sample properties of the condi-
tional logit estimates are described clearly by the regres-
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
JOURNALOF MARKETING AUGUST1982
RESEARCH,
sion results reported in equation 18, the explanatory
power of the model is rather low. The reason is partly
the experimental design used in this Monte Carlo study.
Because only a single replication was performed in each
of the 320 cells in the factorial design, we could not
account for the within-cell variation of the performance
measures (the dependent variables). The source of this
within-cell variation is the probabilistic choice process
itself: the "chosen" alternative in each exploded choice
observation is determined by a probabilistic model and
a single replication of the choice process reflects but a
single point estimate of the parameter values. To mea-
sure the approximate magnitude of this replication error,
five replications of the simulated choice process were
conducted in each of the 16 cells characterized by the
factorial design with N = 4 and 7, I = 100 and 250,
E = 1 and 4, and SCALE = 0.50 and 2.00. The average
standarddeviation of the log RELRSVvalues over these
16 cells is 0.499. Because the standard error of the es-
timate in equation 18 is 0.894, the replication error rep-
resents about 56% of the total unexplained variance in
the regression model in equation 18. If means of repli-
cated choice processes had been used as the dependent
variable in equation 18, we could reasonably expect that
the total unexplained variance would have been reduced
by about 56%. The corresponding R2 value would have
been increased to about 0.85. Such resulting unex-
plained variance (about 15%) is much more in line with
a priori expectations.
Biasedness. Because the conditional logit estimates
are obtained by means of maximum likelihood estima-
tion techniques, the property of unbiasedness is guar-
anteed only for large samples (asymptotically). Conse-
quently, it would be useful to know the degree and
direction of any small sample bias.
For the 320 cells in the experimental design, the mean
BIAS is -0.025 with an associated standard deviation
of 0.311. A two-tailed test of the hypothesis that the
mean BIAS of the conditional logit estimates is zero
leads to the conclusion that the unbiasedness hypothesis
cannot be rejected at the conventional 5% level of sig-
nificance (p = 0.15).
To test for the possible presence of conditional
bias-bias depending on the external factors-BIAS was
regressed against several theoretically plausible combi-
nations of the external factors (NOBS, SIZE, COLL,
SCALE, and N). None of the resulting F-statistics for the
regression models tested yielded significant values.
Though it is not possible to prove unbiasedness via
simulation, these Monte Carlo results may be viewed as
support for an hypothesis of unbiasedness. The possi-
bility that subsets of the parameter estimates could be
biased is not precluded. That is, offsetting biases in in-
dividual parameter estimates could be present.
Computational considerations. The final performance
measure of interest was computational cost. Because the
explosion process yields additional choice sets for anal-
ysis, the decision of how much to explode the available
STOCHASTIC
UTILITY
MODEL 297

rankorderedchoice set datamightbe influencedby the Table 1

extra associatedcomputationalcosts. The performance CALCULATIONSOF RELRSVAND TIME
(CPUEXECUTION
measure for computationalcosts is execution time, TIME,IN SECONDS)IN THEEXAMPLEILLUSTRATING
TIME,which refers to the CPU executiontime for the PRECISION
AND COSTTRADEOFFS
Newton-Raphsonalgorithm(in seconds) used in Man-
ski's (1974) program. N=5 N=10
Theoretically,TIME= f(NOBS,SIZE,N) with TIME E NOBS SIZE RELRSV TIME RELRSV TIME
increasingwith increasesin each of the externalfactors. 1 100 8 0.96 7.7 2.93 23.4
Momentmatricesmustbe assembledat each iterationof 2 200 7.5 0.65 13.6 2.00 35.4
the algorithmand the numberof elementsprocessedat 3 300 7 0.53 18.6 1.61 45.3
each iterationequalsNOBS(SIZE- 1)N. Also, the mo- 4 400 6.5 0.45 22.6 1.39 53.3
ment matrixmust be invertedat each iterationand the
timefor suchan operationdependsapproximatelyon N2.
RegressingTIMEon these factorsyielded the following
OLS results. terizingthe choice set alternativesare expectedto have
some correlationamong them (COLL = 0.25) and a
(19) TIME= -2.2548 + 0.001991 NOBS(SIZE-1)N
SCALEvalue in the area of about 1.00 is anticipated.
(0.9866) (0.000040)
Table 1 lists the predictedvalues of RELRSVand TIME
+ 0.1175 N2 undervariousexplosiondepthsfor five- and 10-variable
(0.0185) multinomiallogit models. The values in Table 1 were
R2 = 0.91 No. of obs. = 320
calculatedby using the resultsin equations18 and 19.
Severalfeaturesof these data may be noted. First, as
wherethe standarderrorsof the parameterestimatesare NOBSincreasesbecauseof explosions, SIZEdecreases
in parentheses.These resultsimply thatTIMEincreases becausechoice sets of smallersize are being added to
linearlywith increasesin NOBSandSIZE,andincreases the availablepool of choice observations.Second, the
approximatelyquadraticallywith increasesin N. These decreasein RELRSVis approximately proportionalto the
resultscan also be used as a guide to the magnitudeof squareroot of the increasein NOBS. Hence, to halve
the computationalcosts associatedwith estimatingthe RELRSV,the choice modelerrequiresabout four times
parametersof the multinomiallogit model. Naturally, as manychoice observations.Third,even for the largest
the absoluteamountof executiontime dependson the model situationin this example (in terms of computa-
specific algorithmused and the particularcomputerin- tional cost), when E = 5 and N = 10, computational
stallationon which the analysisis performed.(All con- costs arenot prohibitive,amountingto slightlyless than
ditionallogit estimationsin the studywere performedon a minutein CPU time.
an IBM/360.) However,relativetimes for variouscom- Fora researcherdecidingon the size of a choice sam-
binationsof NOBS, SIZE, and N should be approxi- ple studyand on subsequentexplosion strategy,the re-
matelyequivalentacross differentalgorithmsand com- sults in Table 1 offer some insightinto the precisionof
puters. the multinomiallogit model parameterestimates. Sup-
For many, if not most, marketingresearchapplica- pose a marketingresearcherwouldlike each of the coef-
tions of the multinomiallogit model, computational ficient estimatesin a five-parametermultinomiallogit
costs apparentlywill not be an inhibitingforce. Com- model to be statisticallysignificant(i.e., have an esti-
putationalcosts tend to be in the orderof secondsrather matedstandarderrorno largerthan about one-half the
thanminutes.For example, with 250 choice sets of av- size of the coefficientestimate).Given the conditionsin
eragesize fourwith fourparametersto be estimated,the this illustrativeexample, explosion to a depth of four
applicationof the regressionresultsreportedin equation wouldbe required.Of course, the choice modelermight
19 yields a predictedexecutiontime value of about5.6 choose to gathermorethan 100 choice sets to startwith
seconds. An eight-parameter model of the same choice so that a lesser explosion depth would be possible. By
set configurationwould requireabout 17.2 seconds of using the resultsin equation18, the thoughtfulmarket-
executiontime. ing researchercan anticipatethe impactof choice study
Precisionand Cost Tradeoffs sample sizes and possible explosion strategieson the
precisionof the parameterestimates before collecting
To illustratethe natureof the tradeoffsbetween re- data.
duced samplingvarianceand additionalcomputational
costs associatedwith analyzing exploded choice sets, AN ILLUSTRATIVE EMPIRICAL APPLICATION OF
and to indicatehow a choice modelermight use the re- THEEXPLOSION PROCEDURE
sults reportedin equations18 and 19, considerthe fol- To illustratethe applicationof the explosionprocedure
lowing example. Supposea researcheris workingwith for exploitingthe informationcontent of rank ordered
completelyrankedchoice sets of size eight for 100 con- choice set data, we use a college choice behaviormod-
sumer decision makers. Further,the variablescharac- eling study. Chapman(1977) has given the full details

This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
298 JOURNALOF MARKETING AUGUST1982
RESEARCH,

of this study and a summary of the empirical results Table 2

(1979). COEFFICIENTS
OF THECOLLEGE CHOICEMODEL
The data base for the study consisted of admitted FOR"LOW"INCOMEENGINEERING AND
freshmen applicants to Carnegie-Mellon University, a SCIENCESTUDENTS (ASYMPTOTIC STANDARD
private independent coeducational university in Pitts- ERRORSIN PARENTHESES) FOREXPLOSIONDEPTHS
burgh, Pennsylvania. A survey research questionnaire OF 1, 2, AND 3
to admitted applicants was used to obtain the required
preference rank order and choice set composition infor-
Explosionstrategies
mation, as well as information on financial aid offered
E= 1 E=2 E=3
by schools to each high school student. These data were
combined with other information contained in Carnegie- FS, (quality/ 0.69555 0.6894 0.6100
Mellon's applicant information system (derived from affluence) (0.13553) (0.1123) (0.1037)
FS2(size/graduate 0.461[1 0.4104 0.3221
each applicant's application forms) and other publicly (0.0974)
orientation) (0.134t8) (0.1057)
available information (college alternative characteristics FS3 (technical 0.21556 0.2351 0.2428
and costs) to yield the required study data. Self-reported orientation) (0.12337) (0.0984) (0.0914)
preference rank order and financial aid information was FS4(ruralness) -0.069 }0 0.0281 0.0359
(0.12337) (0.0952) (0.0902)
subjected to consistency and reliability checks. Across FS5(fine arts 0.02C)5 -0.1208 -0.1503
the various academic areas in this data base, average orientation) (0.144t6) (0.1137) (0.1057)
choice set sizes ranged from 2.9 to 3.5 college alter- FS6(liberalness) 0.14223 0.1044 0.0811
natives, and less than 10% of the students in the sample (0.134t6) (0.1105) (0.1020)
had more than five college alternatives in their choice GRANTS 0.18319 0.2115 0.3608
sets. 1000 (0.19351) (0.1620) (0.1526)
The variables in the college choice model were of DRAIN -1.652 29 -2.0704 -1.3354
three categories: college characteristics variables; price, (0.87237) (0.7459) (0.6707)
INCOME
cost, and financial aid variables; and a miscellaneous
GRANTS 0.35228 0.3701 0.0442
group. To describe the attributes and characteristics of
colleges, 46 different variables were obtained from pub- TOTALAID (0.45C)9) (0.3750) (0.3501)
licly available sources, such as college guidebooks. Fac- (MILES/10)/2 -0.343 39 -0.1397 -0.0822
tor analysis procedures were used to reduce these 46 (0.185)5) (0.1403) (0.1271)
variables to a set of composite indices that would serve COMMUTE -0.09, 78 0.0524 0.1474
as suitable proxies for the college attributes. The six fac- (0.35337) (0.2763) (0.2546)
{SATi - SAT,2 -0.38( 53 -0.4081 -0.3188
tors extracted accounted for about 58% of the variance
in the original 46 raw variables. The results of a varimax 100 J (0.17110) (0.1425) (0.1331)
rotation led to the six factors (FS1, ..., FS6) being in- NOBS(after 303 452 505
terpretedas quality/affluence, size/graduate orientation, explosion)
L(O = 0) -309.18 -448.23 496.53
technical orientation, ruralness, fine arts orientation, and L(O= 0) -240.80 -353.16 399.60
liberalness, respectively. In the financial area, the fol- k2 0.221 0.212 0.195
lowing variables were used in the choice model:
DRAIN GRANTS
GRANTS, , and model with explosion depths of 1, 2, and 3 are reported
INCOME' TOTALAID
in Table 2. These results are for "low" income engi-
(GRANTSis total scholarship aid, DRAIN is total out-of- neering and science students, where "low" is defined
pocket costs, INCOME is parental income, and TOTAL- to be the bottom half of the applicants in terms of pa-
AID is total financial aid.) Collectively, these variables rental income (for those students who had applied for
account for amount and kind of financial aid (mix of financial aid and, hence, had supplied parental income
financial aid), and for costs in relation to income. Within information).
the miscellaneous group of choice model variables are The main result to be noted in Table 2 is that the stan-
MILES (distance from college campus to place of resi- dard errors of the parameter estimates decrease as more
dence), COMMUTE (a dummy variable which equals choice sets become available for analysis due to the ex-
one if the student could commute to a college campus plosion process for exploiting preference rank ordered
and equals zero otherwise), and (SATi - SATj)2(where choice set data. This result illustrates the main value of
SAT refers to mean scholastic aptitude test score, the the explosion procedure: explosion of rank ordered
subscript i refers to the student, and the subscriptj refers choice sets leads to more choice observations being
to a college), which attempts to measure the impact of available for analysis which leads to improved precision
"quality zoning"-the desire to attend a college that is (reduced sampling variance) in the stochastic utility
neither too much above nor below the student in average model parameterestimates.
student quality. The key strategic issue faced by the choice modeler
Statistical results of estimating this college choice is the choice of an explosion strategy. In the college

This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
 MODEL
UTILITY
STOCHASTIC
choice resultsreportedin Table 2, the heuristicof ex-
amining R2 as E increases would probably lead the
choice modelerto concludethat an explosion depth of
2 is reasonablysafe (as R2only declinesslightlyin shift-
ing fromE = 1 to E = 2, namelyfrom0.221 to 0.212),
buttheremightbe some uncertaintyin resolvingwhether
E = 3 should be preferred to E = 2.4
In using the groupingand sequentialhypothesistest-
ing procedurefor statisticallytesting for an "optimal"
explosiondepth, it is useful to begin by notingthat the
criticalchi squarevalue for a 5% level of significance
is 21.03, whereasthe correspondingvaluefor a 1%level
of significanceis 27.69 (for 12 degreesof freedom,the
numberof parametersin this choice model). Recall also
that the null hypothesisbeing tested is whether0(E =
0(E+1). In assessing whether a change from E = 1 to E
= 2 is appropriate,the relevantlog likelihoodvalues are
L(O = 6(+2) = -353.16, L(O = ()) = -240.80, and
L(O = (2) = -105.67. (In notational form, 6(1+20is the
MLE for the pooled explodedrankorderedchoice sets
obtained by exploding to a depth of 2, @(l)is the MLE
for the choice sets basedonly on the firstmost preferred
alternative, and 0(2) is the MLE for the choice sets based
only on the second preferredalternative.)The relevant
chi square test statistic value is therefore equal to
-2(-353.16 - {[-240.80] + [-105.67]}) = 13.38.
Comparingthis calculatedtest statistic value with the
relevantcriticalvalues leads to the conclusionthat the
nullhypothesisshouldnot be rejectedon the basis of the
availablesample evidence. Therefore,pooling the ob-
servationsis feasible, and an explosiondepthof at least
2 is warranted.In iteratingthis test to determinewhether
an explosiondepthof 3 is appropriate,the relevantlog
likelihood values are L( = (1+2+3)) = -399.60,
L(O = o+2)) = -353.16, and L(O = (3) = -32.50.
The relevantchi squaretest statisticvalue is equal to
27.88, which is above the 1% significancecritical chi
squarevalue. Thus, one could conclude that the null
hypothesisshould be rejectedin this situation.As the
availablesample evidence seems inconsistentwith the
pooling hypothesis,an explosion depth of 3 seems in-
appropriate.The grouping and sequential hypothesis
testingprocedurefor choosing an "optimal" explosion
strategyleads to the conclusionthat an explosiondepth
of 2 is most appropriatewith these choice set data.
4R2does not depend on the number of available choice set obser-
vations, so the value of R2 should remain approximately constant as
E increases, if the extra choice sets generated by the explosion process
are characterized by the same 0 vector. If "noisy" choice sets are
added by the explosion process (either due to unreliable rank order
information or because the lower ranked choice sets are characterized
by some 0 vector other than that which described the higher ranked
choice sets), one may observe that the value of R2 will decline with
increases in E. Sampling variation may explain successive values of
R2 that are not identical, so an appropriate statistical test is required
to assess whether the explosion is feasible in the context of any spe-
cific empirical application.
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
299
CONCLUDING
REMARKS
Withinthe context of the stochasticutility model of
choicebehavior,the informationcontentof rankordered
choice sets can be exploitedto estimateefficiently the
parametersof the model. The availabilityof rank or-
deredchoice set data leads to an explosion (decompo-
sition)procedurefor exploitingsuch extra information.
This explosionprocessinvolves the decompositionof a
preferencerankorderedchoice set into a series of un-
rankedandstatisticallyindependentchoice observations.
Becausethe ability to estimatethe model's parameters
dependson the numberof choice sets availablefor anal-
ysis, the explosionprocessrepresentsan intuitivelysim-
ple mechanismfor exploiting rank orderedchoice set
data.
The choice modelermustresolvethe key issue of how
much to explode the availablerankorderedchoice set
data. Several considerationsbear on this issue. First,
because the computationalcosts are ratherminor, the
differencein computationaltime from exploding to a
depthof E + 1 ratherthan a depth of E can probably
be ignored. Second, the problemof reliabilityof the
preferencerank order informationobtainedthrougha
surveyresearchdesign may not be a majorproblem,as
the precisionof the conditionallogit parameterestimates
is relatedto the squareroot of the numberof available
choice set observationsafter explosion. Thus, because
each successive explosion from E to E + 1 results in
incrementallysmallerimprovementsin the precisionof
the parameterestimates,explodingbeyonda depthof 3
or 4 will usually not increaseaccuracygreatly. This is
true partlybecause the precision of the estimates de-
creases by the squareroot of the numberof exploded
choice sets and partlybecause not all decision makers
in mostchoice situationswill have choice set sizes large
enoughto allow for furtherexplosion. Also, as larger
explosionsare more likely to increasethe possibilityof
including"noisy" choice sets, an explosion depth of
morethanabout3 would not normallybe appropriatein
a surveyresearchapplication.
In summary,an explosionstrategywill dependon the
choicemodeler'sknowledgeof the choice processbeing
studiedand the likely reliabilityof the complete pref-
erencerankorderinformation,the use of the grouping
andsequentialhypothesistestingprocedureandthe heu-
ristic tools we describe, and the Mont6 Carlo results
which suggest that large explosions (say, E largerthan
3) are normallynot required.
These resultsalso have implicationsfor surveyques-
tionnairedesignsto elicit rankorderedchoice set infor-
mationfrom decision makers.Attemptingto guarantee
completelyrankorderedchoice set data probablycosts
morethanthe dataare worth. A betterresearchstrategy
would be to ask for the first few rank orderedalterna-
tives (say, up to three), and then ask for all otheralter-
nativesconsideredby the decision makerin the choice
behaviortask. As the statisticalworthof more than the
300
first couple of rank ordered choice set alternatives is
marginal at best (because of reliability and "noisy" data
problems), survey questionnaire respondents need not be
confused and antagonized by having to supply more de-
tailed information than can be used profitably in the sub-
sequent statistical analysis.
APPENDIX
SCALE CONSIDERATIONSAND THE STOCHASTIC
UTILITYMODEL
The rank order of a set of alternatives in a choice set
will be invariant under monotone transformationsof the
form W = axU+ {3, for ot > 0. Hence, the most general
form of the stochastic utility model would be
(Al.1) U,i = a(O Zi + Ei) + 3.
The model in equations 5 and 6 follows from the nor-
malizations ao = 1 and P = 0. Because we are concerned
only with differences in the utilities of various choice set
alternatives, P will be identical for each alternative.
Hence, no generality is lost by assuming that P = 0.
Setting a = 1 causes no difficulties with the ca0 Z.. term
because ac is just a scalar, and in essence just indicates
the size (or scale) of the 0 values. We estimate the vector
a 0, the relative population taste parameters. To obtain
absolute 0 values, an estimate of a somehow determined
independently from the estimate of 0 would be required.
We only need the relative 0 values to draw inferences
about utilities, which are only relative quantities them-
selves. Setting a = 1 in the term acxEdoes, however,
have important ramifications for the variances of the
disturbance terms.
The most general form of the double exponential dis-
tribution is
(A1.2) Prob(E< t) = exp -exp( -- 1)
where m1and Tq2are location and scale parameters, re-
spectively (Johnson and Kotz 1970, p. 276-9). If m1
= 0 and q2 = 1, then equation A1.2 is identical to equa-
tion 4. Assuming 1 = 0 implies that P = 0 in A1.1;
similarly, setting 12 = 1 implies that ca = 1 in A1.1.
Because the variance of the double exponential distri-
1 larger, the systematic component of the stochastic utility
bution is Var(E) = - 72 r12,setting ql2 = 1 implies that
6
Var(e) = ,r2 = 1.645. Hence, fixing ax in the general
6 model, with the probabilities of choosing the alternatives
model in A1.1 implies that the variance of the distur-
bances is also fixed.
The preceding discussion reveals an interesting facet
of the stochastic utility model with double exponentially
distributed error terms-the variance of the disturbance
terms is a fixed number, not dependent on the estimated
0 values. In contrast, in "regression-like" models the
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
All use subject to JSTOR Terms and Conditions
JOURNALOF MARKETING AUGUST1982
RESEARCH,
estimated vector of parameters affects the error terms
and, hence, their variance.
Fixing the variance of the disturbance terms implies
that the proportion of the variance of U represented by
the systematic component of the stochastic utility model
will depend directly on the "size" of the 0 values. In
this way, "large" 0 values will be associated with high
explanatory power of the stochastic utility model be-
cause the variance of the systematic component will be
large in relation to the error component. The following
theorem clarifies these relationships.
Scale theorem. In the stochastic utility model Uu =
OZy + E., where the error terms follow the double ex-
ponentiai distribution in equation 4, let On = coO,where
N=i 02 = 1. This scaling implies that o2 = N=i 0n. The
scaling factor o may be interpreted as the "size" of the
0 values. It follows that
(a) as w --> o, Pio -> 1 for j? alternativesuch that OZio=
max(OZi), and Pij- 0 for all otheralternatives
1
(b) as o -> 0, Pi - forj = 1,2,...,J
i
Proof. One version of the probabilistic choice model in
equations 5 and 6, which can be derived with some sim-
ple manipulations of equation 5, is
Pij*= J,
1+ E exp {0(Z - Zi.)}
j=!
with j* being some alternative in the choice set.
Part(a): For alternativej?, O(Z,i- Zio) < 0 because alter-
nativej? has been defined as the most preferredal-
ternative.Hence, o -> ooimplies that exp{wO(Z,-
-- -- 1.
Zi)} 0. Therefore, o -> ooimplies thatPij
For Pj* such thatj* jo, O(Zio- Z.*) > 0, and o
- ooimpliesthat -
exp{co(Zijo Zi*)}-- ooso thatPi*
--> 0.
Part(b): as -->0, exp{(x(Zi - Z.*)}- 1 forj* = 1,2,...,Ji.
Therefore,co -> 0 implies that Pj*--> , for j* =
1,2,...,Ji.
In summary, as the estimated parameters become
model dominates the stochastic component and the re-
sulting utilities of the alternatives show greater disper-
sion. This follows through to the probabilistic choice
approaching one for the alternative with greatest utility
and approaching zero for all other alternatives. In con-
trast, "small" 0 values would be associated with choice
probabilities approximately equal to 1/Ji for each alter-
native. The latter empirical result implies that the choice
process is essentially random with each alternative in a
choice set having an equivalent chance of being chosen.
 STOCHASTIC
UTILITY
MODEL
REFERENCES
Boskin, MichaelJ. (1974), "A ConditionalLogit Model of
OccupationalChoice," Journal of Political Economy, 82
(April), 389-98.
Chapman,RandallG. (1977), BuyerBehaviorin Higher Ed-
ucation:An Analysis of College Choice Decision Making
Behavior,unpublished doctoraldissertation,GraduateSchool
of IndustrialAdministration,Carnegie-MellonUniversity.
(1979), "PricingPolicy and the College Choice Pro-
cess," Researchin HigherEducation,10(1), 37-57.
(1980), "Retail TradeArea Analysis: Analytics and
Statistics,"in Proceedings:MarketMeasurementandAnal-
ysis, RobertA. Leone, ed. Providence,RI: TIMS College
on Marketingand The Instituteof ManagementSciences,
40-9.
(1981), "A Note on GeneratingCorrelatedData,"
GraduateSchool of Business, Universityof Chicago.
Domencich,T. and Daniel McFadden(1975), Urban Travel
Demand:A BehavioralAnalysis. Amsterdam:North-Hol-
landPublishingCompany.
Gensch, Dennis H. and Wilfred W. Recker (1979), "The
Multinomial,Multiattribute Logit Choice Model," Journal
of MarketingResearch, 16 (February),124-32.
Johnson,NormanL. and Samuel Kotz (1970), Continuous
UnivariateDistributions--1.New York:HoughtonMifflin
Company.
Joskow, Paul L. and FredericS. Mishkin(1977), "Electric
Utility Fuel Choice Behaviorin the United States," Inter-
nationalEconomicReview, 18 (October),719-36.
Kohn, Meir G., CharlesF. Manski, and David S. Mundel
F--
FACE)th
Fr
with
Readings In The Analysis Of
Survey Data
Robert Ferber, editor
249 pp. 1980
$16/member $24/nonmember
Key pieces of the published literature
concerning applications of multi-
variate and related techniques to sur-
vey data, and new, innovative ap-
proaches to the analysis of survey
data are brought together in this book
of previously published articles.
Emphasis is given to recent material
although some of the classics in the
field have also been included. Biblio-
graphies follow each piece, to stimu-
late the researcher to go further in
examining the various techniques.
This content downloaded from 152.3.153.137 on Fri, 20 Mar 2015 15:33:01 UTC
View publication stats
All use subject to JSTOR Terms and Conditions
301
(1976), "An EmpiricalInvestigationof FactorsWhich In-
fluenceCollege-GoingBehavior,"Annalsof Economicand
Social Measurement,5 (Fall), 391-419.
Li, Mingche M. (1977), "A Logit Model of Homeowner-
ship," Econometrica,45 (July), 1081-97.
Luce, R. Duncan(1959), IndividualChoiceBehavior:A The-
oreticalAnalysis. New York:JohnWiley & Sons, Inc.
andP. Suppes(1965), "Preference,Utility, and Sub-
jective Probability,"in Handbookof MathematicalPsy-
chology-Volume3, R. DuncanLuce, RobertR. Bush, and
EugeneGalanter,eds. New York:JohnWiley & Sons, Inc.,
249-410.
Manski, Charles F. (1974), "The Conditional/Polytomous
LogitProgram:Instructionsfor Use," School of Urbanand
PublicAffairs, Carnegie-MellonUniversity.
McFadden,Daniel (1974), "ConditionalLogit Analysis of
QualitativeChoice Behavior," in Frontiers in Economet-
rics, Paul Zarembka,ed. New York: Academic Press,
105-42.
(1980), "EconometricModelsfor ProbabilisticChoice
Among Products,"Journal of Business, 53, No. 3, Pt. 2
(July), S13-29.
Punj,GirishN. andRichardStaelin(1978), "The ChoicePro-
cess for GraduateBusinessSchools," Journalof Marketing
Research, 15 (November),588-98.
Wald,A. (1943), "Tests of StatisticalHypothesesConcerning
Several ParametersWhen the Numberof Observationsis
Large," Transactionsof the AmericanMathematicalSoci-
ety, 54, 426-82.
Watson,PeterL. and RichardB. Westin (1975), "Transfer-
abilityof DisaggregatedMode Choice Models," Regional
Scienceand UrbanEconomics,5 (May), 227-49.
Ferber
Readings In Survey Research
Robert Ferber, editor
604 pp. 1978
$10/member $13/nonmember
A collection of readings which form
an extension of the special issue on
survey research of the August 1977
issue of Journal of Marketing Re-
search. The articles focus on three
aspects of survey research: sampling,
questionnaire preparation and data
collection. An extensive bibliography
is included.
TO ORDER call or write Order
Department, American Market-
ing Association, 250 S. Wacker
AMERICAN
MARKETING
ASOCIATION
Drive, Chicago, IL 60606, (312)
648-0536.