DECISIONBIASESINUSERAGREEMENTWITHINTELLIGENTDECISIONAIDS
ByJacobBennionSolomon
ADISSERTATION
Submitted
toMichiganStateUniversity
inpartialfulÞllmentoftherequirements
forthedegreeof
MediaandInformationStudiesÐDoctorofPhilosophy
2015ABSTRACT
DECISIONBIASESINUSERAGREEMENTWITHINTELLIGENTDECISIONAIDS
ByJacobBennionSolomon
IntelligentDecisionAids(IDAs)areemergingtechnologiesusedinareassuchasmedicine,
Þnance,ande-commercethatleverageartiÞcialintelligence,datamining,orrelatedcompu-
tationalmethodstoproviderecommendationstodecisionmakers.Animportantgoalfor
designersshouldbetohelpusersidentifyandacceptgoodrecommendationsandignorepoor
recommendations.However,considerableresearchhasfoundthatIDAusersfrequentlymake
poordecisionsaboutwhichrecommendationstofollow.
IpresentÞndingsfromthreestudiesthatprovideevidenceoffourdistinctdecision-
makingbiasesrelatedtoIDA-supporteddecisionmaking.Thesebiasesarecharacterizedby
anincreaseinusersÕagreementwithanIDAÕsrecommendationsthatisunassociatedwith
therecommendationsthemselvesbutassociatedwithsomeotheraspectofthedesignofthe
IDAoroftheuser.
InanexperimentthatmanipulatedtheperceivedcustomizabilityofanIDAthatassisted
usersinpredictingtheoutcomesofbaseballgames,Ifoundthatuserswhobelievedtheyhad
customizedtheIDAweremorelikelytofollowbothgoodandpoorrecommendationsthan
otheruserswhoreceivedidenticalrecommendationsfromtheIDAbutdidnotcustomizeits
logic.ThisÞndingisevidenceofa
customizationbias
.Importantly,thisstudyfoundthat
customizationbiasisnotcausedbyusersbelievingtheyhaveimprovedthealgorithmby
customizingit.
Inasecondexperiment,subjectswereencouragedtobelievethatthesystemhadei-
therhighorlowe
!
cacypriortoseeingrecommendations.Thisencouragementcreatedan
expectationsbias
inwhichsubjectsweremorelikelytofollowbothgoodandpoorrecom-
mendationswhentheyhadhigherexpectationsoftheIDAÕse
!
cacythanothersubjectswho
hadexpectedtheIDAÕsalgorithmtoperformpoorly.
Inthethirdexperiment,IassesseddecisionmakingbyusersofanIDAforrecommend-
ingexerciseactivities.SubjectswhousedacustomizableversionofthisIDA,wherethe
recommendationsdependedonhowusersconÞguredtheIDA,weremorelikelytoagree
withtherecommendationsthanuserswhoreceivedrecommendationsofsimilarqualitybut
didnotcustomizetheIDA.ThisÞndingshowsadditionalevidenceofcustomizationbias,
demonstratingthatitextendstoIDAswherethecustomizabilityhasrealinßuenceoverthe
recommendationsratherthanmerelyperceivedcustomizationasintheÞrststudy.Inthis
studyIalsofoundthatwhenusersbelievethatanIDAÕsinternallogicismoreclearand
understandable,theyaremorelikelytofollowrecommendationsregardlessoftheirquality.
ThisÞndingsuggestsa
transparencybias
.Therewasastrongrelationshipbetweenthequal-
ityofrecommendationsthatsubjectsreceivedandthequalityoftheirdecisions,indicating
thatwhendecisionmakersaresupportedbyIDAs,thequalityofrecommendationsisim-
portanttosystemsuccess.However,subjectswhoperformedthedecisiontaskunaidedby
anIDAperformedaswellastheIDA-supportedsubjects.
TheseÞndingsshowthatwhendecisionmakersareaidedbyanIDA,thesystema
"
ects
thedecisionmakingprocessbyrequiringuserstoevaluaterecommendations.IDAusersmay
makebiasedevaluationsduetocharacteristicsoftheinterfaceandinteractiondesignofthe
systemaswellasindividualcharacteristicsoftheusers.IntheconcludingchapterIdiscuss
theimplicationsoftheseÞndingsforthedesignofIDAsandrelatedsocio-technicalsystems,
aswellasforfutureworkoncomputer-supporteddecisionmaking.
ACKNOWLEDGMENTS
Therehavebeenmanycontributionsfrommanydi
"
erentpeoplethathavemadeitpossible
formetopursueaPh.D.andcompletethisdissertation,andIwouldliketoacknowledge
thosecontributionshere.
Ihavebeenfortunatetohavebeensupportedincountlesswaysbymyfamily,andIam
gratefultothembeyondwhatIcanadequatelydescribe.IamgratefultomywifeChalsea
forgivingmeherfullsupportformetocompleteaPh.D.Iamalsogratefultomyparents
whohavebeensupportiveandencouragingofmeunconditionallythroughoutmylife.
MyadvisorRickWashhasspentcountlesshoursprovidingfeedbackandguidanceonmy
workandtomeasayoungresearcher.Hehasgivenmewithopportunitiestoworkasa
researchassistantonhisgrant-fundedresearch,whichprovidednotonlyÞnancialsupport
butalsoinvaluableexperienceatdesigningandexecutingresearch.Thatexperiencehas
beencriticaltomydissertationworkandtomydevelopmentasaresearcher.Rickhas
encouragedmyideas,givenmeconstructivecritiques,andbeenpatientandallowedme
ßexibilitytopursuenewideas(evenbadones)thathavehelpedmelearnandgrow.
Mydissertationcommitteehasmademanyimportantcontributionstomydissertation
andIamgratefulfortheirinputandfeedbackthathaveimprovedthisworksubstantially.I
Þrstcameupwithseveralideasthatweretheseedsforthisdissertationfromthediscussion
andfeedbackIreceivedinWeiPengÕscourseinmyÞrstyearasaPh.D.student,andher
feedbackandsuggestionscontinuedtobehelpfulthroughouttheprocessofdevelopingand
reÞningmydissertation.EmileeRaderhasprovidedimportantfeedbackonthedesignand
executionoftheresearch,andshehasalsogivenmeimportantguidanceforimprovingthe
wayIcommunicatemyideas.Emileehasalsogenerouslyprovidedfundingforsomeof
researchinmydissertation.JoshIntronemadeseveralsuggestionsonboththedesignof
thestudiesandthedesignofthesystemusedinmyresearchthatprovedtobecriticalfor
ivobtaininginterpretableresearchresults.IamalsogratefulfortheguidanceIreceivedfrom
formermembersofmycommitteeGaryHsiehandJosephB.Walther,whocontributedto
theearlystagesofthisresearch.
IhavebeenamemberoftheBITLabatMSU,whereIhavebeenfortunatetohave
receivedhelpfulfeedbackonmanyoccasionsfromthefacultyandstudentsthere.Iam
particularlygratefultoChankyungPak,whohelpedmeÞgureouthowtodosomeofthe
morecomplexstatisticalproceduresthatareusedinmydissertation.
AndIalsowishtoacknowledgetheÞnancialsupportIreceivedforthisworkfromthe
CollegeofCommunicationArtsandSciencesintheformofaDissertationCompletionFel-
lowshipandSummerResearchFellowship.
TABLEOFCONTENTS
LISTOFTABLES
...................................ixLISTOFFIGURES
...................................xiCHAPTER1INTRODUCTION
...........................11.1IntelligentSystemsandIntelligentDecisionAids................2
1.2AgreementwithRecommendations.......................6
CHAPTER2BACKGROUND
............................112.1ComputationalTechniquesforDecisionSupport................12
2.1.1CollaborativeFiltering..........................12
2.1.2Content-BasedRecommendations....................14
2.1.3ArtiÞcialNeuralNetworks........................15
2.1.4GeneticAlgorithms............................15
2.2ApplicationsofIDA................................16
2.3IDAE
"ectiveness.................................17
2.4AgreementwithRecommendations.......................19
2.4.1Trust....................................19
2.4.2AutomationBias.............................21
2.5Transparency...................................23
2.6Customization...................................26
2.6.1CustomizationofInteractiveSystems..................27
2.6.2CustomizationinIDAs..........................29
2.6.2.1TheoreticalBasisforCustomizableIDAs...........29
2.6.2.2CustomizableIDAResearch..................33
2.6.3PotentialProblemswithCustomizationinIDAs............37
2.6.4SummaryofCustomization.......................40
2.7Summary.....................................41
CHAPTER3CUSTOMIZATIONBIAS
.......................443.1Introduction....................................44
3.2Methods......................................48
3.2.1GamePlay................................49
3.2.2IDAandConditions...........................52
3.2.3SubjectMatching.............................53
3.2.4Measures.................................56
3.3Hypotheses....................................59
3.4Results.......................................62
3.4.1DescriptiveStatistics...........................62
3.4.2E
!
cacyBeliefsandAgreement.....................65
vi3.5Discussion.....................................76
3.5.1Limitations................................82
CHAPTER4CAUSALEFFECTOFEFFICACYBELIEFSANDEXPECTA-
TIONSBIAS
..............................854.1Methods......................................86
4.2Results.......................................89
4.2.1Descriptiveresults............................89
4.2.2Manipulationcheck............................89
4.2.3Estimateofe
"ectofe
!cacybeliefsonagreement...........93
4.2.4Causale
"ectofe
!cacybeliefsonagreement..............94
4.2.5DecisionMaking.............................97
4.3Discussion.....................................98
4.3.1Limitations................................101
CHAPTER5CUSTOMIZATION,NON-CUSTOMIZATION,ORBOTHINAN
IDAFOREXERCISEDECISIONS
..................1025.1ResearchQuestions................................103
5.2ExerciseRecommender..............................107
5.2.1IDARecommendationData.......................107
5.2.1.1CrowdsourcingExercise/AttributeEvaluations.......110
5.2.1.2LatentAttributes........................111
5.2.2CustomizingRecommendations.....................113
5.3Methods......................................115
5.3.1DecisionTask...............................115
5.3.2ExperimentConditions..........................118
5.3.3Procedure.................................120
5.3.4Measures.................................122
5.4Results.......................................123
5.4.1Questionnaire...............................123
5.4.2RecommendationQuality........................125
5.4.3DecisionMaking.............................126
5.4.4Agreement.................................133
5.4.5Transparency...............................136
5.5Discussion.....................................138
5.5.1Limitations................................143
CHAPTER6CONCLUSIONS
............................1456.1IDA-SupportedDecisionMaking.........................145
6.2CustomizationasanIDADesign........................148
6.3ImplicationsforIDAsintheWild........................151
6.4Conclusion.....................................154
APPENDICES
......................................155AppendixATrustPropensityScale
........................156AppendixBBaseballKnowledgeScreeningQuiz
.................157AppendixCCategoryRatingsSurvey
.......................158AppendixDExerciseRecommenderSeedDataSurveyExample
.........159AppendixEExerciseActivitiesandLatentFactorScores
............160AppendixFExerciseRecommenderStudyInstructions
.............163AppendixGExerciseRecommenderInterface
...................167AppendixHPost-TestQuestionnaireforExerciseRecommenderStudy
.....171REFERENCES
.....................................178viiiLISTOFTABLES
Table3.1Factoranalysisofe
!
cacybeliefsmeasures..................57
Table3.2Factoranalysisofpropensitytotrustautomateddecisionaidsitems....58
Table3.3Descriptivestatistics..............................63
Table3.4E
!cacybeliefsmodel..............................66
Table3.5MultilevelmodelsofagreementwithIDArecommendations.........68
Table3.6Mediationanalysisforscoreagreement....................71
Table3.7Mediationanalysisforwinneragreement...................71
Table3.8Resultsofpost-testsurvey...........................74
Table3.9Recommendation/conÞgurationconsistencyledtogreateragreement....77
Table3.10Summaryofresults...............................78
Table4.1Descriptivestatisticsandfactoranalysisofdependentvariables.......89
Table4.2E
!cacyvariablesbyexperimentcondition..................91
Table4.3E
"
ectofencouragementone
!
cacybeliefs..................91
Table4.4E
"
ectofe
!
cacybeliefsonagreement.....................92
Table4.5Two-stageleastsquaresestimatesofcausale
"
ectofe
!
cacybeliefson
agreement.E
!
cacybeliefsareinstrumentedbyencouragementinthis
model......................................93
Table4.6Decision-makingquality............................97
Table4.7Summaryofresults...............................98
Table5.1Latentattributesfromexercisesurvey.....................111
Table5.2PercentagesofpeopleansweringyestothesequestionsabouttheExercise
Recommender..................................123
ixTable5.3Questionnairequestionsaboutrecommendations.Answeredon5-point
Likertscale(StronglyDisagreetoStronglyAgree).Standarddeviations
inparentheses..................................123
Table5.4Self-reporteduserexperiencevariables....................124
Table5.5Self-reportedunderstandingofIDAlogic...................125
Table5.6Meansandtandarddeviationsofrecommendationqualitywithinconditions.126
Table5.7Thee
"
ectofeachrecommendationtypeondecisionquality.........128
Table5.8DecisionmakingwithinonlytheIDA-supportedconditions.........130
Table5.9ModelofagreementwithoneoftherecommendationsinCustomonly
andNon-customonlyconditions........................132
Table5.10Transparencymeasuresbycondition.Itemsareona5-pointscale.....135
Table5.11Relationshipbetweenperceivedtransparencyandagreementwithrec-
ommendations..................................136
Table5.12Summaryofresults...............................139
Table6.1DecisionbiasesinIDA-supporteddecisionmaking..............146
TableE.1Exerciseratingsfactorscores..........................162
xLISTOFFIGURES
Figure2.1SomehypotheticalsystemdesignsandtheirplacementwithinParasur-
amanetal.Õsofdecisionmakingstagesandlevelsofautomation......31
Figure3.1HowcustomizationcancreateagreementwithIDArecommendations...46
Figure3.2CustomizableIDA...............................51
Figure3.3Distributionofmatchespercustomizer....................55
Figure3.4PreferenceforconÞguredcategories......................63
Figure3.5Averagewinneragreementbyroundnumber................65
Figure3.6Agreementwithrecommendedscore.....................67
Figure3.7ProbabilityofagreeingwithIDAÕspredictedwinner............69
Figure3.8Totalpointsearnedbysubjectsineachconditionoverthe12rounds...72
Figure3.9Decisionqualitybyrecommendationquality.................73
Figure3.10GreaterconsistencybetweenrecommendationandconÞgurationledto
moreagreement................................76
Figure4.1Winneragreementoverthecourseofthe12rounds.............90
Figure4.2E
!
cacyBeliefsoverthecourseofthe12rounds...............90
Figure4.3Distributionsofe
!
cacybeliefsbyencouragementcondition........92
Figure4.4Associationbetweene
!
cacybeliefsandagreement.............95
Figure5.1ExerciseRecommenderinterface.......................108
Figure5.2Distributionofweightedrecommendationscores..............127
Figure5.3Distributionofrecommendationscoreslistedas#1onthepage......127
Figure5.4Distributionofscoresofbestrecommendationonthepage.........129
Figure5.5E
"
ectofrecommendationqualityondecisionquality...........131
Figure5.6Agreementwithrecommendations......................134
xiFigure5.7Changeinaverageagreementoverthe10roundsoftheexperiment....135
Figure5.8E
"
ectoftransparencyonagreementwithrecommendations........137
Figure6.1IDA-supporteddecisionmakinginvolvesevaluatingrecommendations,
whichdoesnotÞtneatlyintoParasuramanetal.Õsframework.......147
FigureF.1Exerciserecommenderinstructions.....................164
FigureG.1Exerciserecommenderinterface.......................167
FigureH.1Exerciserecommenderpost-testsurvey....................171
xiiCHAPTER1
INTRODUCTION
Advancesinmachinelearning,artiÞcialintelligence,andrelatedcomputationaltechniques
havebeenwidelyappliedtohelppeoplefrommanydi
"
erentindustriesandcircumstances
makebetterdecisions.AclassofsystemscalledIntelligentDecisionAids(IDAs)provide
recommendationstodecisionmakersbyleveraginglargequantitiesofdataandapplying
artiÞcialintelligenceorsophisticatedstatisticalmodelstogeneraterecommendations.
Inthisdissertation,IarguethatusingIDAstoassistindecisionswithhighuncertainty
altersthedecisionmakingprocessbyrequiringuserstoevaluatetherecommendationsthat
thesystemprovides.Iwillshowthatthiscanbeachallengingtaskfordecisionmakersand
thatpeoplearenotalwayscapableofidentifyinggoodandpoorrecommendations.Iwillalso
showthatthewhilequalityofrecommendationsiscriticaltoagreementwithrecommenda-
tions,userscanhavebiasesthatcaninßuenceusersÕagreementindependentlyofthequality
ofrecommendations.Thereisacustomizationbias,whereusersareinclinedtoagreewith
recommendationswhentheyhaveparticipatedincustomizingtheIDAÕsinnerlogic.There
isalsoanexpectationsbias,whereusersaremoreinclinedtoagreewithrecommendations
whentheyexpectthesystemtoperformwellbecausetheybelieveitsprocessforproducing
recommendationsise
!cacious.Aconsistencybiasoccurswhenuserswillbemorelikely
toagreewithrecommendationsthatappeartobeconsistentwiththewaytheIDAwas
conÞgured.Andthereisatransparencybias,whereusersaremoreinclinedtoagreewith
recommendationswhentheyfeeltheyunderstandthelogicthatwasusedtoproducethem.
Thesebiasesareevidencethatthedesignoftheuserexperienceisinßuentialinthedecisions
thatusersmakeandthereforeinthee
"
ectivenessofanIDAasasocio-technicalsystem.
Thesebiaseshaveprovenancebothinthedesignofsystemsandintheusersofsystems.
Byobservingandreportingthesebiases,Imakeacontributiontotheoriesofcomputer-
1supporteddecisionmakingando
"
erknowledgethatadvancesknowledgeabouthowpeople
makedecisionswhensupportedbyintelligentsystems.Additionally,anunderstandingof
thesebiasesmakesapracticalcontributiontothedesignofIDAs.Becausethesebiases
showpredictablebehaviorthatiscausedbythedesignofthesystem,systemdesignerscan
navigatethesebiasesorevenemploythesebiasesinane
"
orttoengineerbetterdecision
makingbyusersoftheirsystems.
1.1IntelligentSystemsandIntelligentDecisionAids
SophisticatedcomputationaltechnologiessuchasartiÞcialintelligenceandÒbigdataÓalgo-
rithmsareincreasinglybeingdevelopedtoaddautomationtoknowledgeworkandtodaily
life(Carr,2014).Thesetechnologieshavebeencalled
intelligentsystems
(Guerlain,Brown,
&Mastrangelo,2000).ExamplesofintelligentsystemsarethealgorithmsthatÞltercontent
onsocialmediaorotheronlinecontentresources(Pariser,2011),systemsthatcollect,ag-
gregate,andanalyzedetailedpersonalinformationabouthealthandbehaviortoindividuals
inwhatareoftencalledsystemsfortheÒquantiÞedselfÓ(Choe,Lee,Lee,Pratt,&Kientz,
2014),searchenginesthatcatalogthewebandusealgorithmstomatchqueriestoinfor-
mationwithinthecatalog,recommendersystemsthatsuggestproductstobuyormovies
tosee,oralgorithmsthatcontrol(underhumansupervision)unmannedvehiclesinmilitary
operations(Clare,Cummings,How,Whitten,&Toupet,2012).
Afeatureofintelligentsystemsisthatdespitesophisticatedautomation,theystillne-
cessitatesomehumaninteractionorcooperation(Guerlainetal.,2000).Forthisreason,
intelligentsystemsaresocio-technicalsystems.Establishedprinciplesofhuman-centered
designargueformakingsystemfunctionsvisibleandcontrollable(Norman,1990a),yetthe
computationalmethodsbehindintelligentsystemsarenotalwaysamenabletovisibilityor
controllabilitybecauseoftheircomplexity.Thishascreatedaconsiderablechallengeforde-
signersofintelligentsystems,whomustdesigntheinterfacesandinteractionsbetweenusers
2andthealgorithmscontainedwithinthesystem.Asaresultofthisdi
!
cultyindesigning
theseinterfaceswithintelligentsystems,manysystemssu
"
erfrompoorusability,utility,or
haveotherpracticalchallenges.ÒFilterbubblesÓ,forexample,resultfromalgorithmsthat
Þlteronlinecontentinsuchawaythatthediversityofcontentthatanyonepersonsees
isslowlyreduced,andthislackofdiversitycangounnoticedbyusers(Pariser,2011).Or
aswillbediscussedinchapter2,intelligentsystemsusedinmedicine,aviation,andother
domainscanleaduserstomakepoordecisionsbecausetheyareill-suitedtoenablingusers
toproperlycalibratetheirtrustinthesystems.
Norman(1990b)argued25yearsago,inresponsetoconcernsregardingtheunfulÞlled
potentialofautomationtechnologies,thattheunfulÞlledpotentialofautomationisthere-
sultofinadequateinterfacesandinteractiondesignbetweenhumansandautomation.Iecho
thisargumenttodayregardingintelligentsystems.Intelligentsystemscannotbesuccessful
unlesstheinterfacesthata
"
ordhumaninteractionandcollaborationwiththemareaspow-
erfulandsophisticatedasthecomputationalmethodsthatmakethemÒintelligent.ÓIargue
forahuman-centereddesignapproachtointelligentsystemsthatfocusesondevelopingan
understandingofusers,designingtechnicala
"
ordanceswithinthissystem,andevaluating
therelationshipbetweenthenatureoftheusersandthetechnicaldesignofthesystemon
thesystemÕsoutcomes.Inthisdissertation,Ipresentresearchthatexaminesindividualand
aggregateddi
"
erencesinhowpeopleuseintelligentsystemstomakedecisions,testsdi
"
erent
designsofintelligentsystemsthatassistindecisionmaking,andevaluateshowthedesign
ofthesystemandthenatureofitsuserscombinetodeterminedecisionmakingoutcomes.
InthisdissertationIhavefocusedontheapplicationofintelligentsystemstodecision
aids.ThestudiesinthisworkevaluatetheusersanddesignofaclassoftechnologyIreferto
asintelligentdecisionaids
.IdeÞneintelligentdecisionaidsascomputationaltechnologies
that:¥ProvidearecommendationorsetofrecommendationsaboutspeciÞcactionsoritems
3thatmaybechosenbyadecisionmaker.
¥GeneratetheserecommendationbymeansofartiÞcialintelligence,statisticalormath-
ematicalmodeling,orsimilarlycomplexcomputationalmethodsthatcannotbeclearly
ore
!cientlyrepresentedtodecisionmakersinentirety.
¥Provideauserinterfacewithwhichadecisionmakerinteractsinordertoaccessthese
recommendations.
¥Refrainfromactuallyselectingorexecutinganactionordecisionwithouttheapproval
oftheuser.
SomeexamplesofsystemsthatareincludedinthisdeÞnitionofanintelligentdecision
aidare:
¥AclinicaldecisionsupportsystemsuchasDxPlain(Barnett,Cimino,Hupp,&Ho
"
er,
1987)whereaclinicianprovidesinformationaboutapatientÕscase,thenclicksabutton
toreceiveasetofpotentialdiagnosesthatshouldbemostseriouslyconsidered.
¥Asystemthatrecommendsmoviesthatusersmayenjoywatchingbyestimatingtheir
interestsincertaintypesofmoviesbasedondatacollectedabouttheuser.
¥Asystemthatrecommendsstocksthataninvestormaywishtopurchase.
¥Asystemthatalertsaluggagescreenerofthepotentialforahazardousitemusing
imagerecognitionsoftware.
Twotypesofintelligentdecisionaidshaveemergedthathavebeeninvestigatedseparately
bydistinctresearchcommunities.
IntelligentDecisionSupportSystems
havebeenthetopic
ofresearchininformationsystemsliterature.Oneoftheprimaryapplicationsofthislineof
researchhasbeentargetedtowardsdevelopingintelligentclinicaldecisionsupportsystems
(Berner,2007).OtherapplicationswithininformationsystemsliteraturehavebeeninÞnance
4andoperationsplanning.
RecommenderSystems
havebeenthetopicofconsiderableresearch
inhuman-computerinteractionandcomputerscience.Recommendersystemsarefrequently
designedfore-commercetohelpcustomersÞnditemsorservicesthatmatchtheirpreferences.
ManyrecommendersystemsuseacollaborativeÞltering(Su&Khoshgoftaar,2009)approach
inwhichrecommendationsaremadetodecisionmakersbyÞndingitemspopularamong
peoplesimilartothedecisionmaker.
User-centeredresearchonbothintelligentdecisionsupportsystemsandrecommender
systemsrevealsconsiderableoverlapinthesocio-technicalissuesthatmustbeconsideredin
thedesignofthesesystems.Issuesoftrust(Wang&Benbasat,2008;OÕDonovan&Smyth,
2005;Massa&Avesani,2007;Muir,1987;Sanchez,Fisk,&Rogers,2004)transparency
(Sinha&Swearingen,2002;Crameretal.,2008;Herlocker,Konstan,&Riedl,2000)and
usability(Lietal.,2012;Herlocker,Konstan,Terveen,&Riedl,2004)pervadetheresearch
frombothresearchcommunities.
IhavefocusedthisdissertationonIDAs,ratherthanothertypesofintelligentsystemsor
onintelligentsystemsmorebroadly,fortwoprimaryreasons.First,theyareanimportant
subclassofintelligentsystemsthatarebeingincreasinglyadoptedtoassistinmedicine,law,
Þnance,andotherprofessions(Carr,2014)wheredecisionmakersmakedi
!cultdecisions
thata
"
ectlives.BygeneratingnewknowledgethatcanbeusedtoimproveIDAs,this
researchcanmakeanimportantcontributionintherealworld.Second,IDAshaveseveral
propertiesthatmakethemidealforhuman-centeredsocio-technicalresearchasdescribed
above.Theycanbeusedtoassistinveryexplicitdecisionsforwhichadherencetothe
systemcanbeclearlyandobjectivelyevaluated.Othertypesofintelligentsystemsmayhave
amultiplicityofintendedpurposesorofdi
"
erentbehaviorsthatareofinterest.Filtering
algorithmsonsocialmedia,forexample,mayhavepurposestocreateanengagingsite
butalsotoselladvertisements,andtheremaybemanydi
"
erentoutcomesorbehaviorsof
interestlikecontentconsumption,contentcreation,orcontentsharing.IDAscan,atleast
inalabsetting,beplausiblypresentedtohaveasingularpurposeandasingularbehavior
5ofinterestwithclearandobjectivelydeÞnedcriteriaforevaluation.Thesepropertiesmake
IDAshighlyusefulforquantitative-basedresearchthatseekstounderstandusersandtest
intelligentsystemdesigns.
However,inspiteofthisfocusonIDAs,thereisanimportantreasonthattheresultsof
thesestudiesmightberelevanttointelligentsystemsmorebroadly.TheissuesthatIhave
assessedinthisdissertationarelargelyfunctionsofthedesignoftheinterfacebetweenusers
andthesophisticatedcomputationalmethodsembeddedinthesystem.Thecustomizability
ofanalgorithm,theconsistencybetweenalgorithminputandoutput,thetransparencyof
thealgorithm,orthesystemÕsself-evaluationofitsowne
!cacyareaspectsofthedesign
that,basedontheresultsofthestudiesdescribedinthefollowingchapters,cana
"
ect
usersÕbehaviorwheninteractingwithintelligentsystems.AlthoughthespeciÞcbiasesand
behaviorsIobserveinthisdissertationmaymanifestthemselvesdi
"
erentlyinnon-IDA
intelligentsystems,theyareworthyofconsiderationinhuman-centereddesignandresearch
onthosesystemsaswellandthisdissertationprovidesabasisforextensionsofthisIDA-
focusedresearchtoothertypesofintelligentsystems.
1.2AgreementwithRecommendations
Consideradoctorwhousesacomputerizedsystemthatsuggeststreatmentoptionsfor
patients.Thedoctormayaddrelevantinformationaboutthepatientandthediagnosis,
andthesystemwillreturnalistofsuggestedtreatments,andmaynotewhichtreatmentor
treatmentsitbelievesaremostlikelytobesuccessful.Thedoctormustthenexaminethislist
andconsidertheseoptions,aswellasconsideroptionsthatheorsheisawareofbutthathave
notbeenrecommendedbythesystem.Imagineifthetoprecommendationisanoptionthat
thedoctorwouldnothaveconsideredorwithwhichthedoctorhaslittleexperienceandwould
bereluctanttoprescribe.Andalsoimagineifthistoprecommendationisthetreatment
thatwouldtrulybemostbeneÞcialtothepatient.Ifthedoctordecidestofollowthis
6recommendationandprescribesthistreatment,thesystemhasbeene
"
ectiveatimproving
decisionmaking.However,ifthedoctordoesnotfollowthisrecommendationandinstead
choosesalessbeneÞcialtreatmentforthepatient,thesystemhasfailedtoimprovedecision
makingdespiteachievingatechnicalsuccessatÞndingthebestpossibletreatmentforthe
patient.Areversesituationisalsoplausible,wherethesystemrecommendsalessthanideal
treatmentbutthedoctorignoresthatrecommendationandchoosesanotherbettertreatment
thathasnotbeensuggestedbytheIDA.Inthiscase,thesystematworsthascausedno
harmandmayevenbeconsideredasuccessiftheprocessofusingthesystemcontributed
insomewaytothedoctormakingthecorrectdecision,eventhoughitsrecommendation
wasnÕtfollowed.Butifthedoctorchosetofollowthepoorrecommendationeventhough
itisnotwhatheorshewouldhaveotherwisechosen,thanthesystemwillhaveactually
causedharm.
ThisscenarioillustrateswhyitisimportanttounderstandwhatcausesusersofIDAsto
followorignorerecommendations.IDAscanonlybee
"
ectiveatimprovingdecisionsifthey
a)providerecommendationsthatarebetterthanwhatdecisionmakerswouldotherwise
choose,andb)persuadedecisionmakerstofollowthosegoodrecommendations.Ifboth
ofthoseconditionsarenotmet,anIDAmaybeine
"ectiveorevenharmful.Designers
mustÞndwaystoimproveboththequalityandtheacceptabilityoftherecommendations
thatthesystemproducessimultaneously,andthiscanbeaseriouschallenge.Itisfurther
complicatedbythefactthatifIDAsarenotsuccessfulorareinconsistentatcreatinghigh
qualityrecommendations,thananye
"
ortstopromoteagreementwithrecommendationsmay
actuallybecounterproductivebecausetheywillleaduserstomakebaddecisions.
IDAs,inspiteoftheire
"
ortstoprovidecriticalinformationtodecisionmakers,may
createnewuncertaintyfordecisionmakersiftheydonotfullyunderstandhowthesystem
hasproduceditsrecommendations.Thedoctorintheexampleabovemaychoosetoignore
thegoodrecommendationbecauseheorshedoesnotunderstandwhyitwassuggested,
insteadoptingforalesseroptionbecauseitisbetterunderstood.IDAshavebeenshown
7tofrequentlysu
"
erfromalackof
transparency
inthatusersdonotfullyunderstandhow
theyworkorwhyspeciÞcrecommendationswereprovided(Sinha&Swearingen,2002).
However,therehasbeenconsiderablee
"
ortinIDAresearchanddesigncommunitiesto
designmoretransparentsystems,primarilybyprovidinguserswithclearexplanationsfor
recommendations(Lim,Dey,&Avrahami,2009;Ehrlichetal.,2011;Tintarev&Mastho
"
,2011).Nevertheless,transparencyisnoteasilyachievedinIDAdesigns(Herlockeretal.,
2000).ResearchontransparencyinIDAshasoftennotmadeadistinctionbetween
understand-
inghowasystemworksand
preferring
howitworks.Iarguethatuserscanlooselybe
categorizedasthosewhodonotunderstandhowasystemworks,thosewhounderstandhow
itworksandbelievetheprocessise
"
ective,orthosewhounderstandanddonotthinkthe
processise
"
ective.InthisdissertationIwillshowthatthesedi
"
erenttypesofuserswill
oftenmakedi
"erentdecisionsinregardtoagreementwithIDArecommendations,evenwhen
therecommendationsthemselvesarenodi
"
erent.Iwillshowthatwhenusersunderstand
howasystemworksandfeelthatthesystemÕsprocesshasgoode
!cacy,theywillbemore
likelytofollowbothgoodandpoorrecommendationsthanotherswhohavethesamelevel
ofunderstandingbutfeelthesystemhaslowe
!cacy.Thisisimportantbecauseitprovides
atargetforsystemdesigners.Aslongasthesystemcanproducegoodrecommendations,
designerscanencouragestrongagreementbytakingmeasurestoincreaseusersÕexpectations
aboutthee
!cacyofthesystemÕsrecommendationlogicsothattheyarelikelytofollowthe
goodrecommendations.Atthesametime,systemsneedtodevelopmethodstohelpusers
understandwhenrecommendationsmaybeunreliable.
Onewaythatdesignersmaybeabletosimultaneouslyimproverecommendationsand
usersÕbeliefsaboutitse
!cacyisbydesigningcustomizablesystemsthatrequirethedecision
makertoprovidespeciÞcinputtothesystemÕslogicoralgorithm.Thisinputmaybeused
toprovidelocalexpertiseorinformationtohelptheIDAprovidearecommendationthat
bestsuitsthespeciÞccircumstancesofthedecision.Andbyallowinguserscontroloverthe
8algorithm,customizationcanallowuserstogiveitaconÞgurationthattheybelievehashigh
e!
cacy.
Ihavepreviouslyfound,however,thatcustomizationleadsuserstomoreagreement
withrecommendationsintheirdecisions(Solomon,2014).Thisisadecisionmakingbias
thatIcall
customizationbias
.Inchapter3,Iwillshowthatusersarebiasedbyperceived
customizationevenwhentheydonotbelieveithasledtoanygreatere
!
cacyoftheIDA.
ThishasanimportantimplicationforIDAresearchanddesign,asitisanexampleofhow
theprocessandexperienceofusinganIDAcanimpactthedecisionmakinginwaysthatare
unrelatedtothequalityofrecommendationsthesystemproduces.Inchapter5Iwillshow
thatthisbiasisobservableunderastateoftruecustomizationwhereusersÕinßuenceover
therecommendationsmorevisiblethaninthestudyinchapter3.Thisadditionalevidence
providessomeexternalvaliditytotheÞndingofcustomizationbias.
WhenusershaveanunderstandingoftheconÞgurationofasystemÕslogic,whetherbe-
causetheyhavecustomizedthatconÞgurationthemselvesorhavesimplybeenmadeaware
ofitinthedesignoftheIDA,thereisanopportunityforthemtoevaluatewhetherthe
recommendationsareconsistentwiththeconÞguration.Forexample,inanIDAforrecom-
mendingstockstopurchase,iftheusercustomizesthesystemtofocusonenergy-related
stocks,andtherecommendationsdonotsuggestanyenergy-relatedstocks,theusermight
thinkthatthesystemhasmalfunctionedorissimplyapoorsystem,andthenmightchoose
toignoretherecommendations.Butthesystemmayhaveagoodreasonfornotsuggest-
ingenergy-relatedstocksinspiteoftheconÞguration(e.g.itsalgorithmthinksallenergy
stocksarepoorinvestmentsatthatmoment),andthereforeignoringthisrecommendation
wouldbeabadchoicebytheuser.Inthisscenario,thelackofconsistencybetweenthe
conÞgurationandtherecommendationsiscausingabias,becausetheuserisignoringthe
recommendationonthebasisofalackofconsistencyeventhoughthesystemismakinga
goodrecommendationthatshouldbefollowed.Iobservedthisbiasinthestudypresented
inchapter3,andIarguethatitisanotherexampleofthedesignoftheinteractionbetween
9userandIDAinßuencingdecisionsindependentlyofthequalityofrecommendations.
10CHAPTER2
BACKGROUND
Designingintelligentdecisionaidspresentsanumberofchallengesduetothesocio-technical
natureofcomputer-supporteddecisionmaking.Computingtechnologiescanbepowerful
andcapableofprovidingvaluableinsighttodecisionmakers.However,decisionmakersare
humanandthereforediverseintheircapabilitiesandcharacteristics,pronetousingheuristics
andhavingbiasesindecisionmaking,andmayhavedi
"
erentdecisionmakinggoalsthan
systemdesigners.Forthisreason,anydesignthatkeepshumansintheloopofdecision
makingmustaccountforthesehumanfactors.
Inthissection,Iwilldiscussexistingresearchonthedesignofintelligentdecisionaids.I
willdescribesomecommontechnicalapproachestoprovidingcomputerizeddecisionsupport.
IwillalsogiveanoverviewofresearchonhumanfactorsinIDAs,withafocusonissuesof
trustinautomation,transparencyofsystemsÕinnerlogic,usercontrolandthedivisionof
laborbetweenuserandsystem.
Fromthisoverview,Iwillmakeatheoreticalargumentthatacompletelytop-down
approachtoIDAdesignwheresystemengineersmakealldeterminationsabouthowthe
systemproducesrecommendationsmaybeinadequateformaximizingthepotentialofIDAs.
Iwillprovideevidencefromtheliteraturethatallowingend-userstohavesomecontrolover
thedesignofanIDAÕsinnerlogic,aprocesscalled
customization,canservetomakeIDA
designslessvulnerabletoknownhumanfactorsproblemswithIDAs.Iwillalsodiscussthe
theoreticalbasisfornewhumanfactorsproblemsthatmayarisefromanIDAdesignthat
a"ordsend-usercustomizationofitsrecommendation-producingprocess.
112.1ComputationalTechniquesforDecisionSupport
Thereisawidevarietyofcomputationaltechniquesthathavebeendevelopedanddeployed
withinIDAstoprovidesupporttodecisionmakers.Forthemostpart,thesecomputational
techniquesareusedtoprocessandanalyzedatathatisavailabletothesysteminorder
toproducerecommendationsfortheuserabouttheirdecision.InthissectionIwillbrießy
describeafewofthemostcommoncomputationaltechniquesthatareusedinIDAs,including
adiscussionofthestrengthsandweaknessesoftheseapproaches.
2.1.1CollaborativeFiltering
ManyIDAsmakeuseofacollaborativeÞltering(Su&Khoshgoftaar,2009)approachto
generatingrecommendations.IncollaborativeÞltering,usersgenerallyprovideratingsor
otherinformationabouttheirpreferenceforitemsinasystemÕscatalog.Collaborative
Þlteringtechniquestypicallycreateamatrixorsetofmatricesthatrepresentusersoritems
withinthesystem.AUser-Itemmatrixrepresentsaratingorotherformofvaluationfor
everycombinationofuseranditemwithinthesystem,althoughmanyofthecellsofthis
matrixmaybeemptybecauseuserstypicallyonlyrateasmallportionofallpossibleitems.
AUser-User(aswellasanItem-Itemmatrix)storesaÔsimilarityÕvaluecalculatedusinga
distancemetricsuchascosinedistanceorEuclideandistancebetweeneachpairofusersor
items.Forexample,considertwouserswhohaveeachratedasetofmovies.Thesimilarity
betweenthesetwouserscanbecalculatedbythecosinedistancebetweeneachusersÕvector
ofratedmovies.Ifthetwousersmostlyagreeontheirratingsofmovies,thecosinedistance
willbesmallandthesimilaritylarge.ThissimilarityvalueisthenstoredintheUser-User
matrix.Fromthesematricesorsimilardatarepresentations,recommendationscanbegenerated
usingoneofavarietyofcomputationaltechniques.Acommonapproachtoproducingrecom-
mendationsfromthesematricesisk-nearestneighbors(Ekstrand,Riedl,&Konstan,2011).
12Inthisapproach,thesystemÞndsasmallgroup(ofsizek)ofusersfromthesystemthatare
mostsimilartothecurrentuserintermsoftheirexplicitvaluationsofitems,thenrecom-
mendstothecurrentuseritemsthataremostpopularwithintheirÔneighborhoodÕ.Another
approachismatrixfactorization.Matrixfactorizationtechniques(Koren,Bell,&Volinsky,
2009)seektoÞndlatentfactorswithinthematricesthatindicatesomeunderlyingconcept
orvariableamongasetofusersÕpreferencesforitems.Forexample,matrixfactorization
techniquesmayrevealthatmoviesmayfallonsomescalebetweenÔseriousÕandÔescapistÕ
andthatusersÕpreferencesforthedi
"erentextremesofthisscaleismanifestedintheirrat-
ingsofmovies(Korenetal.,2009).Thus,byidentifyingsuchlatentfactors,asystemcan
determinebothwhereonthescaleauserÕspreferenceliesandwhereonthescaleeachitem
lies,andrecommenditemsthatareaclosematchtotheinferredpreference.
ArelatedapproachtocollaborativeÞlteringisanetwork-basedapproachtogenerating
recommendations.Inthisapproach,asocialnetworkisconstructedfromthesetofusersand
itemsinthesystem,andvariousattributesofthenetworkareusedtomakerecommendations
suchasthecentralityofnodes.Forexample,Qinetal.(2010)builtarecommendersystem
forYouTubevideosbybuildinganetworkofvideosasnodesandconnectinganytwovideos
iftheyshareatleastonecommonuserwhohascommentedonthevideo.Theythenmake
recommendationsbyusingnetworkpropertiestoproduceexpectedutilitiesforvideosand
usersandrecommendingtousersthevideoswiththehighestexpectedutilities.
Asocialnetworkingapproachtorecommendationscanbeparticularlyusefulwhenso-
cialconnectionsthemselvesaretheitemsbeingrecommended.Socialnetworkingsitescan
recommendotheruserstoeachotherbyevaluatingthenumberofsharednodeswithinthe
networkofusers,aswellasotherfeaturesofthenetworkstructure,andusethisinformation
tosuggestnewconnections(Rothetal.,2010).
OneofthegreateststrengthsofcollaborativeÞlteringisthatitisÒcontent-independentÓ
(Park&Chu,2009).Thismeansthatthesystemdoesnotneedtohaveverymuchexplicit
informationabouttheitemsitrecommends.Forexample,movierecommendationscanbe
13madeusingcollaborativeÞlteringwithjustthetitleofeachmovieandasetofratingsfrom
users,sincetherecommendationsareentirelydeterminedbytheratingsandnotanything
inthecontentoftheitems.Thiscontent-independencecanalsobebeneÞcialinthatitcan
leadtomoreÒserendipitousÓrecommendations.CollaborativeÞlteringcanÞnditemsthat
aresimilartoitemsauserlikesbutinwaystheuserhasnotpreviouslyconsidered(Park&
Chu,2009).
Thecontent-independence,however,comesatacostforcollaborativeÞlteringofbeing
highlyuserdependent.Thisuser-dependenceisknownastheÒcold-startproblemÓ(Lam,Vu,
Le,&Duong,2008).CollaborativeÞlteringrequiresdataaboutusers,particularlyratings
ofitems,inordertobeuseful.Itrequiresarelativelylargeexistinguserbaseinorderto
Þndmeaningfulclustersofusersthatsharepreferences.Andforanygivenuser,itneeds
ratingsorotherinformationaboutthatuserinordertoÞndsimilarusersfromwhichto
makerecommendations.Newsystemsdonothavethedatatheyneedtobeusefultousers,
whichmakesitdi
!
culttobuildthecriticalmassofusersrequiredtoeverbecomeuseful.
2.1.2Content-BasedRecommendations
Content-basedrecommendersystemsproducerecommendationsbyusingknownandexplicit
attributesofitems.ThesesystemsmaintainanItem-Attributematrixwhereallitemshave
beenevaluatedontheattributes,andthesystemsthenelicitpreferencesfromusersfor
attributesandrecommenditemsthataresimilartousersÕstatedpreferencesforattributes
(Leino,2014).
Oneoftheprimaryadvantagesofcontent-basedrecommendersisthattheysu
"
erless
fromthecold-startproblemthatplaguescollaborativeÞltering(Schein,Popescul,Ungar,&
Pennock,2002).Content-basedrecommendersdonotrequireacriticalmassofexistingusers
tobeabletomakerecommendations,meaningthatnewsystemscanbemoreusefulfrom
inception.However,content-basedrecommendersdostillneeddataaboutagivenuserin
ordertomakepersonalizedrecommendations,meaningthecold-startproblemisnotentirely
14eliminated.
Aproblemwithcontent-basedrecommendersisthatcuratingthecontentfeaturescanbe
di
!
cultandcostly.Pandora.com,forexample,usesacontent-basedrecommender(Glaser,
Westergren,Stearns,&Kraft,2006)thatrequireseverysongintheirdatabasetoberated
byamusicalexpertonover400musicalfeatures
1.Developingandcuratingthedatainsuch
awaymaynotbepracticalforalltypesofrecommendations.
2.1.3ArtiÞcialNeuralNetworks
OtherinnovationsinartiÞcialintelligencehavebeenadoptedinsomecontexts,mostnotably
inmedicine.ArtiÞcialNeuralNetworks(ANNÕs)areaformofartiÞcialintelligencethat
triestoÞndpatternsamonglargeandcomplexdata(Hill,Marquez,OÕConnor,&Remus,
1994).ANNÕsareareplicationofbiologicalneurologicalprocessesandhavebeenfound
tobepowerfulinmanyanalytictasks,particularlyincontextswheretraditionalstatistical
approachesbasedonregressionareproblematic(Hilletal.,1994).ANNÕshavebeenfound
tobeparticularlyusefulforIDAthatsupportclinicaldecisionmaking(Berner,2007).A
knownshortcomingofANNÕsisthatunlikemanyothertechniques,ANNÕsdonotprovide
clearreasoningabouthowtheyhavecometoconclusions(Berner,2007).
2.1.4GeneticAlgorithms
GeneticalgorithmsareanothertechniquefromartiÞcialintelligencethathavebeendeployed
withsomesuccessinIDAs(Berner,2007).Geneticalgorithmssimulateanevolutionary
processbasedonthenotionofsurvivaloftheÞttest.Attributesofrecommendableitems
thataredistributedamongadatasetarerandomlycombinedandtheresultingcombinationis
evaluatedaccordingtoanestablishedcriteriaorÒÞtnessfunction.ÓCombinationsthatyield
thebestevaluationsarekept,whileweakercombinationsaredropped,andnewcombinations
1https://www.pandora.com/about/mgp
15formedfromtheremainingset.Thisprocessiscontinueduntilperformancestopsimproving,
atwhichpointasolutionisapparentthatcanbeusedasabasisforrecommendations.
GeneticalgorithmshavethesameshortcomingasANNÕsinthattheystruggletoprovide
clearreasoningfortherecommendationsthatareproduced(Berner,2007).
2.2ApplicationsofIDA
Oneofthechallengesinresearchinganddevelopinge
"
ectiveIDAsisthetremendousvariety
inthedecisioncontextsforwhichtheyareimplemented.TheapplicationsofIDAtechnolo-
giesrangefrome-commercetoÞnancetomedicinetojournalism.E-commercewebsitessuch
asAmazon.comhaveadoptedrecommendersystemstosuggestproductstocustomersand
helpthemchooseitemsfromanenormouscatalog(Linden,Smith,&York,2003).Recom-
mendersystemsonlargesitessuchasAmazonmaymakerecommendationsaboutspeciÞc
itemswithinaclassofalternatives(e.g.whichbooktoread)aswellassuggestingdi
"
erent
classesofitemsforitscustomerstoconsider(e.g.whethertolookforabookoracamera).
Othere-commercerecommendersystemsfocusonamorespeciÞcdecisioncontext,suchas
puttingtogetheranoutÞtorwardrobe(Tu&Dong,2010).
SomehighlysuccessfulIDAimplementationsmakerecommendationsaboutmediacon-
sumption.Netßixhassoughttoimprovetheuserexperienceofitsservicebyresearching
anddeployingarecommendersystem
2.MovieLens(Miller,Albert,Lam,Konstan,&Riedl,
2003)isasimilarmovierecommendationservicewhichhasbeendevelopednotonlytomake
recommendationstodecisionmakersbutasatestinggroundforexperimentalapproachesto
recommendersystemdesign.OnlinemusicservicessuchasPandora
3andLastFM
4have
usedrecommendersystemstoengageusersbyprovidingpersonalizedbutserendipitousmu-
sicrecommendations.SocialnetworkingsitessuchasasFacebook,Twitter,orLinkedIncan
2http://www.netßixprize.com/index
3http://www.pandora.com
4http://www.last.fm
16helppeoplemakedecisionsaboutwhopeopletointeractorcommunicatewith.LinkedIn,
forexample,makesrecommendationstousersaboutwhotheymightconnectwithforpro-
fessionaldevelopmentandnetworking,aswelljobsorothercareeropportunities(Skeels&
Grudin,2009).AnimportantapplicationofIDAsisinrecommendingnewsarticlesorother
typesofweb-contentforuserstoconsume.Largeweb-portalssuchasGoogleandYahoo
usetechniquestakenfromrecommendersystemstoproducepersonalizednewsportalsthat
recommendcontentthatisexpectedtobeofinteresttousers(Liu,Dolan,&Pedersen,
2010).2.3IDAE
!
ectiveness
IDAshavebeenfoundtobee
"
ectiveatreducingthetimeande
"
ortrequiredtomake
decisions,(Hostler,Yoon,&Guimaraes,2005;Amento,Terveen,Hill,Hix,&Schulman,
2003;Chen&Pu,2009;Xiao&Benbasat,2007).IDAscanautomatetheacquisitionand
analysisofinformation,reducingthenumberofalternativesthatadecisionmakermust
consider(H¬aubl&Trifts,2000)andthusreducinge
"
ort.ThisisanimportantbeneÞtthat
IDAsprovidetodecisionmakersthatjustiÞestheirdevelopmentandimplementation.
However,someresearchersarguethatIDAsshouldbeneÞtnotonlydecisionprocesses
suchasthee
!
ciencyofmakingdecisions,butalsothedecisionoutcomesthemselves(Knijnenburg,
Willemsen,&Kobsa,2011;Xiao&Benbasat,2007).Researchershavehaddi
!
cultyines-
tablishingthee
!
cacyofIDAswhenusedinpractice.Forexample,inclinicalsettings,Bright
etal.(2012)conductedasystematicreviewofclinicaltrialsofIDAs.Theyfoundthatonly
20%ofrandomizedtrialsofclinicalIDAsevenevaluateddecisionoutcomes,whereasmost
werefocusedonassessingdecisionprocessmeasuresandtheeconomicjustiÞcationofIDAs.
Amongthetrialsthatdidassessdecisionoutcomes,theyfoundonlylimitedevidencethat
IDAswerebeneÞcial.Theynotedhoweverthatthelackofclearevidencemaybeduetothe
tremendousdi
!
cultythatexistsinexecutingclinicaltrialsforIDAsthatevaluatedecision
17outcomes.Theremaybeaseriousselectionbiasinthatsystemsandcircumstancesforwhich
clinicaltrialscanbeeasilyexecutedarealsothoseforwhichtheparticularsystemsareinef-
fective.Brightetal.concludednotthatIDAswereine
"
ectivebutthatthereisinsu
!
cient
evidencetodrawaconclusion.
Ine-commerce,therearenocomparablereviewsorevenanystudiestomyknowledge
thatassessthebroadimpactofIDAsondecisionqualityordecisionoutcomesthatarebased
ondataoutsideofalaboratorysetting.Thismaybeduetothehorizontaldi
"erentiation
ofmostdecisionsine-commerceinwhichusersmayhavewidelyvaryingcriteriaforwhat
theyprefer,makingitdi
!culttodevelopmeasuresfordecisionqualitythatareobjective.
However,therehasbeenconsiderableworktoevaluatee-commerceIDAsinlabsettings
wheredecisionqualitycanbeobjectivelydeÞnedandmeasured(Xiao&Benbasat,2007).
Someexamplesofsuchmeasuresincludeamatchbetweentheattributesofachosenitem
andsubjectsÕpreferencesforthoseattributes(Pereira,2001)orthefrequencywithwhich
subjectschangetheirdecisionlaterwhengivenacost-freeopportunitytochange(H¬aubl&
Trifts,2000).XiaoandBenbesat(2007)conductedasystematicreviewofIDAresearchin
e-commercethatusedsuchmeasures,andnotedthatthereweremixedÞndings,withsome
studiesÞndingtheIDAshelpeddecisionmakingwhileothersfoundnoe
"
ectoreventhat
IDAsharmeddecisions.However,theyfoundthatvariationsinthedesignoftheIDAsused
inthesestudies,suchasthewaypreferenceswereelicitedandthewayrecommendations
presented,hadanimpactondecisionquality.ThissuggeststhatthedesignoftheIDAis
animportantfactorfordecisionmaking.Italsosuggeststhatthereisastrongneedfor
socio-technicaltheoryofIDAdesigntohelpdesignersengineergooddecisionmaking.
182.4AgreementwithRecommendations
2.4.1Trust
Technologiesthatautomatedecisions,actions,informationprocessing,orotherfunctions
withinahuman-machineinteractionplaceaburdenonuserstodeveloptrustinthesystem.
TrustintechnologieshasbeenakeyareaofresearchonIDAsandinautomationtechnologies
moregenerally.Researchersintheseareashavesoughttounderstandhowusersdeveloptrust
insystems,howuserslosetrustinsystems,andhowusersÕtrustinanautomatedtechnology
inßuencestheirbehaviorsordecisions.
LeeandMoray(1992)arguethatasautomationisincreasedinasystem,thecontrol
thatusersexertoverasystemanditsoutputshiftsfromanactivecontroltoÒsupervisory
control.ÓThisroleofasupervisorratherthanadirectcontrollerdemandsthatusersplace
anincreasedamountoftrustinthesystem.
Inseekingtounderstandhowusersdevelopandmaintaintrustinautomatedtechnolo-
gies,researchershavedisagreedovertheappropriatewaytodeÞnetrustwithinthecontext
ofahuman-machineinteraction.Oneapproachthathasfrequentlybeentakenistoborrow
deÞnitionsoftrustfromresearchoninterpersonalrelationships(Ho
"
man,Johnson,Brad-
shaw,&Underbrink,2013;Muir,1987;Madhavan&Wiegmann,2007;Wang&Benbasat,
2008).InterpersonaltrusthasbeendeÞnedasÒawillingnessofapartytobevulnerableto
theactionsofanotherpartybasedontheexpectationthattheotherwillperformaparticular
actionimportanttothetrustor,irrespectiveoftheabilitytomonitororcontrolthatparty?Ó
(Mayer,Davis,&Schoorman,1995,p.712).Inthissense,trustinautomationcanbeviewed
asauserÕswillingnesstobevulnerabletoasubordinatemachineÕsactionsinthesameway
thatsupervisorsinahierarchicalinterpersonalrelationshiparevulnerabletotheactionof
thosetheysupervise.TheÒComputers-Are-Social-ActorsÓparadigm(Nass&Moon,2000),
whichdemonstratesahumanpropensitytoformsocialrelationshipswithinteractivema-
chines,hasoftenbeenusedasajustiÞcationfortreatingandmeasuringtrustintechnology
19usinganinterpersonalconceptualizationoftrust(Madhavan&Wiegmann,2007).
However,muchworkhasfoundthatpeopletrusttechnologyindi
"
erentwaysthanthey
trustotherpeople.LeeandSee(2004)claimthatwhileinterpersonaltrustissymmetrical
inthatbothpartieshaveaneedtodeveloptrustinoneanotherthroughrepeatedinterac-
tion,trustinautomationisgenerallyasymmetrical.Usersmayneedtodevelopbeliefsand
attitudesthatallowthemtobewillinglyvulnerable,butautomatedsystemsgenerallydo
notneedtodeveloptrustintheirusers.
OtherworkhasfoundthatpeopleÕsleveloftrustinsystemsoftendi
"ersfromtheirlevelof
trustinahumanperformingthesamerolewiththesamereliability.Forexample,Dzindolet
andcolleagues(Dzindolet,Pierce,Beck,&Dawe,2002;Dzindolet,Peterson,Pomranky,
Pierce,&Beck,2003),inexperimentsthatcontrolthereliabilityandutilityofbothan
automateddecisionaidandahumandecisionaid,foundthatpeoplehaveaninitialtendency
totrustandrelyontheautomatedaidmorethantheytrustthehuman.However,theyalso
foundthatincaseswheretheaidmadeanobviouserror,usersÕrelianceontheautomated
aidplummetedconsiderablytoalevelfarbelowboththereliabilityoftheautomatedaidand
belowtherelianceobservedinaconditionusinganotherhumanasthedecisionaidinplaceof
theautomatedsystem.ThissensitivitypresentsanimportantchallengefordesigningIDAs.
DesignersmustÞndwaystoappropriatelysetexpectationsofsystemquality.Expectations
canbeapowerfulinßuenceonbehaviorinsocio-technicalsystems(Wash,2013),andIDA
usersmayhavedi
!
cultyformingaccurateexpectationsofthesystem.
Trust,particularlyinthecontextofautomationuse,istypicallydeÞnedasanattitude
orabelief(J.D.Lee&See,2004).Thisdistinguishestrustfromarelatedconceptknown
asreliance
,whichisabehaviorinwhichusersrelyonordeferactiontoanautomated
system.LeeandSee(2004)arguethatwhiletrustinßuencesreliance,thetwoconceptsare
notidentical.Theyarguethattrustisexplicitlyanattitudeandshouldbetreatedand
measuredassuchinresearchontrustandautomation,whereasrelianceisthebehaviorof
deferringactionsordecisionstoanautomatedsystem.
20Animportantquestionthathasbeenexploredintheresearchontrustinautomation
relatestohowtrustisdeveloped.Onehypothesisisthattrustistheresultofreliabilityover
time.Asasystemdemonstratesthatitisreliable,usersgrowtotrustitmoreandbecome
morereliantontheautomation.LeeandMoray(1992)suggestthattrustinautomationis
enhancedwhenusersunderstandtheprocessthattheautomationusestoproduceitsoutput.
Anumberofindividualdi
"
erencesbetweenuserscanalsoimpacttrustinautomation.
Sanchezetal.(2004)foundthatolderadultsweremoresensitivethanyoungeradultsto
declinesinthereliabilityofanautomateddecisionsupportsystemforadrivingtask,losing
trustinthesystemmorequicklyasthesystembecamelessreliable.Merrittandcolleagues
(Merritt&Ilgen,2008;Merritt,Heimbaugh,LaChapell,&Lee,2012)showedthatsome
peoplehavemoreofageneralpropensitytotrustautomationthanothers.
Someresearchers(OÕDonovan&Smyth,2005;Massa&Avesani,2007)haveusedthe
notionoftrustasabasisformakingrecommendations.InacollaborativeÞlteringsystem,
similaritybetweenusersintermsofpreferencescanbereplacedbyatrust-metric,where
usersÕdegreeoftrustofotherusersiselicitedorestimatedinordertomakerecommendations.
LeeandSee(2004)arguethatanimportantgoalofresearchonautomationistoÞnd
waystodesignforcalibrationbetweenusersÕtrustinasystemanditsreliability.Ifasystem
ishighlyreliable,thanusersshouldtrustandfollowitsrecommendationsandbydoingso
willmakeoptimaldecisions.Ifasystemisonlymoderatelyreliable,usersÕshouldbeless
trustingandcarefullyscrutinizeitsrecommendations.Asadesignobjective,IDAdesigners
shouldÞndwaystohelpusersproperlycalibratetheirtrustinanIDAinordertooptimize
decisionmaking.
2.4.2AutomationBias
Humanfactorsengineeringresearchershaveexaminedwhetherautomateddecisionaidssuch
asalertsystemsandscreeningtoolshelpusersmakebetterdecisions.SkitkaandMosier
(Skitka,Mosier,&Burdick,1999)haveobservedconsiderableevidenceof
automationbias
.21Automationbiasoccurswhenadecisionmakerfailstoseekoutevidencethatcontradicts
recommendationsprovidedbyadecisionaid,leadingthemtofollowpoorrecommendations
andmakepoordecisions(Manzey,Reichenbach,&Onnasch,2012).SkitkaandMosier
(Skitkaetal.,1999)conductedanexperimentonaßightsimulatortaskwhereanautomated
alertsystemwas94%accurate.Theyfoundthatwhenthesystemgaveanincorrectalert,
subjectsfollowedtherecommendation65%ofthetime.Theyalsofoundthatsubjectsfailed
totakeappropriateactionwhentheywereusinganautomatedalertsystemanditincorrectly
failedtoalertthem,ascomparedtoacontrolgroupwhichdidnothaveanautomatedalert
system.SimilarÞndingshavebeenreportedusingotheraviationtasks(Mosier,Skitka,
Heers,&Burdick,1998),luggagescreening(Madhavan&Phillips,2010),processcontrol
(Manzey,Reichenbach,&Onnasch,2008),andmammography(Alberdi,Povyakalo,Strigini,
&Ayton,2004).
Automationbiashasbecomeaseriousprobleminclinicaldecisionmaking.Goddard
etal.(2012)conductedametaanalysisofstudiesonautomationbiasinclinicaldecision
making.Theirreviewincludedautomationbiasinducedbothbyintelligentdecisionaidsas
wellasothertypesofautomatedsystemssuchasalerts.Theyfoundthatwhenadecision
aidgaveincorrectrecommendations,itincreasedtheriskofapoorclinicaldecisionby26%.
Coieraetal.(2006)havearguedthatautomationbiashasbeenoneoftheprimaryculprits
forwhyintelligentsystemshavefailedtomeetexpectationsforimprovingcareinclinical
settings.TheyarguethatwhileintelligentsystemsprovidemanybeneÞts,automationbias
andotherrelatedhumanfactorsproblemsaresimultaneouslycreatedbytheintroductionof
intelligentsystemsthatmayo
"
setthegains.
Carr(2014)providesanumberofexamplesofautomationbiasasacauseofdevastating
decisionmakingincontextssuchastransportation,Þnance,andlaw.Hearguesthatan
over-relianceonintelligentsystemsthatautomateknowledgeandintellectualworkmaydull
theanalyticalskillsrequiredforgooddecisionmakingandcreativeproblemsolving.Partof
this,heargues,isthatsuchsystemmaydeprivepeopleofthemoreenjoyableaspectsoftheir
22workandmakethemindi
"
erenttothedecisionmakingprocessbyremovinganemotional
attachmenttoit.
Fortunately,thereisevidencethatautomationbiasisnotaninevitableconsequence
ofusingintelligentsystems,andthatautomationbiascanbeminimizedthroughsystem
design.Dzindoletetal.(2003)foundthatautomationbiaswasreducedwhenthesystem
explaineditsreasoningprocess.Thetransparencyofasystemmaymakeiteasierforusers
toscrutinizerecommendationsandappropriatelycalibratetheirrelianceandagreementwith
them.Minimizingtheprominenceofaninformationdisplay(Berner,Maisiak,Heudebert,&
YoungJr,2003)canalsohelpusersavoidbecomingtooreliantonanautomateddecisionaid.
Byavoidinginformationoverload,usershavemorecognitiveresourcesavailabletoprocess
informationontheirownandscrutinizerecommendationsmadebyanaid.
2.5Transparency
OneofthemostimportanttopicsinIDAresearchanddesignis
transparency
.Transparency
referstothedegreetowhichusersunderstandwhythesystemgaveitsrecommendations,
orunderstandthesystemlogicforhowtheyweregenerated.Transparencyisoftenaccom-
plishedinIDAbygivingusersexplanationsforrecommendations.Forexample,e-commerce
siteslikeAmazonoftenexplainrecommendationsbystatingÒPeoplewhopurchasedthis
itemalsopurchasedtheseitems.Ó
TransparencyisimportantforIDAsforseveralreasons.Herlocker,Konstan,andReidl
(2000)havearguedthatmostIDAareÒblackboxesÓandthattransparencyisimportant
sothatuserscanhandleerrorsinrecommendations.Sincerecommendersystemsarerarely
perfect,itishelpfulforuserstounderstandwhyarecommendationwasgivensotheycan
determineforthemselveswhetherarecommendationcontainserror.Thisapproachallo-
catesthefunctionofrecognizingerrorstotheuserbutplacesthedemandofarticulatingor
communicatinghowrecommendationsaregeneratedtothesystem.TintarevandMastho
"23(2012)havesuggestedthatprovidingexplanationsinrecommendersystemsservestoestab-
lishusertrustinthesystemandtomaketherecommendationsmorepersuasive.Alackof
transparencyinintelligentsystemsforhomeautomationhasbeenfoundtobeoneofthe
primaryfrustrationsofusersandanimpedimenttomorewidespreadadoption(R.Yang&
Newman,2013).
Insection2.1above,IoutlinedasetoftechnicalapproachesfortheinnerlogicofIDAs.
Thereisconsiderablevariabilityacrossthesetechnicalapproachessuchthattwodi
"
erent
systemsdesignedtoassistthesametypeofdecisionmayreasonablyuseoneofmanydi
"erent
technicaldesigns.Thissuggeststhatfromtheperspectiveofusers,itisunreasonablefor
designerstoexpectevenpeopleexperiencedwithusingIDAstoknoworhaveanexpectation
thatasystemworksinanyparticularway.Usersshouldnotbeexpectedtothinkthatthe
blackboxthatgeneratesrecommendationsinonesystemisthesameastheblackboxof
anothersystem.Therefore,usersÕbeliefsabouttheinnerlogicmustprimarilycomefrom
whattheyaretoldwithinthesysteminterface,documentation,ortrainingaboutthesystem.
MuchoftheworkontransparencyinIDAsorrelatedsystemshasfocusedonevaluating
di"
erentwaystodesignexplanationsintotheinterfaceofanIDA(Tintarev&Mastho
"
,2011).Limetal.(2009)havecomparedexplanationsthatexplainrecommendationsin
termsofÒwhy,ÓÒwhynot,ÓÒhowto,ÓorÒwhatif.ÓTheyfoundthatforenhancinguser
understandingofthesystem,ÒwhyÓandÒwhynotÓexplanationsworkedbestandthat
otherformsofexplanationsdidnothelpusersbetterunderstandasystemthanproviding
noexplanations.OtherworkinthisareahasexploredusingvisualizationstoexplainIDA
recommendations(Verbert,Parra,Brusilovsky,&Duval,2013;Knijnenburg,Bostandjiev,
OÕDonovan,&Kobsa,2012),providingdetailedtutorials(Kulesza,Stumpf,Burnett,&
Kwan,2012),usingandexplainingsimpleformulasastherecommendationlogic(Aksoy,
Bloom,Lurie,&Cooil,2006),describingthetradeo
"sbetweenitemsasawaytojustify
recommendations(Wang&Benbasat,2007),andhierarchicalexplanationsthatallowusers
toÒdrilldownÓfurtherandfurtherdownadecisiontreeuntiltheyaresatisÞedwiththe
24explanation(Kay&Kummerfeld,2012).
ManystudiesofIDAsÞndthatprovidingtransparencyleadstousersatisfactionwith
thesystemandmakesthemshowgreatertrustinrecommendations(Herlockeretal.,2000;
Tintarev&Mastho
",2012,2008;Sinha&Swearingen,2002;Crameretal.,2008).These
studiesshowthatusersprefertransparentIDAinterfacesandreportthattheyareusefulin
thedecisionmakingprocess.However,amorepertinentquestioniswhethertransparency
actuallyhelpsusersmakebetterdecisions.CummingsarguesthattheÒinabilityofthe
humantounderstandcomplexalgorithmsonlyexacerbatesthetendencytowardsautomation
biasÓ(Cummings,2004,p.3).WhenusersdonÕtunderstandhowasystemproducesits
recommendations,theymayÞndithardertoidentifywhenthesystemhasmadeanerror
andthereforemaymakepoordecisionswhenthesystemprovidespoorrecommendations.
Someresearchhasshownthoughthattransparencyalonemaynothelpdecisionmaking,
andinsomecasemayactuallyleadtodecisionerrors.Ehrlichetal.(2011)foundthatina
DSSforhelpingITadminsdetectsecurityattacks,providingexplanationsledsomeusersto
haveunwarrantedconÞdenceintherecommendationsandsubsequentlymakepoordecisions
whenthesystemmadepoorrecommendations.TintarevandMastho
"
(2012)foundthat
inrecommendersystemsformoviesandforcameras,userswhoweregivenexplanations
forrecommendationswereactuallymorelikelytochangetheirmindabouttheirdecision
lateron,suggestingthattheexplanationspromptedthemtomakeadecisionthattheylater
regretted.Theseresultshavesomeimportantimplications.Onereasonthatusersmay
makepoordecisionswhenasystemistransparentisthattheymaybeconvincedbythe
explanationfortherecommendationmorethantherecommendationitself.Inotherwords,
usersmaydevelopapreferenceforhowrecommendationsaregeneratedthatisthebasis
fortheirdecisionsmorethananindependentevaluationofrecommendationquality.This
wouldsuggestthattransparency,incontrasttoCummingÕsargumentmentionedabove,can
increaseautomationbiasifuserpreferenceforsystemlogicsupersedesacriticalanalysisof
therecommendationsbytheuser.
252.6Customization
AnemergingdesignapproachtoIDAsistoa
"
ordend-usercustomizationofthesystem.
Thismeansthatusersaregivensomecontroloversomeaspectsofthesystem.Usersmay
beabletocustomizethevisuallayoutorotheraestheticdetailsoftheinterface,thedata
thatarestoredorusedinthesystem,oreventhealgorithmorreasoningprocesscanbe
controlledorinßuencedbytheuser.
Customizationisawidespreadapproachtodesigningalltypesofinteractivesystems,
includingIDAs.CustomizationinIDAhasoftenbeenconceivedasallowinguserscontrolover
thedatathattheyinput,andÞndingswaystoelicitthebestdata(e.g.itemratingsoruser
proÞleinformation)fromusers(McNee,Lam,Konstan,&Riedl,2003).However,recently
manyexperimentalIDAhaveexploredcustomizationofanIDAÕsalgorithm(Bostandjiev,
OÕDonovan,&H¬ollerer,2012;Han,He,Jiang,&Yue,2013;Schafer,Konstan,&Riedl,2004;
Bostandjiev,OÕDonovan,&H¬ollerer,2013;Parra,2013).Ashortcomingofthepublished
researchonthesekindsofsystemsisthattheyhavenotbeenevaluatedintermsoftheir
e
"
ectondecisionmaking.Rather,user-centeredevaluationsofusabilityorusersatisfaction
havebeenfavoredinmostcases.Thus,itisnotclearfromtheexistingworkoncustomizable
IDAhowtheya
"
ectuserdecisionmaking.
Inthissection,Iwilldiscusstheliteratureoncustomizationininteractivesystemsand
whatisknownaboutmakinghuman-machinesystemscustomizablefromadesignanduser-
experienceperspective.IwillthendiscusstheliteraturespeciÞcallyaboutcustomizableIDA
andmakeanargumentbasedonboththeoryandempiricalÞndingsthatcustomizationcan
bebeneÞcialtoIDA-supporteddecisionmaking.However,Iwillalsopointtoevidencethat
suggestsnewhumanfactorsdesignproblemsthatmayariseasaresultofcustomizableIDA.
262.6.1CustomizationofInteractiveSystems
Muchoftheexistingworkoncustomizationhaslookedatcustomizationwithininteractive
mediasuchaswebportals,newsfeedreaders,andonlinecommunitieswithuser-generated
content.Sundar(Sundar,2008;Sundar,Oh,Bellur,Jia,&Kim,2012)hasarguedthat
interactivemediasuchasthesethata
"
orduserstheabilitytocustomizesomeaspectof
theinterfacenurtureasenseofagencyinusers,andthatthissenseofagencyisapowerful
predictorofusersatisfactionwhenusinginteractivemedia.Sundardescribesthesenseof
agencyexperiencedbyusersofcustomizablemediaasafeelingofÒself-as-source,Ómeaning
thatusersfeelrewardedbysimultaneouslybeingbothaconsumerandaproducerofthe
medium.
Onepurposeforcustomizationininteractivemediais
personalization
.Personalization
meansthatcontentprovidedthroughthemediumispersonallyrelevantorengagingtothe
user.Personalizationmaybeachievedbyasystemtryingtoestimatewhatcontentauserwill
preferorÞndmostrelevantandtheninsertingthatcontentintothemedium.Blom(2000)
describesacontinuumbetween
personalization
andcustomization,where
personalization
issystem-initiatedtailoringofcontentwhereascustomizationisuser-initiatedtailoringofcon-
tent.Personalizedsystemsgatherandprocessinformationautonomouslyaboutusersand
performtailoringofcontentoftenwithoutthedirectionorevenawarenessofusers.Cus-
tomizationprovidesaninterfaceforuserstoexplicitlyinßuencethistailoring.Forexample,
targetedadvertisingthattriestoshowusersthemostrelevantadsbasedoninformationit
collectsthroughcookiesaboutusersbrowsingwouldbeanexampleofpersonalization,as
thetailoringofthecontentisentirelysystem-initiated.However,asystemthatgivesusers
promptsandasksquestionstoexplicitlyallowuserstotailortheiradswouldbeacustomiz-
ablesystem.Thisdistinctionhighlightsadi
"erenceintheprocessofhowcontentbecomes
tailoredtoauser,butnodistinctionintheoutcomeoftheprocessinwhichthecontentis
highlyrelevantandengagingtoaparticularuser.
27Animportantquestionthatfollowsfromthisdistinctioniswhetherornotdi
"
erences
intheprocessofprovidingtailoredcontent,user-controlledvs.systemcontrolled,hasany
impactonusersÕexperienceswithinteractivemedia.Sundaretal.(2010)haveexploredthis
question,seekingtodeterminewhethercustomizationispurelyaformofuser-controlled
personalization,orwhetherithasothere
"
ectsonusers.Theycomparedacustomizable
versionofanRSSnewsreaderwithonethatwaspersonalizedtoproviderelevantcontent
usingacollaborativeÞlteringtechnique.TheyfoundthatÒpowerusersÓwereabletogenerate
contentthattheymostpreferredwhenusingthecustomizableinterfaceratherthanthe
personalizedinterface.However,non-poweruserspreferredthecontentinthepersonalized
interface.Theauthorsperformedafollow-upstudytodeterminewhetheraperceptionof
controlovertheinterfacecanexplainthepreferencethatpowerusershadforthecontent,
butthisstudyfoundthatperceivedcontroldidnotexplaintheresults.Thisresultsuggests
thatthebeneÞtofcustomizationisthatitallowsskilleduserstomaximizeasystemÕsoutput
tomatchtheirpreferencesbetterthanasystem-maintainedalgorithmmight.Whenusers
aregivencontrolofamediumandtheyhavetheskilltouseitproperly,theycantailorits
outputtomaximizeitsquality.
Otherwork,however,hasfoundthatthesenseofagencya
"ordedbycustomization
providesadditionalbeneÞtstousersatisfactionbeyondsimplyenablingasystemtoprovide
high-qualitycontenttobeconsumed.MaratheandSundar(2011)arguethatcustomization
enablesaninterfacetobepredictabletotheuser.Theyalsoarguethatcustomizationfosters
notonlyasenseofcontroloragency,butalsoasenseofidentityforuserswithinthemedium.
Customizationofamediumallowsuserstoexpresswhotheyarethroughthemedium,and
thisa
"
ordanceofself-expressionisbeneÞcialtousersatisfactionanduserexperiencewith
aninteractivesystem.Inanempiricaltestofthishypothesis,theyfoundthatusinga
customizablewebportalleduserstobothhaveagreatersenseofcontrolandtofeelthat
thesystemwasamorepreciseexpressionoftheiridentity.
Thethemeresultingfromthisworkoncustomizationininteractivemediaisthatcus-
28tomizationenablesskilleduserstotailorcontentoroutputtomatchtheirpreferences,and
thatthisa
"
ordancealsoleadstoagenerallymoresatisfactoryuserexperiencebyfostering
asenseofcontrol,identity,andpredictabilitywhenusinginteractivesystems.
2.6.2CustomizationinIDAs
Iscustomizationagooddesignforintelligentdecisionaids?Howdoesitimpactthedecision
outcomes,decisionprocesses,anduserexperienceofIDA?Inthissection,Iwilldetaila
theoreticalargumentforthebeneÞtsofcustomizationinIDAs.Iwillthenexamineempirical
evaluationsofcustomizableIDAanddiscusshowtheysupportthisargument.Iwillalso
reviewsomeresearchthatdemonstratespotentiallynegativeconsequencesofcustomization
inIDAs.
2.6.2.1TheoreticalBasisforCustomizableIDAs
Inanyhuman-machinesystem,thereissomedivisionoflaborbetweentheuserandthe
system,whereusersfulÞllsomefunctionsorsub-tasksnecessarytotheprimarytaskand
thesystemperformsotherfunctions.DesignersofIDAsandotherhuman-machinesystems
mustdeterminewhichfunctionsaretobeperformedbythemachineandwhichfunctions
shouldbelefttousers,andmakingthisdeterminationcanbeadi
!
cultdesignchallenge.In
theÞeldofhumanfactorsengineering,thisdivisionoflaboristhe
functionallocation
,and
ithasbeenanimportantareaofresearchforover60years(deWinter&Dodou,2014).
Astechnologieshaveadvancedincapabilities,thisquestionhasgenerallybecomeeven
morecomplicatedbecausethereisincreasinglyoverlapbetweenthefunctionsthatcanfeasi-
blybeperformedbyeitherhumansormachines.Manytheoriesandframeworkshavebeen
developedtoassistdesignersinallocatingthefunctionsofasystem.Themostprominent
oftheseframeworksisFittsÕlist(Fitts,1951).Thislist,sometimesknownastheMABA-
MABAlist(ÒMenAreBetterAt-MachinesAreBetterAtÓ)makesdeclarationsabouttypes
29offunctionsthateachentityislikelytoexcelatincomparisontotheotherentity.As
examples,thislistsuggeststhathumansarebetteratperceivingpatterns,improvisingor
usingßexibleprocedures,andreasoninginductively.Itsuggeststhatmachinesarebetterat
followingprocedureswithprecisionandperformingmultiplefunctionssimultaneously.
Sinceitsoriginalpublication,FittslisthasfrequentlybeenadaptedandmodiÞedtoapply
tonewtypesofmachinesandtechnologies.However,itsbasicpremiseofcreatingheuristics
thatdistinguishtherelativestrengthsofhumansandmachineshasbeenwellpreservedin
engineeringanddesign(deWinter&Dodou,2014).
IDAdesignersfacemanyfunctionallocationdecisions.Functionssuchasinputingand
validatingdata,conductinganalysisorÞlteringoutputmayconceivablybecompletedby
eithertheuserorthesystem.Forexample,designersofanIDAmayallocatetotheuser
functionofacquiringorcuratingrelevantinformation,andallocatetothemachinethe
functionofcalculatingprobabilitiesorretrievingadocumentfromadatabase.Orthe
designersmayallocatethefunctionofinputingdatatothesystemthroughanautomated
webcrawler,andthefunctionofanalyzingtherelevanceofthedatatotheuserorusers.
Designersmayneedtobalancefactorsofutility,reliability,andusabilityinmakingfunction
allocationchoices.
Parasuraman,Sheridan,andWickens(2000)haveproposedaclassiÞcationframework
todetermineanIDAÕs
levelofautomation
.Thelevelofautomationcanthoughtofas
theproportionoffunctionsthathavebeenallocatedtothesystemratherthentheuser.
Thisframeworkproposesascalebetweencompletelyautonomoussystemsthatdecideand
executealldecisionsandcompletelyunassisteddecisionmaking.Thisframeworkisuseful
fordescribingsystemswherefunctionsareuniquelyallocatedtoeithertheuserorthesystem
inaÒdivide-and-conquerÓdesign.
Figure
??illustratesParasuramanetal.Õsframework,includingsomeexamplesofhypo-
theticalsystemdesignsandwheretheyÞtintotheframework.Parasuramanetal.explain
thattheÞrsttwostages,informationacquisitionandinformationanalysis,canbethoughtof
30Information 
Acquisition
Information 
Analysis
Action 
Selection
Action 
Implementation
Decision  
Stages
Low 
Automation
High 
Automation
Moderate 
Automation
All data input 
 from sensors or  
existing databases
Some data input 
from sensors or 
existing databases. 
Other data input or 
verified by user
All data input by 
user directly
System performs 
calculations or 
makes predictions and 
presents 
analyses to user
User specifies 
parameters or chooses  
analysis technique, then 
system 
performs analysis
System recommends a few 
options, 
user then makes final 
choice
System chooses an 
action, 
allows user to veto
System chooses a final 
action, 
informs user
System chooses a final 
action, 
does not inform user
System implements 
action, 
notifies user of status
User implements action, 
system monitors 
progress
Figure2.1:SomehypotheticalsystemdesignsandtheirplacementwithinParasuramanet
al.Õsofdecisionmakingstagesandlevelsofautomation.
astheÒinputÓtothesystemwhereasthelatterstagesareitsoutput.CustomizableIDAs,as
discussedinthisdissertation,referstousercustomizationofsysteminputs,andthereforeÞts
intotheÞrsttwostagesofdecisionmaking.Byallowinguserstoinßuenceinputs,customiz-
ableIDAswouldbestbedescribedashavingamoderatetolowlevelofautomationinthese
inputstages.Completemanualcontrol,thelowestlevelofautomationintheframework,
wouldfalloutsidethedeÞnitionofanintelligentdecisionaidhowever.
Researchonautomationattheinputstageshasgenerallyfoundthatahigherlevelof
automationleadstoimprovedtaskperformanceaslongastheautomationperformsreliably.
Butwhentheautomationfails,thisleadstoworseperformancethaniftheinputstages
hadbeenallocatedprimarilytotheuser(Onnasch,Wickens,Li,&Manzey,2013;Schuster,
Jentsch,Fincannon,&Ososky,2013).Onnaschetal.(2013)attributedthistoalossof
situationalawareness
.Thismeansthatusersbecomelessawareofalltheconditionsand
31informationthatshouldimpacttheirdecisionandthismakesitharderforthemtorecognize
afailureofthesystem.
AcustomizabledecisionaidmaybeabletotakeadvantageoftheperformancebeneÞts
ofautomationwhileavoidingthelossofsituationalawarenessthatleadstopoordecision
making.BymakingasystempersonalizedtothespeciÞcdecisioncontext,customizationmay
makethesystemmorereliable,i.e.providebetterrecommendations.Buttheinvolvement
requiredoftheusertocustomizethesystemmaydemandthatusersmaintainsituational
awarenessandmaybebetterabletorecognizewhenthesystemhasfailedorprovidedpoor
recommendations.Inotherwords,customizationmaybeabletocreateamiddleground
boththebeneÞtsofmanualcontrolandautomationarerealized.
Insupportofthise
"orttoÞndanidealcompromisebetweenallocatinginputfunctionsto
theuserortothesystem,therehasbeenanincreasingcallamongautomationresearchersto
exploreamorecollaborativestyleofhuman-machineinteraction(Cummings&Bruni,2009)
thantheoneprescribedbytraditionfunctionallocationtheories,includingParasuraman
etal.Õsframework.Inhuman-machinecollaboration,functionsarenotnecessarilydivided
betweenuserandsystemandthelevelofautomationisnotnecessarilyameasureofÒhow
muchÓisdonebyeitherentity.Instead,theinteractionisdesignedtoallowfunctionsand
rolestobesharedtovaryingdegreesandcommunicationencouragedandsimpliÞedtoallow
theuserandsystemtocollaborativelyarriveatanoptimalsolution.
CummingsandBruni(Cummings&Bruni,2009)havebuiltontheParasuramanetal.Õs
levelsofautomationframeworktomakeitmoreinclusiveoftheconceptofhuman-machine
collaboration.Theyidentifythreeimportantrolesinthedecisionprocess,andeachofthese
rolescanbeallocatedeitherinfullorinparttoeitherthehumanuserorthemachine.The
generator
producesasetofdecisionalternativesorrecommendations.The
decider
makesa
Þnalchoicefromamongthegeneratedalternatives.Andthe
moderator
keepsthingsmoving
forwardtowardsmakingandexecutingaÞnaldecision.
CustomizableIDAscreateamixedfunctionallocationforthe
generator
role.Thesys-
32temÕsrecommendationsareinßuencedbothbythesystemÕsinternal(automated)logicas
wellasbytheusersinput.CummingsandBruiniÕsempiricallyevaluatedtheirframework
andexplicitlytestedvaryingdegreesofautomationwithinthe
generator
role.Theyfound
thatdecisionmakingoutcomessu
"
eredinadesignwherethe
generator
rolewasmostly
performedbythemachine,incomparisontotwootherdesignswherethisrolewaseither
mostlyperformedbytheuserorevenlysplitbetweenuserandmachine.Thisexperiment
testedasystemthatwasprimarilyasearchengine,andwecannotsaywhetheritsresult
wouldextendtoIDAs.However,itdoessupportthehypothesisthatcustomizationcould
haveapositivee
"
ectondecisionmakingifimplementedasasystemdesign.
Functionallocationandlevelsofautomationaretheoriesthathavebeendevelopedto
explainhumanbehaviorandperformanceasassistedbyabroadspectrumofassistivetech-
nologies.IDAsareincludedinthetechnologiesbuttheresearchusedtobuildandvalidate
thesetheoriesofteninvolvessystemsthatfalloutsideofthedeÞnitionofanintelligentde-
cisionaid.Forthisreason,thereisnotyetconvincingevidenceinthefunctionallocation
literaturethatcustomizableIDAswillimprovedecisionmaking,eventhoughito
"
erstheo-
reticalsupportforthishypothesisiftheÞndingsfromthebroaderclassofdecisionaidsare
maintainedwithinthesubsetofintelligentaids.
2.6.2.2CustomizableIDAResearch
AdditionalevidencesupportingtheuseofcustomizationinIDAscanbefoundwithinthe
IDAliterature.
AutomationtechnologiessuchasIDAshavebeencriticizedforbeingtoorigid(Norman,
1990b).Rigidsystemsdonotaccountwellenoughforthevariabilityinthecontextsinwhich
theywillbeused,variabilityinthepeoplewhowillusethem,andthisrigidityleadstovarious
errorseitherbythesystemorbyusers.McDonaldandAckerman(2000)haveappliedthis
criticismspeciÞcallytorecommendersystemsthatuseacollaborativeÞlteringapproach
whereitemsarerecommendedbyaggregatinggroupswithsimilaroverallpreferences.They
33arguethatvaryingcontextsoftendemanddi
"erentapproachestomakingrecommendations,
butmostsystemsaredesignedtomaximizetheaccuracyofasingleapproach.
OnewaytocreateßexibilityinanIDAistolettheusercustomizethesystemtomeet
theneedsofthespeciÞcdecisioncontext.Inotherwords,allowingcustomizationenablesa
systemtoreliablybetailoredtoaspeciÞcdecisioncontext.MuchresearchinHCIhassought
todesignandevaluatenewwaystoallowuserstocontrolorcustomizetheoutputofanIDA.
Oneapproachtoa
"ordinguserssomecontrolisforthesystemtoacceptfeedbackfromusers
abouttherecommendations.ChenandPu(2012)describevariouswaystoforasystemto
acceptfeedbackfromusersaboutrecommendations,includingstructuredinterfacesforusers
tocritiquespeciÞcaspectsoftherecommendationsetaswellasnaturallanguagedesigns
thatletusersexpressqualitativefeedbackthatcanbeinterpretedbythesystemandused
tobuildauserproÞle.Thiscritique-basedapproachtofeedbackisdesignedtoa
"
ordsome
controltousersbylettingthemdialoguewiththesystemtohelpitbuildamoreaccurate
user-proÞle.
CustomizationmayalsohelpuserÒbuyinÓtorecommendationsmorethanpersonalized
systemsthatgiveuserslittlecontrol.LeeandLee(2009)showedthatIDAsine-commerce
thatrecommenditemstobuyusingpersonalizationtechniquescancreateasenseofpsy-
chologicalreactanceinwhichusersfeeltheirfreedomtochooseisbeingrestricted.Intheir
study,thesensethattheIDAwasrestrictingfreechoiceleduserstorejectusingthesystem
altogether.Customizationmayprovideasolutiontothisproblem.Bygivinguserschoice
intherecommendation-producingprocess,customizableIDAsmayavoidthispsychological
reactancewillstillgettingpersonalizedrecommendations.
CustomizingIDAshasgenerallybeenshowntoincreaseuserengagementandsatisfaction
withthesystemanditsrecommendations(Hijikata,Kai,&Nishida,2012;Knijnenburgetal.,
2012;Parra,2013;Burkolter,Weyers,Kluge,&Luther,2014).However,despitetheseclear
beneÞtsofcustomizationasadesignchoice,muchlessresearchhaslookedatcustomization
anditsimpactondecisionmaking.ThisisanimportantgapintheIDAresearchbecausefor
34manyIDAs,decision-centeredcriteriaaremoreimportantthanuser-centeredcriteria.For
example,ifdoctorsenjoyusingasystemandcometorelyonitbecausethecustomizationis
engaging,butitleadsthemtomakeworsedecisions,thanthesystemasawholeisharmful
toitstruepurposeofhelpingtoprovidebettercareforpatients.Likewise,ane-commerce
recommendersystemthatuserslovetousebutleadsthemtobuyproductsthatdonÕtactually
matchtheirpreferencesmaybecounter-productive,asuserswilllikelyreturnproductsor
stoppatronizingthesite.
Onereasonthatcustomizationmayhelpusersmakebetterdecisionsisbya
"
ordinga
Òwhat-ifÓstyleofanalysiswhereuserscancanrepeatedlytryoutdi
"
erentconÞgurationsor
inputstothesystemandevaluatetheoutput.What-ifanalysisisacommontechniquein
businessintelligenceasawaytoobtaindecisionsupport(Golfarelli,Rizzi,&Proli,2006).
What-ifanalysisallowsuserstesthowvariationsinimportantparametersinadecision
contextwilla
"
ecttheoutcomeofthedecision(Golfarellietal.,2006).CustomizableIDAs
canenablethisbyallowingthesechangesinparameterstobespeciÞedaspartoftheinput
toasystem,anduserscanthenusetherecommendationsthatareproducedasanestimation
ofthee
"
ectthesechangeswillhave.What-ifanalysiscanbee
"
ectiveatimprovingdecision
makingwhencertainconditionsfavorit(Kottemann,Boyer-Wright,Kincaid,&Davis,2009).
Therefore,asameansofimprovingdecisionmaking,customizationmaybehelpfultodecision
makersbya
"
ordingwhat-ifanalysis.
Bostandjievatal.(2012)havearguedthatacustomizableIDAthatallowsusersto
adjustweightsandotheralgorithmfeaturescanhelpuserslearnabouthowthealgorithm
work.Theyarguethatwhat-ifanalysisisa
"
ordedbythistypeofcustomizationandthatit
canmakethesystemmoretransparent.
Usersofcomputingtechnologieshaveatendencytobefocusedandengagedwiththeim-
mediateinterfacethattheyareusing,andgenerallydonottreatinteractionwithacomputer
asproxyforinteractionwiththepeoplewhodesignedandbuiltthecomputer(Sundar&
Nass,2000;Solomon&Wash,2014).ItisunnaturalforuserstotrytoÒgetinsidetheheadsÓ
35ofsystemdesignerstoreasonabouthowthesystemmightwork.Rather,usersgenerallyare
orientedtowardstheimmediateinterfacewithwhichtheyareinteractingasthesourceof
theinteraction,andarenotorientedtowardsothersourceslikethesystemdesigners.
Customizationofsystemlogicprovidesameansintheinterfaceforuserstonaturally
thinkabouthowasystem
shouldworkbasedontheirownunderstandingofthedecision
problemandtheirexpertiseinitsdomain.Byrequiringuserstothinklike
asystemdesigner
withoutrequiringthemtothinklike
thesystemdesigners,customizationmayhelpusersnot
onlyimprovethesystemÕsrecommendationbutalsoimprovetheirsituationalawareness.
InlinewiththenotionofÒself-as-sourceÓ,thereisevidencethatwhenanIDAworksor
thinksinafashionthatissimilartousers,theywilllikeusingthesystemandÞndituseful.
Aksoyetal.(2006)foundincontrolledexperimentsthatsimilaritybetweenarecommender
systemÕslogicandusersÕdecisionmakingstrategyhelpedusersmakebetterdecisionsand
prefertousetherecommendersystem.Thisresearchsuggeststhatithelpsuserswhenthe
systemÒthinksÓliketheydo.OnewaytoensurethatanIDAthinkslikeitsuseristo
allowuserstoconÞguretomatchtheirpreferredmethodofgeneratingrecommendations.
However,thisstudyhasanimportantlimitationthatnecessitatesfurtherresearch.The
recommendersystemusedwasquitesimpleasitsimplyrankeditemsastheweightedsumof
fourattributescores.Theformulawasdescribedtousersclearly,meaningthatthesystem
hadanexceptionallyhighdegreeoftransparency.Suchahighdegreeoftransparencyat
thistimenotrealisticformanymorecomplexmethodsofgeneratingrecommendations.
Therefore,itisnotclearwhetherthissimilaritye
"
ectwillholdtruewhenuserscanonly
infersimilarityorhaveonlypartialinformationaboutthesimilaritybetweentheirwayof
thinkingandthesystemÕs.Andmoreimportantly,ifuserscanonlychangeapartofthe
algorithmorsystemlogicthroughcustomization,howissimilaritybetweenuserlogicand
systemlogicperceived?
InasimilarstudyAl-Natouretal.(2008)foundthatusersexpressedmorepositive
opinionsaboutarecommendersystemthatusedlogicthatmatchedtheusersÕdecision
36makingstyle.Inthestudy,subjectsmadeadecisionaboutalaptoppurchaseandthen
completedasurveythatmeasuredtheirdecisionmakingstyle.Theythenusedadecision
aidtogetarecommendationforthedecisiontheyhadjustmade,andthissystemexplained
theprocessitusedtoproducetherecommendation.Insomecases,thesystemusedasimilar
decisionprocesstotheonethesubjecthadused,andothertimesitusedadi
"
erentprocess.
Subjectsthenreportedtheirattitudesandintentionstousethedecisionaid,withsubjects
whohadseenasystemthatsharedtheirpersonaldecisionmakingprocessreportingmore
positiveattitudesandintentionstousethesystemforfuturedecisions.
ThesestudiesdemonstratethepotentialforcustomizationinIDAs.Whensystemspro-
videexplanationsoftheirlogicortakeothermeasurestoestablishtransparency,userreac-
tionstothesystemmaylargelydependonwhetherthelogicusedispreferredbytheuser.
Customizationprovidesameanstonotonlyenhancetransparency,buttoalsoensurethat
thelogicispreferredbyusers.EstablishingthisassertionwithinacomplexIDAandevalu-
atingtheactualdecisionmakingthatensuesusingtheIDAhasnotbeenaddressedinthe
literature.
2.6.3PotentialProblemswithCustomizationinIDAs
Inmypreviouswork,IshowedthatwhenuserscustomizeanIDA,theybecomebiased
towardsacceptingitsrecommendations(Solomon,2014).Inthisstudy,subjectsplayeda
fantasybaseballgame,andusedanIDAthatrecommendedlikelyoutcomesforbaseball
gamestohelpthemmakepredictions.Someusersweregiventheopportunitytocustomize
thesystembychoosingsomeattributesofthegamethattheywantedthesystemtoempha-
sizewhenproducingarecommendation.Inthestudy,customizingthealgorithmdidnot
actuallya
"
ecttherecommendations,althoughusersbelievedthatitdid.Theseuserswho
customizedtheIDAweremorelikelytoacceptbothgoodandpoorrecommendationsthan
asetofuserswhowerenotallowedtocustomizethesystem.Thisdemonstratesanewform
ofdecisionmakingbiasthatIhavecalled
customizationbias
.Whenrecommendationsare
37highquality,havingcustomizedasystemmayleadtoimproveddecisionmaking,regardless
oftheiractualinßuenceontherecommendations.However,ifthesystemgivesapoorrec-
ommendation,usersmaynotscrutinizetherecommendationswellenoughandmakepoor
decisions.
Customizationbiasisaformofautomationbiasinthatcustomizationleadsuserstorely
toomuchontheIDAÕsrecommendationsandbecomepoorjudgesofrecommendationquality.
Onereasonforthismaybethatcustomizationcreatesan
illusionofcontrol
inIDAusers.
Langer(1975)hasdemonstratedanillusionofcontrolinhumandecisionmakingwhereby
peoplebecomeoverconÞdentthatrandomeventswillhavepositiveoutcomesbecausethey
havemadesometypeofchoiceassociatedwiththeoutcome.Forexample,Langershowed
thatpeoplewerewillingtobetmoreinacardgamewhentheywereabletoblindlychoose
acardfromthedeckratherthanbedealtthetopcard.Choosingacardfromashu
#
ed
deckdoesnota
"
ecttheprobabilityofagettingagoodcard,butpeoplebelievedtheycould
controlthequalityofcardtheydrew.
However,someotherrecentworkonillusorycontrolhasprovidedadi
"
erentinterpreta-
tionofthisÞnding.Ginoetal.(2011)haveshownthatÞndingsfromillusorycontrolstudies
arenotlikelytheresultofauniversaltendencyforpeopletooverestimatetheircontrol,but
ratherofatendencytohavepoorperceptionoftheiractualcontroloveroutcomes.Intheir
studies,theyvariedthedegreeofactualcontroloveranoutcomebetweennocontrol(acom-
pletelyrandomoutcome)tocompletecontrolwheretheoutcomewasdeterminedentirely
byasubjectÕschoice.Somesubjectsweregivenmoremoderateamountsofactualcontrol,
wheretheirchoicesa
"ectedoutcomeswithsomeprobability.Theyfoundthatwhenactual
controlwaslow,peopletendedtooverestimatetheircontroljustasinillusorycontrolstud-
ies.However,whenactualcontrolwashigh,peopletendedtounderestimatetheircontrol.
Perceptionofcontrolwasonlymildlycorrelatedwithactualcontrol.Thisworksuggests
thatpeopleareperhapspoorjudgesoftheircontrol,ratherthaninherentlybiasedtowards
overestimation.
38Afewstudieshaveexploredwhethertheillusionofcontrolismanifestinusingdecision
supportsystems.KottemanandDavis(1994)foundthatusersofaspreadsheet-basedÞnan-
cialforecastingsystemweremoreconÞdentintheirdecisionswhenthesystemallowedthem
tomakeadjustmentstoitsinputsandvalues.However,theirconÞdencewasnotwarranted
astheyperformedatthesamelevelintheirdecisionmakingasotherswhohadusedalocked-
downversionofthesystemthatcouldnotmakechanges.Kahaietal.(1998)replicated
bothoftheseÞndingsinascenariowhereuserscustomizedadecisionaidbyhelpingtobuild
thestatisticalmodelitusedtogenerateforecasts.
AnotherpotentialproblemwithusingcustomizableinputinIDAsisthatitmayenable
conÞrmationbias.ConÞrmationbiasisaninformationseekingbehaviorinwhichpeopleseek
orinterpretevidencethatispartialtoexistingbeliefsorhypotheses(Nickerson,1998).Con-
ÞrmationbiashasbeenobservedinIDAswhereuserscontrolthesystemÕsinput(Woolley,
2007;Berneretal.,2003;Solomon,2014;MessierJr,Kachelmeier,&Jensen,2001).
RelatedtoconÞrmationbiasisanotherproblemwithcustomization.E
"ectivelycus-
tomizinganIDAmayrequireafairamountofexpertisebothinusingthesystemandin
thedecisiondomain.Berneretal.(2003)studiedaclinicalIDAthatalloweduserstocus-
tomizequeriesforrecommendationsthatthoseuserswhohadtheexpertisetocreategood
customqueriesgenerallyalreadyknewthebestdecision,andmerelyconÞrmeditwiththe
system.ThosewithoutexpertisealsotendedtoconÞrmtheirinitialhypothesesusingthe
system,buttheseinitialhypotheseswereoftenincorrectandthereforepoordecisionswere
made.ThisÞndingsuggestsapossibleparadoxforcustomization.Thepeoplewhohavethe
expertisetouseittogenerategoodrecommendationsmaynotactuallyneedtheIDA,while
thosewhocouldmostbeneÞtfromitmaynotbeabletouseite
"
ectivelytoproducegood
recommendations.
392.6.4SummaryofCustomization
DesigningIDAstobecustomizablebyend-usershasanumberofimportanttheoretical
advantages.Theseinclude:
¥AllowinguserstopersonalizethesystemÕsinnerlogictobestmeettheirspeciÞccir-
cumstances,potentiallyleadingthesystemtomakebetterrecommendations.
¥Addingtransparencytothesystemsothatusershaveanadequateunderstandingof
howitworksandhowitsrecommendationshavebeenproduced.
¥Givingusersanopportunitytothinkcriticallyabouttheirdecisionandbuildsitua-
tionalawareness.
¥HelpsusersbuildtrustintheIDAsothattheywillrelyonitwhenitgivesgood
recommendations.
¥AllowsuserstoconductÒwhat-ifÓanalysisthatcanhelpthemlearnbothaboutthe
systemaswellasthedatathatinformit,potentiallyleadingtoinsightaboutthe
decision.
¥Situatesthesystematamoderatedegreeofcontrolforactivitiesofinformationac-
quisitionandinformationanalysis,reducingthepotentialforautomationbiasand
complacency.
However,customizationmayalsopresentseveralnewchallengesthatneedtobeinves-
tigatedandconsideredintheoriesthatinformIDAdesign.Someofthesenewchallenges
are:¥Creatingabiaswherebyusersarepronetoagreementwithsystemrecommendations
evenwhentherecommendationsarepoor.
40¥Creatingamisplacedsenseofcontrol,whereusersincorrectlyjudgethee
"ectoftheir
inputonthesystemÕsoutput,leadingtopoordecisionmaking.
¥EnablingconÞrmationbiasbylettinguserstailorinformationacquisitionandanalysis
sothatitconÞrmspriorbeliefs
¥Usersmaybeabletocustomizeasystemsothatitmatchestheirsubjectiveprefer-
enceforhowasystemshouldworkorhowdecisionsshouldbemade,whichmaynot
objectivelybethebestapproachforagivensituation.
¥CustomizinganIDAe
"
ectivelymayrequireexpertisethatuserswhocanmostbeneÞt
fromanIDAmaynotpossess.
2.7Summary
IntelligentDecisionAidscanbepowerfulinhelpingtheirusersacquireandanalyzeinforma-
tionandselectactionsaspartofadecisionmakingprocess.Avarietyofpowerfultechnologies
havebeendevelopedthatcanautomatetheprocessingandanalysisoflarge-scaledataand
usethatprocessingtomakerecommendationstousersaboutaparticulardecision.The
applicationofthesetechnologiestodecisionmakinginmedicineandhealthcare,business
ande-commerce,andmanyotherimportantdomainssuggeststhatcomputer-assistedde-
cisionmakingisseenashavingtremendouspotentialforimprovingdecisionprocessesand
outcomes.However,todatethereissomeconcernthatthetheorizedbeneÞtsofthesetechnologies
arenotbeingrealized.Oneoftheconcernsisthatwhilethetechnologiesthatdrivethese
systemsarepowerful,thedesignoftheinteractionbetweenusersandsystemshasnotbeen
perfected.Humanfactorsproblemssuchasmiscalibratedtrust,automationbias,conÞr-
mationbias,andothercognitivebiaseshavebeenshowntolimitthee
"
ectivenessofthese
systemsinhelpingusersanddecisionmakersactuallyimprovetheirdecisionmaking.For
41thisreason,user-centeredapproachtoIDAdesignthatconsidersusersasanintegralpart
ofthesystemisnecessary.Suchauser-centeredapproachtoIDAdesigndemandsastrong
theoreticalunderstandingofhowcharacteristicsandfeaturesofanIDAdesignimpactusersÕ
decisionmaking.
OneofthemostwidelystudieddesigncharacteristicsinIDAresearchisthenotionof
transparency.Sincetheunderlyingsystemlogic,includingbothalgorithmsorstatistical
modelsaswellasdatabases,isoftentoolarge,complex,orvariedforuserstoeasilysee,
IDAsareoftenthoughtofasÒblack-boxes.ÓHowever,muchworkhassoughttodesigntrans-
parentIDAsthato
"
ersomeformofinsighttousersabouthowasystemworkstoproduce
recommendations.ManybeneÞtshavebeendemonstratedtomakingIDAstransparent.
However,itisnotyetclearwhetherimproveddecisionmakingisoneofthosebeneÞts.The
littleresearchthatexaminesthatquestionisinconclusive,withsomeworkevensuggesting
thattransparentIDAmayinhibitsomeaspectsofdecisionmaking.Forthisreason,IDA
designsthatprovidetransparencymustbecarefullyevaluatedinordertodeterminehowthe
designmaybeimpactingdecisionmaking.
Onesuchdesignfeaturethatrequiresacarefulevaluationiscustomization.Customiza-
tionmaymakeanIDAsmoretransparentbyallowinguserstoplayadualroleasboth
systemuserandsystemdesigner.AcustomizablesystemthatallowsuserstoconÞgureitin
awaythattheyexpectwillhelpitpersonalizeitsrecommendationsprovidessomeautomatic
transparencytousers.Sinceusersknowwhattheyhavedonetoasystemincustomizing,
theygainatleastsometransparencyintohowitworks.
However,customizationasadesignfeaturemayhaveotherconsequencesforsystem
usabilityandforcomputer-supporteddecisionmaking.Customizingasystemmaycreate
orenhancesomebiasesthatarisewhenusersrelyonautomatedaidstoassistindecision
making.Usersmaybecomebiasedtowardsagreeingwithrecommendations,theymaytryto
customizeasystemtoconÞrmanexistinghypothesis,theymayhavemiscalibratedtrustor
unrealisticexpectationsforhowwellthesystemwillworkasaresult.Oneofthedi
!
culties
42withcustomizationofanIDAÕsinternallogicisthatinmostcases,customizationonly
providespartialcontroltousers.Therefore,theiractionstoconÞgurethesystemmust
interactwithanumberofotherpotentiallyinvisiblefactorswithinthesystemÕslogic,and
usersmaynotbeabletoeasilyinterprethowtheyhavea
"
ectedthesystemÕsoutputbyway
oftheircustomizationofthesystem.
43CHAPTER3
CUSTOMIZATIONBIAS
3.1Introduction
Customizationisadesignapproachtocreatingpersonalizedrecommendations.Ratherthan
completelyusingartiÞcialintelligenceorcomputationaltechniquestopersonalizeandtailor
recommendationstotheuserandthespeciÞcdecisioncontext,customizableIDAsallocate
someoftheresponsibilityofpersonalizingrecommendationstotheusersthemselves.
Fromtheperspectiveofadesigner,therecanbeseveralgoalsbehindusingthisdesign.
OneimportantgoalcanbetoleverageusersÕknowledgeoftheirlocalcircumstances,pref-
erences,andsituationalawarenesstohelpthesystemproducebetterrecommendationsfor
users.Anothergoalcanbetogiveusersasenseofcontroloragencythatproducesapositive
userexperienceandattitudetowardsthesystem.
Thereisanassumptionbehindthisgoal,whichisthatwhenasystemproducesgood
recommendations,thenuserswillaccepttheserecommendationsandconsequentlymake
gooddecisions.Herlockeretal.(2004)suggestthatrecommendationqualityisoftenthe
fundamentalfocusofIDAdesignersandthatotherevaluationcriteria,particularlyhuman-
centeredcriteria,areoftenignored.Iarguethatinmostcases,IDAshouldbemostcritically
evaluatedbasedonthedecisionsthattheirusersmakeratherthano
#
inetechnicalcriteria
oruser-centeredcriteriasuchasusabilityorusersatisfaction.However,thisisatremen-
douschallengeforIDAresearchbecauseofthecomplexityofevaluatingdecisionmaking,
particularthetypesofdecisionsthatIDAarefrequentlydesignedtoassistthathavehigh
uncertainty,highstakes,timepressure,andvariabilityindecisionmakersÕexpertise(Klein,
2008).Also,formanydecisions,noteveryoneagreesonthebestcriteriaforwhichdecisions
shouldbeevaluated.IDAsmayforexamplebeusedtorecommendproductsthatare
hori-
44zontallydi
!erentiated
(Cremer&Thisse,1991),whichmeansthatagooddecisionforsome
willbeapoordecisionforothers.
Furthermore,sincedecisionqualitymaybedi
!
culttoevaluate,particularlyinrealtime,
thequalityofrecommendationsthatasystemprovidescannotbeeasilycommunicatedto
usersorevenknowntothesystempriortoprovidingrecommendations.Thispresentsachal-
lengeforIDA-supporteddecisionmaking.WhenanIDAgivesarecommendation,should
usersfollowtherecommendation?Howcanusersknowhowhowgoodarecommendationis
andwhetherornotrecommendationsaretrustworthy?UsingIDAtosupportdecisionmak-
ingaddssomecomplexitytodecisionsinthatIDAsprovidenewbutaggregatedinformation
intheformofrecommendationsthatusersmustevaluateaspartofthedecisionprocess.
AsdiscussedinChapter2,thereisanabundanceofevidencethatIDAusersoftenmake
decisionerrors.Usersoftenfollowpoorrecommendationfromasystem(Skitkaetal.,1999),
althoughusersmayalsofailtofollowgoodrecommendationswhichhurtstheirdecision
quality.Fromahumanfactorsengineeringperspective,understandingwhatleadsusers
tofollowordeterfromIDArecommendationsisacriticalaspectofunderstandinghowto
makesystemsthatimprovedecisionmaking.AsarguedbyDzindoletetal.(2003),agoalof
human-centeredIDAdesignshouldbetocalibrateusersÕrelianceonIDAsÕrecommendations
withthesystemÕsÒreliability,Ówhichistosaythequalityoftherecommendationsthey
produce.Ifuserscane
"
ectivelydetectrecommendationquality,theycanuseanIDAto
makegooddecisions.
Inthischapter,IwillpresentastudythatexamineshowIDAdesignsthata
"
ordend-
usercustomizationcanimpactuserdecisionmaking,leadingtoadecisionmakingbiascalled
customizationbias
.Customizationbiasoccurswhenusersbecomepartialtoacceptingrec-
ommendationsfromIDAsasaresultoftheirinvolvementincustomizingitsalgorithm.I
Þrstobservedthisbiasinmypreviouswork(Solomon,2014)oncustomizableIDAs.This
studywillbuildonthatworkbytestingatheoreticalmechanismbywhichcustomization
biasisenabled.
45EfÞcacy Beliefs
Agreement
Customization
Figure3.1:HowcustomizationcancreateagreementwithIDArecommendations.
IDAsthatarecustomizablea
"
orduserstheopportunitytotailorthesystemÕsinner
logictomatchtheirpreferencesforhowrecommendationsareproduced.Theintentofthis
tailoringistoallowtheusertocreateanalgorithmthatworksinawaythattheuserbelieves
ise
!cacious.Inotherwords,customizationallowsuserstomakethesystemworkinaway
thattheybelievewillbesuccessfulatproducinggoodrecommendations.Iwillreferto
usersÕbeliefsaboutthequalityofrecommendationsthatusersexpectasystemtoproduce
ase"
cacybeliefs.
Onepossiblereasonforcustomizationbiasisthatbyallowinguserstotailorthealgo-
rithm,theydevelopinßatedbeliefsaboutthesystemÕse
!cacy.Usersmaybelievethattheir
actionsincustomizingtheIDAwilluniformlyimprovethesystem,ignoringthepossibility
thattheyhaveharmedthealgorithmÕsperformanceorhadlittlee
"
ect.
TheconceptofusersÕe
!
cacybeliefsisanimportantconstructbothforunderstanding
customizationbiasaswellasmoregenerallyforIDAdesign.Howdousersformexpectations
46orbeliefsaboutthee
!
cacyofasystem?ParticularlywhenanIDAlackstransparencyor
whentheuserhaslittleexperiencewithasystem,usersmayhavelittleinformationthat
allowsthemtoadequatelyassesshowwellasystemislikelytoworkatproducinggood
recommendations.Butthisbeliefmaynonethelessimpacttheirdecision.Ifusershavelittle
knowledgeaboutadecisionoraboutasystem,theymayhavelittletofallbackonwhen
makingdecisionsotherthanabelief.
Figure3.1illustratesthetheoreticalrelationshipbetweencustomization,e
!
cacybeliefs,
andagreement.ThisÞgurerepresentsthreerelationships.First,itsuggeststhatcustomiza-
tioncausesincreasede
!
cacybeliefs.Second,itsuggeststhate
!
cacybeliefscauseagree-
ment.Together,thesetworelationshipsareanargumentfor
mediation
,inthatthee
"
ectof
customizationonagreementismediatedbyusersÕe
!
cacybeliefs.ThisÞgurealsosuggestsa
thirdrelationship,whichisadirectrelationshipbetweencustomizationandagreementthat
isnotrelatedtoe
!cacybeliefs.
Thisdiagramsuggestsacausalchainfromcustomizationtoagreement.Customization
causesuserstoincreasetheirbeliefsinsysteme
!
cacy,andbecausetheyhaveincreased
theire
!
cacybeliefstheywillthenbeinclinedtoagreewiththeIDAÕsrecommendation.It
alsosuggeststhatthedirectinßuenceofcustomizationonagreementiscausalinnature.
Inthischapter,Iwillpresentanexperimentthatteststhecausalrelationshipsbetween
customizationande
!
cacybeliefs,aswellasthedirectrelationshipbetweencustomization
andagreement.Thisexperimentwillalsolookattherelationshipbetweene
!
cacybeliefs
andagreement,althoughitwillonlyo
"erweakevidencethattherelationshipiscausal.
However,thecausalityofe
!
cacybeliefsonagreementwillbetestedinasubsequentstudy
reportedinchapter4.
473.2Methods
ToevaluatetheroleofcustomizationindecisionmakingwithIDA,Icreatedanexperiment
whereIDAusersaregivenrecommendationspurportedlygeneratedbyacomplexalgorithm.
SomeusershadthechancetocustomizetheIDAtoinßuenceitsrecommendations,butin
realitythecustomizationhadnoe
"
ectontherecommendations.Thisdesigntestswhether
theactofcustomizinganIDAinßuencesdecisionswhileholdingthequalityofthoserec-
ommendationsconstant.Thisallowsforacomparisonofdecisionmakingbetweenusers
whocustomizeanIDAandthosewhodonotbutwhoreceiveidenticalrecommendations.
Byholdingrecommendationqualityconstantbetweenconditions,thisdesignevaluatesthe
e"ectcustomizationhasdirectlyonthedecisionsthataremadebyusers,ratherthanany
e
"
ectthatisduetothechangeinrecommendationsthatcomesfromcustomizingtheIDA.
Thedecisionintheexperimentwasafantasybaseballpredictiongameinwhichsubjects
triedtopredictthescoresofMajorLeagueBaseballgamesafterbeingshownstatisticsabout
theteamsinvolved.Thistaskhasseveraladvantages.First,itisataskwithalowthreshold
forexpertise,sincemanypeopleinthegeneralpopulationfollowbaseballandplaysimilar
games.Second,itisataskforwhichIDA-liketoolsarefrequentlyusedtohelppeoplemake
choices.Third,itisataskthatrequiresadecisiontobemadewithouttheavailabilityofall
possiblyrelevantinformation.Eventhebeststatisticalsimulatorsarenotperfectlyreliable
inpredictinggameoutcomes,andthereforesomejudgementorextraknowledgefromthe
decisionmakerisrequired.Fourth,itisataskthatcaninvolvebothabinary,Óyes/noÓ
typedecision(ÓWhichteamwillwin?Ó),aswellasacontinuousoutcome(ÓHowmanyruns
willeachteamscore?Ó),andeachoftheseoutcomescanbeobjectivelycomparedtoactual
outcomes.249subjectswererecruitedfromAmazonMechanicalTurktoplaythisgameaspartofa
studytoÒhelpimproveanalgorithmictoolforaidingdecisionsinfantasybaseball.ÓInorder
tocompletetheexperiment,subjectshadtoÞrsttakeatimedtestonthebasicrulesand
48statisticsofbaseball.ThisquizisdescribedingreaterdetailinAppendixB.73ofsubjects
didnotsuccessfullycompletethisquiz.Thesesubjectswerepaid$0.40andscreenedout
offurtherparticipation.Thisbasicknowledgewasequivalenttotheminimumknowledge
requiredtoplayfantasybaseball.Inordertoenrollinthestudy,subjectswererequired
throughMechanicalTurkÕssystemtobeintheUnitedStatesandtohavecompletedatleast
95%oftheirpreviousassignmentsonMechanicalTurk.MechanicalTurkworkerswhohad
participatedinpilotsofthisstudyorinmypreviousstudyusingthistask(Solomon,2014)
werealsonoteligibletoenroll.Subjectswerepaid$2forparticipation.Subjectswerealso
promisedanadditionalpaymentthatwoulddependontheirperformanceinthegame,and
weretoldthattheaverageexpectedpaymentwouldbe$2.25.Subjectstookanaverageof
14.4minutestocompletetheexperiment.SixsubjectswereremovedfromtheÞnaldataset
becausetheycompletedthestudyinlessthanÞveminutes.InpilottestingIdeterminedthat
Þveminuteswasnotsu
!cienttobeabletocompletethestudywhilegivinganythoughtto
thedecisions.Twoadditionalsubjectswereremovedbecausetheyhadnomatchingsubject
fromthenon-customizationcondition(thisisexplainedindetailinsection3.2.3).
TheÞnaldatasetcontained168subjects.Thesubjectpoolwas73%malewithanaverage
ageof33yearsold.
3.2.1GamePlay
SubjectsÞrstreadinstructionsandwererequiredtopassadi
!
cultquizontheinstructions.
Onaverage,subjectsrequired2.1attemptstopassthisquiz.Thedi
!cultyandtimerequired
topassthequizensuredthatthedistributedonlinesamplehadsinceremotivetoparticipate
andthattheyadequatelyunderstoodthegame.Subjectsplayed12roundsofthefantasy
baseballpredictiongame.Inthisgame,subjectswereshownextensivestatisticsabouttwo
teamsandaskedtomakeapredictionaboutthescoreofthegamebetweenthetwoteams.
Toensurethatonlytheavailablestatisticalinformationwasusedtoinformdecisions,the
namesoftheteamswerenotrevealedtosubjects.Additionally,thebaseballgamesthat
49subjectswerepredictingweregamesthathadalreadybeenplayed.Subjectsweretoldthat
eventhoughthegameswerepastgames,allstatisticsandalgorithmsinthestudytreated
thegamesasiftheywereinthefuture.
Iselectedgamesfortheexperimentfromthe2011and2012MajorLeagueBaseball
seasonsusingseveralcriteria.IÞtanexistingstatisticalmodel(T.Y.Yang&Swartz,2004)
forassessingtheprobabilityofahomevictorytogamesfromtheseseasons.Thismodel
estimatestheprobabilitythatahometeamwillwinusingtherelativestrengthofeachteam
inthreecategories:winningpercentage,theEarnedRunAverageofthestartingpitcher,and
BattingAverage.ThemodelalsoincludesanadjustmentforhomeÞeldadvantage.Alldata
aboutMajorLeagueBaseballgamesandplayerswastakenfromBaseball-Reference.com
1.Iselectedgamestomatchtheapproximatedistributionofprobabilitiesestimatedbythe
model.Ichosetwogameswherethepredictedwinnerhadlessthana60%chanceofwinning,
fourgameswherethepredictedwinnerhada60-70%chanceofwinning,fourgameswere
selectedwithprobabilitybetween70and80%,andtwogameswithprobabilitygreaterthan
80%.Onlygameswheretheteamwiththehigherexpectedprobabilityactuallywonthe
gamewereincludedintheÞnalsetof12games.These12gameswerepresentedinarandom
ordertoeachsubject.
Subjectsearnedpointsinthegamebymakingaccuratepredictionsabouttheoutcome
ofthegame.Subjectsearn20pointsiftheypredicttheexactlycorrectscore.Ifthey
choosethewrongwinner,theylose10pointsfromtheirscore.Theyalsoloseonepointfor
theabsolutedi
"
erencebetweenthepredictednumberofrunsforeachteamandtheactual
numberofruns.Forexample,iftheÞnalscoreofagamewasAway5ÑHome3andthe
subjectpredictedAway4ÑHome6,thesubjectwouldlose10pointsforchoosingthe
wrongwinner,lose1pointformissingtheAwayruntotalby1,and3pointsformissingthe
Homeruntotalby3,leavingatotalof6pointsforthegame.Thisbaseballtask,likemany
decisions,hasaclearÒbestpossibleoutcomeÓyetnoclearÒworst-possibleoutcomeÓsince
1http://www.Baseball-Reference.com
50Figure3.2:CustomizableIDA.
onecantheoreticallypredictscoresthatdeviateinÞnitelyfromtheactualscore.Similarly,
doctorscouldprescribethewrongmedicineandalsoprescribeadosageanddurationthat
deviateinÞnitelyfromtheoptionthatintruthwouldbemostbeneÞcialtoapatient.The
presenceofclearÒbestoptionsÓmeansthatanydeviationisalosstothedecisionmaker.
Forthisreason,thescoringsystemidentiÞedaclearbestoptionandtheIDAwascapable
ofrecommendingthisbestoption,anddeviationfromthisoptimaldecisionwasrepresented
asaloss.
513.2.2IDAandConditions
AllsubjectsusedanIDAthatprovidedextensivestatisticalinformationabouttheteams
involvedineachofthegames.Inadditiontoprovidingstatisticalinformation,theIDA
alsorecommendeditsownpredictionaboutthescoreofthegame.Subjectsweretold
thispredictionwasbasedonastatisticalalgorithm.However,therecommendationswere
actuallypre-determinedforeachgame.Thereweretwotypesofrecommendations.Good
recommendationssuggestedtheactualscoreofthegame,yielding20pointsiffollowed
exactly.Poorrecommendationssuggestedthewrongwinner,aswellasascorethatwould
yield5points.Subjectsweregivenpoorrecommendationsforfourgames(onerandomgame
foreachdi
!
cultylevel),andgoodrecommendationsfortheremaininggames.Overthe12
games,theaveragescoreoftheIDAÕsrecommendationswas15.Subjectsweretoldofthis
average,andthattherewasconsiderablevariationinthequalityoftherecommendation.
Thereweretwoconditionsoftheexperiment.Inthecustomizablecondition,subjects
hadtheopportunitytomakeadjustmentstotheIDAÕsrecommendationalgorithmafter
seeingatableofstatisticalcomparisonsbetweentheteams(seeFigure3.2).Subjectswere
askedtochoosebetweenoneandÞvestatisticalcategoriestoreceiveextraemphasisin
thesimulationalgorithm.Forexample,asubjectcouldselectwinningpercentage,home
runs,andstartingpitcherERAandthealgorithmwouldthenemphasizethecontribution
ofthesestatisticswhenestimatingthegameÕsoutcome.Theinstructionsstatedthatgood
customizationimprovestheperformanceofthealgorithm,butpoorcustomizationcould
harmperformance.
Inthenon-customizationcondition,subjectssawthesametableofstatisticalcomparisons
asthecustomizationcondition.Theinterfacehadpre-loadedasetofcategoriesthatwould
beemphasized,andthebuttonsusedtoconÞgurethesystemweredisabledsothatchanges
couldnotbemadetotheconÞguration.Intheinstructions,subjectswereinformedthat
thepre-loadedcategorieswereconÞgurationsthathadbeenusedbyprevioususersofthe
52systemwhowerecompletingthesametask.DetailsofhowtheconÞgurationwasselected
forusersinthenon-customizableconditionaredescribedbelowinthesection3.2.3.
Afterviewingthestatistics(andifinthecustomizationconditioncustomizingtheIDA),
subjectsclickedabuttontogeneratearecommendationabouttheoutcomeofthegame
thattheycouldusetohelpthemmakeadecision.Priortoseeingtherecommendations,
subjectsansweredathreequestionsurveytoassesstheirbeliefabouthowwelltheyexpected
thesystemtoperform.Thissurveyisdescribedinthemeasuressectionbelow.Subjects
werethenshowntheIDAÕsrecommendationandgiventheopportunitytomaketheirown
predictionaboutthegameoutcome.Oncesubjectshadsubmittedtheirdecision,theywere
directedtothenextrounduntilall12roundswerecompleted.Subjectsthentookapost-
testquestionnaire,andweregivenacodetoreturntoMechanicalTurktobesubmitted
forpayment.Aftertheentirestudywascompleted,subjectsweresentamessagewitha
breakdownoftheirscoreforeachgameandadebrieÞngstatementaboutthetruenature
oftheIDA.Subjectswereshownalltheirscoresafterthestudy,ratherthanimmediately
followingeachround,toreducethenuisancefactoroflearningabouttheIDAorthedecision
scenariooverthecourseofthestudy.
3.2.3SubjectMatching
Apotentialconfoundcanariseinanystudythatcomparesacustomizablesystemtoanon-
customizablesystem.Whensubjectscustomizeasystem,theyhavesetitsconÞguration.
Subjectswhodonotcustomizeasystemmustneverthelessuseasystemthathasbeencon-
Þguredinsomeway.IftheconÞgurationforthesystemusedbythenon-customizersis
pre-determined,aswasdoneinmypreviouswork(Solomon,2014),thereisaconfoundin
thatthecustomizationconditiondi
"
ersbothinthe
actofcustomizingaswellasthe
product
ofcustomizing(i.e.theconÞgurationthatisused).Di
"
erencesbetweentheconditionsmay
bebecauseoftheactofcustomizing.ButtheyalsocouldbebecauseofspeciÞcconÞgura-
tionsthatareused.Forexample,ifnon-customizerssimplydonotthinkthatthedefault
53conÞgurationisagoodconÞguration,theymaybelesspronetoagreementnotbecausethey
havenÕtchosentheconÞgurationbutbecausetheybelieveitistrulyapoorconÞguration.
Thepurposeofthisstudyistounderstandhowthe
actofcustomizinganIDAinßuences
decisionmaking,andisnotconcernedwith
product
ofcustomizing.Thisisbecausetheact
ofcustomizingisgeneralizabletoothersystemsandotherdecisionscenariosbesidestheone
usedinthisstudy.TheproductofcustomizingisspeciÞctothetaskandIDAusedinthe
studybutwillnothaveapplicabilitytoothersystems(e.g.knowingwhichcategoriesof
baseballstatisticspeopleselectwonÕttellusanythingaboutwhatparametersaninvestor
shouldusetocustomizeastockrecommender).
Forthisreason,thisstudywasdesignedtoremovethisconfoundandensurethatonlythe
actofcustomizingwasvariedbetweenconditions.Todothis,thestudywasrunonecondi-
tionatatime.ThecustomizationconditionwasrunÞrst,followedbythenon-customization
condition.TheconÞgurationsthatsubjectsinthecustomizationconditionusedwereas-
signedtosubjectsinthenon-customizationcondition.Thisensuredthatsubjectsinthe
non-customizationcontrolconditionwereusingthesameconÞgurationsasthoseinthecus-
tomizationcondition.
RatherthanrandomlyassigningconÞgurationstonon-customizers,acollaborativeÞl-
teringtechniquewasusedtomatchnon-customizerstothemostsimilarsubjectfromthe
customizationconditionintermsoftheirattitudesaboutthespeciÞcbaseballstatisticalcat-
egories.Thiswasdonetoreducedi
"
erencesinsubjectiveopinionsaboutwhichcategories
aremoste
"ectiveforacomputerizedsimulatorinestimatinggameoutcomes.Forexample,
somepeoplemayfeelthatHomeRunsandEarnedRunAveragearethemostinformative
statisticalcategories,whereasothersmaythinkOn-BasePercentageandStolenBasesare
bettercategories.Ifweassumethatcustomizerswillchoosetheirpreferredcategoriesmost
ofthetime(andthisassumptionissupportedbythedatashowninFigure
??),thensubjects
whocustomizemayhavedi
"
erentoutcomesinthestudyasaresultofthispreferencefor
theconÞgurationratherthantheactofcustomizing.ThecollaborativeÞlteringtechnique
54123456789123456891012How many non
!customizers were matched to each customizer
countFigure3.3:Distributionofmatchespercustomizer.
isintendedto,asbestaspossible,assignnon-customizersconÞgurationsthattheyprefer
equivalentlytothecustomizers.
ToexecutethiscollaborativeÞlteringmatching,subjectstookapre-testsurveythat
askedthemtoassesshowinformativeeachofthe27statisticalcategoryoptionsmight
betoacomputerizedbaseballoutcomesimulator.Theyansweredaquestionabouteach
categoryona5-pointscale.Wheneachnon-customizerÞnishedthissurvey,thecosine
similaritybetweentheirresponsesandtheresponsesofeverysubjectfromthecustomization
conditionwascalculated.Then,thenon-customizerwasmatchedwiththecustomizerwith
thehighestcosinesimilarity.Thenon-customizerthendidthebaseballpredictiontaskusing
thesameorderofgames,andforeachgametheIDAÕsconÞgurationwasshownasthesame
conÞgurationthatthecustomizerhadused.
Aconsequenceofthismatchingapproachisthattwosubjectswhocustomizedthesystem
55nevergotmatchedtoanynon-customizers.ThesesubjectswereremovedfromtheÞnaldata
set,leaving49subjectsinthecustomizationgroup.Inane
"
orttoÞndmatchesforthe
majorityofcustomizers,andtoobtainmultiplenon-customizers,119additionalsubjects
wererecruitedforthenon-customizationcondition.Thematchingdistributionwasnot
uniform.9customizersonlyreceivedonematchingnon-customizer.Onecustomizerreceived
12matches,whichwasthemaximum.ThisdistributionisshowninFigure3.3.Theaverage
cosinesimilaritybetweenmatcheswas0.983,withastandarddeviationof0.014.Overall,
subjectsinthenon-customizationconditionhadverysimilarresponsestotheirmatchfrom
thecustomizationconditionandtherewasonlyasmallamountofvarianceacrosssubjectsin
theirsimilarity.Forthisreason,theunevendistributionofmatchesisnotlikelyproblematic
becausenon-customizersallhadverynearlythesamesimilaritytotheirmatchingsubject
fromthecustomizationgroup.
Alimitationofthisstudydesignisthatthematchingtechniquecreatesamildviolation
ofrandomassignmentbecausesubjectswhosignedupearlierforthestudyweremorelikely
tobeinthecustomizationgroup.Ananalysisofalldemographicandpre-treatmentvariables
didnotÞndanystatisticallysigniÞcantorevenpotentiallymeaningfuldi
"
erencesbetween
subjectsineachcondition.Therefore,Ihavenoreasontosuspectthatthisbiasimpacted
theresultsoftheexperiment.
3.2.4Measures
Agreement.
TherearetwomeasuresofagreementbetweensubjectsÕpredictionsforthe
outcomesofagameandtheIDAÕsrecommendation.
WinnerAgreement
isabinarymeasure
ofagreement.Winneragreementiscodedas1ifthesubjectpicksthesameteamtowinas
theIDAÕsrecommendation,andcodedas0iftheychoosetheoppositeteam.
ScoreAgreement
isacontinuousmeasurethatusesthegameÕsscoringmechanism.Score
agreementismeasuredusingthesamemethodassubjectsÕpointtotalsasdescribedabove,
howeverratherthanusingthetrueoutcomeofthegameasthecomparison,itusestheIDA
56recommendation.Forexample,iftheIDApredictsAway2ÑHome6,andthesubject
predictsAway4ÑHome5,scoreagreementwouldbe15sincethesubjectstartswith20
points,thenloses4forthedi
"
erenceinAwayscoresand1forthedi
"
erenceinHomescores.
Thismeasureisincludedtoprovideagranularindicationofagreementandrelianceonan
IDA.Moststudiesthatexaminerelianceondecisionaidsuseonlyabinarytasksothat
relianceoragreementismeasuredasafrequency.However,manyIDA-supporteddecisions
havemoregranularityintheoptionsusershave.Adoctor,forexample,canuseanIDAto
determineadosageordurationofatreatment.TheIDAmaygiveaspeciÞcrecommendation,
andthedoctormaybeinßuencedbytherecommendationbutmaymakeasmalladjustment
toit.Usingonlyabinarymeasureofagreementwouldnotcapturethatinßuence,and
thereforebothacontinuousandbinarymeasurearevaluabletounderstandingagreement.
BeliefsofSystemE
"
cacy.
SubjectsÕbeliefsaboutsysteme
!cacyweremeasuredusing
threeindicators.Subjectsansweredthefollowingquestionona7-pointLikertscale:
Based
onthecategoriesthatarebeingemphasized,howwelldoyouexpectthesimulatortoperform
atmakingitsprediction?
Asecondmeasuredaskedsubjectstoestimatethenumberof
pointsthattheIDAÕspredictionwouldearnifitwerescored.Thethirdmeasureasked
subjectstoassesstheprobabilitythatthesimulatorÕspredictionwouldchoosethecorrect
teamtowinthegame.
Table3.1:Factoranalysisofe
!
cacybeliefsmeasures.
StatisticFactorLoading(Std.Error)MeanSt.Dev.MinMax
7-pointLikert1.000(Ð)4.7731.29117
ExpectedPoints2.272(0.095)12.9313.714020
CorrectWinner9.240(0.370)72.85912.93450100
FitIndices
ComparativeFitIndex1.000
RootMeanSquareError0.000
AconÞrmatoryfactoranalysiswasusedtoevaluatethereliabilitybetweenthesemeasures
57fromthedata.Table3.1describesthisfactoranalysis.Overall,thethreemeasureswere
consistentwitheachotherandloadedontoalatentvariablewithastrongÞt.Intheanalyses
presentedintheresultssection,thefactorscoreforeachobservationthatresultedfromthis
analysiswasusedasasinglevariablecalled
e"
cacybeliefs
.Theanalyseswerelaterrepeated
usingeachoftheindividualindicatorsseparately,andthismadenodi
"
erencetoanyofthe
conclusionsthatweredrawnfromtheanalyses.
Propensitytotrustautomation.
Sincesubjectsmayhaveindividualdi
"
erencesintheir
attitudesaboutautomationordecisionaids,ImeasuredsubjectsÕpropensitytotrustauto-
mateddecisionaidsusingthescaledevelopedandvalidatedbyMerrittetal.(2012).This
scalehas6itemsthatarelistedinAppendixA.CronbachÕsalphafortheseitemswas0.72.
Toobtainafactorscoreforeachsubjectasameasurementofpropensitytotrustautomated
decisionaids,IconductedaconÞrmatoryfactoranalysisthatincludedallsixitemsontoa
singlefactor.ThisanalysisisdescribedinTable3.2.Thefactorscoresforeachsubjectwere
calculatedandusedasasinglevariablecalled
automationtrustpropensity
inthesubsequent
analyses.Theseitemswereadministeredinthepre-test,beforesubjectssawtheIDAorhad
anyotherinformationaboutit.
Table3.2:Factoranalysisofpropensitytotrustautomateddecisionaidsitems.
StatisticFactorLoading(Std.Error)MeanSt.Dev.MinMax
Trust11.00(Ð)3.4940.92215
Trust2(Reverse)-0.733(0.088)2.5890.99315
Trust30.941(0.071)3.1900.92815
Trust41.125(0.066)3.1370.96615
Trust51.126(0.073)3.3271.02415
Trust60.947(0.084)3.1071.03315
FitIndices
ComparativeFitIndex0.977
RootMeanSquareError0.104
Categoryratings.
ForasingleconÞgurationoftheIDA,whichcouldemphasizebetween
58oneandÞvecategories,subjectsÕoverallratingoftheconÞgurationwascalculatedasthe
averageratinggiventoallthecategorieswithinaconÞguration.ThisisameasureofsubjectsÕ
beliefsaboutwhethertheemphasizedcategoriesaregoodindicatorsofgameoutcomes.
Control,helpfulness,accuracy,andimportance.
Inthepost-testquestionnaire,subjects
answeredquestionsandrespondedon7-pointLikertscalesabouthowtheyusedtheIDAfor
theirdecisions.Thesequestionsandtheirassociateconstructswere:
¥Control
-Iwasabletocontroltheaccuracyofthesimulator.
¥Helpful
-Howhelpfulwasthesimulator?
¥Accurate
-Howaccuratedoyouthinkthesimulatorwasatpredictingtheoutcomeof
games?¥Important
-HowimportantwerethesimulatorÕspredictionsininformingyourpredic-
tions?
Decisionquality.
Thenumberofpointsearnedfromadecisionwasusedtomeasure
decisionquality.
3.3Hypotheses
H1.SubjectswhocustomizetheIDAwillbelievethatthesystemhashighere
"
cacythan
thosewhodonotcustomizethesystem.
Theillusionofcontrol(Langer,1975)wouldpredict
thatuserswhomakechoicesabouthowtheIDAsalgorithmworkswillbelieveitworksbetter
thanotheruserswhouseanidenticalalgorithmbutdonotmakethechoicetouseitinthe
IDA.Importantly,thisillusionisstrictlytheresultofmakingthechoiceabouthowtheIDA
isconÞgured,andnottheresultofusersconÞguringitusingcategoriesthattheyfeelare
betterpredictorsofgameoutcomes.
59H2.SubjectsÕbeliefsabouttheIDAÕse
"
cacywillbepredictiveoftheiragreementwith
itsrecommendations.Subjectswillhavemoreagreementwhentheybelievethesystemhas
greatere
"
cacy.
IexpectuserswhobelievetheIDAÕsprocessforproducingrecommenda-
tionsise
"
ectivetoevaluaterecommendationstobemoreaccurateandtrustworthy,and
thereforetheseuserswillagreewithrecommendationsmorethanthosewhofeeltheprocess
isine
"
ective.
H3.SubjectswhocustomizetheIDAwillhavegreateragreementwithitsrecommendations
thanthosewhodonotcustomize.
Iexpecttheresultsofthisstudytobeconsistentwith
mypreviouswork(Solomon,2014)whichfoundthatusersaremorelikelytoagreewith
recommendationswhentheybelievetheyhavecustomizedtheIDA.Thedesignofthisstudy
allowsthishypothesistobebrokenintotwopieces:
¥H3a.Thee
!ectofcustomizationonagreementwillbepartiallymediatedbye
"
cacy
beliefs.SubjectswhocustomizewillagreemorewiththeIDAinpartbecausebycus-
tomizingthesystemtheyincreasetheirbeliefsinitse
"
cacy(H1andH2).
Ifboth
H1andH2aresupported,itwouldfollowthatcustomizationcancauseagreementby
causinganincreaseine
!
cacybeliefs,whichthencausesmoreagreement.
¥H3b.
Therewillbeadirecte
!ectofcustomizationonagreement.Subjectswhocus-
tomizewillagreemorewiththeIDAforreasonsotherthananincreasedbeliefinits
e"
cacy.
Thereareatleasttwomechanismsotherthane
!
cacybeliefsbywhichcus-
tomizationmightplausiblycausegreateragreement.OneisaconÞrmationbias.If,
forexample,auserlooksoverthestatisticsoftheteamsandformsanopinionthatthe
awayteamwillwinbecausetheyhaveabetterpitcher,shemightconÞguretheIDAto
emphasizepitchingstatistics,whichareconsistentwithherinitialopinion.Thenifthe
IDArecommendedtheawayteamtowin,shemaytreattherecommendationascon-
Þrmationbythesystemthatherinitialopinionhasmeritandbeinclinedtoagreewith
60therecommendation.Anotherpotentialmechanismisthee
"
ortrequiredtoconÞgure
theIDA.TheIKEAe
"
ect(Norton,Mochon,&Ariely,2011)illustratesthatpeople
aremoreinclinedtobuyproductswhentheyhaveparticipatedincreatedthem.If
thise
"
ectextendedtofollowingIDArecommendationsitwouldpredictgreateragree-
mentwithrecommendationsbycustomizers.E
"
ortmightalsocauseagreementby
creatingfatigueinusers,suchthatafterexertinge
"orttocustomizetheIDAusersare
fatiguedandnotinterestedinexertingmorecognitivee
"
orttoscrutinizeandevaluate
recommendationsclosely,andinsteadjustchoosetoagreebecausethatiseasierthan
formingtheirownprediction.Andtheremayalsobeotherunconsideredreasonsthat
customizationmayincreaseagreementotherthanthroughe
!
cacybeliefs.Thepur-
poseofthisstudyisnottoidentifyanymechanismsotherthane
!
cacybeliefs.Rather
itisintendedtoidentifywhethere
!cacybeliefsareamechanismforcustomizationÕs
e
"
ectonagreementandwhetherthereisaneedforfutureresearchthatexploreand
identifyothermechanisms.
H4.SubjectswhocustomizetheIDAwillmakebetterdecisions,earningthemmorepoints
inthegame,thanthosewhodonotcustomize.
Iexpectthistohappenbecausecustomizers
areinclinedtoagreewiththeIDAÕsrecommendations(H3),andsincetheIDAgivesmostly
reliablerecommendations,itwillbemoreusefultotheminmakinggooddecisionsinthe
baseballpredictiongame.
H5.SubjectswhocustomizetheIDAwillreportfeelingmorecontroloverthesystemthan
subjectswhodonotcustomize.
Previousworkoncustomizationinwebportalshasfound
thatcustomizationincreasesusersÕsenseofcontrol(Marathe&Sundar,2011),andIexpect
thatthiscanbereplicatedbyusersofanIDA.
H6.SubjectswhocustomizetheIDAwillreportthatthesystemisa)morehelpfultothem
astheymakedecisions,b)generatesmoreaccuraterecommendations,andc)isamore
61importantpartoftheirdecision-makingprocessthansubjectswhodidnotcustomize.
Cus-
tomizationhasbeenfoundtobebeneÞcialtousersÕperceptionsofthesystemanduser
experience(Hijikataetal.,2012;Knijnenburgetal.,2012).Iexpectthistoholdtruein
thisstudyandcauseusersfeelthesystemismoreaccurateandbeneÞcialtotheirdecision
process.
H7.ConsistencybetweentherecommendationsthattheIDAgivesandtheconÞgurations
usedwillleadtogreateragreementbysubjects.Whentheteamthatispredictedtowinthe
gamebytheIDAisstrongerinthecategoriesthatwereusedintheconÞguration,subjects
willbemorelikelytoagreewiththeIDA,eveniftherecommendationispoor.
Inmyprevi-
ouswork(Solomon,2014)Ifoundthatuserswhocustomizedweremorelikelytoagreewith
recommendationswhentheyappearedtobeconsistentwithhowtheyhadconÞguredthe
system.Thatstudycouldnotassesswhetherthiswouldbetrueforuserswhodonotcus-
tomize.Thisstudyhoweverenablesthisassessment,andIexpectsubjectsinbothconditions
toagreemorewhenthesystemÕsrecommendationsareconsistentwithconÞgurations.
¥H7a.Thise
!ectwillbestrongerforsubjectswhohavecustomizedtheIDAthanthose
whodidnotcustomize.
TheconÞrmationbiasdescribedunderH3bwillmakethee
"
ect
ofconsistencystrongerforuserswhocustomizebecausetheyhavetheopportunityto
conÞguretheIDAtomatchtheirinitialopinionmorepreciselythanthosewhodonot
customize.
3.4Results
3.4.1DescriptiveStatistics
Themeansandstandarddeviationsforthethreemeasuresofe
!
cacybeliefsarereported
inTable3.1.Ingeneral,subjectshadfairlypositivebeliefsthattheIDAwouldproduce
62Table3.3:Descriptivestatistics.
StatisticNMeanSt.Dev.MinMax
WinnerAgreement2,0160.7670.42301
Agreement2,01615.4465.877
!220
PointsEarned2,01614.8615.905020
Control1682.9881.67217
Helpful1685.0241.62417
Accurate1684.4971.20217
Important1684.9761.52917
Non!Customized
Customized
02004006001234512345Average Rating of Configured Categories
countDistribution of Preference for Configured Categories
Figure3.4:PreferenceforconÞguredcategories.
goodrecommendations,althoughresponsesacrossthespectrumofthethreevariableswere
observed.
Onaverage,subjectsagreedwiththeIDAÕspredictedwinner77%ofthetime,withan
averagescoreagreementof15.47(aperfectagreementscoreis20).Thesenumbersclosely
matchthereliabilityoftheIDA,whichpredictedthecorrectwinner67%ofthetimeand
scoredanaverageof15points.Thisindicatesthatsubjectshadreasonablywell-calibrated
relianceontheIDAintermsofthefrequencywithwhichtheyfolloweditsrecommendations.
However,aswillbediscussedbelow,calibrationofthefrequencyofrelianceisnotthe
63sameasdiscernmentofgoodandpoorrecommendations.Subjectsdidnotalwaysmake
gooddecisionsbyfollowinggoodrecommendationsorrejectingpoorrecommendations.On
averagesubjectsearned14.86pointsperround,althoughtherangeofpointsearnedspread
betweentheminimumandmaximumpossible(0to20points).
DistributionsoftheothermeasuresarelistedinTable3.3.Figure
??showsthedistri-
butionofthesubjectsÕcategoryratings,brokenintothetwoconditionsofthestudy.As
canbeseenfromthisÞgure,inbothconditionssubjectswereveryfavorabletowardsthe
categoriesthattheIDAusedinaround.Ononlyahandfulofoccasionsdidanyonegivean
averageratingofthecategorieslowerthantheneutralpointof3onthescale.Soitappears
thatinbothconditions,subjectsfeltthecategoriesbeingusedweregoodcategoriesthatare
informativetopredictingtheoutcomeofabaseballgame.Thissupportstheassumption
statedabovethatuserswilltendtochoosecategoriesthattheybelieveworkworkwellin
producingrecommendations.
Itisimportanttoevaluatehowwellthematchingmechanismoftheexperimentworked
atassigningconÞgurationsthatmatchedthepreferencesofsubjectsinthenon-customizable
IDAcondition.AlinearmodelwasÞttotestwhethersubjectsinthecustomizableIDA
conditionratedtheirchosencategorieshigherthanthoseinthenon-customizableIDAcon-
dition,whohadbeenassignedconÞgurationsexpectedtomatchtheirpreferences.This
modelincludedarandome
"
ectforeachsubjecttoaccountforhavingmultipleobservations
foreachsubject.ThemodelfoundastatisticallysigniÞcantdi
"erencebetweenconditions
(p<.05),withcustomizerspreferringtheircategoriesmorethannon-customizers.However,
althoughstatisticallysigniÞcant,themagnitudeofthise
"
ectwasverysmall.Thedi
"
erence
inmeansbetweenconditionswas0.17ontheÞvepointscale.Overallitappearsthatthe
matchingmechanismworkedreasonablywellatpairingsubjectswithsimilarbeliefsabout
howstatisticalcategoriesmighthelpacomputerizedsimulator.
Becausesubjectsplayedthegamerepeatedly,therewasthepossibilityforthemtolearnor
adapttheirdecisionmakingoverthecourseoftheexperiment.Figure3.5showshowsubjects
640.650.700.750.800.85123456789101112Round NumberAverage Winner Agreement
Non!Customized
Customized
Figure3.5:Averagewinneragreementbyroundnumber.
adjustedtheiragreementwiththeIDAovertheroundsonaverage.Therewasoverallaslight
trendagainstagreeingwiththeIDAastheexperimentwenton.However,thiswasnota
smoothdeclinebutrathersomewavesofincreasedanddecreasedagreement.Thispattern
wassimilarforbothconditionsoftheexperiment,suggestingthatcustomizationdidnot
leadtouserslearningabouttheIDAinanywaythatdi
"
eredfromthenon-customization
users.Becauseofthisnegativetrend,theroundnumberforeachobservationwasincluded
inallmodelsofdecisionmakingpresentedbelow.
3.4.2E
"
cacyBeliefsandAgreement
E
"
cacyBeliefs.
Totestthee
"
ectofcustomizationone
!
cacybeliefs(H1),IÞtamultilevel
linearregressionmodeltothedata.Thismodelestimatesthee
!cacybeliefsscorefor
eachroundofthegameforeachsubject.Itincludestheconditionthesubjectwasin
65Table3.4:E
!
cacybeliefsmodel.
Dependentvariable:
E!
cacyBeliefs
Intercept
!1.444"""(0.225)Customized0.040
(0.093)CategoryRatings0.367
"""(0.051)TrustPropensity0.129
""(0.055)RoundNumber
!0.020"""(0.005)RandomE
"
ectsStd.Deviation0.649
LogLikelihood
!2,434.228LogLikelihood
!223.423"""Note:
"p<0.1;""p<0.05;"""p<0.01(customizationornon-customization)aswellassubjectsÕaveragecategoryratingforthe
emphasizedcategories,theirpropensityfortrustinautomateddecisionaids,andtheround
numberascovariates.Becausetherearemultipleobservationspersubject,themultilevel
modelincludedarandome
"
ectthatvariedtheinterceptofthemodelforeachindividual
subject.Thisrandome
"ectaccountsforthelackofindependencebetweenobservationsfrom
thesamesubject.
H1wasnotconÞrmed,astherewasnotastatisticallysigniÞcantdi
"
erencebetween
customizersandnon-customizers.ThemodeldidÞndthatwhentheIDAusedcategories
thatthesubjecthadratedhighly,theyhadahigherbeliefofitse
!cacy.Themodelalso
foundthatpeoplewithahigherpropensitytotrustautomateddecisionaidsbelievedthe
IDAhadhighere
!
cacythanthosewithalowerpropensitytotrustautomateddecisionaids.
6605101520!3!2!1012Efficacy Beliefs ScoreScore AgreementGood RecommendationPoor Recommendation
Non!Customized
Customized
Effect of Customization and Efficacy Beliefs on Agreement
Figure3.6:Agreementwithrecommendedscore.
ThisÞndingindicatesthathavingcontroloveranIDAÕsinnerlogicdoesnotleadusers
toinßatedexpectationsforhowwellthesystemwillwork.Subjectswhocustomizedthe
systemlargelyconÞgureditusingcategoriesthattheybelievedwouldworkwell.Subjects
whodidnotcustomizetheIDAbutwereassignedconÞgurationsthatcloselymatchedtheir
ratingsofthecategoriesreportedthesamebeliefsinthesystemÕse
!cacyasthepeople
whohadchosenthecategories.Thissuggeststhatcustomizationdoesnotpromptusersto
irrationallybelievethatasystemwillworkbettersimplybecausetheyhaveinßuencedit.
Rather,usersbelieveasystemwillworkwellwhentheyhavesomeunderstandingofhowit
worksandtheybelieve,basedontheirdomainknowledgeofthedecision,thatitslogicis
appropriateforthedecisionathand.
Agreement.
Toevaluatetheagreementvariables,twomultilevelmodelswereÞttothe
data.Amultilevelregressionmodelestimatedscoreagreementusingtheconditionthe
67Table3.5:MultilevelmodelsofagreementwithIDArecommendations.
Dependentvariable:
ScoreAgreementWinnerAgreement
Intercept16.248
"""2.406"""(1.403)(0.691)
Customized1.014
""0.507""(0.495)(0.235)
PoorRecommendation
!3.519"""!1.633"""(0.235)(0.126)
CategoryRatings0.114
!0.037(0.320)(0.156)
E!
cacyBeliefs0.726
"""0.282"""(0.145)(0.075)
TrustPropensity0.2430.097
(0.291)(0.132)
RoundNumber
!0.063"!0.039""(0.032)(0.018)
RandomE
"ectsStd.Deviation2.5110.499
LogLikelihood
!6,209.821!950.162LogLikelihood
!2245.75"""206.27"""Pseudo
R20.232Note:
"p<0.1;""p<0.05;"""p<0.01680.250.500.751.00!3!2!1012Efficacy Beliefs ScoreProbability of Agreeing with IDA
Good RecommendationPoor Recommendation
Non!Customized
Customized
Effect of Customization and Efficacy on Agreement
Figure3.7:ProbabilityofagreeingwithIDAÕspredictedwinner.
subjectwasin,thequalityoftherecommendationreceived,thesubjectÕse
!
cacybelief,
theircategoryratingsfortheemphasizedcategories,theirpropensitytotrustdecisionaids,
andarandome
"ectforeachsubject.Toevaluatewinneragreement,amultilevellogistic
regressionmodelwasÞtusingthesameindependentvariables.Thismodelestimatesthelog
oddsofasubjectchoosingthesameteamtowinthegameastheIDA.Thesemodelsare
describedinTable3.5andvisualizedinFigure3.6andFigure3.7.
H2,therelationshipbetweene
!
cacybeliefsandagreement,istestedbythecoe
!
cients
fore
!
cacybeliefsinthesemodels.Inbothmodels,therewasastatisticallysigniÞcant
relationshipbetweensubjectse
!cacybeliefsandtheiragreementwithIDArecommenda-
tions.Whensubjectsbelievedthesystemwouldworkbetterpriortoactuallyseeingits
recommendations,subjectsagreedmorewiththerecommendations.Thise
"
ectcanbeseen
inFigure3.6andFigure3.7.Notethatforthemodelofwinneragreement,the
R2value
69listedinthetableisaPseudo
R2.ThismeasureofÞtwascalculatedusingthemethod
presentedbyTjur(2009)thatisbasedontheaccuracyofthemodelatcorrectlypredicting
theobservedvaluesofeitheragreeordisagree.Itshouldnotbeinterpretedasanindicator
oftheproportionofvarianceexplainedbythemodel.ThisPseudo
R2willbereportedfor
alllogisticregressionmodelsinthisdissertation.
H3isatestofthee
"
ectofcustomizationonagreement,andwastestedusingthesame
models.ThestatisticallysigniÞcantcoe
!cientforcustomizationindicatesthee
"
ectthatthe
typeofIDAhadonagreement.H3wassupported,assubjectswhocustomizedthesystem
overalldidagreemorewiththeIDAÕsrecommendationsthannon-customizers.
H3apredictedthatsomeofthee
"
ectofcustomizationonagreementwouldbehappenby
customizationinßuencinge
!
cacybeliefs,ande
!
cacybeliefsinturninßuencingagreement.
H3bpredictedthatcustomizationwouldalsoinßuenceagreementforotherunobservedrea-
sons.ToevaluateH3aandH3b,Iconductedamediationanalysistoestimateboththe
indirecte
"
ectofcustomizationonagreement(bywayofe
!cacybeliefs)andthedirectef-
fect.Anestimateofthee
"
ectsizeoftheindirecte
"
ectcanbeobtainedbymultiplying
thecoe
!cientsforthee
"
ectfromtreatmenttomediatorbythecoe
!cientfrommediator
tooutcome(Imai,Keele,&Tingley,2010).Inthisstudy,thismeansthecoe
!cientfor
customizationinthee
!
cacybeliefsmodel(Table3.4),andthecoe
!
cientfore
!
cacybeliefs
intheagreementmodels(Table3.5).Imaietal.(2010)developedaproceduretoperform
hypothesistestingonthisestimatebysimulatingthepotentialoutcomeforeachobservation.
Thepotentialoutcomeforanobservationrepresentsthee
!
cacybeliefsandagreementthat
wouldhavebeenobservedwerethegivensubjectassignedtotheoppositecondition.Imai
etal.showedthatthedistributionforthepotentialoutcomeofanobservationcanbetaken
fromtheobserveddataundertheassumptionthatthetreatment(i.e.customization)was
randomlyassignedandthattherearenounobservedpre-treatmentconfoundingvariables.
Thissecondassumptionisuntestable(Imaietal.,2010)andthereforethemediationanalysis
ofthisexperimentcanprovideonlyincompleteevidenceofamediatedcausalrelationship
70betweencustomizationandagreementbywayofe
!
cacybeliefs.
Imaietal.Õsmethodforhypothesistestingamediatedcausalrelationshipinvolvesa
MonteCarlosamplingofsimulateddatabasedontheobservedparameterestimatesofthe
modelstobuildaconÞdenceintervalaboutitssizeanddirection.1000simulationsofthe
datasetareperformedandineachsimulationtheMediatedE
"
ect,theDirectE
"
ect,and
theTotalE
"
ectofthetreatmentontheoutcomearemeasured.Theaveragee
"
ectsfromall
1000simulationsarecalculatedalongwitha95%conÞdenceintervaltotestthehypotheses
ofmediatedanddirecte
"ectsofcustomizationonagreement.Thetotale
"ectthatresults
fromthisanalysiscanbeinterpretedastheaveragechangeinagreementthatsubjectswould
experiencehadtheybeenassignedtotheoppositecondition.Themediatede
"
ectrepresents
theportionofthetotale
"
ectthatwouldbetheresultofachangeine
!cacybeliefs.The
directe
"
ectcanbeinterpretedastheexpectedchangeinagreementthatwouldbeobserved
thatisnotduetoachangeine
!cacybeliefs.
Table3.6:Mediationanalysisforscoreagreement.
Estimate95%CILow95%CIHighSig.
AverageMediatedE
"
ect0.026-0.1080.167
AverageDirectE
"
ect1.0310.0922.022**
TotalE
"
ect1.0580.0892.025**
Table3.7:Mediationanalysisforwinneragreement.
Estimate95%CILow95%CIHighSig.
AverageMediatedE
"
ect0.02-0.0050.009
AverageDirectE
"
ect0.0680.0080.125**
TotalE
"
ect0.0700.0090.129**
Contrarytoexpectations,therewasnotsupportforH3a.Thee
"
ectofcustomization
onagreementwasnotmediatedbysubjectsÕe
!
cacybeliefsusingeithermeasureofagree-
ment(seeTable3.6andTable3.7).ThisÞndingprovidesmoreevidencefortheconclusion
71125150175200Non!Customized
Customized
Total Points Earned
Figure3.8:Totalpointsearnedbysubjectsineachconditionoverthe12rounds.
discussedunderH1,namelythatcustomizationdoesnotcreateanirrationalorinßatedex-
pectationofsysteme
!
cacythatthenleadsuserstoagreewithitmoreoften.Andwhile
beliefsofhighsysteme
!
cacydoleadtomoreagreement,thesebeliefsarenotenhanced
purelybytheactofcustomizingthesystem,butratheraredeterminedbyusersÕpreferences
forhowasystemcanbestworktogeneraterecommendations.
ThemediationanalysisdidÞndsupportforH3binthattherewasastatisticallysigniÞcant
(p<.
05)directe
"ectofcustomizationontreatment.Subjectswhocustomizethesystem
didagreemorewithitsrecommendationsthanthosewhodidnot,butthise
"
ectisunrelated
tosubjectsÕe
!cacybeliefs(Table3.6andTable3.7).
Decisionquality.
H4predictedthatuserswhocustomizewillmakebetterdecisionsas
determinedbythenumberofpointstheyearninthegame.Figure3.8showsthedistributions
ofpointsearnedbysubjectsinthegame.At-testindicatedthattherewasnotastatistically
7203006009001200Good Rec. AgreedPoor Rec. Disagreed
Good Rec. DisagreedPoor Rec. Agreed
Type of Decision
countPoor Decision
Good DecisionNon!customized
Customized
Figure3.9:Decisionqualitybyrecommendationquality.
signiÞcantdi
"erencebetweenthecustomizationandnon-customizationgroupsintermsof
howmanypointstheyearned(
t(166)=
!0.649,p=0.517).Overall,customizationdidnot
leadtobetterdecisionmakingthananon-customizableIDA.
Thenatureofthedecisiontaskallowsfordi
"erenttypesofÒpoorÓdecisionmakingrel-
ativetotherecommendationthatwasprovided.Apoordecisioncouldhappeniftheuser
wasgivenagoodrecommendationbutfailedtofollowit,orifhewasgivenapoorrecom-
mendationandagreedwithit.Conversely,gooddecisionmakingcanbeconsideredtobe
whenauseragreeswithagoodrecommendationordisagreeswithapoorrecommendation.
Figure3.9illustratesthedi
"
erenttypesofdecisionsthatweremadeinthestudy,separated
bythecondition.Chi-squaretestswereperformedtodeterminewhethertherewereanydif-
ferencesinthetypesofdecisionsthatsubjectsmadebasedontheirexperimentcondition.In
termsofanytypeofgoodorpoordecision,therewasnotastatisticallysigniÞcantdi
"
erence
73betweenconditions.Subjectsineitherconditionwereequallylikelytomakegooddecisions.
However,somesmalldi
"
erencesbetweentheconditionswerenotedwhenexaminingthe
typesofdecisions.Subjectsinthecustomizeconditionhadaslightlyhigherproportion
(9%)ofdecisionsinwhichtheyagreedwithapoorrecommendationthannon-customizers.
Andconversely,thenon-customizershadaslightlyhigherproportion(9%)ofdecisionswhere
theydisagreedwithagoodrecommendation.Thesedi
"
erenceswerestatisticallysigniÞcant
(p<.05).Table3.8:Resultsofpost-testsurvey.
Dependentvariable:
ControlHelpfulAccurateImportant
(1)(2)(3)(4)
Intercept2.377
"""4.679"""4.426"""4.253"""(0.360)(0.442)(0.333)(0.400)
CustomizedSystem2.306
"""0.536"0.3340.583
""(0.230)(0.282)(0.213)(0.255)
Age
!0.0010.002
!0.0030.009
(0.010)(0.013)(0.010)(0.012)
Gender(Female)
!0.1690.3500.1590.744
"""(0.233)(0.286)(0.216)(0.259)
AutomationTrustPropensity0.1570.2420.219
"0.319""(0.135)(0.166)(0.125)(0.150)
R20.3900.0520.0420.129
ResidualStd.Error(df=163)1.3371.6411.2361.485
FStatistic(df=4;163)26.081
"""2.238"1.7816.047
"""Note:
"p<0.1;""p<0.05;"""p<0.01H5predictedthatuserswhocustomizethesystemwouldfeelmorecontroloverit.H6
predictedthatuserswhocustomizewouldreportthattheIDAwasmorehelpful,accurate,
andmoreimportanttotheirdecisionthanuserswhodidnotcustomize.H5andH6were
74testedusinglinearregressionmodelsthatestimatedtheresponsetothesurveyquestions
usingthecondition,demographicinformation,andautomationtrustpropensity.These
modelsarepresentedinTable3.8.H5wasconÞrmed.Theuserswhocustomizethesystem
reportedthattheyfeltmoreincontrolthanthosewhocustomizedthesystem.
H6waspartiallyconÞrmed.SubjectswhocustomizedtheIDAfoundthesystemtobe
morehelpfulandmoreimportantintheirdecisionmaking.However,theydidnotÞnd
theIDAÕsrecommendationstobemoreaccuratethanthenon-customizers.Theseresults
supporttheexistingworkoncustomization(Sundaretal.,2012;Hijikataetal.,2012)in
thatcustomizableIDAsarebetterperceivedbyusersthansystemsthatdonothavethis
a"ordance.Alsonoteworthyisthatuserswhohadahigherpropensitytotrustautomationreported
thatthesystemwasmoreaccurateandmoreimportantintheirdecisionthanthosewithlower
trustinautomation.ThisisanexpectedÞnding,butitshouldbenotedasanimportant
individualdi
"
erenceamongusers.Thewidevariabilityintrustinautomationthatwas
observedinthissubjectsamplesuggeststhattheremaybeunderlyingmentalmodelsand
attitudesaboutthee
!
cacyofIDAsthatcanimpacthowpeopleusethemtomakedecisions.
TotestthehypothesisthatgreaterconsistencybetweenconÞgurationandrecommenda-
tionwouldbeassociatedwithgreateragreement(H7)IÞtamultilevellogisticregression
modelsimilartothelogisticregressionmodelreportedinTable3.5.Thismodelincluded
therecommendationconsistencyvariabletodeterminewhethertheconsistencybetweenthe
conÞgurationandtherecommendationwouldinßuencesubjectsÕagreementwithrecommen-
dations,whilecontrollingfortheothervariablesknowntoa
"
ectagreementreportedabove.
Themodelalsotestedforaninteractionbetweentheconditionandrecommendationcon-
sistencytodeterminewhetheranye
"ectofrecommendationconsistencywasequivalentin
bothconditionsofthestudy(H7a).
ThismodelfoundthatthemoreconsistentarecommendationwaswiththeconÞguration
thathadbeenused,themorelikelyasubjectwastoagreewiththerecommendation.This
750.50.60.70.80.90.000.250.500.751.00Recommendation Consistency with Selected Categories
Probability of Agreeing with IDA
Good Rec.Poor Rec.
Non!customized
Customized
Figure3.10:GreaterconsistencybetweenrecommendationandconÞgurationledtomore
agreement.
resultsupportsH7.However,H7awasnotsupported,astheinteractione
"
ectofcustomiza-
tionandrecommendationconsistencywasnotstatisticallysigniÞcant.Itdoesnotappear
thatthistendencytoagreewithrecommendationsthatmatchedupwellwiththeconÞgu-
rationthatwasusedisanydi
"
erentwhenuserscustomizeanIDAthanwhentheydonot
customize.
3.5Discussion
Theseresultssuggestthatcustomizationbiasasfoundinmypreviouswork(Solomon,2014)
isnotentirelytheresultofthepersonalizedaspectofthesystemconÞguration.Inthis
studywherenon-customizerswereassignedconÞgurationsthatwerebelievedtomatchtheir
preferencesforhowasystemshouldwork,subjectsreportedthesamebeliefsabouttheIDAÕs
expectede
!
cacyasuserswhohadtheopportunitytoactuallyselecttheconÞguration.
However,inspiteofthenon-customizershavingthesameexpectationsforthesystemÕs
76Table3.9:Recommendation/conÞgurationconsistencyledtogreateragreement.
Dependentvariable:
WinnerAgreement
Intercept1.309
"""(0.175)Customized0.623
""(0.314)RecommendationConsistency1.408
"""(0.236)PoorRecommendation
!1.530"""(0.128)E!
cacyBeliefs0.297
"""(0.074)CustomizedXRec.Consistency
!0.277(0.469)RandomE
"
ectsStd.Deviation1.065
LogLikelihood
!930.304LogLikelihood
!2245.98"""Psuedo
R20.255Note:
"p<0.1;""p<0.05;"""p<0.01e!
cacypriortousingit,theywerelesslikelytoagreewithitsrecommendations.This
suggeststhatthereisanothercauseofcustomizationbiasthathasnotbeenmeasuredin
thisstudy.
ThisÞndinghasimportantimplicationsforIDAdesign.Firstofall,ithighlightsthe
importantconceptualdi
"
erencebetweenpersonalizationandcustomization,andshowsthat
evenifpersonalizationcanachievethesameintermediateoutcomeofÞndingsomethingthat
highlymatchesusersÕpreferences,itmaystillleadtodi
"
erentoutcomesintermsofthe
decisionsthatusersmakeinresponsetorecommendations.
77Table3.10:Summaryofresults.
Hypothesis
ResultH1SubjectswhocustomizetheIDAwillbelievethatthe
systemhashighere
!
cacythanthosewhodonotcus-
tomizethesystem.
Notsupported
H2SubjectsÕbeliefsabouttheIDAÕse
!
cacywillbepre-
dictiveoftheiragreementwithitsrecommendations.
Subjectswillhavemoreagreementwhentheybelieve
thesystemhasgreatere
!cacy.
Supported
H3SubjectswhocustomizetheIDAwillhavegreateragree-
mentwithitsrecommendationsthanthosewhodonot
customize.
Supported
H3a
Thee
"
ectofcustomizationonagreementwillbepar-
tiallymediatedbye
!cacybeliefs.
Notsupported
H3b
Therewillbeadirecte
"ectofcustomizationonagree-
ment.Subjectswhocustomizewillagreemorewiththe
IDAforreasonsotherthananincreasedbeliefinits
e!
cacy.
Supported
H4SubjectswhocustomizetheIDAwillmakebetterdeci-
sions,earningthemmorepointsinthegame,thanthose
whodonotcustomize.
Notsupported
H5SubjectswhocustomizetheIDAwillreportfeeling
morecontroloverthesystemthansubjectswhodonot
customize.
Supported
H6SubjectswhocustomizetheIDAwillreportthatthe
systemisa)morehelpfultothemastheymakede-
cisions,b)generatesmoreaccuraterecommendations,
andc)isamoreimportantpartoftheirdecision-making
processthansubjectswhodidnotcustomize.
Partiallysupported
H7Consistencybetweentherecommendationsthatthe
IDAgivesandtheconÞgurationsusedwillleadto
greateragreementbysubjects.
Supported
H7a
Thise
"ectwillbestrongerforsubjectswhohavecus-
tomizedtheIDAthanthosewhodidnotcustomize.
Notsupported
78Thisisanimportantconclusionforsystemdesignersthatrelatestothefunctionalloca-
tionproblem.Systemsthatautomatetheinformationacquisitionandinformationanalysis
stagesofdecisionmakingwillproducedi
"erentdecisionmakingoutcomesthansystemsthat
allocatemoreofthosefunctionstotheuser,eveniftheautomationperformsthetaskequiv-
alently.Similarly,systemdesignersthattrytoselectÒdefaultÓsettingsandconÞgurations
thatwillinßuenceasystemÕsoutputmustconsiderthatevenifusersweretochoosethe
sameconÞgurationsontheirown,theymayinterpretthesystemÕsoutputdi
"erentlythan
theoutputgeneratedbythedefaultprocess.
Theconsequencesofcustomizationbiasmaybeeitherpositiveornegative,andthismay
relatelargelytothereliabilityofthesystem.Inthisstudy,therewasnooverallimpactof
customizationondecisionmakingquality.However,customizerswereslightlymorelikely
tomakethepoordecisionofagreeingwithapoorrecommendation,andnon-customizers
mademoreerrorswheretheydisagreedwithagoodrecommendation.Thereliabilityofthe
systemisanimportantfactorfordesignerstoconsiderwhenevaluatingthepotentialcon-
sequencesofcustomizationbias.Ifasystemwillproducehighlyreliablerecommendations,
thencustomizationwilllikelybebeneÞcialtodecision-makingperformanceasuserswillbe
nudgedtowardsagreement.Asystemthatproducesunreliablerecommendationsthatre-
quireconsiderablediscernmentfromuserswouldÞndcustomizationbiastobeproblematic
towardsdecisionmaking.
Thisstudyalsoprovidesanevaluationofadi
"erenttypeofsystemdesignthathas
notbeenexaminedintheexistingliterature.CustomizationofanIDAÕsalgorithmisan
emergingareaofresearchinIDA,andthisstudydemonstratesapotentialproblemthat
end-usercustomizationcancreate.However,italsoshowsthatthisbiasmaybecountered
byusingacrowdsourcingorcollaborativeÞlteringapproachtotailoringtheIDAÕsalgorithm
thatremovestheactofcustomizingfromtheuser.Otherusersmakingsimilardecisions
maybeleveragedtoprovidetheinputtothesystem,leavingtheuseranddecisionmakerto
interprettheoutput.Futureresearchshouldexplorethisapproachfurthertodeterminehow
79itmaycounteractcustomizationbias.Aparticularlyimportantfocusforfutureresearch
iswhetherusersbelievecrowdsourcingcanbee
"
ectiveingeneral.Subjectsinthisstudy
reportedthattheconÞgurationsgeneratedbyotheruserswerejustase
!
caciousasthose
thathadbeecustomized,butthentheyproceededtodisagreewiththemmoreoften.UsersÕ
mentalmodelsofcrowdsourcinganditscapabilitiesmaylargelydeterminewhetherthey
willbereceptivetothisdesign.Itshouldbenotedthatsubjectsinthenon-customizable
conditionknewthattheconÞgurationshadbeenchosenbyotherusers,butdidnotknow
thattheyhadbeenmatchedtothemostsimilarotheruser.Whetherornotthisknowledge
woulda
"
ectusersÕbeliefsofthesystemÕse
!
cacyandagreementwithrecommendations
wouldmakeanimportantcontributioninfutureresearch.
ThisstudyhasanadditionalÞndingthathasimportantimplicationsforIDAdesign.
WhentheIDAÕsrecommendationsappearedtobeconsistentwiththewayithadbeen
conÞgured,subjectsweremorelikelytoagreewiththesystem.Thise
"
ectwasthesame
forcustomizersandnon-customizers.VariationintheconsistencybetweenconÞguration
andrecommendationwaslargelyafunctionofrandomchance.Somecategorieswerealways
emphasized,andoneofthetwoteamsalwayswasstrongerineachcategory,soitalways
hadtoappearthatthesystemhadbeenconsistentwiththeconÞgurationtosomedegree.
Nevertheless,whenbychancethesystemappearedconsistent,itleduserstoagreement.
ThisÞndingisevidenceofaconsistencybiaswhereIDAusersaremoreinclinedtoagree
withrecommendationswhentheyfeelthattherecommendationsareconsistentwiththe
waytheIDAisconÞgured.Thisisabiasbecauseithappenedregardlessofrecommenda-
tionqualityandbecauseallusersreceivedtheexactsamerecommendations.ThisÞnding
replicatestheÞndingfrommyearlierwork(Solomon,2014),whereIpositedthatitcould
betheresultofconÞrmationbiaswhereusersmadeaninitialdecisionintheirminds,chose
categoriesthatwouldbeconsistentwiththatinitialdecision,andthenwereassureditwas
correctafterseeingrecommendationsthatwereconsistentwiththeinitialdecisionandcon-
Þguration.TheÞndingsfromthisstudyrefutethatinterpretationandsuggestthatthis
80consistencybiasisnotanexampleofconÞrmationbias.Thisisbecausetheconsistencybias
wasobservedinbothconditionsofthisstudy,withnostatisticallysigniÞcantdi
"
erencein
thesizeoftherelationship.Subjectsinthenon-customizationconditionhadnoopportunity
tochoosecategoriesandsocouldnotchoosecategoriesthatwereconsistentwithanyinitial
decisiontheymayhavemade.
RatherthanconÞrmationbias,thisbiasmaybeaconßationofalgorithmsuccesswith
outcomesuccess.Whenitappearedthatthealgorithmwasabletosuccessfullyarriveat
asolutionthatmatcheditsinput,subjectsmayhaveinterpretedthisasanindicatorof
recommendationquality.ItispossiblethatusershaveamentalmodeloftheIDAalgorithm
thatpermitsittoÞndsolutionsthatdonotmatchitsinput,butwhenitdoesÞndasolution
itisanindicatorthatthesolutionisreliable.
Animportantconsiderationforunderstandingthisbiasisthatitdependsonthetrans-
parencyofthesystem.IfusersknownothingabouttheconÞgurationofthesystem,than
itisnotpossibleforthemtoassesswhethertherecommendationsareconsistentwithits
conÞguration.Consistencybiasisthereforehighlyrelatedtothetransparencyofthesystem.
Oneinterpretationofthisbiasmaybethatusersaremoreinclinedtoagreewithrecom-
mendationswhentheyunderstandhowitworks,butwhenthesystemisnotconsistentwith
itself,itisconfusingtouserswhothenquestionwhetherthesystemworkshowtheythink.
Therefore,itispossiblethatconsistencybiasisabiasofthetransparencyofthesystem.In
chapter5,Iwilldiscusstransparencybiasanditsrelationshiptoconsistencybiasfurther.
Anotherconsiderationforunderstandingthisbiasisthatbythedesignofthestudy,
usersgenerallyhadhighexpectationsthatthesystemwouldworkwell.Therefore,itis
possiblethatthisbiaswouldnotapplyinasituationwhereusersexpectedthesystemto
performpoorly.IfthesystemwasusingaconÞgurationthatsubjectsexpectedtoproduce
poorrecommendations,andthesystemproducedpoorrecommendations,thesystemwould
beconsistentbutitwouldseemunlikelythatuserswouldagreewithit.However,this
consistencymightleadtousershavingagenerallyhigherexpectationofthesystembecause
81itdemonstratesthatitslogicworksasintended,whichmightleadtogreateragreementin
thefuture.
Thestudywasnotdesignedtoidentifythemechanismbehindthisbias,butthisisan
areaforfuturework.Inparticular,theideathatIDAusersmaybemislediftheyfeel
thatasystemhasworkedÒas-intendedÓhasdesignimplications.Designingtransparent
systemsthatallowuserstoseetheinnerlogicandunderstanditmayleadtosuchabias
ifusersconßateasystemworkingÒas-intendedÓwithworkinge
"
ectivelytoproducegood
recommendations.
Overall,thisstudydemonstratesthatcustomizationbiaspresentsahumanfactorscon-
cernforthedesignofinteractiveIDAs.Usersmaybecomepartialtoagreeingwithrecom-
mendationspurelybecauseoftheirinvolvementinproducingthem,regardlessoftheactual
qualityofrecommendationsoreventheirexpectationsthatthesystemworkswell.This
createsthepotentialfordecision-makingerrorsandbiasesthatmaylimitIDAe
"
ectiveness.
Nevertheless,animportanttakeawayfromthisstudyisthatuserbeliefsaboutsysteme
!
-cacyarerelatedtodecisionmaking.Whatusersbelieveabouthowwellthesystemworks
priortousingitinßuenceshowtheyinterprettheoutput.Thisissuewillbefurtherexamined
inthenextchapter.
3.5.1Limitations
AmajorlimitationofthisstudyisthattheIDAusedwasonlyashellanddidnotactually
useanyintelligenttechnologytoproducerecommendations.TheconÞgurationsthatwere
purportedlybeingusedbythesystemhadnoe
"
ectontherecommendations,andtherefore
thissystemonlymimicsthefunctionalityofanIDA.Themanipulationchecksusedinthe
studysuggestedthatuserswerenotawarethesystemwasfake,butitnonethelessmayhave
providedadi
"
erentexperiencethanatrueIDAthatusesactualintelligenttechnologiesto
producerecommendationsand,ifcustomizable,isresponsivetousers.Thislimitationis
addressedinthestudyreportedinchapter5whichhasasimilarexperimentdesignbutuses
82anIDAthatisactuallyresponsivetoitsconÞguration.
Anotherimportantlimitationforthisstudy,andforallthreestudiesreportedinthis
dissertation,isthattheremaybeimportantdi
"
erencesbetweenthewaypeoplemakedeci-
sionsinalabsettingandhowIDA-supporteddecisionsaremadeintherealworld.Some
humanfactorsscholarshavearguedthatcontrolledlabexperimentssuchasthismustbe
complimentedbyÞeldstudiesinÒnaturalisticÓsettings(Klein,2008)wherepeoplehavereal
pressureandconsequencesfortheirdecisions.Kleinarguesthatstudyingdecisionmaking
throughobservationalÞeldstudiesallowshumanfactorsresearcherstobetterunderstand
cognitiveanddecision-makingprocessesastheyactuallyhappeninreal-worldandthatthis
providesausefulbasisforsystemdesign.Inthisstudy,Io
"
eredincentivestosubjectsto
makegooddecisions.Andsincethesubjectswererecruitedfromaworkforcewhoinmany
casesuseparticipationinstudiesasameansofincome(Ross,Irani,Silberman,Zaldivar,&
Tomlinson,2010),thedecisiontaskhassomeÒnaturalisticÓvalidity.However,thereareun-
doubtedlydi
"
erencesbetweenanunfamiliarbaseballdecisiontaskworthafewextradollars
andhighpressuredecisionslikethoseinahospitaloronabattleÞeld,andthesedi
"
erences
maystronglya
"
ecthowpeopleuseandtrustanIDA.Furthermore,thedesignofthestudy
triedtolimittheIDAasthe
onlysourceofinformationaboutthedecisiontask,whenin
morenaturalsettings,anIDAmaybeoneofmanyinformationtoolsbeingusedbyadeci-
sionmaker.Icannotknowwhetherthee
"ectIhavereportedherewouldbeapplicablein
situationswhereusershaveavarietyoftoolsandinformationresources.
AnotherlimitationisthatthisstudyfocusedonnoviceusersusingatoolfortheÞrst
time,andtheseuserswerelimitedintheirabilitytolearnandadjusttheirdecisionmaking
overtimebecausetheywerenotshowntheirscoresuntilafterallroundshadbeenplayed.
Whilethisprovidessomeinternalvaliditytothestudy,itdoessoattheexpenseofsome
externalvaliditysinceintherealworld,userscanevaluatetheirowndecisionperformance
overtimeandusethattoinßuencetheirfuturedecisionsandfutureinteractionswithIDAs.
Itisnotclearfromthisstudywhethertheseresultswouldapplytomoreexperiencedusers,
83orhowthesebiasesmightchangeovertime.ThisisimportantbecausemanyIDAsinreal-
worldsettingsareusedprimarilybypeoplewithhighexpertiseandalotofexperience.It
ispossiblethatthesebiasesonlya
"
ectnoviceusersandthatwithexperience,usersadjust
theirbehaviortooptimizedecisions.However,itisalsopossiblethatthesebiasesbecome
stronger,asexpertsbegintousemoreshortcutsorheuristicdecisionmaking.Additional
researchisneededtoevaluatehowthesebiasesa
"
ectmoreexperiencedIDAusers.
84CHAPTER4
CAUSALEFFECTOFEFFICACYBELIEFSANDEXPECTATIONSBIAS
BecauseIDAsareusedfordi
!
cultdecisionswithhighuncertainty,andbecausethesys-
temsthemselvesareoftennothighlytransparent,usersmayhavedi
!
cultiesformingwell-
calibratedbeliefsaboutasystemÕse
!
cacy.UsersÕperceptionsofanIDAÕsreliabilitymay
notmatchitsactualreliability,andthiscanbeasourceofautomationbias.Inthestudy
presentedinthepreviouschapter,therewasarelationshipbetweenusersÕe
!cacybeliefs
andtheiragreementwithrecommendations.Whenusersexpectedthatasystemwould
workwell,basedatleastpartiallyonhowitwasconÞgured,theyweremorelikelytoagree
withrecommendationsregardlessoftheirquality.
Theevidencefromthatstudyhowevercanonlysuggestanassociationbetweene
!
cacybeliefsandagreementwithrecommendations.E
!
cacybeliefswerenotrandomlyassigned,
andtheonlyrandomlyassignedvariableinthatstudy(customizabilityoftheIDA)didnot
haveane
"
ectone
!
cacybeliefs.Althoughe
!
cacybeliefsweremeasuredbeforeagreement,
itcannotberuledoutthatagreementactuallycausede
!cacybeliefs.Itispossiblethat
subjectsdecidedwhethertheywouldagreewithrecommendationsbeforetheyactuallysaw
themandthisdecisioninßuencedtheire
!cacybeliefs.Orperhapsmorelikely,therelation-
shipbetweene
!
cacybeliefsandagreementisspurious,withsomeunobservedthirdvariable
causinguserstobothhavehighbeliefsofe
!
cacyandtoagreewithrecommendations.The
designofthepreviousstudy,anditsanalysiswhichsuggestsacausalrelationship,assume
nounobservedvariablesthatcauseaspuriousrelationship,butthisassumptioncannotbe
testedunderthedesignofthatexperiment.Forthatreason,inthischapterIpresenta
studythatisdesignedtoestimatethecausale
"ectofe
!
cacybeliefsonagreementwithIDA
recommendations.
Establishingthate
!
cacybeliefscauseagreementmakesanimportantcontributionto
85IDAresearch.Ifagreementwithrecommendations,andconsequentlyusersÕcalibrationof
trustanddecisionmaking,canbeinßuencedbytheire
!cacybeliefs,thane
!cacybeliefs
presentatargetforsystemdesignerstoengineerusersÕdecisionmaking.Findingwaysto
giveusersclearandpreciseexpectationsabouthowwellasystemworksandwhatqualityof
expectationscanbeexpectedcanleadtobettercalibratedtrustinIDAsandbetterdecisions
bytheirusers.
Ife
!
cacybeliefspriortoseeingrecommendationscauseuserstoagreewithrecommen-
dationsregardlessoftheirquality,itrepresentsabiasthatiscausedbyusersÕexpectations
oftheIDAÕse
!
cacy.Butbecauseexpectationsweremerelyobservedandnotrandomly
assigned,wecannotknowwhetherthebiasistrulyabiascausedbyexpectationsorifthat
relationshipisaby-productofsomeotherunobservedbias.InthischapterIhavedesigned
astudytoidentifywhetherexpectationscauseabiasinagreement.Thisstudyprovides
evidencethattheexpectationsusershaveaboutanIDAÕse
!
cacybiastheirsubsequent
agreementwithitsrecommendations.
4.1Methods
TostudywhetherusersÕbeliefsofsysteme
!
cacycausethemtoagreewithrecommendations,
Iconductedanexperimentusingarandomizedencouragementdesign.Underthisdesign,
subjectswereencouragedbyarandomizedtreatmentvariablewhichisdescribedbelowto
haveeitherhighorlowe
!cacybeliefs.TheythenusedanIDAtogetrecommendationsand
madetheirdecision.Thisdesignusestheencouragementvariableasaninstrumentalvariable
inordertoestimatethecausale
"
ectofe
!
cacybeliefsonagreement.Aninstrumental
variableisavariablewhichhasnocorrelationwiththedependentvariableotherthanthrough
itscorrelationwiththeexplanatoryvariable(i.e.e
!cacybeliefs).Throughthismethod,
variationintheexplanatoryvariablewhichiscausedbytheinstrument(whichhasbeen
randomlyassignedinthiscase)canbeconsideredtoberandomvariationwithregardstothe
86dependentvariable.Therefore,correlationbetweentherandomvariationintheexplanatory
variableandthedependentvariablecanbeinterpretedasacausalrelationshipbetweenthe
explanatoryanddependentvariables(Angrist&Pischke,2008).
Inthisstudy,thesamefantasybaseballtaskwasusedtoimplementthisrandomized
encouragementdesignaswasusedinthecustomizationbiasstudy.Inthisstudy,onlythe
non-customizableversionoftheIDAwasused.Thiswasdonetopreventcustomizationbias
fromcreatingunwantednoiseinthedataset.Asinthepreviousstudy,theIDAprovided8
goodrecommendationsand4poorrecommendations,scoringanaverageof15points.The
gamesusedinthetask,thescreeningquiz,thepre-testsurveyandpost-testsurvey,andthe
scoringprocedureandincentivewerethesameasinthecustomizationbiasstudy.
Encouragementofe
"
cacybeliefs.
Subjectswereassignedtooneoftwoconditions.In
thehighe
"
cacy
condition,subjectsweretoldthatthesystemÕsaverageperformancewas
18points.TheywerealsotoldthatthedataaboutMajorLeagueBaseballusedbytheIDA
wascomprehensiveandcontainednoknownerrors.Inthe
lowe
"
cacy
condition,subjects
weretoldthatthesystemÕsrecommendationsscore12pointsonaverage.Theywerealso
toldthatitsdatasetcontainederrorsandomissions.Thisinformationwaspresentedto
subjectsintheinstructions,buttheinterfaceofthesystemalsocontainedareminderofthis
information.Inordertoproceedtothebaseballpredictiongame,asinthepreviousstudy,
subjectshadtopassaquizontheinstructionsthatveriÞedwhethertheyunderstoodthe
averagequalityofthesystemÕsrecommendations.
Thesedi
"erencesintheinstructionswereintendedtoencouragesubjectstowardsabe-
lievingtheIDAhadhighorlowe
!cacyinproducingaccuraterecommendationsaboutthe
outcomesofbaseballgames.Inadditiontothesedi
"erencesininstructions,therewasalso
adi
"
erenceintheconÞgurationthatwaspurportedlyusedbythesystem.Asinthecus-
tomizationbiasstudy,subjectsratedeachofthe27categoriesaccordingtohowwellitwould
informacomputerinpredictingtheoutcomeofbaseballgames.Usingtheratingsobtained
aboutthesecategoriesinpilottesting,IcreatedconÞgurationsthatusedcategoriesthatwere
87eitherhighlyratedorpoorlyratedonaveragebysubjects.Icreated8conÞgurations,with
4usingmostlypoorlyratedcategoriesandtheother4usingonlyhighlyratedcategories.In
thehighe
!
cacycondition,theIDAwaspresentedasusingoneoftheconÞgurationswith
highlyratedcategories,andinthelowe
!
cacyconditiontheIDAusedonlyconÞgurations
thatwerepoorlyrated.
ThegoaloftheseconditionswastorandomlysetsubjectsÕbeliefsofsysteme
!
cacyto
eitherahighorlowlevel.Iusedthreedi
"erentapproachestoencouragethesee
!
cacybeliefs
(expectedperformance,dataquality,conÞgurationquality)inordertocreateastrongand
e"ectiveinstrument.Adisadvantageofsimultaneouslyusingthreeapproachesisthatthe
e"ectofanyofthethreecannotbeidentiÞedastheyareconfoundedwiththecondition
assignment.However,astheprimarypurposeofthisstudyistoestimateacausale
"
ect,
astronginstrumentisnecessary.Pilotteststhatusedonlythedataqualityandexpected
performanceapproachesproducedonlyamoderatedi
"
erenceine
!
cacybeliefsandthusmay
haveresultedinaweakinstrument.
Atotalof93subjectswererecruitedfromAmazonMechanicalTurk.Subjectswerepaid
$2forparticipationandanaverageof$2.25inbonuspaymentsfordecisionperformance.
Subjectstookthesamebaseballknowledgescreeningquizasinthecustomizationbiasstudy
inChapter2.31subjectsfailedthisquizandwerescreenedoutoffurtherparticipation.
TheÞnaldatasetincluded62subjects,with32inthehighe
!
cacycondition.
Thisstudyhasthreehypotheses.
¥H1.Encouragementofhighorlowe
!
cacybeliefswillcausesubjectstohavehighor
lowe
!
cacybeliefs.
¥H2.E!
cacybeliefswillbeassociatedwithagreement.
¥H3.E!
cacybeliefswillcauseagreementasassessedbytwo-stageleastsquaresre-
gression.884.2Results
Table4.1:Descriptivestatisticsandfactoranalysisofdependentvariables.
StatisticNMeanSt.Dev.FactorLoadingStd.Err.
Likertmeasure7444.8241.3951.000-
Expectedscore74413.7434.2103.0130.121
Expectedprobability
ofcorrectwinner
74475.88214.96111.2530.435
Agreement74415.1458.967
WinnerAgreement7440.7940.404
FactorAnalysisFitIndices
ComparativeFitIndex1.000
RootMeanSquareError0.000
4.2.1Descriptiveresults
AconÞrmatoryfactoranalysiswasperformedtoobtainfactorscoresforthee
!
cacybeliefs
variableusingthethreedi
"
erentmeasures.ThefactorloadingsarereportedinTable4.1.
Figure4.1andFigure4.2showsubjectsÕe
!cacybeliefsandwinneragreementover
thecourseofthe12roundsoftheexperiment.IÞtmultilevelregressionmodelstothe
datatodeterminewhethertherewasanytrendinhowusersmadedecisionsortheire
!cacy
beliefsasaresultofrepeatedlyplayingthebaseballpredictiongame.Thesemodelsfoundno
statisticallysigniÞcanttrend.Subjectsdidnotadjusttheiragreementwithrecommendations
ortheirbeliefsofitse
!cacyasalinearfunctionofhowmanyroundsofthegamehadbeen
played.
4.2.2Manipulationcheck
Todeterminewhethertheencouragementhadtheintendede
"
ectonsubjectsÕbeliefsabout
theIDAÕse
!
cacy,IÞtamultilevelregressionmodelthatestimatedthee
!
cacybeliefsfactor
scorewiththeencouragementvariable,propensitytotrustdecisionaids,andarandome
"ect890.60.70.80.9123456789101112Round NumberAverage Winner Agreement
High Encouragement
Low Encouragement
Figure4.1:Winneragreementoverthecourseofthe12rounds.
!1.0!0.50.00.51.0123456789101112Round NumberAverage Efficacy Beliefs
High Encouragement
Low Encouragement
Figure4.2:E
!
cacyBeliefsoverthecourseofthe12rounds.
90Table4.2:E
!
cacyvariablesbyexperimentcondition.
StatisticHighEncouragementLowEncouragement
MeanSt.Dev.MeanSt.Dev.
Likert5.5051.0674.0971.337
Score16.4643.34010.8422.903
Probability83.52912.63867.72512.791
E!
cacyBeliefsFactor.6580.853-.702.850
Table4.3:E
"
ectofencouragementone
!
cacybeliefs.
Dependentvariable:
E!
cacyFactorLikertScoreProbability
Intercept
!0.662"""4.149"""10.923"""68.337"""(0.106)(0.136)(0.423)(1.707)
Encouraged1.253
"""1.272"""5.405"""14.179"""(0.149)(0.192)(0.598)(2.412)
TrustPropensity0.362
"""0.461"""0.735"5.499"""(0.102)(0.131)(0.409)(1.651)
RandomE
"
ectsStd.Dev.0.5470.6862.2158.981
LogLikelihood
!763.979!1,078.090!1,723.692!2,710.091LogLikelihood
!261.176"""47.885"""59.624"""41.291"""Note:
"p<0.1;""p<0.05;"""p<0.01foreachsubjectasregressors.IÞtthreeadditionalmodelsusingtheseparatee
!cacybeliefs
variables(LikertScale,ExpectedScore,andExpectedProbabilityofCorrectWinner).These
modelsaredetailedinTable4.3.Allfourmodelso
"
erthesameconclusionthatsubjects
intheHighencouragementconditionhadhigherbeliefsaboutthesystemÕse
!
cacythan
subjectsintheLowencouragementcondition.Figure4.3illustratesthedistributionsof
e!
cacybeliefsineachcondition.Fromtheseanalysesitisclearthattheencouragement
variablewassuccessfulinitsintendede
"
ecttomanipulatesubjectsÕbeliefsaboutthesystemÕs
e!
cacy.
91!3!2!1012High Encouragement
Low Encouragement
Efficacy Beliefs Factor Score
Figure4.3:Distributionsofe
!
cacybeliefsbyencouragementcondition.
Table4.4:E
"
ectofe
!
cacybeliefsonagreement.
Dependentvariable:
AgreementWinnerAgreement
Intercept16.591
"""2.308"""(0.775)(0.221)
E!
cacyBeliefs1.315
"""0.321""(0.380)(0.128)
PoorRecommendation
!4.226"""!1.554"""(0.519)(0.213)
TrustPropensity
!0.8980.297
(1.053)(0.238)
RandomE
"
ectsStd.Dev.5.6221.029
LogLikelihood
!2,533.185!329.493LogLikelihood
!272.853"""66.118"""Pseudo
R20.225Note:
"p<0.1;""p<0.05;"""p<0.01924.2.3Estimateofe
!
ectofe
"
cacybeliefsonagreement
Table4.4describesmodelsthatestimateagreementwithrecommendations.Themodels
showastatisticallysigniÞcantrelationshipbetweene
!
cacybeliefsandagreementforboth
measuresofagreement.Whensubjectsbelievedthatthesystemwouldworkwellprior
toreceivingrecommendations,theyagreedwiththoserecommendationsmoreregardlessof
theirquality.Anothernoteworthyresultfromthesemodelsisthattheindividualdi
"
erence
ofgenerallyhavingtrustinautomateddecisionaidsdidnothaveastatisticallysigniÞcant
relationshipwithagreementwhencontrollingfore
!cacybeliefs.Thissuggeststhatthe
individualvariationinunderlyingtrustiswellcapturedbythee
!cacybeliefsmeasures.
Whenuserstrustdecisionaids,theyalsoexpectthemtoperformwellbutbasetheirdecisions
moreontheexpectationofsysteme
!cacythanontheirunderlyingtrustofdecisionaidsin
general.Table4.5:Two-stageleastsquaresestimatesofcausale
"
ectofe
!
cacybeliefsonagreement.
E!
cacybeliefsareinstrumentedbyencouragementinthismodel.
Dependentvariable:
AgreementWinnerAgreement
(1)(2)
Intercept16.578
"""0.870"""(0.487)(0.017)
E!
cacyBeliefs2.77
"""0.087"""(1.210)(0.032)
PoorRecommendation
!4.298"""!.228"""(1.233)(0.034)
RandomE
"ectsStd.Dev.5.8790.135
R20.0530.084
Wald
!213.09"""53.46"""Note:
"p<0.1;""p<0.05;"""p<0.01934.2.4Causale
!
ectofe
"
cacybeliefsonagreement
Encouragementdesignssuchasthisonecanallowforcausalinferenceabouttheactual
treatment,inthiscasee
!cacybeliefs,eventhoughonlytheencouragementvariablehas
beenreliablyrandomized.Theencouragementdesignusestherandomizedencouragement
variableasaninstrumentalvariablethatcreatessomerandomvarianceine
!cacybeliefs.
Therandomvarianceine
!
cacybeliefsthatiscausedbytheencouragementvariablecanbe
usedtoestimatethecausale
"
ectofe
!
cacybeliefsonagreement(Angrist&Pischke,2008).
Iusedtwo-stageleastsquaresasthemethodforestimatingthecausale
"
ectofe
!
cacy
beliefswiththeinstrumentalvariablesmethod.Two-stageleastsquaresperformstwore-
gressions.IntheÞrststage,theexplanatoryvariable(e
!cacybeliefs)isregressedonthe
instrument(encouragement),aswellastherecommendationqualityvariableasanexogenous
covariate.Inthesecondstage,agreementisregressedone
!cacybeliefsandrecommendation
quality,withtheÞttedvaluesfromtheÞrststagereplacingtheobservedvalues.
Becauseoftherepeatedmeasuresofthestudy,thedataweretreatedaspaneldatafor
thepurposesoftwo-stageleastsquaresestimation.Sincetheencouragementvariablewas
assignedatthesubjectlevel,ratherthanattheroundlevel,randome
"
ectsforeachsub-
jectwereestimatedratherthanÞxede
"ectsineachstageofthetwo-stageleastsquares.
Fixede
"
ectscannotbeestimatedbecausethesee
"
ectsarecolinearwiththeencouragement
variable.ThemodelwasrunusingStataÕs
xtivreg
commandwhichperformsageneralized
two-stageleastsquaresregressionthatincludestherandome
"
ects.Robuststandarderrors
werecalculatedusingabootstrappingmethodthatrandomlyresampledthedata1000times
tore-estimatethestandarderrors.BootstrappedstandarderrorswereusedbecauseStataÕs
xtivreg
commandassumesconstantvarianceifregularstandarderrorsareused.Therandom
e
"
ectsinthismodelnecessitateanadditionalassumption,whichisthatanyindividual-level
e"ectsarenotcorrelatedwithanyoftheinstrumentsorcovariatesinthemodel(Mundlak,
1978).Thisassumptioncanreasonablybemadeforthesedatabecauseboththeencourage-
940.50.60.70.80.9!3!2!1012Efficacy Beliefs ScoreProbability of Agreeing with IDA
Good RecommendationPoor Recommendation
Effect of Efficacy on Agreement
Figure4.4:Associationbetweene
!
cacybeliefsandagreement.
mentvariableandrecommendationqualitywererandomlyassigned.
AseparatemodelwasÞtforeachversionofagreement.Forthemodelwiththebinary
WinnerAgreementdependentvariable,IfollowedAngristandPischkeÕs(2008)suggestion
andusedaLinearProbabilityModelthattreatsthebinaryvariableascontinuousrather
thananon-lineartransformationsuchasthelogisticregressionsusedinotherpartsofthis
dissertation.AngristandPischkearguethatthesenon-linearmodelscreateadditionalcom-
plexitywhenusedwithinstrumentalvariablesandthatthisadditionalcomplexityisoften
notjustiÞedbecauseLinearProbabilityModelsperformwellatestimatingmarginale
"
ects,
eveniftheirpredictedvaluesareimprecise.
Theresultsofthetwo-stageleastsquaresarereportedinTable4.5.Forbothmeasures
ofagreement,theanalysisshowedstatisticallysigniÞcante
"
ectsofe
!
cacybeliefsonagree-
ment.TheseestimatessuggestthatbelievinganIDAhashighe
!cacypriortoseeingits
recommendationscausesuserstoagreewiththoserecommendations,evenwhencontrolling
forthequalityoftherecommendation.Thise
"
ectisillustratedinFigure4.4
95Two-stageleastsquareshastwoprimaryrequirementsinordertobeabletointerpret
acausale
"
ect.TheÞrstrequirementisthattheinstrumentisstronglycorrelatedwiththe
explanatoryvariable.Theanalysespresentedintheprevioussectiono
"erevidenceofa
stronge
"ectoftheencouragementone
!
cacybeliefs.Inaddition,StockandYogo(2005)
developedcriteriaforidentifyingweakinstruments.Undertheircriteria,amodelwithone
endogenousexplanatoryvariableandoneinstrumentshouldhaveanF-statisticofgreater
than16.38intheÞrststageregressioninordertobeassuredofonlyaminimalamountof
biasduetoaweakinstrument.TheF-statisticfromtheÞrststageregressions(whichisthe
sameregressionforbothmodels)is75.08.ThereforeIcanconcludethattheinstrumentis
strong,satisfyingtheÞrstrequirement.
Thesecondrequirementisthattheinstrumentshavenoe
"ectontheoutcomevariable
otherthanthroughtheexplanatoryvariable.Thiscanbetestedwhenamodelisoveriden-
tiÞed,meaningthattherearemoreinstrumentsthanexplanatoryvariables(Sargan,1958).
However,thismodelisexactly-identiÞedwithoneinstrumentandoneexplanatoryvariable,
andthereforethisassumptioncannotbeexplicitlytested.Forthisreason,theinterpretation
ofthetwo-stageleastsquaresmodelofacausale
"ectofe
!cacyonagreementisvalidonly
inasmuchasthisassumptionisvalid.Whilethisisalimitationthatshouldbeconsidered
intheinterpretationoftheseresults,Iarguethatthisassumptionisreasonableinthiscir-
cumstance.OneoftheencouragementvariableÕsmanipulationswasasystem-estimationof
itsowne
!cacythatwasmerelycommunicatedtousers.Theothertwomanipulationswere
variationsofhowthesystempurportedlyworkedtoproducerecommendations.Itisdi
!
cult
toconceiveofanywaytheserandomlyassignedmanipulationscoulda
"
ectagreementother
thanbya
"
ectingsubjectsÕbeliefsaboutthesystemÕse
!cacy.Forthisreason,themostlikely
interpretationoftherelationshipthathasbeenfoundbetweene
!cacybeliefsandagreement
isthate
!
cacybeliefscausesomeagreementwithIDArecommendations
964.2.5DecisionMaking
Table4.6:Decision-makingquality.
Dependentvariable:
NumberofPointsEarned
Intercept16.572
"""(0.327)E!
cacyBeliefs0.608
""(0.240)TrustPropensity
!0.730"(0.428)PoorRecommendation
!6.199"""(0.400)RandomE
"ectsStd.Dev.1.816
LogLikelihood
!2,300.606LogLikelihood
!2211.51"""Note:
"p<0.1;""p<0.05;"""p<0.01IexaminedhowsubjectsÕe
!
cacybeliefsimpactedtheirdecisionmakingqualitybyÞtting
amultilevelmodelthatestimatesthenumberofpointsearnedfromaroundofthegamewith
thee
!cacybeliefsfactorscore,recommendationquality,andtrustpropensityasregressors
alongwithrandome
"
ectsforeachsubject.ThismodelisdescribedinTable4.6.
SubjectsmadebetterdecisionswhentheyhadhigherbeliefsabouttheIDAÕse
!
cacy.
Thiscanlikelybeexplainedbythefactthathighe
!
cacybeliefscausedagreementwithrec-
ommendations,andsincethereweremoregoodrecommendationsthanpoorones,andsince
goodrecommendationssuggestedthebestpossibledecisionthatcouldbemade,subjects
whobelievedthesystemwouldworkwelltendedtofollowthosegoodrecommendations.
Althoughthise
"
ectisstatisticallysigniÞcant,itisnotaparticularlylargee
"
ect.Thedi
"
er-
encebetweentheloweste
!
cacybeliefsscoreandthehighestisonlyabout3points,which
97Table4.7:Summaryofresults.
Hypothesis
ResultH1Encouragementofhighorlowe
!
cacybeliefswillcause
subjectstohavehighorlowe
!
cacybeliefs.
Supported
H2E!
cacybeliefswillbeassociatedwithagreement.
Supported
H3E!
cacybeliefswillcauseagreementasassessedbytwo-
stageleastsquaresregression.
Supported
isfairlysmallgiventheoverallvarianceondecisionquality.
AnothernoteworthyÞndingisthatsubjectswithhigherpropensitytotrustdecisionaids
actuallymadeworsedecisionsthanthosewithlowerlevelsoftrust.ThisÞndingmayillus-
trateanindividualdi
"
erenceinautomationbias.Peoplewhohavehigherpropensitytotrust
decisionaidsmaynothavescrutinizedrecommendationscarefully,causingaparticularlyhigh
rateoferrorwheretheyagreedwithapoorrecommendation.
4.3Discussion
ThisstudyhasdemonstratedthatthebeliefsIDAusershaveaboutthee
!
cacyofasystem
priortoseeingitsoutputhasane
"
ectontheirdecisionmakingaftertheyseeitsrecommen-
dations.Whenusersexpectthesystemtoproducegoodrecommendations,theyinterpret
therecommendationsthesystemproducesasbeingmoretrustworthy.Inthisstudy,all
subjectssawidenticalrecommendationsbutwereencouragedtohavevaryingbeliefsabout
thee
!cacyofthesystem,andthesebeliefschangedthelevelofagreementwiththesame
recommendations.
Becauserecommendationswereidentical,thisÞndingthate
!
cacybeliefsinßuenced
agreementisevidenceofanotherdecisionmakingbias.E
!
cacybeliefsweremeasuredprior
tosubjectsseeingtherecommendations,makingthemexpectationsaboutthesystemthat
wereformedbasedonsubjectÕsknowledgeofthesystemandofthestatisticsaboutthe
teamsplayinginthegame.Thegreateragreementbythosewithgreatere
!
cacybeliefsis
anexpectationsbias
.98AnimportantimplicationforthedesignofIDAfromthisÞndingisthatwhenmaking
decisions,evaluatingthee
!cacyofthesystemisapartofthedecisionmakingprocess.
AnIDAÕsrecommendationsareinformationthatmustbeevaluated,andbeliefsaboutthe
processthatproducesrecommendationscanleadpeopletointerpretthesameinformation
indi
"
erentways.ThishasrelevancetothedebateovertransparencyinIDAs.Transpar-
entIDAÕsarewellunderstoodtoprovideabetteruserexperience(Herlockeretal.,2000),
howevertheire
"
ectondecisionmakingislessconclusive,withsomeevidencesuggestinga
negativeimpact(Ehrlichetal.,2011).Thisstudyo
"
erssomeevidenceforwhytransparency
hasnotdemonstratedaclearimpactondecisionmakingperformance.Iftransparentsys-
temsprovideinformationabouthowasystemworks,thisinformationmustbeevaluated
byusers.EvaluatingthesystemÕsinnerlogicandforminganexpectationaboutitse
!
cacy
maydistractusersfromevaluatingrecommendations.Transparencymaythereforepresent
somewhatofaparadoxbecauseusers,whoareseekingadvicefromanautomatedsystem
becausetheyfaceadi
!
cultdecisionwithhighuncertainty,mayhavehighuncertaintyabout
whatconstitutesane
"
ectiveprocessforgeneratingrecommendations.Inmanycasesusers
arelikelybetterqualiÞedtoevaluaterecommendationsdirectly(withoutknowledgeofany
processthatproducedthem)thantoevaluateacomputationalprocessforproducingrec-
ommendations.CalibrationofusersÕexpectedreliabilityofasystemanditsactualreliabilityshouldbe
animportantgoalforsystemdesigners,becausethiscalibrationcanleadtobetterdecisions
(Dzindoletetal.,2003,2002).Poordecisionsarecertaintohappenifusersbelieveasystem
tobeeithermuchmorereliableormuchlessreliablethanitactuallyis.Therefore,IDA
designsshouldconsiderwaystohelpuserscalibratetheire
!cacybeliefs.
Anothercontributionofthisstudyisthatitdemonstratesthate
!
cacybeliefsaremal-
leable,andthatthedesignofthesystemcaninßuenceusersÕe
!cacybeliefs,atleastfor
newuserslikethesubjectsinthisstudy.Theencouragementvariablemanipulatedthree
aspectsofthesystem.Thesystemitselfstateditsexpectede
!cacybyindicatingitsaverage
99performancewithintheinterface.Also,subjectsweregiveninformationaboutthequalityof
thedatathatitusestomakepredictions.AndtheconÞgurationthatthesystemusedwas
eitheronethatmightbeexpectedtoworkwelloronethatusedcategoriesmostpeopledo
notthinkcouldworkwellforpredictinggameoutcomes.Becausetheintentoftheencour-
agementvariablewastorandomlymanipulatee
!cacybeliefs,thesethreedesignfeatures
werevariedtohighorlowdegreessimultaneouslyinane
"
orttomaximizerandomvariation
ine
!cacybeliefs.Itisclearthatatleastoneofthesedesignfeaturesisresponsibleforthe
observeddi
"
erencesine
!
cacybeliefsbetweenconditions.However,thestudydesignmakes
itimpossibletodeterminethespeciÞce
"
ectofeachindividualfeatureone
!cacybeliefs.
Nevertheless,allthreearereasonablecausesofe
!
cacybeliefsthatcanbeincorporatedinto
asystemdesign.FutureworkcanexplorehowthesefeaturesandothersinßuenceusersÕex-
pectationsaboutIDAoutput.Aparticularlyusefulfuturestudywouldexaminetheconcept
ofdataqualitymorepreciselytodeterminehowusersevaluatethequalityofdatathatare
inputintoIDAsandhowthata
"
ectstheirexpectationsoftheire
!
cacy.Thisdissertation
hasprimarilyfocusedoncustomizationofIDAlogic,butcustomizationofdataisanother
waytomakesystemscustomizable.Evaluatingwhethercustomizationofdataa
"
ectse
!
-cacybeliefs,andhowcustomizationandgeneralattitudesaboutdataa
"ectdecisionsand
trustinthesystem,wouldmakeavaluablecontinuationoftheworkpresentedhere.
Theresultthathighe
!
cacybeliefsimproveddecisionmakingwarrantssomeadditional
discussionaswell.Sincehighe
!cacybeliefscauseuserstoagreewithrecommendations,
highe
!cacyworkstoimprovedecisionmakingwhensystemsarelargelyreliableandusers
otherwisecommitomissionerrorsbynotfollowinggoodrecommendations.However,had
theIDAinthisstudybeenfarlessreliable,itispossiblethathighe
!cacybeliefscould
havehurtdecisionmaking.Again,thisillustratestheimportanceofcalibrationbetween
usersÕbeliefsandactualsystemperformance.Withoutthiscalibration,IDAscanleadusers
topoordecisionmaking.DesigningtoallowuserstoformaccuratebeliefsaboutanIDAÕs
e
!
cacycanhaveimportantbeneÞtsformakingIDAsthataree
"
ectiveatimprovinguser
100decisionmaking.
4.3.1Limitations
Thisstudywasdesignedtoidentifythecausale
"ectofe
!
cacybeliefsonagreement,and
o"ersonlyverylimitedinsightintohowusersnaturallyformexpectationsofthesystemÕsef-
Þcacy.Theencouragementvariablemanipulatedthreeaspectsofthesystem(conÞguration,
pastperformance,dataquality)anditsinstructions,buttheseaspectswereintentionally
confoundedwitheachother.ThereforeIcannotassesshowmucheachspeciÞcaspectcon-
tributedtothevarianceine
!cacybeliefs.Sincethestudyfoundthatexpectationscanhave
astronge
"
ectonusersÕbehavior,itisimportanttoexploreaspectsofthesystemdesign
orcharacteristicsofusersthatinßuencee
!cacybeliefssoitcanbedeterminedhowbestto
engineeragreementwithrecommendations.
101CHAPTER5
CUSTOMIZATION,NON-CUSTOMIZATION,ORBOTHINANIDAFOR
EXERCISEDECISIONS
CustomizationinIDAdesignscancreateadi
!
cultsocio-technicalchallengefordesigners.
CustomizableIDAsmaybeabletotakeadvantageofusersÕexpertise,theirlocalknowledgeof
theirdecisionscenario,aswellaspersonalknowledgeoftheirpreferencesanddecisionmaking
stylestoimprovetheIDAÕsalgorithmandprovidebetterrecommendations.However,asI
willshowinthischapter,therearedecision-makingconsequencestothisdesign,andthese
consequencesmaynotbejustiÞedbyanimprovementintheIDAÕsalgorithm.
Inthischapter,IwillpresentastudythatcomparesanIDAdesignutilizingcustomizable
recommendationlogictoadesignthatdoesnotgiveusersanycontrolovertherecommenda-
tion.Additionally,thisstudywillevaluateadesignthatprovidesuserswithbothcustomized
andnon-customizedrecommendationssimultaneously.Thisstudywillevaluateandcompare
thesedi
"
erentdesignsregardingtheire
"
ectonusersÕdecisionmaking.Thepurposeofthis
evaluationistoproduceinsightsthatcanhelpIDAdesignersdeterminehowbesttoin-
corporatecustomization,ifatall,inanIDAdesign.Thisstudywillalsoextendmywork
oncustomizationbiasandreplicatepreviousÞndingsinadi
"erentdecisioncontextusing
adi
"
erenttypeofIDA,notablythatwillactuallyadjustrecommendationsinresponseto
usersÕconÞgurationdecisions.
Inthisstudy,Iwillalsodescribeatestofthee
"ectthatcustomizationhasonthe
transparencyofanIDA.TransparencyinIDAsistypicallythoughtofasafunctionof
providingexplanationsintheinterfaceabouthowrecommendationsweregenerated(see
section5ofChapter2above).However,customizationmaymakeasystemmoretransparent
withoutprovidingexplanationsbecauseitgivessignalsaboutwhataspectsofthedecision
areimportanttothealgorithm.InthisstudyIwillshowthatcustomizationleadstoa
102moderateincreaseinthetransparencyoftheIDA.Iwillalsoshowanassociationbetween
transparencyandusersÕagreementwithrecommendations.Userswhofeelthesystemis
moretransparentaremorelikelytoagreewithit,evenwhencontrollingforrecommendation
qualityandcustomizabilityoftheIDA.Thisassociationsuggestsanotherimportantdecision
makingbiasthatresultsfromIDAuse.
5.1ResearchQuestions
RQ1.
CancustomizationleadIDAuserstomakebetterdecisionsevenifrecommendations
arenobetterthanthoseprovidedbyanon-customizableIDA?
OneargumentthatcanbemadeformakingIDAscustomizableisthatusersprovide
expertise,localknowledgeandsituationalawarenessaboutthedecisionscenarioandtheir
ownpreferencesthatisimpossibleordi
!
cultforthesystemtoobtainfromanyothersource
butthedecisionmaker.Therefore,byincorporatingacustomizablealgorithmintoanIDA
design,designersallowthesystemtocaptureinformationthatmayleadtobetterrecommen-
dationsforspeciÞcdecisions.However,becausecustomizationplacesadditionaldemandson
users,itcreatesadi
"
erentdecisionprocesscomparedtosystemsthatcompletelyautomate
theacquisitionandanalysisofinformation.UsersmustthinkabouthowtoconÞgurethe
IDA,spendtimeande
"
ortdoingso,andinterpretbothhowtheirinputhasinßuenced
thesystemÕsoutputaswellasevaluatetheoutputanditsappropriatenessforthedecision.
Evenifcustomizationleadstobetterrecommendations,theseadditionaldemandsplaced
ontheusermaya
"
ecttheirdecisions.Forexample,theprocessofthinkingabouthow
toconÞgurethesystemmightleaduserstonewinsightsaboutthedecision,helpingthem
makebetterdecisionsinawaythatisnotrelatedtotheactualrecommendationsprovided.
Or,itmaycreatefatiguethatleadsuserstoinsu
!
cientlyevaluateallalternativesortake
cognitiveshortcutswhenselectinganaction.Theseareexamplesofmechanismsbywhich
customizationmighta
"ectdecisionmakingthatareindependentoftherecommendations
103thattheIDAproduces.
InthisstudyIwillconductanexperimentthatcomparesdecisionmakingassupportedby
bothcustomizableandnon-customizableIDAs.Ihavedesignedastudytoseekoutevidence
thatcustomizationcanleaduserstomakebetterdecisions.ThemechanismbywhichI
suspectthiscouldhappeniselaborationoverthedecisionthatmaybestructuredwithinthe
decisionprocessbythedesignoftheIDA.Thismechanismisacognitiveprocessforwhich
thereisnotaclearordirectmeasureatthistime.Therefore,IwillÞrstdeterminewhether
customizationhasanya
"ectondecisionqualityinthisstudysothatIcandeterminewhether
developmentofameasureofcustomizationelaborationisaworthwhileresearchdirection.
Thehypothesesrelatedtothisresearchquestionare:
¥H1UserswhousetheIDAwillmakebetterdecisionsthanotherswhomakethe
decisionunaided.
¥H2Userswhoseecustomizedrecommendationswillmakebetterdecisionsthanusers
whoseeonlynon-customizedrecommendations.
RQ2.
CancustomizationbiasbeobservedinanIDAthatistrulycustomizable?
AmajorlimitationofthestudiesIhavepresentedoncustomizationbias(Chapter3,
(Solomon,2014))isthattheIDAusedinthosestudieswasmerelyashellthatincludeda
customizableinterface,butthesystemÕsrecommendationlogicwasnottrulycustomizable.
Althoughmanipulationchecksfromthosestudiesindicatedthatusersdidperceivetohave
somecontrolovertheIDAwhencustomizingit,thoseresultsmayhavebeentempered
bythefactthatsubjectsmightnotalwayshavebeenabletorecognizeaclearconnection
betweentheiractionsandthesystemÕsoutput.Therefore,thoseresultsmaybeonlya
manifestationoftheIllusionofControl(Langer,1975).However,recentworkonthistheory
hasdemonstratedthattheIllusionofControlcanonlybeobservedinsituationswhenpeople
havelittletonoactualcontrol(Ginoetal.,2011),aswasthecaseinmypreviousstudies.
104Ginoetal.foundthatinsituationswherethereisahighdegreeofactualcontrol,peopletend
tounderestimatetheircontrol.Ifcustomizationbiasfrommypreviousstudieswastheresult
oftheIllusionofControl,theninacontextwhereusershaveahighdegreeofactualcontrol
overtherecommendations,itisplausiblethatcustomizationbiaswouldnotbeobservedor
thatcustomizersmaybebiasedagainstfollowingtheircustomrecommendations.Forthis
reason,itiscriticaltoevaluatecustomizationbiasinacontextwhereusershaveahighdegree
ofactualcontrolovertherecommendationsthatareproducedbytheIDA.Thehypothesis
(H3)isthatuserswhoseeonlycustomrecommendationswillhavemoreagreementthan
userswhoseeonlynon-customrecommendations.
RQ3.
CananIDAdesignutilizingbothcustomizedandnon-customizedrecommendations
reducecustomizationbiasorimprovedecisionmakingovereitherapproachindividually?
Customizationbiasmayhavepositiveornegativee
"ectsonIDAe
"ectiveness.Forex-
ample,inasystemwhereusershavetoolittletrustandoftenignoregoodrecommendations,
customizationmayhelpusersmakebetterdecisionsbyincreasingthelikelihoodthatthey
followthesystemÕsadvice.However,customizationmayalsocreateautomationbiaswhere
usersaretootrustingofthesystemandfollowpoorrecommendations.Becausecustomiza-
tionmayhavenegativee
"
ectsondecisionmaking,developingdesigninterventionsthatcan
reducecustomizationbiasisanimportantcontributiontoIDAresearch.
OnepotentialinterventiononcustomizationbiasisfortheIDAtosimultaneouslyshow
bothcustomizedrecommendationsaswellasrecommendationsproducedbyanalgorithm
thatisnota
"
ectedbyusersÕinputs.Non-customizedrecommendationsmayprovideusers
withacontrastingperspectiveoraÒsecondopinionÓaboutwhatdecisionshouldbemade.
Thisadditionalperspectivemaytriggeruserstoscrutinizetheirrecommendationsmore
closelyandbasetheirdecisiononabroadersetofinformation.Ifusersconsidermore
alternatives,theymaybemoreinclinedtoacceptsomealternativesthatwerenotgenerated
bytheircustomizedalgorithm.
105Thehypothesesare:
¥H4Userswhoseebothcustomandnon-customrecommendationswillmakethebest
decisionsoverall.
¥H5Userswhoseebothcustomandnon-customrecommendationswillhavelessagree-
mentwithrecommendationsthanuserswhoseeonlycustomrecommendations.
RQ4.
CancustomizationcreatetransparencyinanIDA?
TransparencyiswellunderstoodtobeimportantfortheuserexperienceofIDAs(Cramer
etal.,2008;Wang&Benbasat,2007).IDAresearchhasoftenusedexplanationsasthedesign
mechanismtocreatetransparency(Herlockeretal.,2000).Notalltypesofexplanationsare
e
"
ectivehowever(2009),andsomeworkhasfoundthatexplanationsneedtobetailored
toindividualuserstobee
"
ective(Tintarev&Mastho
"
,2007).Tailoringexplanationsto
individualusersmaybeadi
!culttechnicalchallenge,andIamnotawareofanyworkthat
hasreportedsucceedingatthis.AndotherworkhasfoundthatsomeusersÕdecisionmaking
isinhibitedbyexplanations(Ehrlichetal.,2011).
Customization,however,mayprovideanopportunitytoaddtransparencywithoutneed-
ingtoprovideexplanations.Acustomizablesystem,byvirtueofgivinguserscontrolsthat
describesomee
"
ectonthealgorithm,giveusersasignalaboutwhatisimportanttothe
algorithm.Thissignalmayhelpusersunderstandthesystemandgiveittransparency.In
thisstudythatcomparescustomizableandnon-customizableversionsofanIDA,Iwillmea-
suredi
"erencesinhowwellusersfeeltheyunderstandthealgorithmandthelogicbehind
therecommendations.Ihypothesizethatcustomizingthesystemwillmakeusersfeelthe
systemismoretransparent(
H6)andthatthemoretransparentusersfeelthesystemÕslogic
is,themoretheywillagreewithrecommendations(
H7).1065.2ExerciseRecommender
Toanswertheseresearchquestions,IbuiltanIDAthatmakesrecommendationstousers
aboutÞtnessorexerciseactivities.InthissectionIwilldescribethissystemanditsdesign
indetail.Figure5.1andAppendixGshowtheinterfaceforthissystem.
5.2.1IDARecommendationData
Icreatedalistof50exerciseactivitiesbyreferencingexistingcatalogsandselectingexercises
oractivitiesthatarewellknown.Inselectingexercises,Itriedtocreateabalancebetween
havingdiversityinthetypesofactivitiesthatcouldbechosenÐsothattheactivitiescould
bedi
"
erentiatedfromeachotherbydecisionmakersÐ,andcreatinganexhaustivelistof
activitiesthatwouldbeburdensomeforsubjectstobrowsewhenparticipatinginthestudy.
Aftercreatinganinitiallistofactivities,Ithenchoseasetofattributesbywhichthese
activitiescouldbeevaluated.Theseattributesare:
¥Cardio-Theamountofaerobicexerciserequiredbyanactivity.
¥Intensity-Howphysicallyorpsychologicallyintenseanactivityis.Lowintensity
activitiescanbethoughtofasrelaxing.
¥Group-Howmanypeopleareideallyneededforanactivity.Thehighestvaluemeans
atleast15people
¥LowerBody/Core-Thedegreetowhichanactivityprovidesexerciseforthelegs,
abdomen,orlowerback.
¥UpperBody-Thedegreetowhichanactivityprovidesexerciseforthearms,neck,
shoulders,orupperback.
¥Convenience-Theresourcesrequiredforanactivity.Inconvenientactivitiesrequirea
lotofmoney,equipment,time,orotherresources.
107Figure5.1:ExerciseRecommenderinterface.
108¥Di
!
culty-Theamountofskillorexperiencerequiredtoperformanactivity.Easy
activitiescanbecompletedbyanovice.Di
!
cultactivitiesrequiretrainingorexpertise
toperformoptimally.
¥Fun-Whetheranactivityisenjoyableornot.
Idevelopedthislistofattributesfortwopurposes.Onepurposewastocreateasetof
preferenceproÞlesthatrepresentasetofthingsthatauserofanIDAsuchasthismight
careaboutwhendecidingonanactivity.Thesecondpurposewastocreateattributesfor
theIDAÕscontent-basedrecommendersystemtousewhenmakingrecommendations.As
discussedinChapter2,content-basedrecommendersystemsmakerecommendationsbased
onexplicitly-knownattributesoftheitemsinthesystemÕscatalog.Bycreatingthislistof
attributes,IwasabletogivetheIDAexplicitcontentinformationabouttheexercisesin
ordertomakerecommendations.
Ichoseattributesthatwouldcreatesomediversitywithintheframeworkof
search
at-tributesversus
experience
attributes.Nelson(1970)developedthisframework,whichhasbe-
comewidelyadoptedinmarketingofconsumergoods,thatdistinguishesbetweenattributes
ofproductsforwhichinformationcanbeobtainedthroughsearchcomparedtothingsfor
whichinformationcanonlybeobtainedthroughexperience.Forexample,amattresshas
somesearchcharacteristicssuchasitssize,price,orbrandname.Thisinformationcanbe
obtainedeasilythroughasearch.However,amattressalsocontainsexperienceattributes
(e.g.comfort)whichforanyparticularpersonmayonlybeobtainedbyexperiencingthe
mattress.Thedistinctionbetweenthesetwotypesofattributesisimportantforrecom-
mendersystems(Ochi,Rao,Takayama,&Nass,2010).Usersmaybeabletoreasonably
expectasystemtohaveanduseinformationaboutsearchattributesoftherecommendable
items,butexperienceattributesmaybe,oratleastappeartousers,tobelesscompatible
withautomatedrecommendations.Sinceusersmaycareaboutmanyexperienceattributes,
butmayexpectthesystemtomaintainandusesearchattributes,Iincludedattributesthat
109Þtacrossthecontinuumofthisframework.The
Fun
attributeisastrongexampleofan
experienceattribute,whereasthemusclegroupsworkedbyanactivitycanbeeasilyob-
tainedthroughasearchorthroughbasicknowledgeofanactivity.Otherattributeslikethe
intensityordi
!cultymaybeunderstoodthroughsearchbutlikelyrequiresomeexperience
aswelltoobtaincompleteinformation.IDAusers,particularlyofcustomizablesystems
thatdemanduserinput,mayÞnditchallengingtocoercethesystemintoconsideringboth
searchandexperiencecharacteristics.Yetthischallengeisinherentinmanycontextsof
IDA-supporteddecisionmaking,andforthisreasonIchosetodiversifytheattributesthat
gointothepreferenceproÞlesandintothesystemÕscontent-basedrecommendationsasan
acknowledgementofthischallengingaspectofIDA-supporteddecisionmaking.
5.2.1.1CrowdsourcingExercise/AttributeEvaluations
Becausetheattributelistcontainedbothsearchandexperienceattributes,andbecauseI
wasunabletolocateareliablesourceofobjectiveevaluationsofexercisesforallsearch
attributes,Iconductedasurveytocrowdsourcetheevaluationofallexercisesagainstall
attributes.ThiscrowdsourcingservesasawaytobootstraptheIDAbyobtainingsome
initialcontentratingsbywhichtomakerecommendations.
Forthesurvey,112crowdworkerswereenlistedthroughAmazonMechanicalTurk,and
eachwaspaid$0.50forcompletingthesurvey.Thesurveyaskedthemtorate40exercise
activitiesthatweredividedinto4groupsof10,witheachgroupbeingratedaccordingtoa
di"
erentattribute.Forexample,asubjectwouldrateonegroupoftenexercisesonthe
Fun
attribute,thenanothergroupof10onthe
Di
"
cultyattributeandsoon.Thesurveywas
rununtilatleast10ratingshadbeenrecordedforeachofthe400exercise/attributepairs
(50exercisesby8attributes).Workersweregiventhedescriptionforeachattributethatis
listedabove.Theratingsforeachattributewereassessedona10-pointscale.AppendixD
showsanexampleofthissurvey.
Thesurveycontainedseveralmeasuresintendedtoensurehighdataquality.Workershad
110toanswerthreeÒattentioncheckÓquestionswheretheyweregivenspeciÞcinstructionsinthe
questionprompttoverifythattheywerereadingtheprompts.Anyworkerwhofailedanyof
theseattentioncheckquestionswasremovedfromtheÞnaldataset.Also,eachsubjectwas
showntwoÒrepeatsÓofquestionstheyhadpreviouslyansweredtodeterminewhetherthey
wouldbeconsistentabouttheirratings.Anyworkerwhoseanswertoarepeatedquestion
deviatedbymorethanonepointfromtheiroriginalanswerwasremovedfromtheÞnaldata
set.The10-pointscalealsoincludedanadditionaloptiontoindicatethattheworkerwas
notfamiliarwiththeactivityinquestion,andexercisesthatfrequentlyreceivedthisresponse
wereremoved.Afteralldatacleaning,83workersand44exercisesremainedinthedataset,
foratotalof3320ratings.
5.2.1.2LatentAttributes
Table5.1:Latentattributesfromexercisesurvey.
AttributeWorkoutIntensityWorkoutAtmosphereMuscleGroup
FactorLoadings
Cardio0.6250.343-0.213
Convenience-0.659
Di!culty0.520
Fun0.7090.196
Group0.534-0.207
Intensity0.8120.111
LowerBody0.531-0.415
UpperBody0.823
FactorFitMeasures
Sum-of-SquaredLoadings1.6091.3450.991
ProportionofVariance0.2010.1680.124
InsomeinitialusabilitytestingoftheExerciseRecommenderandthestudytask,I
determinedthateightattributesweretoomanyforuserstosimultaneouslyconsiderinan
explicitwaybothforusingandconÞguringtheExerciseRecommenderandformakinga
decisionthatmatchedthepreferenceproÞleforallattributes.Torespondtothisproblem,
111Iconductedanexploratoryfactoranalysisontheratingssurveydatatoderiveasmaller
numberoflatentfactors.
Inthisfactoranalysis,Icreatedanexercise-by-attributematrix,andineachcellofthe
matrixIinsertedthemedianratingforthatcombination.Theintentofthefactoranalysis
wastoderivesomelatentfactorsthatrepresentcombinationsoftheexplicitattributessuch
thatthecorrelationsbetweentheattributescouldbetranslatedintoameaningfulconstruct
orlabel.Inanexploratoryfactoranalysisusingvarimaxrotationsuchasthisone,the
researcherspeciÞesanumberoflatentfactorsforthealgorithmtoÞndsuchthatthevariance
betweenthedi
"
erentfactorsismaximized.IinitiallyspeciÞedfourfactors,andfoundfour
factorswithsu
!
cientfactorloadings.However,uponexaminingthespeciÞccorrelations
withinthesefactorsandthelistofexercisesthatwouldscorehighorlowonthesefactors,
thefourthfactor(whichbydesignofthefactoranalysisprocessexplainstheleastvariance)
didnotappeartohaveanidentiÞabletheme.Iwasunabletoputaclearlabeltothisfourth
factor,whichwouldmakeitconfusingtousersoftheExerciseRecommendertounderstand.
TheÞrstthreecategoriesdidhaveamoreclearthemehowever.Ithenranasecondfactor
analysissearchingforonlythreefactors,theanalysisreturnedthreefactorswiththesame
themeandgeneralcorrelationsastheoriginalanalysiswithfourfactors.Thisanalysisis
describedinTable5.1.IusedthissecondanalysistogeneratefactorscoresusingBartlettÕs
method(DiStefano,Zhu,&Mindrila,2009)foreachexercise/latentattributecombination.
Iaddedthelabel
WorkoutIntensity
totheÞrstfactor.Theintensityattributehadthe
highestloading,followedbycardio,lowerbody,anddi
!
culty.Activitiesthatscoredhighon
thisfactorwerethingslikeClimbingStairs,Jumprope,Basketball,Joggingforonehour,and
Boxing.ActivitieslikeStretching,Bowling,GolfandYogascoredlowonthisfactor.The
ExerciseRecommenderdescribesthisfactorasÒHowmuchwilltheactivitymakeyouwork
hard,breathehardandsweat.ÓUserscanspecifyalevelforthislatentattributebetween
ÒtakeiteasyÓandÒmakemesweat.Ó
Igavethesecondfactorthelabel
WorkoutAtmosphere
.Thisfactorhadhighloadingsfor
112theFun
andGroup
attributes,aswellasstrongnegativeloadingfor
Convenience
.Thisfac-
torappearstoidentifyactivitiesthatarefuntodoinagroupofpeoplesuchasrecreational
activities(Snorkeling,WhitewaterRafting,SquareDancing)andteamsports(Soccer,Ul-
timateFrisbee).Activitiesthataretypicallydonebyoneselfinagymscoredverylowin
thisfactor,suchasLunges,Squats,Planks,andBicepCurls.Tohelpusersunderstandthis
latentattribute,thehighlevelofthisattributewaslabeledintheExerciseRecommender
interfaceasÒhavefunwithfriendsÓandthelowlevelasÒlistentomusic.ÓThedescriptionof
thelatentattributewithintheinterfacewasÒFunsocialrecreationactivitiesversussolitary
workouts.Ó
ThestrongestfactorloadingsfortheÞnallatentattributewerethe
UpperBody
andLowerBody/Core
attributes,whichhadoppositesigns.Forthisreason,Ilabeledthislatent
attributeas
MuscleGroup
andlabeledthehighlevelasÒupperbodyÓandthelowlevelas
Òlowerbody/core.Ó
5.2.2CustomizingRecommendations
Oneofthemostimportantgoalsforthedesignofthissystemwastoallowuserstocustomize
therecommendations.Userscancustomizerecommendationsbyspecifyingwhatlevelof
eachofthethreelatentattributestheyprefer.Forexample,ifuserswantanexercisethat
willbeintenseandalsofun,theymightset
WorkoutIntensity
andWorkoutAtmosphere
tothehighersettings.Andiftheypreferacoreexercise,theycouldset
MuscleGroup
toalowersetting.TheExerciseRecommenderalsoallowsuserstoprioritizetheattributes.
Therecommenderalgorithmdescribedbelowcangivemoreweighttoitemsthatmatchsome
attributesthanmatchingonothers.Tofacilitatethisintheuserinterface,userscanmovethe
blocksthatcontaineachattributeupanddownthroughthelistandrearrangetheirorder.
Theattributeatthetopofthelistisgiventhehighestweightinthealgorithm,making
activitiesthatmatchcloselyonthisattributemorelikelytoappearintherecommendations.
Therecommenderalgorithmisbasedonaformulaforweightingqueriesinrecommender
113systemspresentedbySchafer,KonstanandReidl(2004).Thisformulacalculatesasimilarity
scorebetweenauserÕsqueryandtheitemsinthesystemÕscatalog.AuserÕsqueryconsists
oftwovaluesforeachofthethreeattributesforwhichuserscanspecifytheirpreferences.
TheÞrstvalueisthelevelthattheuserspeciÞesforthatattribute.Thesevaluesareona
scalebetween-2and2,whichapproximatelymatchestherangeofthecenteredfactorscores
storedforeachactivity/attributepairwithinthesystemÕsdatabase.Thesecondvalueisthe
rank(1,2,or3)thattheattributeissettoaspriorityforthatattribute.Thissecondvalue
issubtractedfrom4inordertogivethehighestrankedattributeavalueof3andthelowest
avalueof1.
Takingthisquery,theExerciseRecommendercalculatesthesimilaritybetweenthisquery
andeveryactivityinthesystemÕsdatabaseusingEquation5.1.
Similarity
(Activity,Query
)=1!!"""""#$a#attributes
w2a(1!da)2$a#attributes
w2a(5.1)Inthisequation,
drepresentsthedegreeofmatchbetweenthequeryandtheactivityÕs
rating(factorscore)forthatattribute.
discalculatedas
d=|QueryLevel
!Rating
|/3.Dividingby3sets
donascalebetween1and0whichisrequiredbytheequation.
wisthe
weightthattheattributeisbeinggivenbasedonthepriorityranking.Forexample,ifauser
setthe
MuscleGroup
attributetolevel2,andtheactivitybeingcalculatedhasascoreof1
forthatattribute,
dwouldbecalculatedas
|2!1|/3=0
.333.Ifthatattributewasranked
asmostimportant,
wwouldequal3intheequation.
Afterallsimilarityscoreshavebeencalculated,theactivitiesarethensortedbythese
scoresandthetopÞvearereturnedtobepresentedintheuserinterface.TheExercise
Recommenderwasbuiltasawebapplicationthatwasaccessedthroughawebbrowser.I
builtanddeployedthesystemusingtheweb.py
1framework,HTML5,JQuery1.10.2and
1http://webpy.org
114jQueryUI1.11.4
2,andtheSkeletonCSSframework
3.Thesetechnologiesenabledthe
systemtobeusablebothonatraditionaldesktopwebbrowseraswellasotherdeviceswith
smallerscreensandtouchdevices.
5.3Methods
IconductedanexperimentinwhichsubjectswoulduseaversionoftheExerciseRecom-
mendertohelpthemmakeadecisionaboutwhichexerciseactivitywouldbemostappropri-
ategiventheirpreferencesforthekindofactivitytheywant.Thestudycompareddecision
makingusingdi
"
erentversionsoftheExerciseRecommender,aswellasacontrolcondition
ofunaideddecisionmaking,inordertoanswertheResearchQuestionsdescribedabove.
5.3.1DecisionTask
InordertoevaluatethequalityofdecisionsmadebyusersoftheExerciseRecommender,I
createdadecisiontaskinwhichsubjectsweregivenasetofpreferencesforsomeattributes
ofexerciseactivitiesandwereaskedtochooseanactivitythatmostcloselymatchedthose
preferences.Animportantfeatureofthistaskisthatsubjectswerenotchoosingactivities
thattheywouldprefertodooutsideofthecontextofthestudyontheirown.Instead,they
werechoosingexercisesthatmatchedanassignedpreferenceproÞle.
Underthismethod,ratherthanhavingeachsubjectsupplytheirownpreferencesaswould
bemostnaturalistic,preferencesareassignedtosubjectsbytheresearchers.Igavesubjects
anincentiveintheformofadditionalpaymenttomakedecisionsthatmatchtheassigned
preferencesratherthantheirownpersonalpreferences.Byassigningallusersthesame
preferences,decisionmakingaboutwhatmightotherwisebeahorizontallydi
"
erentiated
setofpreferencescanbemadeverticallydi
"erentiatedandobjectivelyevaluated.This
2http://jqueryui.com
3http://getskeleton.com
115approachdrawsonmethodologyfromexperimentaleconomics,andspeciÞcallyonInduced
ValueTheory(Smith,1976).Smithshowedthatintheabsenceofotherincentives,subjects
inadecisionmakingexperimentcanbeassignedpreferencesforsomedecisionalternatives
overothersbyvaryingaÞnancialrewardfromtheexperimentfromchoosingthosemore
valuedalternatives.Forexample,ifsubjectsaretoldthatiftheoutcomeofagameisA,
theywillearn1pointandifitisoutcomeB,theywillearntwopoints,andthepointsare
laterexchangedforrealcash,subjectswillactuallypreferthegametohaveoutcomeBover
A,andwillmakedecisionsthattheybelievewillresultinB.
ThedecisiontaskIcreatedforthisstudymakesuseofInducedValueTheorybygiving
subjectsasetofpreferencesrepresentedaspointvaluesthatcanbeearnedfromachosen
exerciseactivity.SubjectswereshownÞveattributesofexercises.TheseattributeswereÞve
oftheeightattributesthatwereevaluatedintheexerciseratingsurveydescribedabove.The
attributesthatmadeupapreferenceproÞlewere
Cardio
,Convenience
,Fun
,Di
"
culty,and
Group
.Foreachoftheseattributes,subjectsweretoldtheyprefereitherahighoralowlevel,
andassignedapointvalueforthatattribute.Theratingsgiventotheactivitiesforthese
attributesestablishagroundtruth.Foreachattribute,amediansplitoftheratingsforthat
attributedeterminedwhethertheexercisewouldbeclassiÞedasÒhighÓorÒÕlowÓforthat
attribute.Whenasubjectselectedanactivity,thatactivityÕsgroundtruthhigh/lowratings
werecomparedtothesubjectÕspreferenceproÞle.SubjectsearnedthespeciÞednumberof
pointsiftheirselectedexercisehadthesamehigh/lowlevelastheirpreferredlevelfromthe
proÞle.
Forexample,ifasubjecthadthefollowingpreferenceproÞle:
¥HighinFun-30points
¥LowinCardio-20points
¥LowinGroup-15points
116¥HighinConvenience-10points
¥LowinDi
!culty-5points
andthesubjectchoseBowling,thesubjectwouldearn55pointsbecauseBowlingisratedas
Highin
Fun
(30points),Lowin
Cardio
(20points),Highin
Group
(0points,notmatched),
Lowin
Convenience
(0points,notmatched)andLowin
Di
"
culty(5points).
PreviousworkonIDAsine-commercehasusedasimilarapproachtoevaluatingdeci-
sionmaking(Pereira,2001)thatappropriatespointstousersbasedonthematchbetween
theirpreferenceforattributesandtheitemtheyhaveselected.Thereareafewimportant
di"
erencesbetweenPereiraÕsmethodandmyown.Pereiradidnoto
"
eranincentivefor
subjectstoearnmorepoints.Iconsiderthisincentivetobecriticalforastudyconducted
onlinebyMechanicalTurkworkers,whootherwisehavegreatincentivetoperformthetask
quicklywithouttryingexceptionallyhardtomakegooddecisions.Theseconddi
"
erenceis
thatPereiraaskedsubjectstocreatetheirownpreferenceproÞlespriortousingadecision
aidbyselectingtheirpreferredvaluesforeachattributethenweightingtheimportanceof
theattributes.Thismethodcreatessomeexternalvalidityforthedecisiontaskbecausesub-
jectsaremakingdecisionsfortheirownpreferences.However,therearesomeseriousthreats
tointernalvaliditywiththisapproachthatmakeitinappropriateformystudy.First,it
eliminatesthepossibilitytoevaluatewhethersubjectsimprovetheirdecisionmakingover
timebecauseasubjectcanhaveonlyoneproÞle.Second,itcreatesapotentialselection
bias.Sincesubjectschoosetheirownpreferences,itispossiblethatsomesubjectshave
preferencesthatareinherentlyabettermatchtothelimitedcatalogofoptionsthanothers.
Thismeansthatsomesubjectsmayhavemoreoptionsavailabletothemthanothersthat
wouldgivehighscores,andthemaximumpossiblescoremaybedi
"
erentforeachsubject.
Forthesereasons,Ichosetoassignusersthesamepreferences,leaningonInducedValue
TheorytoprovideassurancethattheassignedproÞlesdoinfactcreateatruepreferencefor
selectingactivitiesthatmatchthem.
117Alongwithothercolleagues,IhaveusedthisapproachbasedonInducedValueTheory
toassessdecisionmakingcrowdfundingsystems(Wash&Solomon,2014;Solomon,Ma,&
Wash,2015).ThatresearchtookadvantageofanotherimportantcharacteristicofInduced
ValueTheory,whichisthatitcanbeusedtoinducepreferencesinordertoevaluatestrategic
decisionmakingingamesplayedbetweenmultipleparties.Theapplicationofthisapproach
inthisstudyabouttheExerciseRecommenderismoresimple,asitinvolvesnostrategic
behavior.SubjectsÕpayoutsaredeterminedentirelybytheirowndecisions.Thisisauseful
featureofthistaskdesignbecauseitremovessomeexternalincentivesthatcancreepinto
inducedvalueexperiments,suchaslearningandnegotiatione
"
ectsthatcancomefrom
repeatedgames(Andreoni,1988).Instead,thequalityofsubjectsÕdecisionscanbequite
objectivelymeasuredasthenumberofpointsearnedfromtheiractivityselection.
TocreatethetenpreferenceproÞles,IÞrstclusteredtheexerciseactivitiesaccordingto
theirratingsfortheÞveattributesinthepreferencesusinghierarchicalagglomerativecluster
analysis.Icreatedtenclusters,thenchoseoneactivityfromeachclustertosetasaÒtop
choice.ÓIthensetthepreferredlevelofeachattributeintheproÞletoequaltheratingofthe
chosenexercise,sothateverypreferenceproÞlehadatleastoneactivitythatwouldscore
themaximumof80points.TheclusteranalysiscreateddiversityinthepreferenceproÞles,
andensuredthatalltypesofactivitiesthatcouldbechosenwererepresentedamongthe
possiblebestchoices.
5.3.2ExperimentConditions
Therewerefourconditionsofthestudy,whicheachrelatetodi
"erentversionsoftheEx-
erciseRecommender.Inthe
UnaidedDecision
condition,subjectsdidnotusetheExercise
Recommendertohelpthemmakeadecision.Inthe
CustomOnly
version,subjectscould
customizetheExerciseRecommenderÕsrecommendationalgorithmasdescribedabove,and
wereshownonlyrecommendationsthatwereproducedbytheircustomizedalgorithm.In
theNon-CustomOnly
condition,subjectsusedaversionoftheExerciseRecommenderthat
118wasnotcustomizable,butprovidedrecommendationsofequivalentqualitytousers.These
non-customizedrecommendationswillbedescribedindetailbelow.Inthe
Bothalgorithms
version,theExerciseRecommenderdisplayedrecommendationsproducedbytheuserscus-
tomizedalgorithm,aswellasrecommendationsproducedbythenon-customizedalgorithm.
Non-customizedrecommendations.
Customizationmaya
"ectdecisionmakingifusers
aree
"
ectiveatcustomizingthealgorithmandthereforeimprovetherecommendationsthat
areproduced.Animportantgoalofthisexperiment,andinunderstandingcustomization
biasanddecisionmaking,istoaccountforthee
"
ectthatrecommendationqualitymight
have.Recommendationqualitycanbeeasilymeasuredinthisdesignduetothenatureofthe
decisionmakingtask,andthereforeanalysescanaccountforit.However,itwasnevertheless
desirablethatrecommendationqualitybeasclosetoconstantaspossiblebetweenconditions.
Becauserecommendationqualityisadependentvariableandcannotbeassignedtosubjects
whocustomizetheIDA,thisisnotastraightforwardmanipulationintheexperimentdesign.
Tobalancethequalityofrecommendationsbetweenconditions,thenon-customizable
algorithmusedapoolofrecommendationsthathadbeenproducedinapilotstudyby
usersoftheCustom-onlysystem.Inthispilot,37subjectsparticipatedinthethe
CustomOnlyconditionofthestudy.WhentheExerciseRecommenderneededtodisplaysome
non-customizedrecommendationstoauser,itrandomlysampledarecommendationset
fromthispool.TheExerciseRecommenderselectedarecommendationsetthathadbeen
producedbyapilotsubjectusingthesamepreferenceproÞleasthecurrentsubjectrequesting
therecommendations.Thismethodofgivingnon-customizedrecommendationsresembles
collaborativeÞlteringinthatthesystemgivesrecommendationstoauserbasedonthe
actionsofothersimilarusers.
TheExerciseRecommenderdidnotgiveanexplanationforhowthenon-customized
algorithmworked,otherthantoindicatetousersinthe
Bothalgorithms
conditionthat
theywerenota
"
ectedbytheuserÕsinput.Therewereseveralreasonsfornotexplaining
thenon-customizedrecommendations.First,asadesignchoiceforanIDAthereislittle
119justiÞcationintheliteratureonexplanationsforgivingexplanationsinregardstodecision
making.ExplanationsdonothaveaclearbeneÞttodecisionmaking,andinsomecasesmay
harmfultodecisionmaking(Ehrlichetal.,2011).Asecondreasonisthatifanyexplanation
weretobegiven,itwouldcreateaconßictforexperimentalvalidity.Forexampleifusersin
thenon-customonlyconditionweretoldthatothersimilarusershadcustomizedtheIDA,it
wouldrevealthenatureoftheexperimenttothoseusersandcreateapotentialfordemand
e"ects.Ifafalseordeceptiveexplanationweregiven,itwouldviolatetheestablishednorms
ofexperimentaleconomicsuponwhichthedecisiontaskisbased.Furthermore,di
"erent
explanationsmayresultindi
"erente
"ects,ashasbeenfoundinseveralpreviousstudies(Lim
etal.,2009;Tintarev&Mastho
",2008,2012),whichmeansthatthespeciÞcexplanation
chosencouldpotentiallyhaveadirecte
"
ectontheresultsoftheexperiment.Evaluating
speciÞcexplanationdesignsisnotoneoftheresearchquestionsforthisstudy,althoughit
isanimportantareaforfutureresearchthatbuildso
"
oftheexperimentprotocolIhave
developedforthisstudy.
5.3.3Procedure
IrecruitedsubjectsfromAmazonMechanicalTurktoparticipateinastudyaboutExercise
Choices.Io
"ered$2.00asaguaranteedpaymentforparticipation,andtoldtheminthe
recruitmentpostthatthattheycouldearnabonuspaymentdependingontheirdecisions
inthestudy,withtheaveragebonuspaymentexpectedtobe$3.00.Afterenrollinginthe
study,subjectswereassignedtoaconditionandgiveninstructionsabouttheirtask.These
instructionscanbeseeninAppendixF.Theseinstructionsdescribedthedecisiontask,
includinghowmanypointswouldbeearnedanddescriptionsoftheattributestheywould
begivenpreferencesfor.Afterviewingtheinstructions,subjectstookaquizthatevaluated
theirunderstandingofthekeypointsoftheinstructions.Inparticular,thisquizrequired
subjectstodemonstratethattheyunderstoodhowtheirscoreswerecalculated,understood
theExerciseRecommenderandthewayitproducedrecommendations,andveriÞedthat
120subjectsunderstoodtheirincentivetomakegooddecisionsfortheirpreferenceproÞle.If
subjectsansweredanyquestionincorrectly,theywereshownthecorrectanswersandthen
redirectedtotheinstructionsforreview.Theythenhadtoretakethequizwiththequestions
andanswersslightlyalteredandreordered.Theyhadtopassthequizbeforemovingon,but
werefreetotakeitasmanytimesasneeded.Themediannumberofquizattemptswas2.
Aftercompletingthequiz,subjectswereshownapagewiththeirpreferences.Inthe
unaidedcondition,theywerealsoshownamenuwheretheycouldselectanexercisefor
thegivenpreferenceproÞle.InthethreeExerciseRecommenderconditions,afterviewing
theirpreferencestheyhadtoloadtheExerciseRecommenderintoaframeonthescreen.
Theythenusedthesystemtogetrecommendations,andoncerecommendationshadbeen
generatedthemenuappearedwheretheycouldmaketheirdecision.Recommendations
couldonlybeproducedonetimeperround,sothatalldecisionswerebasedonusingthe
IDAtoproduceonesetofrecommendations,ratherthanallowingforrepeatedrequestsfor
recommendations.
Aftermakingtheirdecision,subjectswereshowntheirscore,andremindedthatthe
maximumpossiblescorewas80pointssotheycouldgagehowclosetothebestpossible
decisiontheyhadmade.Subjectsthenproceededtothenextroundwhereanewsetof
preferenceswasshown.SubjectsusedthesameversionoftheExerciseRecommender(or
usednoIDA)forallroundsofthestudy.AfterÞnishingthetenroundsofthedecisiontask,
subjectsweredirectedtoapost-testquestionnaire.TheorderofpreferenceproÞlesthat
subjectsmadedecisionsforwasrandomizedforeachsubject.
Atotalof155subjectscompletedthestudy.Afteraninitialanalysisofthedata,I
removedtwosubjectswhocompletedthestudyfromthesameIPaddress,andÞvesubjects
whocompletedthedecision-makingportionofthestudyinlessthan3minutes.Inpilot
testingIdeterminedthatthreeminuteswastoofastforsubjectstohavebeenseriously
lookingattheirproÞlesandmakingthoughtfuldecisions,soIremovedfoursubjectswho
completedthestudyinlessthanthreeminutes.48%femalewithamedianagecategoryof
12126-34.Subjectstookanaverageof27minutestocompletethestudy.
5.3.4Measures
Decisionquality.
Imeasureddecisionqualityusingthenumberofpointsearnedbythechosenactivityfor
thesubjectÕspreferenceproÞle.
Recommendationquality.
TheExerciseRecommendergaveÞverecommendationsineachround,plusanadditional
ÞvetosubjectsintheBothAlgorithmcondition.Iusedthreedi
"
erentmeasurestoassess
thequalityofrecommendationsthatasubjectreceivedinagivenround.
AverageRecom-
mendationQuality
wastheaveragenumberofpointsthatallrecommendationsshownwould
earn.
WeightedRecommendationQuality
gaveextraweighttorecommendationshigheron
thelist.Theaveragewascalculatedbyaddingcopiesofeachrecommendationscoretothe
vector,withthenumberofcopiesdeterminedbytheactivityÕsrankintherecommendation
list.Thetopactivitywasrepeated5times,whiletheÞfthactivityappearedonlyonce.The
thirdmeasureisthescoreofthebestrecommendationonthepage.
Post-testmeasures.
Thepost-testquestionnaireassessedseveraladditionalmeasures.AppendixHshowsthe
fullpost-testquestionnaire.ThisquestionnaireassessedsubjectsÕknowledgeaboutÞtness
andexerciseactivities,theiruserexperiencewiththeExerciseRecommender,theirpercep-
tionoftransparencyoftheExerciseRecommender,theirpropensitytotrustdecisionaids,
anddemographicinformation.Thisquestionnairealsocontainedquestionsaboutwhether
theylookedat,considered,trustedorignoredrecommendationsfromtheExerciseRecom-
mender.
122Table5.2:PercentagesofpeopleansweringyestothesequestionsabouttheExerciseRec-
ommender.CustomOnlyNoncustomOnlyBoth
IusedtheExerciseRecommendertohelpme
makemydecision
92%66%94%
IcouldconÞguretheExerciseRecommender
toadjusttherecommendationsitgaveme
73%3%77%
TheExerciseRecommendergavemesome
recommendationsthatIhadnocontrolover
41%18%74%
TheExerciseRecommendergavemeonly
recommendationsthatIhadnocontrolover
11%45%3%
Table5.3:Questionnairequestionsaboutrecommendations.Answeredon5-pointLikert
scale(StronglyDisagreetoStronglyAgree).Standarddeviationsinparentheses.
CustomOnlyNoncustomOnlyBoth
Lookedatrecommendations4.2164.0264.429
(0.584)(0.944)(0.502)
Trustedrecommendations3.2972.9213.229
(0.996)(1.148)(0.843)
Ignoredrecommendations2.1892.5792.229
(0.811)(1.056)(0.843)
Lookedatcustomrecommendations4.054-3.971
(0.815)-(0.664)
Lookedatnon-customrecommendations-3.5523.857
-(0.921)(0.879)
5.4Results
5.4.1Questionnaire
Table5.2describessubjectsÕanswerstoquestionsabouttheExerciseRecommender.This
tableillustratesthatsubjectsintheBothAlgorithmsconditionunderstoodthenatureof
theirrecommendations,astheyfoundthesystemtobeconÞgurable,andthatsomebutnot
alloftheirrecommendationswerenotinßuencedbytheirconÞguration.
Table5.3describesthequestionnairequestionsaboutattentiontotheExerciseRecom-
123Table5.4:Self-reporteduserexperiencevariables.
CustomOnlyNoncustomOnlyBoth
Useful3.4593.0003.571
(1.016)(1.040)(0.884)
Accurate3.4323.0003.324
(0.987)(1.040)(0.843)
Easytouse4.1354.4214.118
(0.918)(0.793)(1.094)
ConÞgurable3.5411.9733.029
(0.931)(1.000(0.937)
menderÕsrecommendations.TherewerenostatisticallysigniÞcantdi
"
erencesbetweenany
conditionsonanyofthesemeasures.Almostallsubjectsreportedthattheylookedatand
consideredtheExerciseRecommenderÕssuggestions,withanaveragehigherthantheÒagreeÓ
pointandonlyafewsubjectsreportinganydisagreementwiththestatementaboutconsid-
eringrecommendations.Therefore,thedatainthethreeExerciseRecommenderconditions
canbelargelyinterpretedasIDA-supporteddecisionmaking.Furthermore,subjectsinthe
BothAlgorithmsconditionindicatedequivalentamountsoflookingandconsideringboth
typesofrecommendations.Thesmalldi
"
erenceinmeansbetweenlookingatcustomand
non-customrecommendationswithintheBothAlgorithmsgroupwasnotstatisticallysig-
niÞcant.Thissuggeststhatsubjectswereawareofallrecommendationsandlookedover
thembeforemakingdecisions,andthatthereisnodi
"
erencebetweenanyoftheExercise
Recommenderconditionsintermsofseeingandconsideringrecommendations.
Table5.4describesresponsestoquestionsabouttheuserexperienceoftheExercise
Recommender.ANOVAandTukeyHonestSigniÞcantDi
"
erencetestsfoundthatsubjects
intheNon-customconditionthoughttheIDAwaslessusefulthansubjectsintheBoth
Algorithmcondition.Also,subjectsintheNon-customconditionfoundthesystemtobe
lessconÞgurablethantheothertwoconditionswiththeIDA.Noothercomparisonsshowed
statisticallysigniÞcantdi
"erencesbetweenconditions.
ThequestionnaireaskedthreequestionstoassessthetransparencyoftheExerciseRec-
124Table5.5:Self-reportedunderstandingofIDAlogic.
CustomOnlyNoncustomOnlyBoth
Understoodwhy3.9193.3683.829
(0.682)(0.942)(0.664)
Madesense3.7843.2893.629
(0.672)(1.011)(0.770)
Logicwasclear3.7843.2113.600
(0.787)(1.094)(1.094)
ommender.Subjectsstatedona5-pointLikertScalewhethertheyunderstoodwhytheIDA
gavecertainrecommendations,whetherthoserecommendationsmadesense,andwhetherthe
systemÕslogicwasclear.Table5.5describedtheseresults.ANOVAandTukeyÕsHSDtests
wereruntodetermineanydi
"
erencesbetweenconditionsontheseitems.TheNon-custom
OnlyconditionwasratedashavinglesstransparencyonallthreemeasuresthantheCustom
Onlycondition,and(
p<.05),andtheNon-customOnlyconditionwasalsoratedasless
transparentthentheBothAlgorithmconditionontheÒunderstoodwhyÓitem.ThisÞnding
isnotsurprisingastheNon-customrecommendationsweregivennoexplanationabouttheir
process,howeveritisnoteworthyinthatito
"
erssomeevidencethatusersperceivetheact
ofcustomizingspeciÞcinputsasaformoftransparency.Itshouldbenotedhoweverthatthe
e"ectsizesinthesetestsarerelativelysmall,andallmeanssatbetweentheÒneutralÓand
ÒagreeÓpointsonthescale,suggestingthatsubjectsdidnotÞndtheIDAhighlytransparent
overall.
5.4.2RecommendationQuality
Table5.6describesthemeansandstandarddeviationsofthethreerecommendationquality
measureswithineachcondition.IconductedANOVAÕsoneachofthesemeasuresbycon-
dition.ThesetestsfoundnostatisticallysigniÞcantdi
"
erencesbetweenconditionsonthe
measuresofaveragerecommendationquality,weightedaveragerecommendationquality,and
number1recommendationscore.TherewasastatisticallysigniÞcantdi
"
erence(
p<.05)125Table5.6:Meansandtandarddeviationsofrecommendationqualitywithinconditions.
Non-customonlyCustomonlyBothalgorithms
AverageRec.Quality45.013(15.238)46.486(15.210)46.279(14.843)
WeightedAverageRec.Quality43.757(16.250)45.014(16.088)44.710(15.999)
TopRec.Score64.854(15.546)65.897(15.210)72.500(11.959)
onthemeasureofbestrecommendationscore,whichevaluatedhowmanypointsthebest
recommendation(appearinganywhereonthepage)wouldearn.Furtheranalysisusinga
multilevelregressionmodelindicatedthattheBothAlgorithmsconditionhadahigherav-
erageforthebestrecommendationthatappearedonthepage.Thiscanlikelybeexplained
bythefactthattheBothAlgorithmsconditionshowedtenrecommendationscomparedto
ÞveintheotherIDAdecisions.Figure5.4showsthedistributionofthismeasure.Inmost
rounds,thesystemgaveatleastonerecommendationthatscored80points,whichwasthe
maximumpossibleineveryround.
Overall,thestudydesignwassuccessfulatprovidingrecommendationsofequivalent
qualitytosubjectsinallthreeIDA-supportedconditions.Customizedrecommendationswere
equivalentinbothconditionsthatusedthem,aswerenon-customizedrecommendations.
5.4.3DecisionMaking
Toevaluatedecisionmaking,IÞtamultilevellinearmodelthatincludeddummyvariables
foreachrecommendationtype,theinteractionbetweenrecommendationtypes,andrandom
e"ectsforeachsubjecttoaccountfortherepeatedobservationspersubject.Themodel
alsoincludedaÞxede
"ectforeachproÞle.IusedÞxede
"ectsbecauseIwasinterested
inthespeciÞce
"
ectofeachproÞleinthisstudy,andnotanye
"
ectsthatmightgeneralize
toabroaderpopulationofpreferenceproÞlesforexerciseactivities.However,inapost-
hoctestIcheckedthismodel(andallothersinthischapter)usingrandome
"
ectsforthe
proÞles,andtherewasnosubstantivedi
"
erencethatwouldalteranyconclusions.The
interceptofthismodelrepresentsadecisionmadeintheUnaidedconditioninproÞlenumber
126050100150020406080Weighted Recommendation Quality
countBoth Algorithms
Custom OnlyNoncustom OnlyFigure5.2:Distributionofweightedrecommendationscores.
0501001502002500255075Score of #1 RecommendationcountBoth Algorithms
Custom OnlyNoncustom OnlyFigure5.3:Distributionofrecommendationscoreslistedas#1onthepage.
127Table5.7:Thee
"
ectofeachrecommendationtypeondecisionquality.
Dependentvariable
Points
Intercept59.302
"""(1.815)CustomizedRecs.Shown0.462
(1.492)Non-customizedRecs.Shown
!0.500(1.482)BothRecommendationTypes
!0.811(2.125)ProÞle2
!7.230"""(2.209)ProÞle33.074
(2.209)ProÞle4
!14.662"""(2.209)ProÞle5
!10.541"""(2.209)ProÞle64.527
""(2.209)ProÞle7
!4.932""(2.209)ProÞle8
!16.385"""(2.209)ProÞle9
!8.108"""(2.209)ProÞle10
!7.838"""(2.209)RandomE
"ectStandardDeviation2.370
LogLikelihood
!6,444.970LogLikelihood
!20.8921Note:
"p<0.1;""p<0.05;"""p<0.01128010020030040050020406080Top Recommendation Quality
countBoth Algorithms
Custom OnlyNoncustom OnlyFigure5.4:Distributionofscoresofbestrecommendationonthepage.
1.Thecoe
!cientlabeledÒBothRecommendationTypesÓrepresentstheinteractione
"ectbetweenseeingcustomizedandnon-customizedrecommendations.Thismodelisdescribed
inTable5.7.
ThemodelsuggeststhattheExerciseRecommenderingeneraldidnothaveane
"
ecton
theaveragedecisionqualityforsubjects,astherewerenostatisticallysigniÞcantdi
"
erences
betweentheIDA-supporteddecisionsandtheUnaidedDecisioncondition.Whilesome
proÞlesappearedtobemorechallengingthanothers,theExerciseRecommenderinanyform
didnothelpsubjectsmakebetterdecisionsonaverage.Iconductedarepeatedmeasures
ANOVAonthismodeltotestwhethertherewereanydi
"
erencesbetweenanyconditionsin
decisionquality.ThistestfoundnostatisticallysigniÞcantdi
"
erencesbetweentheconditions
ofthestudy(
F(3,144)=0
.287)ondecisionquality.Forthisreason,thereisnosupport
forH1(theIDAwillimprovedecisionmaking),H2(customizationwillimprovedecision
making),orH4(showingbothtypesofrecommendationswillleadtothebestdecisions).
AlthoughtheExerciseRecommenderdidnotimprovedecisionsonaverage,itdidinßu-
129Table5.8:DecisionmakingwithinonlytheIDA-supportedconditions.
Dependentvariable:
Pointsearnedfromchosenactivity
Intercept(Bothcondition)28.994
"""(4.358)CustomOnlyCondition5.442
(4.553)Non-customOnlyCondition7.327
(4.464)Avg.RecommendationQuality0.530
"""(0.075)CustomxRec.Quality
!0.089(0.093)Non-customxRec.Quality
!0.141(0.092)ProÞle2
!1.235(2.554)ProÞle32.858
(2.513)ProÞle4
!3.869(2.792)ProÞle5
!3.886(2.611)ProÞle62.945
(2.510)ProÞle7
!0.895(2.517)ProÞle8
!7.568"""(2.594)ProÞle9
!1.076(2.665)ProÞle100.358
(2.763)RandomE
"ectsStandardDeviation1.774
LogLikelihood
!4,726.461LogLikelihood
!2267.16"""Note:
"p<0.1;""p<0.05;"""p<0.0113002040608020406080Average Recommendation Quality
ScoreFigure5.5:E
"
ectofrecommendationqualityondecisionquality
encedecisions.ThequalityofrecommendationsgivenbytheIDAhadastronginßuence
onthequalityofdecisionsforusersinallthreeIDA-supportedconditions.IÞtamultilevel
regressionmodeltothedatafromthethreeIDA-supportedconditionsandincludedthe
averagequalityofrecommendationsonthepageasacovariate.Thismodelisdescribedin
Table5.8.Theresultsweresimilarforallmeasuresofrecommendationquality.Figure5.5
illustratestherelationshipbetweenrecommendationqualityanddecisionquality.Forevery
additionalpointofrecommendationquality,subjectscouldexpecttoearnanadditionalhalf
apointfromtheirdecision.
ThisÞndingsuggeststhattheIDAchangedthedecisionprocessforsubjectswhousedit
whencomparedtothosewhomadeunaideddecisions.Evaluatingrecommendationsmade
bytheIDAisanimportantpartofdecisionmaking,eventhoughitmaynotactuallyimprove
overalldecisions.
131Table5.9:ModelofagreementwithoneoftherecommendationsinCustomonlyandNon-
customonlyconditions.
Dependentvariable:
Logoddsofagreement
Intercept(Customonly)
!2.316"""(0.555)Non-customonly
!0.536""(0.270)TrustPropensity0.496
"""(0.158)Avg.Rec.Quality0.055
"""(0.008)ProÞle21.441
"""(0.409)ProÞle30.641
(0.395)ProÞle40.237
(0.416)ProÞle51.025
""(0.404)ProÞle60.407
(0.390)ProÞle71.228
"""(0.414)ProÞle82.328
"""(0.459)ProÞle91.685
"""(0.420)ProÞle100.713
"(0.412)RandomE
"ectsStandardDeviation0.801
LogLikelihood
!393.757LogLikelihood
!2137.78"""Psuedo
R20.289Note:
"p<0.1;""p<0.05;"""p<0.011325.4.4Agreement
Table5.9describesamodelthatestimatesthelikelihoodofagreeingwithoneoftherec-
ommendationsgivenbytheIDAbysubjectsintheCustomonlyandNon-customonly
conditions(H3).TherewasastatisticallysigniÞcantdi
"
erencebetweenthesetwoconditions
inagreementwithrecommendations.Subjectswhocustomizedthesystemweremorelikely
tofollowoneofitsrecommendationsthansubjectswhodidnotcustomizetherecommen-
dations.TheseresultssupportH3,ando
"
erfurthersupportthatcustomizationcreatesa
decisionmakingbiasbyIDAusers.
Presentingbothcustomizedandnon-customizedrecommendationsdidleadtoastatis-
ticallysigniÞcantreductionincustomizationbias,whichmeansthereisnosupportforH5.
Inamodelthatestimatedthelikelihoodofagreeingwithcustomrecommendationsinthe
CustomOnlyandBothAlgorithmscondition,therewasnostatisticallysigniÞcantdi
"
erence
betweentheconditionsinagreementwiththecustomrecommendations.
Recommendationqualityalsohadane
"ectonagreementwithrecommendations.Sub-
jectsweremorelikelytoagreewithrecommendationswhentheyhadhigherquality.This
relationshipheldtrueforallmeasuresofrecommendationquality.Likewise,subjectswho
reportedahigherdegreeofoveralltrustindecisionaidsweremorelikelytoagreewith
recommendations.TheseÞndingsarenotsurprising,buttheyarenoteworthybecauseI
observedcustomizationbiaswhilecontrollingfortheseotherimportantfactors.Asubject
whocustomizedtheIDAwouldbemorelikelytoagreewithoneofthemthansomeoneelse
whohasthesamepropensitytotrustdecisionaidsandreceivedrecommendationsofequal
quality.
TheBothalgorithmsconditiono
"
ersawithin-subjectsversionofthistestofcustomiza-
tionbias,assubjectssawbothtypesofrecommendationsimultaneously.Ascanbeseenin
Figure5.6,subjectsweremuchmorelikelytoagreewiththeircustomizedrecommendations
thanthenon-customizedrecommendations.Subjectschoseanactivityfromthecustomized
1330%20%40%60%80%Custom onlyNon!custom onlyBoth (custom rec.)Both (non!custom rec.)Agreement with RecommendationsFigure5.6:Agreementwithrecommendations.
listin67%ofallrounds,comparedtochoosinganactivityonthenon-customizedlistonly
37%ofthetime.ApairedMann-WhitneyU-testconÞrmedthisdi
"
erencewasstatistically
signiÞcant(
p<.001).Itshouldbenotedthatin27%ofrounds,thechosenactivityappeared
onbothlistsatthesametime.Whensubjectschoseanactivitythatappearedononlyone
list,79%ofthetimeitwasacustomrecommendationand21%itwasfromthenon-custom
list.
Figure5.7showshowsubjectsineachoftheIDAconditionsadjustedtheiragreement
withtheIDAoverthecourseofthetenroundsoftheexperiment.InthisÞgure,thebolded
linerepresentsamovingaveragesmoothedusingLOESS(LocalRegression)smoothing.The
lightdottedlinesaretheactualaveragesforeachconditionateachroundnumber.
ThisÞgureillustratesthateachconditionhadadistinctpatternofagreementovertheten
rounds.SubjectsintheNon-customconditionmaintainedastationaryamountofagreement,
withtheaveragebarelymovingatallandnottrendinginanydirection.Subjectsinthe
CustomonlyconditionstartedathigheragreementthantheNon-customsubjects,andtheir
1340.50.60.70.812345678910Round number
Average agreement
BothCustomNon!customFigure5.7:Changeinaverageagreementoverthe10roundsoftheexperiment.
Table5.10:Transparencymeasuresbycondition.Itemsareona5-pointscale.
Non-customonlyCustomonlyBothalgorithms
1.Iunderstoodwhy
thesuggestionsweremade
3.368(0.942)3.919(0.682)3.829(0.664)
2.Ithoughtthe
suggestionsmadesense
3.289(1.011)3.784(0.672)3.629(0.770)
3.Thelogicbehind
therecommendationswasclear
3.211(1.094)3.784(0.787)3.600(0.775)
agreementincreasedfortheÞrstseveralroundsbeforebeginningagradualdecline.Subjects
intheBothAlgorithmsconditionhadanoppositepatternfromtheCustomonlygroup.Their
initialagreementwasalsohigh,buttheyimmediatelyshowedadeclineintrust.However,
inthelaterroundstheyregainedthattrustandbytheendoftheexperiment,thesesubjects
wereshowingthemostagreementwiththeIDA.
1355.4.5Transparency
Table5.10showstheresponsestoquestionsrelatedtothetransparencyoftheExercise
Recommender.IconductedanANOVAforeachitemfollowedbyTukeyÕsPost-Hocteststo
testforpairwisedi
"
erencesbetweenconditions.AllthreeANOVAtestsindicatedthatthe
modelwasstatisticallysigniÞcant(
p<.
05).Subjectsreportedless
understandingwhythe
suggestionsweremade
(Item1)thanbothoftheothertwoconditions(
p<.05)accordingto
theTukeyÕstestwithHonestSigniÞcantDi
"
erencestoaccountforthemultiplecomparisons.
Foritems2and3,onlythedi
"
erencebetweentheNon-customonlyandtheCustomonly
conditionswasstatisticallysigniÞcant(
p>.05).TheseresultssupportH6andsuggestthat
customizationcanservetoincreasethetransparencyofanIDA.Thesedi
"
erencesarenot
particularlylargehowever,suggestingthatcustomizationmayonlyhaveamodeste
"
ecton
transparency.
Table5.11:Relationshipbetweenperceivedtransparencyandagreementwithrecommenda-
tions.
Dependentvariable:
Logoddsofagreement
(1)(2)(3)
Item1(understoodwhy)0.401
"""(0.143)Item2(madesense)0.314
""(0.134)Item3(logicclear)0.335
"""(0.124)RandomE
"
ectsStandardDeviation0.8340.8510.837
LogLikelihood
!594.605!595.604!594.737LogLikelihood
!2155.85"""153.87"""155.59"""Pseudo
R20.2600.2600.260
Note:
"p<0.1;""p<0.05;"""p<0.011360.250.500.751.0012345Transparency item response
Probability of agreementLogic clearMade senseUnderstood why
Figure5.8:E
"
ectoftransparencyonagreementwithrecommendations.
TheseitemsarenottestsofsubjectsÕactualunderstandingofhowtheExerciseRecom-
menderworks,ratherameasureoftheirperceptionofitstransparency.Toevaluatehow
subjectsÕperceptionoftransparencya
"
ectedtheiragreementwithrecommendations,IÞt
threemodelsthatestimatedthelogoddsofagreementwithrecommendationswithoneof
thetransparencyitemsasanexplanatoryvariable.Thesemodelsalsoincludedtheexperi-
mentconditionandaveragerecommendationqualityascovariates,alongwithÞxede
"
ects
fortheproÞlesandarandome
"ectforeachsubject.IÞtaseparatemodelusingeachofthe
threeitems,ratherthanincludingthemallinonemodel,becausethethreeitemsarehighly
correlatedwitheachotherandwouldcreatemulticollinearity.Similarly,propensitytotrust
decisionaidswasmoderatelycorrelatedwiththethreetransparencyitems(
r=0.448,0.449,
and0.427)andsoitwasnotincludedinthesemodelseventhoughinpreviousmodelsI
foundittobeastrongestimatorofagreement.
Table5.11describesthesemodels.Notethatthistableexcludestheothercovariates
inthemodelforbrevity,asthoserelationshipsarereportedinTable5.9.Thesemodels
137indicatethatholdingallothervariablesconstant,subjectswhoreportedperceivinghigher
transparencyoftheExerciseRecommenderweremorelikelytoagreewithrecommendations.
Thecoe
!
cientofeachofthethreetransparencyitemswasstatisticallysigniÞcant,evenwhen
controllingfortheconditionthesubjectwasin.Fromthesemodels,wecanconcludethat
thereissupportforH7inthatthereisanimportantrelationshipbetweenusersÕperception
oftransparencyofanIDAandtheirlikelihoodoffollowingitsrecommendations.
Thisrelationshipisimportantbecauseitisobservableaftercontrollingforrecommenda-
tionqualityandforcustomization.Thissuggeststhataperceptionoftransparencycreates
anadditionaldecisionmakingbias.Twouserswiththesamepreferencesandgivenequiva-
lentrecommendationsshouldhaveequivalentagreementwiththerecommendations.Yetthis
modeldemonstratesthatthoseuserswhofelttherecommendationsweremoretransparent
weremorelikelytofollowthem.
5.5Discussion
TosummarizetheÞndingsofthisstudy:
¥Therewerenodi
"erencesindecisionmakingquality(pointsearned)betweenanyof
thefourconditions.TheExerciseRecommenderdidnothelpsubjectsmakebetter
decisions,norwasthereanye
"
ectofthetypeofrecommendationdisplayed(orthe
combinationofbothtypes).
¥SubjectswhocustomizedtheIDAweremorelikelytochooseoneofitssuggestions
thanthosewhodidnotcustomize,evenwhencontrollingforthequalityoftherec-
ommendations.Thisillustratesthatcustomizationbiasextendstosystemswithhigh
controllability.
¥Thethreedi
"
erentIDAinterfacesledtodi
"
erentpatternsinagreementovertime.
Non-customonlyrecommendationsledtoastationarypattern,wheresubjectsdid
138Table5.12:Summaryofresults.
Hypothesis
ResultH1UserswhousetheIDAwillmakebetterdecisionsthan
otherswhomakethedecisionunaided.
Notsupported
H2Userswhoseecustomizedrecommendationswillmake
betterdecisionsthanuserswhoseeonlynon-customized
recommendations.
Notsupported
H3Userswhoseeonlycustomrecommendationswillhave
moreagreementthanuserswhoseeonlynon-custom
recommendations.
Supported
H4Userswhoseebothcustomandnon-customrecommen-
dationswillmakethebestdecisionsoverall.
Notsupported
H5Userswhoseebothcustomandnon-customrecommen-
dationswillhavelessagreementwithrecommendations
thanuserswhoseeonlycustomrecommendations
Notsupported
H6Customizingthesystemwillmakeusersfeelthesystem
ismoretransparent
Supported
H7ThemoretransparentusersfeelthesystemÕslogicis,
themoretheywillagreewithrecommendations
Supported
notchangetheiraverageagreementovertime.Customonlyrecommendationsled
toinitialgrowthinagreement,followedbyaslowdecline.Showingbothtypesof
recommendationledsubjectstoslowlyreducetheiragreementinearlyrounds,then
increasetheiragreementlateron.
¥SubjectswhousedtheNon-customonlyversionoftheIDAperceivedlesstransparency
thanthosewhohadcustomizedtheIDA.
¥SubjectswhoreportedthattheybetterunderstoodthelogicbehindtheIDAÕsrecom-
mendationsweremorelikelytoagreewiththerecommendations,regardlessofwhich
conditiontheywerein.
Thelackofanydi
"
erenceindecisionqualitybetweentheUnaidedandtheIDA-supported
groupswassurprising,althoughitisconsistentwithmanyappliedstudiesofIDAe
"
ectiveness
(Brightetal.,2012).Demonstratingthee
!
cacyofIDAsforimprovingaggregatemeasures
139ofdecisionmakinghasproventobeachallenge,andtheresultsofthisstudyarefurther
evidencethatdecisionaidswillnotautomaticallyleadtobetterdecisions.
Whatisapparent,bothfromthisstudyandotherworkondecisionaids,isthatIDAs
doalterthedecisionmakingprocess,eventhoughtheymaynotimproveit.Thestrong
e"ectofrecommendationqualityondecisionqualityshowsthatwhengivenadecisionaid,
usersdouseittoinformtheirdecisions.SubjectsinalltheIDA-supportedconditionsrelied
onthedecisionaidtomakedecisions,andthereforethequalityofrecommendationswas
astrongdeterminantoftheirscorefortheround.Bygivingsuggestionsforthedecision,
IDAusersmustevaluatethosesuggestionsaspartofmakingtheirdecision.Evaluating
thesesuggestionsmayaltertheoveralldecisionprocess.Forexample,itmayleadtousers
toevaluatejustthesuggestedalternativesandnotconsiderothers,usingtheIDAasan
initialÞlter.Inthiscase,suggestedalternativesmaybegivengreaterscrutinythanthose
alternativeswouldreceiveinunaideddecisionmaking,andthiscouldmakeusersbetter
discerningofthequalityofthosesuggestions.However,ifthesystemdoesnotrecommend
thebestpossibleoptions,usersmaynotconsiderthem,andthereforeoverallthesystem
wouldnotbeoptimallye
"
ectiveinrelationtounaideddecisionmaking.Alimitationofthis
studyisthatitdidnotexploredecisionmakingstrategiesorrecommendationevaluation
processindetail.Thatisanimportantdirectionforfutureresearchthatwouldbeavaluable
contributiontounderstandingIDAe
"ectivenessandinformingdesign.Thisstudyo
"ers
supportforahypothesisthatIDA-supportedandUnaideddecisionmakinginvolveseparate
processes.IntroneandIandoli(2014)foundthattherewasnorelationshipbetweensubjects
performanceinaIDA-supporteddecisiontaskandthesameindividualÕsperformancewhen
makingthedecisionunaided.Thissuggeststhattheremaybeseparateskillsandseparate
processesthatdeterminewhetherapersonwillmakegoodevaluationsofrecommendations
thanwhethertheywillmakegooddecisions.ThefactthatIDA-supportedusersrelied
heavilyontherecommendations,yetperformednobetterthanunaidedsubjects,further
supportsthisÞnding.Inordertobuildmoree
"
ectiveIDAs,futureresearchisneededto
140betterunderstandtheprocessesinvolvedinevaluatingIDAoutputandhowthatÞtsinto
theoveralldecision-makingprocess.
Thisstudyo
"ersmoreevidencethattheactofcustomizinganIDAalgorithmbiasesusers
toacceptingitsrecommendations.Again,subjectswhocustomizedtheIDAweremorelikely
tofollowoneofitssuggestionsthanthosewhoreceivedonlynon-customrecommendations,
evenwhenaccountingforthequalityoftherecommendations.Thisisanimportantaddition
totheÞndingsdiscussedinmypreviouswork(Solomon,2014)andinChapter3,asitshows
thatthiscustomizationbiascanhappenwhenusingamoreresponsiveIDAthanwasused
inthosestudies.ItalsogivesevidenceagainsttheIllusionofControlasamechanismof
customizationbias.AstheIllusionofControlcanonlybeobservedincaseswherethereis
littleactualcontrol(Ginoetal.,2011),thisstudyrefutesthattheoryasanexplanation.In
fact,thisstudy,alongwiththeÞndingsofthepreviousstudies,suggestthatcustomization
biasismostlikelytohappenwhentheIDAappearstobetrulyresponsivetousersÕinput.
Thissuggeststhatcustomizationbiasmaybetheresultofusersfeelingthattheyhave
successfullyexertedcontroloverthesystem,orthattheywereabletosuccessfullyproduce
reasonable-lookingrecommendations.
Thedi
"erentpatternsofagreementshownbetweenconditionsovertimeillustratesan-
otherdi
!
cultywithdesigninge
"
ectiveIDAs.AscustomizationledtoaninvertedU-shape
pattern,withsubjectsinitiallyincreasingtheiragreement,followedbyadeclinelateron.If
customizationweretobeusedasane
"orttoincreaseagreementwithrecommendations,the
e
"
ectmaybeshort-lived.Itisunclearwhyusersofthesystemwithbothrecommendation
typesshowedtheoppositepattern,wheretheyinitiallydecreasedtheiragreementbutlater
begantoagreemoreoftenwiththeIDA.Thisdi
"
erenceisunclear,butnoteworthybecause
inbothconditionssubjectscustomizedthesystem.Itsuggeststhatthedecliningpatternof
agreementwithcustomizedrecommendationscanbealteredbydi
"erentsystemdesigns.
AnimportantÞndingfromthisstudyistheobservedrelationshipbetweentransparency
andagreementwithrecommendations.Themorethatusersbelievedthattheyunderstand
141howthesystemproducesrecommendations,themorelikelytheyweretoagreewiththem.
ThisisanotherexampleofadecisionmakingbiasbyIDAusers.Thisrelationshipheldtrue
evenwhenaccountingforrecommendationqualityandforthecustomizabilityofthesystem
(whichIhaveshowna
"
ectbothagreementandtransparency).Therefore,Iarguethatthis
isevidenceofatransparencybias.SomeusersmayhavebeenactivelytryingÞgureout
howtheIDAproducesrecommendations,whereasothersmaynothavetriedtoactivelysee
intothesystemorgivenanythoughttohowrecommendationsareproduced.Someother
possibilitiesarethatdi
"
erentusersbringdi
"
erentmentalmodelsandexperienceusingthese
kindsofsystems,andtheymayusethatexperienceorthosementalmodelstoinferhowthe
systemworks,withthisexperiencealsomakingthemmoreorlesstrustingofthesystem.
Thecorrelationthatwasfoundbetweenpropensitytotrustautomateddecisionaidsand
perceivedtransparencyo
"
erssomeevidencethatcouldsupportthisexplanation.
ItisalsopossiblethatusersÕpre-dispositiontotrustingordistrustingdecisionaidsmay
leadthemtoe
"
ectivelydecidewhethertheywillagreewiththeIDAbeforetheyreceivethe
recommendation,andtheirevaluationofrecommendationscouldconsistofÞndingreasons
toÒrationalizeÓthisdecision.IftheycanÞndareasontojustifytheirdecision,theymay
claimthatthesystemÕsrecommendationsaremoreclear.
Overall,thisÞndingisconsistentwithexistingliteratureontransparency(Sinha&
Swearingen,2002;Tintarev&Mastho
"
,2008;Crameretal.,2008).However,animpor-
tantadditionhereisthatunlikepreviousstudies,transparencywasnotcreatedbyusing
explanations.Infacttherewaslittlee
"
orttocreatetransparencyinthedesignoftheIDA,
suggestingthattherecouldbeanimportantindividualdi
"erenceinhowpeopleinterpretthe
logicbehindanIDAsrecommendations.Somepeoplemayfeeltheycaneasilyunderstand
recommendationswithoutexplicitexplanations,whereasothersmayneedexplanationsor
otherindicatorswithinthedesign.
Andunlikepreviouswork,thee
"
ectoftransparencyinthisstudycanbeconsideredto
beabiasbecauseitwasobservedevenwhenholdingthequalityofrecommendationsand
142andtheexperimentconditionconstantinthestatisticalmodel.Althoughthisisonlyevi-
denceofanassociationbetweentransparencyandagreement,thereissomeintuitiontoa
hypothesisthatgreatertransparencywouldcausegreateragreementwithrecommendations.
Ifusersbelievetheyunderstandwhyarecommendationwasgiven,itmayprovideconÞdence
thattherecommendationisnottheoutputofarandomprocessbutthatitrepresentstrue
insightintothedecision.However,alimitationofthisstudyisthatitwasnotdesigned
explicitlydeterminewhethertransparencycausesagreement.Analternateexplanationmay
bethatsubjectswhoagreedmoreoftenwithrecommendationsmayhavetriedtojustifythis
agreementwhenpromptedabouttransparency,evenifitdidnota
"
ecttheirdecisionsduring
theexperiment.Animportantareaforfutureresearchwillbetodeterminewhethertrans-
parencycausesagreementandwhethervariationsinthedesignofanIDAa
"
ectagreement
bywayofalteringitstransparency.Iftransparencycausesuserstofollowrecommendations,
ito
"
ersdesignersanothertargetforengineeringagreementinIDAs.
Asdiscussedinchapter3,thereisapparentsimilaritybetweenconsistencybiasand
transparencybiasduetothefactthatconsistencybiasdependsonthesystemhavingsome
transparency.Itispossiblethatthesebiasesarebothmanifestationsofasingleunderlying
e"ect.Oneimportantconsiderationhoweveristhatinthisstudy,therewasanassoci-
ationbetweentransparencyandagreement
withinthenon-customonlycondition.Since
consistencybiasasdescribedinchapter3requiresthatusershaveknowledgeaboutthecon-
Þgurationofthesystem,theÞndingfromthepresentstudywouldappeartobedistinctfrom
thate
"
ectbecausesubjectsinthenon-customonlyconditionhadnoinformationaboutthe
conÞgurationofthesystem.
5.5.1Limitations
Onelimitationofthisstudyisthatunlikemostrecommendersystems,subjectswerenot
selectingitemsthatmatchtheirownpreferencesbutrathertomatchanassignedpreference
proÞlethatincludedonlyasmallnumberofattributes.Therefore,thisdecisiontaskmay
143engageadi
"erenttypeofdecisionmakingprocessthanwhatisdonebyusersinnaturalistic
settings.Thisstudyusedadecisiontaskthatwouldnaturallyhavealotofhorizontal
di"
erentiationinthatdi
"
erentuserswouldhavewidelyvaryingpreferences.Someworkon
recommendersystems(H¬aubl&Murray,2003)hasshownthatthesesystemscannotonly
Þnditemsthatmatchusersexistingpreferences,butthatthesystemitselfcanpersuade
userstoformnewpreferences.Thedesignofthisstudydoesnotallowforthataspectof
interactionwithIDAsbecausepreferenceswerepre-determinedandÞxed.
Anotherlimitationofthisstudyisthatuserswereshowntheirscoresbetweenrounds,
andtheninapost-testwereaskedaboutthetransparencyofthesystem.Oneexplanation
forthetransparencybiasresultisthatuserssawhowwelltheyhadperformed,andthen
basedtheirreportedunderstandingofthesystemÕslogiconthescoretheyhadreceived.
Inthiscase,thetransparencymeasuresmayprimarilyreßectusersÕinterpretationsoftheir
performanceratherthanthetransparencyofthesystem.
144CHAPTER6
CONCLUSIONS
IhaveshownthatIDAscanimpactthedecisionmakingprocessbyprovidingrecommen-
dationstousersthatrequiresomeevaluation.IDAusersmustassessrecommendations
anddeterminewhethertheactionstheysuggestwillmeetthegoalsoftheirdecisiontask.
Givingusersgoodrecommendationsisacriticalfactortoobtainingagreementwithrecom-
mendations,buttherearealsoseveralbiasesthata
"
ectagreementandareunrelatedto
recommendationquality.Usersarebiasedtowardsagreeingwithrecommendationswhen
theyhavecustomizedtheIDAslogic,evenwhentheircustomizationhashadnoactual
impactontherecommendations.Usersarealsobiasedtowardsagreementwhenpriorto
seeingrecommendationstheyhavehighexpectationsofthesystemÕse
!
cacyatproducing
recommendations.TheconsistencybetweenrecommendationsandtheIDAsconÞguration
alsocreatesabias,withusersbeingmoreinclinedtoagreewithrecommendationswhenthey
appeartobeconsistentwiththewaytheIDAwasconÞgured.Andthereisalsoabiasof
transparency,withuserswhoreportedhigherunderstandingandclarityregardingtheway
theIDAproducedrecommendationsalsobeingmorelikelytoagreewithrecommendations
thanthosewhoreportedlowerunderstanding.InthissectionIwilldiscusstheimplicationsof
theseÞndingsbothforIDAdesignandforfutureresearchonIDAsandcomputer-supported
decisionmaking.
6.1IDA-SupportedDecisionMaking
WhenIntelligentDecisionAidsprovidespeciÞcrecommendationstodecisionmakersaction
andalternativesthatshouldbeconsidered,theyplaceademandonuserstoevaluatethese
recommendationsbeforemakingadecision.Ihavepresentedevidencethatthisrequirement
145Table6.1:DecisionbiasesinIDA-supporteddecisionmaking.
Bias
Description
Evidence
CustomizationBias
Usersagreemorewithrecommendations
whentheycustomizetheIDA
Table3.5,Table5.9,
Figure3.6,Figure3.7,
Figure5.6
ExpectationsBias
Usersagreemorewithrecommendations
whentheyexpectedittoworkwell
beforeseeingrecommendations
Table4.5
ConsistencyBias
Usersagreemorewithrecommendations
whentheyappeartobeconsistent
withtheIDAsconÞguration
Table3.9,Figure3.10
TransparencyBias
Usersagreemorewithrecommendations
whentheyfeelthattheIDAÕslogic
ismoreclearandunderstandable
Table5.11
ofevaluatingrecommendationsisanimportantpartofIDA-supporteddecisionmakingthat
canbeinßuencedbythedesignofthesystemsuchthatdi
"erentsystemdesignscanleadto
di
"
erentevaluationsofthesamerecommendations.Inparticular,theprocessofgenerating
recommendationsandusersÕbeliefsabouttheIDAÕse
!
cacyandtheirinvolvementinthat
processcanleaduserstorespondindi
"
erentwaystorecommendations.
Thisrequirementtoevaluaterecommendationsdoesnothaveaclearplacementwithin
Parasuramanetal.Õs(2000)frameworkofautomationindecisionsupportthatwasdescribed
inChapter2(seeFigure
??).Evaluatingrecommendationsinvolvessomeanalysisofinfor-
mationandmayÞtthere.However,whenconsideredwithinthedomainofhuman-machine
cooperation,evaluationsofrecommendationsmustbeperformedbytheentity(userorma-
chine)thatismostresponsiblefortheactionselectionstage.Itmakesnosense,forexample,
forasystemtoproviderecommendationstoauseraboutactionstoselectifthesystemwill
betheoneselectingtheaction.Indesignsthathavealowlevelofautomationattheaction
selectionstage,whichistosaysystemssuchasIDAswhereahumanisprimarilyresponsible
forchoosingtheaction,evaluatingrecommendationsshouldbeconsideredadistinctand
criticalstageinthedecisionprocess(seeFigure6.1).
Theresultsofthesestudiesprovidesomeevidencethatthisrecommendationevaluation
146Information Acquisition
Information Analysis
Evaluate RecommendationsAction SelectionAction ImplementationInputOutputEvaluating recommendations 
involves information analysis, but 
must be performed by the entity 
doing the action selectionFigure6.1:IDA-supporteddecisionmakinginvolvesevaluatingrecommendations,which
doesnotÞtneatlyintoParasuramanetal.Õsframework.
stageiscritical.Firstofall,inallthreeexperiments,thequalityofrecommendationswas
oneofthemostimportantfactorsthata
"
ecteddecisionperformance.Thebettertherecom-
mendationsthatsubjectsreceivedfromtheIDA,thebetterthedecisionsthattheymade.
Thisshowsthattherecommendationsdoinfactmatter.Anargumentcouldbemadethat
recommendationsprovidejustanadditionalpieceofinformationtoconsider,andthatusers
mightlearnsomethingusefulfromanytypeofrecommendation,evenpoorones.Whilethis
ispossible,theevidencefromthestudiesabovesuggeststhatusersdotendtofollowrecom-
mendations,eventhough,asinthecaseoftheExerciseRecommenderstudy,theymaynot
necessarilybeneÞtfromusingrecommendations.
ThetransparencybiasobservedintheExerciseRecommenderstudyshowedthatusers
weremorelikelytoagreewithrecommendationswhentheyhadabetterideaofhowthe
systemÕslogicproducedthoserecommendations,i.e.whentheyfeltthesystemwasmore
transparent.Andinthebaseballstudies,usersweremorelikelytoagreewithrecommenda-
147tionswhenpriortoactuallyseeingtherecommendations,theyfeltthatthelogicthatwould
beusedwase
!
caciousforproducinggoodrecommendations(expectationsbias).These
Þndingssuggestthatevaluatingrecommendationsisachallengingstageandthatmanyfac-
tors,fromindividualdi
"
erencesintrustofautomatedaidstothedesignofexplanations,
customization,orotherfeaturesunrelatedtoactualrecommendationqualitycanimpacthow
usersevaluateIDArecommendations.Inotherwords,whileprovidinggoodrecommenda-
tionsisimportant,itmaynotalwaysbesu
!
cienttoenablegooddecisionmaking.Users
mustlearnhowtocalibratetheirtrustofIDArecommendationswiththeiractualreliability
andlearnhowtoidentifyrecommendationquality.Animportantconclusionfrommystudies
hereisthatthistaskisdi
!cultandpronetobias.Characteristicsoftherecommendations,
thesystem,ortheusersthemselvescanleadpeopletoperformpoorlyatidentifyingrecom-
mendationquality,ultimatelyleadingtopoordecisions.Animportantareaforfuturework
willbetodeveloptheoryabouthowusersevaluatethequalityofrecommendationsandhow
theseevaluationscanbeengineeredthroughthedesignofthesystemsothatusersareable
tomakegooddecisionsaboutwhentofollowornotfollowrecommendations.
6.2CustomizationasanIDADesign
TheÞndingsfromthesestudieso
"
ersomepracticaldesignadviceregardingtheuseofcus-
tomizationinIDAs.InChapter2,IoutlinedanargumentforusingcustomizationinIDAs
thatwasbasedonthebeneÞtsofpersonalization,transparency,andtheoriesoffunction
allocation.IarguedthatcustomizationcouldbeneÞtdecisionmakingbyimprovingrecom-
mendations,helpingusersbuildsituationalawareness,andincreasingthetransparencyof
theIDA.However,Ialsoarguedthatcustomizationmaycreatesomehazardsfordecision
makers,notablythatitmaymakeusersbiasedtowardstheirrecommendations,itmayen-
ableconÞrmationbiasbylettingpeopleproducerecommendationsthatsupportadecision
theyhavealreadymade,orthatitmayrequiresomuchskilltoproducerecommendations
148thattheusersmostinneedofthedecisionaidmaynotpossess.
BecauserecommendationqualityissocriticaltodecisionqualityinIDA-supportedde-
cisionmaking,customizationcanbeverybeneÞcialtosystemsforwhichuserscanimprove
recommendationsthroughcustomization.IfusersÕcustomrecommendationsarebetterthan
personalizedrecommendationsachievedthroughintelligentalgorithmsorcollaborativeÞl-
tering,thancustomizationwillbeneÞtdecisionmakingforthatsystem.Animportantpart
ofauser-centereddesignprocessforIDAsthatareconsideringusingcustomizationisto
evaluatehowwellusersperformatproducingrecommendationscomparedtothebestnon-
customizablealternatives.Ifusersareunabletoproducerecommendationsofequivalent
qualitytonon-customizablealternatives,thancustomizationwilllikelybeapoordesign
forthatIDA.Theimportanceofrecommendationqualityinmanycaseswilloutweighany
otherbeneÞtsthatmaycomefromcustomization,suchasimproveduserexperienceand
useracceptanceofrecommendations,andthereforecustomizationwillbeveryharmfulto
systeme
"
ectivenessifitleadstosuboptimalrecommendations.Likewise,thepotentialdetri-
mentsofcustomizationsuchasbiaseddecisionmakingmaybetolerableifoverallusersare
muchmoree
"
ectiveatproducingrecommendationsinacustomizablesystemthanthebest
non-customizablealternative.
Customizationbiasisanimportantconsiderationforsystemdesigners,andshouldbe
carefullyconsideredalongsideevaluationsofrecommendationqualityaswellasevaluations
oftrustcalibration.SystemsforwhichusersÕrelianceonrecommendationsisinconsistent
withthequalityoftherecommendationsthatitproducesshouldcarefullyconsiderhow
customizationbiasmighta
"
ecttheIDAÕse
"
ectiveness.Inasystemwhereusersoverrely
onrecommendationsandarepronetomakingcommissionerrors,whichistosayproneto
followingpoorrecommendations,customizationbiasshouldbeastrongconcernasitislikely
toonlyexacerbatetheproblemoffollowingpoorrecommendations.However,systemsthat
su
"
erfromanunder-reliancecouldbeneÞtfromcustomizationbias.Ifusersarelargely
ignoringgoodrecommendationsandmakingerrorsofomission,thebiascreatedbygiving
149userssomecontroloverthesystemÕslogiccouldleadtobetterdecisionperformance.Itis
possibleeventhatgivingonlytheperceptionofcontrolcouldbebeneÞcialinsuchcases.
Adar,Tan,andTeevan(2013)describethebeneÞtsofÒbenevolentdeceptionÓinuserin-
terfacesusingexamplessuchascrosswalksignalbuttonsorthermostatcontrolsthatdonot
actuallydoanythingtothesystembutwhichgiveusersasenseofcontrolthatbeneÞtsthem
inthecontextinwhichthesystemisbeingused.Customizationmaybeabletoprovidea
similarbeneÞttousersbyincreasingbuy-intorecommendationstogoodrecommendations,
evenifusersdonotchangetherecommendationsorareunabletoimprovethem.Incases
wheremoreagreementisneeded,thiscouldbeneÞtIDAe
"ectiveness.
AnotherimportantconsiderationfordesigningcustomizableIDAsisthefeedbackthat
usersreceiveabouthowtheircustomizationhasimpactedrecommendations.Inthesestud-
ies,therewasnoe
"
ortmadetoprovideexplicitfeedbacktousersabouthowtheiractions
hada
"
ectedthealgorithmortherecommendations.However,therewassomeevidencethat
userssoughtoutthisinformationanduseditintheirdecisionmaking.Theconsistency
biasdescribedinChapter3showedthatwhenthesystemappearedtogivearecommen-
dationthatwasconsistentwithhowtheIDAhadbeenconÞgured,usersweremorelikely
toagreewithrecommendations.Andthiswastrueforusersofboththecustomizableand
non-customizableIDAs.ItsuggeststhatwhenusersbelievethataconÞgurationhasbeen
successful,userswillbeinclinedtofollowrecommendations,includingpoorrecommenda-
tions.WhenuserscustomizeanIDA,theyareawareoftheconÞgurationandmayseek
evidenceintherecommendationsthattheconÞgurationhasbeensuccessfulorthatthe
recommendationsareconsistentwiththeconÞguration.Iftheydonotappearconsistent,
usersmaybelievethatthesystemdoesnotworkwellandpotentiallyignoregoodrecom-
mendations.OriftherecommendationsappearconsistentwiththeconÞguration,usersmay
misinterpretthesuccessofthealgorithmasagoodrecommendation.Systemdesignsshould
seekoutwaystocommunicatetousershowtherecommendationstiebackspeciÞcallytothe
conÞgurationssousersofcustomizableIDAscandeterminetheextentoftheirimpact.
1506.3ImplicationsforIDAsintheWild
Industriesandsocietiesareadvancinge
"ortstouseartiÞcialintelligenceandautomationin
knowledgework(Carr,2014),andthesee
"
ortsmeanthatIDAswillimpactanincreasing
numberofdecisions.DesigningIDAsthataree
"
ectiveatleadinguserstomakebetter
decisionsthantheywouldunaidedshouldbetheprimarygoalforsystemdesigners.Ihave
presentedseveralÞndingsinthisdissertationabouthowIDAsa
"
ectdecisionmakingina
laboratorycontext.Anincorrectdiagnosisbyadoctor,apoorinvestmentbyabanker,ora
missedsecuritythreatbyananalystcouldhavefarmoresevereconsequencesthanthefew
dollarsthatthatgooddecisionswereworthtosubjectsinthesestudies.Yetifappliedtoreal
worldsetting,wecanseethatthereareveryimportantconsiderationsforIDA-supported
decisionmakingthatrelatetotheÞndingsIhavepresentedhere.
Forexample,theclinicalIDADxPlain(Barnettetal.,1987)assistscliniciansinmaking
diagnosesbyextractinginformationfromapatientÕshealthrecordandusinganextensive
databaseandartiÞcialintelligencetorecommendpossiblediagnoses.Itallowsuserstocus-
tomizethealgorithmbyemphasizingcertainaspectsofthehealthrecord,similartohowthe
customizableIDAsinthisdissertationallowedforemphasizingcertaincategoriesofbaseball
statisticsorcertainattributesofexerciseactivities.Ifthisa
"ordanceofcustomizationcre-
atesabiastowardsfollowingitsrecommendations,thanpatientscouldbeadverselya
"ectedbyanypoorrecommendations.ErrorsinthedatabaseortechnicallimitationsoftheartiÞcial
intelligencecouldleadtopoordiagnosesthatarefollowedasaresultofcustomizationbias.
Andsinceevaluatingtheaccuracyofmedicaldiagnosesisslow,di
!
cultandexpensive,it
couldbedi
!
culttoeverattributepoorhealthoutcomestothedesignofDxPlainifinfact
suchproblemsexisted.
Ine-commerce,givingusersgreatercontroloverthealgorithmsthatproviderecommen-
dationscouldincreasesalesbyelicitingcustomizationbias.Ifusersfeeltheyhavegreater
controloverhowasystemlikeAmazonorNetßixprovidesrecommendations,itcouldlead
151userstotrustthoserecommendationmorefrequentlyandpurchasemorerecommendeditems.
Theseresultsmayhavesomeapplicationtoothertypesofintelligentsystemsbesides
IDAs.SocialnetworkingsiteslikeFacebookusealgorithmstocontrolandÞlterwhatcontent
usersseefromwithintheirnetwork,andusershavewidelyvaryingunderstandingsofhowthis
algorithmworksandtheirownroleininßuencingit(Rader&Gray,2015).Facebookrecently
madechangestothenewsfeedinterfacetoallowuserstocustomizewhatcontenttheysee,
whichFacebookstateswasintendedtobetterpersonalizethenewfeed
1.Theresultsofthis
dissertationsuggestthataslongasusersdoagoodjobofpersonalizingtheirnewsfeedsfor
themselves,Facebookmayseeanincreaseinengagementwithitscontentthatexceedseven
anyincreasethatwouldfollowmorepersonalizedcontent.InthestudiesIhavepresented
here,recommendationswerepersonalizedequallyaswellforusersofbothcustomizableand
non-customizablesystems,yetusersofthecustomizablesystemfolloweditsrecommendations
morefrequentlyandclosely.Ifthise
"
ectextendstoconsumptionofcontentonFacebook,
itwouldfollowthatuserswillconsumemorecontentfromtheirnewsfeedbecausetheyhave
customizedthealgorithm.Systemsthatenableuserstocollectandanalyzeextensivedata
aboutthemselveslikeÞtnesstrackerscouldallowuserstoconÞgurehowdataarecollected
andanalyzed.Customizationbiasmightleaduserstobelievethedeviceismoreaccurate
thanitreallyisbecauseoftheircustomization.PerceivedaccuracyinquantiÞedselfsystems
orconsumptionofcontentinsocialmediamaybedi
"erentfromagreementwithexplicit
decisionrecommendationsprovidedbyIDAs.Butiftheinterfacesforinteractionwiththe
underlyinglogicoralgorithmsaresimilar,itisreasonabletosuspectthatothertypesof
intelligentsystemscanbea
"
ectedbythebiasesIhaveobservedinthesestudies.
TheimportanceofusersÕbeliefsaboutanIDAÕse
!
cacypriortousingthesystemhas
implicationsformanytypesofIDAs.Ifclinicians,forexample,arenotwillingtofollowa
goodrecommendationbasedonabeliefthatthesystemthatproduceditusespoorlogicor
haserroneousinformation,thanthesystemmaybeharmingclinicaldecisionmakingand
1http://newsroom.fb.com/news/2015/07/updated-controls-for-news-feed/
152harmingpatients.Similarly,investorsusinganIDAmaybepersuadedbyrecommendations
tomakeaninvestmentbecausetheybelieveintheprocessthatthesystemused,suchasa
particulartypeofstatisticalmodelthattheinvestorbelievesispowerful.Thisdissertation
didnotseektounderstandclearlyhowusersformtheire
!
cacybeliefs,andratherfocused
ontheconsequencesofthosebeliefsonagreementwithrecommendationsandondecision
making.Butitisclearthattheseconsequencesaremeaningfulandthereforeanimportant
directionforfutureresearchwillbetounderstandhowuserscometoforme
!cacybeliefs
andhowappropriatee
!cacybeliefsÐonesthatareconsistentwiththeactuale
!cacyofthe
IDAÐcanbeengineered.
Animportantlimitationofofthestudiesinthisdissertationisthatalluserswereinex-
periencedwiththeIDAtheywereusing.Theseresultsmayonlyapplytonewusersandas
theygainexperience,potentiallyoveryearsofusingthesystem,thebiasesIhaveobserved
heremaydisappearorchange.
TheseÞndingspresentanotherpotentialchallengeforIDAsinapplication.Itisnotclear
whethertheskillsthatmakeforagooddecisionmakerinagivencontext,suchasdomain
knowledge,experience,andsounddecisionprocesses,willnecessarilytranslateintoskillin
customizinganIDAwellorintointerpretingrecommendations.Itappearsthatsomedegree
ofliteracywithIDAsmayberequired,andthismayplaceanundesireableburdenonusers
todevelopthisliteracy.Forexample,theskillsandcapabilitiesrequiredtobecomeagood
doctormaynotbethesamecapabilitiesthatwillenableapersontoe
"
ectivelyuseaclinical
IDA.Anddevelopingthisliteracymightrequiretimeande
"
ortthatmightotherwisebe
spentsharpeningtheirskillsintheirpractice.Thisisanotherimportantconsiderationfor
futureresearchthatfollowsfrommyworkinthisdissertation.Howdi
!
cultisitforpeople
tolearntocontrolcomplexalgorithmsgivenana
"
ordanceofcustomization,andwhatare
theconsequencesinpracticeifsubstantialliteracyisrequiredinorderforcustomizableIDAs
tobee
"
ective?
1536.4Conclusion
IntelligentDecisionAidshavetremendouspotentialtoimprovedecisionoutcomesinmany
di"
erentcontexts,buttheyalsopresentdi
!cultsocio-technicalchallengesasusersmust
learntointeracte
"
ectivelywithpowerfulbutoftencomplexandopaquetechnologies.Eval-
uatingtherecommendationsthatareproducedcreatesanewtypeofuncertaintyfordecision
makers,andIhavefoundthatusersoftendonotmakecorrectevaluationsofIDArecommen-
dationsandfrequentlymakebiaseddecisionstofollowornotfollowrecommendations.The
designoftheIDA,andparticularlytheuseofcustomizationtoa
"
orduserincreasedcontrol
overthesecomplextechnologies,caninßuencehowusersnavigatethischallenge.Ibelieve
thatIDAdesignsmustaccountforthewaythatinteractionwiththesystema
"
ectsdecisions
andseekdesignsthathelpuserslearnhowtoidentifyrecommendationqualityandmake
gooddecisionsaboutfollowinggoodrecommendationsandignoringpoorrecommendations.
154APPENDICES
155APPENDIXA
TRUSTPROPENSITYSCALE
Propensity to Trust Automation Scale 
!!Scale adapted from Merritt et al. (2012) 
!Items responses are on a 5 point scale. 
Strongly Disagree Ñ Disagree Ñ Neither Agree nor Disagree Ñ Agree Ñ Strongly Agree 
!1.I usually trust automated decision aids until there is a reason not to. 
2.For the most part I distrust automated decision aids. 
3.In general I rely on automated decision aids to assist me when they are available. 
4.My tendency to trust automation decision aids is high. 
5.It is easy for me to trust automated decision aids to do their job. 
6.I am likely to trust an automated decision aid even when I have little knowledge about it.
156APPENDIXB
BASEBALLKNOWLEDGESCREENINGQUIZ
1. How many innings are there in a typical Major League Baseball game?a. 7b. 3c. 9   d. 14
2. Which of the following best describes a designated hitter?    a. A player who bats in place of his team's pitcher
   b. A player who is substituted into the lineup when a team 
really needs a hit    c. The player who has the best batting average on the team.
    d. A player who bats in place of any other outfield player on 
his team.3. Which of the following pitchers has a better Earned Run Average (ERA)?    a. Pitcher A 3.13
    b. Pitcher B 6.49
4. Which of the following best describes a triple in baseball?a. When a player gets three hits in a gameb. When a player gets a hit and makes it safely to third basec. When a player gets three hits in one at batd. When a pitcher throws three strikes to the same batter5. Which of the following best describes a walk in baseball?a. When a pitcher is replaced and walks to the dugoutb. When a batter must walk back to home plate because his hit has gone foulc. When a batter can round the bases at a slow pace because he has hit the ball over the fence.d. When a pitcher throws four balls to a batter so he can walk to first base157APPENDIXC
CATEGORYRATINGSSURVEY
Category Rating Survey 
!Subjects answered this question about the 27 categories listed below: 
!How important do you believe the following statistical categories to be in helping a computer predict the 
outcome of baseball games? 
!Not at all important Ñ Very Unimportant Ñ Neither important nor unimportant Ñ Very important Ñ Extremely important 
!Team stats 
Winning Percentages 
!Team Hitting 
Batting Average 
Walks 
Home Runs 
Hits Triples 
Doubles 
On Base Percentage 
Slugging Percentage 
Runs 
Stolen Bases 
!Team Pitching 
Strikeouts 
Home Runs 
Earned Run Average 
Walks 
!Starting Pitcher Stats 
Innings Pitched season-to-date 
Earned Run Average 
Strikeouts 
Wins 
Hits Home Runs 
Losses
158APPENDIXD
EXERCISERECOMMENDERSEEDDATASURVEYEXAMPLE
6/15/15, 1:23 PM
Qualtrics Survey Software
Page 10 of 37
https://msuccas.us2.qualtrics.com/ControlPanel/Ajax.php?action=GetSurveyPrintPreview&T=X0PmbNLACyPXIQb9W1s7K
!"#$%&'"()&*
!"#$%#&$%%#%%&'(#)(#*&)(#&+,"",'-./&$0)-1-)-#%&2*,1-3#&$&(-/(&,*&",'&3#/*##
,+&!"#$%
&'"()&+,(&-"%$&./%$,0/1
&)*$-.-./4&5(-%&6#$.%&)(#&$0)-1-)7&2*,1-3#%&#8#*0-%#
),&)(#&9#/%:&;<3,6#.:&$.3&9,'#*&=$0>4&?+&7,@&$*#&@.+$6-"-$*&'-)(&$.&$0)-1-)7:
0(,,%#&
2345
&&&A,&9,'#*
=,37BC,*#
D)*#./)(
5*$-.-./
EEEEEEEEF-/(
9,'#*=,37BC,*#
D)*#./)(
5*$-.-./
AB;C$"-%)(#.-0%
&&5#..-%
&&D>--./
&&=,8-./
&&G(-)#'$)#*
H$+)-./
&&H,'-./&I$0(-.#
&&G*#%)"-./
&&J""-2)-0$"
I$0(-.#&KL&6-.4
&&M,//-./&NO&6-.4
&&5#..-%
&&M,//-./&P&(*4
&&159APPENDIXE
EXERCISEACTIVITIESANDLATENTFACTORSCORES
ActivityWorkoutIntensityWorkoutAtmosphereMuscleGroup
1Basketball1
.4690
.352-0
.1252BenchPress0
.889-0
.5262
.3243BicepCurls-0
.701-1
.2502
.2464Biking1Hr0
.4060
.438-1
.3755Bowling-2
.7521
.1290
.7066Boxing1
.3240
.0721
.3837Calisthenics0
.800-0
.804-1
.1638Canoeing-0
.0740
.6200
.8609Curling-0
.990-1
.0430
.87110Deadlift0
.402-0
.846-1
.54111Diving-1
.3690
.1420
.20412DumbellFly0
.418-1
.0401
.46913EllipticalMachine0
.283-1
.085-0
.77614Golf-1
.3680
.011-0
.26515Hiking1
.0570
.462-0
.44116InlineSkating0
.1911
.157-1
.87517Jogging1Hr1
.082-1
.003-1
.19918Jogging30Min0
.114-1
.325-1
.79319Jumprope1
.346-1
.4060
.111TableE.1Exerciseratingsfactorsscores
160TableE.1(contÕd)
20Lunges-0
.209-1
.781-1
.33721MedicineBall-0
.708-0
.8190
.30122MountainClimbing0
.1411
.7751
.37023PaddleBoarding0
.0860
.2121
.36724Pilates-1
.0400
.3230
.33425Planks-0
.379-1
.4961
.02426Plyometrics0
.979-0
.5891
.49127Pushups1
.038-1
.1701
.89728RockClimbing1
.4850
.9670
.87129RowingMachine0
.330-0
.8830
.83130ScubaDiving-0
.6952
.0380
.53031ShoulderPress-1
.644-0
.6181
.45432Skiing0
.9540
.7930
.18533Snorkeling-1
.3541
.937-1
.08334Soccer1
.0561
.274-1
.56435SquareDancing-0
.8041
.294-2
.10636Squats0
.513-1
.651-1
.29337Stairs30Min1
.396-2
.013-1
.95338Stretching-3
.328-1
.691-0
.53439SurÞng0
.4691
.746-0
.44640Swimming1
.0220
.5420
.42541TableTennis-1
.6590
.851-0
.11342Tennis0
.3271
.028-0
.29243UltimateFrisbee0
.0970
.485-0
.02744Volleyball0
.5471
.974-0
.324161TableE.1(contÕd)
45Walking1Hr-2
.2410
.998-1
.41546Wallsits0
.052-1
.532-0
.98547WhitewaterRafting0
.4771
.5551
.24648Wrestling1
.006-0
.0860
.72449Yoga-1
.085-1
.1400
.11550Zumba0
.6471
.624-0
.314TableE.1:Exerciseratingsfactorscores.
162APPENDIXF
EXERCISERECOMMENDERSTUDYINSTRUCTIONS
163FigureF.1:Exerciserecommenderinstructions
Click Here to Return
Thank you for participating in the Exercise Decision Study. Read these instructions carefully, as
you will be required to pass a quiz on them in order to continue to the study task. In this study, you
will select an exercise activity that best matches a hypothetical scenario. You will repeat this task
for 10 rounds, in each round using a different scenario. You will earn points for your decisions.
The more points you earn, the larger your bonus payment will be.
Each scenario lists the number of points you will earn for choosing an activity that meets certain
criteria. Each scenario has slightly different criteria, and different criteria will earn a different
number of points. These preferences will change each round.
Here are the criteria that are used in the scenarios:
Attribute
Description
Cardio
High Cardio activities demand a lot of oxygen and make you breathe hard. 
LowCardio Activities do not require lots of additional oxygen.
Group
Group Activities
 are best done with more than one person. 
Individual activities
can be done alone easily.
Resources
Resource-intensive
 activities require money, equipment, or a lot of space or take
a long time. 
Convenient activities
 can be done at home or a park and don't
require a gym, expensive equipment or a lot of time or space.
Difficulty
Challenging activities require a lot skill, experience, or training. 
Not challenging
activities can be completed by a novice.
FunFun activities make you forget that you are exercising. 
Boring activities are not
engaging but may be more efficient and not distracting.
Here is an example scenario. If you had this scenario, you would want an exercise that is:
High Cardio (earns 25 points)
Group activity (earns 25 points)
Convenient (earns 12 points)
Challenging (earns 10 points)
Fun (earns 5 points)
All the activities that you can choose from have been rated on each of the attributes. For example,
Volleyball
 has been rated as being Low Cardio, a Group Activity, Resource-intensive,
Challenging, and Fun. If you had the example scenario and you chose Volleyball, you would earn
40 points (25 for choosing a Group activity, 10 for choosing a Challenging activity, and 5 for
choosing a Fun activity). You would earn no points from the Cardio and Resources attributes
because Volleyball does not match your preference for those attributes.
In the task, you will not be given any information about how different activities have been rated.
You must use your own judgment to choose the activity that will earn the most points. After you
make your choice, you will be shown your score before moving on to the next scenario.
To help you make your decision you will use an experimental tool called the 
Exercise
Recommender for recommending exercise activities. 
The purpose of this research is to evaluate
how well the tool works at helping users choose exercise activities that they will like.
164FigureF.1(contÕd)
This system will provide some recommendations for exercises that you can select. 
You do not
have to choose one of the activities. You are free to choose any activity even if the
Exercise Recommender has not suggested it.
You will be able to adjust the settings of the Exercise Recommender to help its algorithm suggest
activities for you. You can adjust three settings:
Workout Intensity
 is how much will the activity make you work hard, breathe hard and
sweat.
Workout Atmosphere
 looks for either Fun social recreation activities or solitary workouts
depending on your setting
Muscle Group
 looks for activities that focus more on upper body muscles or an lower body
and core muscles.
The system will prioritize each setting category in the order they are listed. If Muscle Group is
listed at number 1, the Exercise Recommender will try hardest suggest activities that match your
specification for Muscle Group. 
You can adjust the priority order by clicking the arrow buttons to
move the tiles up and down, or by dragging tiles up and down.
165FigureF.1(contÕd)
The system will provide two sets of recommendations. One set is based on your configured
settings and the priority order you have specified. These suggestions apear on top.
The second set of recommendations is based on a different experimental algorithm. This
algorithm does not use your settings to produce its suggestions. Both of the algorithms can
produce good recommendations, but sometimes they might produce different results. 
You should
evaluate and consider all recommendations when making your decision. Again, you do not have
to choose an activity that has been suggested by the Exercise Recommender.
Important Points:
Each round, after looking over your preferences, click the link that says "Load Exercise
Recommender" to open up the system to help you make your decision. You must load the
Exercise Recommender and generate recommendations before you can select an activity.
The Exercise Recommender does not have any information about your points. Its
suggestions are based entirely on the settings you enter or on an intelligent algorithm that
does not know either your settings or your point preferences.
You can only generate recommendations one time per round. Be sure your settings are
correct when you click the "Generate Recommendations" button.
Reminder- You are free to choose an activity that has not been suggested. Your objective is
to choose the activity that 
you believe will score the most possible points, regardless of
what the system suggests. 
The system works well, but it may not always suggest the best
possible activity.
Be sure you fully understand the instructions. Before beginning the task, you must pass a quiz on these
instructions
Click here to continue to the quiz
If you have already passed the quiz, 
click here to return to the game
166APPENDIXG
EXERCISERECOMMENDERINTERFACE
FigureG.1:Exerciserecommenderinterface
Non-custom only interface
!167FigureG.1(contÕd).
Custom only interface
!168FigureG.1(contÕd).
Both algorithms interface 
 169FigureG.1(contÕd).
Experiment Interface 
 170APPENDIXH
POST-TESTQUESTIONNAIREFOREXERCISERECOMMENDER
STUDY
FigureH.1:Exerciserecommenderpost-testsurvey.
7/30/15, 12:40 AMQualtrics Survey Software
Page 1 of 7
https://msuccas.us2.qualtrics.com/ControlPanel/Ajax.php?action=GetSurveyPrintPreview&T=5Gb2DNrVs0nsAbr0DQkklQ
Male
Female
Under 13
13-17
18-25
26-34
35-54
55-64
DemographicsBrowser Meta Info
This question will not be displayed to the recipient.
Browser: 
Safari
Version: 
7.1.5Operating System: 
Macintosh
Screen Resolution: 
1440x900Flash Version: 
18.0.0Java Support: 
1User Agent: 
Mozilla/5.0 (Macintosh; Intel Mac OS X 10_9_5) AppleWebKit/600.5.17
(KHTML, like Gecko) Version/7.1.5 Safari/537.85.14
GenderHow old are you?
171FigureH.1(contÕd).
7/30/15, 12:40 AMQualtrics Survey Software
Page 2 of 7
https://msuccas.us2.qualtrics.com/ControlPanel/Ajax.php?action=GetSurveyPrintPreview&T=5Gb2DNrVs0nsAbr0DQkklQ
65 or over
Fitness Experience
Please indicate whether the following activities are High Cardio or Low Cardio in your
opinion.   Low Cardio
High Cardio
Paddle Boarding
  Swimming
  Please indicate whether the following activities are Individual Activities or Group
Activities in your opinion.
   Individual
Group
Walking for 1 hour
  Lunges  Please indicate whether the following activities are Challenging or Not Challenging in
your opinion.
   Not ChallengingChallengingScuba Diving
  Stretching
  Please indicate whether the following activities are Convenient or Resource Intensive
in your opinion.
   Convenient
Resource Intensive
Squats
  Pushups
  172FigureH.1(contÕd).
7/30/15, 12:40 AMQualtrics Survey Software
Page 3 of 7
https://msuccas.us2.qualtrics.com/ControlPanel/Ajax.php?action=GetSurveyPrintPreview&T=5Gb2DNrVs0nsAbr0DQkklQ
Far more knowledgable
Somewhat more knowledgable
About average
Somewhat less knowledgable
Far less knowledgable
I used the Exercise Recommender to help me make my decision
I could configure the Exercise Recommender to adjust the
recommendations it gave me
The Exercise Recommender gave me 
some
 recommendations that I had no
control over
The Exercise Recommender gave me 
only recommendations that I had no
control over
I do not agree with any of these statements
Please indicate whether the following activities are Boring or Fun in your opinion.
   Boring
FunDeadlift  Whitewater Rafting
  What is your level of knowledge about fitness and exercise compared to the general
population?Manipulation Checks
Check all statements below that you agree with
In the last round of the task, you selected ${e://Field/decision}. Please explain why
you chose ${e://Field/decision} in this round. Your answer must have at least 100
characters. 
173FigureH.1(contÕd).
7/30/15, 12:40 AMQualtrics Survey Software
Page 4 of 7
https://msuccas.us2.qualtrics.com/ControlPanel/Ajax.php?action=GetSurveyPrintPreview&T=5Gb2DNrVs0nsAbr0DQkklQ
 As a reminder, in that round you had the following preferences:
${e://Field/prefs}
IDA questions
Overall, the Exercise Recommender was ...
Not at all useful
 Extremely useful
Inaccurate
 Accurate
Difficult to use
 Easy to use
Not at all configurable
 Highly Configurable
When the Exercise Recommender made suggestions ...
   Strongly
Disagree
Disagree
NeitherAgree nor
Disagree
Agree
Strongly
Agree
I understood why the
suggestions were
made
  I thought thesuggestions made
sense
  the logic behind the
recommendations
was clear
  When making my decision about which activity to choose, I ...
174FigureH.1(contÕd).
7/30/15, 12:40 AMQualtrics Survey Software
Page 5 of 7
https://msuccas.us2.qualtrics.com/ControlPanel/Ajax.php?action=GetSurveyPrintPreview&T=5Gb2DNrVs0nsAbr0DQkklQ
   Strongly
Disagree
Disagree
NeitherAgree nor
Disagree
Agree
Strongly
Agree
Looked at the
Exercise
Recommender's
suggestions and
considered them
  Trusted the
suggestions made by
the Exercise
Recommender
  Ignored the Exercise
Recommender's
suggestions
  Looked at the
suggestions that I had
configured by
adjusting the settings
  Looked at the
suggestions that were
made by the
Intelligent Algorithm
(not affected by your
settings)
  When making my decision about which activity to choose, I ...
   Strongly
Disagree
Disagree
NeitherAgree nor
Disagree
Agree
Strongly
Agree
Looked at the
Exercise
Recommender's
suggestions and
considered them
  Trusted the
suggestions made by
the Exercise
Recommender
  Ignored the Exercise
175FigureH.1(contÕd).
7/30/15, 12:40 AMQualtrics Survey Software
Page 6 of 7
https://msuccas.us2.qualtrics.com/ControlPanel/Ajax.php?action=GetSurveyPrintPreview&T=5Gb2DNrVs0nsAbr0DQkklQ
Recommender's
suggestions
  Looked at the
suggestions that I had
configured by
adjusting the settings
  When making my decision about which activity to choose, I ...
   Strongly
Disagree
Disagree
NeitherAgree nor
Disagree
Agree
Strongly
Agree
Looked at the
Exercise
Recommender's
suggestions and
considered them
  Trusted the
suggestions made by
the Exercise
Recommender
  Ignored the Exercise
Recommender's
suggestions
  Looked at the
suggestions that were
made by the
Intelligent Algorithm
(not affected by your
settings)
  Trust Propensity
In the following questions, the term 
automated decision aids
 refers to any kind
computer system that makes recommendations about making a decision or taking an
action. Examples of such systems are:
1. Recommendations listed on e-commerce sites such as Amazon suggesting
products you might be interested in
2. Alerts on a car dashboard telling the driver that the fuel is low or maintanence is
required
176FigureH.1(contÕd).
7/30/15, 12:40 AMQualtrics Survey Software
Page 7 of 7
https://msuccas.us2.qualtrics.com/ControlPanel/Ajax.php?action=GetSurveyPrintPreview&T=5Gb2DNrVs0nsAbr0DQkklQ
Survey Powered By 
Qualtrics
3. Suggestions about movies or videos to watch on sites like Netflix or YouTube
4. Suggestions about people to connect with on social networking sites like
Facebook or LinkedIn
5. Automated recommendations made on an online stock trading website about
which stocks or investments to buy
    Strongly
Disagree
Disagree
NeitherAgree nor
Disagree
Agree
Strongly
Agree
I usually trust
automated decision
aids until there is a
reason not to
  For the most part I
distrust automated
decision aids
  In general I rely on
automated decision
aids to assist me
when they are
available.
  My tendency to trust
automated decision
aids is high
  It is easy for me to
trust automated
decision aids to do
their job
  I am likely to trust an
automated decision
aid even when I have
little knowledge about
it  177REFERENCES
178REFERENCES
Adar,E.,Tan,D.S.,&Teevan,J.(2013).Benevolentdeceptioninhumancomputer
interaction.In
Proceedingsofthesigchiconferenceonhumanfactorsincomputing
systems(pp.1863Ð1872).
Aksoy,L.,Bloom,P.N.,Lurie,N.H.,&Cooil,B.(2006).Shouldrecommendationagents
thinklikepeople?
JournalofServiceResearch
,8(4),297Ð315.
Alberdi,E.,Povyakalo,A.,Strigini,L.,&Ayton,P.(2004).E
"
ectsof
incorrectcomputer-aideddetection(CAD)outputonhumandecision-making
inmammography.
AcademicRadiology
,11(8),909Ð918.Retrievedfrom
http://www.ncbi.nlm.nih.gov/pubmed/15354301
Al-Natour,S.,Benbasat,I.,&Cenfetelli,R.T.(2008).Thee
"
ectsofprocessandoutcome
similarityonusersÕevaluationsofdecisionaids*.
DecisionSciences
,39(2),175Ð211.
Amento,B.,Terveen,L.,Hill,W.,Hix,D.,&Schulman,R.(2003).Experimentsinso-
cialdatamining:Thetopicshopsystem.
ACMTransactionsonComputer-Human
Interaction(TOCHI)
,10(1),54Ð85.
Andreoni,J.(1988).Whyfreeride?:Strategiesandlearninginpublicgoodsex-
periments.
JournalofPublicEconomics
,37(3),291-304.Retrievedfrom
http://www.sciencedirect.com/science/article/pii/0047272788900436
doi:10.1016/0047-2727(88)90043-6
Angrist,J.D.,&Pischke,J.-S.(2008).
Mostlyharmlesseconometrics:AnempiricistÕs
companion
.Princetonuniversitypress.
Barnett,G.O.,Cimino,J.J.,Hupp,J.A.,&Ho
"
er,E.P.(1987).Dxplain:anevolving
diagnosticdecision-supportsystem.
Jama
,258(1),67Ð74.
Berner,E.S.(2007).
Clinicaldecisionsupportsystems:theoryandpractice
.SpringerScience
&BusinessMedia.
Berner,E.S.,Maisiak,R.S.,Heudebert,G.R.,&YoungJr,K.R.(2003).Clinician
performanceandprominenceofdiagnosesdisplayedbyaclinicaldiagnosticdecision
supportsystem.In
Amiaannualsymposiumproceedings
(Vol.2003,p.76).
Blom,J.(2000).Personalization:ataxonomy.In
ChiÕ00extendedabstractsonhuman
factorsincomputingsystems
(pp.313Ð314).
179Bostandjiev,S.,OÕDonovan,J.,&H¬ollerer,T.(2012).Tasteweights:avisualinteractivehy-
bridrecommendersystem.In
Proceedingsofthesixthacmconferenceonrecommender
systems(pp.35Ð42).
Bostandjiev,S.,OÕDonovan,J.,&H¬ollerer,T.(2013).Linkedvis:exploringsocialand
semanticcareerrecommendations.In
Proceedingsofthe2013internationalconference
onintelligentuserinterfaces
(pp.107Ð116).
Bright,T.J.,Wong,A.,Dhurjati,R.,Bristow,E.,Bastian,L.,Coeytaux,
R.R.,...Lobach,D.(2012).E
"ectofclinicaldecision-supportsys-
tems:Asystematicreview.
AnnalsofInternalMedicine
.Retrievedfrom
http://www.ncbi.nlm.nih.gov/pubmed/22529043
doi:10.1059/0003-4819-157-1-
201207030-00450Burkolter,D.,Weyers,B.,Kluge,A.,&Luther,W.(2014).Customizationofuserinterfaces
toreduceerrorsandenhanceuseracceptance.
Appliedergonomics
,45(2),346Ð353.
Carr,N.(2014).
Theglasscage:Automationandus
.NewYork:W.W.Nortonand
Company.
Chen,L.,&Pu,P.(2009).Interactiondesignguidelinesoncritiquing-basedrecommender
systems.
UserModelingandUser-AdaptedInteraction
,19(3),167Ð206.
Chen,L.,&Pu,P.(2012).Critiquing-basedrecommenders:surveyandemergingtrends.
UserModelingandUser-AdaptedInteraction
,22(1-2),125Ð150.
Choe,E.K.,Lee,N.B.,Lee,B.,Pratt,W.,&Kientz,J.A.(2014).UnderstandingquantiÞed-
selfersÕpracticesincollectingandexploringpersonaldata.In
Proceedingsofthe32nd
annualacmconferenceonhumanfactorsincomputingsystems
(pp.1143Ð1152).
Clare,A.S.,Cummings,M.L.,How,J.P.,Whitten,A.K.,&Toupet,O.(2012).Oper-
atorobjectfunctionguidanceforareal-timeunmannedvehicleschedulingalgorithm.
JournalofAerospaceComputing,Information,andCommunication
,9(4),161Ð173.
Coiera,E.,Westbrook,J.,&Wyatt,J.(2006).Thesafetyandqualityofdecisionsupport
systems.
MethodsofInformationinMedicine
,45(1),S20Ð5.doi:10.1.1.111.9326
Cramer,H.,Evers,V.,Ramlal,S.,VanSomeren,M.,Rutledge,L.,Stash,N.,...Wielinga,
B.(2008).Thee
"ectsoftransparencyontrustinandacceptanceofacontent-based
artrecommender.
UserModelingandUser-AdaptedInteraction
,18(5),455Ð496.
Cremer,H.,&Thisse,J.-F.(1991).Locationmodelsofhorizontaldi
"
erentiation:aspecial
caseofverticaldi
"erentiationmodels.
TheJournalofIndustrialEconomics
,383Ð390.
180Cummings,M.L.(2004).Automationbiasinintelligenttimecriticaldecisionsupport
systems.In
Aiaa1stintelligentsystemstechnicalconference
(Vol.2,pp.557Ð562).
Cummings,M.L.,&Bruni,S.(2009).CollaborativehumanÐautomationdecisionmaking.
InSpringerhandbookofautomation
(pp.437Ð447).Springer.
deWinter,J.,&Dodou,D.(2014).WhytheÞttslisthaspersistedthroughoutthehistory
offunctionallocation.
Cognition,Technology&Work
,1Ð11.
DiStefano,C.,Zhu,M.,&Mindrila,D.(2009).Understandingandusingfactorscores:Con-
siderationsfortheappliedresearcher.
PracticalAssessment,Research&Evaluation
,14(20),1Ð11.
Dzindolet,M.T.,Peterson,S.A.,Pomranky,R.A.,Pierce,L.G.,&Beck,H.P.(2003).
Theroleoftrustinautomationreliance.
InternationalJournalofHuman-Computer
Studies,58(6),697Ð718.
Dzindolet,M.T.,Pierce,L.G.,Beck,H.P.,&Dawe,L.A.(2002).Theperceivedutilityof
humanandautomatedaidsinavisualdetectiontask.
HumanFactors:TheJournal
oftheHumanFactorsandErgonomicsSociety
,44(1),79Ð94.
Ehrlich,K.,Kirk,S.E.,Patterson,J.,Rasmussen,J.C.,Ross,S.I.,&Gruen,D.M.
(2011,February).Takingadvicefromintelligentsystems:Thedoubleedged
swordofexplanations.In
Proceedingsofthe15thinternationalconferenceonin-
telligentuserinterfaces-iuiÕ11
(p.125).NewYork,NewYork,USA:ACM
Press.Retrievedfrom
http://dl.acm.org/citation.cfm?id=1943403.1943424
doi:10.1145/1943403.1943424Ekstrand,M.D.,Riedl,J.T.,&Konstan,J.A.(2011).CollaborativeÞlteringrecommender
systems.
FoundationsandTrendstextregisteredinHuman-ComputerInteraction
,4,175-243.Fitts,P.M.(1951).Humanengineeringforane
"
ectiveair-navigationandtra
!
c-control
system.
Gino,F.,Sharek,Z.,&Moore,D.A.(2011).Keepingtheillusionofcontrolundercon-
trol:Ceilings,ßoors,andimperfectcalibration.
OrganizationalBehaviorandHuman
DecisionProcesses
,114(2),104Ð114.
Glaser,W.T.,Westergren,T.B.,Stearns,J.P.,&Kraft,J.M.(2006,February21).
Consumeritemmatchingmethodandsystem.
GooglePatents.(USPatent7,003,515)
Goddard,K.,Roudsari,A.,&Wyatt,J.C.(2012).Automationbias:asystematicre-
viewoffrequency,e
"
ectmediators,andmitigators.
JournaloftheAmericanMedical
181InformaticsAssociation
,19(1),121Ð127.
Golfarelli,M.,Rizzi,S.,&Proli,A.(2006).Designingwhat-ifanalysis:towardsamethod-
ology.In
Proceedingsofthe9thacminternationalworkshopondatawarehousingand
olap(pp.51Ð58).
Guerlain,S.,Brown,D.E.,&Mastrangelo,C.(2000).Intelligentdecisionsupportsystems.
InSystems,man,andcybernetics,2000ieeeinternationalconferenceon
(Vol.3,pp.
1934Ð1938).Han,S.,He,D.,Jiang,J.,&Yue,Z.(2013).Supportingexploratorypeoplesearch:astudy
offactortransparencyandusercontrol.In
Proceedingsofthe22ndacminternational
conferenceonconferenceoninformation&knowledgemanagement
(pp.449Ð458).
H¬aubl,G.,&Murray,K.B.(2003).Preferenceconstructionandpersistenceindigital
marketplaces:Theroleofelectronicrecommendationagents.
JournalofConsumer
Psychology
,13(1-2),75Ð91.
H¬aubl,G.,&Trifts,V.(2000).Consumerdecisionmakinginonlineshoppingenvironments:
Thee
"ectsofinteractivedecisionaids.
Marketingscience
,19(1),4Ð21.
Herlocker,J.L.,Konstan,J.A.,&Riedl,J.(2000).ExplainingcollaborativeÞltering
recommendations.In
Proceedingsofthe2000acmconferenceoncomputersupported
cooperativework
(pp.241Ð250).
Herlocker,J.L.,Konstan,J.A.,Terveen,L.G.,&Riedl,J.T.(2004).Evaluatingcollab-
orativeÞlteringrecommendersystems.
ACMTransactionsonInformationSystems
,22(1),5Ð53.doi:10.1145/963770.963772
Hijikata,Y.,Kai,Y.,&Nishida,S.(2012).Therelationbetweenuserintervention
andusersatisfactionforinformationrecommendation.In
SACÕ12Proceedingsof
the27thAnnualACMSymposiumonAppliedComputing
(pp.2002Ð2007).doi:
10.1145/2231936.2232109Hill,T.,Marquez,L.,OÕConnor,M.,&Remus,W.(1994).ArtiÞcialneuralnetworkmodels
forforecastinganddecisionmaking.
InternationalJournalofForecasting
,10(1),5Ð15.
Ho
"
man,R.R.,Johnson,M.,Bradshaw,J.M.,&Underbrink,A.(2013).Trustinautoma-
tion.IntelligentSystems,IEEE
,28(1),84Ð88.
Hostler,R.E.,Yoon,V.Y.,&Guimaraes,T.(2005).Assessingtheimpactofinternetagent
onendusersÕperformance.
DecisionSupportSystems
,41(1),313Ð323.
Imai,K.,Keele,L.,&Tingley,D.(2010).Ageneralapproachtocausalmediationanalysis.
182Psychologicalmethods
,15(4),309.
Introne,J.,&Iandoli,L.(2014).Improvingdecision-makingperformancethroughargu-
mentation:Anargument-baseddecisionsupportsystemtocomputewithevidence.
DecisionSupportSystems
,64,79Ð89.
Kahai,S.S.,Solieri,S.A.,&Felo,A.J.(1998).Activeinvolvement,familiarity,framing,and
theillusionofcontrolduringdecisionsupportsystemuse.
DecisionSupportSystems
,23(2),133Ð148.
Kay,J.,&Kummerfeld,B.(2012).Creatingpersonalizedsystemsthatpeoplecanscrutinize
andcontrol:Drivers,principlesandexperience.
ACMTransactionsonInteractive
IntelligentSystems(TiiS)
,2(4),24.
Klein,G.(2008).Naturalisticdecisionmaking.
HumanFactors:TheJournaloftheHuman
FactorsandErgonomicsSociety
,50(3),456Ð460.doi:10.1518/001872008X288385
Knijnenburg,B.P.,Bostandjiev,S.,OÕDonovan,J.,&Kobsa,A.(2012).Inspectability
andcontrolinsocialrecommenders.In
Proceedingsofthesixthacmconferenceon
recommendersystems
(pp.43Ð50).
Knijnenburg,B.P.,Willemsen,M.C.,&Kobsa,A.(2011).Apragmaticprocedure
tosupporttheuser-centricevaluationofrecommendersystems.In
Proceedingsof
theÞfthacmconferenceonrecommendersystems
(pp.321Ð324).NewYork,NY,
USA:ACM.Retrievedfrom
http://doi.acm.org/10.1145/2043932.2043993
doi:10.1145/2043932.2043993Koren,Y.,Bell,R.,&Volinsky,C.(2009).Matrixfactorizationtechniquesforrecommender
systems.
Computer(8),30Ð37.
Kottemann,J.E.,Boyer-Wright,K.M.,Kincaid,J.F.,&Davis,F.D.(2009).Understanding
decision-supporte
"
ectiveness:Acomputersimulationapproach.
Systems,Manand
Cybernetics,PartA:SystemsandHumans,IEEETransactionson
,39(1),57Ð65.
Kottemann,J.E.,Davis,F.D.,&Remus,W.E.(1994).Computer-assisteddecisionmaking:
Performance,beliefs,andtheillusionofcontrol.
OrganizationalBehaviorandHuman
DecisionProcesses
,57(1),26Ð37.
Kulesza,T.,Stumpf,S.,Burnett,M.,&Kwan,I.(2012).Tellmemore?:thee
"ectsof
mentalmodelsoundnessonpersonalizinganintelligentagent.In
Proceedingsofthe
sigchiconferenceonhumanfactorsincomputingsystems
(pp.1Ð10).
Lam,X.N.,Vu,T.,Le,T.D.,&Duong,A.D.(2008).Addressingcold-startproblem
inrecommendationsystems.In
Proceedingsofthe2ndinternationalconferenceon
183ubiquitousinformationmanagementandcommunication
(pp.208Ð211).
Langer,E.J.(1975).Theillusionofcontrol.
JournalofPersonalityandSocialPsychology
,32(2),311Ð328.doi:10.1037/0022-3514.32.2.311
Lee,G.,&Lee,W.J.(2009).Psychologicalreactancetoonlinerecommenda-
tionservices.
InformationandManagement
,46(8),448-452.Retrievedfrom
http://www.sciencedirect.com/science/article/pii/S0378720609000962
doi:http://dx.doi.org/10.1016/j.im.2009.07.005
Lee,J.,&Moray,N.(1992).Trust,controlstrategiesandallocationoffunctioninhuman-
machinesystems.
Ergonomics
,35(10),1243Ð1270.
Lee,J.D.,&See,K.A.(2004).Trustinautomation:Designingforappropriatereliance.
HumanFactors:TheJournaloftheHumanFactorsandErgonomicsSociety
,46(1),50Ð80.Leino,J.(2014).
Userfactorsinrecommendersystems:Casestudiesine-commerce,news
recommending,ande-learning
(Unpublisheddoctoraldissertation).UniversityofTam-
pere.
Li,A.C.,Kannry,J.L.,Kushniruk,A.,Chrimes,D.,McGinn,T.G.,Edonyabo,D.,&Mann,
D.M.(2012).Integratingusabilitytestingandthink-aloudprotocolanalysiswith
Ònear-liveÓclinicalsimulationsinevaluatingclinicaldecisionsupport.
International
journalofmedicalinformatics
,81(11),761Ð772.
Lim,B.Y.,Dey,A.K.,&Avrahami,D.(2009).Whyandwhynotexplanationsimprove
theintelligibilityofcontext-awareintelligentsystems.In
Proceedingsofthesigchi
conferenceonhumanfactorsincomputingsystems
(pp.2119Ð2128).
Linden,G.,Smith,B.,&York,J.(2003).Amazon.comrecommendations:Item-to-item
collaborativeÞltering.
InternetComputing,IEEE
,7(1),76Ð80.
Liu,J.,Dolan,P.,&Pedersen,E.R.(2010).Personalizednewsrecommendationbasedon
clickbehavior.In
Proceedingsofthe15thinternationalconferenceonintelligentuser
interfaces
(pp.31Ð40).
Madhavan,P.,&Phillips,R.R.(2010).E
"ectsofcomputerself-e
!
cacyandsystemreliability
onuserinteractionwithdecisionsupportsystems.
ComputersinHumanBehavior
,26(2),199-204.doi:10.1016/j.chb.2009.10.005
Madhavan,P.,&Wiegmann,D.A.(2007,July).Similaritiesanddi
"
erencesbetween
humanÐhumanandhumanÐautomationtrust:anintegrativereview.
TheoreticalIssues
inErgonomicsScience
,8(4),277Ð301.doi:10.1080/14639220500337708
184Manzey,D.,Reichenbach,J.,&Onnasch,L.(2008).Performanceconsequencesofautomated
aidsinsupervisorycontrol:Theimpactoffunctionallocation.In
Proceedingsofthe
HumanFactorsandErgonomicsSocietyAnnualMeeting
(Vol.52,pp.297Ð301).doi:
10.1177/154193120805200421
Manzey,D.,Reichenbach,J.,&Onnasch,L.(2012).Humanperformanceconsequencesof
automateddecisionaids:Theimpactofdegreeofautomationandsystemexperience.
JournalofCognitiveEngineeringandDecisionMaking
,1555343411433844.
Marathe,S.,&Sundar,S.S.(2011,May).Whatdrivescustomization?In
Pro-
ceedingsofthe2011annualconferenceonhumanfactorsincomputingsystems
-chiÕ11
(p.781).NewYork,NewYork,USA:ACMPress.Retrievedfrom
http://dl.acm.org.proxy2.cl.msu.edu/citation.cfm?id=1978942.1979056
doi:10.1145/1978942.1979056Massa,P.,&Avesani,P.(2007).Trust-awarerecommendersystems.In
Proceedingsofthe
2007acmconferenceonrecommendersystems
(pp.17Ð24).
Mayer,R.C.,Davis,J.H.,&Schoorman,F.D.(1995).Anintegrativemodeloforganiza-
tionaltrust.
Academyofmanagementreview
,20(3),709Ð734.
McDonald,D.W.,&Ackerman,M.S.(2000).Expertiserecommender:aßexiblerecom-
mendationsystemandarchitecture.In
Proceedingsofthe2000acmconferenceon
computersupportedcooperativework
(pp.231Ð240).
McNee,S.M.,Lam,S.K.,Konstan,J.A.,&Riedl,J.(2003).Interfacesforelicitingnewuser
preferencesinrecommendersystems.In
Usermodeling2003
(pp.178Ð187).Springer.
Merritt,S.M.,Heimbaugh,H.,LaChapell,J.,&Lee,D.(2012).Itrustit,butidonÕt
knowwhye
"
ectsofimplicitattitudestowardautomationontrustinanautomated
system.
HumanFactors:TheJournaloftheHumanFactorsandErgonomicsSociety
,0018720812465081.Merritt,S.M.,&Ilgen,D.R.(2008).Notalltrustiscreatedequal:Dispositionaland
history-basedtrustinhuman-automationinteractions.
HumanFactors:TheJournal
oftheHumanFactorsandErgonomicsSociety
,50(2),194Ð210.
MessierJr,W.F.,Kachelmeier,S.J.,&Jensen,K.L.(2001).Anexperimentalassessmentof
recentprofessionaldevelopmentsinnonstatisticalauditsamplingguidance.
Auditing:
AJournalofPractice&Theory
,20(1),81Ð96.
Miller,B.N.,Albert,I.,Lam,S.K.,Konstan,J.A.,&Riedl,J.(2003).Movielensunplugged:
experienceswithanoccasionallyconnectedrecommendersystem.In
Proceedingsofthe
8thinternationalconferenceonintelligentuserinterfaces
(pp.263Ð266).
185Mosier,K.L.,Skitka,L.J.,Heers,S.,&Burdick,M.(1998).Automationbias:Decision
makingandperformanceinhigh-techcockpits.
TheInternationalJournalofAviation
Psychology
,8(1),47-63.doi:10.1207/s15327108ijap0801
3Muir,B.M.(1987).Trustbetweenhumansandmachines,andthedesignofdecisionaids.
InternationalJournalofMan-MachineStudies
,27(5),527Ð539.
Mundlak,Y.(1978).Onthepoolingoftimeseriesandcrosssectiondata.
Econometrica:
journaloftheEconometricSociety
,69Ð85.
Nass,C.,&Moon,Y.(2000).Machinesandmindlessness:Socialresponses
tocomputers.
JournalofSocialIssues
,56,81Ñ-103.Retrievedfrom
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.87.2456
Nelson,P.(1970).Informationandconsumerbehavior.
TheJournalofPoliticalEconomy
,311Ð329.Nickerson,R.S.(1998).ConÞrmationbias:Aubiquitousphenomenoninmanyguises.
ReviewofGeneralPsychology
,2(2),175Ð220.doi:10.1037/1089-2680.2.2.175
Norman,D.A.(1990a).
Thedesignofeverydaythings
.NewYork:Doubleday.
Norman,D.A.(1990b).TheÕproblemÕwithautomation:inappropriatefeedbackandinterac-
tion,notÕover-automationÕ.
PhilosophicalTransactionsoftheRoyalSocietyofLondon.
B,BiologicalSciences
,327(1241),585Ð593.
Norton,M.,Mochon,D.,&Ariely,D.(2011).TheÕIKEAe
"
ectÕ:Whenlaborleadsto
love.
HarvardBusinessSchoolMarketingUnitWorkingPaper
(11-091).Retrieved
from
http://papers.ssrn.com/sol3/papers.cfm?abstract
id=1777100Ochi,P.,Rao,S.,Takayama,L.,&Nass,C.(2010).Predictorsofuserperceptionsofweb
recommendersystems:Howthebasisforgeneratingexperienceandsearchproduct
recommendationsa
"
ectsuserresponses.
InternationalJournalofHuman-Computer
Studies,68(8),472Ð482.
OÕDonovan,J.,&Smyth,B.(2005).Trustinrecommendersystems.In
Proceedingsofthe
10thinternationalconferenceonintelligentuserinterfaces
(pp.167Ð174).
Onnasch,L.,Wickens,C.D.,Li,H.,&Manzey,D.(2013).Humanperformanceconsequences
ofstagesandlevelsofautomationanintegratedmeta-analysis.
HumanFactors:The
JournaloftheHumanFactorsandErgonomicsSociety
,0018720813501549.
Parasuraman,R.,Sheridan,T.B.,&Wickens,C.D.(2000).Amodelfortypesandlevelsof
humaninteractionwithautomation.
Systems,ManandCybernetics,PartA:Systems
186andHumans,IEEETransactionson
,30(3),286Ð297.
Pariser,E.(2011).
TheÞlterbubble:Whattheinternetishidingfromyou
.PenguinUK.
Park,S.-T.,&Chu,W.(2009).Pairwisepreferenceregressionforcold-startrecommendation.
InProceedingsofthethirdacmconferenceonrecommendersystems
(pp.21Ð28).
Parra,D.(2013).Usercontrollabilityinahybridrecommendersystem.
Pereira,R.E.(2001).Inßuenceofquery-baseddecisionaidsonconsumerdecisionmaking
inelectroniccommerce.
InformationResourcesManagementJournal
,14(1),31.
Qin,S.,Menezes,R.,&Silaghi,M.(2010).Arecommendersystemforyoutubebasedonits
networkofreviewers.In
Socialcomputing(socialcom),2010ieeesecondinternational
conferenceon
(pp.323Ð328).
Rader,E.,&Gray,R.(2015).Understandinguserbeliefsaboutalgorithmiccurationin
thefacebooknewsfeed.In
Proceedingsofthe33rdannualacmconferenceonhuman
factorsincomputingsystems
(pp.173Ð182).
Ross,J.,Irani,L.,Silberman,M.,Zaldivar,A.,&Tomlinson,B.(2010).Whoarethe
crowdworkers?:shiftingdemographicsinmechanicalturk.In
ChiÕ10extendedabstracts
onhumanfactorsincomputingsystems
(pp.2863Ð2872).
Roth,M.,Ben-David,A.,Deutscher,D.,Flysher,G.,Horn,I.,Leichtberg,A.,...Merom,
R.(2010).Suggestingfriendsusingtheimplicitsocialgraph.In
Proceedingsofthe
16thacmsigkddinternationalconferenceonknowledgediscoveryanddatamining
(pp.233Ð242).Sanchez,J.,Fisk,A.D.,&Rogers,W.A.(2004).Reliabilityandage-relatede
"
ectson
trustandrelianceofadecisionsupportaid.In
Proceedingsofthehumanfactorsand
ergonomicssocietyannualmeeting
(Vol.48,pp.586Ð589).
Sargan,J.D.(1958,July).Theestimationofeconomicrelationshipsusinginstrumental
variables.
Econometrica
,26(3),393-415.
Schafer,J.B.,Konstan,J.A.,&Riedl,J.(2004).Viewthroughmetalens:usagepatterns
forameta-recommendationsystem.
IEEProceedings-Software
,151(6),267Ð279.
Schein,A.I.,Popescul,A.,Ungar,L.H.,&Pennock,D.M.(2002).Methodsandmetrics
forcold-startrecommendations.In
Proceedingsofthe25thannualinternationalacm
sigirconferenceonresearchanddevelopmentininformationretrieval
(pp.253Ð260).
Schuster,D.,Jentsch,F.,Fincannon,T.,&Ososky,S.(2013).Theimpactoftypeandlevel
187ofautomationonsituationawarenessandperformanceinhuman-robotinteraction.In
Engineeringpsychologyandcognitiveergonomics.understandinghumancognition
(pp.252Ð260).Springer.
Sinha,R.,&Swearingen,K.(2002,April).Theroleoftransparencyinrec-
ommendersystems.In
ChiÕ02extendedabstractsonhumanfactorsincom-
putingsystems-chiÕ02
(p.830).NewYork,NewYork,USA:ACM
Press.Retrievedfrom
http://dl.acm.org/citation.cfm?id=506443.506619
doi:10.1145/506443.506619Skeels,M.M.,&Grudin,J.(2009).Whensocialnetworkscrossboundaries:acasestudyof
workplaceuseoffacebookandlinkedin.In
Proceedingsoftheacm2009international
conferenceonsupportinggroupwork
(pp.95Ð104).
Skitka,L.J.,Mosier,K.L.,&Burdick,M.(1999).Doesautomationbiasdecision-
making?InternationalJournalofHuman-ComputerStudies
,51(5),991-1006.doi:
10.1006/ijhc.1999.0252Smith,V.L.(1976,May).ExperimentalEconomics:InducedValueTheory.
TheAmerican
EconomicReview
,66(2),274Ð279.
Solomon,J.(2014).Customizationbiasindecisionsupportsystems.In
Proceedingsofthe
32ndannualacmconferenceonhumanfactorsincomputingsystems
(pp.3065Ð3074).
Solomon,J.,Ma,W.,&Wash,R.(2015).DonÕtwait!:Howtiminga
"
ectscoordination
ofcrowdfundingdonations.In
Proceedingsofthe18thacmconferenceoncomputer
supportedcooperativework&socialcomputing
(pp.547Ð556).
Solomon,J.,&Wash,R.(2014,October).Human-whatinteraction?understandinguser
sourceorientation.In
Humanfactorsandergonomicssocietyannualmeetingproceed-
ings.Stock,J.H.,&Yogo,M.(2005).Testingforweakinstrumentsinlinearivregression.
Identi-
Þcationandinferenceforeconometricmodels:EssaysinhonorofThomasRothenberg
.Su,X.,&Khoshgoftaar,T.M.(2009).AsurveyofcollaborativeÞlteringtechniques.
AdvancesinartiÞcialintelligence
,2009,4.
Sundar,S.S.(2008).Self-as-source.InE.Konijn,S.Utz,M.Tanis,&S.B.Barnes(Eds.),
Mediatedinterpersonalcommunication
(pp.58Ð74).NewYork:Routledge.
Sundar,S.S.,&Marathe,S.S.(2010).Personalizationversuscustomization:Theimpor-
tanceofagency,privacy,andpowerusage.
HumanCommunicationResearch
,36(3),298Ð322.188Sundar,S.S.,&Nass,C.(2000,December).SourceOrientationinHuman-
ComputerInteraction.
CommunicationResearch
,27(6),683Ð703.Re-
trievedfrom
http://crx.sagepub.com/content/27/6/683.abstract
doi:10.1177/009365000027006001
Sundar,S.S.,Oh,J.,Bellur,S.,Jia,H.,&Kim,H.-S.(2012,May).Interactivityasself-
expression.In
Proceedingsofthe2012ACMAnnualConferenceonHumanFactorsin
ComputingSystems-CHIÕ12
(p.395).NewYork,NewYork,USA:ACMPress.doi:
10.1145/2207676.2207731Tintarev,N.,&Mastho
"
,J.(2007).E
"
ectiveexplanationsofrecommendations:user-
centereddesign.In
Proceedingsofthe2007acmconferenceonrecommendersystems
(pp.153Ð156).
Tintarev,N.,&Mastho
"
,J.(2008).Thee
"
ectivenessofpersonalizedmovieexplanations:
Anexperimentusingcommercialmeta-data.In
Adaptivehypermediaandadaptive
web-basedsystems
(pp.204Ð213).
Tintarev,N.,&Mastho
"
,J.(2011).Designingandevaluatingexplanationsforrecommender
systems.In
Recommendersystemshandbook
(pp.479Ð510).Springer.
Tintarev,N.,&Mastho
"
,J.(2012,February).Evaluatingthee
"
ectivenessofexpla-
nationsforrecommendersystems.
UserModelingandUser-AdaptedInteraction
.Retrievedfrom
http://www.springerlink.com/content/9087j6w55r654255/
doi:10.1007/s11257-011-9117-5Tjur,T.(2009).Coe
!
cientsofdeterminationinlogisticregressionmodelsÑanewproposal:
Thecoe
!
cientofdiscrimination.
TheAmericanStatistician
,63(4),366Ð372.
Tu,Q.,&Dong,L.(2010).Anintelligentpersonalizedfashionrecommendationsystem.In
Communications,circuitsandsystems(icccas),2010internationalconferenceon
(pp.479Ð485).Verbert,K.,Parra,D.,Brusilovsky,P.,&Duval,E.(2013).Visualizingrecommendations
tosupportexploration,transparencyandcontrollability.In
Proceedingsofthe2013
internationalconferenceonintelligentuserinterfaces
(pp.351Ð362).
Wang,W.,&Benbasat,I.(2007).Recommendationagentsforelectroniccommerce:Ef-
fectsofexplanationfacilitiesontrustingbeliefs.
JournalofManagementInformation
Systems,23(4),217Ð246.
Wang,W.,&Benbasat,I.(2008).Attributionsoftrustindecisionsupporttechnologies:A
studyofrecommendationagentsfore-commerce.
JournalofManagementInformation
Systems,24(4),249Ð273.
189Wash,R.(2013).
Expectationsandcriticalmassinonlinecommunities.
(NSFProposal)
Wash,R.,&Solomon,J.(2014).Coordinatingdonorsoncrowdfundingwebsites.In
Pro-
ceedingsofthe17thacmconferenceoncomputersupportedcooperativework&social
computing
(pp.38Ð48).
Woolley,D.J.(2007).Theinßuenceofdecisioncommitmentanddecisionguidanceon
directingdecisionaidrecommendations.
AcademyofInformationandManagement
SciencesJournal
,10(2),39.
Xiao,B.,&Benbasat,I.(2007).E-commerceproductrecommendationagents:use,charac-
teristics,andimpact.
MisQuarterly
,31(1),137Ð209.
Yang,R.,&Newman,M.W.(2013).Learningfromalearningthermostat:lessonsfor
intelligentsystemsforthehome.In
Proceedingsofthe2013acminternationaljoint
conferenceonpervasiveandubiquitouscomputing
(pp.93Ð102).
Yang,T.Y.,&Swartz,T.(2004).Atwo-stagebayesianmodelforpredictingwinnersin
majorleaguebaseball.
JournalofDataScience
,2(1),61Ð73.
190