ContentslistsavailableatScienceDirect
Social Networks
j ourna l h o me pa g e :w w w . e l s e v i e r . c o m / l o c a te / s o c n e t
Referrals and information flow in networks increase discrimination: A laboratory experiment
Károly Takács
a,∗, Giangiacomo Bravo
b, Flaminio Squazzoni
caMTATK“Lendület”ResearchCenterforEducationalandNetworkStudies(RECENS),CentreforSocialSciences,HungarianAcademyofSciences,Tóth Kálmánu.4.,1097Budapest,Hungary
bDepartmentofSocialStudiesandCentreforDataIntensiveSciencesandApplications,LinnaeusUniversity,Universitetsplatsen1,35195Växjö,Sweden
cDepartmentofEconomicsandManagement,UniversityofBrescia,ViaSanFaustino74B,25122Brescia,Italy
a r t i c l e i n f o
Articlehistory:
Received8August2017
Receivedinrevisedform19March2018 Accepted22March2018
Availableonline30March2018
Keywords:
Hiringdiscrimination Referrals
Recommendations Informationnetworks Labormarket Laboratoryexperiment Qualitystandards
a b s t r a c t
Referralsandinformationflowdistortmarketmechanismsofhiringinthelabormarket,buttheymight assistemployersunderasymmetricinformationinfindingbetteralternatives.Thispaperinvestigates whetheranimpartialinformationflowbetweenemployersinacyclicnetworkstructurecouldgenerate morediscriminationthanwhennoinformationisexchangedbetweenemployers.Wesetupanartificial labormarketinwhichtherewasnoaveragequalitydifferencebetweentwocategoriesofworkers.We askedparticipantstoplaytheroleofemployersandexaminedthepartialityoftheirhiringchoices.Results showedthatdiscriminationwasprevalentinallconditions.Higherstandardsbytheemployersforthe qualityofworkersincreaseddiscriminationasdidthepresenceofreferralsfromworkers.Unexpectedly, impartialinformationflowinacyclicnetworkofemployersdidnothelptodecreasediscrimination.We alsoshowedthatthesemechanismsinteractwithandsubdueeachotherincomplexways.
©2018TheAuthors.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-ND license(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Introduction
Hiringdiscriminationmeansdifferentialtreatmentofacertain socialcategory,basedoncategorymembershipratherthanindi- vidualmerit.Differentialtreatmentisrecurrentinhiringchoices characterizedbyasymmetryofinformationbetweentheorganiza- tionandtheapplicant(PetersenandSaporta,2004;Rooth,2010).
Giventhatthetrueworker’squalitycannotbeaccuratelypredicted duringhiringdecisions,organizationsmightuserecognizabletraits (e.g.,raceandgender)‘asinexpensivescreeningdeviceswhenhir- ingforjobs,particularlyskilledjobs,inthebelief(correctornot) thatraceandsexstatusare,onaverage,relatedtoproductivity’
(Kaufman2002,p.550).Whenthereisnoorlittlestatisticalbasis todistinguishthequalityofmembersofdifferentcategories,fol- lowingrecognizabletraitscannothelptoestimatetheapplicant’s quality.Inthesecases,understandingwhydiscriminationcould persistisofparamountimportance(Bertrandetal.,2005).
∗ Correspondingauthorat:MTATK“Lendület”ResearchCenterforEducational andNetworkStudies(RECENS),CentreforSocialSciences,HungarianAcademyof Sciences,TóthKálmánu.4.,1097,Budapest,Hungary.
E-mailaddresses:takacs.karoly@tk.mta.hu(K.Takács),
giangiacomo.bravo@lnu.se(G.Bravo),flaminio.squazzoni@unibs.it(F.Squazzoni).
Hiringdecisionsmightreflectsignalsandinformationthatchan- nelthrough social networks. The important role of referrals in particulariswelldocumentedforgettingajob(Granovetter,1973, 1974;Linetal.,1981;Wegener,1991;Elliott,2001;Mouw,2002;
FernandezandFernandez-Mateo,2006;PonzoandScoppa,2010;
FountainandStovel,2014).Manystudiesarguedthatgettingajob viareferralsmightdistorttheperfectmarketlogicandreplacemer- itocraticprocessesinhiring(IoannidesandLoury,2004;Petersen etal.,2000;TassierandMenczer,2008).Forinstance,theextended useofinformaljobsearchmethodsmayhaveanegativeeffecton therateofmobilityfromlowstatustohighstatusjobs(McBrier 2003;1212).Ifoneofthegroupshasabetteraccesstoinformal jobsearch,thisisdetrimentalfortheothergroup,asinthecaseof referralsfromthe“oldboys”networkinawiderangeofcontexts (Rogers,2000;McBrier,2003;McDonald,2011;Bianetal.,2015).
Research concerning referrals highlighted how the hiring mechanismcouldenhanceinequalityofemploymentandwages (Montgomery,1991;Krauth,2004;Fontaine,2008).Giventhatcon- tactsmightbehomophilouswithregardtointernal quality,the extensiveuseofreferralsleadnecessarilytogrowinginequality (Montgomery,1991;Beamanand Magruder,2012).Considering that contactsare homophilous alsowithregard to social char- acteristics that are uncorrelated with ability, research showed that initial differences in the employment rate could result in greaterwage inequalitiesovertime (Montgomery,1991;Arrow
https://doi.org/10.1016/j.socnet.2018.03.005
0378-8733/©2018TheAuthors.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense(http://creativecommons.org/licenses/by-nc-nd/4.
0/).
and Borzekowski,2004).Exogenously provided job information thatispassedonvianetworktiesalsoenlargessmallinitialdif- ferences(Calvó-ArmengolandJackson,2004,2007).
Afewstudieslookedatreferrals(Engströmetal.,2012;Beaman andMagruder,2012;Beamanetal.,2013;CariaandHassen,2013;
FernandezandGreenberg,2013)andotherstructuralmechanisms thataffectdiscrimination(seeOlianetal.,1988foranearliermeta- analysis).Ourpaperaimstostudystructuralmechanismsthatcan influencedifferentialhiringpracticesexperimentally.
Themainadvantagesoftheexperimentalmethodologyarethat:
(a)thehypothesizedcorrelationscanbetestedunambiguouslyin afullycontrolledenvironmentthatexcludesconfoundingeffects (e.g.,Smith,1991;Roth,1993;WebsterandSell,2007);(b)genera- tiverelationscanbeidentified;(c)replicationoffindingsispossible (e.g.,Chapin,1932;FalkandFehr,2003;WillerandWalker,2007;
Fehr and Gintis,2007;Bohnet,2009; Falk and Heckman,2009;
Smith,2010).Itisworthnotingthatexperimentalstudiesmight helpustoidentifymechanismsthatcanbeempiricallyexamined indifferentcontexts(Camerer,2003;Ostrom,2010;Ariely,2008;
FalkandHeckman,2009).
Classicaland recent small group experiments in social psy- chologytestifiedtothehumantendencytodiscriminateunknown partnersbasedoncategorymembership(e.g.,Brewer,1979,1996;
Dovidioetal.,2002;Fiske,2009).Similarly,recentlaboratoryand fieldstudiesineconomicsandsociologyconfirmedtheexistenceof discriminativepractices(e.g.,SolnickandSchweitzer,1999;Pager etal.,2009;Jackson,2009;Midtbøen,2014;Agerström,2014;Lee etal.,2015).Somelaboratorystudieswereabletoisolateimportant behavioral effectsand interactionalinfluencesin discrimination (e.g.,Keuschniggand Wolbring,2015; Takácsetal.,2015; Lane, 2016).
Veryfewstudieshavetriedtoanalyzefactorsrelatedtosocial capital in hiring experimentally (Godechot, 2016). This can be explainedasitisverydifficulttodepictthecomplexcharacteris- ticsofsocialcapitalinthelaboratory,causingconcernsofexternal validity. But exactlydue to thecomplex nature of social capi- talrelatedprocesses,fieldresearchcannotfullydisentanglethe informationalaspectsofsocialcapitalfromothermechanismson discrimination.Bycontrast,carefullydesignedexperimentsusing simplenetworkstructurescanprovideusatrulycausalaccountby focusingonspecificmechanismsinherentintherelationalstruc- ture(Kosfeld,2004;WillerandWalker,2007;Gërxhanietal.,2013;
BrashearsandQuintane,2015;BrashearsandGladstone,2016).In ourcase,theexperimentaldesigncanconcentrateandrelyonsome elementaryandempiricallyrelevantmechanismsthatpotentially determinediscriminationinhiring.Oneofthesemechanismscov- ersreferralscomingfromworkers.Anotheronesummarizesthe informationflowcomingfromotheremployerswhoareverymuch inthesamesituationandhavesimilargoals.Acquiring,passingon, andexchanginginformation betweenemployersaboutemploy- eesisverydifficulttotraceinfieldstudies.Tagsandsignalsthat characterizeworkersarealsomulti-dimensional,somecorrelate withinternalqualitiesandskills,whileothersdonot.Asourstudy demonstrates,thesemechanismscanbeabstractedandusedin thelab.Inordertoallowforcausalinference,ourlaboratoryexper- imentsexcludeconcernsaboutstrategicchoicesandendogeneity inrecommendationsandreferralsbydesign.
Itanalyzeshiringdecisionsinacontrolledsettingandisable toreducethehighdimensionalityofrealityintoastraightforward model.
Given thecomplexity of hiring choices in the labormarket, empiricalfieldresearchisunabletotestunivocallywhetherrefer- ralsincreasediscriminationcomparedtoabaselinecasewithout referrals or not. Furthermore, it cannot be explored whether referralsmakeadifferencealsowithoutanyinitialbiasesoralter- natively,observedinequalitiesaretherebecauseofhistoricalpath
dependence.Besides,inexistingfieldstudies,workerreferralsand informationflowamongemployersareconsideredjointlyandtheir impactsarehardlyseparated.
In order to overcome these empirical difficulties, following Takácsetal.(2015),wehavedesignedalabormarketexperiment whereparticipants(universitystudents)wereaskedtoplay the roleofemployersandselectfictiveworkersbelongingtotwocate- gories.Inourexperiment,byexcludingcontextualeffectsandother importantaspectsofthehiringprocess,wetestedthenetandthe jointeffectsofreferrals,informationflowfromotheremployers,and qualitystandardsondiscrimination.
Referralshavebeendefinedasrecommendationsforhiringby workersin-house(Montgomery,1991;FountainandStovel,2014).
Referralsarenaturallybiasedtowardsmembersofthein-group.
Thischaracteristicfeaturehasbeendepictedinourexperimental design.Informationflowhasbeenconceptualizedasanautomated process in a simple directed network of employers. This con- ceptualizationcovers multiple mechanisms accordingto which employers gettoknow thetruequalitiesof workersemployed atconnectedfirms;suchasrecommendationletters,information exchange,andobservations thattakeplaceasa resultofestab- lishedcontactbetweentheorganizations.Qualitystandardswere evaluationthresholdssetupexogenouslytodeterminewhetherit iseconomicaltokeepworkersinhouseornot.
Inourexperiment,representinganidealisticworld,therewas nodifferenceinthemeananddistributionofqualityofworkersin thetwocategories;hencetheimpactofhistoricalpathdependence andinitialbiasescouldbeexcluded.Notethatsignificantelements ofeverydayinteractionswereneglectedinourlabormarketlab- oratory.Forinstance,therewasnorecruitmentprocedureinthe experimentaswewereinterestedindiscriminationwhenhiring decisionsaremade.Recruitmentitselfcanaddanextralayerofdis- criminationbyselectivelytargetingcertaingroupsorusingbiased information channels.Moreover, in reality,workers themselves couldapplyselectively byexpectingdiscrimination.Thesecom- plicationsarepresentinthefieldandwoulddistorttheevaluation ofimpartialityofhiringdecisionsinthelab.
Ourresearchquestionswereasfollows.
1.Doesdiscriminationoccurinanartificiallabormarketwithbal- ancedandfairconditions?
2.Dohigherqualitystandardscreatemorediscrimination?Orin otherwords,ifemployersarerewarded onlyforhighquality workers,willdiscriminationincrease?
3.Doworkerreferralsincreasediscrimination?
4.Doesflowofaccurateinformation inanetworkofemployers decreasediscrimination?
Thesequestionsconcernprimarilythebehavior ofindividual employers.Inaddition,wewerealsoabletoanalyzewhetherocca- sionalindividualbiasesbalanceeachotheroutortheyaddupto inequalityofemploymentbetweengroupsinourartificiallabor marketwithbalancedandfairconditions.
Hypotheses
Aninclinationtowardsdiscrimination
AssuggestedbyTakácsandSquazzoni(2015),whobuiltasimple modelofanidealizedlabormarketinwhichtherewasnoobjective differenceinaveragequalitybetweengroupsandhiringdecisions werenotbiased,acertainlevelofdiscriminationcouldbeexpected alsoinanidealworldwithimpartialemployers.Judgmenterrorsof thiskindcouldbetheconsequenceof“rational”adaptivesampling ofavailableinformation(cf.Simon,1955;Denrell,2005;Fiedler
andJuslin,2006;LeMensandDenrell,2011;DenrellandLeMens, 2011).Thatis,peoplemakesystematic judgmenterrors asthey naïvelyextrapolate fromlimited information availabletothem.
Moreover,psychologicalstudiesindicatedthatindividualsareto largeextentalsoprocessinformationinaccuratelyortheyrelyon biasedheuristics.Theyareconstrainedbyselectiveattentionand tendtoretaininformationconfirmingtheirbeliefs, whileignor- inginformationthatcontradictstheirexpectations(e.g.,Hamilton, 1981).Aninitialbiaswithsequential,path-dependenthiringdeci- sionsmayresultinpersistentandself-reinforcingdiscriminative choices(e.g.,Hoefferetal.,2006).Theseexpectancyconfirmation sequenceshavebeenfoundinexperimentalresearch(Bergeretal., 1980;DarleyandFazio,1980)andareexpectedtoinfluencealso employerdecisionsinhiringprocesses.Theseconsiderationsledus toformulatethefollowinghypothesis:
Hypothesis1. Anon-zerolevelofdiscriminationisexpectedalso inthebaselineconditionofnoreferralsintheidealisticexperimen- tallabormarket.
Higherstandardsanddiscrimination
Higherqualitystandardsforapplicant qualitiescanoriginate inhigherneedforthebestqualitylaborforce,inmoredemand- ingtasks,inintensemarketcompetition,andinsimplegreediness.
Employerswhohavehighstandardsforapplicantssortoutcan- didateswhowould otherwisebe abletoconductthejob.After employment,employerswithhigherstandardsarenoteasilysatis- fiedandexperimentmorewithnewhires.Thatis,higherstandards areusedforselectionand alsofor keepingthelaborforce(e.g., TakácsandSquazzoni,2015).Higherstandardsaretypicalforhigh statusjobsandforjobswherespecializedknowledgeoradvanced skillsarerequired.Advancedskillscouldbelearntafteremploy- mentwithinhouse,butthatrequiresexpensiveinvestmentsfrom theemployer.Theseinvestmentsareeasilylostiftheemployee quitsthejob.Turnovercoststhereforearemuchhigherinjobsthat requireadvancedskillsthaninjobsthatdonot.
Greedysearch(“over-searching”)thattriestoseekbetteralter- nativesthananemployeewiththequalitystandardoftheoptimal reservationlevelconveysaloss(e.g.,McCall,1970;Stigler,1961, 1962;Mortensen,1986).Asearchthatisextendedbeyondthealter- nativethatisatleastasgoodasthereservationlevelresultsin anexpectedrelativelossnotjustbecauseofsearchcosts,butalso becausetheaverageexpectedqualityofnewworkersissmaller thanthatofthecurrentalternative.Extendedsearchandrepeated failuresimplythatthehigherqualitystandardsalsoresultinlower employerprofits.
Forthesakeofsimplicity,inourexperiment,weassumedthat employershavenoopportunitytotraintheirworkersandtheydo notfacedifferentialturnovercosts.Wewereprimarilyinterestedin theconsequencesofsettingahigherqualitystandardexogenously forworkerselectionforlabormarketinequality.
Ahigherqualitystandardimpliesmoreexperimentationwith new workers.Information cues therefore could have a greater importanceforemployerdecisions.Inourexperiment,theonly availableinformationcue wascategorymembership.Itsimpor- tance,therefore,is expectedtoincreasefor employerswho are motivatedtohirenewworkerswithhighquality.Hence,weexpect thatemployerslookingforhighqualityworkersconsidergroup markersmore closely and, dependingonavailable information, areinclinedtodevelopbiasedgroupreputations.Thefewhighly skilledworkerswhomtheyaresatisfiedwitharekeptinhouseand continuetobiastheemployer’sjudgmentaboutavailableskillsin thegroups.Becauseofthelargerperceivedroleofsupplementary informationandtheover-representationofskilledworkerskeptin houseforgroupreputationformation,higherdiscriminationrates
couldoccurforemployerswithhigherqualitystandards.Thisled ustooursecondhypothesis.
Hypothesis2. Higherqualitystandardsleadtohigherdiscrimi- nationrates.
Referralsanddiscrimination
Socialnetworkscanbeusedinjobhiringfortworeasons.First, duetoitsaffectivecontent,asocialtiecreatesanobligationforthe workerin-houseononeendandanopportunityforthejob-seeker attheotherendoftheconnection.Second,networktiesarechan- nelsofgathering,conveying,andsignalinginformationonhidden individualqualities.Dependingonwhichaspectismoreprevalent, referralsmighthavedifferentconsequencesfordiscrimination(cf.
RubineauandFernandez,2015).
Workerreferralsbetweencurrentand prospectiveemployees buildontheaffectivecontentofrelationshipsandaremorecon- cerned with the welfare of the applicant, less with employer benefits.Thismeansthatworkerreferralsdonotnecessarilyreflect informationonqualityanddonotallowtheemployertofindopti- malmatchesinhiring.Consequently,ifconsideredatall,worker referralsarenottakenintoaccountbecauseofdirectprofit-seeking motives.
Workerreferralsarebasedonsocialtiesthatarehomophilous toalargeextentintheethnicand otherimportantdimensions (McPhersonetal.,2001; RubineauandFernandez,2013).When referralnetworksare usedin whichties are typicallybased on homophily, theycause labormarket segregation(Model, 1993;
Tilly,1998;Elliott1999,2001;Kugler,2003;StovelandFountain, 2009). With homophilous referrals, labor market segregation develops endogenouslyeven withnon-prejudicedagents (Barr, 2009).
Empiricalworkshowedthat membersof a particularethnic grouptendtorecommendotherswiththesameethnicbackground forajob(Elliott,2001;FernandezandFernandez-Mateo,2006).This canreinforcetheirdisadvantagedpositionandexcludethemfrom betterjobs(Wilson,1987).Therefore,thedeficitofdisadvantaged groupsdoesnotdependonthefactthattheywouldrelylessor moreonnetworksinfindingajob.Rather,itdependsonthefact thattheyextensivelyrelyon“wrongnetworks”thatcannotoffer themgoodjobs(FernandezandFernandez-Mateo,2006;Petersen etal.,2000).
Thecharacteristicfeatureofworkerreferralsthatweenteredin thelaboratoryistheirhomophilouscharacterwithregardtocat- egorymembership.Weassumethattheirpresencereinforcesthe initialrandomdelusionsofemployersaboutgroupdifferences(cf.
FernandezandGreenberg,2013).Consequently,wecanhypoth- esizethatdiscriminationisstrongerwithworkerreferralsthanin caseofisolated,unconnectedemployers.Asweareinterestedinthe discriminationtendenciesofemployers,referralsbyfictiveworkers inourexperimentconsideredtobearandomwithin-groupprocess.
Hypothesis3. Homophilousworkerreferralsincreasediscrimi- nationrates.
Anetworkofinformationflowanddiscrimination
In contrasttoaffectivecontent ofworker referrals, theflow ofaccurateand relevant information betweenemployersabout workers coulddecrease information asymmetry in hiring deci- sions(Fernandezetal.,2000;Elliott,2001).Witha correctview onindividualqualitiesinalargerpoolofworkers,employerscould arriveatbetterinformeddecisions.Thisisinlinewiththeempirical factthatrecommendationsfrombusinesspartnersandrespected employersareconsideredseriouslyatjobinterviews.Surveysof personnelofficersfoundthatrecommendationsfromamanager
weremoreimportantforhiringthanobjectivesignals,suchashigh schoolgrades(Rosenbaumetal.,1990;Spoonley2008:27).
Whenemployersareisolatedandbasedecisionsonlyontheir ownexperience,decisionscouldbebiasedinfavorofonegroupor another.Ifadditionalinformationisreceivedaboutthequalitiesof individualworkersandgroups,thentheincreasedamountofinfor- mationisexpectedtodecreasethestatisticalbiasfromindividual sampling.Inordertoprovidestandardizedconditionsforhiring decisions,informationflowisconsideredasanautomatedprocess inourexperiment.Pleasenotethatwedesignedourexperimentto guaranteethattheinformationexchangedisaccurate.Hence,we avoidedthestrategiccomplexityofrecommendationsinbusiness lifeforthesakeofsimplicity.
Employerswerearrangedinasimplecircularnetwork.Related experimentalresearchoncoordination incircular networkshas foundmixedevidenceaboutconvergencetowardsasinglecon- sensus(Keseretal.,1998;Berninghausetal.,2002).
Biasesofindividualemployersfavoringonegrouporanotherare expectedtobalanceeachotheroutattheaggregatedlevelofthe fairlabormarket.Informationflowbetweenemployers,however, createsnetworkexternalitiesasemployersmightbepronetosocial influence.Inaperfectlybalancedidealisticworldasrepresentedin ourexperiment,socialinfluence(Takácsetal.,2016;Flacheetal., 2017)andthe“wisdomofthecrowd”(Galton,1907;Surowiecki, 2005;Lorenzetal.,2011)areexpectedtodrivethegrouptowards consensus.Flowoftrustedinformationbetweenemployerscould beviewedasaninsurancedevice(cf.GemkowandNeugart,2011) thatbringmarketsclosertoperfection,therebydecreasinginequal- ityinemployment.Experimentsshow,however,thatanaverage initialbiasandevenmildsocialinfluencecouldunderminethewis- domofthecrowd(Lorenzetal.,2011),implyingthatinourcase socialinfluencecouldenlargeratherthandiminishinequalityin employment.
Thisledustoformulatethefollowinghypothesisatthelevelof individualemployersthatcould,butnotnecessarilytranslatedto morebalancedemploymentchancesattheaggregatedlevel.
Hypothesis4. Accesstoreliableinformationinthenetworkof employersaboutindividualworkerqualitiesdecreasesdiscrimi- nationrates.
Interactioneffects
Higherqualitystandardsareexpectedtoincreasediscrimina- tionwhenthereisasymmetryofinformation,becauseemployers aremoredisappointedwithnewlaborforceandrelyextensively onfewhighqualityworkerskeptinhouse(TakácsandSquazzoni, 2015).Atthesametime,additionalandtrustedinformationfrom otheremployerscouldreducethisbiasandevenimprovethesitu- ation.Thisisbecauseemployersaremoremotivatedtohirequality workersandcouldovercometheirintrinsicpartialityforoneofthe groupsassoonasadequateinformationbecomesavailable.
Fortheemployers,workerreferralsenlargethepoolofworkers fromthecategorythatisalreadyoverrepresented.Ashighstan- dardsincreasetheturnoverrateandemployerscontinuetofollow theirworkers’advices,thiscouldpotentiallyresultina positive interactioneffect(e.g.,Takácsetal.,2015).Wehaveexaminedthese potentialinteractioneffectsinourexperimentbytheinclusionof treatmentsinwhichamixtureofourmanipulationswaspresent.
Method
Experimentaldesign
We designed an experiment where participants played the roleofemployersandaskedtohireworkersfortheirfirm. Par-
ticipantsplayed anonymouslyingroupsofsix,withtwogroups playingsimultaneouslyinthesamelaboratorytoavoididentifica- tionofothergroupmembers.Theyinteractedthroughacomputer networkrunningtheexperimentalsoftwarez-Tree(Fischbacher, 2007).Participantswereaskedtoimaginethattheywereemploy- ers and were invited to hire 10 workers per period, which representedonecontractyear.
WorkerswerevirtualagentsandhadIDnumbersfrom1to200.
Eachworkerhadafixedqualitydrawnfromauniformintegerdis- tributioninthe[0,19]intervalatthebeginningofthegame.Halfof theworkerswerelabeledas‘blues’,halfas‘greens’.Colorswerenot relatedtothequalityofworkers:thedistributionofqualitywas, onaverage,thesameforbluesandgreens.Subjectswerenotaware oftheexactdistributionofqualities.Theirinstructions,however, containedthat“ThereareplentyofBLUEandGREENworkersinthe marketwithamaximumqualityvalue”(Supp.Mat.).Inshort,sub- jectswerenotawareofanyinitialdifferencebetweentheworkers’
groupsthatcouldberesponsibleforindividualdiscriminationand suchdifferenceinfactdidnotexist.
Inallexperimentalconditions,participantscouldselecteach worker using one of thefollowing options: (i) hiring a worker randomly;(ii)hiringablueworkerrandomly;(iii)hiringagreen workerrandomly;(iv)hiringoneoftheworkerstheyhiredinthe previousperiod.Toallowparticipantstouse(iv),thelistofworkers hiredinthepreviousperiod,includingtheirqualityandcolor,was displayedontheirscreen(Supp.Mat.).Theseoptionsweresupple- mentedintherespectiveexperimentalconditionswith(v)hiring workersthatwerehiredbytheconnectedemployerintheprevi- ousperiod(informationflow)and(vi)hiringworkersreferredby workersin-house.
Afterthehiringstage(i.e.,attheendofeachperiod),participants wereaskedtoprovideanestimationoftheaveragequalityofboth bluesandgreensintheentirepool,whichservedtomeasuretheir actualbeliefsandprejudice.Wesimulatedturnover(randomfluc- tuation)byexcluding10percentofworkersfromthelistdisplayed tosubjectsineachround.Excludedworkerswereunavailablefor hiringwithprocedure(iv).
Eachexperimentincluded25periods.Theprofitofemployers dependedonthequalityofhiredworkers,whichwasunknown beforehiringexceptthatofworkershiredinthepreviousperiod.
Lowandhighqualitystandardswereimplementedwiththefol- lowingexogenousthresholdrule.Thethresholdwas=12inthe caseoflowqualitystandards(−S)and=17inthecaseofhigh qualitystandards(+S).Thresholdswereimperativeforindividual payoffs.Specifically,profitwascalculatedbysummingthequality ofworkerswithqi≥,anddividingtheresultbyten(thenumberof jobsatthefirm).Therefore,qualitystandardsweresetupbydesign intheexperiment.
Attheendofeachperiod,theaveragequalityofblueandgreen workershiredbytheparticipantwascalculatedanddisplayedwith thenumberofpointsearned.Attheendoftheexperiment,partic- ipantswereaskedtocompletea shortquestionnaire.Individual profitswereaveragedacrossall25periods,witheachpointbeing exchangedwithoneEuro.Earningswerecashedimmediatelyafter theendoftheexperiment.
Wefollowedabetween-subjects2×2×2fullfactorialdesign.
We manipulated:a)highorlow qualityemployerstandards,b) thepresenceofinformationflowbetweenemployers,andc)the presenceofworkerreferrals(Table1).Therefore,weincludedeight treatments,i.e.,fourtestingforeffectsofthesemanipulationsalone andfourexaminingtheirinteractions.
Incaseofinformationflowbetweenemployers(+E),eachpar- ticipantwaslinkedtoanotheremployerinadirectedcirclenetwork withsixnodes(Fig.1).Instructionsreferredtoemployernetwork tiesas“friends”(Supp.Mat.).Thisreferencehasbeenmadeinorder toincreasethecredibilityofinformationconveyedandtomakethe
Table1
Treatmentoverview.
Treatment High standards
Information flow
Worker referrals Pureeffects −S−E−W
+S−E−W 䊏
−S+E−W 䊏
−S−E+W 䊏
Interaction effects
+S+E−W 䊏 䊏
−S+E+W 䊏 䊏
+S−E+W 䊏 䊏
+S+E+W 䊏 䊏 䊏
experimentalsituationmorerealistic.Intheinformationflowcon- dition(+E),ownworkershiredinthepreviousperiod,andthose hiredinthepreviousperiodbythefriendweredisplayedautomat- ically,includingcolorandqualityofworkerswithoutanyerror.
Theaveragequalityofblueandgreenworkershiredbythefriend wasalsodisplayed.Toenhanceinformationflowfurther,subjects receivedtheirfriends’estimatesabouttheaveragequalityofblues andgreens.Intheinformationflowcondition(+E),subjectscould alsoselectaworkerwhowashiredbythebusinessfriendinthe previousperiod(selectionprocedurev).Notethatthismeantthat thesameworkercouldbehiredbytwoormoreparticipantsinthe sameperiodandbestworkerswerenotsubjecttocompetitionby businessfriends.Thisexcludedthepossibilitythatdecisionspeed woulddeterminewhichemployerhiresaworkerwithsuperbqual- ity.Participantsdidnotknowtheentirenetworkstructure;hence strategicconsiderationswereunimportantforwhocouldhirethe bestworkers. In case of noinformation flow (−E),no network existedandnoneofthesubjectsknewtheworkershiredbyother subjectsorothers’estimationsofworkers’quality.
Intheworker referralscondition(+W),eachworker hiredin thepreviousperiod ‘recommended’a friend of the samecolor, randomlyselected,withoutrevealing itsquality.The suggested workerswereshowninaspecificlistontheparticipants’screen.
Incaseof workerreferrals(+W),subjectscouldselecta worker fromthislist(selectionprocedurevi).Incaseofnoworkerrefer- rals(−W),participantsdidnothavealistofrecommendedworkers.
Participantswerefullyawarethatworkerswillgiveahomophilous referral(seefullinstructionsinSupp.Mat.).
Subjects
Atotalof144subjects(56percentfemales)forming24groups participatedintheexperiment,whichwasheldintheGECScom- puterlaboftheDepartmentofEconomicsandManagementofthe UniversityofBrescia,Italy.Subjectswereuniversitystudentvol- unteersrecruitedacrossalluniversityfacultiesusingtheonline platformORSEE(Greiner,2004).Theygaveinformedconsent.Par- ticipantearningsaveragedtoD12.49(std.dev.D3.65),plusafixed show-upfee ofD5.Theexperiment, includinginstructionread- inganda finalquestionnaire,tookapproximatelyonehour.The fullinstructionsforparticipantsareincludedintheSupplementary Material.
Discriminationindexes
Wedefinedmicroleveldiscriminationastheaverageextentto whichindividualparticipantshiredworkersfromasinglegroupas ourmaindependentvariable.Themicroleveldiscriminationindex ıittakesanindividualvalueforeachparticipanti∈{1,...,144}in eachperiodt∈{1,...,25}.Itwasdefinedasoneminustheratioof thenumberofworkersofthemorediscriminatedgrouphiredbyi
Fig.1. Theinformationflownetworkintheexperiment.
Note:Nonetworkinformationwasgiventothesubjects.Subjectshavereceived thefollowinginstructionintreatment(+E):“Duringtheexperiment,someofthe otheremployerswillbeconsideredas‘yourfriends’.Asfriendscommunicatewith eachother,yourfriendemployerswillshareinformationabouttheworkersthey employ.Hence,inthesesituations,youwillgettoknowthequalityoftheworkers employedbyyourfriendsandyoucanalsoselectaworkerfromthislist,whichwill bedisplayedinthemiddleofyourscreen.Inthiscase,youwillbeabletoselect workerswhomyourfriendsemployedinthepreviousyear...”
inperiodttothenumberofhiredworkersofthelessdiscriminated group.Formally:
ıit={ 1−Hitg
Hitb if Hitg≤Hbit 1−Hitb
Hitg if Hitg>Hitb
(1)
whereHitbwasthenumberofhiredbluesandHitgwasthenumber ofhiredgreensbyparticipantiinperiodt.Eq.(1)showsthatıitwas 0incaseofnodiscriminationand1ifallworkerswereofthesame color,withintermediatecasesofdiscriminationfallinginbetween.
Toanalyzewhetheronegroupisfavoredoveranotheroneand discriminationtendenciesspreadorbalanceeachother,wealso definedmacroleveldiscriminationastheobjectiveextenttowhich groupsweredisproportionally hiredinthegivenperiod.Thisis importantbecause,althoughindividualemployersmightbeper- fectdiscriminators,thismaynotgenerateemploymentinequality ifmutualdiscriminationtendenciesacrossdifferentfirmsarebal- anced.
Themacro-leveldiscriminationindextissimilartothemicro- levelindexbutisdefinedatthegrouplevel.Morespecifically,itwas calculatedasoneminustheratioofthenumberofworkersofthe morediscriminatedgrouphiredbyallthesixemployerstothework- ersofthelessdiscriminatedgrouphired.Thismeansthatttook asinglevalueineachperiodtanditsmathematicalformulation wasequivalenttoEq.(1)butHtbandHtg werecomputedasthe sumofallbluesandgreenhiredinperiodt.Asforthemicro-level index,t was0fornoinequalityinemploymentandincreased withincreasingdiscriminationupto1.
Results
Testingforpureeffects
Wewillpresentourresultsinthreesteps.First,wewilldis- cussresultsfromtheno-interactiontreatments,inwhichonlyone manipulationwasintroducedatonce.Thiswillhelptohighlight thepureeffectof highquality standards,informationflow, and workerreferralsondiscrimination.Second,wewilldiscussresults ondiscriminationfromalltreatmentsthatincludedalsointerac- tions.Third,wewillanalyzeemployerearnings.Althoughpayoffs werenotessentialtoourmainresearchquestions,theiranalysis hasimportanteconomicimplications.
Table2
Thenumberofworkershiredthroughdifferentoptions.
Treatment Newworkersrandomly(i-iii) Re-hiringworkers(iv) Hiringfromemployercontact(v) Hiringviaworkerreferrals(vi) Total(i-vi) no-interactiontreatments
−S−E−W 1827 2651 4478
+S−E−W 3093 1405 4498
−S+E−W 782 2542 1155 4479
−S−E+W 1212 2493 792 4497
interactiontreatments
+S+E−W 820 2708 951 4479
−S+E+W 719 2136 1280 351 4486
+S−E+W 1439 2043 985 4467
+S+E+W 745 2005 1209 510 4469
Total 10,637 17,983 4595 2638 35,853
Note:N=3600hiringdecisionswithamaximumof10workersperdecision.
Table3
Meansandstandarddeviationsofdiscriminationindexesinno-interactiontreatments.
Treatment micro–leveldiscriminationindex discriminationindexinrehiring macro-leveldiscriminationindex
N Mean SD SE N Mean SD SE N Mean SD SE
−S−E−W 450 0.283 0.275 0.013 400 0.367 0.275 0.014 75 0.173 0.060 0.034
+S−E−W 450 0.426 0.301 0.014 361 0.514 0.500 0.017 75 0.238 0.015 0.009
−S+E−W 450 0.366 0.227 0.011 402 0.466 0.293 0.015 75 0.322 0.116 0.067
−S−E+W 450 0.438 0.313 0.015 402 0.516 0.325 0.016 75 0.204 0.031 0.018
Note:Themacro-leveldiscriminationindexcouldonlybeaffectedlogicallyinthe−S+E-Wtreatment.
Table4
Mixed-effectsregressiononthemicro-leveldiscriminationindexinno-interactiontreatments(N=1800),withrandomeffectsforindividualsandgroups.Thedegreesof freedomwerecomputedusingSatterthwaite’sapproximation.
(Intercept) Estimate SE t p Estimate SE t p
0.283 0.042 6.736 0.000 0.239 0.083 2.892 0.006
treatmenteffects
+S 0.143 0.060 2.406 0.043 0.150 0.063 2.392 0.045
+E 0.082 0.060 1.384 0.204 0.090 0.064 1.406 0.196
+W 0.154 0.060 2.595 0.032 0.161 0.063 2.552 0.035
subjectcharacteristics
male 0.018 0.036 0.499 0.619
religious 0.006 0.038 0.166 0.868
economics 0.006 0.043 0.140 0.889
studyyear 0.011 0.013 0.859 0.394
part2 −0.016 0.012 −1.385 0.166
F 35.507 0.000 8.479 0.001
Note:Thedummyvariable“economics”hadavalueofoneiftheparticipantwasstudyingeconomicsormanagement.
Theupper four rows in Table2 show thatall available hir- ingmethods(i–vi)wereselectedfrequentlybyparticipantsinthe no-interactiontreatments.Re-hiringofworkers(iv)wasfavored tohiring workersfromemployercontacts(v) andto hiringvia workerreferrals(vi).Re-hiringwasmorecommoninthelowqual- itystandardscondition.Higherqualitystandardsresultedinmore churnandexperimentationwithnewworkers.Table3 displays anoverviewofmeanıit valuesinno-interactiontreatments.All factormanipulationsproducedsignificantlyhigherdiscrimination thanthebaseline(Wilcoxonranksumtestsonindividualaverages:
W=83,p=0.006for+S;W=110,p=0.052for+E;W=89,p=0.010 for+W;allpvaluesareonetailed).
Notethatdiscriminationindexvaluescannotbefullyattributed toonemethodofhiringoranotherasparticipantscouldre-hire workers or select other hiring methods in every round. They typicallyused combinedstrategies and mighthave individually compensatedfortheirownbiasin onetype ofhiringby favor- inganothergroupwhenchoosinganotherhiringmethod.Still,in ordertoseetheextentofdiscriminationthatcouldpotentiallybe linkedtokeepingworkersinhouseselectively,wecalculatedıit valuesforre-hiresonly.Table3showsthesemeanvalues(seemid- dlecolumns).Itisremarkablethatparticipantsweremorebiased
infavorofonegroupwhenre-hiringworkers.Thediscrimination indexforre-hireswashigherineveryexperimentalconditionthan themeandiscriminationvalueforotherhiringmethods.Although thesefactorsareintertwined,itisprobablethatdiscriminationby groupmembershipwasmoreinfluentialinkeepingworkersthan hiringnewones.
Asourdataincludedrepeatedobservationsofthesameindi- vidual,weestimatedamixed-effectsmodelwithrandomeffectsat theindividualandgrouplevel(hereafterMEmodel).Therecould befurtherdependenciesinourdataascurrentchoicesmighthave beeninfluencedbyearlierexperiencesindifferentways(cf.e.g., Ule,2008; Corten,2009).Asweare notinterestedin theexact nature of these dependencies,they are covered by individual- leveleffects,theexperimentalpartdummyandtheothercontrols includedinthemodel.Table4showstheresultsofthemodel.The positiveinterceptindicatesthatasignificantlevelofdiscrimination occurredinalltreatments.Resultssupportedourfirsthypothesis.
Results confirmedalsoHypothesis 2,as higher qualitystan- dards resultedin higherdiscrimination rates. Asformulated in Hypothesis3,workerreferrals(+W)increaseddiscriminationrates.
Meanwhile,Hypothesis4wasnotsupportedastheinformation flow(+E)coefficientwasnotsignificantand,ifanything,thisfactor
Fig.2.Socialinfluenceondiscriminationintheinformationflow(+E)treatment.
Notes:N=1800decisions.Positivevaluesindicatebiastowardsgreensandnegativevaluesindicatebiastowardsblues.Casesatthezeropointareimpartialjudgments.Fitted smoothregressionlineisindicated.
Fig.3.Confirmationbiasintheinformationflow(+E)treatment.
Notes:N=1800decisions.Positivevaluesindicatebiastowardsgreensandnegativevaluesindicatebiastowardsblues.Casesatthezeropointareimpartialjudgments.
tendedtoincreaseratherthandecreasediscrimination.Itisimpor- tanttonotethatcontrolvariablesincludedinthemodel,i.e.,gender, religion,facultystudyyearhadnosignificanteffect.Wealsocon- trolledforapossibleexperimentallearning(orfatigue)effectby includingadummyandfoundnosignificantdifferencesbetween thefirstandsecondhalfoftheexperiment.
Regardingmacro-leveldiscrimination,indexvaluesshowthat individualtendenciestodiscriminatehavebeenbalancedouttoa largeextentatthemacrolevel(Table3).Thetvalueof0.173inthe
−S−E−Wbaselineconditioncorrespondstoanimbalanceofhiring atotalof32.84workersfromonegroupand27.16workersfrom theothergrouponaverage(outof60).Theinformationflowtreat- ment(+E),however,ledtohigherdiscriminationthanthebaseline (Table 3).This was themanipulation under which participants couldhavebeenaffectedbytheselectionofanotheremployer.The highertvalueimpliesthatindividualdiscriminationtendencies ruledeachotherouttoalesserextent.Thesomewhathighermacro- leveldiscriminationindexintheinformationflow(+E)condition
impliesthatsomeemployershaveadoptedthebiasesoforhave beeninfluencedbytheircontacts.Fig.2displaysthatwhilethe mostdecisionswerefreeofbias,participantsslightlyincreasedthe partialityoftheirdecisionfavoringthegroupofworkersthathad ahigheraveragequalityattheconnectedemployer.Hence,par- ticipantspartlyadaptedtheirsamplingrationallygiventhenew availableinformation(cf.LeMensandDenrell,2011).Fig.3high- lightsmoreevidentlythatdiscriminationwascontagiousinthe informationflowtreatment(+E).Alargerbiasbytheconnected employerresultedinlargerdiscriminationinfavor ofthesame groupofworkers,andthereforediscriminationwassubjecttosocial influenceintheexperiment.
Interactioneffects
The lowerfour rows in Table 2 showthat all available hir- ingmethods(i-vi)wereselectedfrequentlybyparticipantsinthe interactiontreatments.Re-hiringofworkers(iv)wasalwaysthe
Table5
Meansandstandarddeviationsofdiscriminationindexesininteractiontreatments.
Treatment micro–leveldiscriminationindex discriminationindexinrehiring macro-leveldiscriminationindex
N Mean SD SE N Mean SD SE N Mean SD SE
+S+E−W 450 0.224 0.206 0.010 421 0.356 0.300 0.015 75 0.146 0.110 0.013
−S+E+W 450 0.413 0.284 0.013 403 0.553 0.336 0.017 75 0.264 0.202 0.023
+S−E+W 450 0.477 0.334 0.016 403 0.633 0.336 0.017 75 0.198 0.132 0.015
+S+E+W 450 0.483 0.292 0.014 370 0.592 0.328 0.017 75 0.385 0.178 0.021
Note:Themacro-leveldiscriminationindexcouldonlybeaffectedlogicallyinthe+Etreatments.
Table6
Mixed-effectsregressiononthemicro-leveldiscriminationindexincludinginteractioneffects(N=3600),withrandomeffectsforindividualsandgroups.Thedegreesof freedomwerecomputedusingSatterthwaite’sapproximation.
(Intercept) Estimate SE t p Estimate SE t p
0.283 0.040 7.064 0.000 0.229 0.062 3.662 0.001
treatmenteffects
+S 0.143 0.057 2.523 0.023 0.150 0.060 2.484 0.025
+E 0.082 0.057 1.451 0.166 0.084 0.061 1.379 0.186
+W 0.154 0.057 2.721 0.015 0.163 0.061 2.699 0.016
+S*+E −0.284 0.080 −3.546 0.003 −0.292 0.086 −3.396 0.004
+S*+W −0.104 0.080 −1.292 0.215 −0.117 0.086 −1.354 0.195
+E*+W −0.107 0.080 −1.329 0.203 −0.118 0.086 −1.376 0.188
+S*+E*+W 0.314 0.113 2.769 0.014 0.335 0.123 2.732 0.015
subjectcharacteristics
male 0.008 0.024 0.349 0.728
religious 0.019 0.026 0.716 0.475
economics 0.027 0.028 0.950 0.344
studyyear 0.006 0.009 0.599 0.550
part2 −0.000 0.008 −0.056 0.956
F 144.831 0.000 105.864 0.000
mostfavoredoption.Table5showsthatdiscriminationinre-hiring washigherineveryinteractiontreatmentthan micro-leveldis- crimination in general.Hiring workersfrom employercontacts (v) and new hires altogether (i-iii) weremore frequentlyused thanworkerreferrals(vi)whentheywereavailable.Treatments includinginteractionbetweenthemanipulatedfactorsgenerally ledtohighermicro-leveldiscriminationthanthebaseline,except when employers with high standards benefited from informa- tionfromotheremployersandworkerreferralswerenotpresent (Table5).Alldifferenceswiththebaseline(−S−E−W)weresignif- icant:−S+E+W,+S−E+Wand+S+E+Wledtohigherdiscrimination (W=82, p=0.005,W=70,p=0.002and W=43,p<0.001respec- tively,allpvalueswereonetailed),while+S+E−Wledtolower discrimination(W=226,p=0.022,onetailed).
We examinedinteraction effectsby estimatinga furtherME model.Notethat,unlikethepreviousmodels,alltreatmentswere includedintheanalysistocontrolforpureeffects.Severalpure andinteractioneffectsweresignificantinthenewmodel(Table6).
Thesignsofallpurefactorcoefficientswerepositive,whichisin linewiththeresultspresentedintheprevioussection.Allbinary interactionsledtonegativecoefficientsbutonlytheinteractionof highqualitystandardsandinformationflow(+S*+E)wassignifi- cant.ConsistentlywithHypotheses2and3,highqualitystandards (+S)and workerreferrals (+W)increaseddiscrimination.In line withHypothesis4,informationflow(+E)decreaseddiscrimination, althoughthathappenedonlyinconjunctionwithhighstandards.
Thecoefficientfortheinteractionofhighqualitystandardsand informationflow (+S*+E)wassufficientlylargetoovercomethe pureeffectsofthefactorsinvolved.Thismeansthatinthespecific caseofhighqualitystandards,informationflowactuallydecreased discrimination,despitetheoppositepureeffectsofthetwofactors.
Thiscouldbeexplainedasfollows.Ontheonehand,incaseof lowqualitystandards,informationflowincreaseddiscrimination viaselectiveattentionorconfirmationbias.Ontheotherhand,in caseofhighqualitystandards,whenavailableinformationreally
mattered, information flow decreased employer partiality. This impliesthatwhentherewasashortageofworkerswithaccept- ablequality,participantshaverealizedthebenefitsofadditional informationandacteduponaccordinglytoavoidprofitloss.
Theinteractionofhighqualitystandardsandworkerreferrals had a similareffect,although weaker.Given that there wasno informationbenefitinthecaseofworkerreferrals,anincreased alertness of employers withhigh quality standards couldhave reduced thepartiality of their judgment. Our results also sug- gestthateventhemutualpresenceofinformationflowbetween employersandworkerreferralscouldhavemadeemployersmore alertaboutpartiality.
Analternativeexplanationofpairwiseinteractioneffectsisthat informationflowandworkerreferralsoffseteachother.Thiscould haveoccurredwhenindividualdiscriminationdidnothaveacoun- terpartintermsofmacroinequalityandneighboringemployers favoreddifferentgroups.Asthethree-wayinteractiontermshows, this balancingtendencyoccurredlesslikely foremployers with highstandards,whowerepickierintheirchoicesandtrappedmore likelybyconfirmationbias.
Exceptofthe+S+E−Wtreatmentthatshowedlowermacro- leveldiscriminationthanthebaseline,informationflow(+E)ledtoa highervalueoft(Table5).Thisimpliesthatalsointheinteraction treatments,individualbiasesdidnotruleeachotheroutcompletely inthepresenceofinformationflowbetweenemployers.Thiscon- firmssomespread,butnotaninevitabledisseminationofbiased evaluationsinthenetworkofemployers.
Effectsonparticipants’earnings
Participants’payoffscouldbeseenasaproxyforthegeneral efficiencyintheallocationofworkerquality.Table7showsanME modelwherethedependentvariablewasthepayoffearnedineach periodofthegame.Followingtheobservationthathigherqual- itystandardsresultedinmoreexperimentationwithnewworkers
Table7
Mixed-effectsregressiononparticipant’speriodearnings(N=600),with randomeffectsforgroups.ThedegreesoffreedomwerecomputedusingSatterthwaite’s approximation.
(Intercept) Estimate SE t p Estimate SE t p
12.388 0.639 19.375 0.000 10.415 0.941 11.074 0.000
treatmenteffects
+S −5.659 0.904 −6.259 0.000 −5.609 0.915 −6.129 0.000
+E 3.497 0.904 3.868 0.001 3.671 0.926 3.964 0.001
+W −0.337 0.904 −0.373 0.714 −0.321 0.916 −0.351 0.730
+S*+E 4.603 1.279 3.600 0.002 4.361 1.303 3.346 0.004
+S*+W 2.654 1.279 2.075 0.054 2.549 1.303 1.956 0.068
+E*+W −0.447 1.279 −0.349 0.731 −0.603 1.302 −0.463 0.649
+S*+E*+W −2.820 1.808 −1.559 0.138 −2.461 1.854 −1.327 0.202
subjectcharacteristics
male 0.123 0.360 0.341 0.734
religious −0.042 0.393 −0.108 0.914
economics −0.278 0.422 −0.660 0.511
studyyear 0.150 0.140 1.074 0.285
part2 3.218 0.105 30.521 0.000
F 472.233 0.000 440.719 0.000
1Itisworthoutliningthatmixedeffectsmodelswithindividualand/orgroup-levelrandomeffects,althoughtheyarenotalwaysstandardforexperimentaldataanalysis, representanefficientmethodthatcanbeprofitablyusedtoanalyzetheoutcomeofrepeatedgamespresentinginteractionsamongparticipants(e.g.,BarreraandBuskens, 2009;Boeroetal.2009a,b;Bravoetal.,2015).
(Table2),qualitystandardshadadirectimpactonpayoffs.Coher- entlywiththerulesofthegame,higherqualitystandardsledto lowerprofits.Incontrast,theinformationflow(+E)treatmentled tohigherearningsthatdemonstratesthedirectpayoffbenefitsof additionalavailableinformation.
Amongtheinteractioneffects,onlytheonebetween+Sand+E onewaspositiveandhighlysignificant,whiletheonebetween+S and+W,alsopositive,wasonlysignificantatthe10%level.None oftheindividualvariableswassignificant,whileearningsclearly increasedduringthegame.Thiswasexpectedasparticipantscould relyonalargersampleofworkerswhosequalityhasalreadybeen discoveredovertime.
Discussionandconclusions
Duetothecomplexityofmechanismsandproblemsoftheir identificationinthefield,wehavedesignedalaboratoryexperi- mentthatismeanttotesttheexistenceofabaselinetendencyfor impartiality,theimpactofahighernecessityofhighqualitywork- ers,theinfluenceofinformationflowinemployernetworks,and theroleofreferralsnetworksondiscriminationinhiringchoices.
Intheexperiment,subjectsplayedtheroleofemployers.Repre- sentingasymmetricinformationintherealworld,subjectswere notawareofworkerqualitiesinadvanceandhadnopriorknowl- edgeaboutthedistributionofqualityforthetwocategories. In fact,workerqualitiesforthetwocategoriesweredrawnfromthe samedistributionandhencetherewasnoreasonforanyimpartial- ityinemployerdecisions.Thisneutralsituationprovidedthebest contrasttodemonstratetheinevitabilityofdiscrimination.
Wefoundthata substantiallevelof discriminationoccurred acrossalltreatments,confirmingthepresenceofabaselineincli- nationfordiscriminationalsoinanimpartialidealworld(Takács andSquazzoni,2015;Takácsetal.,2015).Thisisagainsttypical predictionsbymainstreameconomists,accordingtowhichbelief- relateddiscriminationisnotviableasemployersnotsharingfalse beliefswouldgainacompetitiveadvantage(Arrow,1973;Aigner andCain, 1977).Accordingly,discriminationis costly:themost beneficialemployment strategyinourexperiment wastocom- pletelyneglectcolormarkers.Thereislittleevidence,however,that employerpracticesreflectefficientandrationalresponsestodif- ferencesinskillsandturnovercostsinreality(BielbyandBaron, 1986;Kaufmannetal.,2015).Thisisconfirmedbythepersistence ofdiscriminationinallconditionsofourexperiment.
Wehypothesizedthatmorediscriminationcouldbeinduced byhigherstandardstowardsworkers(Hypothesis2).Whenstan- dardsare sethighexogenously,lower qualityworkers arealso employedbymereexperimentation,and hencepayoffsarealso lower.Ourfindingssuggestthathigherstandardsgeneratealarger biasandinequalityinhiring,confirmingprevioussimulationfind- ings(TakácsandSquazzoni,2015).Subjectsactingasemployers withhighstandards reliedmoreonsupplementaryinformation ofgroupmembershipofworkersalreadyinhouseratherthanon theabundanceofinformationonthequalitiesofworkersscreened before.Inourexperiment,employerswithhighstandardskepta fewhighlyskilledworkersinhouseandthisbiasedtheirjudgment aboutthemeanqualityofsocialcategories.
Wedo notwant toconveythemessagethat thereis agen- eralcognitivebiasofhumandecisionmakers,whichiscorrelated withhigherdesires andmore selectivepreferences.Our results rathersuggestthatanunintentionaladaptivesamplingbiasexists thatoriginatesfromtheover-generalizationofthequalityofthe smallsampleofemployeesinhouse(Denrell,2005;LeMensand Denrell,2011).Thisadaptive samplingbias wasmore probable whenemployerswithhighstandardswereinterestedinthemost qualifiedworkersonly(cf.Polman,2010;TakácsandSquazzoni, 2015).
Consideringcomplexcognitivemechanismsisnotrequiredto accountfor theemergence ofdiscrimination,althoughcomplex cognitivemechanismscouldin fact make thingsworse(Takács etal.,2015).Theeffectofhighstandardsondiscriminationcouldbe enlargedbyjudgmentbias,inparticularbyself-perceivedobjectiv- ity(UhlmannandCohen,2007),cherrypicking,andconfirmation bias(e.g.,Simon,1955,1956).Anotherpotentiallyrelevantmech- anismconcernsexpectancyconfirmationsequences.Inthiscase, individuals wouldbeconditionedbyselectiveattentionleading themtoconsideronlyinformationthatconfirmstheirbeliefs,while ignoringinformationthatcontradictstheirexpectations(Hamilton, 1981;Bergeretal.,1980;DarleyandFazio,1980).
Previousstudiesemphasizedtherelevanceofsocialnetworks forinequalitiesofemployment(e.g.,Montgomery,1991;Stoveland Fountain,2009; Beamanand Magruder,2012).Wages,jobposi- tions, or workconditions coulddepend onsocial networks (cf.
Montgomery,1991,1992;TianandLiu,2018).Weemphasizedthat socialnetworkeffectsinhiringcannotbeconsideredasasingle mechanism.Atleasttwoverydifferentmechanismsareatplay(cf.
Gërxhanietal.,2013).Oneisworkerreferralthatfacilitateslabor
marketsegregationgiventheirhomophilouscharacter.Anotheris informationflowaboutworkersvianetworktiesofemployers,for instance,intheformofrecommendationsoractivesearchsuchas man-hunting.Accordingtothefirstnetworkmechanism,theaffec- tivecontentof socialrelationshipscouldcreate obligationsthat makerecommendingandhiringfriendsmorelikely.Thisdisrupts thebasiclogicofa‘perfect’marketandresultsinsuboptimalallo- cationoflaborforce(e.g.,IoannidesandLoury,2004;Tassierand Menczer,2008).
Ourresultsconfirmedthatworkerreferralsincreasediscrimina- tion(Hypothesis3).Wefoundincreaseddiscriminationwithworker referralsinourexperiment, althoughwe assumednoclustering ofworkerrelationsbyquality.Inamorerealisticscenario,where workersaremorelikelytogivereferralstoworkersofthesame groupandofsimilarquality,onecouldprobablyobserveaneven strongerbiasfavoringonegroupovertheother(cf.Montgomery, 1991). When contacts are homophilous, referral networks can reduceinequalityviaaninvertedadvantagemechanismonlyifini- tialadvantageisnegativelycorrelatedwiththenewstatusvalue (DiMaggioandGarip,2012).
Contrarytoourexpectations(Hypothesis4),informationflow between employers increased discrimination. Previous studies showedthatbusinesscontactsareimportanttoreducetheeffectof informationasymmetryandhelpemployerstoevaluateproperly thelabormarketpotentialofanemployee(e.g.,Rosenbaumetal., 1990;Uzzi,1996).Ontheotherhand,incaseofdifficultevaluation ofquality,networkscanenlargefalsegeneralizationsandmight magnifytherandominitialbias.Evenunprejudicedemployersseek and rely on evaluationsof others. In case of racial discrimina- tion,thisleadstowhatisknownasracialprofiling(Groggerand Ridgeway,2006;FernandezandGreenberg,2013).
Thereisempiricalevidencethatpeoplearelargelyinfluenced bytheirsocialnetwork(e.g.,MarsdenandFriedkin,1993)andfol- lowadvicesometimesevenwhentheyhavedirectobservations (Sommerfeldetal.,2007;Gilbertetal.,2009).Weobtainedsome evidencethatintheinformationflowcondition(+E)ofourexperi- ment,subjectsplayingemployerswereinfluencedtosomeextent byinformation they received. Adifferencein themean quality valuesofworkershiredbytheconnectedemployermadehiring fromthebettergroupmore likely. Participants,therefore, have learntfromtheexperiencesintheirnetwork,whichisconsistent withearlierresearchonlearninginnetworks(e.g.,Barreraandvan deBunt,2009;Hofstraetal.,2015;MasonandWatts,2012).At thesametime,discriminationwasprobablynotfullyconscious andwasnotonlyduetorationalsociallearning,asimpartiality inhiringwasoftencopiedfromtheconnectedemployer.Follow- ingthepreferencesofothersisindicativeofsocialinfluencethat resultedinsomebiascontagioninthenetwork.Furthermore,sub- jectscouldhavedevotedselectiveattentiontothegroupreputation estimatesoftheircontacts.Theypotentiallyconfirmedtheirbelief thathadfavoredoneofthegroupsifthiswasinaccordancewith theestimateoftheotheremployer.Otherwise,iftheestimateof theotheremployerwasnotinlinewiththeirs,theyprobablyhan- dledthisinformationwithlessattention.Thesefactorsaltogether haveresultedinsomespreadofdiscriminationinthenetwork.Indi- vidualbiasesdidnotbalanceeachotheroutcompletely,causinga certainextentofinequalityintheinformationflowcondition.
Furthermore, our resultsshowed that these factors interact witheach otherin complex ways. It is important tonote that thecombinedeffectofthesefactorswasnotastatisticalillusion duetotheparticularmethodused.Wefoundthatemployerswith highstandardshavemadeuseofinformationflowinawaythat decreaseddiscrimination.Theinteractionofhighqualitystandards andworkerreferralshadasimilardiminishingimpact.
Itseemsthatnetworkscontributedtomorepartialdecisionsof employerswithhighstandards,whohadtobemorecarefulabout
whomtheyemploy(TakácsandSquazzoni,2015).Moreover,the jointpresenceofinformationflowbetweenemployersandworker referralsdecreaseddiscrimination,probablybecausegroupquality informationoriginatingfromthesesourcesoffseteachother.We alsofoundathree-wayinteractioneffectofhighqualitystandards, informationflow,andworkerreferrals.Thisinteractioncouldhave occurredbecauseemployerswithhighstandardsweremorelikely tofallintoconfirmationbiasandneglectedthemorebalancedjoint perspectiveofinformationflowandworkerreferrals.
Obviously,ourresultsneedtobeinterpretedcautiouslygiven the high level of abstraction of the study from the real labor market.Severalimportantfeaturesofhiring processeswerenot coveredinourexperimentaldesign.Therewasnoapplicationpro- cedureandhiringbydifferentemployerstookplacesimultaneously andwasrepeatedinthesamefashionseveraltimes.Participants werearrangedaccordingtoadirectedcirclenetworkandnoother informationflowstructureswereinvestigated.Workerswerenot decisionmakers,theyinfactdidnotwork,andtheircharacteristics didnotchangeovertime.Therewerenodifferencesinwagesand everyjobwasacceptabletoworkers.Theseandotherempirically relevantcharacteristicfeaturesofthehiringprocesswereexcluded onpurposeinordertoilluminatetheimpactofqualitystandards, informationflowinemployernetworks,andhomophilousworker referralsonhiringdiscrimination.
Participants themselvesplayingtherole of employersmight havediscountedtheadequatenessofthelabormarketframingfor thedecisionsituation.Theymighthaveperceivedthedecisiontask asanoptimizationproblemthatcanbeapproachedusingacertain hiringstrategy.Infact,thisisnotunrealisticatall,asmanyemploy- eesarestrategic intheirhumanrelations policyandlearn from theexperienceofothers.Conclusionsfromourexperiment,more- over,couldberelevanttocontextsotherthanthelabormarket.
Qualitystandards,recommendations,andinformationflowarerel- evantfactorsinanyrepeatedselectiontask.Thesamemechanisms potentiallycontribute,forinstance,toconsumerdiscrimination.
Theidealisticconditionsinthelabwereaimedtohighlightsome fundamentaltendenciesofdiscriminationthatcouldbestrength- ened,altered,orevencuredinacomplexinstitutionalizedsetting.
Inordertounderstandtheseaspectsbetter,furtherempiricaland fieldanalysisisneeded.Here,forinstance,empiricalstudiesthat lookatcross-culturalcomparisonscouldenrichourunderstand- ing of social and institutional embeddedness of discriminative outcomes(e.g.,Zschirntand Ruedin, 2016; Tianand Liu,2018).
Providinga‘clean’testofcertainsimplehypotheseswhileleaving contextualeffectsinthebackgroundisanimportantsteptounder- standemployerdiscriminationandexplorecountermeasuresthat couldhelpusreducethesedistortions.
Funding
Financial support for the conduct of the research has been providedbytheIntra-EuropeanFellowshipProgramoftheEuro- peanUnion(“0Discrimination”,GA-2009-236953).Thefirstauthor acknowledgessupportfromtheEuropeanResearchCouncil(ERC) undertheEuropeanUnion’sHorizon2020researchandinnova- tionprogramme(GrantagreementNo648693).WethankMarco CastellaniandNiccolòCasnicifortheirpracticalhelp.
AppendixA. Supplementarydata
Supplementarydataassociatedwiththisarticlecanbefound, intheonlineversion,athttps://doi.org/10.1016/j.socnet.2018.03.
005.