ContentslistsavailableatScienceDirect
Labour
Economics
journalhomepage:www.elsevier.com/locate/labeco
Hiring
Discrimination
Against
Transgender
People:
Evidence
from
a
Field
Experiment
Mark
Granberg
∗,
Per
A.
Andersson
,
Ali
Ahmed
Division of Economics, Department of Management and Engineering, Linköping University, 581 83 Linköping, Sweden
a r t i c l e
i n f o
JEL Classifications: C93 J15 J16 J23 J71 Keywords: Field experiment Correspondence test Transgender people Discrimination Labor marketa b s t r a c t
Thispaperpresentstheresultsofthefirstcorrespondencestudythatexaminedhiringdiscriminationagainst transgenderpeople.Fictitiousjobapplications(N=2,224)weresenttoemployerswithjobpostingsin12 low-skilloccupationsinSweden.Overall,40percentofcisgenderapplicantsand34percentoftransgenderapplicants receivedapositiveemployerresponsetotheirapplications.ThisresultwasnotrobusttotheHeckman-Siegelman critique.However,whentransgenderapplicantswerecomparedtothedominantgenderinmale-and female-dominatedoccupations,estimatesofdiscriminationwerelargerandrobusttothecritique.Therewasnoclear supportforthestatisticaldiscriminationhypothesis.
1. Introduction
Transgenderpeople report experiencinghighrates of discrimina-tion.1 Forexample,inanEU-widesurvey,54 percentoftransgender respondents reported that they “have felt personally discriminated againstorharassedbecauseofbeingperceivedastrans” withinthe pre-cedingyear(European UnionAgencyforFundamentalRights,2014). Inasurveyof6,436transgendermenandwomenfrom theU.S.,50 percentreportedexperiencingharassmentatworkand44percent re-portedhavingexperiencedhiringdiscriminationbasedontheiridentity (Grantetal.,2011).Hiringdiscriminationcouldleadtoincreasedrates of unemployment, adversely affecting individual mental health and wellbeing(PaulandMoser,2009).Thismeansthathiring discrimina-tioncouldbeapartialexplanationfortheincreasedratesofpsychiatric hospitalizationandsuicidalityamongtransgenderpeople(Dhejneetal., 2011).Yet,thereisalackofexperimentalstudiesgoingbeyond self-reportedexperiencesthatidentifytheactual discriminationfaced by
∗Correspondingauthor.
E-mailaddresses:mark.granberg@liu.se(M.Granberg),per.a.andersson@liu.se (P.A.Andersson),ali.ahmed@liu.se(A.Ahmed).
1 Inthispapertransgenderdesignatesapersonwhosegenderidentity
mis-matcheswiththesexthepersonwasidentifiedashavingatbirth.Incontrast, cisgenderdesignatesapersonwhosegenderidentitymatcheswiththesexthe personwasidentifiedashavingatbirth.
transgenderpeopleinthelabormarket.Weconductedapreregistered fieldexperimenttodocumentcausalevidenceofhiringdiscrimination againsttransgenderpeoplein12labormarketsectorsinSweden.2
Therehavealreadybeensomeeffortstostudytransgenderpeople in thelabormarketthroughnonexperimental methods.Forexample, Geijtenbeek and Plug (2018) found that Dutch transgender people suffered an approximate11 percent wagepenalty. There are, how-ever, some well-known difficulties in using registry data to study discrimination. First, these nonexperimental statistical approaches cannot fully separate discriminationfrom other explanationsdue to issues with omissionof unobservable productivity-relevant variables (Neumark,2018) which preventsidentificationof average treatment effects. Second,andmorespecifictothecase oftransgenderpeople, registrydatainmostcasescaptures asubsetofindividualswhohave gone through, orstarted,theprocessoflegally changingtheirname and/orgender.InSweden,thefirstpointofcontactforanindividual wishingtostartthisprocesswouldbeapsychologisttodeliveragender
2The preregistration includes a preset data collection period
and an outline of hypotheses, variables and methods (available at: https://aspredicted.org/sn8sj.pdf). For more on preregistration, see Noseketal.(2018).
https://doi.org/10.1016/j.labeco.2020.101860
Received4December2019;Receivedinrevisedform20May2020;Accepted26May2020 Availableonline31May2020
0927-5371/© 2020TheAuthors.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense. (http://creativecommons.org/licenses/by-nc-nd/4.0/)
dysphoriadiagnosis.3Thenumberoftreatmentreferralsforyouth gen-derdysphoriahasgrowndrasticallyinSwedeninrecentyears(Frisén et al., 2017; Socialstyrelsen, 2020).4 As this increase is especially pronouncedamongyoungpeoplewhoaresoontoenterthelaborforce itisimportanttoexplorethehiringsituationtheymightface.
Inthepresentstudy,weusedabroadandmoreinclusivedefinitionof transgenderpeople.Wedefinedtransgenderpeopleasthosewholegally havechangedtheirgenderaswellasthosewhosociallyidentifyasthe oppositegenderassignedtothematbirth.Itisdifficulttosayfor cer-tainhowmanypeoplefitthisbroaderdefinition;accordingonestudy only34 percentSwedish transgendermenandwomenreported hav-inglegallychangedtheirgender(Zelufetal.2016).5Thissuggeststhat alargeproportionoftransgenderpeoplemaynotappearinregistries butdoidentifywiththeoppositegender.Researchhasshownthat26 percentoftransgenderpeopleinSwedenhaveexperienced discrimina-tioninaccesstohealthcare(EuropeanUnionAgencyforFundamental Rights,2014)and69percenthaveexperiencednegativetreatmentin receivinghealthcare(Zelufetal.,2016).Theseexperiencescouldmake itdifficultfortransgenderpeopletostartorcompletetheprocessof for-malandlegalgendertransition,buttheymaystillbeatriskoffacing discriminationinthelabormarket.
Asanalternativetononexperimentalmethods,Bertrandand Mul-lainathan(2004) 6 popularizedtheuse of correspondencetesting to detect discrimination. This method has become the standard way of testing for discrimination in housing and labor markets. Corre-spondencetests have been used to document discriminationagainst many groups in Sweden (e.g., Ahmed et al., 2012; Ahmed et al., 2009; Ahmedand Lång,2017, 2019; Bursell, 2014; Carlsson, 2011; CarlssonandRooth,2007;Rooth,2004;CarlssonandEriksson,2019). Correspondencestudiesarefieldexperimentsinwhichfictitious appli-cationsaresenttorealemployerswithjobpostings.Theassociated cur-riculavitae(CVs)arekeptthesameforallapplicantsbuttherelevant groupsignalisrandomizedbytheresearcher.Ratesofpositiveemployer responsestojobapplicationsfromeachgroupcanthenbecomparedto acquireaveragetreatmenteffects.Themethodhasprovidedcredible evidenceof hiringdiscrimination,especiallyagainstminoritygroups inmanydifferentcountries.(Forrecentreviewsofthisliterature,see BertrandandDuflo,2017;Baert,2018;andNeumark,2018)
BetweenJanuaryandMayof2019,wesent2,224fictitious appli-cationstoemployerswithjobvacanciesin12low-skilledsectorsinthe Swedishlabormarket.WeexploitedthecommonpracticeinSwedenof prefacingone’sCVwithashortcoverletter.Inthisletter,ourfictitious applicantsintroducedthemselvesbyname(e.g.,“Hello,mynameisA Eriksson).” Inthesecondparagraph,theythenmentioned:“Forfuture contactIwouldliketoclarifythatIhavechangedmyname(fromB Eriksson)soafewofmyolddocumentsareinmypreviousname.” For transgenderapplicants,thenamesAandBweredrawnfromapoolof commondifferentlygenderednamesandforcisgenderapplicantsthe nameswheredrawnfromthesamegender.Ourdesignmakesuse of
3 InSweden,5,841peoplewerediagnosedwithgenderdysphoriaasof2018
(Socialstyrelsen,2020).Withapopulationof10,230,185,theproportionwith thediagnosiswas0.057percent.
4 TherewerenearzeroreferralsataStockholmyouthclinicin2006,anumber
whichincreasedto197in2016(Figure1inFrisénetal.,2017).Theincreaseis especiallymarkedamongyoungpeople,andnatalgirls(Socialstyrelsen2020).
5 OwncalculationbasedonnumbersreportedbyZelufetal.(2016).Weused
thenumberofrespondentswhohadchangedtheirlegalgender(114)and di-videditbythenumberofrespondentswhowereeither“transfeminine” or“trans masculine” (149+187)toarriveatthe33.9figure.Includinganyoftheother categoriesinthesurvey,gendernonbinary(346)andcross-dresser(112),would decreasethepercentageandonlyservetostrengthenourpresentargument.
6 This study started a virtual explosion in use of correspondence
stud-iesin recentyears,butuseof the methodgoesas farback asJowelland Prescott-Clarke(1970)anditsawsemi-frequentusebeforeBertrandand Mul-lainathan(2004),forinstance,Neumarketal.(1996).
thefactthatchangingone’snameisfairlycommonplaceinSweden. Be-tween2013and2018,95,090peopleappliedtochangetheirfirstname inSweden(owncalculationsusingdataobtainedfromtheSwedishTax Agency).Itis,therefore,reasonabletoassumethatcisgenderapplicants werenotseenasparticularlyspecialbyevaluatorsbecauseoftheirname change.
Fromaninternationalperspective,studyingthesituationfor trans-genderpeopleinSwedenmeansthatweestimatealowerboundof hir-ingdiscriminationagainstthisgroup.Swedenisconsideredoneofthe mostprogressivecountriesintheworldwhenitcomestogender equal-ity(WorldEconomicForum,2018)andLGBTacceptance(Floresand Park,2018).Furthermore,Swedishtransgenderpeoplereportsomeof thelowestratesofexperienceddiscriminationintheEUwhensearching forwork(EuropeanUnionAgencyforFundamentalRights,2014).
Followingourpreregistrationplan,webeganbyexaminingoverall discriminationagainsttransgenderapplicantsacrossall12occupations in theexperiment.Besidesestimatingtheaveragetreatmenteffectof beingtransgenderonpositiveemployerresponserateswhenapplying forwork, ourdesignallowedustotestapredictionofthestatistical discriminationtheory(Arrow,1973;Phelps,1972);thatdiscrimination shoulddiminishasinformationaboutotherrelevantskillindicators in-creases. Wealsoexamined whetheremployers’based theirdecisions ontheircustomers’tastes(Becker,1957)byexaminingdifferencesin discriminationbetweenoccupationswhichtendtorequireface-to-face customerinteraction(FFCI)andthosewhichdonot.7Wethenapplied Neumark’s(2012)methodtocheckwhetherourestimatesof discrimi-nationwererobusttothecritiquepresentedbyHeckmanandSiegelman (1993).8
Additionally,weexploredtheroleofoccupationalgender distribu-tion.Thiswasmainlymotivatedbyevidencefromcorrespondence stud-iesofhiringdiscriminationwhichtendstoshowheterogeneityacross oc-cupationswithdifferentgenderratios(Rich,2014;Ahmedetal.,2013a; Carlsson,2011).Anotherreasonwasthatsomesurveydatasuggestthat menmaybemoreprejudicedagainsttransgenderpeoplethanwomen (Nagoshietal.,2008).Weperformedsubsampleanalyseswhere occu-pationalsectorswerecategorizedbyoccupationalgenderratio,splitting thetotalsampleintomale-dominated,female-dominated,andmixed oc-cupations. Discriminationin occupationswithdifferentgender ratios wasevaluated inasimilarwaytooveralldiscrimination(i.e., check-ingforsignsofstatisticalandcustomerdiscrimination).Inaneffortto betterunderstandtheheterogeneousdiscriminationuncoveredacross occupationswithdifferentgendercompositions,wealsoexamined dif-ferences inthetreatmentoftransgendermaleandfemaleapplicants. Lastly,wetestedtherobustnesstotheHeckman-Siegelmancritiqueof theestimateddiscriminationinthesegenderratiosubsamples.
2. CorrespondenceStudyDesign
Weconductedourpreregisteredcorrespondencetestbetween Jan-uaryandMay2019.During thistime,wesent fictitiousapplications to2,224employerswhohadpostedvacanciesontheSwedishPublic EmploymentOfficewebsite,thelargestrepositoryforjobvacanciesin Sweden.Thefirsthalfof2019wasarelativelygoodtimefortheSwedish labormarket.Thejobvacancyratewas2.8percentinSwedenwhilethe
7Customerdiscrimination,whichwasalsoproposedbyBecker(1957),has
beenshowntobeimportantinthecontextofotherminorities(e.g.Holzerand Ihlanfeldt,1998).
8This method isanincreasinglycommonrobustness checkin the
litera-ture(forrecentapplications,seeNeumark,2012;Neumarketal.,2016,2019; CarlssonandEriksson,2019)
Europeanjobvacancyratewas2.4percent.9Also,theunemployment ratewas6.8percent,amongthelowestratesforSwedenin15years.10 We sent only one job application to each employer because paireddesignsriskviolatingtheStableUnitTreatmentValue Assump-tion (SUTVA) and can introduce bias into discrimination estimates (Phillips,2019).Also,anunmatcheddesignallowedustouseaprecise signalingoftransgenderidentity.Ascommonaschangingone’sname isinSweden,employerswouldhavelikelygrownsuspiciousiftheyhad receivedtwoapplicationsfromapplicantswhohadbothchangedtheir name.Wefocusedon12occupationalcategorieswhereeducational re-quirementswererarelyhigherthanasecondaryeducation;hencewe refertothemaslow-skilljobs.Weexaminedlow-skilljobsbecausewith thegrowingratesofgenderdysphoriaamongyoungpeople(Frisénet al.,2017;Socialstyrelsen,2020)wewantedtoassessthesituationthey wouldfacewhenenteringthelaborforce.Anotheradvantageto focus-ingonlow-skilljobswasthatthehiringfirmstendedtobesmall busi-nessesandthereforedecision-makingisnotasdispersedthroughoutthe organization.Wewerethereforemorelikelytoelicitthetastesofthe businessownersthemselvesratherthanthoseofanotheremployee.
Theoccupationalcategorieswerechosensothattherewere suffi-cientnumbersofvacanciestoaccommodatestatisticalanalyseswhile representingdifferentoccupationalgenderratios.Occupationalgender ratioswasanimportantcriteriontoconsiderbecausetheseratioshave beenfoundtobeakeyfactorinearlierstudiesofconventionalgender discrimination(Carlsson,2011;Rich,2014).11Ofthe12low-skill occu-pationsincludedinourexperiment,fourweremale-dominated(3.9to 21.9percentfemale),fourweremixed(43.3to62.5percentfemale), andfourwerefemale-dominated(70.1to91.6percentfemale) occupa-tions.Dataongenderratiosinoccupationswerefrom2016andwere basedoneveryemployedpersoninSwedenaged16to64.12
Table1summarizesthedatabyshowingtheoccupationalcategories, theoccupationlevelcharacteristics(genderratioandamountof face-to-faceinteraction),thenumberofapplicationssenttovacanciesineach occupation,andthenumbersofpositiveemployerresponsesthese ap-plicationsgenerated.
Wecodedanyresponsefromemployerswhichseemedpositively in-clinedtocontinuethehiringprocessasa“positiveemployerresponse,” thisincludedbothdirectinvitationstointerviewsandpositivequeries ofmoreinformationabouttheapplicants.Positiveemployerresponses werepolitely declined,assoonaspossible,bytellingemployers that theapplicantalreadyhadfoundanotherjob.Someemployerresponses weredifficulttocategorizeaspositiveornegative.These50responses werecategorizedaccordingtoourbestjudgmentandweremarkedas uncertain.IntheSupplementaryMaterial(TableA8)weshowthatour resultsarerobusttorecodingthese50casesaseitherallpositiveorall negative.
9 Source: Eurostat.Available at: https://appsso.eurostat.ec.europa.eu/nui/
show.do?dataset=jvs_q_isco_r2&lang=en
10 Source: Statistics Sweden. Available at:
https://www.scb.se/hitta-statistik/statistik-efter-amne/arbetsmarknad/arbetskraftsundersokningar/ arbetskraftsundersokningarna-aku
11 Wealsotestedthisonpre-existingdatacollectedin2016(Thedatawere
originallyusedinAhmed,Granberg,andLång2017;AhmedandLång2017, 2019).Thisanalysisconfirmsthatoccupationalgenderratioisimportantwhen studyingconventionalsexdiscriminationinSweden.Resultsfromthatanalysis: Femaleapplicantswerefavoredingeneral;thiseffectwasdrivenby female-dominatedoccupationswhiletherewasnosignificantmale-femaledifference inpositiveemployerresponserateseitherinmixedorinmale-dominated occu-pations.SeeSupplementaryMaterialformoredetails.
12 Source:StatisticsSweden.Availableat:https://www.statistikdatabasen.scb.
se/pxweb/sv/ssd/START__AM__AM0208__AM0208B/YREG61/?rxid= 0e62353a-d331-4916-8730-8997ab1b72af
2.1. Procedureandexperimentalmaterials
InSweden,jobsaretypicallyappliedtousingashortletterof in-terest(acoverletter)followedbyaCV,aformatwhichwefollowed (seetheSupplementaryMaterial,Example A2,foratranslated exam-pleofanapplicationletter).Eachprospectiveemployerwascontacted throughemailandwassentasingleapplicationwheregenderidentity oftheapplicantandfourattributesrelatingtoobservableworkerskills wererandomized.13 These randomizedelementswillbe discussedin Sections2.2and2.3.
Thefictitiousapplicationscontainednospecificaddressandthe con-tactinformationconsistedsolelyofanemailaddressandamobile tele-phonenumber.Allapplicantshadthesamebirthdate(thisiscommon toreportinSwedishjobapplications)andstatedthattheywere28years oldintheintroductiontotheircoverletter.Wedecidedtonotgivea postaladdressbecauseearlierstudieshavefoundthatemployerstend tonotrespondtoapplicantsviaordinarypostinSweden(Ahmedetal., 2012).Theapplicationincludedinformationabouteducationandwork experiencewhichalsoimpliedthattheapplicantshadgrownup and livedaroundStockholm.Tobe abletoapply tojobpostingsallover Sweden,andnotjustinthegreaterStockholmarea,theapplicantsstated earlyintheircoverletterthat“…IthinkIwouldbeagreatfitforthejob, andIamabouttomovetothearea” wheretheword“area” was man-uallychangedtothenearestmajorpopulationcenterbasedonthejob beingappliedto(IntheSupplementaryMaterial,TableA14,weshow thatresultsareunaffectedbyusingdifferentcontrolsforjoblocation).14 Anyreferencetojobtitleintheapplicationletterwascustomized man-uallyforeachjob.
2.2. Experimentalmanipulationoftransgenderidentity
Eachapplicant’sgenderidentity(i.e.,cisgenderortransgender)was signaledthroughastatednamechangeintheapplication.Transgender applicantsstatedthattheyhadchangedtheirnamefromafemalename toamalenameorviceversa.Cisgenderapplicantsstatedthattheyhad changedtheirnamefromonemalenametoanotherorfromonefemale nametoanother.Hence,oursignalofbeingacisgenderortransgender personwascleartoanyonereadingtheapplicationwithoutthefictitious applicant explicitlystatingtheir genderidentityorhavingtoprofess affiliationwithsome LGBTinterestgroup(awaysometimesused to signalhomosexualityinthesetypesofstudies,see,e.g.,Drydakis,2015). WeconsultedhumanresourcespersonnelatalargeSwedish univer-sitytogettheirinputonhowtomakethename-changesignalfeelas naturalaspossible.Wealsoconsultedwithsmallbusinessownersinthe restaurantindustrywhoprovideduswithsomeexamplesoftypical ap-plicationstheyhadreceived.Followingthisfeedback,thenamesofthe fictitiousapplicantswereshownthreetimes;intheinitialandclosing greetingsinthecoverletterandatthetopoftheCV.Thebirthname wasfirstpresentedinaparentheticalsentenceinthesecondparagraph ofthecoverletter.Thepassageread:“ForfuturecontactIwouldliketo clarifythatIhavechangedmyname(fromBEriksson)soafewofmy olddocumentsareinmypreviousname,butwiththesamesocial secu-ritynumber.” Theinformationaboutthenamechangewasreiteratedin parenthesesatthetopoftheCVnexttothecurrentnameforinstance, “AEriksson(prev.BEriksson).” Weareconfidentthatthesignaledname changewascleartomostrecruiters,notleastbecauseafewemployers explicitlyacknowledgeditduringtheexperiment.
13Weavoidedapplyingforjobsthatrequirefillinginstandardizedonline
ap-plicationformsfortworeasons:(i)theytypicallyrequirepersonalinformation whichisdifficulttofalsifysuchasSwedishsocialsecuritynumbers;and(ii) theyoftenrequireacceptingthetermsofservicerelatingtothestorageof per-sonalinformation.Withfictitiousapplications,thistypeofacceptanceis ethi-callyquestionable.
14OwntranslationfromSwedish.SeeSupplementaryMaterialforatranslated
Table1
OccupationalCategories
Occupation Gender ratio % FFCI CVs sent Positive responses (%) Male-dominated
Vehicle mechanic 3.9 Low 170 62 (36.5)
Forklift operator 6.2 Low 35 9 (25.7)
Delivery/Truck driver 8.4 High 230 117 (50.9)
Warehouse worker 21.9 Low 185 33 (17.8)
Mixed
Telemarketing 43.3 Low 92 61 (66.3)
Chef 52.5 Low 343 151 (44.0)
Waitstaff 58.9 High 274 107 (39.1)
Store clerk 62.5 High 226 35 (15.5)
Female-dominated
Customer service 70.1 Low 116 42 (36.2)
Cleaner 75.1 High 303 96 (31.7)
Childcare 82.1 High 127 46 (36.2)
Enrolled nurse 91.6 High 123 65 (52.9)
Total (Mean) (50.2) - 2,224 824 (37.1)
Note.—“FFCI” is short for face-to-facecustomer interaction.“Gender ratio %” is the percentageoffemaleworkersinthegivenoccupationin2016(datafromStatistics Sweden)themeansinparenthesisonthebottomrowreportsthesamplemean,notthe meanofthecolumnvalues.TheoccupationalcategoriesofStatisticsSwedenandthe onesusedontheSwedishPublicEmploymentOfficewebsitedonotmatchcompletely, andassuchwehaveusedthenumbersfortheoccupationalcategorywhichwefound tobethemostsimilar.Theupperfouroccupationsareclassifiedasmale-dominated becausethefractionofwomenislessthanonethird,likewisethelastfourwereclassified asfemale-dominatedbecausethefractionofmenislessthanonethirdandlastlythe middlefourwereclassifiedasmixedoccupations.
Figure1. MeanPositiveEmployerResponseRatesforDifferentApplicants.Note.—Thesefiguresshowthemeanpositiveemployerresponseratesforapplicants withdifferenttypesofnamechangesinoursample.Theblacklinesindicate95percentconfidenceintervals.Thetypesofnamechangesare:FemaletoFemale(FF), FemaletoMale(FM),MaletoFemale(MF),andMaletoMale(MM).GraphAaggregatesFMwithMFintoatransgendercategoryandFFwithMMintoacisgender category.
Both cisgender and transgender applicants disclosed the name changeintheircoverletter.Thestatedreasonfordisclosingthis infor-mationwasthatsomeoftheapplicants’documentation,suchas refer-encesorgrades,wouldbeintheoldnameandtheprospectiveemployer shouldbeawareofthisifgoingforwardwiththeapplication.The de-signmimickedtheway,forexample,newlymarriedpeopledealwitha surnamechangeinofficialmattersuntiltheirdocumentationisupdated withthenewsurname—theysimplyprovideboththeformerand cur-rentname.Inourcase,theapplicantschangedtheirgivennamesand, therefore,informedtheemployeraboutboththeformerandcurrent givenname.
Atthis pointitisworth mentioning thatitis notuncommon for Swedish people to change their given names, even for reasons that havenothingtodowith genderchange.Accordingtodatafromthe Swedishtaxagency,95,090peopleinSwedenofficiallychangedtheir
givennamebetween2013and2018(owncalculationsbasedondata re-ceivedfromtheSwedishTaxAgency).Somepeopleinformallychange thenametheyuseforeverydaypurposes,prénomusuel,toanotherof theirgivennames.Itis,therefore,altogetherreasonabletoarguethat employersareaccustomedtopeoplechangingtheirnames.Alongwith thefactthatapplicantsinourstudygenerallyexperiencedhigherrates ofpositiveemployerresponsesthandidapplicantsinpreviousstudies, makesusconfidentthatourcisgenderapplicantswerenotseenasless employableorodderbecauseoftheirnamechange.
Acritiquesometimesleviedatcorrespondencestudiesisthatnames carrymoresignalsthantheresearchermayintend.Anexampleisthat socioeconomicsignalsmayaffecttheresults ofcorrespondence stud-iesofracediscriminationifnamesareusedtosignalgroupbelonging (FryerandLevitt,2004).Toaddressthispotential issue,eachof our applicationshadtworandomlyassignedcommonSwedishfirstnames
drawnfromalistoffiveforeachgender(femalenames:Amanda,Elin, Emma,Hanna,Julia;malenames:Anton,Erik,Filip,Oscar,Simon).The nameswerepickedfromalistofthefivemostcommonSwedishbaby namesforeach genderin1998.15 Thisensuredthatany signals con-tainedwithinthesenameswerenotheldconstantwithinan experimen-talcondition(addingfixedeffectsforthesenamesdoesnotaffectoverall results,seeSupplementaryMaterial,TableA14).Forexample,boththe cisgenderandtransgenderapplicantcouldhavehadthesameSwedish name,forinstance,Simon.Theonlydifferencewasthatthenamethey statedtohavechangedfromwasoneoftheothermale(forcisgender) orfemale(fortransgender)names.Thismeansourexperimentdidnot restontheuseofsomespecificnametosignalgroupbelongingasin many otherstudies. Yet,we exploitedthegender saliencyof names toclearlycommunicateapplicants’genderidentity.Bothcisgenderand transgenderapplicantsweregiventhefifthmostcommonSwedish sur-name:Eriksson.Usingthesecommoncombinationsofnamesensured thatourapplicantsweredifficulttofindin searchesonsocialmedia platformsandinotheronlinesourcesofpersonalinformation.
2.3. TheHeckman-Siegelmancritiqueandrandomizationofobservables
Besidesrandomizingapplicants’genderidentity,fourother observ-ableswererandomizedintheCVs:(i)thenumberofyearsofeach ap-plicant’srelevantworkexperience,rangingfromoneyeartonine;(ii) whethertheapplicantwascapableofusingacomputer,(iii)whetherthe applicantwasproficientinbothSwedishandEnglish;and(iv)whether theapplicanthadinterestsandhobbiesthatimplyphysicalactivity out-sideofwork. Previousstudieshaveincludedvariationsinadditional informationsuchaslanguageskills(LeeandKhalid,2016),personality characteristics(Drydakis,2014),computerskills,andworkexperience (BertrandandMullainathan,2004).Inthisstudywerandomizedallfour ofthesevariablessothatanapplicationcouldcontainanycombination ofskills.
VariationinskillswasrequiredtouseNeumark’smethodtorespond toHeckmanandSiegelman’s(1993)critique.Theessentialpartofthe critiqueis thatestimates fromfieldexperimentalstudiesinthelabor market can be biasedby differences in thevariances of unobserved productivityacross groupsevenifobservableproductivityindicators arekeptconstantacrossexperimentalconditions.Weacknowledgethat differentialtreatmentbased onassumptions aboutdifferences in the variancesbetweengroupsshouldstillbeconsidereddiscrimination,as stereotyping isessentially assumingavery peakedprobability distri-butionfor theproductivityof astereotyped group.Indeed, such dif-ferentialtreatmentbasedonthevarianceinunobservedproductivity issometimescalledsecond-momentstatisticaldiscriminationHowever, theproblemisthatcorrespondencestudyestimatesarebiased depend-ingonthequalityoftheCVsusedinthestudyandhowthatquality relatestothethresholdofacceptanceagainstwhichitisbeingjudged. Bothoftheseareverydifficulttocontrolfor(forsomeintuitive exam-plesseeHeckman,1998).
NeumarkandRich(2018)reanalyzeddatafromsomeearlier cor-respondence studies and found that accounting for the Heckman-SiegelmancritiqueusingNeumark’smethodchangedtheresultsfora fewofthere-examinedstudies.Neumark’smethod disaggregatesthe estimateofdiscriminationintoalevelpartwhichincludesboth taste-basedandfirst-moment statisticaldiscriminationandavariancepart whichisduetosecond-momentstatisticaldiscrimination.Weusedthis method as a robustness check. We used the same Stata code as in Neumarketal.(2016).
Thesetofskillswasgeneralenough tohave someapplicationto eachoccupationinthestudy.Wedidnotwanttousedifferentskillsfor
15 Source: Statistics Sweden. Available at:
https://www.scb.se/hitta-statistik/statistik-efter-amne/befolkning/amnesovergripande-statistik/ namnstatistik/
Table2
PositiveEmployerResponseRates– FullSample
Positive response FF MM FM MF No 302 337 380 381 (57.97) (61.27) (66.43) (65.58) Yes 432 213 192 200 (40.34) (38.73) (33.57) (34.42) 𝜒2 (1, 𝑁 = 2 , 224 ) = 10.908, 𝑝 = . 012 Cisgender Transgender No 639 761 (59.66) (66.00) Yes 432 392 (40.34) (34.00) 𝜒2 (1, 𝑁 = 2 , 224 ) = 9.562, 𝑝 = . 002 Note.—Percentagesin parentheses.Thep-valueswere pro-ducedwiththePearson𝜒2-test.
differentoccupationsbecauseitwouldmakeitdifficulttoaccountfor theHeckman-Siegelmancritiqueacrossoccupations.Further,thereare notenoughjobpostingswithinmostoccupationsinSwedentoallowfor useofNeumark’smethodwithinsingleoccupations,unlessdatawereto becollectedforaverylongtime.
3. AnalysisandResults
AswecanseeinPanelAofFigure1,therewasadifferencein posi-tiveemployerresponsesbetweencisgenderandtransgenderapplicants. PanelBshowsthepositiveemployerresponseratesforallfourtypesof namechanges:male-to-female(MF),female-to-male(FM),male-to-male (MM)andfemale-to-female(FF).Itisevidentthattherewaslittletono differenceinmeanpositiveemployerresponsesbetweentheMFandFM categoriesandsomedifferencesbetweenMMandFFcategories.The dif-ferencesbetweenMMandFFwasnotunexpected;conventionalsex dis-criminationisalargefieldofstudybutoutsideourscope.Inthisstudy, however,weneededtodecidewhichoftheconventionalgender cate-gories,MMandFF,weshoulduseasareferencegroupintheanalyses. WecouldincludeacontrolforeitherMMorFF,leavingtheotherasthe referencegroup,andarriveatdifferentlevelsfordifferencesinpositive employerresponsestotransgenderapplicants.Alternatively,wecould abstainfromincludingsuchcontrolsandhaveourestimatereflectthe differencebetweenthemeanpositiveemployerresponserateforMM andFFapplicantstogether,compared totransgenderapplicants.Our solutionwastodobothwheneverpossible:Wegeneratedresultsusing bothabinarycisgender/transgenderspecificationandafour-category MM/FF/MF/FMspecificationwiththehighestmeanresponserategroup asthereferencecategory,whichwasusuallyFFexceptwhenwe subdi-videdthedatabygenderratioinoccupation.
3.1. Overalldiscriminationagainsttransgenderapplicants
Table2showstherawpositiveemployerresponseratesforalltypes of applicantsandtheresultsof 𝜒2-tests ofno differencesin positive
employerresponseratesacrossapplicantswithdifferent gender iden-tities.16Ifourrandomizationofgenderidentitywassuccessful,lower positiveemployerresponseratesfortransgenderapplicantsshouldbe anestimateoftheaverageeffectoftransgenderstatusonreceivinga positive employerresponse.Theprocedureweusedwhengenerating theCVsandassigningthemtovacanciesensured thattreatmentwas randomlyassigned,ascorroboratedbyatestforbalanceofcovariates betweenourtreatmentandtheobservablesinourdata(see Supplemen-taryMaterial,TableA1).
16Therewasatypoinourpreregistrationstatingthatwewouldbeperforming t-testsofourhypotheses.Thiswasneverourintention,butweperformedall relevanttestsast-testsaswell.Resultswerenotdifferenttoanyappreciable degreeandarepresentedintheSupplementaryMaterial(TableA8).
Table3 Variables
Variables Explanation
Outcome variable
Positive response 1 if the applicant received a positive employer response, 0 otherwise Treatment variable
Transgender 1 if the applicant changed gender with name change, 0 otherwise
Skill variables
Work experience 1-9 indicating the years of relevant work experience in application
Square of work exp. Work experience times work experience
Computer skills 1 if the applicant claimed to have computer skills, 0 otherwise
Language skills 1 if the applicant claimed to have skills in Swedish and English, 0 otherwise Active lifestyle 1 if the applicant claimed to have an active lifestyle, 0 otherwise
Occupation level control variables
Occupation fixed effects Fixed effect dummies for each of the twelve occupations Gender ratio Takes the values from Table 1 , column 2 for each occupation
High face-to-face customer interaction (FFCI) 1 if occupation was classified as typically requiring high amounts of face-to-face customer interaction, 0 otherwise Vacancy controls
Full time work 1 if job posting was offering full time work, 0 otherwise
Indefinite contract length 1 if job posting was offering work with no determined period length, 0 otherwise Month fixed effects Fixed effect dummies for each month of data collection
Urban 1 if job posting was for work in one of the three greater metropolitan areas of Sweden, 0 otherwise
Note.—Thistablecontainsdefinitionsofthevariablesusedinregressions.Naturally,occupationFEandotheroccupation-levelvariables(genderratioandFFCI)are neversimultaneouslyincludedinregressions.
Figure2. PredictedMarginalPositiveEmployerResponseRatesAcrossWorkExperience.Note.—GraphAshowsthepredictedprobabilityofapositiveemployer responseforcisgenderandtransgenderapplicantsgivendifferentamountsofworkexperience.Theunderlyingprobitmodelincludesthefullsetofcontrolsdetailed inTable3.GraphBshowsdifferencesinmarginaleffectsbetweencisgenderandtransgenderapplicantsandthe95percentconfidenceintervalforthatdifference. Notethedifferencesiny-axisscales.SeeSupplementaryMaterialTableA3forprobitestimates.
Overallandinrawfigures,transgenderapplicantsreceivedpositive employerresponsesin34.0percentofthecasesandcisgenderapplicants receivedpositiveemployerresponsesin40.3percentofthecases.There wasastatisticallysignificant6.3percentagepointpenaltyinpositive employerresponserateforbeingtransgender,𝜒2(1,𝑁=2,224)=9.56, 𝑝=.002.17 Inotherwords, anapplication fromacisgenderapplicant was18.6percentmorelikelytoreceiveapositiveemployerresponse thananapplicationfromatransgenderapplicant.
Tocontrolforpotentialconfoundingfactors,weusedaprobitmodel toestimatethemarginaleffectofbeingtransgenderontheprobability
17 Asthisisthefirsttimewereportedstatisticalsignificanceinthispaper,it
isworthemphasizingthatweusedstricterthresholdsforstatisticalsignificance thanusual(inthespiritofBenjaminetal.,2018).Wheneverp-valuesarenot reported,oneasteriskwill,therefore,indicatesignificanceatthe5percentlevel, twoasteriskswillindicatesignificanceatthe1percentlevel,andthreeasterisks willindicatesignificanceatthe0.1percentlevelofsignificance.
ofreceivingapositiveemployerresponse.Differentsetsofcontrol vari-ableswereusedineachmodeltoexaminetherobustnessofourresults. Insomemodels,fixedeffectswereusedtofullycontrolfortheoretically importantconfoundingvariablesattheoccupationlevelsuchas gen-derratioandface-to-facecustomerinteraction.Theprobitwasofthe followingform:
Pr(𝑦𝑖=1|𝑥𝑖,zi,𝛾𝑖)=Φ(𝑥′𝑖𝛽 +𝑧′𝑖𝛼 +γ′𝑖𝜃),
wherethedependentvariableyiwasadummytakingthevalue1ifthe applicationreceivedapositiveemployerresponseand0otherwise.xi wasavectorofexplanatoryvariables,ziwasavectorofproductivity relevantskills,and𝛾iwasavectorofothercontrolvariables.Thefull listofvariableswiththeirdefinitionsispresentedinTable3.
Table4presentsprobitestimatesofthemarginaleffect(atmeans) from disclosing transgender statusin applications. The estimates in panelAshowthattransgenderapplicantshadbetween6.3and7.1 per-centage pointslowerprobability ofreceivingapositiveemployer
re-Table4
ProbitEstimatesofDiscriminationAgainstTransgenderPeople
(1) (2) (3) (4)
Panel A: Cisgender and transgender specification
Trans -0.0632 ∗∗ (0.0206) -0.0646 ∗∗ (0.0207) -0.0700 ∗∗∗ (0.0210) -0.0707 ∗∗∗ (0.0211) Panel B: Four gender identities specification
FF REF REF REF REF
MM -0.0319 (0.0294) -0.0305 (0.0295) -0.0339 (0.0298) -0.0310 (0.0299) MF -0.0789 ∗∗ (0.0292) -0.0793 ∗∗ (0.0293) -0.0839 ∗∗ (0.0298) -0.0827 ∗∗ (0.0299) FM -0.0800 ∗∗ (0.0294) -0.0809 ∗∗ (0.0294) -0.0907 ∗∗ (0.0299) -0.0903 ∗∗ (0.0300) Skills Experience 0.0359 ∗ (0.0182) 0.0362 ∗ (0.0183) 0.0355 (0.0186) 0.0362 (0.0186) Experience 2 -0.0027 (0.0018) -0.0027 (0.0018) -0.0027 (0.0018) -0.0028 (0.0018) Computer -0.0026 (0.0018) (0.0018) -0.0027 (0.0018) -0.0027 (0.0018) -0.0028 Language 0.0063 (0.0206) 0.0055 (0.0207) 0.0042 (0.0210) 0.0034 (0.0210) Active 0.0089 (0.0206) 0.0109 (0.0207) 0.0208 (0.0210) 0.0226 (0.0211) Other controls
FFCI and gender ratio No Yes No No
Month FE No Yes No Yes
Occupation FE No No Yes Yes
Observations 2,224 2,224 2,224 2,224
Note.—Thistablereportstheestimatesofdiscriminationagainsttransgenderpeopleusingprobitmodels withmarginaleffects(atmeans)anddifferentsetsofcontrolvariables.PanelAshowsthemaineffectof interestformodelswithacisgenderandtransgendercategorization.PanelBshowsboththemainandskill effectswhenusingmodelswithfourgendercategories,withFFasthereferencegroup.Vacancycontrols areincludedinallmodels.Standarderrorsinparentheses.“FE” isshortforfixedeffectsand“FFCI” isshort forface-to-facecustomerinteraction.VariablesareexplainedinmoredetailinTable3.
∗p< .05. ∗∗p<.01. ∗∗∗p<.001.
sponsethancisgenderapplicants. Additionally,inthemostcomplete models(Models3–4)theeffectisstatisticallysignificantatthe0.1 per-centlevel.Theresultsforthefour-gendercategorizationinpanelBare alsostableacrossmodelspecifications.Becausethereferencegroupin panelBisthemostfavoredgender,FF,thepointestimatesareofcourse slightlylargerinabsolutetermsthaninpanelA.TheestimatesinPanel BforMFandFMareallstatisticallysignificantatthe1percentlevel.
TheresultsinTable4lineupwellwiththesimpletestspresented inTable2andreaffirmthatcisgenderapplicantsenjoyedatleasta20 percentadvantageinpositiveemployerresponserates over transgen-derapplicants.AnotherinterestingfindinginTable4isthattheonly skillemployersseemedtohavevaluedwasworkexperience,whichhad asmallbutstatisticallysignificantpositivemarginaleffectonpositive employerresponseratesinsomemodelspecifications(skillestimatesare onlyshownforthefour-gendercategorizationspecificationsinthetable, butweresimilarforthebinarymodel).Thethreeotherskills—computer skills,languageskills,andanactivelifestyle—allshowedstatistically in-significanteffectsonpositiveemployerresponserates.
Weproceededwithouranalysisbydividingthesampleintothree categoriesbased ontherandomizedworkexperiencefor each appli-cant,low(1–3years),medium(4–6years),andhigh(7–9years).We wanted toexamine whether discriminationdiminished as the infor-mation provided about the applicants’ qualificationsincreased. This wouldbethecaseiftheobserveddifferencesinpositiveemployer re-sponserateswereduetostatisticaldiscrimination.Wefoundthat trans-genderapplicantswithlowandmediumamountsofexperiencefaced no statistically significant discrimination—𝜒2(1,𝑁=713)=2.45,𝑝= .118,and𝜒2(1,𝑁=756)=2.54,𝑝=.111,respectively—while
transgen-derapplicantswithhighexperiencedid,𝜒2(1,𝑁=755)=4.75,𝑝=.029.
InFigure2,weplottedthepredictedprobabilitiesofpositiveemployer responsesforcisgenderandtransgenderapplicantsfordifferentamounts ofworkexperience,estimatedusingaprobitmodel.Thefigureclearly showsthatthedisparatetreatmentwhichtransgenderapplicantsfaced asregardspositiveemployerresponserates didnot diminishastheir reportedworkexperienceincreased.Takentogether,boththe𝜒2-tests
andtheprobitmodelshowninFigure2providenoevidenceof statisti-caldiscrimination.18
Next,we investigatedifthere wereanypatternsin thedatathat wereconsistentwithtaste-baseddiscrimination.Wedidthisby split-tingthesampleintotwosubsamples,onewithoccupationswhich re-quiredagreatdealofface-to-facecustomerinteraction(FFCI)and an-otherwithlowFFCIoccupations(Table1).Weobtainedsometentative evidencethatemployersmayhaveactedontheircustomers’tastesas discriminationwas onlystatistically significantin highFFCI occupa-tions,𝜒2(1,𝑁=1,283)=7.65,𝑝=.006,andnotinlowFFCI
occupa-tions,𝜒2(1,𝑁=941)=2.36,𝑝=.125.19However,whenweestimated alinearprobabilitymodel(LPM)ofaspecificationsimilartoModel2
18SeeSupplementaryMaterial,TableA3,forregressionestimates,TableA4for 𝜒2-tests,andFigureA10foragraphicalillustrationofpositiveresponseratesfor
allfourgendercategoriesacrossworkexperience.Allinlinewiththeconclusion thatoveralldiscriminationwasnotstatistical.
19TableA5(PanelA)intheSupplementaryMaterialshowsthese𝜒2-testsmore
Table5
Neumark’sMethod,AddressingtheHeckman-SiegelmanCritique
(1) Transgender and Cisgender (2) FF control included (3) MM control included
Reference category MM and FF MM FF
Panel A: Probit estimates
Trans -0.0707 ∗∗∗ -0.0555 ∗ -0.0865 ∗∗∗
(0.0211) (0.0257) (0.0260)
Panel B: Heteroskedastic probit estimates
Trans -0.0770 ∗∗∗ -0.0594 ∗ -0.0953 ∗∗∗ (0.0215) (0.0273) (0.0280) Trans-level -0.0375 -0.0199 -0.0558 (0.0277) (0.0333) (0.0322) Trans-variance -0.0395 ∗ -0.0395 ∗ -0.0395 ∗ (0.0201) (0.0201) (0.0201) Panel C: Tests
Ratio S.D of unobservables (Trans/cis) 0.7439 ∗ 0.7440 ∗ 0.7440 ∗
Diagnostic tests (p-values)
Test S.D. ratio = 1 0.0230 ∗ 0.0232 ∗ 0.0232 ∗
Overidentification test 0.6531 0.6512 0.6512
LR test: Probit vs. Heteroskedastic Probit 0.0460 ∗ 0.0462 ∗ 0.0462 ∗
Controls
Skills and vacancy Yes Yes Yes
Month FE Yes Yes Yes
Occupation FE Yes Yes Yes
Observations 2,224 2,224 2,224
Note.—InpanelAandBmarginaleffects(atmeans)arereported.PanelBreportstheestimatesofdiscriminationagainst transgenderpeopleusingNeumark’smethodofaddressingtheHeckman-Siegelmancritique.Theincludedskill,vacancy,and FEcontrolsarethesameinallmodelsandequivalenttothemodelspecificationinModel4ofTable4.Models2and3inthis tableshowthattheNeumarkresultsarerobusttoincludingcontrolsforeitherMMorFFapplicants,i.e.changingthereference group.Standarderrorsinparentheses.“FE” isshortforfixedeffects,“LR” isshortforLikelihoodRatio,and“FFCI” isshortfor face-to-facecustomerinteraction.VariablesareexplainedinmoredetailinTable3.
∗p<.05. ∗∗∗p< .001.
inTable4wefoundnosupportforaninteractioneffectbetweenthe transgenderdummyandtheFFCIdummy(𝛽 =.023,𝑝=.579).20,21
Table5presentstheresultsofapplyingNeumark’smethodtoour datafocusingonthemostcompletespecificationinTable4(Model4). Neumark’smethodusesaheteroskedasticprobitmodelanda decompo-sitiontoestimatehowmuchoftheheteroskedasticprobitestimateis at-tributabletodifferencesinthevarianceofunobservablesandhowmuch isnot(“level”).FormoredetailsonthemethodseeNeumark(2012). PanelAreiteratestheestimatesshowninTable4(PanelA).Inpanel BweshowtheestimatedmarginaleffectswhenemployingNeumark’s method.Theestimatesoftheeffectofvariancewerestatistically signif-icantatthe5percentlevel,andunaffectedbywhetherwecompared tocisgendermaleandfemaleapplicants,cisgendermaleapplicants,or cisgenderfemaleapplicants.Becausethelevelestimateswerenot sig-nificant,we concludethatourestimate ofoveralldiscriminationwas notrobusttotheHeckman-Siegelmancritique.Wefoundthatthe vari-anceofunobservableswasconsistentlysmallerfortransgender appli-cants(PanelC,firstrow),whichsuggestedthattheestimatesofoverall discriminationwerebiased.Indeed,PanelCofTable5showsthatthe likelihoodratiotestrejectedtheconventionalprobitmodelinfavorof theheteroskedasticprobitmodelatthe5percentlevelofsignificance. WeconcludethatourestimatesofdiscriminationinTable4werebiased.
20 Followingareviewercomment,weperformedsimilaranalysiscategorizing
occupationssubjectivelybywhatlevelofteamworktheytendtorequiretesting forco-workerdiscrimination.Theresultsof𝜒2-testsindicatedsignificant
dis-criminationinbothtypesofoccupationswhenclassifiedbyteamworkandthe interactioninalinearprobabilitymodelwasnotsignificant(𝛽 =.032,𝑝=.440). Weconcludethatwefoundnosupportforthisexploratoryhypothesis,which couldbebecausetheoccupationswerenotaseasilydividedbyteamworkasthey werebyFFCI(FFCIwaspartofourpre-registrationandaconsiderationwhen pickingoccupations,teamworkwasnot).Moreextensiveresultsareavailable uponrequest.
21 AllresultsbasedonLPMsinthisstudyareavailablefromthecorresponding
authoruponrequest.
Therobustmarginalchangeintheprobabilityofreceivingapositive employerresponsewas3.75percentagepointslowerforatransgender applicant thanacisgenderapplicant. Thiseffectwas notstatistically significant.
3.2. Discriminationagainsttransgenderapplicantsinoccupationswith differentgenderratios
Wenowcontinueinaccordancewithourpreregistrationplanand explore any heterogeneityin discrimination across occupationswith differentgenderratios.Wedivideoursampleintothreecategoriesby occupationalgenderratio,maledominated,mixed,andfemale domi-nated(categorizationbasedondatafromStatisticsSwedenandshown in Table1).Wedidnotfindasignificant differencebetween cisgen-derandtransgenderapplicantsinpositiveemployerresponseratesin female-dominatedoccupations,𝜒2(1,𝑁=669)=.61,𝑝=.435,norin
mixed-gender occupations,𝜒2(1,𝑁=935) =2.35,𝑝=.126).The
dif-ferenceinmale-dominatedoccupationswas,however,clear. Transgen-der applicantswere12.3percentagepointslesslikely thancisgender applicantstoreceiveapositiveemployerresponseinmale-dominated occupations,𝜒2(1,𝑁=620)=10.15,𝑝=.001.22 Inotherwords,this differencesuggeststhatcisgenderapplicantswereabout41.3percent morelikelythantransgenderapplicantstoreceiveapositiveemployer responseinmale-dominatedoccupations.Theheterogeneityin discrim-inationacrossoccupationswithdifferentgenderratioswasevidenteven whenwecontrolledforotherfactors.Thisisshowngraphicallyforthe cisgender/transgender specificationusing probit modelsin Figure3. Figure3showsthatthepredictedprobability ofapositive employer responseforcisgenderandtransgenderapplicantsconvergedasthe fe-maleproportionacrossoccupationsmovedtowardsunity.
22TableA5(PanelB)intheSupplementaryMaterialshowsthese𝜒2-testsmore
Figure3.PositiveEmployerResponseRatesAcross OccupationalGenderRatio.Note.—GraphAshowsthe predictedprobabilityofapositiveemployerresponse forcisgenderandtransgenderapplicantsgiventhe dif-ferentpercentagesofwomenintheoccupation.The underlyingprobitmodelincludesthefullsetof con-trolsdetailedinTable3,includingasquaretermfor femaleworkforcebutexcludingoccupationfixed ef-fects.Graph Bshowsdifferencesinmarginaleffects betweencisgenderandtransgenderapplicantsandthe 95percentconfidenceintervalforthatdifference.Note thedifferencesiny-axisscales.Onthex-axesthe dif-ferentoccupationsinthe studyare indicated.Refer toTable1forwhattheseabbreviationsmean.Three occupationsstartwiththeletterCandtherefore war-rantmorecarefulexplanationhere,theyareChef(C), Cleaner(Cl),andChildcare(Chi).SeeSupplementary MaterialTableA3forprobitestimates.
Figure4.PositiveEmployerResponseRatesforDifferent GenderIdentities AcrossOccupationalGenderRatio.Note.—Thisgraphshowsthepredicted prob-abilityofapositiveemployerresponseforapplicantsgiventhedifferent percent-agesofwomeninanoccupation.Theunderlyingprobitmodelincludesthefull setofcontrolsdetailedinTable3,includingasquaretermforfemaleworkforce butexcludingoccupationfixedeffects.Onthex-axisthedifferentoccupations inthestudyareindicated.RefertoTable1forwhattheseabbreviationsmean. ThreeoccupationsstartwiththeletterCandthereforewarrantmorecareful explanationhere,theyareChef(C),Cleaner(Cl),andChildcare(Chi).See Sup-plementaryMaterialTableA3forprobitestimates.
Based on Figure 3 we may conclude that employers in female-dominatedoccupationsdidnotdiscriminateagainsttransgender appli-cants.Therealitywas,however,more nuanced.Whenwe examined thefour-categoryspecification,wefoundconsiderableheterogeneityin howemployerstreatedtheapplicantsacrossoccupationsdependingon whethertheyweremale-orfemale-dominated.Figure4displaysthe pre-dictedprobabilitiesofapositiveemployerresponseforallfourtypesof applicantsacrossoccupationalgenderratios.Inmale-dominated occu-pations,cisgendermaleapplicantsstoodoutasthemostpreferred. Cis-genderfemaleandtransgendermaleapplicantswerelesslikelythan cis-gendermaleapplicantstoreceiveapositiveemployerresponse. Trans-genderfemaleapplicantswereleastlikelytoreceiveapositiveemployer response.Infemale-dominatedoccupations,discriminationseemedto bemostlybasedonapplicants’gender.Femaleapplicants(FFandMF)
didbetterintermsofpositiveemployerresponsesthandidmale appli-cants(MMandFM).
We also estimated separate probit models for male-dominated, mixed, and female-dominated occupations. Panel A of Table 6 cor-responds to what was visually shown in Figure 3, and Panel B of Table6correspondstoFigure4.Table6showsstatisticallysignificant discriminationagainsttransgenderfemaleapplicantsinmale-dominated occupationsandagainstcisgenderandtransgendermaleapplicantsin female-dominatedoccupations.Notethatthetypeofapplicantwechose tosetasareferencepointintheregressionspresentedinTable6was theapplicantthathadhadthehighestprobabilityofreceivinga posi-tiveemployerresponseinagiventypeofoccupation(i.e.,MMin male-dominatedandFFinfemale-dominatedoccupations).
Next, we examined whether there were any signs of statistical andcustomer discriminationin male-andfemale-dominated occupa-tions. Mixedoccupationswerenotexaminedsinceno discrimination wasfoundintheseoccupationsintheprecedingstepof theanalysis. Wefoundnoevidenceofstatisticaldiscriminationinmale-dominated occupations as disparate treatment did not seem to diminish with increased experience. There was no statistically significant discrimi-nation among less experiencedapplicants, 𝜒2(1,𝑁=206)=1.74,𝑝= .187, while therewas discriminationamongapplicantswithmedium andhigh amountsof experience, 𝜒2(1,𝑁=210)=4.31,𝑝=.038, and 𝜒2(1,𝑁=204)=5.31,𝑝=.021, respectively. The interactionbetween
the transgender dummy and experience was not significant in an LPMspecification(𝛽 =−.007,𝑝=.634).Similarresults werefoundin female-dominatedoccupations.Theinteractionbetweenthe transgen-derdummyandlevelofexperiencewasnotsignificantinanLPM spec-ification(𝛽 =−.010,𝑝=.471).Foranintuitivegraphicalillustrationof theseresults,seeFigureA11intheSupplementaryMaterial.Insum,we foundthatthedisparatetreatmentfacedbytransgenderapplicantsdoes notdiminishasreportedworkexperienceincreased,eitherinmale-or infemale-dominatedoccupations.
Intermsofcustomerdiscrimination,resultsformale-dominated oc-cupationsweresimilartothosewefoundearlierforoverall discrimi-nation.DiscriminationwassignificantonlyinhighFFCIoccupations,
𝜒2(1,𝑁=230)=7.67,𝑝=.006,andnotinlowFFCIoccupations,𝜒2(1, 𝑁=390)=3.250,𝑝=.071.Yet,usinganLPM,wefoundnosupportfor aninteractionbetweenthetransgenderdummyandtheFFCIdummy (𝛽 =−.083,𝑝=.294).Similarly,therewasnointeractioneffectbetween thetransgenderdummyandtheFFCIdummyinfemale-dominated oc-cupations(𝛽 =.051,𝑝=.626).
Lastly,weusedNeumark’smethodtocheckiftheestimatesof dis-criminationagainsttransgenderpeople inoccupations withdifferent genderratioswererobusttotheHeckman-Siegelmancritique.Table7 showsthatthevarianceofunobservablesdidnotbiastheestimatesin male-dominatedoccupations.Infemale-dominatedoccupationswecan
Table6
ProbitEstimatesofDiscriminationAcrossOccupationswithDifferentGenderRatios (1) Male-dominated (2) Mixed (3) Female-dominated
Reference category MM MM FF
Panel A: Cisgender and transgender
Trans -0.1629 ∗∗∗ -0.0472 -0.1313 ∗∗
(0.0474) (0.0400) (0.0466)
Panel B: Four gender identities
FF -0.0726 -0.0190 REF (0.0561) (0.0469) MM REF REF -0.1905 ∗∗∗ (0.0566) MF -0.2339 ∗∗∗ -0.0318 -0.1036 (0.0575) (0.0460) (0.0530) FM -0.1051 -0.0638 -0.1614 ∗∗ (0.0537) (0.0470) (0.0543) Controls
Skills and vacancy Yes Yes Yes
FFCI and gender ratio No No No
Month FE Yes Yes Yes
Occupation FE Yes Yes Yes
Observations 620 935 669
Note.—Probitmarginaleffects(atmeans).Thistablereportstheestimatesof dis-criminationagainsttransgenderpeopleinoccupationswithdifferentgenderratios. Allmodelsusethesamecontrolspecificationasthefullmodelin(Model4)inTable4. Standarderrorsinparentheses.“FE” isshortforfixedeffectsand“FFCI” isshortfor face-to-facecustomerinteraction.VariablesareexplainedinmoredetailinTable3.
∗∗p<.01. ∗∗∗p< .001.
Table7
Neumark’sMethodforOccupationswithDifferentGenderRatios
(1) Male-dominated (2) Mixed (3) Female-dominated
Reference category MM MM FF
Panel A: Probit estimates
Trans -0.1629 ∗∗∗ -0.0472 -0.1313 ∗∗
(0.0474) (0.0400) (0.0466)
Panel B: Heteroskedastic probit estimates
Trans -0.1618 ∗∗ -0.0632 -0.1799 ∗∗ (0.0493) (0.0437) (0.0634) Trans-level -0.1651 ∗∗ -0.0055 -0.0828 (0.0550) (0.0485) (0.0652) Trans-variance 0.0032 -0.0577 ∗ -0.0971 (0.0424) (0.0277) (0.0612) Panel C: Tests
Ratio S.D of unobservables (Trans/cis) 1.0213 0.6471 ∗∗ 0.4848 ∗ Diagnostic tests (p-values)
Test S.D. ratio = 1 0.9396 0.0074 ∗∗ 0.0109 ∗
Overidentification test 0.9195 0.6615 0.6078
LR test: Probit vs. Heteroskedastic Probit 0.9389 0.0258 ∗ 0.0505
Observations 620 935 669
Note.—InpanelAandBmarginaleffects(atmeans)arereported.Thistablereportstheestimatesof dis-criminationagainsttransgenderpeopleusingNeumark’smethodofaddressingtheHeckman-Siegelman critiqueforeachgenderratiosubsample.Vacancy,occupationfixedeffects,monthfixedeffects,and skillcontrolsareincludedinallmodels.Standarderrorsinparentheses.“LR” isshortforLikelihood Ratio.
∗p< .05. ∗∗p<.01.
seeinpanelCofTable7,thereisevidenceforadifferenceinvariance be-tweenfemaleandtransgenderapplicants(firstandsecondrowofpanel Cincolumn3).However,inpanelBweseefromthedecomposition thatthevariancetermisnotsignificantandthereforedidnotaffectthe results,leadingustothesameconclusionasinmale-dominated occupa-tions.However,inmixedoccupationsthevarianceestimatewas signif-icant,indicatingthatthereeitherwasdiscriminationonthevarianceof unobservablesinthosesectorsorourCVsmayhavebeenless represen-tativeintermsofskillsinthoseoccupations.Weconcludethatwhilethe levelofourCVspotentiallyintroducedabiasattheaggregatelevel(as showninTable5)theywererepresentativeenoughwithinmale-and
female-dominatedoccupationsnottoinfluencetheestimates.Itseems thatthefailureoftheestimatetowithstandtheHeckman-Siegelman cri-tiquemaybeduetomixedoccupations,wheretherewasless discrimina-tiongenerally.TheLR-testsdonotrejectthehomoskedasticrestriction inthefemale-ormale-dominatedoccupations,sowebaseour conclu-sionsontheestimatesinpanelA.Weconcludethattransgendermale andfemaleapplicantssuffered a16.3percentagepointpenalty com-paredtocisgendermaleapplicantsinmale-dominatedoccupationsanda 13.1percentagepointpenaltycomparedtocisgenderfemaleapplicants infemale-dominatedoccupations.Thismeansthat,inmale-dominated occupations, anapplication froma cisgendermalewas 82.6percent
morelikelytoreceiveapositiveemployerresponsethanan applica-tionfromatransgendermaleorfemale.Thecorrespondingfigurefor applicationsfromacisgenderfemaleinfemale-dominatedoccupations was52.8comparedtoatransgendermaleorfemale.
3.3. Robustnesschecksandcomplementaryanalyses
IntheSupplementaryMaterialswereportsomeadditional robust-nesschecks. Wetested forbalanceofcovariatestovalidatethe ran-domizationofgenderidentityvariableandfoundthatitwassuccessful (TableA1).Weshowanexampleofanapplicationletter(ExampleA2). WereporttheprobitestimatesunderlyingFigures2–4(TableA3).In TablesA4–A5,weshowmorecompleteoutputformanyofthe𝜒2-tests
wediscussedearlier.Weperformedsurvivalanalysestoexplorewhether positiveemployerresponsetimingdifferedbetweencisgenderand trans-genderapplicantsandfoundnodifferencesinresponsetimes(TableA6 andFigureA7).Further,weverifiedtherobustnessofthe𝜒2-testsby
re-examiningthetestswithanyambiguousresponseallcodedaspositive andthenallasnegative.Wefoundthatthisdidnotaffecttheresults (Ta-bleA8PanelA)nordidsuchrecodingaffecttheprobitestimatesmuch (TableA8PanelB).Weshow estimatesofdiscriminationindifferent workexperiencesubsamples(TableA9),andproducedafiguresimilar toFigure4,butforworkexperience(FigureA10),bothofwhich(Table A9andFigureA10)corroboratethestoryofFigure2.Inotherwords, thereis nostatistical discrimination.FigureA11reaffirmsthatthere wasnosupportforstatisticaldiscriminationwithinmale-dominatedor female-dominatedoccupations.FiguresA12andA13depictthe signif-icanceofdifferencesacrossworkexperienceandgenderratio respec-tively.TableA14showshowestimateswereunaffectedbyusing differ-entspatialcontrolspecifications,andfixedeffectsforapplicantname (currentorpast).
4. DiscussionandConclusion
Thispaperhaspresentedtheresultsofthefirstcorrespondencetest ofdiscriminationagainsttransgenderpeopleinthelabor market.We havedocumentedseveralimportantfindings.Wedidnotfind signifi-cantevidenceforoveralldiscriminationagainsttransgenderapplicants whenweconsideredtheHeckman-Siegelmancritique.However,when weexaminedoccupationswithdifferentgenderratios,anotherpicture emerged.Wefoundthatbothmale-andfemale-dominatedoccupations exhibited highratesof discriminationagainsttransgenderapplicants whilemixedoccupationsdidnot.Therewasa16.3percentagepoint penaltyfortransgendermaleandfemaleapplicantscomparedto cis-gendermaleapplicantsinmale-dominatedoccupationsandtherewas a13.1percentagepointpenaltyfortransgendermaleandfemale ap-plicantscomparedtocisgenderfemaleapplicantsinfemale-dominated occupationsasregardspositiveemployerresponserate.Theseestimates wererobusttotheHeckman-Siegelmancritique.
Inmale-dominatedoccupations,Neumark’smethodshowedusthat employers act as ifthere was no difference in the variance of un-observablesbetween cisgender menand transgenderapplicants (i.e.
𝜎T/𝜎MM≈1),wealsofindnosupportforstatisticalmotivations leav-ingonlytaste-basedones.Thisfindingisinlinewithsomesurvey evi-dencethathaveshownthatmenmaybemoreprejudicedthanwomen againsttransgenderpeople(Nagoshietal.,2008).Employersin female-dominatedoccupationsinsteadactedasiftherewasadifferenceinthe varianceof unobservables(i.e.𝜎T/𝜎FF ≈.49),butevidentlyour CVs wererepresentativeenoughthatthisdidnotaffectourestimatesof dis-criminationinthissubsample.Thismeansthatwefoundnoevidence thattransgenderapplicantswereseenasparticularlyrisky(ineither male-orfemale-dominatedoccupations),astheywouldhavethenbeen treatedasiftheyhadahighervarianceofunobservablesthan cisgen-derapplicants.Insteaditseemedasifemployersinfemale-dominated occupationsweremorecertainoftheunderlyingproductivityof trans-genderapplicantsthantheywereofthatoffemaleapplicants,butstill
preferredtohirefemaleapplicants.Thissuggeststhatemployersmay havehadstereotypedperceptionsabouttheunobservedproductivityof transgenderapplicants.
The heterogeneouseffectsfordifferent transgender(MF andFM) andcisgender(FFandMM)applicantsacrossoccupationswith differ-entgender ratiosprovided somesuggestiveevidencethat employers infemale-dominatedoccupationsseemedlessacceptingofmale appli-cants,whereasemployersinmale-dominatedoccupationsseemedless acceptingoftransgenderapplicants.Infemale-dominatedoccupations, employersweremorelikelytorespondpositivelytocisgenderand trans-genderfemaleapplicantsthantocisgenderandtransgendermale appli-cants.Inmale-dominatedoccupationsemployersseemedtohave posi-tivelyrespondedtoapplicantsinfollowingdescendingorder:cisgender maleapplicants,cisgenderfemaleapplicants,transgendermale appli-cants,andlastlytransgenderfemaleapplicants.
Wealsofoundthatdiscriminationdidnotdiminishwiththeamount ofexperiencereportedintheCVs,eitherinthefullsampleorinany sub-sample,whichsuggeststhatdiscriminationwaslesslikelytobedriven bystatisticalmotivations.Oneimportantnoteonstatistical discrimina-tionisthatstatisticalmotivescouldstillplayaroleinindividualCV evaluationseventhoughwedidnotdetectthemattheaggregatelevel, becausestatisticalmotivesarenotnecessarilyunidirectional.For exam-ple,VanBormandBaert(2018)andVanBormetal.,2020use hypothet-icalhiringdecisionsofhumanresourcesstudentstoexamine discrim-inatorymotivationsagainsttransgenderapplicants.Theyfoundmixed evidenceforstatisticaldiscrimination,withbothpositiveandnegative effectsbasedonthreestatisticalmotivations:(i)thelikelihoodofsick leave(negativeforbothtransgendermalesandfemales),(ii)the prob-abilityof maternityleave(heterogeneousbetweentransgendermales andfemales),and(iii)perceptionsaboutassertivenessandautonomy (positiveforbothtransgendermalesandfemales).Ourresultthat dis-paratetreatmentwaslikelynotduetostatisticaldiscriminationcould beinlinewiththeseresultsbecausethepositiveandnegativeeffects fromdifferentmotivationsmaycanceleachotheroutattheaggregate level.
Thecontextof ourstudymustbe discussedat thispoint,bothin termsoflocaleandtiming.Thereareseveralreasonswhyestimates re-portedininthispapermightreflectalowerboundofhiring discrimina-tionfacedbytransgenderpeople.First,weconductedthisexperimentin Sweden,aparticularlyprogressiveandtolerantcountrywhenitcomes togenderequality(WorldEconomicForum,2018)andLGBTacceptance (FloresandPark,2018).Second,theexperimentwasconductedwhen theSwedish economywasdoing well,which meansthatthemarket forlow-skilljobsshouldberelativelytight.Reportsofdiscrimination tendtobelowerundersuchconditions(BoulwareandKuttner,2019) whichmayhaveloweredourestimatesofdiscrimination.Third,even thoughwefocusexplicitlyonprivatesectorjobs,somesectorsclearly interactmorewiththepublicsector;thisismostlytrueforthe female-dominatedoccupationsin oursample(e.g.,enrollednurseand child-care)andmayhaveaffectedtheratesofdiscriminationseeninthose sectors(Ahmedetal.,2013b).Fourth,ourexperimentaldesignclearly signalstransgenderstatusandseparatesitfrom namechanges. How-ever,wemayhaveunderstateddiscriminationifthecisgender appli-cantshavereceivedfewerpositiveemployerresponsesbecauseoftheir namechangecomparedtoapplicantswhohavenotgonethroughaname change.Wefindthisunlikelybecausetheresponserateswerehigh(40 percent)andchangingone’sgivennameisfairlycommoninSweden.
Itisalsoimportanttonotethatourmethodcapturesdiscrimination onlyintheinitialstepofthehiringprocessandusesonlyofficialjob searchchannels.Obtainingajobthroughsocialcontactsandnonofficial channelsiscommoninSweden,andtheuseofsocialchannelsismore commonamongsmallfirms(Eliasonetal.,2017).Ourstudy, unfortu-nately,didnotcapturethisinformalsideofthemarket.Evenifthereare fewreasonstobelievethatthesechannelsshouldexhibitless discrimi-nationthantraditionalchannels,moreresearchisclearlyneededhere. Itisalsolikelythatotheravenuesfordiscriminationsuchasworkplace
