ChallengesofChange:AnExperimentPromotingWomentoManagerialRolesintheBangladeshiGarmentSector

September25th2018

RoccoMacchiavello,LondonSchoolofEconomics

AndreasMenzel,Cerge,Prague

AtonuRabbani,UniversityofDhaka

ChristopherWoodruff,UniversityofOxford#

Abstract:Women underrepresentation in managerial positions, particularly in developingcountries, could lead to misallocation of managerial talent. This paper designs andimplements an experiment to test for suchmisallocation in the Bangladeshi garmentsector,acontextinwhichwomenrepresent90%ofthelabourforcebutlessthan10%ofthelinesupervisors(thelowestmanageriallevel).Toovercomeempiricalchallengestotest for misallocation, the experiment i) identifies the marginal male and femalecandidatesupervisor,ii)allocatesthecandidatesupervisorstorandomproductionlinesforatrialperiod,iii)collectdetaileddataonlineperformanceaswellassubordinatesand co-workers attitudes and perceptions. Results reveal that women initiallyunderperformrelativetomalecandidatesbutthenquicklycatch-up.Wefindevidenceofresistance from male operators against women candidates and initially wrongperceptionsaboutwomen’sabilityasmanagers.Eventuallyfactoriesappoint53%(66%)of female (male) candidates to supervisory role. The results suggest that womenrepresent a largely untapped pool of managerial talent in these factories but thatimplementingchangealsoentailcosts.NB:Preliminary&WIP.Wearecurrentlyre-driftingthepaper.Slidesattachedattheendincludethenewmaterial.

#Corresponding author: christopher.woodruff@qeh.oc.ac.uk. The project has benefitted fromcommentsfromseminarsandconferencesatUCSanDiego,theUniversityofWashington,NotreDame,Duke,Leuven,EcolePolytechnique,MIT/Harvard,PUC-Chile,theCEPR-IMOandCEPR-IZAworkshops,AEA-ASSA2015andLSE.Remainingfailingsaretheresponsibilityoftheauthors.We are grateful for the cooperation and financial support of Deutsche Gesellschaft fuerInternationaleZusammenarbeit(GIZ),whodevelopedthetrainingprogramthatweimplementin the project.We are also grateful for financial and logistical support from the InternationalGrowth Centre, and financial support from the IPA SME initiative, the ERSC – DFID GrowthResearch Programme and IFC-Bangladesh, and for the cooperation of the large number ofparticipatingworkersandfactoriesinBangladesh.

1

1. INTRODUCTION

Management of large firms in low-income countries is highly variable and, onaverage, poor (Bloom et al. [2012]). The recent literature has focused on theimplementationofabroadsetofmanagementpracticespioneeredbyBloomandReenen[2007].However,effectivemanagement,includingtheadoptionofsuchpractices,restsonsuccessfullymanagingrelationshipsandperceptionsintheworkplace(GibbonsandHenderson [2012]).Thisobservation shifts our attention frompractices tomanagers.Shortages of qualified managers are perceived to be an important barrier to bettermanagement in developing countries (e.g.,McKinsey [2011]); yet,we still know littleabouthowcompaniesinthesecountriesdevelopandselectmanagerialtalent.

We study low-level management in the ready-made garment industry inBangladesh,asectorwithmorethan4,000factories,employingaround4millionworkersandaccounting foranestimated12percentofBangladesh'sGDP.Besides its intrinsicrelevance,thesectorprovidesanidealcontexttostudylow-levelmanagers.Thesewingsection in a typical factory is organized along several production lines employingbetween20and80workers(operators)directlymanagedbylinesupervisors,thelowestlevelofmanagement.Wefocusononedistinctivefeatureoftheindustry:whilewomenaccountforabout75to80percentofworkersinthesewingoperations,menaccountforaround 95 percent of supervisors and higher-level managers. The situation is stark:Figure1contrastsemploymentpatternsinBangladeshwiththehistoricalevolutionintheUnitedStatesandshowsjusthowstrongthegenderimbalanceisinBangladesh.

Whyaretheresofewfemalesupervisors?Doesthisgenderimbalanceresultinameaningfulmisallocationofmanagerialtalentinthesector?Toaddressthesequestions,westartfromasimpleobservation:inastaticsense,managerialcapitalismisallocatedifthemarginalfemalesupervisorismoreeffectivethanthemarginalmalesupervisor.1Ifthiswasthecase,factoriescouldimproveefficiencybypromotingadditionalwomenandfewermen.2

1Notethattheobservationiscorrectforanydistributionofpotentialsupervisor'seffectivenessacrossgenders.Inparticular,itispossiblethatinthecurrentindustryequilibriummenself-selectand/orinvestinadditionalskillswiththeexpectationofbecomingsupervisors.Thiscouldresultin thepoolofmenavailable forpromotionbeingonaveragebetter than thepoolofavailablewomenforpromotion.2Large inefficiencieswouldbeatoddswith the fact that all factories inour sampleare largeexporters operating in highly competitive product markets. A large literature shows thatcompetitionincreasesefficiency(Syverson[2004];Fosteretal.[2008];Backus[2014]),improves

2

Empirically,we face twomajor challenges in answering these questions. First,given there are so few female supervisors to beginwith, it is difficult to identify themarginal female supervisor in observational data. Second, we need to observe theperformanceofthesemarginalmalesandfemalesworkingaslinesupervisors.

To overcome these issues, we implement a six-week operator-to-supervisortrainingprogramwithmore than500workers in80 factories.3Theprogram inducesfactoriestotryout(andpossiblypromote)morefemalesupervisorsthantheyotherwisewould. The design of the project evolved both because we learned from theimplementation in the first factories with which we worked, and because factoriesbecame more flexible in how we implemented the program as the initial evidencesuggestedthatfemalesupervisorsweremoreeffectivethanmanagershadexpected.Wefocusheremainlyonresultsfrom150traineesinthefinal24factories,whereboththedesignandthedataarethecleanest.

We implementanexperimentaldesign inwhichreturning traineesare triedas

assistantsupervisorsonrandomlyassignedproductionlines.Thisallowsustoidentifythecausalimpactoffemalesupervisorsonperformance.Wecomparetheperformanceoffemalesandmalestrainedintheprogram,andtheresponseofoperatorsworkingforthem,usingbothverydetailedproductiondataandin-factorysurveys.

Weshowfoursetsofresults.First,weaskwhatsupervisors(aresupposedto)do,andwhattheperceivedweaknessesoffemalesassupervisorsare.Wefindremarkableagreementacrosshierarchical layers in the factories about the relative importanceofeachofeightsetsofbroadtasksperformedbysupervisors.Thereisalsoalmostuniversalagreement inthe factorythatwomenareweakerthanmeninalleightdimensions. Inparticular, women are perceived to be less competent than men in understandingmachinesandoperations-crucially,themostimportanttaskforasupervisorfromthepoint of view of operators. These negative perceptions are weaker, but nevertheless

managementpractices(BloomandReenen[2007];Bloometal.[2012])andthatexportstatusisassociatedwithhigherproductivity(Bernardetal.[2007])andbettermanagement(Bloometal.[2015]).Ontheotherhandthefactoriesinoursettingaretypicallyownedbyasmallgroupofinvestorsandmightfacelowerpressureonthefinancialmarketside.3ThetrainingprogramwasdesignedbytheGermanbilateralaidagency,DeutscheGesellschaftfuerInternationaleZusammenarbeit(GIZ),togetherwithlocaltrainingcompanies.

3

present, even among female operators and among those operators with experienceworkingunderafemalesupervisor.

Second,wecomparetheseperceptionstoreality.Beforethetrainingbegan,weconductedanextensiveskillsassessmentwiththetrainees.Threeresultsemerge.First,there is no difference between female and male trainees in technical knowledge ofmachinesandoperations -despite thewidelyheldopinionto thecontrary.Second, insimple leadership exerciseswomen are less likely to be selected by their team for aleadership position and women perform slightly worse in an exercise in which theyinstruct other teammembers to perform a simple task. Third, in essentially all eightbroad tasks, females rate themselvesasbeingweaker thanexistingsupervisorswhilemaletraineesdonot.

Third,weexaminetheperformanceoffemaleandmaletraineesoncetheyreturnfrom the training. Immediately upon returning from training female traineesunderperformrelativetomaletrainees.Thisinitialgapinperformanceisobservedbothin surveys of operators supervised by the trainees and in detailed daily line-levelproductiondata.Thegapinperformance,however,completelyclosesafterfewmonthsworkingonthelineassupervisors.Insimulatedmanagementexercises,femaletraineesoutperformmaletraineesonaveragebutnotwhenmanagingsmallteamsthatincludeamaleoperator.

Finally,weexploreattitudesofmaleoperatorsexposedtotheprogram.Theseareof particular importance given that the bulk of future line supervisors is currentlyrecruited from among male operators. Two results stand out: first, male operatorsexposedtofemaletraineesimprovetheirviewoffemalesassupervisors.Second,maleoperatorsexposedtofemaletraineesaremorepessimisticabouttheirownprospectsofbeinglaterpromotedtosupervisorrolesandexpecttoworkforashorterperiodoftimeinthefactory.Inshort,thepromotionoffemalesupervisorappearstodemotivatemaleworkers.

Takenalltogether,theseresultsportrayanuancedbutcomprehensivepictureofthe causes and consequences of gender imbalance in the sector. The evidence isconsistentwithanindustryequilibriuminwhichfactorieshavenotexperimentedwithfemale supervisors due to misperceptions about their relative effectiveness. Theequilibrium is supported by the fact that misperceptions are widespread across theorganization, including amongworkers and potential female supervisors themselves.

4

Shiftingtoanewequilibriumrequirescoordinatedchangesinbeliefs.Inastaticsense,evenaprofitmaximizingmanagerwithcorrectbeliefsmightnotpromotewomenif-inourcase-hebelievesotherco-workerswon'trespondadequatelyduetotheirbeliefs.Dynamically, such a manager might believe workers' perceptions can be aligned toreality,butatthecostofalienatinganddemotivatingmaleoperators-fromwhichthebulkofmanagerialtalentisstilllikelytobesuppliedtothefactoryintheshort-run.Intheconclusions,wedistilsomeimplicationsofthisinterpretationforourunderstandingoforganization's failure to adopt adequate management practices, the sources andconsequences of gender imbalances in general, and the design of policies that couldamelioratethose.

This paper contributes to different strands of literature. It complements aliterature examining the causes and consequences of the (lack of) female leadership.Althoughtherearenumerouscontributionsstudyingthegendergapinlabourmarketsand in the private sector (see, e.g., Bertrand et al. [2014]; Matsa and Miller [2013];BertrandandHallock[2000];DezsoandRoss[2012];Gloveretal.[2015]),ourworkisconceptuallyclosertostudiesoffemalepoliticiansinIndiabyChattopadhyayandDuflo[2004]andBeamanetal.[2009].4

AsChattopadhyayandDuflo[2004]wefocusonestablishingthecausalimpactoffemaleleadershipsonoutcomes.AsBeamanetal.[2009]weemphasizetheimportanceandevolutionofperceptionsof female leadership.Ouranalysis,however,needs tobeadaptedtoreflecttheoperationsandincentivesoflargefirmsoperatinginacompetitiveexport sector.First, theperformance -not just theappointment -of female leaders isaffected by beliefs and perceptions of co-workers. Second, we investigate the costsassociatedwithappointingfemaleleaders.

In so doing, the paper also contributes to the literature on management andproductivity(see,e.g.,BloomandReenen[2007];Bloometal.[2012,2013];Bruhnetal.[2012];McKenzieandWoodruff [2015]).5TheworkbyBloomandvariousco-authors

4 Some of our results are also related to a large experimental literature documenting genderdifferences in attitudes andpreferences, see, e.g.,GneezyandRustichini [2004];Niederle andVesterlund[2007];Niederle,Segal,andVesterlund[2013].5 There are two additional methodological contributions of the paper. With respect to theproductivity literature, the paper uses a physicalmeasure of productivity in amulti-productindustrywithproductdifferentiation.Line-levelproductivity ismeasured takingadvantageof“standardminute values" which allow to convert units of differentiated garment pieces intostandardizedmeasuresoftimevalueofoutput.Withrespecttotheliteratureontheevaluationof

5

raises a puzzle: themanagement practices they study arewell-known and seeminglysimple to implement.Why do firms fail to implement them?Gibbons andHenderson[2012]arguethatchangingpracticesisactuallyquitecomplex,bothbecauseindividualpractices are complementary to one another (see also Ichniowski et al. [1997]) andbecausemanagement involves both formal rules and informal norms.Managersmayknowwhatiswrong,knowhowtofixwhatiswrong,butyetbeunabletoimplementtherequiredchangesbecausetheyareunabletoshifttheequilibriumofthegamebetweenmanagersandworkers(orbetweenmanagersatdifferentlevelsofthehierarchy).Ourresearchdesignandemphasisonunderstandingmisalignmentofperceptionswithinthefirmborrowsfromthisperspective.ThedifficultiesofimplementingchangeechorecentworkbyAtkinetal.[2015]inthesoccerballindustryinPakistan.Theyshowthatfirmsmay fail to adopt productivity-increasing changes in technology because the paystructure of production workers encourages them to misreport to management theproductivity of the technology. We instead highlight how resistance to change isembeddedinasetofnormsandperceptionswesetouttomeasure.

Finally, the paper contributes to our understanding of the garment sector inBangladesh and elsewhere. Historically, the sector has represented one of the firstopportunities forwomentoenter the formal labour force.HeathandMobarak[2015]study the relationship between garments, female labour force participation andschoolinginBangladeshrespectively.LinesupervisorsinthegarmentindustryarealsostudiedbySchoar[2011]andAchyutaetal.[2014]withadifferentfocusandresearchdesign.

The pool of female andmale trainees for the program are selected by factorymanagement.Theinitiallackoffemalesupervisorsmayalsoposeachallengetofactorymanagement because the managers have little experience selecting females forpromotion. We examine the selection environment using uniquely detailed baselinesurveysanddiagnosticstoolsimplementedwithworkersandmanagersatall levelsinthe factories. Data from these exercises help us understand what supervisors areexpectedtodo,andhow–inbothperceptionandinreality–theskillsoffemalescomparewiththeskillsofmales.

trainingprogram,wedirectlyinvestigatetheimpactofthetrainingonproductivity,notjustonthewagespaidtotrainees.Thisisimportantasforavarietyofreasonswagesmightfailtoreflectthemarginalvalueoflabour.

6

{AdditionalPapers}

2. INDUSTRYBACKGROUNDWestudyproductionlinesupervisorsintheBangladeshigarmentindustry.Bangladeshisnowthesecondlargestexporterofgarments,afterChina.Theindustrystarteddevelopinginthe1980sandnowcountsmorethan4,000factories,employsaround4millionworkersandaccountsforabout80percentofBangladeshexportsandanestimated12percentofBangladesh'sGDP.

Besidesitsintrinsicrelevance,thesectorprovidesanidealcontexttostudyhowlow-levelmanagersimpactproductivity.Thesewingsectioninatypicalfactoryisorganizedalongseveralproductionlinesemployingbetween20and80workers(operators)directlymanagedbylinesupervisors,thelowestlevelofmanagement.Garmentfactoriesareverylargerelativetodevelopingcountriesstandards.Thefactoriesinthisstudyhaveanaverageofaround1000workersemployedinthesewingsectionalone.

[RM:doweaddhereaparagraphhereonmeasuringproductivity?]Shortagesofqualifiedmanagersareperceivedtobeanimportantbarriertobetter

management in developing countries (e.g.,McKinsey [2011]), yet,we still know littleabouthowcompaniesinthesecountriesdevelopandselectmanagerialtalent.Industryobserversperceivelinesupervisorstobecriticalforbothproductivityandquality(see,e.g.,[addreferences])inthisindustryaswell.Accordingtodatawehavecollectedinaseriesofdifferentprojects,approximatelyhalfoftheexistingsupervisorsareinternallypromoted from within the factory with the rest recruited from other factories. Linesupervisorstypicallystartworkinginthefactoryaseitherhelpersonthelineorinthequalitydepartment.

Inthispaperwefocusononedistinctivefeatureoftheindustry:whilewomenaccountforabout75to80percentofworkersinthesewingoperations,menaccountforaround95percentofsupervisorsandhigher-levelmanagers.Thesituationisstark:Figure1contrastsemploymentpatternsinBangladeshwiththehistoricalevolutionintheUnitedStatesandshowsjusthowstrongthegenderimbalanceisinBangladesh.Atthebeginningofthetwentiethcenturyapproximately30%ofmanagersinthetextile

7

andapparelsectorintheU.S.werewomen,asignificantlylargersharethanintoday’sBangladesh.Industryobserversperceivethisimbalancetobeapotentialobstacletothefurtherexpansionandupgradingoftheindustry.First,astheeconomydevelops,thesectorappearstobelessandlessabletoattractthemostqualifiedmen,whichnowhaveotherjobopportunities.Consistentlywiththishypothesis,thelevelofeducationhasdeclinedovertimeacrosssuccessivecohortsofsupervisors.Second,thegenderimbalancemightalsoberesponsibleforrelativelypoorindustrialrelations,socialcomplianceandpossiblyproductivity(addreferences),furtherhamperingtheindustryupgradingprocess.

[Addhereaparagraphonthecross-sectionalcorrelationbtwefficiencyandfemalesupervisors]+[anoteonthegenderwagegap]

3. CONCEPTUALFRAMEWORK

Whyaretheresofewfemalesupervisors?Doesthisgenderimbalanceresultinalargemisallocationofmanagerialtalentinthesector?Toanswerthesequestionswelayoutasimpleconceptualframeworkthatclarifiestheempiricalchallengesweface,howourempiricalstrategyaddressesthem,aswellasanumberofinterpretationissues.

DefiningMisallocation

Considerafactoryemployingbothmenandwomen.Eachworkerisendowedwithapotentiallevelofmanagerialtalent,!.Potentialmanagerialtalentisdistributedamongmaleworkersaccordingtoacumulativefunction"(!).Managerialtalentisdistributedamongfemaleworkersaccordingtoacumulativefunction%(!).Thetwodistributionscanpotentiallybedifferent,reflectingactualdifferencesininitialendowmentsofmanagerialtalentbetweenmenandwomen,or–perhapsmorelikely-simplyduetodifferentialselectionateithertheindustry(orevenfactory)level.Forinstance,itispossiblethatinthecurrentindustryequilibriuminwhichmenexpecttobepromotedtolinesupervisors(butwomendonot)onlymenwithrelativelyhigh!eitherself-selectintotheindustryorinvestinskillstoacquirealarger!.

Wedefinemanagerialtalenttobesuchthat,allelseequal,factoryefficiencyandprofitability(inshort,performance)increasebyappointinglinesupervisorswithhigher!.Performance,however,alsodependsonotherfactors,includingthesupervisorowneffort(&),peersandsubordinateactions('),managersactions((),andotherfactors

8

()).Definingactionsandfactorsappropriately,performanceofthelineisafunction* =(!, &, ',(, ))increasinginallitsarguments.Asimpleobservationfollows:allelseequal,managerialtalentismisallocatedifthemarginalfemalesupervisorismoreeffectivethanthemarginalmalesupervisor.Ifthiswasthecase,factoriescouldimproveperformancebyreplacingtheleastcapablemansupervisorwiththebestavailablewoman.Weareinterestedintestingwhetherthereislargemisallocationofmanagerialtalentaccordingtothisdefinition.

EmpiricalChallengesandStrategy

Giventhedefinitionabove,testingformisallocationofmanagerialtalentpresentsthreesetsofchallenges.First,weneedtoidentifythemarginalmaleandfemalesupervisor.Identifyingmarginalinputs(workers,capital,etc.)isdifficultingeneral.Inourset-upthetaskisfurthercomplicatedbythefactthat,inmanyfactories,thereisnofemalesupervisortobeginwith.Second,weneedtomeasureperformanceontheline* = (!, &, ',(, )).Finally,weneedtotakeintoaccountthatperformanceisn’tsolelyafunctionofsupervisor’stalent,!.Duetothenatureoftheproductionprocess,performancecruciallydependsontheactionofmanyothersintheorganization.Inturn,theactionsofallthoseinvolved–fromlineoperators,uptothemanagersandincludingthelinesupervisorhim/herself–dependontheirpreferencesandbeliefs.Thisisanimportantdistinction(andcomplication)relativetoexistingpapersintheliteratureandonethatweseeabsolutelycentraltothecontributionofthepaper.Forinstance,inboth(Pallais)studyofdiscriminationamongcashierataFrenchsupermarketor(Atkinetal)studyonresistancetotechnologicalchangeamongcuttersinthesoccer-ballsectorinPakistanworkersperformessentiallyindividualtasks. Wetacklethesechallengesasfollows.First,we“engineer”asetofmarginalmaleandfemalesupervisorsbyofferingfactoriesthepossibilityofselectingandthentrainingequalnumbersofmaleandfemaleoperatorstobecomesupervisors.Subjecttocaveatwediscussbelow,thisallowsustoobtainfromthefactoriesthe“nextinline”forpromotiontotheroleoflinesupervisorbygender.Whencandidatesreturnfromthetraining,weaskfactoriestorandomlyallocatethemtoworkaslinesupervisors.Therandomallocationinducedbytheexperiment,allowsustocontrolforfactors)thatmightaffectperformance.

9

Second,wemeasureperformancethroughacombinationofdifferenttypesofdata.First,wemeasureevaluationoftheperformanceofthecandidatelinesupervisorsamongsubordinatesandco-workers.Second,wecollectextremelydetaileddailyproduction-lineleveldataonefficiencyandotheroutcomes.Third,wemeasureactualperformanceandmanagementstyleinsimulatedmanagementexercises.

Thethirdchallengeismoredifficulttoaddressbutliesattheheartofunderstandingwhatorganizationalchangeentails.Observingperformanceontherandomlyassignedlineisn’tsufficienttoinferdifferencesinpotentialtalent!.Workersactions–whicharedrivenbybothbeliefsandpreferencesandmightthereforebegenderspecific–donotallowustoinvertobservedperformance*toinfer!.Thiscallsforadesigninwhichinitialbeliefsandattitudes,aswellastheirevolution,areadequatelymeasuredbeforetheinterventionandcontrolledforwhenstudyingperformance.Ofcourse,totheextentthatbeliefsandattitudesarethemselvesinfluencedbyexposuretofemalesupervisors,theybecomeoutcomesofinterestinthemselves.Throughouttheproject,wedistinguishandmeasurethreesetsofbeliefs:self-confidenceoftraineethemselves(!/ and!0forfemaleandmalerespectively),thoseofsubordinatesandpeers(!1),andthoseofmanagers(!0).HypothesisTestedandFurtherInterpretationChallengesWeareinterestedinunderstandingtheextenttowhichapossiblywrongsetofbeliefswithintheorganizationprecludestalentedwomenfrombecominglinesupervisors.Withthisgoalinmind,theframeworksuggeststhefollowinghypotheses:H1:oneormorelayerswithintheorganizationhaveexcessivelypessimisticbeliefsaboutfemalerelativemanagerialtalent;H2:whentestedaslinesupervisorsonrandomlyallocatedlineswomenperformrelativelybetterthanexpected(butnotnecessarilybetterthanmen);H3:performanceisdifferentiallyaffectedbyorganizationalfactors,includingbeliefs/attitudesfromothers,acrossgenderofthesupervisor.

Evidenceinfavourofthethreehypothesesabovewouldprovidestrongsupporttotheviewthatorganizationalfactors–ratherthanalackofavailablemanagerialtalentamongpotentialsupervisors–liebehindthescarcityoffemalesupervisorsinthe

10

industry.Beforeturningtotheempiricaldesignandresults,wediscussanumberofchallengestotheinterpretationofthehypothesis.

First,thetrainingpotentiallyconfoundstheinterpretationoftheresults.Suppose,inparticular,thatthetrainingincreasesmanagerialtalentby2/forfemaletraineesand20formaletraineesandthat2/ > 20.Ifthisisthecase,then,webiasthetest:intheabsenceoftraining,womenwouldperformasexpectedand,therefore,definitelyworsethanmen.Webelieveinpractice,thiswasn’tthecase.Wewillpresentevidencethatthetraininghasincreasedfemaleself-confidence(!/)alotmorethanmaleself-confidence(!0)butdoesn’tappeartoendowedfemalewithrelativelymorenewskills.6

Second,itispossiblethateitherdoesn’twantto,ordoesn’tknowhow,toselectthemarginalcandidateforthetrainingprogram.Althoughinprinciplethiscouldbiasourtesteitherway,wepresentevidencethat–ifanything–itbiasthetestagainstfemaletrainees.Thefactorymightgivemalecandidatessignificantlyworsethanthemarginalones–e.g.,ifthefactorybelievesthatthetrainingisnoteffectiveandsparingthemarginalmalecandidatehashigheropportunitycosts.Inpracticethisisnotaconcerngiventhatalmostallmaletraineesareeventuallypromotedbythefactories.Ifanything,theyareinfra-marginalsupervisors.

Moreplausibly,themanagementofthefactorydoesn’twantordoesn’tknowhowtoselectthefemalemarginaltrainees.Leavingasidethepossibilityoftaste-baseddiscrimination,itispossiblethatinthecurrentindustryequilibriummanagersarebetterableatpickingtherightmarginalmaleratherthanfemale.First,theyhavemoreexperienceatit.Second,differentialselectionintotheindustryandthesheerdifferenceinthesizeoftheavailablepoolmakefindingthebestcandidateamongmalesmucheasier.7Wewillprovidedirectevidencethat,indeed,managersaresubstantiallylessconfidentabouttheirselectiondecisionsforfemalethanformalecandidates.

Finally,thethreehypothesisdonotdistinguishwhichorganizationalfactorsare“necessary”or“sufficient”fortheinefficiencytoarise.Forinstance,arewrongbeliefsatalllevelsofthehierarchynecessarytogeneratethemisallocationofmanagerialtalent?

6Infact,althoughevaluatingthetrainingisn’tthepurposeofthisstudy,wewillpresentevidencethatbesides

self-confidencetrainingperse(i.e.,beforeactualworkonthelines)hasnotimprovedanyskillforbothfemale

andmaletrainees.7Thebestavailablewomenmightalsofaceadditionalconstraintsfromthefamilywhichwouldpreventher–if

selected–toparticipateintheprogramandeventuallybepromoted.

11

Theanswerdepends,interalia,onhigherorderbeliefswhicharedifficulttomeasure.Forinstance,managerscouldhavethecorrectbeliefsaboutmarginalfemalerelativeabilitiesbutcouldbeawarethatworkershaveincorrectbeliefsornegativeattitudestowardsfemalesupervisors.Themanagermightthennotfinditprofitabletoinitiatethechange.Conceptually,however,anequilibriumentirelysupportedbyworkerswrongbeliefsappearsimplausible.Byinitiatingchangeamanagerwouldsignaltoworkersherbeliefsthatfemalecanbegoodsupervisors.Forworkersnottorespondadequately,theymusteitherdislikefemalesupervisorsforotherreasons(e.g.,worsefutureprospects,orsimplypreferences)ortheymustnotbelievethemanager,e.g.,becauseofaconflictofinterest(i.e.,themanagerwouldlikemorewomenthantheworkersdo).

Wenowdescribetheexperimentalinterventionandthedataandthenteststhethreehypothesissuggestedbytheframework.

4. DESIGNANDDATA

AtthecoreofourstudyisatrainingprogramdesignedbytheGermanbilateralaidagency(GIZ)and local trainingcompanieswhichaims toprovidesewingmachineoperators skills necessary to be sewing line supervisors. GIZ's goal in developing theprogramwastoincreasethenumberofwomenworkingassupervisorsinthesector.Thetrainingwasviewedas important tobuild skills of femaleoperators, and to convincefactoriesthatwomenwereequippedtobesupervisors.Thetraininglastssixweeks,witheight-hoursessionsheldattheclassroomsatthetrainingprovider’sofficesonsixdaysperweek.Thecurriculumwasdividedmoreorlessequallyintomodulesonproductionplanningandtechnicalknowledge,qualitycontrol,andleadershipandsocialcompliance.Weinitiallycontractedwiththreetrainingprovidersandthenlaterselectedoneofthemwiththecapacitytoconductallofthesessions.

Ourprojectwascarriedoutintwophases.Phase1beganinNovember2011andcontinued throughFebruary2013,with56 factories sending fiveparticipants each totraining.Afteranalysingthedatafromthefirstphase,wemadeseveralchangestotheprojectdesignandlaunchedthesecondphaseinFebruary2014.Lessonsfromthefirstphasewereincorporatedintothedesignofthesecondphase.Asaresultofincorporatingtheselessons,thequalityofthedataaregenerallyhigherinthesecondphase.AsidefromamanagementsimulationexercisethatweconductedonlyinPhase1,werelyonthedatafromthesecondphaseforthispaper.Wedescribethedesignforthesecondphasehere,

12

andreferthereadertoAppendixAforadescriptionofthedesignofthefirstphase,andacomparisonofresultswheretheyoverlap.

InthesecondphaseoftheprojectweworkedwithdirectandindirectsuppliersofalargeUK-basedbuyer.Westartedwithapoolof26suppliersofwovenandlight-knitproductslocatedintheDhakaarea.8ThebuyerinvitedthesesupplierstoaninformationsessioninFebruary2014.Attheendoftheinformationsession,24factoriesexpressedinterestintheproject,allofwhomultimatelyparticipated.9

Weaskedeachfactorytoconsidertheexpecteddemandfornewsupervisorsinthefactoryinthemonthsfollowingtraining,andselectanumberoftraineesmatchingthatdemand.Because the sizeof the factoriesvariedandbecause, for example, somefactorieswereplanningtoopennewproductionlines,thenumberoftraineesvariedfromasfewasfourtoasmanyas24.Whereanevennumberoftraineeswasprovided,weasked factories toselectanequalnumberofmaleand female trainees.Whereanoddnumberoftraineeswasselected,weaskedthemtoselectonemorefemalethanmale.

We informed the factories thatmuchof the trainingmaterialwaswritten, andthereforethetraineesneededtohaveatleastbasicliteracyskills.Wegavethemnoothercriteria,butdidencouragethemtoinvolveintheselectiondecisionsmanagersdowntoat least the level of the line chief – the immediate superior of line supervisors. Thefactories sent 99males and 100 female trainees to the training centre for the initialdiagnostic.Notethatthisrepresentsasignificantmovementtowardfemalesupervisors,because in the typical factory at baseline only around 4 percent of supervisorswerefemale.

Wescheduledfourtrainingsessions,thefirstbeginningMarch9th,2014,andthelastbeginningJune1st,2014.Inordertostaggerthereturnoftraineestothefactory,halfthenomineesfromeachfactorywererandomlyallocatedtooneoftwotrainingrounds,eitherrounds1and3orrounds2and4.Withinthefactory,thetraineeswererandomlyassignedtoreceiveearlyorlatetraining.Randomizationatthetraineelevelwasstratified

8We limited the sample to theDhaka area for logistical reasons and towoven and light knitbecauseproductionintheseproductsisorganizedbysewinglinesinBangladeshifactories.Directsuppliers are managed by employees working directly for the buyer; indirect suppliers aremanagedonbehalfofthebuyerbyintermediaries.9Fiveofthefactoriessentoperatorstothefirsttrainingsession,butdroppedoutinthesecondhalfoftheprogram.

13

ongendersothatanearlyequalnumberoffemaleandmaletraineesweretrainedineachsession.

Factoriesagreedtogiveeachtraineeasix-toeight-weektrialasanassistantlinesupervisor immediately after the end of the training program.We asked factories toidentifythelineswhichweresuitableforthetraineetrialsandtoidentifyanexperiencedsupervisorworkingoneachofthoselineswhocouldactasamentorforthetrainee.Onthepenultimatedayoftraining,weinvitedthementorsupervisorstothetrainingcentreandmatchedthemrandomlywithoneofthetraineesfromtheirfactory-thusassigningthe traineerandomlytoaproduction line for the trialperiod.Over twodayswith thementors in attendance, we conducted a series of team building exercises betweentraineesandmentors.Afterthesix-toeight-weektrial,factorieswerefreetoreturnthetraineetoapositionasoperator,leavethemasanassistantsupervisor,orpromotethemtosupervisor.

Therewasdropoutoftraineesatvariouspoints,detailedinFigure2.Thefactoriesinitiallyselected121femalesand96malesfortraining.Allwereinvitedtothetrainingcentrefortheinitialassessment.Ontheallocatedday,100femalesand99malesactuallyshowedup.Twenty-onefemalesdeclinedtocometothetrainingcentre,eitherbecausetheydecided theydidnotwant tobe supervisorsorbecauseof resistance from theirfamilies.Meanwhile,threeadditionalmalescameassomefactoriesreplacedthefemaleswhodeclinedtoattend.Admissiontothefulltrainingprogramdependedonpassingtheliteracyandnumeracytestadministeredat thetrainingcentre.The literacyexamwasdeveloped in conjunction with researchers at BRAC University.10 Nominees weredisqualifiediftheyscoredzerooneithertheliteracyornumeracyexam,oriftheyscoredbelow25percentonbothpartsoftheexam.Elevenfemalesand18malesdidnotpassthe literacy / numeracy threshold. An additional three females and five males weredisqualified for other reasons, mainly because the factory sent a male rather than afemale.11Finally,aftertheassessmentday,13femalesandfourmalesdecidedtheydidnotwanttocompletetraininganddroppedoutoftheprogram.Theremainingsample,

10 The literacy/numeracy test was developed by Sameeo Sheesh and Badrul Alam of BRACUniversity'sInstituteofEducationDevelopment(IED).Thecontentisbasedontheskillsrequiredto benefit from theOperator to Supervisor Trainingmaterial, and content taught in grades 5through8.11Inacoupleofcases,theliteracyexamwasmismarkedsothatafailingscorewasgivenwhentheexamwasamarginalpass.

14

allofwhomcompletedthetrainingcourse,was73femalesand72males.Figure2alsoshowsthenumberoftraineesworkingasasupervisoratvariouspointsaftertraining,whichwediscussinmoredetailbelow.2.1Data

Weconductedsurveysonsixseparateoccasions.First,priortothestartoftrainingweconductedacombinedsurveyandskillsassessmentforthetraineesatthetrainingcentre. The survey and assessment lasted a full day. In addition to gathering basicinformation on demographics,work history and attitudes,we assessed knowledge ofmachineandproductionprocesses,conductedcommunication,teachingandleadershipexercises,andtestednumeracy,literacyandnon-verbalreasoningskills.Theassessmentisdescribedinmoredetailbelow.

Second, near the end of the six-week training program, we asked factories tonominateproductionlinesandmentorsupervisorsinanumbermatchingthenumberoftrainees.Withthelistoflinesandmentorsinhand,weconductedabaselinesurveyinthefactorypriortotheendoftrainingandthestartofthetrail.Forthefactorysurvey,wesurveyed line operators, line supervisors, line chiefs, floor supervisors assistantproductionmanagers(floormanagers),productionmanagersandHRmanagers.Threeoperators andall of the supervisors and line chiefswere surveyedat the lineswheretraineeswereassignedtohavetheirtrial.Linechiefsfromthelineswheretraineeswereworking at the start of the training were also surveyed. The three operators wererandomly selected from the line in a way which ensured that at least two of theseoperators work directly under the mentor supervisor, and we select both male andfemaleoperatorswhereverpossible.

Third, on thepenultimateday of the training, thementorswere invited to thetraining centre and paired with their matched trainee. We conducted team buildingexercisesandalsoconductedasurveyandskillsassessmentwithboththetraineesandthementors. The survey and assessment was designed to capture any effects of thetrainingonthetraineesandtomeasuretheskillsofexperiencedmentorsupervisorsforcomparativepurposes.Fourth,attheendofthesix-toeight-weektrialperiod,weagaininvited the trainees back to the training centre for refresher sessions and groupdiscussionsoftheirexperienceduringthetrial.Ontherefresherdaywealsoconductedafinalskillsassessmentfortraineestomeasuretheeffectofthefactorytrial.

15

Thefifthsurveywasconductedinthefactoryjustafterthetrialperiodended.Weagainsurveyedthreerandomlyselectedoperators,thesupervisorsandlinechiefsofthelines that were nominated for the trial, and the assistant production managers,production managers, and HR managers. In addition, where there was either non-compliancewiththeassignmentoftraineestolines,orwheretraineeshadmovedfromthe assigned line to another line after the trail began, we surveyed operators andsupervisorsonthelineswhichwerenotnominatedforthetrial,butwheretraineeswereactuallyworkingasassistantsupervisors.

Finally, we conducted a second follow-up survey in the factory in October(trainingrounds1and2)andNovember(trainingrounds3and4).Thelastfollow-upwasthusaboutfourandahalfmonthsafterthetrailendedforthosetrainedintheearlyrounds,andtwoandahalfmothsaftertheendofthetrailforthosetrainedinthelaterounds.Thesurveysamplewasselectedusingthesamecriteriaasinthepreviousfactorysurvey, but because of time constraints, we were able to survey operators andsupervisorsonlyfromthelineswhereatraineewasworkingaseitheranassistantlinesupervisororalinesupervisor.Inaddition,allofthetraineesweresurveyedin-personiftheywerestillworkingatthesamefactory,andoverthephone,iftheyhadleft.

Inadditiontotheface-to-facesurveys,weconductedtelephonefollow-upsurveyswithtraineesatregularintervals.Duringthesix-toeight-weektrial,wecontactedthetraineeseveryweektotrackthelinetheywereworkingon,andthelevelofresponsibilitygiven to them. We also asked the trainees to keep a daily diary of their experienceworkingasanassistantsupervisororsupervisor.Afterthetrialended,wecontactedthetrainees everymonth untilMarch 2015 (four to ninemonths after the trial) to trackwheretheywereworking,andtheirdesignation.

Inadditiontothesurveydata,wealsocollecteddailyline-levelproductiondatafromeachfactory.WedescribethesedatainmoredetailinAppendixBandinSection6below.

5. EMPIRICALEVIDENCEAddroadmaphere:i)initialperceptions,ii)reality,iii)interventionandresults,

iv)mechanisms.

2.2CharacteristicsofTrainees

16

Table 1 shows basic demographic and skills data for the pool of trainees,comparedtoexistingsupervisorsandrandomoperatorswherethecomparisondataareavailable.Comparedwithasampleofrandomoperators,thetraineeshavetwoadditionalyearsofschoolingandjustmorethanhalfayearmoretenureinthefactory.Age,maritalstatusandexperienceinthegarmentsectoraresimilartootheroperators.Wesplitthesupervisorsampleintomentorsandnon-mentorsforthepurposesofcomparingtraineeswithexistingsupervisors.Weseethat,whilethetraineeshavemuchmoreschoolingthantypicaloperators,theyhavealmostayearlessschoolingthantypicalsupervisors.Theyarealso4.7yearsyoungerwith2.3yearlessexperienceinthesector.However,theageofthetraineesisstatisticallyidenticaltotheageoftherandomsupervisorsatthetimeoftheirpromotiontosupervisor.

Withregardtotherelativeskillsoffemaleandmaletrainees(notshownontable),wefindthatfemalesarejustoverayearyounger(p=0.05),buttherearenodifferencesin schooling or experience. Whether the trainees have less schooling than existingsupervisorsbecausefactoriesfaceashortageofworkerswithhigherschoolinglevels,orwhetherthefactorieshavenotselectedtheverybestsupervisorytalentforthetrainingprogramisnotclear.Butwhile62percentofexistingsupervisorshaveatleastalowersecondarycertificate(thatis,theyhavepassedO-levelexams),only14of430randomoperators(3percent)haveachievedthislevelofeducation.Thissuggeststhatfactoriesdofaceaverylimitedpoolofworkerswitheducationlevelscomparabletothepoolofexistingsupervisors.This,combinedwiththeageandexperienceprofilesofthetraineessuggeststhatthefactoriesselectedtraineesinamannersimilartothoseselectedintheusualpromotionroutine.

Wecanalsocomparetheskillsoftraineesandthementorsupervisorusingtestsadministeredatthetrainingcentreduringtheskillsassessment,thoughwelacksimilardata for other operators and supervisors. The bottom half of Table 1 shows that theliteracyandnumeracyscoresofthetraineesaresignificantlybelowthoseofthementorsupervisors.Thesedataprovidefurtherevidencethattheskillsofthetraineesarebelowthoseofthementorsupervisors.5.1. InitialPerceptions:FemalesasSupervisors

Atypicalfactoryinoursamplehasonlyoneortwofemalesupervisorsatbaseline.Therefore, operators andmanagers have little direct experienceworkingwith female

17

supervisors.Nevertheless,theyhaveperceptionsabouttherelativeabilityoffemalesandmalesassupervisors.Asafirststepinexploringtheseperception,weaskedemployeesat all levels of the factories to tell us which tasks are the most important for linesupervisors.Weconstructedalistofeightmaintasksfromaninitialsetofopen-endedconversationswithmanagers.Wethengaveeachrespondent10tokensandaskedhimorhertoplacethe10tokensonthelistoftheeighttasksplusan“other”categoryinawaywhichindicatedtherelativeimportanceofeach.Respondentsweretoldtheycouldplace all 10 tokens on a single task if they thought that it was the only task that isimportant,orspreadthetokensacrossthetasksastheywished.SurveyswereconductedwithHRManagers,ProductionManagers,AssistantProductionManagers,LineChiefs,LineSupervisorsandOperators.

Figure 3 shows the percentage of tokens placed on each of the eight tasks byrespondents holding different positions at the factory. The characteristics given thehighestweightsareshowntotheleftofthegraph.Onepatternthatemergesisthatalllevels of managers agree about which characteristics are important. Teaching andmotivatingoperatorsaregiven the largestweightsbyallmanagers.Operators,on theother hand give somewhat differentweights. They appear to prefer problem solvers,giving higher weights to understanding machines and correcting mistakes. There isagreement across the hierarchy that organizing resources, corresponding withmanagement,andgivingorderarelessimportanttasksofsupervisors.

We then asked the same set of respondents whether, based on their ownexperience,theythoughtfemalesormaleswerebetterateachoftheeighttasksofbeingasupervisor.Theallowedresponses includedtheoptionof“bothareequal".Wecodethesedatainawaythatindicatestheperceiveddeficitthatfemalesfaceineachofthetasks.Aresponse“malesarebetter"iscodedas-1,“femalesarebetter"iscodedas+1and“bothareequal"iscodedas0.ThescoresareshowninFigure4,againbytypeofrespondent.12Thefirsttakeawayfromthetableisthatmalesareoverwhelminglyseenashavinganadvantageineverysupervisorytask.Lineoperatorsandlinesupervisorsratemalesbetterinalleighttasks,linechiefsandproductionmanagersratemalesbetterinsevenoftheeighttasks,andHRmanagersseemalesasbeingbetterinfiveofthem.

12WedidnotasktheAssistantProductionManagerstomakethiscomparisonbecauseoftimeconstraintsonthesurveyinstrument.

18

Wealsofindaveryhighlevelofagreementaboutthespecifictaskswherefemalesaremost lacking.According toeverycategoryof respondent, femaleshave the largestdeficitsinunderstandingmachinesandorganizingresources.Allrespondentsalsoagreethat the threeareaswhere females are closest tomales are teachingnew techniques,motivating operators, and corresponding with management, though there is somedisagreementabouttherankingofthesethree.Noticethatthetwotasksratedasmostimportantbymanagersaretwoofthosewherethegapbetweenfemalesandmales isperceivedtobethesmallest.Ontheotherhand,machineknowledge,ratedhighestbyoperators,istheareawherefemalesareperceivedtobetheweakest.

Thesampleofoperatorsisthelargestandmostdiverse,soinFigure5,weshowthesamecomparisonsfordifferentsubgroupsofoperators.Firstwesplittherandomlyselectedoperatorsbygender.Therelativerankingsareverysimilarforfemaleandmaleoperators-thecorrelationis0.87-thoughfemaleoperatorsuniformlydescribeasmallergap.Nextwesplittheoperatorsintothosewhohaveandthosewhohavenotworkedforafemalesupervisoratsomepointintheircareer.Pastexperienceworkingforafemalesupervisorhasnosignificanteffectontheperceivedgapinfemaleskills.Finally,whenwe asked the trainees the same comparisons between generic male and femalesupervisors,theresponsesareveryclosetothoseofotheroperators.AsFigure5shows,femaletraineesdoratewomensomewhathigherthandootheroperators.

Wealsoaskedtraineesabouttheirownabilityrelativetotypicalsupervisorsintheirfactory.Wefirstaskedthetraineestoratethetypicalsupervisoronascaleof1-10withregardtoeachoftheeightsupervisoryroles,andthenaskedthetraineetorateher-orhimselfonthesamescale.Femaletraineesratethemselvesasworsethanthetypicalsupervisor on each of the eight characteristics,whilemales rate themselves better atmotivatingworkersandgivingorders.Theaveragegapformalesisonly0.09,whileforfemalesitis0.45.Acrossskills,thefemales'self-assessmentslargelymatchthepatternofthe gender perceptionsmore generally. The correlation between the gaps the femaletrainees perceive in themselves and the gaps that operators perceive in femalesupervisorsis0.68.

Weaggregatetheratingsofmalesandfemalesonalleightskillstocreateasinglevariable indicating each respondent's beliefs about the relative skills of males andfemales.Fortheaggregation,weassignavalueof1to“femalesarebetter",0to“malesarebetter"and0.5totheindifferentresponse.ThefirstcolumnofTable2showshowthe

19

averagedeficitforfemalesacrosstheeighttasksisaffectedbythegenderoftheoperatorandpastexperienceworkingwithfemalesupervisors.ConsistentwiththedatainFigure5,we find that female operators have slightly higher opinions of female supervisors,beingabout12percentmore likely tochoose “female isbetter"over “male isbetter".Previous reported experience working for a female supervisor does not change theperceivedskillleveloffemalesandmales.Inthesecondcolumn,wespilttheexperienceeffectbythegenderoftheoperator.Thereisnoeffectforfemaleoperators,whilethereisasmalleffectformaleoperators(p-value0.101).

We also asked operators whether they prefer to work for a female or malesupervisor.Similartothecodingforskills,wecodetheresponsesas1for“preferfemale",0for“prefermale"and0.5forindifferent.Asagroup,theoperatorssaytheyprefertoworkformalesupervisorsbyamarginofabouttwotoone.However,femaleoperatorsare17percentmorelikelytosaytheypreferfemales,andthosewithpreviousexperienceworkingforfemalesupervisorsare12percentmorelikelytosaytheypreferworkingfora female supervisor (Table 2, column 3). Again there appears to be, if anything, asomewhat stronger effect for male operators (column 4) - though as with the skillsassessment, thegapbetweenfemaleandmaleoperators isnotstatisticallysignificant.Amongthe140femaleoperatorsreportingexperienceworkingforafemalesupervisor,40percentsaytheyprefertoworkformales,30percentforfemalesand30percentareindifferent.Amongmaleswithnoexperienceworkingforfemales,thepercentagesare81,16,and3.

In sum, the skills assessment provides little evidence that perceptions areinfluencedbyexperience.However,whenaskedtoexpressapreferencetoworkformaleorfemalesupervisors,previousexperienceworkingforwomendoesappeartomatter,especiallyformaleoperators.

5.2. Reality:Domeasuredskillsmatchtheperceptions?

Thesurveysindicatethatfemalesupervisorsareviewedaslessskilledthanmalesupervisors in each of eight supervisory tasks. The female trainees see similarweaknesses in themselves. Do these perceptions match reality? We conducted anextensiveskillsassessmentofthefemaleandmaletraineesselectedbytheparticipatingfactoriesduringtheirfirstdayatthetrainingcentre.Weadministeredtestsofnumeracy,literacy,andnon-verbalreasoning.Wealsoassessedtechnicalskillsandknowledgeof

20

machines,andconductedteaching,communication,andleadershipexercises.Thedatafromthisassessmentprovideevidenceonseveraldimensionsof theactualskillsgapsbetweenfemalesandmalesselectedbyfactoriesashavingsupervisorypotential.Weusethesedatafortwopurposes.Thefirstistoassesstheextenttowhichperceptionsmatchrealityatthebaseline.Thesecondistomeasuretheeffectsoftrainingandthetrialperiodworking as an assistant supervisor on the trainees' skills. For the latter purpose,werepeatsomeoftheexercisesattheendoftrainingandafterthefactorytrialperiod.

4.1 Baselinemeasures:Dotheskillsgapsmatchtheperceptions?

Themostdirectandextensivecomparisonwecanmakebetweenperceptionsandrealityisonthequestionoftechnicalandmachineknowledge.Theassessmentaskedthetraineestonamedifferentpartsofsewingmachines,andtotelluswhichtypeofmachine(e.g.,atlock,singleneedle,etc.)wouldbeusedfordifferentsewingprocesses.Weshowedthetraineesgarmentsofthetypetheytypicallyproducewithfaultsinthem,andaskedthemto identifywhatmachineproblem(e.g., loose thread tension)wouldmost likelycausetheparticulartypeoffault.Weshowedthetraineespicturesofproductionlinesandaskedthemtoidentifyissueswhereworkersafetywasbeingcompromised.Inall,thediagnosticincluded86questions.Weconductedaverysimilarexerciseaftertrainingandthenagainafterthetraineescompletedthetrialinthefactory.

WeexaminedifferencesbetweenthefemaleandmaletraineesinTable3.Thefirstcolumnofthetableshowsresultsoffactoryfixedeffectregressionsusingallthreeroundsof the assessment. For now, we focus on the top line of the table, which shows thedifferencebetweenfemalesandmalesonthebaselineassessment.Inrawscores,malesoutperformfemalesbyasinglepercentagepoint,scoring65percentcomparedwith64percentforfemales.Theregressionshowsasimilargap,withfemalesonaveragescoreonepointloweronthe86-pointscale.Thefemale–maledifferenceishighlyinsignificant.Inotherwords, even thoughclose to90percentof survey respondents say thatmalesupervisors have more technical knowledge than female supervisors, we find nostatisticaldifferencebetweenthefemaleandmaletraineesselectedbythefactories.

Wealsoconductedexercisestomeasureteaching,communicationandleadership.Intheteachingexercise,wedividedthetraineesintogroupsoffourtosix.Weassignedeach trainee the role of teacher in one round of the exercise, with the others beingstudents.Theteacherwasgivenanabstractfigure,whichmightbeforexampleseveral

21

triangles and circles with some coloured in. The teacher's task was to instruct thestudentstoreproducethefigureusingonlyverbalinstructions.Shecouldnotshowthefiguretothestudentsoruseherhands.Weexaminetwooutcomemeasures.Thesimplestisthenumberofdrawingthatwerecorrect.Thefirstrowofcolumn2onTable3showsthatmalesobtainaslightlyhigherpercentageofcorrectdrawings,withthegapbeingmarginallyinsignificantwithap-valueof0.10.

The second outcome from the teaching assessment comes from observationsrecorded by two enumerators observing the exercise. For example, the enumeratorsrecordedwhethertheinstructionwasgivenatanappropriatepace,andthenumberoftimestheteacherexplainedthetaskinmorethanoneway.Wetakesixsuchobservationsandconstructstandardizedmeasuresforeachassessmentround.Wethensumthesixstandardizedindicatorvariablestocreateanindexof“softteachingskills".Column3onTable3showsafactoryfixed-effectregressionwiththisindexasthedependentvariable.We see no significant difference between females and males in baseline teachingtechniques,thoughthestandarderrorsarelargerthanwemightlike.Wealsonotethatthe soft skills measure is not significantly associated with the harder outcome - thepercentageofcorrectdrawings-thoughthemeasuredeffectispositive(p=0.22).

We create similar ‘soft'measures as ourmain outcome in the communicationexerciseandtheleadershipexercise.Inthecommunicationsexercise,thetraineeswereaskedtogiveashortspeechonatopicrelatedtorulesinthefactory,suchas:“Describetoanewoperator,allthethingsthatyouneedtodowhenyourmachinebreaks".Duringthespeech,thetraineewasinterruptedwithquestionsontwooccasions.(Forexample,“WhatshouldIdoifIthinkIcanfixthemachinemyself?").Twoenumeratorsrecordedjudgementsonwhetherthetraineesspokeclearly,atareasonablepace,whethershehadconfidence,etc.Thetoprowofcolumn4inTable3showsthatfemaletraineesperforminsignificantlyworsebythesemeasures.Finally,intheleadershipexerciseweaskedthegrouptocreateaproductionhierarchy,andthenaskedthemtoproducesome‘products'usingLegos.Theprecisehierarchydependedonthesizeofthegroup,butwemeasurewhethertherearedifferencesacrossthegendersintheprobabilityofbeingappointedamanagement role, and in softmeasures reflecting the extent to which the individualparticipatedactivelyinthediscussion.Thetoprowofcolumn5showsthatfemalesarescored insignificantly loweron the soft skillsmeasure.Butwedo find thatmales are

22

significantlymore likely to be appointed tomanagement (75 percent vs. 32 percent,p<0.001).

Theteaching,communicationandleadershipexerciseswereintendedtomeasureimportant aspects of confidence andpreparedness to lead a production line.We alsoelicitedadirectmeasureofself-confidenceofthetrainees.Wefirstaskedeachtraineetoratetheaveragesupervisorinher/hisfactoryonascaleof1-10.Wethenaskedthemtoratethemselvesasasupervisor,twomonthsafterbeginningthejob.Thetakethegapbetweenthetypicalsupervisorandthetrainee’sownexpectedperformanceasameasureofself-confidence.Atbaseline,wefindthatthemaletraineesexpressmoreconfidenceintheir ability. In the raw data, they rate themselves 0.33 points lower than a typicalsupervisor,whilethefemaletraineesratethemselves0.79pointslower.Thetoprowofcolumn6 inTable3 showsa similardeficit forwomenof 0.47points, controlling forfactory fixed effects, significant at the .10 level. Thus, while we find no significantdifferences between the female andmale trainees in the technical assessment or theteachingandleadershipexercises,wedoseedifferencesintheirself-reportedconfidencelevels.

5.3. InterventionResults:Promotion,ComplianceandPerformance

4.2 TrainingandTriallingeffects

Werepeatedtheteaching,communicationandleadershipassessmentsattheendofthesix-weektrainingperiodandagainattheendofthefactorytrialperiod.Forthelatterassessment,thetraineesreturnedtothetrainingcentreforareviewdayduringwhichweconductedtheseassessmentsaswell.Rows2and4ofTable3showthevariouspost-trainingmeasures,allmeasuredrelativetobaseline.Atthebottomofthetable,weshowthep-value fortestsofequivalenceof femaleandmaletraineesateachpoint intime.Thetableisanunbalancedpanel,astherewereseveralnominatedtraineeswhoeitherfailedtheliteracy/numeracyexamordroppedoutforotherreasons,andtherewasfurtherattritionbeforethepost-trialreviewday.However,resultsfromregressionsusingthebalancedpanelareverysimilar,suggestingthatthepatternsweobservearenotdrivenbyselection.

Looking first at the scores on the assessment of technical knowledge, we seeinsignificantimprovementsinbothmalesandfemalesafterthetraining,andnochangeintherelativeperformanceacrossgender.Relativemalesatbaseline–thebasegroupfor

23

theregressions–femalesperform0.5pointsbetterandmales0.2pointsbetterfollowingthetraining.Afterthefactorytrial(rows3and5),however,weseetheperformanceoffemalesappearstodeterioratesomewhat.Indeed,comparingfemaleandmaletrainees,thepost-trialtechnicalassessmentistheonlymeasureshowingasignificantdifferencebygender.Weareunsurewhatmightexplainthisdipinperformance,butwefindthesame effect in the balanced panel. The other outcome worth noting is that the self-confidence increases significantly after training for both males (by 0.5 points) andfemales(by0.71points).Themagnitudeoftheincreaseisslightlylargerforfemales,butnotsignificantlyso.However,theself-confidencegapbetweenfemalesandmalesshrinksandbecomestatisticallyinsignificant.

6. Howareoperatorperceptionschangedbyexperience?Theskillsdiagnosticsindicatethatthefemaletraineeshaveonlyaverysmalland

statistically insignificant gap in technical skills. On the other hand, there are moresignificantgaps inself-confidenceand in theoutcomesof the teachingand leadershipexercises.Thetrainingclosesthesegaps.Buttheoutcomesontheproductionfloorareofmoreinterestthantheoutcomesofthediagnostics.Weexaminetheseusingbothsurveysofoperatorsworkingforthetraineesandusingadministrativedataonproductivityofthelineswherethetraineesareassigned(ITT)orwork(OLS).

Weconductedafirstfollow-upsurveyinthefactoryjustaftertheendoftheinitialtrialperiod.Duringthesix-toeightweeksbetweentheendofthetrainingandthisfirstfollow-up survey, trainees were meant to be working as assistant line supervisors,togetherwiththeirmentor.Compliancewiththisagreementwasveryhigh.Ofthe135operatorscompletingtraining129weretrialledasanassistantsupervisor.Fourofthesixnottrialled(threefemalesandonemale)leftthefactorybeforethetrialstarted.Recallthatwerandomlyallocatedthetraineestooneofthelinesselectedbythefactoryforthetrials.Wecanmeasurenon-compliancewiththisassignmenteitherattheindividuallevelorat thegender level. Some individualswere trialledonadifferent line than theoneassigned.Inmanycases,thefactoryswitchedtwofemalesortwomales,leadingtonon-complianceattheindividuallevelbutcomplianceatthegenderlevel.Weareprimarilyconcernedwithcomplianceatthegenderlevel–that is, thatfactoriesplacedafemale(male)traineeonalineassignedtoafemale(male).Withingendernon-compliance,therearetwotypes.Themajorityofthenon-complianceinvolvedthetraineebeingplacedon

24

alinenotincludedintheoriginalpooloftriallines. Indeed,34percentofthetraineesweretrialledonalinenotselectedasatrialline.Butinthe77caseswheretraineesweretrialledonthelinesdesignatedfortrials,therewascomplianceonthegenderlevel–i.e.,femalesonlinesassignedtofemales–in71cases.

At the first follow-up, conducted at the end of the trial period, we surveyedrandomlyselectedoperatorsworkingonallofthelineswherethetraineeswereassigned(ITTlines)andonallofthelineswherethetraineeswereactuallyworking.However,atthesecond follow-up,wewerenotable tosurveyallof the ITT lines,andhencehaveinformationonlyonthelineswherethetraineeswereworkingatthattime.Thereasonforthisislogistical.Recallthatforeachfactory,thetrainingwasconductedintworoundsapproximatelytwomonthsapart.Thefirstfollow-upsurveywasthenconductedontwodifferentdaysineachfactory,attheendofthefactorytrialforeachtraininground.Thesecondfollow-upsurvey,however,wasconductedonbothtraininggroupsatthesametime.Thismeantthatweweresurveyingtwiceasmanylinesonthedayofthesecondfollow-up, limiting our flexibility with regard to ITT lines. The simultaneousimplementationofthesecondfollow-upsurveyalsoimpliesthatthetimegapbetweentheendofthetrialandthesecondfollow-upsurveywasabouttwomonthslongerforthetraineesinthefirsttrainingroundthanforthoseinthesecondround.

Asaresult,weareabletoreportbothITTandOLSregressionsforthefirstfollow-upsurveydata,butonlyOLSregressionsforthesecondfollow-upsurvey.Ateachfollow-upsurvey,weselectedthreeoperatorsatrandomfromeachofthesurveyedlines.Wefocusontwooutcomes.First,weaskedtheoperatorstorankonascaleof1-10bothatypicalsupervisorinthefactoryandthetraineeontheirlinebasedontheirknowledgeofher/him.Weregresstherankingofthetraineeonanindicatorforhisorhergenderand the gender of the surveyed operator, controlling for the ranking of the typicalsupervisorbytheoperator.Second,weaskedtheoperatorswhethertheyprefertoworkforafemaleormalesupervisor,andasbeforecodetheresponsesas1for“preferfemale",0 for “prefermale", and 0.5 for “indifferent". For the first of these outcomes,we areinterestedintherankingoffemaletraineesrelativetomaletrainees,andforthesecond,we are interested inwhether exposure to a female trainee affects the preference forsupervisors.

ThefirstthreecolumnsofTable4belowshowtheITTregressionsfortherelativeranking(columns1and2)andthepreferenceforfemalesupervisors(column3).Wefind

25

thatthefemaletraineesareratedalmostapoint-about0.4standarddeviations-lowerthanthemaletrainees.Incolumn2,weallowtherelativerankingtodifferforfemaleandmaleoperators.We find, ifanything,malesrate the femalesmoreharshly, though thedifferenceisnotstatisticallysignificant(p=0.26).Incolumn3,weseethatexposuretofemale trainees has the effect ofmakingmale operators significantly less opposed toworkingwithfemalesupervisors.Whilefemaleoperatorsaremoreinclinedthanmaleoperators to say they prefer to work for female supervisors, their opinion is notinfluencedbyexposuretothefemaletrainees.

Columns4-7ofTable4repeatthesameregressionsusingtheactualplacementofthetrainees.Wefindalmostidenticaleffectsintherankingregressions(columns4and5),butslightlyweakereffectsinthepreferenceregressions(column7).Finally,columns8-10showtheresultsofOLSregressionsusingthesecondfollow-upsurveydata.Becauseweusethesampleoftraineesworkingasassistantsupervisorsorfullsupervisorsatthetimeofthesecondfollow-up,incolumn6weshowthefirstfollow-upresultsusingthesampleoftraineesworkingassupervisorsatthetimeofthesecondfollow-up.Weseethattheresultsformaleoperatorsareverysimilartothoseinthefullsample(comparecolumn6withcolumn5),thoughthesmallersampleyieldshigherstandarderrorsandaninsignificanteffect.Theresultsforfemaleoperatorsappearslightlydifferent,andlessnegative,forthesampleoftraineesthatcontinuetoworkassupervisorsatthesecondfollow-up. This indicates that the weaker female trainees may be those who do notcontinueassupervisors.

Inthesecondfollow-upsurvey,thedeficitforfemaletraineesiserasedcompletely(Seecolumns8and9ofTable4).Femaletraineesareratedasequaltomaletrainees,bybothfemaleandmaleoperators.Moreover,operatorsofeithergenderswhoareexposedtothefemaletraineesexpresshigherpreferencesforworkingwithfemalesupervisors.Noteaswellthatthetraineesasawholearenowratedasslightlybetterthanthetypicalsupervisor in the factory. The improvement in the relative ranking of the trainees isconsistentwithstatementsbyproductionmanagersthatnewsupervisorsrequirefourtosixmonthsofexperiencetoreachtheirfullpotential.

7. TraineePerformancemeasuredbyProductionData6.1Productionoutcomes

26

Theliteraturemeasuringtheeffectsofjobtrainingprogramshastypicallyreliedonoutcomemeasuressuchasemploymentorearningsoftrainees.13Thisisreasonableifwagesequalthevalueofthemarginalproductoflabour.Inourcontext,webelievethisapproachhasdrawbacks.First,thefactoriestypicallyhaveveryspecificwagesforeachworkergrade.Manyormostofthesearedeterminedbyminimumwagelevels,whicharesetnationallyandvaryattheworkergradelevel.Thus,wagesmaynotreflectmarginalproducts.Second,factorieswillattempttomakepromotiontrainingdecisionsbasedontheirbeliefsaboutactualproductivityoftheworkersratherthanthechangesinwages.Managerswereuniformlypuzzledbytheclaimthatwecouldlearnanythinglookingatchangesinearningsfollowingtraining.

Withthisinmind,wehaveattemptedtogatherverydetailedproductiondataforeachofthefactories.Forthesecondphaseoftheproject,wehavedaily,line-leveldatafor12or13months,typicallystartingtwomonthspriortothebeginningoftrainingandextendingseventoninemonthsaftertheendofthetraining(seeAppendixBforamoredetaileddescriptionofthedataanditscollectionprocess).Therearethreeoutcomesofinterest:productivity,qualitydefects,andabsenteeism.Byfocusingonsewing,weareable tocaptureameasureofoutputwhich isveryclose thepurequantitymeasure.Atrainedindustrialengineercantakeanygarmentandestimatethenumberofminutesafully-efficientworkerwill take toproduce thegarment.Thesecalculationscome fromsummingtherequiredtimeforeachstichtomakethegarment.Thetimescomefromacombination of international databases and in-factory time-and-motion studies. Bymultiplying these ‘standard minute values' - SMVs (or standard allowable minutes -SAMs)bythenumberofunitsofagivengarmentwhichareproducedduringtheday,weobtain a measure of output - output minutes - which is highly comparable acrossproducts. For example, a line producing 1,000 shirtswith an SMV of 15minutes hasproductionof15,000outputminutes.

For productivity,we divide the outputminutes by inputminutes - the sum ofminutesworkedbyoperatorsandhelpersonthelineoverthesametimeperiod14-toobtaintheindustrystandardmeasureofefficiency.ThisisessentiallyameasureofQ/L:

13Muchofthisliteraturefocusesonprogramsaimedatindividualswhoareoutofwork.See,forexampleCardetal2011;Attanasioetal2011.14Wecouldimprovetheinputminutesmeasurebyastepifwehadthewagebillforthewholeline.However,theindustrytypicallyusesthreedifferentwagegradesforoperators,andwemostoftenknowonlythetotalnumberofoperators,notthenumberbygrade.

27

Output*SMV/[(Operators+Helpers)*hours*60mins] (1)Theaverageefficiencyinthesamplewearecurrentlyusingis53percent,which

ishigherthanthe38-40percentthatthoseintheindustrytypicallyquote.15Asecondmeasureofinterestisthenumberofqualitydefects.Factoriestypicallyreportboththenumber or percentage of garments that require some re-work and the number orpercentagethatmustberejected.Rejectratesaretypicallyverylow,averaginglessthan0.5percentinoursample.Reworkratesaremuchhigher,averagingaround7percent(withamedianofalmost5percent).Becausethere-worktimeisincludedinthemeasureof“inputminutes",theefficiencymeasureincorporatesimprovementsinquality.

We construct a panel at the line level, with dummy variables indicating thepresence of a trainee working on the line either as an assistant supervisor or a fullsupervisor. We begin with an ITT specification, using the gendered assignment of atrainee on the line during the trail period, and then assuming this initial assignmentpredictsthelineonwhichthetraineewillbepromoted.

45678 = 97 + ;68 + <5=>?@A6785∈C,D + E5FGH=_=>?@A6785∈C,D +J5678 (2)whereg={0,1}representsmaleorfemaletrainees,fisfactory,lline,dtheweekof

production,andytheoutcomeofinterest.TRIALreflectstheassignmentofthelinetoafemale/maletraineeduringthetrialweeksandPOST_TRIALtheassignmentofthelinetoafemale/maletraineeduringtheperiodafterthetrial.

WealsopresentOLSresultsontheactualplacementandrolesoftrainees.Thesemay suffer from both the endogenous placement of trainees and the endogenousdecisionstopromote.AswiththeITTregressions,weincludebothlineandfactory/weekfixed effects, which mitigates to some degree the issue of endogenous placement.However,someofthetraineesleavethefactoryandsomereturntobeingoperatorsafter

15Thehigherefficiencyinoursamplemaycomefromhavingamoreefficientsampleoffactories.However, thedata across factories arenot always comparablebecause the international SMVvaluesareoftenadjustedupwardsbyfactoriestoaccountforsomeexpectedlevelofinefficiency.Weare currentlyworking toensure thedataare comparableacross factories,butwe includefactoryfixedeffectsinalloftheregressionsusingproductiondata,whichwillabsorbsystematicmeasurementdifferencesacrossfactories.

28

thetrial.Sincetheseoutcomesaremorefrequentforfemalesthanformales,weshouldclearlybeconcernedwiththeendogenouspromotiondecisionsininterpretingtheOLSregressions.WeneverthelessthinkthattheOLSresultsarepotentiallyinterestinginspiteoftheseselectionissues,becausepromotionofalmostanyfemalesrepresentsachangerelativetowhatlikelywouldhavehappenedintheabsenceoftheexperiment.

The first three columns of Table 5 report the ITT regressions for efficiency,absenteeismanddefectrates.Thesamples foreachof theregressionsvarysomewhatbecausedataonsomemeasuresarenotavailableinsomefactories.16Thecleanestresultsrelate toefficiency.Compared to lineswithout trainees,wesee that lineswheremaletrainees were assigned are about 2.3 percentage points - roughly 5 percent - moreefficientduringthetrialperiod.Duringthetrial,thetraineesrepresentextrasupervisorylabour on the line. Hence, even though they are least experienced at this point, it isperhapsnotsurprisingthattheyhaveapositiveeffectonefficiency.Thereisnoincreaseinefficiencyduringthetrialperiodonthelinesassignedafemaletrainee,suggestingthateventhoughthefemaletraineesareadditionalsupervisorylabour,theyarenoteffectivein increasing efficiency. However, the situation changes during the post-trail period.ThosetraineesremainingassupervisorsmayeitherbeclassedasAssistantSupervisorsorasfullLineSupervisorsduringthisperiod.Inthelattercase,andperhapsevenintheformer, they are replacing an existing line supervisor, andhenceno longer representincrementalsupervision.Duringthisperiod,thefemaletraineescatchuptothemales.Weseethatbothfemaleandmaletraineeshaveverysimilareffectsonefficiency,withpositivecoefficientswhichareeconomically importantbutstatistically insignificantatconventionallevels.

Columns4 through7 presentOLS results based on actual assignment.Weuseactual assignmentbecause the initial line assignmentwas agreed to only for the trialperiod.Wedidnotnecessarilyexpectthefactoriestopromotethetraineestothesamelines.17ThepatternsareverysimilartotheITTregressions,thoughthecoefficientsaregenerallyofslightly largermagnitude.Theregression incolumn4showsthat femalesperformsignificantlyworseduringthetrialperiod,performequallywellasmaleswhen

16Thesamplesizedropsbyabout75percentwhenweuselinesforwhichallthreevariablesarenotmissing.17Inafutureversionofthepaper,wewillmatchlinesbasedonpre-trialcharacteristicstoobtainasomewhatcleanercomparison.

29

both are assistant supervisors, and perform insignificantly better thanmale traineeswhenbothhavebeenpromoted to full supervisor. Incolumn5,we limit tosample toobservationsfromdayswhenthetraineewasworkingononeoftheoriginalITTlines.Thepatternsaresimilar,thoughnowthebetterperformanceoffemaletraineesasfullsupervisors is marginally significant (p=.096). Columns 6 and 7 report results forabsenteeism and defect rates, respectively. Again the patterns are similar to the ITTregressionsexceptthatunderperformanceoffemaletraineesrelativetomaletraineesonqualityissuesisalmostsignificantwhenworkingasassistantsupervisors(p=0.111;seebottomoftable).

The efficiency results mirror the opinions of operators working on the lines.Femaletraineesstartslower;theyperformsignificantlyworsethanmalesduringthetrialperiod.However, they catchup in themonthsafter the trialperiod.Wesee the samepatternintheITTandOLSregressions.IntheITTregressions,thegainmadebyfemaletraineesrelativetomaletraineesissignificantatthe0.10level,whilethegainindefectratesismarginallyinsignificant(p=0.125).18IntheOLSdata,wefindsignificantrelativeperformancegainsbetweenthetrialandpromotiontolinesupervisorsforefficiency,andbetweentheperiodworkingasanassistantsupervisorandpromotiontolinesupervisorfordefectrates.6.2Doattitudesadjust?

Both the surveydata and theproductiondata suggest that the female traineesstartmoreslowlythantheirmalecounterparts,butcatchupthreetofivemonthsafterreturning fromtraining.Theattitudesofoperatorswithdirectexposuretothe femaletrainees adjust over this time.Wemight askwhether there is any evidence that theattitudeadjustmentismoregeneral.Thatis,doestheincreaseinfemalesupervisioninthefactoryhaveindirecteffectsonoperatorattitudestowardsfemalesupervisors?Thedatasuggest there isnochange toattitudesofotherworkers:Thesumof thegenericfemale/malerankings-coded,asbefore,1/0/-1-is-5.06,-4.97and-5.19forthemaleoperatorssurveyedatbaseline,firstfollow-upandsecondfollow-up,respectively,and-

18Thisisthep-valuecomparingthegapbetweenfemaleperformanceinthepost-trialperiodandfemaleperformanceinthetrialperiodwiththesamegapformales.

30

3.33,-3.46and-3.28forfemaleoperatorsatthesamesurveys.Noneofthedifferencesacrosssurveyperiodsarestatisticallysignificant.

Directexposuretofemaletrainees,ontheotherhand,hasasignificanteffectonthese rankings by the time of the second follow-up survey: Female (male) operatorsworkingonlineswithfemaletraineeshaveacumulativerankingacrosstheeighttasksof-2.86 (-4.71), comparedwith -3.71 (-5.28) for thoseon lineswithamale trainee.Thefemaleoperatorgap issignificantwithap-valueof0.07.Thus,genericattitudesshowsomeevidenceofmovementwithdirectexposure,butthereisnoevidenceofanybroadereffectinthefactory.

Survey data available from the first phase of the project indicate that directexposureleadstoadifferentsortofattitudechange,particularlyamongmaleoperators.Weaskedoperatorshowlongtheyexpectedtocontinuetoworkatthefactorywherethey were currently working, and whether they expected to be promoted to linesupervisor one day. Two-thirds (69%) ofmale operatorsworking on lineswithout afemaletraineesaidtheyexceptedtoremainatthefactoryformorethanfiveyears,and95percentsaidtheyexpectedtobepromotedtosupervisor.19Thesepercentagesbothfallsignificantlyamongmalesworkingonlineswithafemaletrainee;only38percentexpecttostayatleastfiveyearsand83percentexpecttobecomeasupervisor(insomefactory).Theattitudesoffemalesmoveintheoppositedirection,butthemovementsaremuchsmallerandstatisticallyinsignificant.Thepercentageexpectingtostayfiveyearsormoreincreasesfrom39percentto44percentamongfemaleswithoutandwithfemaletraineesontheline.Theself-reportedlikelihoodtheywillbecomeasupervisorincreasesfrom58percentto64percentwithexposuretoafemaletrainee.

These two sets of results both underscore the potential costs to individualfactoriesofmakingthetransitionfroman(almost)all-malesupervisoryforcetoamixed-gender supervisory force. The attitudes toward ability of the female supervisors arechangedonlywithdirectexposure,implyingthateachlineneedstobeexposedtofemalesupervision before beliefs aboutwomen’s skills are increased. And the initial changeleadstoare-assessmentofprospectsformaleoperators,andpotentiallythelossoftalent.

19Certainlythesenumbersareexaggerated,butperhapsnotbyasmuchasitseems.Givensupervisor/operatorratiosandthegrowthofthesector,itislikelythathalfofthemaleoperatorswouldbecomesupervisorsinthe“almostallmale”world.Itisreasonablethatsomewhowillnotmakeitstillexpect,exante,tobecomeasupervisor.

31

7.InitialUnderperformance:beliefsvs.skills

Both the operator opinions and the production data suggest that the femaletraineesunderperforminitially.Thisinitialunder-performancemightariseforeitheroftworeasons.Thefemaletraineesmaysimplybeweakersupervisorsaftertrainingwhentheyareassignedtoaline.Theymightbeeitherbecausetheirskillslagorbecausetheylack self-confidence, and hence are less effective as leaders. Alternatively, they mayperformequallywell as supervisors, but those theyworkwith– either theoperatorsworkingunderthem,theirpeersassupervisors,ortheirsuperiors–maybebelievetheyareweakersupervisors,andhencebemorelikelytoquestiontheirleadership.Inotherwords,theymakelackauthoritynotbecauseofskillsbutbecauseothersbelievetheylackauthority, for example because of their own prior beliefs that women are weakersupervisorsthanmen.Welackdefinitiveevidencewhichwouldclearlyseparatethesetwo possibilities, but two pieces of data provide at least a suggestion that theunderperformancearisesfromalackofauthorityarisingfromthebeliefsofco-workers.Thefirstdatapointissimplythattheskillsassessmentconductedaftertrainingrevealednodifferencesbetweenfemaleandmaletrainees.(SeeTable3.)Moreover,althoughtherewasaninitialconfidencegap,thatgapwasalsoclosedfollowingtraining. A somewhat stronger piece of evidence comes fromamanagement simulationexerciseweconductedduringthefirstphaseoftheproject.Theexercisewasconductedduring a follow-up survey around four months after the completion of training. Thesimulationinvolvedthetraineesandeightrandomlyselectedoperators.Theoperatorswere placed into four teams of two each and played two “production" games, oneinvolvingLegosandoneinvolvingbuttons.Werandomizedtheorderinwhichthegameswere played at the factory level. Each teamwas assigned a leaderwhose jobwas toexplaintheparticularexerciseandmanagetheoperatorsastheyperformedtheirtasks.Fortheresultswepresenthere,theteamleaderwaseitherafemaleormaletrainee.20Eachpairofoperatorsplayedtheproductiongametwice,oncewithLegosandoncewithbuttons.Eachteamleaderplayedonlyonesession-eitherLegosorbutton-sotherewere

20Thefullexercisealsoinvolvedteamleaderswhowereoperatorsfromthecontrolgroup,themostrecentlypromotedsupervisorwhowasnota trainee,anothersupervisor fromthesameproductionlineasoneofthetraineesselectedatrandom(a“matched"supervisor),oranothersupervisorselectedatrandom(a”random"supervisor).

32

eightteamleadersineachfactory,andeachpairofworkersplayedwithtwodifferentteamleaders.

ForeachoftheLegoandbuttonexercises,theteamsplayedfiveseparatesessions.Thefirstwasasimplesortingexerciseineachcase,sortingeitherbuttonsorLegosbycolour.ForLegos,thesecond,thirdandfourthsessionsinvolvedconstructingchainsofLegoswithaparticularcolourpattern-blue,yellow,green,blue,yellow,green,etc.Thethreegamesweredifferentiatedby theirpayoffs: thepayoff in the second roundwasdeterminedbysummingthelengthofthechainsproducedbythetwooperators;inthethirdroundwepaidbasedonthelengthofthelongestchainproducedbyeitherworker;andthefourthroundwepaidbasedontheshortestchainproducedbyeitheroperator.Theteamleadersweregivenincentivepaymentsaccordingtothepayofffunction.

Hereweassesstheperformanceofteamsledbyfemaletraineeswiththatofteamsledbymale traineesmeasuredbythepayoffs.Wecombineeachof the five individualgamesintoasingleregressionbystandardizingthepayoffsonthegame-roundlevel.Wethen run regressionswith the standardizedpayoffs on the left-hand side anda set ofcontrols for characteristics of the team leader on the right hand side. We focus thediscussionhereonthesubsetofgameswheretraineesareteamleaders,comparingtheperformanceoffemaleandmaleteamleaders.

Eachpairofoperatorsplaysthesetoffivegamestwice,withoneteamleaderinthefirstsessionandadifferentteamleaderinthesecondsession.Theorderofthegames(Lego-buttons,orbuttons-Lego)israndom,andwithinaroundtheassignmentofteamleaderstooperatorpairswasrandom.Buttheassignmentofteamleaderstosession1orround2dependedonthe(non-random)order inwhich leaderswereprovidedbythefactories.Logisticalcomplexitiesworkinginthefactorypreventedusfromrandomizingthe session in which any team leader participated. In particular, because factoriesanticipated thatwewanted to talkwith trainees, the traineesweremore likely to beassigned to the first session, and theexisting supervisors and controloperatorsweremorelikelytobeassignedtothesecond.Thismatters,becauseevencontrollingfortheteam leader and game types (Lego vs. buttons), operators were significantly moreproductiveduringthesecondsession.Thisislogicalbecauseweexpectsomelearningbytheoperatorsfromthefirsttothesecondsession-eventhoughtheyplaydifferentgamesineachsession.Wecontrolforthesessionordereffectsinregressions.

33

Table6 showshow the standardizedpayoffsvarywith thegenderof the teamleaderinthesampleofgamesinvolvingfemaleandmaletrainees.Thespecificationincolumn1includescontrolsforfactory,session(firstorsecond)andgamefixedeffects.We find that teams led by female trainees have payoffs which are 0.29 standarddeviationshigherthanteamsledbymaletrainees,adifferencewhichishighlysignificant.Inotherwords,femaletraineesappeartobemoreeffectiveasteamleadersthanmaletrainees.Column2addsteamleaderdemographics-age,education,industryexperienceandfactorytenure-andColumn3addsoperatorteamfixedeffects.Notethatthethirdregression then isolates the caseswhere a single teamwas ledbyboth amale and afemaletrainee.Only19teamshadthispairofteamleaders,sowhilethetableshowsasample size of 600, the effective size is much smaller. Nevertheless, the additionalproduction by female-led teams is statistically the same, increasing slightly to 0.42standarddeviations.

Incolumns4and5weexaminewhethertraineeswhobeforethesurveyvisithadbeentriedoutassupervisorsorpromotedtosupervisorperformbetterthanthosenottriedoutorpromoted.Wefindthatthosepromotedtosupervisorperformsignificantlybetterthanthosenotpromoted.Sincepromotionisnotrandom,weareunabletosaywhetherthis isdueentirelytoselection-moreabletraineesarepromoted,while lessableonesarenot-orwhethertheexperienceasasupervisoralsomakestheindividualmoreeffectiveasaleader.Buttheresultdoesprovidesomevalidationfortheexerciseitself,showingthatthosewithmoreabilityorexperienceperformsignificantlybetterinthegame.

Finally,incolumn6weexplorewhetherthegendercompositionoftheoperatorteaminteractswiththegenderoftheteamleader.Wecomparetheperformancewhenbothoperatorsarefemalewiththatwhereatleastoneoperatorisamale.21Thesuperiorperformanceofthefemale-ledteamsissignificantonlyfortheall-femaleteams.Whenbothoperatorsarewomen,theoutputis0.41standarddeviationshigherwithafemaleleader thanwith amale leader.But female leaders obtainnohigherproduction frommixedteamsthanmaleleadersdo.

We find, then, that the female trainees were significantly more effective ingeneratingpayoffsthanwerethemaletrainees.Traineeswhowerepromotedbeforethe

21Sinceonly20percentofoperatorsaremale,all-maleteamsareveryrare.

34

timeofthefirstfollow-upsurveyperformsignificantlybetterthanthosenotpromoted.Crucially, female trainees only perform betterwhen they arematchedwith a pair offemaleoperators,andperformnobetterthanmaletraineeswhentheyleadoperatorswhoaremale.Sincetheteamleaderswererandomlyassignedtotheoperatorpairs,thequalityofleadershipprovidedbythefemaletraineesisindependentofthecompositionoftheteam,eventhoughtheoutcomesdiffersignificantly.

Therearetwofurtheroutcomesfromthegames.Thefirstinvolvesthestrategychoicesoftheteamleaders.Recallthatthepayoffschangedfromonegametothenext.Inthesecondround,payoffswerebasedonthesumoftheoutputofthetwooperators,butthethird(fourth)game,wepaidonthemaximum(minimum)outputofeitheroperator.Afterthesecondgame,weaskedeachteamleaderwhichofthetwooperatorswasbetteratthegame.Wethenrecordedwhethertheteamleaderfocusedmoreattentiononthestrongeroperator in game3 andon theweakeroperator in game4, as these are theoperatorswhoseperformanceisexpectedtodeterminethepayoffsinthesegames.22

Thefifthgameinvolvedacomplexfigurethatwasmostefficientlymadeina“line",witheachoperatorspecializingononecomponent.Werecordwhethertheteamleaderorganizedproductioninthatmanner.Wethensumthenumberoftimestheteamleaderadoptedthe“correct"strategyineachofthesethreegames.Column7regressesthissumonthegenderoftheteamleader,demographicsoftheteamleaderandtheoperators,andfactory fixed effects. We find that the male leaders adopted the correct strategysignificantlymoreoften,inspiteofthefemaleleadersinducinghigheroutput.

Finally,afterthesecondsession,theoperatorsontheproductionteamwereaskedtocomparethemanagementstyleofthetwoteamleaderstheyworkedwith.Theywereaskedwhetherthefirstorsecondteamleadertheyworkedwithwasbetteratexplainingthegame,betteratansweringquestions,betteratmotivatingthem,alwayspressuringthem,andsoforth.Focusingontheresponsesofthe19teamsledbybothafemaleandamaletrainee,wefindthatoperatorsmoreoftensaidthatthemaletraineeswerebetteratansweringquestion,atmotivating,andatencouraging,thoughthegapsacrossgenderareneverquitesignificantatthe.10level.Femaletraineeswereselectedmoreoftenonly

22Weaskedtheteamleadersbothwhichoperatorwasfasterandwhichmadefewermistakes.Wereportresultsbasedonthefastestoperator,buttheresultsarequalitativelythesameifwedefinebestasbothfastestandmostaccurate,andsettomissingthoseobservationswheretheteamleaderoffersasplitdecision.

35

as “always pressuring", an effect which is significant at the .05 level. The last twooutcomes,onstrategyandoperatoropinions,areinterestinginthelightofthesuperiorperformanceoffemaletrainees.

In sum, the post-training skill assessment and the management simulationexercisebothindicatethatfemaletraineeswereasskilledasmaletraineeswhentheseexerciseswereconducted.Wereadtheseassuggestivethatthepriorbeliefsofoperators,and perhaps of other co-workers as well, explain at least partly the initial under-performanceoffemaletrainees.

8Conclusion

Weexaminetheimbalancebetweenthepercentageoffemalesamongproductionworkers and the percentage of females among supervisors in a set of large garmentfactoriesinBangladesh.Surveyresponsesshowthat,atbaseline,factoryemployeesatalllevelsperceivemalestobemoreeffectivesupervisors.Butadetailedskillsassessmentconductedwith thewomenandmenselectedby the factories for training shows thatperceptionssometimesdeviatefromreality.Thisismoststrikingwithregardtomachineandtechnicalknowledge,whichshowsthelargestperceivedadvantageformalesandnoactualdifferencebetweenfemalesandmales.

Wepartneredwithlocaltrainingcentrestoprovidetrainingforfemaleandmaleoperatorsselectedbythe factories.Ourpurpose for implementingthetrainingwastoinduce factories to promote more females to supervisory positions, to allow us tocomparetheactualperformanceoffemalesandmalesassupervisors.Theprojectwassuccessfulinthisregard,increasingthenumberoffemalesworkingasassistantorfullsupervisors in theparticipating factories.Duringa six- to eight-week trail period, thefemaleandmaletraineeswereassignedlinesrandomly.

Theinitialunderperformanceoffemaletraineesraisesacrucialquestion:Wastheunderperformancecausedbyalackofskills,orbyalackofcooperationonthepartofoperatorsbeingmanagedbythesupervisors?Supervisorsmay lackauthoritybecausetheylessskills,orsimplybecausethosetheysupervisedonotbelievetheypossesstheauthority,andhencetheydonottaketheirinstruction.Thestrongbaselinebeliefsintherelative ineffectiveness of females as supervisors raises the possibility that operatorswouldresponddifferentlytoexactlythesamelevelofsupervisiongivenbymalesandbyfemales.

36

We present two pieces of evidence that suggest the channel from baselineperceptionstothereactionofoperatorsandothermanagersisanimportantone.First,thepost-training skills assessment showednodifferencesby gender in themeasuredskillsandonlyasmallandinsignificantdifferenceintheself-reportedconfidenceofthetrainees. Second, the controlledmanagement simulation showed that female trainees(fromthefirstphaseoftheproject)outperformedmaletrainees,butonlywhenworkingwith all-female teams of operators. Since the worker pairs in the simulation wererandomly assigned, this suggests that the same quality of supervision delivered byfemaleswasmore effectivewith female productionworkers. Asmales expressmuchstrongerpreferences toworkwithmale supervisors, this is consistentwith the ‘priorbeliefs/resistance’story.

Roughlyfourmonthsafterthefirstassessment,wefindthattheperformanceoffemaletraineeshascaughtuptothatofmaletrainees.Thereareseveralreasonstothinkthattheseresultsunderstatetheeffectivenessoffemalesupervisorsrelativetothelong-run steady state. First, we compare the top males with the top females. The propercomparisonwouldbethetopfemalesagainstthemarginalmales,thosewhojustmissedbeingselectedfortraining.Second,aswenoted,managershavelessexperienceselectingfemales as supervisors, and express less confidence in their predictions about theperformanceofthefemalesselected.Moreexperiencemightbeexpectedtoleadtomoreefficientselection.Third,malesenterthefactoryexpectingtobecomesupervisorswithveryhighprobability.Theythereforehaveanincentivetoinvestintheskillsrequiredtobeasupervisor.Females,withlittleprospectforpromotion,lackthatincentive.Womenshouldbeexpectedtoincreasetheinvestmentinthenecessaryskillsiftheirprospectsfor futurepromotionare increased.Finally, thewomenpromotedtosupervisorare insomesensepioneers,orat leastveryearlyentrants, in their factories.Manyqualifiedwomen may not want to go against such a strong current, and so may decline toparticipate.Nevertheless, as the female traineesdonot significantly outperformmaletrainees, the results to not give a definitive answer to the question of whether theindustryis,fromtheperspectiveofefficiency,promotingtoofewwomen.

Whatwecansaywithmorecertaintyisthattheexperimentpointstoasubstantialcost to an individual factory in making the transition from an (essentially) all malesupervisorystafftoamixed-gendersupervisorystaff.Theinitialunder-performanceoffemales,whateverthecause,isapartofthatcost.Butanothersubstantialpartofthecost

37

istheshiftinattitudesofmaleoperatorswhoareexposedtofemalesupervisors.Thoseoperatorsreportlowerprobabilitiesofbeingpromotedandshorterexpectedtenureatthefactory.Trainingmaybeathirdcost.Practicalitiesdidnotallowadesignenablingaseparationiftrainingfrompromotionoffemales.Buttraininghadanimportanteffectontheself-confidenceoffemaletrainees.Presumablyovertime,withmorerolemodelsintheposition,womenwouldstartwithhigherlevelsofself-confidence.Butintheinitialtransition,thetrainingmaybeimportant.

Anecdotally,wehavehadconversationsaboutayearafterthefinalsurveywithnineofthe24factoriesparticipatingintheproject.Amongthosenine,twohavereportedpromoting significant number of additional female supervisors. In both cases, factorymanagementstatedthatan importantmotivation for thechange is thatmaleworkerscausemore troubleandunrest in the factories,andhence theyarenowhiringalmostexclusivelywomen for all positions.Whether justified or not, these beliefs provide areasonforbeingwillingtopaythecostsassociatedwithtransitioningtoanequilibriumwithfemalesupervisors.

38

References

Achyuta, A., N. Kala, and A. Nyshadham (2014). Management and Shocks to Worker

Productivity:EvidencefromAirPollutionExposureinanIndianGarmentFactory.WorkingPaper,UniversityofMichigan.

Atkin, D., A. Chaudhry, S. Chaudry, A. Khandelwal, and E. Verhoogen (2015).Organizational Barriers to Technology Adoption: Evidence from Soccer-BallProducersinPakistan.NBERWorkingPaper21417.

Backus,M. (2014).Why isProductivityCorrelatedwithCompetition?WorkingPaper,ColumbiaBusinessSchool.

Beaman, L., R. Chattopadhyay, E. Duflo, R. Pande, and P. Topalova (2009). PowerfulWomen:Does ExposureReduceBias?Quarterly Journal of Economics 124 (4),1497-1540.

Bernard,A., B. Jensen, S.Redding, andP. Schott (2007). Firms in InternationalTrade.JournalofEconomicPerspectives21(3),105-130.

Bertrand,M.,S.Black,S.Jensen,andA.Lleras-Muney(2014).BreakingtheGlassCeiling?TheEffectofBoardQuotasonFemaleLabourMarketOutcomesinNorway.NHHDept.ofEconomicsDiscussionPaperNo.28/2014.

Bertrand, M. and K. Hallock (2000). The Gender Gap in Top Corporate Jobs. NBERWorkingPapers7931.

Bloom,N.,B.Eifert,D.McKenzie,A.Mahajan,andJ.Roberts(2013).DoesManagementMatter:EvidencefromIndia.QuarterlyJournalofEconomics128(1),1-51.

Bloom,N.,K.Manova,J.V.Reenen,andZ.Yu(2015).Management,ProductQualityandTrade:EvidencefromChina.WorkinProgress,StanfordUniversity.

Bloom, N. and J. V. Reenen (2007). Measuring and ExplainingManagement PracticesAcrossFirmsandCountries.QuarterlyJournalofEconomics122(4),351-1408.

Bloom,N.,R.Sadun,andJ.V.Reenen(2012).TheOrganizationofFirmsacrossCountries.QuarterlyJournalofEconomics127(4),1663-1705.

Bruhn,M.,D.Karlan,andA.Schoar(2012).TheImpactofConsultingServicesonSmallandMediumEnterprises:EvidencefromaRandomizedTrial inMexico.C.E.P.R.DiscussionPaper8887.

Chattopadhyay,R.andEstherDuflo(2004).WomenasPolicyMakers:EvidencefromaRandomizedPolicyExperimentinIndia.Econometrica72(5),1409-1443.

Dezso, C. and D. G. Ross (2012). Does Female Representation in Top Managementimprove FirmPerformance?A PanelData Investigation. StrategicManagementJournal33(9),1072-1089.

Foster, L., J. Haltiwanger, and C. Syverson (2008). Reallocation, Firm Turnover andEfficiency:SelectiononProductivityorProfitability.AmericanEconomicReview98(1).

Gibbons, R. and R. Henderson (2012). Relational Contracts and OrganizationalCapabilities.OrganizationScience23(5),1350-1364.

Glover,D.,A.Pallais,andW.Pariente(2015).DiscriminationasaSelf-FulfillingProphecy:EvidencefromFrenchGroceryStores.WorkingPaper,Harvard.

Gneezy,U.andA.Rustichini(2004).GenderandCompetitionataYoungAge.AmericanEconomicReview94(2),377-381.

Heath,R.andM.Mobarak (2015).Manufacturinggrowthand the livesofBangladeshiwomen.JournalofDevelopmentEconomics115,1-15.

39

Ichniowski, C., K. Shaw, and G. Prennushi (1997). The Effects of Human ResourceManagementPracticesonProductivity:AStudyofSteelFinishingLines.AmericanEconomicReview87(3),291-313.

Matsa,D.andA.Miller(2013).AFemaleStyleinCorporateLeadership?EvidencefromQuotas.AmericanEconomicJournal:AppliedEconomics5(3),136-69.

McKenzie,D.andC.Woodruff(2015).BusinessPracticesinSmallFirmsinDevelopingCountries.NBERWorkingPaper21505.

McKinsey (2011). Bangladesh's ready made garments landscape: The challenge ofgrowth.McKinsey&Company,Apparel,Fashion&LuxuryPractice.

Niederle,M.,C.Segal,andL.Vesterlund(2013).Howcostlyisdiversity?Affirmativeactioninlightofgenderdifferencesincompetitiveness.

Niederle,M.andL.Vesterlund(2007).DoWomenShyAwayFromCompetition?DoMenCompeteTooMuch?QuarterlyJournalofEconomics122(3),1067-1101.

Ruggles, S., T. Alexander, K. Genadek, R. Goeken,M. Schroeder, andM. Sobek (2010).Integratedpublicusemicrodataseries:Version5.0[machine-readabledatabase].Minneapolis,MN:MinnesotaPopulationCenter[produceranddistributor].

Schoar,A.(2011).TheImportanceofBeingNice:EvidencefromaSupervisoryTrainingPrograminCambodia.mimeo,MIT.

Syverson, C. (2004). Product Substitutability and Productivity Dispersion. Review ofEconomicsandStatistics.

40

AppendixA:Description,ProjectPhase1

A.1DifferencesinDesign

ThefirstphaseoftheprojectbeganinNovember2011.Thetrainingprogramwasdesignedwiththegoalofincreasingthenumberoffemalesupervisorsinfactories,andGIZexpressedapreferencethatwetrainonly femaleoperatorsaspartof theproject.Recognizingthevalueofhavingsomecomparisonsampleofmaleoperators,weagreedwithGIZtotrainfourfemalesandonemalefromeachoftheparticipatingfactories.Webegancontactingpotentialfactories,withaletterofintroductionfromalargeUK-basedbuyer, in August 2011. The first training session began in November 2011. After sixroundsoftraining,westeppedbackinJanuary2013toassessthedesign.

Our aimwas to select a sample of factories capable of selling directly to large

international buyers.We obtained an initial sample using data from transaction-leveltrade data obtained from the Bangladeshi National Bureau of Revenue. These dataprovidevolume(netweight)andvalueofexportsattheshipmentlevel.Thedatahaveidentifierswhichallowdatafromindividualexporterstobeaggregated.Weaggregateddatabyexporterandcalculatedtheunitvalue(USDperkilogram)foreachexporter/product/year.Wealsosummedtotalexportsbyexporter.Usingthesetwomeasures,weselectedasampleoffirmswithannualshipmentvolumeslargeenoughtoselldirectlytolarge foreignbuyers,withunit values in the rangeofmid-level buyers.This selectionprocessyieldedaninitialsampleof665exporters.Wethenselectedthegroupofaround20 suppliers to oneparticularmid-range buyer based in theUK. For each of the 665exportersontheinitiallist,wecreatedascorebasedonexportvolumeandunitvaluesindicatinghowclosetheexporterwastothe20suppliersoftheinitialUK-basedbuyer.Weselectedaround400exporters,andsearchedlocaldirectoriesandthe internet forcontactinformation.Thisyieldedasampleof230factories,whichwebegancontactinginAugust2011.

ByNovember2011,aftercontactingabout200ofthefactoriesonthelist,wehadreceivedaninitialcommitmenttoparticipateintheprojectfrom96unitsof85distinctfactories.TableA1showshowthecharacteristicsofthe85factoriesdifferfromtheinitiallist.Thetableshowsbothacomparisoncharacteristicbycharacteristic,andthep-valuesfromaprobitregressionincludingseveralofthecharacteristics.Wefindthatthatthosefactoriesagreeingtoparticipateselltomorebuyers,andselltohigher-endbuyers.The

41

qualityofbuyers ismeasuredby theaverageunitpricepaidbyeachbuyer. Foreachseller,wethenorderedthebuyersbyunitprice,andmeasuredtheunitvaluepaidbythebuyer at the 90th percentile in the ranking. We also find some evidence that theparticipatingfactorieshadhigherratesofrecentgrowthandexportproductstoalargernumberofcountries.

Participatingfactorieswererandomlyplacedintooneofeighttreatmentroundsof 12 factories each. In practicewe allowed factories to defer participation to a laterroundonce,and in theend, several factoriesdecidednot toparticipate.ByDecember2012,whentraininground6began,wehadexhaustedtheinitiallistof96factories.Notethatallofthecomparisonswewillmakewithtraineescontrolforfactoryfixedeffects,sowe view the factory-level attrition issue as mainly one of external, but not internal,validity.Duringthesecondroundoftheprogram,discussionswiththelocalofficeoftheInternationalFinanceCorporationledtoinclusionofsevenfactorieslocatedintheDhakaEPZintheproject.Thesefactorieswereaddedintrainingrounds4and5.

TableA1:Take-upoftheProgram

Signed-Up NotSigned-Up

p-value

p-value

N=85 N=145 Probit

Size(Export,1000Kgs) 830.4 683.8 0.11 0.44

Avg.UnitValue(perKg) 925.9 883.8 0.15 0.01

Growth(Sales2009-10)* 1.89 1.46 0.08 --

NumberofDestinations 10.1 8.3 0.09 0.18

NumberofBuyers 9.75 8.3 0.06 0.02

NumberofProducts 3.01 2.91 0.32 0.31

MainProductinWoven 0.59 0.54 0.26 --

Yearoffirstexport 2006 2006.2 0.2 --

MedianBuyer 560 631 0.18 0.44

90thPercentileBuyer 183.6 283 0.03 0.01

Notes:*Onasampleof80and135exportersrespectively.

TableA2showscharacteristicsof thefactoriesparticipating inrounds1-6.The

factoriesare large,averaging19production linesand2,100workers.Somewhatmore

42

than half of the employees in a typical factory work in the sewing section. Thedistributionsare slightly right-skewed,with themedian factoryhaving15productionlines,with 2,000workers in total, ofwhich 59% are in the sewing section. A typicalfactoryhadbeenoperatingfor12years.Giventherapidgrowthofthesector,thisisverylikelyolderthantheindustryaverage.

TableA2:Description,FactoriesPhase1

MeanMe-dian

Numberofsewinglines 19 14

Numberofemployees,total 2116 2000

Numberofemployees,Sewing 1171 1000

Operatorspersewingline 48 47

Numberofsewingsupervisors 48 36

Percentagefemalesupervisors 10.80% 5.60%

Percentconductingtraining 68.10% NA

Percenttrainingoutsidefactory 8.90% NA

Yearfactoryestablished 1999 2001

A.1.1SelectionofTrainees

Ouraimwastoselectfromeachfactoryfourfemaleandonemaleoperatorfortraining,andavalidcomparisongroupagainstwhichtomeasurethetrainees.Thedetailsofselectingworkersevolvedabitacrosstrainingrounds,aswedescribebelow,butinallroundstheprocessstartedwithfactoriesselectingapoolofpotentialtraineestowhichweadministeredadiagnostictest.ThetestwasbasedononedesignedbyGIZtomeasureliteracy, a requirement for the training, and technical knowledge. We also gave thepotential trainees a short non-verbal reasoning test and asked them questions aboutaspirationstoworkasalinesupervisor.Becausewomenweresometimesforbiddentoparticipateinthetrainingbytheirfamilies,wealsoaskedthepotentialtraineesiftheirfamilieswouldallowandsupport themtoattendthe training.Potential traineeswereexcludediftheydidnotpasstheliteracytestorsaidtheirfamilieswouldnotallowthemtoparticipateinthetraining.

43

Fortrainingrounds1to3,weaskedthefactoriestoidentify16femaleand4maleoperatorswhoweregoodcandidatesforthetraining.Werankedthenomineesaccordingto their diagnostic score and then selected the two females with top marks on thediagnostictestastrainees.Wethenassignedarandomnumbertothefemaletraineesranked3rdto6thonthediagnostictest,andassignedthetwowiththehighestrandomnumberstotraining,andthetwowiththelowestrandomnumberstocontrol.Amongthemales,wefollowedasimilarprocedurebytakingthemaleswiththetoptwomarksandrandomlyassigningonetotreatmentandonetocontrol.

Inround4,wemodifiedtheselectionprocesstoallowthefactorytochoosetwofemalestheywantedtosendtotraining,conditionalonlyonthemdemonstratingabasiclevelofliteracy.Wethentookthetopfourfemalesafterexcludingthetwoselectedbythefactoryandrandomlyselectedtwofortreatmentandtwoforcontrol.Wealsoalteredthemethodof replacing traineeswhen the selected individualsdeclined toparticipate. Inround5,wemodifiedtheprocessfurtherbyreducingthenumberofoperatorsthefactoryidentifiedascandidatestoeightfemalesandfourmales.Thefactorythenselectedtwooftheeight females for training; the remaining two femalesand themalewere selectedrandomlyinthesamemannerasthepreviousrounds.

Wefurthermodifiedthemethodforselecting“replacement"trainees,asdescribedbelow. There was a non-trivial amount of noncompliance. Over the six rounds, 50workersassignedtotrainingdidnotattendatall,andanadditionaleightattendedforless than one fullweek. Factoriesmost often reported that theseworkers either haddecided theydidnotwant to attend, or their familieshad said they couldnot attend.However,thefamilywasmostlikelytointerveneinthecaseoffemaletrainees,whilewenotethatthepercentageofnon-complyingmalesassignedtotraining(21.2percent)washigher than the percentage of non-complying females assigned to training (15.2percent).23Thesenon-complierswerereplacedby40workersreceiving trainingeventhough theywerenotassigned to training including19workersassignedas controls.Thus,non-complianceisaconcerninthePhaseIdatawhenwecomparetheoutcomesofthoseassignedtotreatmentagainstthecontrols.

As with the selection of trainees, the protocol for selecting replacements alsoevolved over the training rounds. In training round 1, the factories chose the

23We interpret this as suggesting that factories cared more about which males received training than they did about which females received training, perhaps because they did not plan to promote all of the females.

44

replacementsthemselves,aswehadnotanticipatedtheseverityofthisnon-compliance.Beginninginround2,weinsistedthatthefactorysendthenextfemaleormaleonthediagnosticrankingifaselectedtraineedeclinedtoattend.Then,beginninginround5,wealtered the initial selection process to add a third female control - selecting 2 of thefemalesranked3to7bydiagnosticscore-andasecondmalecontrol-selectingoneofthemales in the top three diagnostic scores as the trainee. Replacementswere thenselectedatrandomfromamongthecontrols.

Over the first six training rounds, 271 operators (213 females and 58males)receivedtraining.Weexcludefromthistotaleightworkerswhoattendedforfivedaysorfewer.Conditionalonattendingatleastoneweek,attendancewasveryhigh.Outofthe36trainingdays,malesattended34.4daysonaverageandfemales34.5days.Allbuttwoofthemenattendedatleastfourofthesixtrainingweeks,asdid96percentofthewomen.Afterthesixthtraininground,wedecidedtosuspendthetrainingtemporarily.Havingalreadygatheredasubstantialamountofdataandinformation,wefeltwewouldgainbyanalysingthosedataandperhapsadjustingthedesignfortheremainingfactories.WeresumedthetrainingwiththestartofPhase2oftheprojectinFebruary2014,which'sdetailsaredescribedinthemainbodyofthepaper.A.2Comparisonofsamplesandoutcomes

Wereportfirst-phaseresultsfromthemanagementsimulationandthepromotionprospectsofoperatorsexposedtofemaleandmaletraineesinthemainbodyofthepaper.We did not conduct the management simulation in the second phase of the projectbecauseitwasseenascostlytoparticipatingfactoriesgiventheamountoftimerequired.Wedidnotaskaboutpromotionprospectsinthesecondphaseintheinterestoftimeandbecause we had not analysed those results when we designed the second phase.However,thereareseveralquestionswhichwererepeatedinbothphasesoftheproject,and we report here on comparisons of outcomes in the two phases where we havecomparabledataineachphase. TableA3 shows the demographic characteristics of the trainees and randomlyselectedpoolsofoperatorsandexistingsupervisors.Asinthephase2sample,wefindthatthetraineeshavemoreeducationandexperiencethanatypicaloperator,butlowereducationandexperiencethantypicalsupervisors.Thetabledoesnotseparatethemaleandfemaletrainees,butinthefirstphasewefindthatthemaletraineeshavejustover

45

halfayearofadditionalschoolingandare1.7yearsolder.Theschoolingdifferencesmayreflectthefactthatweselectedfourfemalesandonlyonemaleineachfactory; inthesecondphase,wherewehaveequalnumbersoffemalesandmales,weseenodifferenceinschoolinglevels.TableA3:Phase1DemographicCharacteristics

Mean Comparisons

Trainee

Pool

Opera-

tors

Supervi-

sors

Traineevs.

Operators

Traineevs.

Supervisors

N=539 N=301 N=292

Gender(female=1) 0.75 0.76 0.12 -0.02 0.63***

Age 23.54 23.53 28.21 0.01 -4.67***

Migrant 0.71 0.75 0.86 -0.04 -0.15***

Married 0.59 0.64 0.75 -0.05* -0.16***

NumberofChildren 0.46 0.60 N/A -0.14*** N/A

Education(=yearsinschool) 8.29 6.46 9.48 1.83*** -1.18***

RavenScore 3.84 2.02 2.74 1.83*** 1.10***

ExperienceinGarments 3.47 2.82 4.64 0.65*** -1.18***

TenureinFactory 3.10 2.52 4.15 0.57*** -1.06***

EverReceivedTraining 0.15 0.08 0.15 0.07*** 0.00

ReceivedTraininginFactory 0.06 0.12 0.13 -0.058*** -0.013

TableA4showsthat,similartophase2,thefemaletraineesweremorelikelytohaveleftthefactorybythefirstfollow-upsurvey,fourmonthsaftertraining.Womenarealsolesslikelytohavebeentriedoutandlesslikelytobeworkingasasupervisoratthetimeofthefollow-upsurvey.Thegendergapsaresimilartothoseinthesecondphase.Finally,TableA5presentstheresultsfromtheproductiondatafromthefirstphase.Therearenosignificantdifferencesbetweenfemalesandmalesinefficiency,qualityorabsenteeism,withtheANCOVAresultsshowingapositiveeffectonefficiencyforfemaletraineeswhichiseconomicallyimportantbutstatisticallyinsignificant,asinthesecondphase.24TableA4:Outcomes4MonthsafterTraining

[I] [II] [III] [V]

24Wehavenotyetcombinedtheproductiondatafromthetwophasesoftheproject,bothbecausetheyare

ofdifferentfrequencyandbecausewehaveonlyrecentlyorganizedtheproductiondatafromphase2.Butwe

plantodothisinafutureversionofthepaper,alongwiththeconstructionofamatchedsampleof

comparisonlines.

46

Dep.Variable:Workingaf-

ter4months

OLS:attritionatFU1

OLS:Infac-toryatFU1

OLS:TriedoutbyFU1

OLS:work-ingasSVat

FU1

Training.008 014*** 0.560*** 0.592***

(0.031) (0.03) (0.086) (0.083)

Female0.064** -0.160** -0.175**

(0.026) (0.068) (0.073)

TrainingXFemale-0.084* -0.0002 -0.029

(0.038) (0.090) (0.093)

Dep.Variable,malecontrols 0.1 0.87 0.43 0.25

Dep.Variable,femalecon-

trols0.18 0.8 0.168 0.114

DemographicControls No No No No

FactoryFixedEffects yes yes yes yes

Sample All All All All

N.Observations 523 523 454 360

Notes:AttritionmeasurebyhavinginformationoneverbeingtriedasanSV.Infac-

toryissettozerowhenwehavenoinformationontheworkerafterbaseline.Where

interactioneffectsarenowshown,theyarehighlyinsignificant.Controllingforde-

mographiccharacteristicsoftheoperatorsmakesverylittledifference,butreduces

samplesize.TheIVregressionsareonthetreatment/controlsample.

47

AppendixB:ProductionData,DescriptionandCollection

AspartofbothPhase1and2oftheproject,wecollecteddailyproductiondatafrom all factories on the sewing line level. The data is similar in its format andorganization across the two project rounds. However, in Phase 1 of the project wecollecteddatainatwoweekintervaleveryothermonth,whileinPhase2wecollecteddata for each day between January 2014 and March 2015. Given the continuity andgreateramountofdata,webasetheanalysisinthemainpartofthepaperonthedatafromPhase2,whichwedescribeinmoredetailinthisappendix.

We collected the data with three main outcome variables in mind: line-levelproductivity,thequalitydefectrate,andworkerabsenteeism.Weaskedfactoriestoshareall internal data needed to construct these variables. The standard measure ofproductivityintheBangladeshigarmentindustryis(piecewiseoutput*SMV)/(workers*dailyhours*60mins),whereSMVis theStandardMinuteValueof thegarmentbeingproduced.TheSMVisthetimeindustrialengineersestimateafullyefficientproductionline would take to produce one unit of the garment. When estimated to a commonstandard, theSMV thusallowsus to compare theefficiencyofproductionofdifferentproducts-e.g.,theefficiencyofalineproducingatanktopwithanSMVofsixminutes

48

canbecomparedwiththeefficiencyofalineproducingadressshirtwithanSMVof18minutes.

Weaskedthefactoriestoprovideproductivityrecordsforeachsewinglineanddaydetailingondailyoutput,thenumberofdefectiveunits,theSMVofgarmentbeingproduced,thenumberofhourseachlineoperated,anddailynumberofworkerspresentandabsentontheline.Notallfactoriesrecordinformationonallofthevariables.Insomeinstances, the factoriesrecorddata,butdeclined toprovide it forcertainoutputs.Forexample, one factory declined to provide SMV data, and a few others do not haveindustrial engineeringdepartments, andhencedonot estimate SMVsbyproduct. Forothervariables,therearesometimesdifferencesinthespecificdatathefactoriesrecord,though often these differences are not consequential. For example, for defects, wesometimesreceiveddefectrates(defectiveunits/output)andsometimesthenumberofdefectivegarments.Recordsonabsenteeismwouldsometimescontaininformationonthenumbersofworkersassignedto the line,allowingtostandardize theabsenteeismnumbers.Atfactorieswherethisinformationwasnotincluded,weinsteadstandardizedthe number of absent workers by the number of present workers provided in theproductivitydata.

In almost all factories, the three types of data (on productivity, defects, andabsenteeism)wasprovidedbydifferentdepartmentswithin the factories (usually theproduction, quality, andHR departments), and thus came in different formats,whichrequiredtoenterthedataseparatelyandsubsequentlymergethemtoonedocument.Likewise, in most factories, the data we requested was provided in a digital format,usuallyaspreadsheetmaintainedbythefactories,whichallowedforeasycollectionandentering.Atsomefactorieshowever,datawasprovidedascopiesofpaperfiles,requiringthedatabedigitised.Ultimately, though,weharmonise thedata so that variables arecomparableacrossfactories.

As we noted, the data from some factories did not contain the informationnecessary to calculate all of the outcomes of interest. This is especially the case forefficiency,whereourstandardcalculationreliesontheavailabilityoftheSMVdata.SomeofthefactoriesthatdonotmeasureSMVhaveotherdatawhichcanbeusedtoestimatea roughlyequivalentmeasureofefficiency.Forexample, four factories in thePhase2samplehave informationondaily targets for their sewing lines.Byassuming that the

49

targetsaresetsuchthatlineefficiencywouldbe100%,wecanbackouta‘syntheticSMV'bysettingDailyTarget*SMV=workers*dailyhours*60mins.25

From17ofthe19factoriesremainingintheprojectthroughout,wewereabletocollect data for at least one of our three outcome variables of interest; productivity,defects,andabsenteeism.TableB1showsfromhowmanyofthese17factorieswecouldcollectenoughdatatoconstructeachofthethreevariables,andforhowmanywecouldconstructallthree.Whiletheavailabilityofdefectsdataismostwidespread,productivitydataisreducedbyanumberoffactoriesrecordingneitherSMVsnortargets.Finally,theavailability of absenteeism data for our analysis is limited by a number of factoriesrecording only daily absenteeism numbers for the whole factory (or sometimes thesewingfloor),butnotrecordingdataonthesewinglinesonwhichworkersareassigned.

TableB1:ProductionDataAvailability

Outcomevariable #FactoriesProductivity 14

Productivity,excludingsyntheticSMV 10

Defectsrate 16

Absenteeism 10

Productivity+Defects+Absenteeism 7

Productivity(excl.synth.SMV)+Defects+Absentee-

ism 4

TotalNumberFactorieswithsomeProd.Data 17

Notes:TableshowsforhowmanyfactoriesparticipatinginPhase2,usable

dailydataonline-wiseproductivity,defectsrates,andabsenteeismcouldbe

collected.

25Factories from which both SMV and targets are available show that targets are usually not set such that efficiency, in case the target is met, is 100%. Rather efficiency in these cases would be around 50%, which is in line with the typical average efficiency in almost all Bangladeshi garment factories. Thus, the ‘synthetic SMVs' which we back out using targets are likely to overstate actual SMVs by a factor of two. And indeed, efficiency values at those factories where we use ‘synthetic SMVs' are on average twice as high as in the other factories (93% vs 47%). However, note that all analysis we conduct with the production data uses factory fixed effects, therefore relying only on within factory variation in productivity. Given that for each factory we use either only productivity based on original or synthetic SMVs, the productivity data is consistent within each factory.

50

Figure1:FemaleWorkerandManagerinGarmentIndustry,U.S.vs.Bangladesh

Notes:FigureshowsthehistoricalevolutionoftheshareoffemaleworkersandmanagersintheUSgarmentindustry,andcomparesitagainstthecurrentsharesintheBangladeshiindustry.USDatafromRugglesetal.[2010],Bangladeshidataowncalculations.

Figure2:Selection,Training,TrialandPromotionofTrainees

Notes:Figureshowsthenumberoffemale(redbars)andmale(bluebars)traineeswhichparticipatedordroppedoutindifferentstagesoftheproject,andwhichgotpromotedtosupervisorlevelsintheirfactories.

51

Figure3:TasksofSupervisors,AttachedImportance

Notes:Workersonvariouslevelsin26factorieswereaskedtoplace10tokensonalistofeightgeneraltasksoflinesupervisors(generatedafteropenendedconversationswithseveralfactorymanagers),accordingtotherelativeimportancetheyattachtothetask.

Figure4:TasksofSupervisors:PerceivedAbilitybyGender

Notes:Workersonvariouslevelsin26factorieswereaskedforeachoftheeightmainsupervisortasks, whether they perceive female or male supervisor as more capable. Answers wereaggregatedonthetaskanddesignationofrespondentlevel,withanswersbeingcodedas-1for“malesaremorecapable",0for“bothareequallycapable",and1for“femalesaremorecapable".

52

Figure5:TasksofSupervisors:PerceivedAbilitybyGender,Extended

Notes:Workersonvariouslevels,ofdifferentgender,andwithvaryingexperienceofworkingunderfemalesupervisorsin26factorieswereaskedforeachoftheeightmainsupervisortasks,whethertheyperceivefemaleormalesupervisorasmorecapable.Answerswereaggregatedonthetaskandgroupofrespondentlevel,withanswersbeingcodedas-1for“malesaremorecapable",0for“bothareequallycapable",and1for“femalesaremorecapable".

53

Table1:DemographicCharacteristicsPanelA:Traineesvs.OperatorsandSupervisors

Mean Comparisons TraineePool Operators RandomSVs MentorSVs Traineevs.

OperatorsTraineevs.RandomSVs N=199 N=430 N=92 N=142

Gender(female=1) 0.50 0.73 0.04 0.04 -0.23*** 0.46***Age(current) 24.4 24.1 29.1 29.3 0.30 -4.66***AgeattimeofpromotiontoSV 24.4 NA 25.3 24.3 NA 0.84Married 0.71 0.77 0.85 0.89 0.06 -0.14**Education(=yearsinschool) 7.83 5.80 8.77 9.56 2.03*** -0.95***ExperienceinGarments(years) 6.53 6.10 8.83 9.21 0.43 -2.30***TenureinFactory(years) 3.41 2.78 3.30 3.80 0.63*** 0.11*** PanelB:Traineesvs.MentorSupervisors

Mean TraineePool MentorSVs Traineevs.

MentorSVs

N=197 N=113 Literacy 7.14 9.54 -2.39*** Numeracy 3.4 5.2 -1.72*** Non-verbalreasoning 2.75 2.78 -0.03

Notes:Thetablereportsmeancharacteristicsoftraineeandrandomsewingoperators,randomlinesupervisors,andthesuper-visorsselectedasmentorsupervisors,allselectedfromthelineswherethetraineesweretobeassignedtoworkasassistantsupervisorsfollowingtraining.Statisticaldifferencesincomparisons:***p<.01;**p<.05;*p<.10

54

Table2:AttitudestowardfemaleSVs:Baselinedata (1) (2) (3) (4) Femalesbet-

terthanmales,8tasks

Preferfe-maleSVtomaleSV

Preferfe-maleSVtomaleSV

Operatorisfemale 0.122*** 0.142*** 0.175*** 0.208*** (0.017) (0.022) (0.043) (0.046)ExperienceworkingforfemaleSV

0.011 0.122*** (0.017) (0.039)

Experience*femaleop-erator

-0.002 0.099* (0.019) (0.051)

Experience*maleoper-ator

0.047 0.182*** (0.028) (0.057)

Obs 428 428 426 426R-squared 0.19 0.19 0.18 0.18FactoryFE YES YES YES YESMean 0.25 0.25 0.32 0.32

Notes:Coefficientsshowtheeffectofeachvariableontheprobabilityofsayingfe-malesaremoreskilledorpreferredassupervisors.Standarderrorsclusteredattheproductionlinelevel;regressionsincludeage,educationandmaritalstatusoftherespondent.Statisticaldifferences:***p<.01;**p<.05;*p<.10

55

Table3:TrainingandTrialEffectsbyGender (1) (2) (3) (4) (5) (6) Percent-

agedraw-ingscor-rect

Drawing,"soft"score

Com-munica-

tion,"soft"score

Leader-ship,"soft"

score

Aptitudescore

Selfconfi-denceVARIABLES

Femaletrainee,baseline -1.19 -0.09 0.22 -0.27 -0.60 -0.47* (0.90) (0.06) (0.56) (0.72) (0.53) (0.24)Femaletrainee,aftertraining -0.69 0.13** 0.22 -0.70 -0.80 0.24 (0.86) (0.06) (0.60) (0.81) (0.67) (0.25)Maletrainee,aftertraining 0.20 0.18*** 0.10 0.21 0.09 0.50** (0.80) (0.06) (0.64) (0.79) (0.79) (0.25)Femaletrainee,afterfactorytrial -4.93*** 0.09 -0.43 -0.01 -1.09 0.32 (1.82) (0.08) (0.58) (0.85) (0.69) (0.24)Maletrainee,afterfactorytrial -0.37 0.05 0.24 -0.75 -0.07 0.58** (1.25) (0.07) (0.69) (0.96) (0.76) (0.24) Baselinemean,maletrainees 55.4 0.33 -0.15 0.16 0.35 -0.33Observations 470 420 421 423 329 470R-squared 0.191 0.125 0.098 0.068 0.125 0.133FactoryFE YES YES YES YES YES YESFemalevs.Male,aftertraining(p) 0.28 0.50 0.86 0.30 0.30 0.32Femalevs.Male,aftertrial(p) 0.02 0.66 0.32 0.48 0.20 0.30

Notes:Unbalancedpaneloftraineesmeasuredonfirstdayoftraining,lastdayoftraining,andattheendofthefac-torytrialperiod.Statisticaldifferences:***p<.01;**p<.05;*p<.10.

56

57

58

59

Table6:ManagementSimulationExercises TraineesFemalesVs.Males (1) (2) (3) (4) (5) (6) (7) Outcome:StandardizedPay-OffinGames

CorrectStrategy

Outcome:Pay-OffinGames(Standardized) OLS OLS TeamFE OLS OLS OLS

FemaleTeamLeader0.290*** 0.255** 0.420** 0.305** 0.332*** 0.412*** -0.371**(0.109) (0.122) (0.190) (0.131) (0.124) (0.521) (0.167)

Mixedgender/maleteam -0.092 (0.725) Mixed*FemaleTL -0.433* (0.240)

TriedasLineSupervisor 0.329 (0.212)

PromotedtoLineSupervisor 0.508** (0.206)

TeamFixedEffects no no yes no no no noGameFixedEffects yes yes yes yes yes yes yesTeamLeaderDemogr. no yes yes yes yes yes yesNumberofObservations 676 612 612 612 608 612 612

Notes:Thedependenetvariableisthestandardizedpayofffromthegame.Standarderrorsareclusteredatthegamelevel.***p<0.01,**p<0.05,*p<0.1.

Challenges of Change:

Training Women to Manage

in the Bangladeshi Garment Sector

Rocco Macchiavello Andreas Menzel

LSE CERGE-EI

Atonu Rabbani Chris Woodru↵

University of Dhaka Oxford

September 21, 2018

Andreas Menzel Challenges of Change September 21, 2018 1 / 10

Introduction

Women under-represented in managerial positions throughout the world

Duflo (2000), Bertrand and Duflo (2015), Flabbi et al. (2016)

Manager practices and identities key driver of productivity

Bloom et al. (2012), Bandiera et al. (2018)

Raises at least three questions:

1What are e↵ects of promoting women to supervisor position?

Bertrand et al. (2017)

2Do these barriers lead to a misallocation of managerial talent?

3What are the barriers to female access to managerial positions?

Bertrand (2018)

Andreas Menzel Challenges of Change September 21, 2018 2 / 10

Context

RCT in 25 garment factories in Bangladesh

⇠ 3 Milion sewing workers in sector

⇠ 80% of workers female, but only ⇠ 5% of supervisors

Andreas Menzel Challenges of Change September 21, 2018 3 / 10

Introduction

Methodological challenges:

1. Identify marginal supervisors (Becker, 1957)

Managers identify equal numbers of male and female workers “next in line” for

promotion

F Incentivised (free 6 weeks training course)F Previous experiment showed ⇠ 60% female promoted

2. Sorting : e.g., allocation to di↵erent tasks/team

Trainees work as supervisors on randomly allocated production lines

F 2-3 months trial periodF Factory decides post-trial promotion and allocation

3. Subordinate/co-workers attitudes (supervisor ‘authority’)

Extensive surveys

F Baseline/Midline/Endline of subordinates and co-supervisors.

Andreas Menzel Challenges of Change September 21, 2018 4 / 10

Introduction

Methodological challenges:

1. Identify marginal supervisors (Becker, 1957)

Managers identify equal numbers of male and female workers “next in line” for

promotion

F Incentivised (free 6 weeks training course)F Previous experiment showed ⇠ 60% female promoted

2. Sorting : e.g., allocation to di↵erent tasks/team

Trainees work as supervisors on randomly allocated production lines

F 2-3 months trial periodF Factory decides post-trial promotion and allocation

3. Subordinate/co-workers attitudes (supervisor ‘authority’)

Extensive surveys

F Baseline/Midline/Endline of subordinates and co-supervisors.

Andreas Menzel Challenges of Change September 21, 2018 4 / 10

Introduction

Methodological challenges:

1. Identify marginal supervisors (Becker, 1957)

Managers identify equal numbers of male and female workers “next in line” for

promotion

F Incentivised (free 6 weeks training course)F Previous experiment showed ⇠ 60% female promoted

2. Sorting : e.g., allocation to di↵erent tasks/team

Trainees work as supervisors on randomly allocated production lines

F 2-3 months trial periodF Factory decides post-trial promotion and allocation

3. Subordinate/co-workers attitudes (supervisor ‘authority’)

Extensive surveys

F Baseline/Midline/Endline of subordinates and co-supervisors.

Andreas Menzel Challenges of Change September 21, 2018 4 / 10

Introduction

Methodological challenges:

1. Identify marginal supervisors (Becker, 1957)

Managers identify equal numbers of male and female workers “next in line” for

promotion

F Incentivised (free 6 weeks training course)F Previous experiment showed ⇠ 60% female promoted

2. Sorting : e.g., allocation to di↵erent tasks/team

Trainees work as supervisors on randomly allocated production lines

F 2-3 months trial periodF Factory decides post-trial promotion and allocation

3. Subordinate/co-workers attitudes (supervisor ‘authority’)

Extensive surveys

F Baseline/Midline/Endline of subordinates and co-supervisors.

Andreas Menzel Challenges of Change September 21, 2018 4 / 10

Introduction

Methodological challenges:

1. Identify marginal supervisors (Becker, 1957)

Managers identify equal numbers of male and female workers “next in line” for

promotion

F Incentivised (free 6 weeks training course)F Previous experiment showed ⇠ 60% female promoted

2. Sorting : e.g., allocation to di↵erent tasks/team

Trainees work as supervisors on randomly allocated production lines

F 2-3 months trial periodF Factory decides post-trial promotion and allocation

3. Subordinate/co-workers attitudes (supervisor ‘authority’)

Extensive surveys

F Baseline/Midline/Endline of subordinates and co-supervisors.

Andreas Menzel Challenges of Change September 21, 2018 4 / 10

Introduction

Methodological challenges:

1. Identify marginal supervisors (Becker, 1957)

Managers identify equal numbers of male and female workers “next in line” for

promotion

F Incentivised (free 6 weeks training course)F Previous experiment showed ⇠ 60% female promoted

2. Sorting : e.g., allocation to di↵erent tasks/team

Trainees work as supervisors on randomly allocated production lines

F 2-3 months trial periodF Factory decides post-trial promotion and allocation

3. Subordinate/co-workers attitudes (supervisor ‘authority’)

Extensive surveys

F Baseline/Midline/Endline of subordinates and co-supervisors.

Andreas Menzel Challenges of Change September 21, 2018 4 / 10

Trainee Gender Comparison: 72 Male, 73 Female

Mean Males Female SE N

Age 24.7 -1.00 (0.70) 145Married 0.61 0.14* (0.08) 145Years Educ. 8.48 -0.42 (0.29) 143Years in Garment Sec. 6.44 -0.24 (0.55) 145Years in Factory 3.65 -0.44 (0.45) 145Nbr Factories 2.06 -0.06 (0.33) 145Literacy 8.84 -0.99 (0.86) 143Numeracy 4.75 -1.37*** (0.39) 143Non-verbal Reason. 3.11 -0.10 (0.37) 143Technical Knowledge 55.8 -1.5 (1.21) 143Drawing 0.37 -0.11* (0.07) 126Drawing - Soft -0.56 1.09* (0.64) 127Communic. - Soft 0.14 -0.26 (0.88) 127Leadersh. - Soft 0.33 -0.66 (0.66) 124Confidence, Baseline -0.14 -0.81*** (0.29) 143Belief Best, Baseline 0.65 -0.22 *** (0.08) 145

Confidence, after Training 0.20 -0.25 (0.28) 144Belief Best, after Training 0.79 -0.13 * (0.07) 145

Compare to Avg. Superv. and Worker Line Balance

Andreas Menzel Challenges of Change September 21, 2018 5 / 10

Individual Trainee Outcomes:

Trial Trial Trial SV Post-TrialAny Line Correct Correct Any Line

Gender Line

Female -0.001 0.066 -0.009 -0.134*(0.04) (0.07) (0.07) (0.07)

Mean Males 0.833 0.597 0.542 0.667

N 145 145 145 145Factory FE YES YES YES YES

Female 0.010 0.079 0.043 -0.092(0.05) (0.08) (0.08) (0.08)

N 141 141 141 141Controls YES YES YES YESFactory FE YES YES YES YES

Andreas Menzel Challenges of Change September 21, 2018 6 / 10

Line Productivity Response

ITT ITT ITTBasel. Trial Post-Trial Basel. Trial Post-Trial Trial

VARIABLES SV at FU2

PDS-Lasso:Female Trainee 1.172 -5.279* 0.312 -2.741 -13.56*** -0.803 -6.958**

(1.88) (2.78) (1.98) (1.94) (2.78) (2.11) (2.71)

Mean Males 57.04 46.91 60.98 60.06 54.11 51.65 53.92

OLS, No Contrl.Female Trainee 1.785 -6.551** 0.312 0.071 -13.47*** -0.891 -8.414**

N 87 91 97 79 93 124 55Factory FE YES YES YES YES YES YES YES

Defects Rate Response Absenteeism Response

Andreas Menzel Challenges of Change September 21, 2018 7 / 10

Line Worker Response - Grading (1-10) of New SV.

Trial - ITT Trial Post-Trial

PDS-Lasso:Female Trainee -0.674*** -0.824*** -0.249

(0.212) (0.195) (0.231)

Mean, Males 6.951 7.315 7.887

OLS, No contr.Female Traine -0.695 -0.979 -0.129p-val. 0.007 0.00 0.623p-val., OLS-PDS 0.951 0.600 0.732OLS All -0.655 -0.718 -0.147p all 0.0117 0.00134 0.587

Nbr Male Tr.s 48 60 47Nbr Female Tr.s 55 62 36N 264 325 233Factory FE YES YES YES

Supervisor Response Similar, but less Stat. Significant

Andreas Menzel Challenges of Change September 21, 2018 8 / 10

What Could Explain Initial Under-performance?

1Trainee observables: Confidence, Numeracy...: Some Evidence

Link

Reduces negative ITT productivity response and worker grading

2Negative Attitudes against Female SVs ?

Employees universally see female supervisors as less capable at baseline.

Link

1.a Co-Supervisor baseline attitudes: No direct evidence

Link

1.b Worker baseline attitudes: Suggestive evidence

Link

Presence of male workers hurt female productivity

F Male Workers have more negative baseline attitudes to female supervisors

F Grade Female Trainees worse (Di↵. meaningful but not st.sign)

F Management simulation:

Andreas Menzel Challenges of Change September 21, 2018 9 / 10

Conclusion

1RCT comparing new male and female line superv. allocated to random lines.

IFemale new superv. initially perform less well than male counterparts

ITransition cost for factories when promoting more women.

2Female superv. that continue post-trial catch up quickly with male peers

IGraded not worse by peer supervisor during Trial

Link

3Exposure to Female Supervisors improves Worker Preference for Female SVs

IEspecially by male workers: Not under-performing as bad as expected?

IMirrors political leadership results: Beaman et al. (2010, 2012), Miller (2014)

Link

Andreas Menzel Challenges of Change September 21, 2018 10 / 10

Thanks very much!

Andreas Menzel Challenges of Change September 21, 2018 1 / 19

Line Supervisor Response: Preference Female SV.

Trial - ITT Trial Post-Trial

PDS-Lasso:Female Trainee 0.005 0.013 -0.014

(0.048) (0.042) (0.057)

OLS, No contr.Femle Trainee 0.008 0.013 -0.014p 0.878 0.775 0.821

Nbr Male Tr.s 43 51 37Nbr Female Tr.s 52 57 29N 150 173 85Factory FE YES YES YES

Andreas Menzel Challenges of Change September 21, 2018 2 / 19

Demographics Trainees, Supervisors & Workers

Opr. Opr. Tr. Tr. SVMale Fem Male Fem Tr vs.LO Tr vs.LO SV vs. Male vs.

N 115 313 72 73 92 (Male) (Female) Male Tr Fem. Tr

Age 24.3 23.9 24.7 23.7 29.0 .396 -.23 -4.3*** -1.0Married .655 .808 .611 .753 .847 -.04 -.05 -.23*** .142*Years Educ. 6.09 5.69 8.48 8.05 8.77 2.39*** 2.35 *** -.28 -.42Years in Sector 6.63 5.90 6.44 6.20 8.83 -.19 .294 -2.3*** -.24Years in Factory 2.38 2.92 3.65 3.20 3.30 1.27*** .280 .351 -.44Nbr. Factories 2.41 1.65 2.06 2 2.77 -.34 .347 -.70** -.06Age Promoted 24.7 23.7 25.2 -.54 -1.0

Back

Andreas Menzel Challenges of Change September 21, 2018 3 / 19

Line BalanceFemale S E Mean Males N N Lines

Line Operators: Gender -.017 ( .0406 ) .7330 416 141Age -.288 ( .4000 ) 24.04 414 141Married .0511 ( .0418 ) 1.762 416 141Years Education -.195 ( .2645 ) 6.660 416 141Months Factory 4.556 ( 2.983 ) 30.65 415 141Months Industry 3.547 ( 3.902 ) 70.80 415 141Months Designation 2.051 ( 3.640 ) 57.77 416 141Months Line .2211 ( 2.236 ) 16.15 414 141Nbr Factories -.105 ( .1461 ) 1.893 416 141Prev Fem SV -.013 ( .0508 ) .4466 416 141

F-Test: p-val 0.541

Line Supervisors: Gender .0208 ( .0262 ) .0360 226 136Age .6863 ( .5858 ) 28.63 225 136Married .0522 ( .0445 ) 1.846 226 136Years Education .9936 ( 2.313 ) 7.117 226 136Months Factory 4.886 ( 4.871 ) 39.89 226 136Months Industry 6.676 ( 5.202 ) 103.7 226 136Months Designation 3.342 ( 4.111 ) 51.32 226 136Months Line 2.466 ( 3.171 ) 20.90 226 136Nbr Factories .3997 ( .2822 ) 2.645 225 136Prev Fem SV -.073 ( .0589 ) .3243 226 136Spouse Works .0345 ( .0699 ) .3655 196 128

F-Test: p-val 0.448

Andreas Menzel Challenges of Change September 21, 2018 4 / 19

Line Balance cont.

Female S E Mean Males N N Lines

Production Data available:

SMV .0128 ( .0202 ) .1911 144 144Order Quantity .0267 ( .0252 ) .1506 144 144Running Days .0318 ( .0209 ) .1609 144 144Hourly Target .0303 ( .0235 ) .2235 144 144Total Target .0284 ( .0238 ) .2428 144 144Daily Hours .0356 ( .0264 ) .2611 144 144Nbr Operators .0342 * ( .0191 ) .3129 144 144Nbr Helpers .0252 ( .0181 ) .3096 144 144E�ciency .0139 ( .0217 ) .2159 144 144Absenteeism .0232 ( .0157 ) .1733 144 144Defects Rate -.020 ( .0224 ) .2853 144 144Spot Rate -.014 ( .0160 ) .1470 144 144Reject Rate .0150 ( .0188 ) .1811 144 144F-Test: p-val 0.036

Back

Andreas Menzel Challenges of Change September 21, 2018 5 / 19

Line Defects Rate Response

ITT ITT ITT OLS OLS OLS OLSBasel. Trial Post-Trial Basel. Trial Post-Trial Trial

VARIABLES SV at FU2

Female Trainee 0.101 0.813 0.154 0.119 0.774 0.682 0.094(0.533) (0.581) (0.392) (0.659) (0.671) (0.530) (0.796)

Mean Males 4.583 5.759 5.938 4.994 5.628 6.831 7.294

N 104 94 96 87 103 129 63Factory FE YES YES YES YES YES YES YES

OLS 0.813 0.813 0.154 0.119 0.774 0.694 0.200p-val: 0.206 0.206 0.718 1.131 0.292 0.214 0.820p OLS-PDS 0.390 1.00 1.00 1.00 1.00 0.988 0.928

Back

Andreas Menzel Challenges of Change September 21, 2018 6 / 19

Line Worker Absenteeism Response

ITT ITT ITT OLS OLS OLS OLSBasel. Trial Post-Trial Basel. Trial Post-Trial Trial

VARIABLES SV at FU2

Female Trainee 0.079 -0.883 -0.657 -0.194 -0.277 -0.762 3.346**(0.733) (0.740) (0.478) (0.588) (0.970) (0.897) (1.459)

Mean Males 5.290 8.456 8.378 51.03 6.900 6.857 4.886

N 52 59 59 35 50 69 24Factory FE YES YES YES YES YES YES YES

OLS 0.0788 -0.883 -0.657 -0.672 -0.547 -0.787 0.270p-val: 0.923 0.288 0.221 0.339 0.607 0.425 1.127p OLS-PDS 1.00 1.00 1.00 0.600 0.851 0.985 0.170

Back

Andreas Menzel Challenges of Change September 21, 2018 7 / 19

Does Performance Predict Survival as SV?

SV SV SV O�cial O�cial O�cialAny Line Any Line Any Line SV SV SV

Female -0.109 -0.110 -0.098 -0.104 -0.098 -0.097(0.09) (0.09) (0.10) (0.11) (0.11) (0.11)

PCA Eval.s 0.007 -0.028(0.04) (0.04)

Avg. Ranking 0.015 0.024(0.03) (0.03)

Avg. Gender Pref. -0.043 -0.109(0.19) (0.22)

Avg. Prod. Trial -0.051 -0.690(0.38) (0.44)

Avg. Prod. missing -0.012 -0.465*(0.22) (0.25)

N 119 119 119 119 119 119

Same non-e↵ect when using Co-Supervisor Evaluations,

or when interacting predictors with the Female Trainee dummy.

Back

Andreas Menzel Challenges of Change September 21, 2018 8 / 19

Line Superv. Response - Grading (1-10) of New SV.

Trial - ITT Trial Post-Trial

PDS-Lasso:Female Trainee -0.075 -0.352 -0.329

(0.268) (0.230) (0.309)

Mean Males 6.279 6.561 7.227

OLS, No controls:Female Trainee -0.155 -0.553** -0.329

Nbr Male Tr.s 43 51 37Nbr Female Tr.s 51 57 29N 149 173 86Factory FE YES YES YES

Back

Andreas Menzel Challenges of Change September 21, 2018 9 / 19

Why Initial Underperformance: Biased Attitudes?

By Operator Gender & Prev. Exposure to Female SVs

Andreas Menzel Challenges of Change September 21, 2018 10 / 19

Initial Beliefs on SV Ability: Male vs. Female SVs

Back

Andreas Menzel Challenges of Change September 21, 2018 11 / 19

Baseline Attitudes to Female Line Superv:Male vs. Female Worker

Overall initial evaluation (stand.) SV Gender Preference (stand.)

(1) (2) (3) (4) (5) (6)

Male -0.602*** -0.678*** -0.733*** -0.617*** -0.579*** -0.794***(0.0851) (0.0929) (0.174) (0.0852) (0.0924) (0.159)

Constant 0.205*** -0.0295 1.050 0.210*** -0.0163 2.032(0.0487) (0.261) (1.408) (0.0487) (0.259) (1.287)

N 627 612 194 624 610 194R-squared 0.077 0.137 0.317 0.035 0.101 0.289Factory FE YES YES YES YES YES YESDemogr.Controls NO YES YES NO YES YESLocus & Ability NO NO YES NO NO YES

Back

Andreas Menzel Challenges of Change September 21, 2018 12 / 19

Line Worker Response - Grading (1-10) of New SV.s

Trial - ITT Trial Post-Trial

PDS-Lasso:Fem. Tr.ee x Fem. Operator -0.584** -0.762*** -0.349

(0.269) (0.242) (0.256)Fem. Tr.ee x Male Operator -0.931** -0.989*** 0.012

(0.431) (0.373) (0.477)Female Respondent -0.069 0.009 0.384

(0.463) (0.354) (0.294)

Mean, Males 6.951 7.315 7.887p-val., Male vs. Fem 0.525 0.619 0.492

OLS, No controls:Fem. Tr.ee x Fem. Operator -0.606* -0.935*** -0.229Fem. Tr.ee x Male Operator -0.892* -1.103** 0.0752

Nbr Male Tr.s 48 60 47Nbr Female Tr.s 55 62 36N 264 325 233Factory FE YES YES YES

Back

Andreas Menzel Challenges of Change September 21, 2018 13 / 19

Management Simulations: Male Workers hurt Female SVPeformance

Variable: Pay-O↵ in Games (Standardized)

Female Team Leader 0.290*** 0.255** 0.420** 0.305** 0.332*** 0.412***(0.109) (0.122) (0.190) (0.131) (0.124) (0.521)

Mixed gender / male team -0.092(0.725)

Mixed * Female TL -0.433*(0.240)

Tried as Line Supervisor 0.329(0.212)

Promoted to Line Supervisor 0.508**(0.206)

Team Fixed E↵ects no no yes no no noGame Fixed E↵ects yes yes yes yes yes yesTeam Leader Demogr. no yes yes yes yes yesNumber of Observations 676 612 612 612 608 612

Back

Andreas Menzel Challenges of Change September 21, 2018 14 / 19

Why Initial Underperformance:Fellow SV Baseline Attitudes?

Productivity Grade (1-10)

Trial-ITT Trial Trial Trial-ITT TrialSV Post-Tr.

Female Trainee -5.278* -14.046*** -7.692*** -0.672*** -0.829***(2.779) (2.841) (2.824) (0.277) (0.196)

N 91 93 55 264 325Factory FE YES YES YES YES YES

Back

Andreas Menzel Challenges of Change September 21, 2018 15 / 19

Why Initial Underperformance:(Inbalanced) Trainee Characteristics?

Productivity Grade (1-10)

Trial-ITT Trial Trial Trial-ITT TrialSV Post-Tr.

Female Trainee -0.885 -14.922*** -5.259*(3.524) (3.585) (2.860)

Fem. Tr.ee x Fem. Operator -0.563* -0.587**(0.289) (0.253)

Fem. Tr.ee x Male Operator -0.861* -0.723*(0.468) (0.405)

Locus o.C. -2.135 22.095*** -0.118 -0.423*(1.805) (7.137) (0.105) (0.218)

Confidence 0.919 -1.466* 0.162 0.181(1.315) (0.826) (0.119) (0.120)

Age2 4.966*** -68.649***(1.849) (24.355)

Age 129.94***(45.664)

Techn. Knowledge -1.067(4.349)

Years Educaction2 1.967(2.109)

Numeracy -0.049 0.103(0.107) (0.122)

N 91 93 55 264 325Factory FE YES YES YES YES YES

Back

Andreas Menzel Challenges of Change September 21, 2018 16 / 19

Line Worker Response - Grading (1-10) of New SV.s

Trial - ITT Trial Post-Trial

PDS-Lasso:Fem. Tr.ee x Fem. Operator -0.584** -0.762*** -0.349

(0.269) (0.242) (0.256)Fem. Tr.ee x Male Operator -0.931** -0.989*** 0.012

(0.431) (0.373) (0.477)Female Respondent -0.069 0.009 0.384

(0.463) (0.354) (0.294)

Mean, Males 6.951 7.315 7.887p-val., Male vs. Fem 0.525 0.619 0.492

OLS, No controls:Fem. Tr.ee x Fem. Operator -0.606* -0.935*** -0.229Fem. Tr.ee x Male Operator -0.892* -1.103** 0.0752

Nbr Male Tr.s 48 60 47Nbr Female Tr.s 55 62 36N 264 325 233Factory FE YES YES YES

Back

Andreas Menzel Challenges of Change September 21, 2018 17 / 19

Grading and Continuation as SV

Operator Fellow Supervisor

OLS PDS-Lasso OLS PDS-Lasso

Fem. Tr., NOT SV Post-Tr. -1.072** -1.115*** -1.087** -0.793**(0.437) (0.424) (0.449) (0.355)

Fem. Tr., SV Post-Tr. -0.917*** -0.731*** -0.251 -0.052(0.298) (0.242) (0.330) (0.308)

SV Post-Tr. 0.018 -0.261 -0.389 -0.158(0.419) (0.429) (0.459) (0.371)

p Stay-Non Stay 0.785 0.458 0.148 0.125

Nbr Female Tr.s 62 62 57 57Nbr Male Tr.s 60 60 51 51N 325 325 173 173Factory FE YES YES YES YES

Back

Andreas Menzel Challenges of Change September 21, 2018 18 / 19

Exposure Increases Preference for Female SVs

Trial - ITT Trial Post-Trial

PDS-Lasso:Fem. Tr.ee x Fem. Operator 0.014 0.039 0.172***

(0.067) (0.052) (0.048)Fem. Tr.ee x Male Operator 0.141** 0.113* 0.206**

(0.070) (0.066) (0.094)Female Respondent 0.226*** 0.181*** 0.124**

(0.068) (0.058) (0.063)

Mean, Males 0.287 0.293 0.277p Male Fem 0.177 0.395 0.749

OLS, No controlsFem. Tr.ee x Fem. Operator 0.0183 0.0417 0.158***Fem. Tr.ee x Male Operator 0.157** 0.121* 0.191*

N 263 324 232Factory FE YES YES YES

0: Prefer Male SV; 0.5: Indi↵erence; 1 Prefer Female SV

No similar adjustment among peer supervisors <=> Better view from beginning?

Back

Andreas Menzel Challenges of Change September 21, 2018 19 / 19