University of Pennsylvania ScholarlyCommons

Management Papers Wharton Faculty Research

2011

Field Experiments With Firms

Oriana Bandiera

Iwan Barankay University of Pennsylvania

Imran Rasul

Follow this and additional works at: https://repository.upenn.edu/mgmt_papers

Part of the Business Administration, Management, and Operations Commons

Recommended Citation Bandiera, O., Barankay, I., & Rasul, I. (2011). Field Experiments With Firms. Journal of Economic Perspectives, 25 (3), 63-82. http://dx.doi.org/10.1257/jep.25.3.63

This paper is posted at ScholarlyCommons. https://repository.upenn.edu/mgmt_papers/99 For more information, please contact [email protected]. Field Experiments With Firms

Abstract We discuss how the use of field experiments sheds light on long-standing esearr ch questions relating to firm behavior. We present insights from two classes of experiments—within and across firms—and draw common lessons from both sets. Field experiments within firms generally aim to shed light on the nature of agency problems. Along these lines, we discuss how field experiments have provided new insights on shirking behavior and the provision of monetary and nonmonetary incentives. Field experiments across firms generally aim to uncover firms' binding constraints by exogenously varying the availability of key inputs such as labor, physical capital, and managerial capital. We conclude by discussing some of the practical issues researchers face when designing experiments and by highlighting areas for further research.

Disciplines Business Administration, Management, and Operations

This journal article is available at ScholarlyCommons: https://repository.upenn.edu/mgmt_papers/99 Journal of Economic Perspectives—Volume 25, Number 3—Summer 2011—Pages 63–82

Field Experiments with Firms

Oriana Bandiera, Iwan Barankay, and Imran Rasul

iirmsrms operateoperate inin complexcomplex environments:environments: a listlist ofof thethe categoriescategories inin whichwhich ttheyhey needneed toto makemake iinterrelatednterrelated choiceschoices wouldwould includeinclude employeeemployee pay,pay, F ppricing,ricing, productproduct attributes,attributes, productionproduction technologies,technologies, andand management.management. IInn tturn,urn, thesethese decisionsdecisions involveinvolve respondingresponding toto characteristicscharacteristics thatthat areare oftenoften hardhard toto mmeasureeasure oror uncertain,uncertain, suchsuch asas thosethose relatedrelated toto marketmarket characteristics,characteristics, thethe produc-produc- ttivityivity ofof individualindividual inputs,inputs, andand entrepreneurialentrepreneurial ability.ability. DueDue toto thethe complexitycomplexity ofof tthehe eenvironment,nvironment, researchresearch thatthat seeksseeks toto understandunderstand thethe behaviorbehavior ofof fi rmsrms basedbased onon oobservationalbservational ddataata ffacesaces mmanyany cchallengeshallenges aatt uuncoveringncovering ccausalausal rrelationships.elationships. IInn tthishis ppaper,aper, wwee illustrateillustrate howhow fi eldeld experiments,experiments, guidedguided byby economiceconomic theory,theory, cancan addressaddress tthesehese challengeschallenges andand provideprovide newnew answersanswers toto long-standinglong-standing questionsquestions aboutabout fi rms:rms: DDoo fi rmrm choiceschoices maximizemaximize profiprofi tsts subjectsubject toto constraints?constraints? IfIf so,so, whichwhich constraintsconstraints bbindind andand informinform ddecisionecision makingmaking iinn fi rms?rms? IfIf not,not, whywhy areare fi rmsrms operatingoperating insideinside tthehe ffrontier?rontier? IInn tthishis ppaper,aper, wwee rrevieweview fi eldeld experimentsexperiments thatthat provideprovide preliminarypreliminary answersanswers toto tthesehese qquestionsuestions andand mapmap ddirectionsirections forfor furtherfurther research.research. WeWe organizeorganize ourour discus-discus- ssionion intointo twotwo classesclasses ofof work.work. TheThe fi rstrst isis fi eldeld experimentsexperiments conductedconducted within fi rms,rms, iinn whichwhich thethe unitsunits ofof observationobservation areare workersworkers oror divisionsdivisions ofof a fi rm.rm. TheThe theorytheory bbehindehind manymany ofof thesethese experimentsexperiments viewsviews thethe fi rmrm asas anan organization,organization, emphasizingemphasizing aagencygency problems.problems. WeWe discussdiscuss fi eldeld experimentsexperiments thatthat shedshed lightlight onon solutionssolutions toto thethe aagencygency problem,problem, fromfrom iincentivencentive paypay toto socialsocial pressurepressure andand nonmonetarynonmonetary rewards.rewards.

■ OOrianariana BandieraBandiera isis ProfessorProfessor ofof Economics,Economics, LondonLondon SchoolSchool ofof EconomicsEconomics andand PoliticalPolitical SScience,cience, LLondon,ondon, UUnitednited KKingdom.ingdom. IIwanwan BBarankayarankay iiss AAssociatessociate PProfessorrofessor ooff MManagement,anagement, WWhartonharton SSchool,chool, UniversityUniversity ofof Pennsylvania,Pennsylvania, Philadelphia,Philadelphia, PPennsylvania.ennsylvania. IImranmran RRasulasul iiss PProfessorrofessor ooff EEconomics,conomics, UUniversityniversity CCollegeollege LLondon,ondon, London,London, UnitedUnited Kingdom.Kingdom. TheirTheir e-maile-mail addressesaddresses areare 〈[email protected]@lse.ac.uk〉, 〈[email protected]@wharton.upenn.edu〉, aandnd 〈[email protected]@ucl.ac.uk〉. doi=10.1257/jep.25.3.63 64 Journal of Economic Perspectives

TThehe ssecondecond strandstrand coverscovers fi eldeld experimentsexperiments conductedconducted between fi rms,rms, inin whichwhich thethe fi rrmm iiss tthehe unitunit ofof observation.observation. TheThe theorytheory behindbehind mostmost ofof thesethese experimentsexperiments viewsviews tthehe fi rrmm tthroughhrough tthehe llensens ofof neoclassicalneoclassical productionproduction theory,theory, andand soso wewe discussdiscuss hhowow fi eldeld experimentsexperiments havehave exogenouslyexogenously variedvaried inputinput availabilityavailability toto shedshed lightlight onon cconstraintsonstraints fi rmsrms face.face. TThroughout,hroughout, wewe ffocusocus oonn eexperimentsxperiments ddesignedesigned ttoo sshedhed llightight oonn fi rms’rms’ behavior.behavior. TThishis stillstill leavesleaves outout a largelarge classclass ofof fi eldeld experimentsexperiments thatthat areare runrun inin collaborationcollaboration wwithith fi rmsrms ttoo pproviderovide evidenceevidence onon otherother issuesissues suchsuch asas consumerconsumer behaviorbehavior oror optimaloptimal aauctionuction ddesign.esign.1 BBeyondeyond thethe resultsresults ofof specifispecifi c fi eldeld experimentexperiment studies,studies, wewe alsoalso believebelieve thatthat eeconomistsconomists cancan rreapeap eenormousnormous bbenefienefi tsts fromfrom establishingestablishing workingworking partnershipspartnerships wwithith fi rmsrms aandnd eengagingngaging iinn pprimaryrimary ddataata ccollection.ollection. TThus,hus, wwee cconcludeonclude tthehe ppaperaper bbyy oofferingffering ssomeome ddiscussioniscussion ooff tthehe ppracticalractical iissuesssues rresearchersesearchers ffaceace iinn ddesigningesigning aandnd iimplementingmplementing fi eldeld experimentsexperiments inin fi rms,rms, andand byby highlightinghighlighting researchresearch questionsquestions tthathat rremainemain rrelativelyelatively untoucheduntouched byby fi eldeld experiments.experiments. WeWe hopehope thatthat byby thethe endend ofof oourur discussion,discussion, readersreaders havehave a clearclear sensesense ofof thethe costscosts andand benefibenefi tsts ofof fi eldeld experi-experi- mmentsents iinn fi rmrm settings,settings, andand areare motivatedmotivated ttoo cconsideronsider thisthis approachapproach themselves.themselves.

FFieldield EExperimentsxperiments withinwithin FirmsFirms

FFieldield experimentsexperiments withinwithin fi rmsrms areare generallygenerally designeddesigned toto shedshed lightlight onon howhow fi rrmsms ccanan solvesolve agencyagency problemsproblems andand motivatemotivate theirtheir employees.employees. InIn thisthis section,section, wewe rrevieweview evidenceevidence onon thethe twotwo classicalclassical solutionssolutions toto thisthis agencyagency problem—monitoringproblem—monitoring aandnd paypay forfor performance—asperformance—as wellwell asas moremore recentrecent workwork onon nonmonetarynonmonetary determi-determi- nnantsants ooff mmotivationotivation ssuchuch aass ssocialocial rrelationselations oorr sstatustatus rrewards.ewards. AAlthoughlthough fi eldeld experimentsexperiments withinwithin fi rmsrms havehave experiencedexperienced a recentrecent resurgence,resurgence, ttheyhey areare farfar fromfrom new.new. OneOne ofof thethe fi rstrst seriesseries ofof fi eldeld experimentsexperiments waswas conductedconducted atat tthehe HawthorneHawthorne plantplant ofof thethe WesternWestern ElectricElectric Company,Company, nearnear Chicago,Chicago, inin thethe 1920s.1920s. WWhilehile tthehe vvalidityalidity ooff ttheirheir sspecifipecifi c fi ndingsndings hashas beenbeen questioned,questioned,2 tthesehese experimentsexperiments llayay tthehe ggroundworkroundwork forfor manymany issuesissues thatthat areare nownow consideredconsidered partpart ofof mainstreammainstream ppersonnelersonnel economicseconomics (Bloom(Bloom aandnd VVanan RReenen,eenen, 22010a).010a). FForor eexample,xample, ttheyhey lleded MMayoayo ((1933)1933) toto stressstress thatthat workersworkers areare motivatedmotivated byby bothboth monetarymonetary andand nonmonetarynonmonetary rrewardsewards fromfrom work,work, anan ideaidea thatthat isis beingbeing testedtested byby thethe newestnewest generationgeneration ofof fi eldeld eexperimentsxperiments rreviewedeviewed bbelow.elow.

1 Further discussion of fi eld experiments on fi rms is provided in Levitt and List (2009) where they discuss fi eld experiments related to how consumers respond to product attributes and pricing. On auctions, a nascent literature is now emerging that uses fi eld experiments to measure, for example, reserve price effects (Reiley, 2006; Brown and Morgan, 2009; Ostrovsky and Schwarz, 2009). 2 As one example, between 1924 and 1927 the level of lighting was systematically changed for experi- mental groups in different departments (Mayo, 1933). Levitt and List (2011) recently recovered the thought-to-be-lost data from this experiment, and fi nd little evidence that workers reacted to the differ- ences in lighting. Oriana Bandiera, Iwan Barankay, and Imran Rasul 65

MMonitoringonitoring aandnd SShirkinghirking TThehe sstandardtandard aagencygency fframeworkramework wwithith aasymmetricsymmetric iinformationnformation vviewsiews eemployeesmployees aass rrationalational sshirkers:hirkers: tthathat iis,s, eemployeesmployees cconsideronsider tthehe mmarginalarginal ccostsosts aandnd mmarginalarginal bbenefienefi tsts ofof shirking,shirking, andand decidedecide onon theirtheir levellevel ofof effort.effort. FirmsFirms thusthus choosechoose compen-compen- ssationation andand monitoringmonitoring policiespolicies withwith shirkingshirking inin mind.mind. TheThe theorytheory suggestssuggests thatthat a rreductioneduction iinn mmonitoringonitoring wwillill ttendend ttoo aann iincreasencrease iinn sshirking.hirking. MMoreover,oreover, aann iincreasencrease iinn sshirkinghirking rresultingesulting ffromrom rreducededuced mmonitoringonitoring sshouldhould bbee ggreatestreatest aamongmong iindividualsndividuals fforor wwhomhom tthehe oongoingngoing employmentemployment relationshiprelationship isis leastleast valuable.valuable. ThreeThree concernsconcerns hhaveave pplaguedlagued nnonexperimentalonexperimental approachesapproaches toto testingtesting thesethese ideas:ideas: 1)1) shirkingshirking bbehaviorehavior isis hardhard toto detect;detect; 2)2) thethe abilityability ofof thethe econometricianeconometrician toto detectdetect shirkersshirkers mmightight bbee eendogenouslyndogenously relatedrelated toto thethe employer’semployer’s mmonitoringonitoring practices;practices; andand 33)) unobservedunobserved factors,factors, suchsuch asas hiringhiring policies,policies, maymay leadlead monitoringmonitoring andand shirkingshirking ooutcomesutcomes ttoo bbee ccorrelated.orrelated. NNagin,agin, Rebitzer,Rebitzer, Sanders,Sanders, andand TaylorTaylor (2002)(2002) addressaddress thesethese challengeschallenges usingusing a fi eldeld experimentexperiment runrun byby a telephonetelephone solicitationsolicitation fi rmrm acrossacross fourfour ofof itsits 1616 callcall ccenters.enters. AtAt eacheach callcall center,center, ttelephoneelephone solicitorssolicitors werewere paidpaid a piecepiece raterate inin whichwhich salarysalary iincreasedncreased wwithith tthehe nnumberumber ooff ““successful”successful” ssolicitations—whereolicitations—where ssuccessuccess wwasas rreportedeported bbyy thethe employeesemployees themselves.themselves. ThisThis piecepiece raterate createdcreated incentivesincentives forfor employeesemployees toto ooverstateverstate whetherwhether a donationdonation hadhad beenbeen promised.promised. ToTo curbcurb opportunisticopportunistic behavior,behavior, tthehe employeremployer monitoredmonitored byby callingcalling backback a fractionfraction ofof thosethose whowho werewere reportedreported toto hhaveave respondedresponded positivelypositively toto a solicitation.solicitation. EmployeesEmployees werewere informedinformed whenwhen hiredhired tthathat ttheirheir aactivitiesctivities wouldwould bebe checkedchecked byby callbacks.callbacks. TheThe resultsresults ofof eacheach week’sweek’s call-call- bbacksacks wwereere ccommunicatedommunicated ttoo bbothoth eemployeesmployees aandnd ttheirheir iimmediatemmediate ssupervisors,upervisors, aandnd ccallsalls ffoundound bbyy tthehe mmonitorsonitors ttoo bbee uunsuccessfulnsuccessful wwereere ddeductededucted ffromrom eeachach iindividual’sndividual’s wweeklyeekly incentiveincentive pay.pay. StrongerStronger sanctionssanctions forfor suchsuch callscalls werewere notnot generallygenerally imposedimposed oonn eemployeesmployees bbecauseecause iitt wwasas uunderstoodnderstood tthathat ddonorsonors ssometimesometimes cchangehange ttheirheir mmindsinds aafterfter aagreeinggreeing ttoo ppledgeledge mmoney.oney. TToo seesee ifif thethe costscosts ofof thisthis monitoringmonitoring systemsystem couldcould bebe reduced,reduced, thethe companycompany ssimulatedimulated a lowerlower auditaudit raterate byby experimentallyexperimentally varyingvarying thethe fractionfraction ofof badbad callscalls thatthat wwereere reportedreported backback toto employeesemployees andand supervisorssupervisors inin thethe fourfour experimentalexperimental sitessites ffromrom 0 toto 2 toto 5 toto 1010 percent,percent, whilewhile keepingkeeping thethe actualactual auditaudit raterate atat 2525 percentpercent inin aallll ffourour ssites.ites. BByy wworkingorking withwith thethe fi rm,rm, thethe researchersresearchers werewere ableable toto collectcollect surveysurvey ddataata oonn eemployeemployee aattitudesttitudes towardtoward thethe job,job, theirtheir expectedexpected jobjob tenure,tenure, andand theirtheir pperceivederceived ddiffiiffi cultyculty ofof fi ndingnding a comparablecomparable job.job. TThehe fi ndingsndings indicateindicate thatthat workers’workers’ responsesresponses areare veryvery heterogeneous.heterogeneous. BetweenBetween 1100 andand 4141 percentpercent ofof thethe employeesemployees inin thethe fourfour experimentalexperimental sitessites behavebehave asas ““rationalrational cheaters”—thatcheaters”—that is,is, theythey respondrespond toto a reductionreduction inin thethe perceivedperceived costcost ofof oopportunisticpportunistic behaviorbehavior byby increasingincreasing thethe raterate atat whichwhich theythey shirk.shirk. TheThe remainingremaining 5599 toto 9090 percentpercent ofof employees,employees, however,however, dodo notnot increaseincrease shirkingshirking followingfollowing thethe rreductioneduction iinn mmonitoringonitoring rrates.ates. UUsingsing tthehe ssurveyurvey ddataata ccollected,ollected, tthehe aauthorsuthors fi ndnd thatthat tthosehose eemployeesmployees wwhoho rrespondedesponded toto rreductionseductions inin mmonitoringonitoring ttendedended ttoo bbee tthosehose wwhoho pperceivederceived tthehe eemployermployer aass bbeingeing uunfairnfair aandnd uuncaringncaring aandnd tthat,hat, iinn ccontrastontrast wwithith tthehe rrationalational cheatercheater model,model, individualsindividuals withwith goodgood outsideoutside optionsoptions diddid notnot increaseincrease sshirkinghirking byby moremore thanthan otherother workersworkers whenwhen thethe raterate ofof monitoringmonitoring declined.declined. ThisThis 66 Journal of Economic Perspectives

hheterogeneityeterogeneity impliesimplies thatthat tthehe ooptimalptimal mmonitoringonitoring sschemecheme wwillill nneedeed ttoo bbalancealance tthehe rrequirementequirement toto rreduceeduce tthehe sshirkinghirking bbehaviorehavior ooff ssomeome wworkersorkers iinclinednclined ttoo rrationallyationally ccheat,heat, againstagainst thethe ddesireesire ttoo aavoidvoid mmonitoringonitoring ccostsosts fforor tthosehose uunlikelynlikely ttoo ccheatheat uundernder nnormalormal ccircumstances.ircumstances.

MMonetaryonetary IIncentivesncentives IInn tthehe mmanyany ccircumstancesircumstances inin whichwhich monitoringmonitoring isis notnot practicallypractically feasible,feasible, tthehe agencyagency problemproblem cancan bebe addressedaddressed byby designingdesigning incentiveincentive schemesschemes thatthat seekseek toto aalignlign thethe employees’employees’ interestsinterests withwith thethe principal’s.principal’s. A widewide classclass ofof schemesschemes suchsuch asas ppieceiece rates,rates, bonuses,bonuses, andand prizesprizes achieveachieve thisthis goalgoal byby makingmaking thethe employees’employees’ paypay anan iincreasingncreasing ffunctionunction ooff ttheirheir pperformance.erformance. AAgencygency theorytheory makesmakes preciseprecise thatthat suchsuch “pay“pay forfor performance”performance” schemesschemes affectaffect pproductivityroductivity bothboth byby increasingincreasing thethe productivityproductivity ofof existingexisting employeesemployees (the(the incen-incen- ttiveive eeffect)ffect) aandnd bbyy aattractingttracting mmore-productiveore-productive eemployeesmployees ttoo tthehe fi rmrm (the(the selectionselection eeffect).ffect). TheThe incentiveincentive effecteffect arisesarises becausebecause pay-for-performancepay-for-performance schemesschemes increaseincrease tthehe marginalmarginal benefibenefi t ofof effort,effort, whichwhich leadsleads employeesemployees toto workwork harder,harder, otherother thingsthings eequal.qual. TheThe selectionselection effecteffect arisesarises becausebecause high-abilityhigh-ability employeesemployees whowho areare capablecapable ooff aachievingchieving highhigh performanceperformance cancan achieveachieve higherhigher paypay andand are,are, thus,thus, attractedattracted byby sschemeschemes tthathat rrewardeward pperformance,erformance, ootherther tthingshings eequal.qual. BBothoth tthehe iincentivencentive aandnd sselec-elec- ttionion eeffectsffects increaseincrease thethe variancevariance asas wellwell asas thethe meanmean ofof productivityproductivity andand pay,pay, asas mmore-ableore-able wworkersorkers ccanan rrespondespond mmoreore ttoo tthehe iincreasencrease iinn iincentivencentive ppower.ower. TTestingesting thethe richrich predictionspredictions ofof agencyagency theorytheory usingusing nonexperimentalnonexperimental datadata facesfaces sseriouserious eeconometricconometric cchallenges,hallenges, mostmost iimportantlymportantly tthathat oobservedbserved iincentivencentive ccontractsontracts mmightight bbee eendogenousndogenous toto fi rmrm performanceperformance (Prendergast,(Prendergast, 1999;1999; ChiapporiChiappori andand SSalanié,alanié, 2003).2003). FieldField experimentersexperimenters cancan ttackleackle tthishis cchallengehallenge ddirectlyirectly bbyy eengineeringngineering eexogenousxogenous changeschanges inin incentiveincentive schemesschemes across,across, oror mostmost commonly,commonly, withinwithin fi rms.rms. MMoreover,oreover, hhigh-poweredigh-powered incentiveincentive schemesschemes mightmight affectaffect productivityproductivity throughthrough chan-chan- nnelsels thatthat areare typicallytypically notnot measuredmeasured inin secondarysecondary data.data. ForFor instance,instance, thethe increaseincrease iinn ppayay iinequalitynequality cancan reducereduce workers’workers’ moralemorale andand leadlead toto sabotage,sabotage, oror thethe changechange iinn tthehe ccompositionomposition ofof thethe peerpeer groupgroup duedue toto thethe selectionselection effecteffect mightmight affectaffect wworkers’orkers’ behaviorbehavior overover andand aboveabove thethe effecteffect ofof thethe compensationcompensation scheme.scheme. FieldField eexperimentersxperimenters areare wellwell placedplaced toto workwork withwith fi rmsrms toto collectcollect primaryprimary datadata onon thethe rrelevanceelevance ofof thesethese mechanisms,mechanisms, fforor iinstancenstance bbyy mmeasuringeasuring ssocialocial ttiesies wwithinithin tthehe fi rm.rm. AAmongmong tthehe fi rstrst ofof thethe fi eldeld experimentsexperiments designeddesigned toto measuremeasure incentiveincentive effectseffects ooff monetarymonetary compensationcompensation schemesschemes isis thatthat ofof ShearerShearer (2004),(2004), whowho estimatesestimates tthehe pproductivityroductivity gainsgains movingmoving fromfrom a fi xedxed wagewage toto a piecepiece raterate schemescheme forfor treetree pplanterslanters inin BritishBritish Columbia,Columbia, Canada.Canada. WorkersWorkers werewere randomlyrandomly assignedassigned toto plantplant uundernder oonene ooff tthehe iincentivencentive sschemeschemes aatt tthehe sstarttart ooff a wworkork dday.ay. WWorkers’orkers’ pproductivityroductivity iincreasedncreased byby 2020 percentpercent movingmoving fromfrom fi xedxed wageswages toto piecepiece rates.rates. InIn lineline withwith thethe ppredictionrediction ofof agencyagency theory,theory, thethe standardstandard deviationdeviation ofof outputoutput acrossacross workersworkers waswas hhigherigher underunder piecepiece rates.rates. ShearerShearer alsoalso developsdevelops andand estimatesestimates a structuralstructural modelmodel ooff wworkers’orkers’ behaviorbehavior toto shedshed lightlight on:on: 1)1) whatwhat wouldwould havehave beenbeen thethe productivityproductivity ggainsains ifif mmanagementanagement hhadad bbeeneen iimperfectlymperfectly iinformednformed aaboutbout pplantinglanting cconditions;onditions; aandnd 2)2) howhow workersworkers wouldwould havehave respondedresponded toto anan effieffi ciencyciency wagewage scheme.scheme. ThisThis isis a Field Experiments with Firms 67

““naturalnatural fi eldeld experiment,”experiment,” toto useuse thethe terminologyterminology explainedexplained inin thethe introductionintroduction toto tthishis ssymposiumymposium aandnd iinn HHarrisonarrison aandnd LListist ((2004),2004), bbecauseecause tthehe wworkersorkers ddidid nnotot kknownow tthathat tthehe eexperimentxperiment wwasas ooccurring.ccurring. CCombiningombining rresearchesearch mmethodsethods llikeike tthishis iiss ttoo bbee eencouraged.ncouraged. IInn mmanyany sscenarios,cenarios, iitt wouldwould bebe idealideal toto combinecombine evidenceevidence fromfrom fi eldeld experimentsexperiments withwith structuralstructural mmodelingodeling toto ppositosit aann uunderlyingnderlying bbehavioralehavioral mechanismmechanism bbehindehind tthehe eeffects,ffects, ttoo aassessssess tthehe sensitivitysensitivity ofof thethe estimatesestimates toto slightslight alterationsalterations inin thethe economiceconomic environment,environment, aandnd toto makemake headwayheadway inin understandingunderstanding thethe optimaloptimal compensationcompensation structures.structures. OfOf ccourse,ourse, tthehe vvalidityalidity ooff tthehe sstructuraltructural mmodelodel ccanan iitselftself bbee ttestedested bbyy eexploringxploring wwhetherhether iitt ppredictsredicts tthehe rresponsesesponses observedobserved toto tthehe eexogenousxogenous vvariationariation engineeredengineered byby thethe fi eldeld experiment.experiment. ForFor example,example, ChoCho andand RustRust (2010)(2010) followfollow thisthis approachapproach inin usingusing a fi eeldld eexperimentxperiment ttoo vvalidatealidate a sstructuraltructural mmodelodel ooff rrentalental rratesates fforor aautomobiles.utomobiles. WWhilehile rrecentecent fi eldeld experimentsexperiments havehave mademade substantialsubstantial progressprogress inin identifyingidentifying thethe ccausalausal effecteffect ofof performanceperformance paypay onon workers’workers’ effort,effort, theythey havehave beenbeen lessless successfulsuccessful atat ppinninginning ddownown sselectionelection eeffects.ffects.3 TThishis iiss a ppriorityriority fforor ffutureuture rresearchesearch aass tthehe aavailablevailable nnonexperimentalonexperimental eevidencevidence ssuggestsuggests tthathat sselectionelection eeffectsffects aarere aatt lleasteast aass iimportantmportant aass iincentivencentive eeffects.ffects. IInn a nnonexperimentalonexperimental sstudytudy tthathat eexploitsxploits tthehe rroll-outoll-out ooff a nnewew ppieceiece rrateate sschemecheme iinn a mmanufacturinganufacturing fi rm,rm, LazearLazear ((2000)2000) sshowshows tthathat sselectionelection eeffectsffects eexplainxplain hhalfalf ooff tthehe 4444 ppercentercent iincreasencrease iinn wworkerorker pproductivityroductivity tthathat ffollowedollowed tthehe iintroductionntroduction ooff ppieceiece rrates.ates. HHowever,owever, iidentifyingdentifying sselectionelection eeffectsffects pposesoses a ddiffiiffi cultcult cchallengehallenge fforor fi eeldld eexperimentersxperimenters bbothoth bbecauseecause iitt rrequiresequires iinformationnformation oonn tthehe eentirentire ppoolool ooff ppotentialotential eemployeesmployees aandnd bbecauseecause tthehe ttimeime hhorizonorizon ooff fi eldeld eexperimentsxperiments iiss ooftenften cconsiderablyonsiderably sshorterhorter tthanhan tthathat nneededeeded fforor eexistingxisting wworkersorkers ttoo qquituit aandnd nnewew wworkersorkers ttoo jjoinoin a fi rrm.m. VVaryingarying iincentivesncentives aacrosscross ddivisionsivisions oorr pplantslants ooff tthehe ssameame fi rmrm wwhilehile aallowingllowing eemployeesmployees ttoo mmoveove aacrosscross ddivisionsivisions mmightight bbee a wwayay ttoo aaddressddress bbothoth iissues.ssues. GGreaterreater kknowledgenowledge aaboutbout sselectionelection eeffectsffects wwouldould aalsolso hhelpelp iinn uunderstandingnderstanding wwhetherhether aandnd hhowow tthehe ccompensationompensation ppoliciesolicies ooff a ggiveniven fi rmrm havehave spilloverspillover eeffectsffects oonn ootherther fi rrmsms tthathat ccompeteompete fforor ssimilarimilar wworkers—whichorkers—which iinn tturnurn wwouldould hhelpelp ttieie ttogetherogether tthehe ttwowo ddisparateisparate lliteraturesiteratures oonn wwithin-fiithin-fi rmrm compensationcompensation ppoliciesolicies aandnd eequilibriumquilibrium wwage-settingage-setting bbehavior.ehavior.

MMonetaryonetary IncentivesIncentives andand thethe SocialSocial OrganizationOrganization ooff tthehe WWorkplaceorkplace UUntilntil rrecently,ecently, tthehe iimportancemportance ooff tthehe iinteractionnteraction bbetweenetween ssocialocial rrelationselations aandnd mmonetaryonetary iincentivesncentives iinn tthehe wworkplaceorkplace hhasas bbeeneen aaddressedddressed mmainlyainly iinn tthehe oorganiza-rganiza- ttionalional aandnd bbusinessusiness ssociologyociology lliteratures.iteratures. HHowever,owever, ssuchuch cconcernsoncerns hhaveave bbegunegun ttoo bbee iincorporatedncorporated iinn eeconomicconomic ttheory;heory; fforor eexample,xample, KKandelandel aandnd LLazearazear ((1992)1992) aandnd RRotembergotemberg ((1994)1994) eextendxtend sstandardtandard aagencygency mmodelsodels ttoo ttakeake iintonto aaccountccount ppeereer eeffectsffects aandnd ssocialocial cconcerns.oncerns. MMoreover,oreover, ccredibleredible eevidencevidence oonn tthehe eexistencexistence aandnd mmagnitudeagnitude ooff ssuchuch ssocialocial mmechanismsechanisms hhasas bbegunegun ttoo eemergemerge uusingsing nnonexperimentalonexperimental mmethodsethods iinn ccombinationombination wwithith ppersonnelersonnel ddataata ((MasMas aandnd MMoretti,oretti, 22009;009; BBandiera,andiera, BBarankay,arankay, aandnd RRasul,asul, 22010).010). IIdentifyingdentifying tthehe ccausalausal eeffectsffects ooff ssocialocial rrelationselations uusingsing nnonexperimentalonexperimental

3 Niederle and Vesterlund (2007) explore selection effects of incentives in a laboratory setting. 68 Journal of Economic Perspectives

ddataata ffacesaces ttwowo mmainain cchallenges:hallenges: 11)) tthehe oobservedbserved vvariationariation iinn iincentivesncentives mmightight bbee ccorrelatedorrelated wwithith ootherther uunobservablenobservable ddeterminantseterminants ooff pperformance;erformance; aandnd 22)) fi rmrm ppersonnelersonnel recordsrecords rarelyrarely containcontain informationinformation onon socialsocial connectionsconnections withinwithin thethe fi rrm.m.4 IInn thisthis section,section, wewe discussdiscuss a seriesseries ofof threethree ofof ourour ownown fi eldeld experimentsexperiments ((Bandiera,Bandiera, Barankay,Barankay, aandnd RRasul,asul, 22005,005, 22007,007, 22009,009, 22011)011) tthathat pproviderovide novelnovel evidenceevidence oonn thethe interplayinterplay ofof incentivesincentives andand thethe socialsocial organizationorganization ofof thethe workplace,workplace, namelynamely tthehe socialsocial relationsrelations thatthat existexist betweenbetween a groupgroup ofof coworkerscoworkers oror betweenbetween workersworkers aandnd managers.managers. TheThe fi rmrm wewe studystudy isis a leadingleading U.K.U.K. producerproducer ofof “soft“soft fruit”—afruit”—a bbroadroad categorycategory thatthat mightmight includeinclude fruitsfruits likelike plums,plums, cherries,cherries, peaches,peaches, strawberries,strawberries, rraspberries,aspberries, andand grapes.grapes. TheThe fi rm’srm’s hierarchyhierarchy hashas fourfour layers:layers: thethe ownerowner andand chiefchief eexecutivexecutive ooffiffi cer,cer, thethe generalgeneral manager,manager, fi eldeld managers,managers, andand workers.workers. TheThe mainmain tasktask ooff tthehe bbottom-tierottom-tier wworkersorkers iiss ttoo ppickick ffruit.ruit. AAroundround 4400 wworkersorkers ppickick oonn aanyny ggiveniven fi eldeld oonn a givengiven day.day. WithinWithin a fi eld,eld, workersworkers areare allocatedallocated theirtheir ownown rowrow ofof fruitfruit toto pick,pick, aandnd wworker’sorker’s pproductivityroductivity isis defidefi nedned asas kilogramskilograms pickedpicked perper hour.hour. ManagersManagers orga-orga- nnizeize fi eldeld logistics:logistics: forfor example,example, theythey assignassign workersworkers toto rowsrows andand makemake suresure workers’workers’ ffullull cratescrates areare replacedreplaced withwith emptyempty ones.ones. ManagersManagers alsoalso choosechoose howhow muchmuch eeffortffort ttoo eexertxert aandnd hhowow ttoo aallocatellocate thisthis efforteffort amongamong differentdifferent workers.workers. InIn thisthis setting,setting, mmanagerialanagerial efforteffort isis complementarycomplementary toto worker’sworker’s effort;effort; forfor example,example, ifif a managermanager rreassignseassigns a wworkerorker ttoo a nnewew rrowow aass ssoonoon aass tthathat wworkerorker iiss ddoneone wwithith tthehe ppreviousrevious oonene aandnd removesremoves thatthat worker’sworker’s ffullull ccratesrates qquickly,uickly, tthenhen eevenven fforor a ggiveniven eeffortffort llevelevel ooff tthehe wworker,orker, tthehe wworkerorker wwillill bebe moremore pproductive.roductive. TThehe ggeneraleneral mmanager,anager, wwhoho iiss a ppermanentermanent eemployeemployee ooff tthehe ffarm,arm, ddecidesecides wwhichhich ooff tthehe workersworkers presentpresent onon thethe farmfarm areare selectedselected toto pickpick fruitfruit eacheach day,day, whichwhich areare aassignedssigned ttoo nnon-pickingon-picking ttasks,asks, aandnd wwhichhich aarere llefteft uunemployednemployed fforor tthehe dday.ay. IInn ourour setting,setting, workersworkers andand managersmanagers areare hiredhired seasonallyseasonally fromfrom EasternEastern EEuropeurope andand livelive onon thethe farmfarm forfor thethe durationduration ofof theirtheir stay;stay; thus,thus, theythey havehave oppor-oppor- ttunitiesunities toto formform socialsocial connections.connections. TheseThese connectionsconnections cancan bebe measuredmeasured byby askingasking wworkersorkers toto reportreport colleaguescolleagues toto whomwhom theythey areare sociallysocially linked,linked, oror indirectlyindirectly byby uusingsing commoncommon characteristics—likecharacteristics—like a commoncommon language—thatlanguage—that predictpredict socialsocial llinks.inks. InIn twotwo ofof thethe threethree experimentalexperimental seasons,seasons, thethe groupgroup ofof coworkerscoworkers thatthat a ggiveniven workerworker isis assignedassigned toto workwork withwith variesvaries acrossacross fi eldselds andand days,days, andand thisthis varia-varia- ttionion isis orthogonalorthogonal toto otherother determinantsdeterminants ofof productivity.productivity. ThisThis createscreates plausiblyplausibly eexogenousxogenous variationvariation inin thethe presencepresence ofof sociallysocially connectedconnected workersworkers thatthat cancan bebe uusedsed toto identifyidentify thethe effecteffect ofof socialsocial connectionsconnections onon behaviorbehavior andand productivity.productivity. OOurur threethree experimentsexperiments ranran duringduring thethe 2002,2002, 2003,2003, andand 20052005 pickingpicking seasons.seasons. TThehe workforceworkforce changeschanges annually,annually, soso thatthat workersworkers andand managersmanagers areare exposedexposed ttoo oneone experimentexperiment only.only. EachEach experimentexperiment inducesinduces exogenouslyexogenously timedtimed changeschanges toto tthehe incentiveincentive structurestructure ofof oneone layerlayer ofof thethe hierarchy.hierarchy. InIn allall casescases thethe experimentalexperimental ttreatmentsreatments areare appliedapplied simultaneouslysimultaneously toto allall relevantrelevant agents.agents. TheThe rationalerationale behindbehind tthehe wwithin-subjectithin-subject designdesign isis that,that, likelike mostmost ootherther fi rmrm settings,settings, itit waswas impossibleimpossible toto ppreventrevent informationinformation spilloversspillovers betweenbetween treatedtreated andand controlcontrol groups.groups. Moreover,Moreover, tthehe

4 List and Rasul (2011) provide a more comprehensive review of this literature. Oriana Bandiera, Iwan Barankay, and Imran Rasul 69

ccompositionomposition ofof tthehe wworkforceorkforce isis fi xedxed forfor thethe durationduration ofof eacheach season,season, thusthus wewe focusfocus oonn incentiveincentive ratherrather thanthan selectionselection effectseffects throughout.throughout. OneOne aspectaspect ofof thethe experi-experi- mmentalental ddesignesign isis thatthat time-varyingtime-varying unobservablesunobservables representrepresent thethe mainmain identifiidentifi cationcation tthreat.hreat. TThishis iiss aaddressedddressed bbyy aallowingllowing fl exibleexible interactionsinteractions withwith timetime effectseffects andand byby uusingsing ddifference-in-differencesifference-in-differences estimatorsestimators thatthat ccombineombine ddataata ffromrom tthehe eexperimentalxperimental sseasonseasons aandnd ffromrom 22004,004, dduringuring wwhichhich nnoo eexperimentsxperiments wwereere iimplemented.mplemented. IInn ourour fi rstrst experiment,experiment, wewe exogenouslyexogenously variedvaried thethe workers’workers’ compensationcompensation sschemecheme ffromrom ““relativerelative ccompensation”ompensation” ttoo ppieceiece rrates.ates. UUndernder rrelativeelative ccompensation,ompensation, wworkersorkers areare paidpaid a unitunit priceprice forfor eacheach kilogramkilogram ofof fruitfruit pickedpicked thatthat isis negativelynegatively rrelatedelated toto averageaverage productivityproductivity onon thethe fi eld-day—thus,eld-day—thus, ifif averageaverage productivityproductivity onon a ccertainertain dayday waswas high,high, thethe unitunit priceprice paidpaid toto workersworkers isis lower.lower. UnderUnder piecepiece rates,rates, wworkersorkers areare paidpaid forfor eeachach kkilogramilogram ooff ffruitruit ppicked.icked. UUndernder rrelativeelative ccompensation,ompensation, eeachach wworkerorker imposesimposes a negativenegative externalityexternality onon colleagues:colleagues: anyany workerworker whowho increasesincreases eeffortffort willwill raiseraise averageaverage productivityproductivity andand soso reducereduce thethe unitunit wagewage forfor allall coworkerscoworkers oonn thethe fi eld.eld. UnderUnder piecepiece rates,rates, thisthis externalityexternality doesdoes notnot exist.exist. AsAs a consequence,consequence, uundernder relativerelative compensation,compensation, thethe sociallysocially optimaloptimal levellevel ofof efforteffort isis lowerlower thanthan thethe pprivaterivate optimum,optimum, wwhereashereas underunder piecepiece rates,rates, thethe twotwo coincide.coincide. TheThe comparisoncomparison ofof pproductivityroductivity underunder thethe ttwowo sschemeschemes rrevealseveals wwhether,hether, aandnd ttoo wwhathat eextent,xtent, wworkersorkers aarere aableble ttoo ccooperate.ooperate. IInn oourur 22005005 ppaper,aper, wwee sshowhow tthathat pproductivityroductivity iiss 5500 ppercentercent hhigherigher uundernder ppieceiece rrates.ates. CCalibrationalibration ooff tthehe mmaximizationaximization cconditionsonditions ooff wworker’sorker’s iindividualndividual eeffortffort choicechoice problemproblem revealsreveals thisthis productivityproductivity differentialdifferential toto bebe consistentconsistent withwith tthehe assumptionassumption thatthat underunder relativerelative incentives,incentives, workersworkers internalizeinternalize thethe negativenegative eexternalityxternality theirtheir efforteffort imposesimposes onon coworkers.coworkers. ThisThis socialsocial incentiveincentive isis equiva-equiva- llentent ttoo tthemhem pplacinglacing a wweighteight ooff ttwo-thirdswo-thirds oonn aallll ccoworkers’oworkers’ ppay.ay. WWee fi ndnd thatthat wworkersorkers internalizeinternalize thethe externalityexternality moremore whenwhen theythey workwork alongsidealongside theirtheir friendsfriends aass opposedopposed toto colleaguescolleagues toto whomwhom theythey areare notnot sociallysocially linked,linked, andand thisthis effecteffect isis llargerarger inin smallersmaller groups.groups. Yet,Yet, thethe effectseffects disappeardisappear underunder piecepiece rates.rates. Finally,Finally, wewe fi ndnd thatthat cooperationcooperation collapsescollapses whenwhen workersworkers cannotcannot monitormonitor eacheach other;other; specifispecifi - ccally,ally, movingmoving toto piecepiece ratesrates doesdoes notnot increaseincrease workers’workers’ productivityproductivity whenwhen theythey ppickick fromfrom a plantplant whosewhose physicalphysical characteristicscharacteristics makemake itit diffidiffi cultcult forfor a workerworker toto sseeee colleagues.colleagues. ThisThis fi ndingnding rulesrules outout purepure altruismaltruism asas a mechanismmechanism toto sustainsustain ccooperationooperation inin thisthis setting.setting. OOurur ssecondecond eexperimentxperiment eexogenouslyxogenously vvariedaried mmanagers’anagers’ ppayay ffromrom fi xedxed wageswages ttoo fi xedxed wageswages plusplus a performanceperformance bonusbonus thatthat increasesincreases inin thethe averageaverage produc-produc- ttivityivity ooff tthehe wworkersorkers mmanaged.anaged. IInn tthishis eexperiment,xperiment, wworkersorkers wwereere ppaidaid ppieceiece rratesates tthroughout.hroughout. OOurur 22007007 ppaperaper sshowshows tthathat tthehe iintroductionntroduction ooff mmanagerialanagerial bbonusesonuses iincreasesncreases bbothoth tthehe mmeanean aandnd tthehe ddispersionispersion ooff wworkers’orkers’ pproductivity.roductivity. OOnene rreason,eason, aass ttheoryheory ssuggests,uggests, iiss tthathat aafterfter tthehe iintroductionntroduction ooff pperformanceerformance ppay,ay, mmanagersanagers ttargetedargeted ttheirheir eeffortffort ttowardsowards mmore-ableore-able wworkers.orkers. AAnothernother rreason,eason, aagaingain ssuggesteduggested bbyy ttheory,heory, iiss tthathat wworkersorkers wwithith tthehe hhighestighest pproductivityroductivity wwereere mmoreore llikelyikely ttoo bbee sselectedelected iintonto tthehe wworkforceorkforce wwhenhen mmanagersanagers wwereere ppaidaid pperformanceerformance bbonuses.onuses. LLeasteast aableble workersworkers werewere employedemployed lessless oftenoften andand workersworkers atat thethe bottombottom ofof thethe produc-produc- ttivityivity distributiondistribution werewere nevernever selectedselected toto pick.pick. 70 Journal of Economic Perspectives

OOurur 20092009 paperpaper furtherfurther analyzedanalyzed thethe datadata fromfrom thisthis secondsecond experiment,experiment, withwith a focusfocus onon thethe interplayinterplay betweenbetween managerialmanagerial incentivesincentives andand thethe socialsocial connectionsconnections bbetweenetween workersworkers andand managers.managers. WeWe fi ndnd evidenceevidence thatthat whenwhen managersmanagers werewere paidpaid fi xedxed wages,wages, theythey targetedtargeted workersworkers toto whomwhom theythey werewere sociallysocially connected,connected, regardlessregardless ooff abilityability level.level. WhenWhen managersmanagers werewere paidpaid asas a functionfunction ofof fi rmrm performance,performance, theythey ttargetedargeted hhigh-abilityigh-ability wworkers,orkers, rregardlessegardless ooff ssocialocial cconnections.onnections. TThehe fi ndingsndings suggestsuggest tthathat socialsocial connectionsconnections havehave a sizablesizable impactimpact onon productivity:productivity: whenwhen managersmanagers werewere ppaidaid fi xedxed wages,wages, thethe averageaverage workerworker waswas 9 percentpercent moremore productiveproductive onon daysdays whenwhen ttheyhey wwereere mmanagedanaged bbyy ssomeoneomeone ttoo wwhomhom ttheyhey wwereere ssociallyocially cconnected.onnected. OOurur tthirdhird eexperimentxperiment eevaluatesvaluates ddifferentifferent ccompensationompensation sschemeschemes uundernder tteameam pproduction.roduction. FromFrom 20052005 onwards,onwards, workersworkers werewere organizedorganized inin teamsteams ofof fi ve.ve. WorkersWorkers wwereere allowedallowed toto choosechoose thethe compositioncomposition ofof theirtheir teamteam asas longlong asas allall fi veve membersmembers aagreed.greed. ComparedCompared toto previousprevious seasons,seasons, thethe socialsocial organizationorganization ofof thethe workplaceworkplace iiss tthushus endogenous.endogenous. InIn Bandiera,Bandiera, Barankay,Barankay, andand RasulRasul (2011),(2011), wewe comparecompare threethree fformsorms ooff tteameam iincentives:ncentives: teamteam ppieceiece rates,rates, rankrank incentivesincentives (namely(namely publiclypublicly pprovidedrovided informationinformation onon eacheach team’steam’s productivity),productivity), andand monetarymonetary prizesprizes assignedassigned ttoo tthehe mmostost pproductiveroductive tteams.eams. TThehe eexperimentxperiment iiss aagaingain ccloselylosely ttiedied ttoo aann uunderlyingnderlying mmodelodel thatthat suggestssuggests twotwo keykey forcesforces thatthat drivedrive teamteam formation:formation: workers’workers’ abilityability andand ssocialocial connections.connections. OnOn oneone hand,hand, workersworkers havehave incentivesincentives toto matchmatch byby ability.ability. OnOn tthehe otherother hand,hand, workersworkers mightmight preferprefer toto formform teamsteams withwith friends,friends, bothboth becausebecause tthishis mmightight llimitimit ffree-ridingree-riding wwithinithin tteamseams aandnd aalsolso bbecauseecause ttheyhey eenjoynjoy nnonpecuniaryonpecuniary bbenefienefi tsts fromfrom interactinginteracting withwith coworkerscoworkers toto whomwhom theythey areare sociallysocially connected.connected. OurOur eexperimentxperiment isis ddesignedesigned toto eexogenouslyxogenously alteralter thethe trade-offtrade-off toto sortingsorting intointo teamsteams byby ffriendshipriendship rrelativeelative ttoo aability.bility. WWee sshowhow thatthat strengtheningstrengthening incentives,incentives, eithereither throughthrough rankrank incentivesincentives oror mmonetaryonetary pprizes,rizes, mmakesakes wworkersorkers moremore likelylikely toto formform teamsteams withwith othersothers ofof similarsimilar aabilitybility insteadinstead ofof withwith theirtheir friends.friends. However,However, rankrank incentivesincentives andand monetarymonetary prizesprizes hhaveave oppositeopposite effectseffects onon averageaverage productivity:productivity: rankrank incentivesincentives signifisignifi cantlycantly reducereduce iitt bbyy 1414 percent,percent, whilewhile monetarymonetary prizesprizes signifisignifi cantlycantly increaseincrease itit byby 2424 percent.percent. BothBoth eeffectsffects aarere hheterogeneous:eterogeneous: rankrank incentivesincentives onlyonly reducereduce thethe productivityproductivity ofof teamsteams atat tthehe bbottomottom ooff tthehe pproductivityroductivity distribution,distribution, andand monetarymonetary prizesprizes onlyonly increaseincrease thethe pproductivityroductivity ofof teamsteams aatt tthehe ttop.op. FFocusingocusing oonn tthehe tteamseams tthathat rremainemain iintactntact aafterfter eeachach cchangehange inin iincentives,ncentives, wewe showshow thatthat thethe documenteddocumented negativenegative effecteffect ofof rankrank incen-incen- ttivesives isis primarilyprimarily duedue toto thethe endogenousendogenous changeschanges inin teamteam composition,composition, ratherrather thanthan cchangeshanges iinn bbehaviorehavior ooff tthehe ssameame tteam.eam. IInn ccontrast,ontrast, tthehe pprovisionrovision ooff mmonetaryonetary pprizesrizes aaffectsffects fi rmrm performanceperformance throughthrough bothboth thethe endogenousendogenous changeschanges inin teamteam composi-composi- ttionion aandnd cchangeshanges iinn bbehaviorehavior wwithinithin tthehe ssameame tteam.eam. TTakenaken together,together, thisthis setset ofof fi eldeld experimentsexperiments yieldsyields severalseveral lessons.lessons. SocialSocial cconnectionsonnections cancan ddriverive bbehaviorehavior iinn tthehe wworkplace:orkplace: wworkersorkers aandnd mmanagersanagers iinternalizenternalize tthehe effecteffect ofof ttheirheir eeffortffort oonn ccolleaguesolleagues toto wwhomhom ttheyhey aarere ssociallyocially cconnected.onnected. More-More- oover,ver, socialsocial connectionsconnections andand monetarymonetary incentivesincentives interact,interact, andand thethe extentextent toto whichwhich aagentsgents internalizeinternalize thethe externalityexternality dependsdepends onon thethe strengthstrength ofof monetarymonetary incentives.incentives. FFirmsirms sshouldhould taketake thesethese potentialpotential interactionsinteractions intointo account.account. Indeed,Indeed, suchsuch differ-differ- eencesnces iinn tthehe ssocialocial oorganizationrganization ofof workplacesworkplaces mightmight helphelp explainexplain somesome partpart ofof Field Experiments with Firms 71

tthehe pproductivityroductivity ddifferencesifferences inin ootherwisetherwise oobservationallybservationally ssimilarimilar fi rms.rms. ThisThis setset ofof eexperimentsxperiments alsoalso illustratesillustrates thethe advantagesadvantages ofof combiningcombining thethe variationvariation exogenouslyexogenously ccreatedreated byby thethe fi eldeld experimenterexperimenter (the(the incentiveincentive scheme)scheme) withwith otherother sourcessources ofof vvariationariation thatthat occuroccur naturallynaturally inin a real-worldreal-world environmentenvironment (social(social ties,ties, monitoringmonitoring ppossibilities,ossibilities, peerpeer groups).groups). ThisThis combinationcombination ofof experimentalexperimental rigorrigor andand collectingcollecting pprimaryrimary ddataata iiss pperhapserhaps tthehe mmostost aattractivettractive ffeatureeature ooff fi eldeld experiments.experiments. TThehe fi ndingsndings openopen upup newnew questionsquestions forfor theoreticaltheoretical researchresearch onon organizations.organizations. FForor instance,instance, relativerelative incentivesincentives ledled toto lowerlower productivityproductivity becausebecause workersworkers internal-internal- iizedzed thethe nnegativeegative externalityexternality toto somesome eextent.xtent. TThishis fi ndingnding speaksspeaks directlydirectly toto Lazear’sLazear’s ((1989)1989) observationobservation thatthat workersworkers areare rarelyrarely compensatedcompensated accordingaccording toto rank-orderrank-order ttournaments,ournaments, aandnd iitt ppointsoints ttoo nnewew aandnd iinterestingnteresting ddirectionsirections fforor tthehe ddevelopmentevelopment ooff ttheoryheory cconcerningoncerning thethe optimaloptimal provisionprovision ofof incentivesincentives underunder more-robustmore-robust assump-assump- ttionsions onon workerworker preferences.preferences. MoreMore broadly,broadly, thethe fi ndingsndings raiseraise thethe issueissue ofof whetherwhether iincentivencentive policiespolicies areare indeedindeed chosenchosen optimally,optimally, oror whetherwhether fi rmsrms areare effectivelyeffectively wwithinithin tthehe eeffiffi ciencyciency frontier.frontier. AsAs wewe shallshall see,see, otherother fi eldeld experimentsexperiments alsoalso castcast doubtdoubt oonn whetherwhether fi rmsrms makemake optimaloptimal choices.choices. TheThe concludingconcluding sectionsection willwill bringbring thesethese ttogetherogether aandnd ddiscussiscuss implicationsimplications fforor ffutureuture rresearch.esearch.

NNewew TTopicsopics iinn WWithin-Firmithin-Firm FFieldield EExperimentsxperiments A vigorousvigorous literatureliterature basedbased onon within-fiwithin-fi rmrm fi eldeld experimentsexperiments isis beginningbeginning toto eemerge.merge. SomeSome ofof thethe topicstopics focusfocus onon organizationalorganizational featuresfeatures ofof fi rms,rms, extendingextending tthehe kkindind ooff aapproachespproaches ddiscussediscussed aabove.bove. FForor eexample,xample, nnewew fi eldeld experimentsexperiments areare bbeingeing designeddesigned toto testtest alternativealternative waysways toto motivatemotivate employees.employees. A naturalnatural candidatecandidate iiss nonmonetarynonmonetary incentivesincentives inin thethe formform ofof statusstatus oror socialsocial recognitionrecognition rewards,rewards, suchsuch aass ““employeeemployee ooff tthehe mmonth”onth” jjobob ttitles.itles. TThehe nnotionotion tthathat iindividualsndividuals ccraverave sstatustatus hhasas llongong beenbeen studiedstudied andand moremore recentlyrecently formalizedformalized (Moldovanu,(Moldovanu, Sela,Sela, andand Shi,Shi, 2007;2007; BBesleyesley andand Ghatak,Ghatak, 2008).2008). InIn a fi eldeld experimentexperiment runrun inin collaborationcollaboration withwith a publicpublic hhealthealth organization,organization, Ashraf,Ashraf, Bandiera,Bandiera, andand JackJack (2011)(2011) randomizerandomize 800800 commu-commu- nnityity agentsagents hiredhired toto sellsell condomscondoms inin urbanurban compoundscompounds intointo fourfour monetarymonetary andand nnonmonetaryonmonetary rrewardsewards treatments.treatments. AgentsAgents whowho areare assignedassigned toto thethe nonmonetarynonmonetary rrewardsewards treatment—namely,treatment—namely, sstarstars fforor pperformanceerformance plusplus a publicpublic ceremonyceremony forfor ttopop pperformers—sellerformers—sell ttwicewice aass mmanyany ccondomsondoms aass aagentsgents wwhoho aarere oofferedffered a fi nancialnancial mmarginargin oonn eeachach ppackack ssold.old. AAnothernother ppersonnelersonnel ppolicyolicy tthathat iiss bbeingeing ssubjectubject ttoo eexperimentalxperimental sscrutinycrutiny iiss tthehe pprovisionrovision ooff pperformanceerformance ffeedback.eedback. IInn a rrecentecent fi eeldld eexperimentxperiment conductedconducted withwith 333030 eemployeesmployees rrecruitedecruited vviaia MMechanicalechanical TTurk,urk, a pplatformlatform rrunun bbyy AAmazon.commazon.com fforor wworkork ssubmittedubmitted oonline,nline, BBarankayarankay ((2010)2010) fi ndsnds thatthat thethe provisionprovision ofof indi-indi- vvidualidual pperformanceerformance ffeedbackeedback aaboutbout rrelativeelative pperformanceerformance rreduceseduces tthehe pproductivityroductivity ooff wworkers.orkers. DDespiteespite rapidrapid progressprogress inin thesethese areas,areas, evidenceevidence onon otherother keykey organizationalorganizational ffeatureseatures iiss sstilltill llaggingagging bbehindehind ttheory.heory. TThehe bbestest eexamplexample iiss pperhapserhaps tthehe ddistributionistribution ooff aauthorityuthority withinwithin fi rms,rms, whichwhich hashas beenbeen atat thethe corecore ofof theoreticaltheoretical studiesstudies ofof thethe fi rmsrms ssinceince Coase’sCoase’s (1937)(1937) seminalseminal contribution.contribution. ModernModern theoreticaltheoretical workwork high-high- llightsights thethe rolerole ofof authorityauthority asas a determinantdeterminant ofof incentivesincentives (Aghion(Aghion andand Tirole,Tirole, 1997)1997) 72 Journal of Economic Perspectives

oorr asas a coordinationcoordination devicedevice (Garicano,(Garicano, 2000;2000; Alonso,Alonso, Dessein,Dessein, andand Matouschek,Matouschek, 22008).008). WWee eenvisagenvisage ffutureuture fi eldeld experimentsexperiments designeddesigned toto shedshed lightlight onon thethe effecteffect ofof oorganizationalrganizational ddesignesign oonn fi rmrm performance.performance.5

FFieldield EExperimentsxperiments acrossacross FirmsFirms

FFieldield eexperimentsxperiments tthathat ttakeake tthehe fi rmrm asas thethe unitunit ofof observationobservation oftenoften seekseek toto eexogenouslyxogenously varyvary thethe availabilityavailability ofof keykey inputsinputs andand inin thisthis wayway seekseek toto uncoveruncover thethe cconstraintsonstraints facedfaced byby fi rms.rms. ManyMany ofof thethe experimentsexperiments reviewedreviewed inin thisthis sectionsection areare iimplementedmplemented iinn ddevelopingeveloping countriescountries fromfrom SouthSouth AsiaAsia toto LatinLatin America,America, bothboth bbecauseecause identifyingidentifying thethe constraintsconstraints facedfaced byby thesethese fi rmsrms isis keykey toto understandingunderstanding thethe ddevelopmentevelopment processprocess andand because,because, inin practicalpractical terms,terms, itit isis cheapercheaper toto createcreate sizablesizable sshockshocks ttoo iinputsnputs wwhenhen tthehe vvaluealue ooff a fi rms’rms’ stockstock ofof inputsinputs isis small.small.

PPhysicalhysical CCapitalapital aandnd AccessAccess toto FFinanceinance A fi eldeld experimentexperiment byby dede Mel,Mel, McKenzie,McKenzie, andand WoodruffWoodruff (2008)(2008) amongamong smallsmall aandnd mediummedium eenterprisesnterprises inin SriSri LankaLanka illustratesillustrates howhow fi eldeld experimentsexperiments cancan illumi-illumi- nnateate tthehe llong-standingong-standing questionquestion ofof howhow creditcredit marketmarket imperfectionsimperfections andand liquidityliquidity cconstraintsonstraints maymay affectaffect fi rmrm growth.growth. TheyThey samplesample 408408 enterprises,enterprises, equallyequally splitsplit bbetweenetween retailretail salessales andand mmanufacturing/services,anufacturing/services, tthathat hhaveave llessess tthanhan $$1,0001,000 iinvestednvested iinn capital.capital. AroundAround halfhalf werewere randomlyrandomly assignedassigned toto receivereceive oneone ofof fourfour treatments:treatments: $$100100 iinn ccash,ash, $$200200 iinn ccash,ash, $$100100 iinn iinventoriesnventories oorr eequipment,quipment, oorr $$200200 iinn iinventoriesnventories oorr eequipment—quipment— iinn tthehe llastast ttwowo ttreatments,reatments, aass cchosenhosen bbyy tthehe fi rmrm owner.owner. TheseThese trans-trans- ffersers aarere llargearge ccomparedompared ttoo tthehe eexistingxisting ccapitalapital sstocktock aandnd mmedianedian mmonthlyonthly profiprofi ts.ts. TThehe experimentexperiment waswas accompaniedaccompanied byby a quarterlyquarterly panelpanel surveysurvey onon investmentinvestment deci-deci- ssions,ions, pprofirofi ts,ts, andand alsoalso personalpersonal characteristicscharacteristics ofof thethe ownerowner suchsuch asas wealth,wealth, riskrisk aaversion,version, andand cognitivecognitive ability.ability. TheThe experimentexperiment waswas framedframed asas randomrandom compensa-compensa- ttionion fforor pparticipationarticipation inin thethe survey.survey. HenceHence ownersowners diddid notnot knowknow aboutabout thethe existenceexistence oorr sscopecope ooff tthehe eexperiment.xperiment. TThehe eexperimentxperiment yyieldsields tthreehree kkeyey fi nndings.dings. FFirst,irst, bbothoth ttypesypes ooff ttransfersransfers iincreasencrease ccapitalapital sstocktock aandnd pprofirofi ts.ts. TTransfersransfers aalsolso iincreasencrease tthehe hhoursours wworkedorked bbyy tthehe oowner,wner, iindicatingndicating ccomplementarityomplementarity bbetweenetween ccapitalapital aandnd llaborabor hhours.ours. SSecond,econd, tthehe rreturneturn ttoo tthehe aadditionaldditional ccapitalapital iiss aaroundround 55–6–6 ppercentercent pperer mmonth,onth, ggivingiving a rrealeal aannualnnual rreturneturn wwellell aabovebove tthehe mmarketarket llendingending rrate.ate. TThird,hird, tthehe aauthorsuthors ddevelopevelop a mmodelodel tthathat mmakesakes

5 Nonexperimental evidence on the determinants of the distribution of authority and its effects on fi rm performance is also being established. Bloom, Sadun, and Van Reenen (2010) develop a survey instrument to measure the level of centralization of decision making in 4,000 fi rms across 12 countries, which shows that highly centralized structures are more likely to occur in developing countries and where product market competition is low. Wu (2011) exploits a natural experiment and detailed personnel data from a Chinese newspaper to provide evidence on the effect of centralizing decision-making authority on the effort and performance of managerial editors and reporters. In his setting, centralizing authority reduces the effort of managers and increases the effort of workers, overall increasing the quality of the average article. More importantly, authority has a signifi cant impact on incentives despite the fact that reporters were already on high-powered performance pay. Oriana Bandiera, Iwan Barankay, and Imran Rasul 73

ppreciserecise hhowow mmissingissing ccreditredit oorr iinsurancensurance mmarketsarkets ccanan ggenerateenerate tthehe oobservedbserved ddiscrep-iscrep- aancyncy betweenbetween returnsreturns toto capitalcapital andand lendinglending rates.rates. TheThe balancebalance ofof evidenceevidence indicatesindicates tthathat rresultsesults aarere ddrivenriven bbyy mmissingissing ccreditredit mmarkets,arkets, nnotot bbyy rriskisk aaversionversion ttoo bborrowing.orrowing. SSeveraleveral methodologicalmethodological pointspoints areare ofof note.note. First,First, usingusing GPSGPS coordinates,coordinates, thethe aauthorsuthors showshow thatthat thethe treatmenttreatment hashas spilloverspillover effectseffects onon nearbynearby fi rms.rms. Interestingly,Interestingly, tthehe aauthorsuthors sshowhow tthathat sspilloverspillovers aarere eentirelyntirely ddrivenriven bbyy fi rmsrms inin thethe bamboobamboo industry,industry, wwherehere tthehe hharvestingarvesting ooff bbambooamboo iiss ssubjectubject ttoo ggovernmentovernment rrestrictionsestrictions aandnd ttreatedreated fi rmsrms ccrowdrowd outout othersothers byby purchasingpurchasing allall ofof thethe availableavailable supply.supply. TheseThese fi ndingsndings illus-illus- ttraterate tthehe ppotentialotential ooff fi eldeld experimentsexperiments toto shedshed lightlight oonn tthehe ffunctioningunctioning ooff mmarkets,arkets, nnotot jjustust fi rmsrms inin isolation.isolation. Second,Second, thethe authorsauthors showshow howhow toto dealdeal withwith anan attritionattrition rrateate tthathat wwasas 5 ppercentercent hhigherigher fforor ccontrolontrol fi rmsrms thanthan forfor treatmenttreatment fi rmsrms byby usingusing thethe mmethodologyethodology pproposedroposed bbyy LLeeee ((2009),2009), wwhichhich eestimatesstimates uupperpper aandnd llowerower bboundsounds fforor tthehe treatmenttreatment effecteffect andand allowsallows themthem ttoo sshowhow tthathat tthehe eestimatedstimated ttreatmentreatment effectseffects aandnd rreturneturn toto capitalcapital areare robustrobust toto attrition.attrition. Third,Third, thethe authorsauthors alsoalso comparecompare thethe ddifferenceifference betweenbetween experimentalexperimental andand nonexperimentalnonexperimental methods.methods. ComparedCompared toto a 55.3.3 ppercentercent mmonthlyonthly rreturneturn ttoo ccapitalapital eestimatedstimated vviaia eexperimentalxperimental mmethods,ethods, oordinaryrdinary lleasteast squares,squares, randomrandom effects,effects, andand fi rmrm fi xedxed effectseffects modelsmodels yieldyield estimatesestimates ofof 2.6,2.6, 11.7,.7, aandnd 00.07.07 ppercent,ercent, rrespectively.espectively. MMcKenziecKenzie andand WoodruffWoodruff (2008)(2008) implementedimplemented thethe samesame experimentexperiment withwith 113737 ssmallmall aandnd mmedium-sizededium-sized rretailetail fi rmsrms inin Mexico,Mexico, andand alsoalso fi ndnd highhigh ratesrates ofof rreturnseturns toto capital,capital, especiallyespecially forfor fi rmsrms thatthat reportreport beingbeing creditcredit constrained/notconstrained/not hhavingaving accessaccess toto formalformal loans.loans. Overall,Overall, thesethese fi eldeld experimentsexperiments provideprovide a usefuluseful aanswernswer toto anan importantimportant question,question, asas wellwell asas providingproviding guidanceguidance forfor futurefuture fi eldeld eexperiments.xperiments. InIn particularparticular theythey illustrateillustrate howhow experimentsexperiments cancan gogo beyondbeyond under-under- sstandingtanding thethe decisiondecision processprocess ofof a singlesingle fi rrmm aandnd pproviderovide eevidencevidence onon thethe naturenature ooff sspilloverspillovers aandnd mmarketarket iinteractions.nteractions.

MManagerialanagerial CCapitalapital aandnd EEnterprisenterprise TTrainingraining EEconomicconomic ttheoryheory hhasas llongong ttaughtaught tthathat mmanagerialanagerial ccapitalapital iiss aann iimportantmportant iinputnput iinn pproductionroduction (Lucas,(Lucas, 1978;1978; Rosen,Rosen, 1982).1982). However,However, systematicsystematic empiricalempirical evidenceevidence aaboutbout thethe effectseffects ofof managerialmanagerial capitalcapital was,was, untiluntil recently,recently, almostalmost nonexistent—innonexistent—in ppartart bbecauseecause ofof thethe diffidiffi cultiesculties inin measuringmeasuring managerialmanagerial capital.capital. InIn addition,addition, mana-mana- ggerialerial capitalcapital seemsseems likelylikely toto bebe relatedrelated toto characteristicscharacteristics ofof thethe fi rmrm andand perhapsperhaps ttoo uunobservablenobservable traitstraits aboutabout managementmanagement andand workersworkers inin waysways thatthat mademade itit veryvery ddiffiiffi ccultult ttoo ddrawraw ccausalausal inferences.inferences. TTwowo recentrecent methodologicalmethodological improvementsimprovements havehave ledled toto substantialsubstantial progressprogress inin tthishis area.area. First,First, somesome sstudiestudies havehave focusedfocused onon managementmanagement practicespractices thatthat cancan bebe mmeasuredeasured systematically.systematically. AsAs theythey discusseddiscussed inin thisthis journal,journal, BloomBloom andand VanVan ReenenReenen ((2010b)2010b) havehave developeddeveloped a standardstandard measuremeasure ofof managementmanagement practicespractices andand usedused itit ttoo ssurveyurvey mmanagersanagers iinn 66,000,000 mmanufacturinganufacturing fi rmsrms inin 1717 developeddeveloped andand developingdeveloping ccountries.ountries. TheirTheir workwork showsshows a robustrobust correlationcorrelation betweenbetween thethe qualityquality ofof manage-manage- mmentent ppracticesractices andand fi rmrm performanceperformance bothboth acrossacross andand withinwithin countries.countries. Second,Second, rresearchersesearchers havehave begunbegun designingdesigning fi eldeld experimentsexperiments aimedaimed atat evaluatingevaluating thethe causalcausal iimpactmpact ooff mmanagerialanagerial ppracticesractices oonn fi rmrm performance.performance. 74 Journal of Economic Perspectives

TToo evaluateevaluate thethe effecteffect ofof businessbusiness trainingtraining onon thethe performanceperformance ofof microenter-microenter- pprises,rises, KarlanKarlan andand VValdiviaaldivia ((2011)2011) aandnd DDrexler,rexler, FFischer,ischer, aandnd SSchoarchoar ((2011)2011) eexaminexamine tthehe experienceexperience ofof clientsclients ofof a microfimicrofi nancenance institutioninstitution inin PeruPeru andand thethe DominicanDominican RRepublic,epublic, respectively.respectively. BothBoth studiesstudies randomizerandomize microentrepreneursmicroentrepreneurs intointo a treat-treat- mmentent ggrouproup tthathat rreceiveseceives fi nancialnancial training,training, includingincluding basicbasic accounting,accounting, marketing,marketing, aandnd pricing,pricing, andand intointo a controlcontrol groupgroup thatthat doesdoes not.not. CommonCommon fi ndingsndings emergeemerge ffromrom tthesehese sstudies.tudies. NNeithereither fi ndsnds anan effecteffect ofof businessbusiness trainingtraining onon sales,sales, profiprofi ts,ts, oorr eemployment.mployment. BBothoth fi ndnd thatthat trainingtraining reducesreduces businessbusiness seasonalityseasonality byby increasingincreasing ssalesales iinn ““bad”bad” months.months. TheseThese areare intriguingintriguing resultsresults andand shouldshould leadlead toto signifisignifi cantcant aadvancesdvances iinn uunderstandingnderstanding tthehe cconstraintsonstraints ffacedaced byby microenterprises.microenterprises. BBloom,loom, EEifert,ifert, Mahajan,Mahajan, McKenzie,McKenzie, andand RobertsRoberts (2011)(2011) developdevelop a fi eldeld experi-experi- mmentent toto evaluateevaluate thethe effecteffect ofof modernmodern managementmanagement practicespractices onon thethe performanceperformance ooff llargearge IIndianndian fi rmsrms inin tthehe ttextileextile iindustry.ndustry. WWorkingorking iinn ccollaborationollaboration withwith a lleadingeading iinternationalnternational managementmanagement cconsultingonsulting fi rm,rm, thethe researchersresearchers offeredoffered freefree manage-manage- mmentent cconsultingonsulting toto a randomlyrandomly selectedselected groupgroup ofof 1414 outout ofof 2020 plantsplants belongingbelonging toto 1177 largelarge manufacturers.manufacturers. A furtherfurther eighteight plantsplants belongingbelonging toto thethe samesame fi rmsrms werewere aalsolso surveyed.surveyed. TheThe consultingconsulting interventionintervention targetedtargeted 3838 keykey practicespractices thatthat capturecapture sstandardtandard mmanufacturinganufacturing pprinciplesrinciples iinn hhigh-incomeigh-income ccountries.ountries. TThehe designdesign ofof thisthis fi eldeld experimentexperiment illustratesillustrates thethe trade-offtrade-off betweenbetween samplesample ssizeize aandnd tthehe ccomplexityomplexity ooff tthehe iinterventionntervention nneededeeded ttoo sstudytudy a ccross-sectionross-section ooff llargearge fi rms.rms. ConsultancyConsultancy andand datadata collectioncollection costs—$75,000costs—$75,000 perper treatedtreated fi rm,rm, $$20,00020,000 perper controlcontrol plant—limitedplant—limited thethe samplesample sizesize toto 2020 units.units. StandardStandard statisticalstatistical ttestsests tthathat rrelyely oonn aasymptoticsymptotic ppropertiesroperties ccannotannot bbee uusedsed iinn tthishis ccontextontext bbecauseecause tthehe numbernumber ofof observationsobservations isis tootoo small.small. CollectingCollecting datadata overover a longlong timetime horizonhorizon ppartiallyartially hhelpselps aass ttherehere aarere pproceduresrocedures tthathat rrelyely oonn aasymptoticsymptotic aapproximationspproximations aalonglong tthehe ttimeime ddimensionimension ((IbragimovIbragimov aandnd MMüller,üller, 2010).2010). InIn addition,addition, statis-statis- tticalical ppowerower ccanan bbee ssustainedustained iinn ssuchuch ssmallmall ssamplesamples bbyy ccollectingollecting ddataata ddirectlyirectly ffromrom machinemachine logs,logs, focusingfocusing onon similarsimilar fi rmsrms usingusing identicalidentical technologies,technologies, andand ccollectingollecting high-frequencyhigh-frequency repeatedrepeated measuresmeasures (McKenzie,(McKenzie, 2011).2011). TheThe keykey fi ndingnding iiss thatthat managerialmanagerial capitalcapital improvedimproved qualityquality andand effieffi ciency,ciency, reducedreduced inventory,inventory, andand rraisedaised averageaverage productivityproductivity byby 1111 percent.percent. TheThe resultingresulting increaseincrease inin yearlyyearly profiprofi tsts wwasas estimatedestimated toto bebe overover 9090 percentpercent ofof thethe marketmarket costcost ofof thethe consultingconsulting servicesservices tthathat fi rmsrms wouldwould havehave paidpaid inin thethe fi rstrst year.year. AAgain,gain, ssuchuch fi nndingsdings rraiseaise tthehe oobviousbvious qquestionsuestions ooff wwhyhy pprofirofi tabletable practicespractices wwereere notnot adoptedadopted beforebefore thethe intervention.intervention. InterviewsInterviews withwith ownersowners andand seniorsenior mmanagersanagers revealreveal thatthat incorrectincorrect beliefsbeliefs aboutabout thethe profiprofi ttabilityability ofof thethe practicespractices wwereere thethe mainmain causecause ofof non-adoption,non-adoption, butbut eveneven whenwhen thesethese incorrectincorrect beliefsbeliefs werewere ppointedointed out,out, fi rmsrms werewere veryvery slowslow toto adoptadopt thethe newnew practices.practices. MostMost ofof thethe ownersowners aattributedttributed thisthis toto lacklack ofof time.time. However,However, itit alsoalso seemedseemed thatthat mostmost IndianIndian textiletextile fi rrmsms diddid notnot needneed toto adoptadopt thethe newnew practicespractices inin orderorder toto competecompete withwith otherother ddomesticomestic fi rmsrms asas highhigh tariffstariffs shieldedshielded themthem fromfrom internationalinternational competition.competition. TheThe sstudytudy thusthus raisesraises whatwhat seemsseems toto bebe a recurrentrecurrent questionquestion inin fi eldeld experimentsexperiments withwith fi rrms:ms: WhatWhat areare thethe constraintsconstraints thatthat discouragediscourage fi rrmsms fromfrom makingmaking changeschanges thatthat cclearlylearly seemseem toto bebe profiprofi table?table? Field Experiments with Firms 75

LLaborabor AAnn alternativealternative kindkind ofof between-fibetween-fi rmsrms study,study, moremore commoncommon inin high-incomehigh-income ccountries,ountries, areare fi eldeld studiesstudies designeddesigned toto measuremeasure discriminatorydiscriminatory practicespractices byby exog-exog- eenouslynously varyingvarying thethe aapplicantpplicant poolpool availableavailable toto fi rms.rms. TheThe typicaltypical “audit”“audit” studystudy ppresentsresents employersemployers withwith twotwo setssets ofof jjobob aapplicantspplicants whowho areare identicalidentical alongalong allall rrelevantelevant employmentemployment characteristicscharacteristics exceptexcept thethe characteristiccharacteristic ofof interest,interest, likelike race,race, ggender,ender, oorr aagege ((HeckmanHeckman aandnd SSiegelman,iegelman, 11993;993; RRiachiach aandnd RRich,ich, 22002).002). BBertrandertrand aandnd MMullainathanullainathan ((2004)2004) ddesignesign a fi eldeld experimentexperiment alongalong thesethese lineslines bbyy ssendingending rrésumésésumés wwithith rrandomlyandomly aassignedssigned wwhite-hite- oorr bblack-soundinglack-sounding nnamesames ttoo ooverver 11,300,300 hhelp-wantedelp-wanted adsads iinn BBostonoston aandnd CChicagohicago nnewspapers.ewspapers. TTheyhey aalsolso rrandomlyandomly vvaryary tthehe qqualityuality ofof thethe résumérésumé byby addingadding experience,experience, skills,skills, oror honors.honors. TheyThey fi ndnd thatthat wwhitehite namesnames receivereceive 5050 percentpercent moremore callbackscallbacks forfor anan interviewinterview thanthan blackblack names.names. TThishis racialracial gapgap iiss uuniformniform aacrosscross ooccupation,ccupation, iindustry,ndustry, aandnd eemployermployer ssize.ize. AAddition-ddition- aally,lly, tthehe returnreturn toto higher-qualityhigher-quality résumésrésumés isis higherhigher forfor whiteswhites thanthan blacks,blacks, implyingimplying tthathat tthehe rracialacial ggapap isis largerlarger forfor more-qualifimore-qualifi eded applicants.applicants. TThehe authorsauthors fi ndingsndings areare consistentconsistent withwith a modelmodel ofof lexicographiclexicographic searchsearch wwherebyhereby tthehe eemployersmployers sstoptop rreadingeading ((andand hhenceence ffailail ttoo sseeee aallll ccredentials)redentials) ooncence ttheyhey sseeee a blackblack name.name. WhileWhile suchsuch practicepractice iiss iindeedndeed cconsistentonsistent wwithith tthehe fi ndings,ndings, moremore wworkork iiss nneededeeded ttoo pproviderovide ddirectirect eevidencevidence oonn iitsts rrelevanceelevance ccomparedompared ttoo sstatisticaltatistical aandnd ttaste-basedaste-based ddiscriminationiscrimination mmodels.odels. IIff tthehe ssearchearch pprocessrocess iiss llexicographic,exicographic, ffutureuture wworkork nneedseeds ttoo uunderstandnderstand wwhyhy ssuchuch mmethodsethods eemergedmerged iinn tthehe fi rstrst place.place. TimeTime cconstraintsonstraints aarere oonene ppossibleossible iingredientngredient hhere;ere; iindeed,ndeed, aalthoughlthough ttimeime cconstraintsonstraints aarere rrarelyarely mmadeade eexplicitxplicit iinn eeconomicconomic mmodelingodeling ooff fi rms,rms, ttheyhey mmayay hhelpelp ttoo eexplainxplain eexperimentalxperimental rresultsesults iinn mmanyany ddiverseiverse ccontexts.ontexts. MMoreore ggenerally,enerally, tthesehese ttypesypes ooff aauditudit eexperimentsxperiments ccouldould bbee ffruitfullyruitfully aappliedpplied ttoo aanalyzenalyze tthehe ppracticalractical rrelevanceelevance ooff ootherther aapplicantpplicant ttraitsraits ssuchuch aass ggender,ender, eeducation,ducation, aandnd ppastast eemploymentmployment hhistory.istory. TThishis wwouldould aallowllow oonene ttoo pproviderovide a cconsistentonsistent ppictureicture ooff fi rms’rms’ preferencespreferences andand constraintsconstraints onon hiringhiring decisions.decisions.

PPracticalractical ConsiderationsConsiderations

TThehe ggrowingrowing literatureliterature usingusing fi eldeld experimentsexperiments onon fi rmsrms hashas begunbegun toto provideprovide iinsightsnsights onon long-standinglong-standing areasareas ofof economiceconomic researchresearch relatedrelated toto fi rmrm behavior.behavior. InIn tthehe hopehope thatthat somesome rreaderseaders mmayay bbee iinterestednterested iinn uundertakingndertaking tthishis kkindind ooff rresearch,esearch, iinn tthishis ssectionection wwee ddiscussiscuss ttwowo ppracticalractical cconsiderationsonsiderations tthathat aariserise iinn ddoingoing ssuchuch eexper-xper- iiments:ments: ddesignesign aandnd eethicalthical issues.issues. InIn tthehe fi nalnal discussion,discussion, wewe thenthen drawdraw togethertogether ssomeome ccommonommon llessonsessons ffromrom tthehe sstudiestudies ppresentedresented ttoo hhighlightighlight ssomeome aareasreas ttoo wwhichhich wwee tthinkhink ffutureuture rresearchesearch sshouldhould bebe directed.directed.

DDesignesign TThehe designdesign ofof a fi eldeld experimentexperiment oughtought toto bebe groundedgrounded inin economiceconomic theorytheory soso tthathat nnullull aandnd aalternativelternative hhypothesesypotheses ccanan bbee mmappedapped bbackack ttoo aann uunderlyingnderlying mmodelodel ooff fi rmrm behavior.behavior. OnceOnce nullnull andand alternativealternative hypotheseshypotheses areare wellwell defidefi ned,ned, thisthis willwill guideguide tthehe ccollectionollection ooff pprimaryrimary aandnd ssecondaryecondary ddata.ata. 76 Journal of Economic Perspectives

RResearchersesearchers thenthen needneed toto choosechoose thethe unitunit overover whichwhich toto introduceintroduce experi-experi- mmentalental vvariation:ariation: fforor fi eldeld experimentsexperiments involvinginvolving fi rms,rms, thisthis cancan involveinvolve engineeringengineering eexperimentalxperimental variationvariation acrossacross fi rmsrms oror withinwithin a fi rm.rm. ForFor thethe former,former, thethe unitunit mightmight bbee iinputs,nputs, wwhosehose ccharacteristicsharacteristics areare exogenouslyexogenously variedvaried acrossacross fi rms,rms, oror thethe unitunit mmightight bebe fi rmsrms themselvesthemselves whowho areare thenthen exogenouslyexogenously assignedassigned toto differentdifferent environ-environ- mments.ents. ThisThis latterlatter typetype ofof fi eldeld experimentexperiment designdesign remainsremains scarce,scarce, butbut wewe discussdiscuss ppotentialotential developmentsdevelopments below.below. ForFor experimentationexperimentation withinwithin fi rms,rms, thethe unitunit isis mostmost ooftenften wworkers,orkers, aalthoughlthough pplantslants oorr fi rmrm divisionsdivisions mightmight alsoalso bebe used.used. ForFor experimen-experimen- ttationation atat thethe workerworker level,level, atat oneone endend ofof thethe spectrum,spectrum, workersworkers areare simultaneouslysimultaneously rrandomlyandomly aassignedssigned toto controlcontrol andand treatmenttreatment groups,groups, asas isis ttypicallyypically donedone inin ppolicyolicy eevaluationvaluation andand inin randomizedrandomized controlledcontrolled trials.trials. AtAt thethe otherother endend ofof thethe spectrum,spectrum, aallll wworkersorkers aarere ttreatedreated bbutut tthehe ttimingiming ooff tthehe ttreatmentreatment iiss eexogenouslyxogenously cchosenhosen bbyy tthehe rresearchers.esearchers. EachEach approachapproach hashas itsits ownown costscosts andand benefibenefi ts,ts, andand thesethese areare likelylikely toto ddifferiffer aacrosscross fi rmrm settings.settings. TThehe mainmain bbenefienefi t ofof havinghaving a controlcontrol groupgroup isis thatthat commoncommon trendstrends cancan bebe wweededeeded outout byby usingusing a difference-in-differencesdifference-in-differences estimator.estimator. TheThe mainmain costcost isis thatthat tthehe estimatedestimated effecteffect ofof thethe interventionintervention mightmight bebe biasedbiased becausebecause thethe controlcontrol groupgroup mmightight rreacteact ttoo not hhavingaving receivedreceived thethe treatment.treatment. WhetherWhether thisthis createscreates a positivepositive oorr nnegativeegative biasbias dependsdepends onon whetherwhether thethe controlcontrol groupgroup triestries toto differentiatedifferentiate them-them- sselveselves from,from, oror toto emulate,emulate, thethe treatedtreated group.group. ThisThis isis a fi rst-orderrst-order issueissue inin fi eldeld eexperimentsxperiments withwith fi rms,rms, wherewhere itit isis oftenoften harderharder toto isolateisolate treatmenttreatment andand controlcontrol ggroupsroups geographicallygeographically oror informationally,informationally, soso thosethose inin a controlcontrol groupgroup areare likelylikely toto fi nndd ooutut aaboutbout thethe experiments.experiments. TheThe threatthreat ofof contaminationcontamination cancan bebe eliminatedeliminated bbyy separatingseparating thethe groups,groups, butbut thisthis typicallytypically causescauses themthem toto bebe subjectsubject toto differentdifferent wworkplaceorkplace conditions,conditions, makingmaking thethe controlcontrol groupgroup a weakerweaker counterfactualcounterfactual forfor whatwhat wwouldould hhaveave hhappenedappened toto tthehe ttreatedreated iinn tthehe aabsencebsence ofof treatment.treatment. SSwitchingwitching allall agentsagents betweenbetween controlcontrol andand treatmenttreatment groupsgroups atat anan exogenouslyexogenously cchosenhosen timetime hhasas tthehe bbenefienefi t ofof eliminatingeliminating thethe contaminationcontamination bias;bias; also,also, toto increaseincrease sstatisticaltatistical power,power, thethe effecteffect ofof treatmenttreatment cancan bebe estimatedestimated byby comparingcomparing eacheach agentagent ttoo tthathat samesame agentagent withoutwithout thethe treatment,treatment, thusthus eliminatingeliminating allall sourcessources ofof unobserv-unobserv- aableble heterogeneityheterogeneity (List,(List, Sadoff,Sadoff, andand Wagner,Wagner, 2010).2010). TheThe costcost ofof thisthis approachapproach isis tthathat tthehe eestimatedstimated eeffectffect ooff tthehe ttreatmentreatment mmightight bbee bbiasediased bbecauseecause ooff uunobserv-nobserv- aableble determinantsdeterminants ofof changeschanges inin behavior.behavior. ThisThis concernconcern mightmight bebe addressedaddressed byby ccollectingollecting a suffisuffi cientlyciently longlong timetime seriesseries duringduring bothboth treatmenttreatment andand controlcontrol periods,periods, oor,r, iiff seasonalityseasonality oror cyclicalitycyclicality isis a potentialpotential threat,threat, byby collectingcollecting informationinformation fromfrom a ddifferentifferent periodperiod duringduring whichwhich nono treatmenttreatment waswas introduced,introduced, toto purgepurge estimatesestimates ofof vvariationariation ddueue ttoo ssuchuch nnaturallyaturally ooccurringccurring fl uctuations.uctuations. SwitchingSwitching allall agentsagents betweenbetween ccontrolontrol andand treatmenttreatment groupsgroups isis a goodgood approachapproach forfor another,another, practicalpractical reason:reason: fi rmsrms oftenoften expressexpress unwillingnessunwillingness toto treattreat similarsimilar workersworkers inin thethe samesame plantplant oror fi rmrm llocationocation iinn ddifferentifferent wways.ays. A rrecentecent fi eldeld experimentexperiment illustratesillustrates howhow differentdifferent randomizationrandomization strategiesstrategies ccanan lleadead ttoo ddifferentifferent eestimates.stimates. SShihi ((2010)2010) ccomparesompares pproductivityroductivity uundernder fi xedxed wageswages aandnd piecepiece ratesrates forfor workersworkers engagedengaged inin treetree thinningthinning inin a fruitfruit orchardorchard inin thethe statestate ooff Washington.Washington. AsAs thethe fi rmrm hashas multiplemultiple sites,sites, inin oneone sitesite sheshe switchedswitched tenten workersworkers Oriana Bandiera, Iwan Barankay, and Imran Rasul 77

ssimultaneouslyimultaneously fromfrom wageswages toto piecepiece ratesrates andand observedobserved themthem forfor threethree daysdays underunder eeachach treatment,treatment, whereaswhereas inin anotheranother sitesite sheshe switchedswitched sevenseven workersworkers toto piecepiece ratesrates aandnd kkeptept aanothernother sseveneven aass ccontrolsontrols wwithith fi xedxed wageswages forfor oneone day,day, identifyingidentifying thethe effecteffect ffromrom tthehe ddifference-in-differencesifference-in-differences betweenbetween thethe twotwo days.days. TheThe estimatedestimated produc-produc- ttivityivity increaseincrease isis 2323 percentpercent inin thethe fi rstrst design—similardesign—similar toto thethe estimatesestimates inin ShearerShearer ((2004)—while2004)—while thethe estimatedestimated effecteffect increasesincreases toto 4343 percentpercent inin thethe secondsecond design.design. SShihi (2010)(2010) reportsreports thatthat workersworkers inin thethe controlcontrol groupgroup learnedlearned ofof thethe existenceexistence ofof a ttreatmentreatment groupgroup andand werewere notnot pleased.pleased. ThisThis couldcould havehave reducedreduced theirtheir productivity,productivity, lleadingeading toto aann ooverestimateverestimate ooff tthehe eeffect,ffect, bbutut hherer ddataata iiss nnotot rrichich eenoughnough ttoo sshedhed llightight oonn tthehe mmechanism.echanism. UUnderstandingnderstanding ssuchuch ccontaminationontamination eeffectsffects iiss ccrucialrucial ttoo bbeingeing aableble ttoo ccompareompare fi ndingsndings acrossacross experiments,experiments, andand suchsuch issuesissues areare beginningbeginning toto bebe eexploredxplored bbyy fi eldeld experimenters.experimenters.

EEthicsthics FFieldield experimentsexperiments involveinvolve humanhuman subjectssubjects andand thusthus typicallytypically fallfall underunder thethe ooversightversight ofof anan institutionalinstitutional reviewreview boardboard runrun byby anan institutioninstitution ofof higherhigher educa-educa- ttionion oror a fundingfunding agency.agency. InIn turn,turn, thesethese boardsboards areare typicallytypically guidedguided byby thethe ethicalethical pprinciplesrinciples setset forthforth byby a UU.S..S. ggovernmentovernment rreporteport ccalledalled tthehe BBelmontelmont RReporteport ((namednamed aafterfter thethe conferenceconference centercenter wherewhere itit waswas drafteddrafted inin 1978).1978). TheThe threethree fundamentalfundamental eethicalthical pprinciplesrinciples iinn tthehe BBelmontelmont RReporteport fforor aallll hhumanuman ssubjectsubjects rresearchesearch aarere rrespectespect fforor ppersons,ersons, bbenefienefi cence,cence, andand justice.justice.6 RecentRecent yearsyears havehave seenseen somesome convergenceconvergence inin iinstitutionalnstitutional rrevieweview bboardoard practicespractices acrossacross institutions.institutions. OOnene qquestionuestion ofof particularparticular interestinterest toto economistseconomists isis whetherwhether allall participantsparticipants iinn anan experimentexperiment mustmust consentconsent inin advance,advance, asas thethe knowledgeknowledge ofof participatingparticipating inin aann experimentexperiment maymay biasbias thethe results.results. A commoncommon rulerule isis thatthat institutionalinstitutional reviewreview bboardsoards mmayay wwaiveaive iinformednformed cconsentonsent aandnd aallowllow tthehe uusese ooff ssuchuch ““deception”deception” if ccertainertain cconditionsonditions areare mmet:et: FFirst,irst, tthehe rresearchesearch iinvolvesnvolves nnoo mmoreore tthanhan mminimalinimal rrisk.isk. SSecond,econd, tthehe wwaiveraiver willwill notnot impactimpact aadverselydversely subjects’subjects’ rightsrights andand welfare,welfare, whichwhich includesincludes nono rreductioneduction inin ccompensation,ompensation, eemploymentmployment bbenefienefi ts,ts, oror mentalmental well-being.well-being. (Moreover,(Moreover, iiff ssubjectsubjects dodo notnot wishwish toto participateparticipate inin thethe researchresearch therethere isis toto bebe nono adverseadverse effecteffect oonn tthemhem eeither.)ither.) TThird,hird, thethe researchresearch couldcould notnot practicallypractically bebe carriedcarried outout withoutwithout tthehe waiver.waiver. Finally,Finally, subjectssubjects willwill bebe providedprovided withwith additionaladditional pertinentpertinent informationinformation aaboutbout pparticipation—aarticipation—a conditioncondition whichwhich cancan oftenoften bebe metmet byby debriefidebriefi ngng subjectssubjects atat tthehe eendnd ooff tthehe sstudy.tudy. TThehe grantinggranting ofof waiverswaivers ofof informedinformed consentconsent variesvaries acrossacross institutionalinstitutional reviewreview bboards.oards. SomeSome simplysimply allowallow thethe waiver.waiver. OtherOther boardsboards requirerequire subjectssubjects toto bebe informedinformed tthathat ttheyhey areare partpart ofof a researchresearch studystudy (rather(rather thanthan anan “experiment”),“experiment”), thatthat theythey cancan ooptpt ooutut ooff tthehe sstudytudy withoutwithout anyany consequence,consequence, andand thatthat theythey areare providedprovided withwith thethe ccontactontact detailsdetails ofof thethe researcher.researcher. ThisThis needneed notnot compromisecompromise thethe conductconduct ofof fi eldeld eexperimentsxperiments wwithinithin fi rmsrms asas longlong asas suchsuch requirementsrequirements areare equallyequally appliedapplied toto treat-treat- mmentent aandnd ccontrolontrol ggroups.roups.

6 The report (National Commission for the Protection of Human Subjects of Biomedical and Behavioral Research, 1979) is located at 〈http://www.hhs.gov/ohrp/humansubjects/guidance/〉. 78 Journal of Economic Perspectives

A ddistinctiveistinctive ccharacteristicharacteristic ooff fi eeldld eexperimentsxperiments wwithith fi rmsrms tthathat hhasas iimportantmportant iimplicationsmplications fforor hhowow fi eldeld experimentsexperiments shouldshould bebe judgedjudged byby institutionalinstitutional reviewreview bboardsoards iiss tthathat fi rmsrms areare likelylikely toto experimentexperiment onon theirtheir own,own, oror bebe advisedadvised byby for-profifor-profi t cconsultancies,onsultancies, aandnd tthehe eethicalthical gguidelinesuidelines ttheyhey mmustust aadheredhere ttoo iinn ssuchuch ccasesases aarere llikelyikely ttoo bbee mmuchuch llessess sstringenttringent tthanhan tthosehose ffacedaced bbyy aacademics.cademics.7 AAcademiccademic rresearchersesearchers mmightight tthenhen bbee crowdedcrowded ooutut bbyy ffor-profior-profi t evaluationevaluation consultanciesconsultancies thatthat areare notnot subjectsubject toto thethe ssameame ethicalethical requirements,requirements, andand thisthis mightmight reducereduce thethe involvementinvolvement ofof academicsacademics inin fi eldeld eexperimentsxperiments aandnd ccreatereate a sstrongtrong sselectionelection aamongmong tthehe ttypeype ooff eexperimentsxperiments tthathat ccanan bbee ssubjectubject ttoo sscientificientifi c sscrutiny.crutiny. TThehe sseverityeverity ooff tthishis cconcernoncern wwillill ddependepend oonn tthehe eevolvingvolving aattitudettitude ooff iinstitutionalnstitutional rrevieweview bboardsoards wwithith rregardegard ttoo aacademics’cademics’ iinvolvementnvolvement wwithith fi rms.rms. ForFor instance,instance, wwillill rresearchersesearchers bbee aallowedllowed ttoo aadvisedvise fi rmsrms onon eexperimentsxperiments tthathat ddoo nnotot mmeeteet aacademiccademic eethicalthical gguidelinesuidelines bbutut wwouldould bbee iimplementedmplemented rregardlessegardless ooff aacademiccademic iinvolvement?nvolvement? WWillill rresearchersesearchers bbee aallowedllowed ttoo aanalyzenalyze ddataata ffromrom eexperi-xperi- mmentsents tthathat ddoo nnotot mmeeteet aacademiccademic eethicalthical gguidelinesuidelines bbutut wwereere iinitiatednitiated bbyy tthehe fi rmsrms tthemselves?hemselves? TThehe aanswersnswers ttoo tthesehese qquestionsuestions wwillill hhaveave pprofoundrofound iimplicationsmplications fforor tthehe ffutureuture ooff fi eldeld eexperimentsxperiments wwithith fi rms.rms.

CCommonommon LessonsLessons andand FutureFuture DirectionsDirections

FFieldield experimentsexperiments areare atat thethe heartheart ofof a growinggrowing empiricalempirical literatureliterature thatthat isis eexpandingxpanding economists’economists’ understandingunderstanding ofof fi rmrm behavior.behavior. InIn thisthis concludingconcluding discus-discus- ssion,ion, wwee ddrawraw ttogetherogether ssomeome ccommonommon llessonsessons ffromrom tthesehese sstudiestudies aandnd ssuggestuggest ssomeome ffutureuture ddirections.irections. A ccommonommon ffeatureeature ofof mostmost fi eldeld experimentsexperiments withwith fi rmsrms isis thatthat theythey identifyidentify ppartialartial equilibriumequilibrium effects,effects, asas theythey typicallytypically affectaffect onlyonly oneone oror atat mostmost a samplesample ofof fi rrmsms wwithinithin oonene iindustry.ndustry. GGeneraleneral eequilibriumquilibrium eeffectsffects mmightight ooff coursecourse differdiffer fromfrom tthehe partialpartial equilibriumequilibrium effects,effects, dependingdepending inin partpart onon whetherwhether thethe returnsreturns toto thethe eexperimentalxperimental innovationinnovation cancan bebe competedcompeted away.away. ForFor instance,instance, ifif high-poweredhigh-powered iincentivesncentives bbenefienefi t thethe fi rmrm byby attractingattracting betterbetter workers,workers, nono fi rmrm cancan gaingain byby offeringoffering hhigh-poweredigh-powered incentivesincentives whenwhen allall fi rmsrms dodo thethe same.same. InIn contrast,contrast, ifif high-poweredhigh-powered iincentivesncentives increaseincrease workers’workers’ productivity,productivity, thenthen allall fi rmsrms cancan benefibenefi t fromfrom offeringoffering hhigh-poweredigh-powered iincentives,ncentives, tthushus iincreasingncreasing aaggregateggregate productivity.productivity. TThehe questionquestion thatthat thenthen naturallynaturally arisesarises is:is: ifif thesethese aggregateaggregate gainsgains areare possible,possible, wwhyhy don’tdon’t fi rmsrms reapreap them?them? Indeed,Indeed, oneone puzzlingpuzzling fi ndingnding isis thatthat almostalmost allall thethe fi eldeld eexperimentsxperiments reviewedreviewed havehave broughtbrought largelarge benefibenefi tsts toto thethe fi rm.rm. InIn part,part, thisthis patternpattern aarisesrises becausebecause fi rrmsms wwouldould notnot agreeagree toto implementimplement experimentsexperiments thatthat areare expectedexpected ttoo hhaveave ddetrimentaletrimental cconsequences!onsequences! YetYet thethe factfact thatthat iinn ssoo mmanyany ccasesases rresearchersesearchers hhaveave managedmanaged toto increaseincrease profiprofi tsts appearsappears atat oddsodds withwith thethe commoncommon assumptionassumption tthathat fi rrmsms aarere ppressuredressured bbyy ccompetitiveompetitive fforcesorces ttoo mmakeake aatt lleasteast cclose-to-optimallose-to-optimal cchoices.hoices. ForFor example,example, thethe incentiveincentive schemesschemes introducedintroduced inin ourour fi eldeld experimentsexperiments

7 We thank the editor, David Autor, for raising this point. Field Experiments with Firms 79

aatt thethe fruit-pickingfruit-picking fi rmrm increasedincreased productivityproductivity andand profiprofi ts,ts, andand werewere laterlater keptkept iinn pplacelace bbyy tthehe fi rm.rm. Likewise,Likewise, thethe modernmodern managerialmanagerial practicespractices introducedintroduced byby BBloom,loom, Eifert,Eifert, Mahajan,Mahajan, McKenzie,McKenzie, andand RobertsRoberts (2011)(2011) atat largelarge textiletextile fi rmsrms inin IIndiandia increasedincreased profiprofi tsts andand werewere keptkept inin placeplace afterafter thethe endend ofof thethe experiment.experiment. IInn bothboth cases,cases, thethe ownersowners attributedattributed thethe failurefailure toto exploreexplore thesethese optionsoptions earlierearlier asas ddueue ttoo tthehe hhighigh oopportunitypportunity ccostost ooff ttheirheir ttime.ime. TThehe ssameame cconstraintonstraint sseemseems ttoo bbee bbindinginding bbothoth iinn vveryery ccompetitiveompetitive eenvironments,nvironments, aass eexperiencedxperienced bbyy tthehe ssoftoft ffruitruit fi rmrm wewe analyzedanalyzed inin Bandiera,Bandiera, Barankay,Barankay, andand RasulRasul (2007),(2007), andand inin settingssettings wherewhere ccompetitionompetition isis veryvery mild,mild, asas forfor thethe fi rmsrms surveyedsurveyed byby Bloom,Bloom, Eifert,Eifert, Mahajan,Mahajan, MMcKenzie,cKenzie, andand RobertsRoberts (2011).(2011). TThishis fi ndingnding suggestssuggests promisingpromising newnew avenuesavenues forfor futurefuture researchresearch onon thethe iimportancemportance ofof timetime constraintsconstraints inin fi rms.rms. TheThe importanceimportance ofof timetime constraintsconstraints atat tthehe ttopop ttiersiers ooff oorganizationalrganizational hhierarchiesierarchies hhasas bbeeneen rrecognizedecognized iinn ttheoryheory ((BoltonBolton aandnd DDewatripont,ewatripont, 1994;1994; Garicano,Garicano, 2000),2000), butbut evidenceevidence onon whetherwhether andand howhow managersmanagers aallocatellocate theirtheir timetime ttoo mmaximizeaximize fi rmrm performanceperformance isis scant.scant. Bandiera,Bandiera, Guiso,Guiso, Prat,Prat, aandnd SadunSadun (2011)(2011) havehave developeddeveloped a surveysurvey methodologymethodology toto measuremeasure howhow chiefchief eexecutivexecutive ooffiffi cerscers spendspend theirtheir time,time, andand showshow howhow thethe patternpattern ofof timetime allocationallocation cancan bbee uusedsed ttoo pproviderovide observationalobservational evidenceevidence onon thethe internalinternal andand externalexternal constraintsconstraints ffacedaced byby fi rms.rms. CombiningCombining thisthis surveysurvey methodologymethodology withwith fi eldeld experimentsexperiments holdsholds ppotentialotential fforor mmakingaking ssubstantialubstantial pprogressrogress iinn uunderstandingnderstanding tthesehese iissues.ssues. IInn tturn,urn, sscarcitycarcity ooff mmanagerialanagerial ttimeime ccanan bbee ssymptomaticymptomatic ooff ttwowo ootherther pproblems.roblems. FFirst,irst, iiff tthehe oownerwner oorr cchiefhief eexecutivexecutive ooffiffi cercer mustmust controlcontrol allall aspectsaspects ofof thethe business,business, tthehe scopescope andand sizesize ofof thethe fi rmrm isis necessarilynecessarily limited.limited. DelegationDelegation ofof authorityauthority andand ddecisionecision makingmaking iiss aann eessentialssential iingredientngredient fforor fi rmrm expansion,expansion, andand yetyet wewe havehave a veryvery llimitedimited eempiricalmpirical uunderstandingnderstanding ooff wwhyhy ssomeome oownerswners ffailail ttoo ddelegate.elegate. AAgencygency pprob-rob- llemsems aandnd tthehe iinabilitynability toto motivatemotivate lower-tierlower-tier managersmanagers intuitivelyintuitively seemseem important,important, aass rreportedeported byby thethe fi rmrm ownersowners surveyedsurveyed byby Bloom,Bloom, Eifert,Eifert, Mahajan,Mahajan, McKenzie,McKenzie, andand RRobertsoberts (2011),(2011), butbut moremore evidenceevidence isis neededneeded onon howhow thesethese cancan bebe tackled.tackled. FieldField eexperimentsxperiments thatthat varyvary thethe distributiondistribution ofof authorityauthority oror thethe agencyagency constraintsconstraints couldcould ppotentiallyotentially provideprovide this.this. Second,Second, therethere isis oftenoften a lacklack ofof managersmanagers withwith adequateadequate hhumanuman ccapitalapital andand talent,talent, whetherwhether duedue toto a marketmarket failurefailure inin educationeducation oror a skewedskewed ddistributionistribution ofof ttalent.alent. TThesehese cconstraintsonstraints aarere nnotot aamenablemenable ttoo eexperimentalxperimental vvariation,ariation, bbutut byby providingproviding evidenceevidence onon thethe internalinternal constraintsconstraints ofof fi rms,rms, fi eldeld experimentsexperiments ccanan hhelpelp gguideuide researchresearch onon thesethese ttopics,opics, ttoo.oo. OOurur discussiondiscussion soso farfar hashas beenbeen basedbased onon thethe assumptionassumption thatthat fi rmsrms maximizemaximize pprofirofi tsts subjectsubject toto constraints.constraints. ButBut ratherrather thanthan focusingfocusing onon constraintsconstraints toto optimizingoptimizing bbehavior,ehavior, a growinggrowing bodybody ofof observationalobservational studiesstudies suggestssuggests thethe alternativealternative viewview thatthat fi rmsrms mmightight nnotot mmaximizeaximize pprofirofi ts,ts, eithereither becausebecause theythey areare ledled byby managersmanagers whowho eenjoynjoy tthehe ““quietquiet llife”ife” ((BertrandBertrand aandnd MMullainathan,ullainathan, 22003)003) oorr bbecauseecause ttheyhey aarere oownedwned bbyy familiesfamilies whosewhose objectiveobjective functionfunction hashas a nonmonetarynonmonetary componentcomponent derivingderiving fromfrom ddirectirect controlcontrol (Bertrand(Bertrand andand Schoar,Schoar, 2006).2006). A burgeoningburgeoning bodybody ofof workwork usesuses fi eldeld eexperimentsxperiments toto uunderstandnderstand consumerconsumer preferencespreferences andand optimizingoptimizing behavior.behavior. ItIt iiss wworthorth eexploringxploring wwhetherhether ssimilarimilar sstrategiestrategies ccouldould bbee aadopteddopted ttoo sshedhed llightight oonn ddepar-epar- tturesures ffromrom tthehe aassumptionssumption ooff pprofirofi t maximizationmaximization byby fi rms.rms. 80 Journal of Economic Perspectives

FFieldield eexperimentersxperimenters hhaveave aalsolso bbegunegun ttoo eexplorexplore tthehe bbehaviorehavior ooff nnot-for-profiot-for-profi ts.ts. WWhilehile ttherehere existsexists a vibrantvibrant literatureliterature usingusing fi eldeld experimentsexperiments onon thethe fund-raisingfund-raising aactivitiesctivities ofof suchsuch organizations,organizations, manymany otherother issuesissues remainremain unexplored.unexplored. TheoreticalTheoretical ccontributionsontributions fromfrom BenabouBenabou andand TiroleTirole ((2003)2003) aandnd BBesleyesley aandnd GGhatakhatak (2005)(2005) makemake cclearlear thatthat thethe provisionprovision ofof incentivesincentives forfor pro-socialpro-social taskstasks raisesraises differentdifferent issuesissues thanthan fforor privateprivate taskstasks onon atat leastleast twotwo dimensions:dimensions: First,First, toto tthehe eextentxtent tthathat aagentsgents eengagedngaged iinn pro-socialpro-social taskstasks areare intrinsicallyintrinsically motivated,motivated, fi nancialnancial rewardsrewards thatthat mightmight success-success- ffullyully eelicitlicit efforteffort forfor privateprivate taskstasks couldcould actuallyactually reducereduce efforteffort ifif thethe rewardsrewards crowdcrowd ooutut intrinsicintrinsic motivation.motivation. Second,Second, thethe typetype ofof incentiveincentive mechanismmechanism mightmight affectaffect tthehe compositioncomposition ofof thethe poolpool ofof agentsagents whowho choosechoose toto participateparticipate inin thethe activity.activity. InIn pparticular,articular, high-poweredhigh-powered fi nancialnancial incentivesincentives mightmight attractattract individualsindividuals whowho areare moti-moti- vvatedated byby fi nancialnancial returnsreturns insteadinstead ofof individualsindividuals whowho shareshare thethe pro-socialpro-social orientationorientation ooff tthehe oorganization,rganization, withwith undesirableundesirable consequences.consequences. LaboratoryLaboratory experimentsexperiments showshow tthathat thethe effecteffect ofof fi nancialnancial rewardsrewards differsdiffers whenwhen thethe tasktask hashas socialsocial value.value. ButBut fi eldeld eevidencevidence oonn tthesehese iissuesssues iiss sscant.cant. AAnothernother question,question, withwith manymany practicalpractical implications,implications, isis howhow dodo fi rmsrms andand iindividualsndividuals atat differentdifferent layerslayers ofof thethe hierarchyhierarchy matchmatch endogenouslyendogenously basedbased onon theirtheir rrespectiveespective characteristics?characteristics? TheThe factfact tthathat mmanyany oobservedbserved ooutcomesutcomes ccanan bbee aascribedscribed ttoo eendogenousndogenous matchingmatching isis oftenoften a limitationlimitation ofof observationalobservational studiesstudies (Ackerberg(Ackerberg andand BBotticini,otticini, 2002).2002). FieldField experimentsexperiments onon ddiscriminationiscrimination pproviderovide ssomeome eevidencevidence onon hhowow fi rrmsms hhireire wworkers,orkers, bbutut mmanyany qquestionsuestions rremainemain oopen.pen. WWee eenvisagenvisage fi eldeld experi-experi- mmentsents tthathat ccreatereate eexogenousxogenous vvariationariation inin thethe parametersparameters ofof thethe matchingmatching process,process, bby,y, forfor instance,instance, varyingvarying thethe informationinformation setset availableavailable toto employersemployers andand employees,employees, oorr bbyy rreducingeducing searchsearch costscosts throughthrough thethe introductionintroduction ofof electronicelectronic marketmarket placesplaces wwherehere employersemployers andand employeesemployees cancan meet.meet. AdvancingAdvancing thethe methodologymethodology inin thesethese ddirections,irections, perhapsperhaps inin somesome casescases interveningintervening atat thethe levellevel ofof marketsmarkets asas a whole,whole, wwouldould ttakeake fi eldeld experimentsexperiments intointo anan excitingexciting newnew realm.realm.

■ We have benefi ted from discussions with Nava Ashraf, David Autor, Stefano DellaVigna, Chad Jones, John List, Stephan Meier, Andrea Prat, David Reiley, Timothy Taylor, and seminar participants at the Econometric Society Meeting in Denver, 2011.

References

Ackerberg, Dan, and Maristella Botticini. 2002. Aghion, Philip, and Jean Tirole. 1997. “Formal “Endogenous Matching and the Empirical Deter- and Real Authority in Organizations.” Journal of minants of Contract Form.” Journal of Political Political Economy, 105(1): 1–29. Economy, 110(3): 564–91. Alonso, Ricardo, Wouter Dessein, and Niko Oriana Bandiera, Iwan Barankay, and Imran Rasul 81

Matouschek. 2008. “When Does Coordination David McKenzie, and John Roberts. 2011. “Does Require Centralization?” American Economic Management Matter? Evidence from India.” NBER Review, 98(1): 145–79. Working Paper 16658. Ashraf, Nava, Oriana Bandiera, and Kelsey Jack. Bloom, Nicholas, Raffaella Sadun, and John van 2011. No Margin, No Mission? A Field Experiment on Reenen. 2010. “Does Product Market Competition Incentives for Pro-Social Tasks. Unpublished paper, Lead Firms to Decentralize?” American Economic London School of Economics. Review, 100(2): 434–38. Bandiera, Oriana, Iwan Barankay, and Imran Bloom, Nicholas, and John Van Reenen. 2010a. Rasul. 2005. “Social Preferences and the Response “Human Resource Management and Productivity.” to Incentives: Evidence from Personnel Data.” Forthcoming in the Handbook of Labor Economics, Quarterly Journal of Economics, 120(3): 917–62. ed. O. Ashenfelter and D. Card. Elsevier. Bandiera, Oriana, Iwan Barankay, and Imran Bloom, Nicholas, and John Van Reenen. Rasul. 2007. “Incentives for Managers and 2010b. “Why Do Management Practices Differ Inequality among Workers: Evidence from a Firm across Firms and Countries?” Journal of Economic Level Experiment.” Quarterly Journal of Economics, Perspectives, 24(1): 203–24. 122(2): 729–74. Bolton, Patrick, and Mathias Dewatripont. Bandiera, Oriana, Iwan Barankay, and Imran 1994. “The Firm as a Communication Network.” Rasul. 2009. “Social Connections and Incentives Quarterly Journal of Economics, 109(4): 809–39. in the Workplace: Evidence from Personnel Data.” Brown, Jennifer, and John Morgan. 2009. “How Econometrica, 77(4): 1047–94. Much is a Dollar Worth? Tipping versus Equilib- Bandiera, Oriana, Iwan Barankay, and Imran rium Coexistence on Competing Online Auction Rasul. 2010. “Social Incentives in the Workplace.” Sites.” Journal of Political Economy, 117(4): 668–700. Review of Economic Studies, 77(2): 417–58. Chiappori, Pierra-Andre, and Bernard Salanié. Bandiera, Oriana, Iwan Barankay, and Imran 2003. “Testing Contract Theory: A Survey of Some Rasul. 2011. “Team Incentives: Evidence from a Recent Work.” Chap. 4 in Advances in Economics Field Experiment.” http://www.homepages.ucl and Econometrics Theory and Applications: Eighth .ac.uk/~uctpimr/research/teams.pdf. World Congress Volume 1, ed. by M. Dewatripont, Bandiera, Oriana, Luigi Guiso, Andrea Prat, L. P. Hansen, and S. J. Turnovsky. Cambridge: and Raffaella Sadun. 2011. “What Do CEOs Do?” Cambridge University Press. CEPR Discussion Paper 8235. Cho, Sungjin, and John Rust. 2010. “The Flat- Barankay, Iwan. 2010. “Rankings and Social Rental Puzzle.” Review of Economic Studies, 77(2): Tournaments: Evidence from a Field Experiment.” 560–94. http://www.management.wharton.upenn.edu Coase, R. H. 1937. “The Nature of the Firm.” /barankay/Documents/Rankings_and_Social Economica, 4(16): 386–405. _Tournaments.pdf. de Mel, Suresh, David McKenzie, and Chris- Benabou, Roland, and Jean Tirole. 2003. topher Woodruff. 2008. “Returns to Capital in “Intrinsic and Extrinsic Motivation.” Review of Microenterprises: Evidence from a Field Experi- Economic Studies, 70(3): 489–520. ment.” Quarterly Journal of Economics, 123(4): Bertrand, Marianne, and Sendhil Mullainathan. 1329–72. 2003. “Enjoying the Quiet Life? Corporate Gover- Drexler, Alejandro, Greg Fischer, and Antoi- nance and Managerial Preferences.” Journal of nette Schoar. 2011. “Keeping It Simple: Financial Political Economy, 111(5): 1043–75. Literacy and Rules of Thumb.” http://personal Bertrand, Marianne, and Sendhil Mullainathan. .lse.ac.uk/fi scherg/Assets/KIS-DFS-Jan2011.pdf. 2004. “Are Emily and Greg More Employable Than Garicano, Luis. 2000. “Hierarchies and the Lakisha and Jamal? A Field Experiment on Labor Organization of Knowledge in Production.” Market Discrimination.” American Economic Review, Journal of Political Economy, 108(5): 874–904. 94(4): 991–1013. Harrison, Glenn W., and John A. List. 2004. Bertrand, Marianne, and Antoinette Schoar. “Field Experiments.” Journal of Economic Literature, 2006. “The Role of Family in Family Firms.” Journal 42(4): 1009–1055. of Economic Perspectives, 20(2): 73–96. Heckman, James J., and Peter Siegelman. 1993. Besley, Timothy J., and Maitreesh Ghatak. “The Urban Institute Audit Studies: Their Methods 2005. “Competition and Incentives with Motivated and Findings.” In Clear and Convincing Evidence: Agents.” American Economic Review, 95(3): 616–36. Measurement of Discrimination in America, ed. M. Fix Besley, Timothy J., and Maitreesh Ghatak. 2008. and R. Struyk, 187–258. Washington, D.C.: Urban “Status Incentives.” American Economic Review, Institute Press. 98(2): 206–11. Ibragimov, Rustam, and Ulrich K. Müller. 2010. Bloom, Nicholas, Benn Eifert, Aprajit Mahajan, “t-Statistic Based Correlation and Heterogeneity 82 Journal of Economic Perspectives

Robust Inference.” Journal of Business and Economic Moldovanu, Benny, Aner Sela, and Xianwen Statistics, 28(4): 453–68. Shi. 2007. “Contests for Status.” Journal of Political Kandel, Eugene, and Edward Lazear. 1992. Economy, 115(2): 338–63. “Peer Pressure and Partnerships.” Journal of Nagin, Daniel, James B. Rebitzer, Seth Sanders, Political Economy, 100(4): 801–13. and Lowell J. Taylor. 2002. “Monitoring, Motiva- Karlan, Dean, and Martin Valdivia. 2011. tion, and Management: The Determinants of “Teaching Entrepreneurship: Impact of Business Opportunistic Behavior in a Field Experiment.” Training on Microfi nance Clients and Institutions.” American Economic Review, 92(4): 850–73. Review of Economics and Statistics, 93(2): 510–27. National Commission for the Protection of Lazear, Edward P. 1989. “Pay Equality and Human Subjects of Biomedical and Behavioral Industrial Politics.” Journal of Political Economy, Research. 1979. “The Belmont Report” aka: 97(3): 561–80. “Ethical Principles and Guidelines for the Protec- Lazear, Edward P. 2000. Performance Pay and tion of Human Subjects of Research.” April 18. Productivity.” American Economic Review, 90(5): http://www.hhs.gov/ohrp/humansubjects 1346–61. /guidance/belmont.html. Lee, David S. 2009. “Training, Wages and Niederle, Muriel, and Lisa Vesterlund. 2007. Sample Selection: Estimating Sharp Bounds on “Do Women Shy Away from Competition? Do Treatment Effects.” Review of Economic Studies, Men Compete Too Much?” Quarterly Journal of 76(3): 1071–1102. Economics, 122(3): 1067–1101. Levitt, Steven D., and John A. List. 2009. “Field Ostrovsky, Michael, and Michael Schwarz. 2009. Experiments in Economics: The Past, the Present, “Reserve Prices in Internet Advertising Auctions: and the Future.” European Economic Review, 53(1): A Field Experiment.” http://faculty-gsb.stanford 1–18. .edu/ostrovsky/papers/rp.pdf. Levitt, Steven D. and John A. List. 2011. “Was Prendergast, Canice. 1999. “The Provision of There Really a Hawthorne Effect at the Hawthorne Incentives in Firms.” Journal of Economic Literature, Plant? An Analysis of the Original Illumination 37(1): 7–63. Experiments.” American Economic Journal: Applied Reiley, David H. 2006. “Field Experiments on Economics, 3(1): 224–38. the Effects of Reserve Prices in Auctions: More List, John A., and Imran Rasul. 2011. “Field Magic on the Internet.” RAND Journal of Economics, Experiments in Labor Economics.” Chap. 2 37(1): 195–211. in Handbook of Labor Economics, Vol. 4A, ed. O. Riach, Peter A., and Judith Rich. 2002. “Field Ashenfelter and D. Card. Elsevier. Experiments of Discrimination in the Market List, John A., Sally Sadoff, and Mathis Wagner. Place.” Economic Journal, 112(483): F480–F518. 2010. “So You Want to Run an Experiment, Now Rosen, Sherwin. 1982. “Authority, Control, What? Some Simple Rules of Thumb for Optimal and the Distribution of Earnings.” Bell Journal of Experimental Design.” NBER Working Paper 15701. Economics, 13(2): 311–23. Lucas, Robert E. 1978. “On the Size Distribu- Rotemberg, Julio J. 1994. “Human Relations tion of Business Firms.” Bell Journal of Economics, in the Workplace.” Journal of Political Economy, 9(2): 508–23. 102(4): 684–717. Mas, ALexandre, and Enrico Moretti. 2009. Shearer, Bruce S. 2004. “Piece Rates, Fixed “Peers at Work.” American Economic Review, 99(1): Wages and Incentives: Evidence from a Field 112–45. Experiment.” Review of Economic Studies, 71(2): Mayo, Elton. 1933. The Human Problems of an 513–34. Industrial Civilization. New York: Macmillan. Shi, Lan. 2010. “Incentive Effect of Piece-Rate McKenzie, David. 2011. “Beyond Baseline and Contracts: Evidence from Two Small Field Experi- Follow-up: The Case for More T in Experiments.” ments.” B.E. Journal of Economic Analysis & Policy, Policy Research Working Paper 5639, World Bank. vol. 10, issue 1 (Topics), Article 61. McKenzie, David, and Christopher M. Wood- Wu, Yanhui. 2011. “Authority, Incentives ruff. 2008. “Experimental Evidence on Returns to and Performance: Theory and Evidence from a Capital and Access to Finance in Mexico.” World Chinese Newspaper.” http://personal.lse.ac.uk Bank Economic Review, 22(3): 457–82. /WUY9/JMP.pdf.