nber-為什么受過教育的工人的工資增長更快

NBERWORKINGPAPERSERIESWHYDOWAGESGROWFASTERFOREDUCATEDWORKERS?avidJDemingWorkingPaper31373http//papers/w31373NATIONALBUREAUOFECONOMICRESEARCHCambridgeMA8June3Ihavenothingtodisclose.TheviewsexpressedhereinarethoseoftheauthoranddonotnecessarilyreflecttheviewsoftheNationalBureauofEconomicResearch.NBERworkingpapersarecirculatedfordiscussionandcommentpurposes.Theyhavenotbeenpeer-reviewedorbeensubjecttothereviewbytheNBERBoardofDirectorsthataccompaniesofficialNBERpublications.?2023byDavidJ.Deming.Allrightsreserved.Shortsectionsoftext,nottoexceedtwoparagraphs,maybequotedwithoutexplicitpermissionprovidedthatfullcredit,including?notice,isgiventothesource.WhyDoWagesGrowFasterforEducatedWorkers?avidJDemingNBERWorkingPaperNo.31373June3JELNoJ4STRACTTheU.S.collegewagepremiumdoublesoverthelifecycle,from27percentatage25to60percentatage55.Usingapanelsurveyofworkersfollowedthroughage60,Ishowthatgrowthinthecollegewagepremiumisprimarilyexplainedbyoccupationalsorting.Shortlyaftergraduating,workerswithcollegedegreesshiftintoprofessional,nonroutineoccupationswithmuchgreaterreturnstotenure.Nearly90percentoflifecyclewagegrowthoccurswithinratherthanbetweenjobs.Tounderstandthesepatterns,Idevelopamodelofhumancapitalinvestmentwhereworkersdifferinlearningabilityandjobsvaryincomplexity.Fasterlearnerscompletemoreeducationandsortintocomplexjobswithgreaterreturnstoinvestment.Collegeactsasagatewaytoprofessionaloccupations,whichoffermoreopportunityforwagegrowththroughon-the-joblearning.avidJDemingHarvardKennedySchoolMalcolmWienerCenterforSocialPolicyJFKStCambridgeMA8arvardKennedySchoolsoNBERdemingharvardeduAdataappendixisavailableat/data-appendix/w3137311IntroductionThecollegewagepremiumroughlydoublesoverthelifecycle,bothintheU.S.andinotherdevelopedcountries(Lemieux2006b,RubinsteinandWeiss2006,Bhulleretal.2017,Lagakosetal.2018).Whydowagesgrowfasterforeducatedworkers?Wagegrowthvariestremendously,withannuallaborearningsintheU.S.risingby60percentfromage25toage55forthemedianworker,butbarelyatallforworkersatthe25thpercentileandroughlydoublingatthe75thpercentile(Guvenenetal.2021).Lifecyclewagegrowthispotentiallyasimportantasschoolinginexplainingcross-countryvariationinGDPperworker,withwagesgrowingtwiceasmuchinrichcountriescomparedtopoorcountries(Rossi2020,Lagakosetal.2018,Jedwabetal.2021).Forallthesereasons,itisimportanttounderstandtherelationshipbetweenwagegrowthandhumancapitalaccumulation.Yetrelativelyfewstudiesinvestigatethedeterminantsofon-the-joblearning,especiallycomparedtothevastliteratureconcerningimpactsofpre-marketschoolinginterventions(e.g.RubinsteinandWeiss2006,SandersandTaber2012,Deming2022).Thispapershowshowoccupationalsortingleadstogreaterwagegrowthforeducatedworkers.IfirstshowthattheU.S.collegewagepremiumincreasessteadilyoverthelifecy-cle,morethandoublingbetweentheagesof25and55.Thisbasicpatternholdsevenaftercontrollingforcognitiveskillandwhenusingdirectmeasuresoflabormarketexperience.1SimilartoBhulleretal.(2017),IfindthattheMincerearningsfunctionsubstantiallyunder-statesthelifecyclereturntoeducation,andIshowthatamodifiedearningsfunctionwhereeducationinteractslinearlywithworkexperiencefitsthedatamuchbetter.1Severalotherstudieshavefoundthatearningsincreasefasterforeducatedworkers,althoughtheyaretypicallylimitedtoshorterpanelsorcannotseparateearningsfromwages(e.g.RubinsteinandWeiss2006,Bhulleretal.2017,Lagakosetal.2018).Theestimatesinthispaperusepaneldatafollowingindividualworkers,areconsistentacrossbothNLSYwaves,andarerobusttoalternativemeasuresofhoursandearn-ings.Thisfindingcontrastswithmanyexistingstudiesthatestimateflatordecliningreturnstopotentialexperienceformid-andlate-careerworkers(e.g.Mincer1974,MurphyandWelch1990,Heckmanetal.2006,Lemieux2006c,Lagakosetal.2018,Jedwabetal.2021).Ishowthatwagegrowthestimatesusingpotentialexperiencearebiaseddownwardbecauseeducatedandhigher-wageworkersaccumulatemorehoursandaremoreconsistentlyattachedtothelaborforce.2Ithenstudytherelationshipbetweenearlycareerjobmobilityandwagegrowth,ex-tendingtheanalysesofTopelandWard(1992)andvonWachterandBender(2006)tothe1979and1997cohortsoftheNationalLongitudinalSurveyofYouth(NLSY)butfocusingondifferencesinjobmobilitypatternsbyeducation.IexploitdetailedemploymentandwagehistorydatafromtheNLSYtodecomposewagegrowthintowithin-andbetween-jobcom-ponents.Forcollegegraduates,85to90percentoftotalwagegrowthiswithinratherthanbetweenjobs,andtheimportanceofwithin-jobwagegrowthincreasesoverthelifecycle.Thewithin-jobcomponentalsobecomesmoreimportantovertimeforlesseducatedworkers.Collegegraduatesswitchjobsfrequentlywhileenrolledandinthefirsttwoyearsaftercompletingschooling,butarelessmobilethanhighschooleducatedworkersfortherestoftheircareers.Aftergraduating,workerswithcollegedegreessortoutofadministrativesupportandservicesoccupationsandintoprofessionaloccupationsinbusiness,engineering,medicine,education,andrelatedfields.Agoodsummarystatisticforoccupationalsortingisroutineness,followingseminalworkbyAutoretal.(2003),Goosetal.(2014)andoth-ers.Collegeeducatedworkerssortintononroutineoccupationsshortlyaftertheygraduate,whereastheoccupationmixchangesverylittleforhighschooleducatedworkersaftertheycompleteschooling.Usingdetailedpaneldata,Iestimatewagegrowthwithjobtenureandshowhowitvariesbyoccupation.Ifindthatwagegrowthisslowerinroutineoccupations.Realwagesaremorethan20percenthigherafter15yearsofjobtenureforworkersinoccupationsatthe25thpercentileofroutineness.Incontrast,Ifindalmostnowagegrowthbeyond4yearsoftenureinjobsatthe75thpercentileofroutineness.Theseconclusionsarerobusttocontrollingforoccupationfixedeffectsandindividualfixedeffects,whichaccountsforleveldifferencesinwagesacrossoccupationsandidentifiesthereturnstotenurefromwithin-workerjobswitching.Howmuchofthelifecyclewideningofeducationalwagedifferentialscanbeexplainedbyoccupationalsorting?Iapproachthisquestionintwoways.First,usingdatafromthe3MarchCurrentPopulationSurvey(CPS),IassignNLSYrespondentstheaveragewageateachagefortheirfirstpost-schoolingoccupation.Thisprojectedcollegewagepremiumstilldoublesbetweenage25andage55andthemagnitudesimplythatabouthalfoflifecyclewagegrowthcanbepredictedbyaperson’sfirstjobafterfinishingschool.Second,Ishowthatthecontrollingforoccupation-by-tenurefixedeffectsreducesthetotalgrowthinthecollegewagepremiumbyabout50percent.Overall,Iestimatethatearlycareeroccupationalsortingexplainsatleasthalfofthelifecyclegrowthinthecollegewagepremium.Tounderstandthesefacts,Idevelopasimplemodelofhumancapitalinvestmentandon-the-joblearninginthespiritofBen-Porath(1967)andCavounidisandLang(2020).Workersvaryintheirlearningability,andtheproductivityofon-the-joblearningisgreaterincomplex(nonroutine)occupations.2Returnstoon-the-joblearningdiminishmorerapidlyinroutinejobs,whicharesimplerandmorepredictable(e.g.Autoretal.2003).Sincelearningabilityandjobcomplexityarecomplements,fasterlearnerswillobtainmoreschoolingandwillsortintocomplexoccupations.Thecollegewagepremiumincreasesbecauseofsorting,butalsobecauseofthegreaterreturnstoscaleinskillinvestmentaffordedbycomplexoccupations.Iclosethemodelbyintroducingimperfectsubstitutionacrossoccupations,whichcauseswagestoadjustuntilthemarginalworkerisindifferentbetweensectors.3Withreasonableassumptionsaboutthedistributionoflearningabilityandtheassignmentofworkerstotasks,themodeldeliversasimplecharacterizationoflife-cyclecollegewagepremiaandoccupationalwagedifferentials.Thispaperestablishestheprimaryroleofoccupationalsortinginexplaininglifecyclegrowthofthecollegewagepremium.Itcomplementsalargeliteraturestudyingthesourcesofwagegrowthoverworkers’careers,includingheterogeneitybyfirmandindustry(e.g.Dust-mannandMeghir2005,Gregory2020,Arellano-BoverandSaltiel2021,AddaandDustmann2InBen-Porath(1967),thehumancapitalproductionfunctionincludesaproductivityparameterthataugmentslearningoutputandacurvatureparameterthatdeterminesthereturnstoscaleinhumancapitalinvestment.IinterprettheproductivityparameterasindividuallearningabilityfollowingNealandRosen(2000)andHuggettetal.(2006),andthecurvatureparameterasoccupationcomplexity(nonroutineness).3Theequilibriumofthemodelischaracterizedbyathresholdlearningability,withsortingbelowandaboveintoroutineandcomplexoccupationsrespectively.42023).Thesestudiestypicallyfindthatgeneralhumancapitalaccumulationisthemostim-portantsourceofwagegrowth(e.g.ConnollyandGottschalk2006,Sch?nberg2007,Addaetal.2013,Baggeretal.2014).Agrowingbodyofworkdocumentssortingofhigh-wageworkerstohigh-wagefirms,whichispotentiallyconsistentwithoccupationalsortinggiventheriseoffirmspecializationanddomesticoutsourcing(e.g.Cardetal.2013,GoldschmidtandSchmieder2017,Songetal.2019).Morebroadly,myfindingshighlighttheimportanceoftheschool-to-worktransitionandearlycareermobilityforlifetimewagegrowth(e.g.Ryan2001,Oreopoulosetal.2012,Wachter2020).Intuitively,acollegedegreeprovidesaccesstobetterjobs,whichallowformoreskilldevelopmentandon-the-joblearning.Thismayseemobvious.Yetitisunexplainedbymanyseminalmodelsofhumancapitalandwagedetermination.4TheMincer(1974)earningsfunc-tionimpliesaconstantreturntoschoolingoverthelife-cycle,althoughmorerecentevidencesuggestsotherwise(Heckmanetal.2003,Lemieux2006c,Heckmanetal.2006,Lagakosetal.2018).Themostinfluentialmodelofhowhumancapitalaffectsthewagestructureofaneconomyisthesupply-demand-institutions(SDI)framework,orthecanonicalmodel,whichshowshowtechnologyincreasesrelativedemandforhigh-skilledlabor(Tinbergen1975,GoldinandKatz2007,AcemogluandAutor2011).InthecanonicalmodelandthetaskframeworkofAcemogluandAutor(2011),thecollegewagepremiumincreasesovertimebecauseoftechnologicalchange.Thispaperexplainswhythecollegewagepremiumincreasesasindividualsgainworkexperience,independentoftheeconomicenvironment.Itdoessobynestinghumancapitalinvestmentandwagedynamicsintothestructureofthecanonicalmodel.4Baggeretal.(2014)developanequilibriummodelwithon-the-joblearning,heterogeneousemployers,andindividualshocks.Theyfindfasterwagegrowthforeducatedworkers,andtheydecomposeitintohumancapitalaccumulationversushigherreturnstojobsearch.Theyfindthatjobsearchisanimportantcontributor,butonlyinthefirst10yearsofaworker’scareer.LiseandPostel-Vinay(2020)estimateastructuralsearch-and-matchingmodelwithcognitive,manual,andinterpersonalskillsandfindthatthereturntocognitiveskills-butnottheothertwo-increaseswithworkexperience.Neitherpaperexplicitlystudiesheterogeneousreturnstoexperiencebyjobroutinenessorotherdimensionsofoccupation.Blundelletal.(2016)estimateadynamiclifecyclemodelandfindthatwagegrowthforwomenintheUKisgreateramongthecollege-educated,andthatpart-timeworkdoesnotcontributetowagegrowth.5Thispapercontributestoourunderstandingofhowroutineworkaffectslifecyclewageinequality.Routinejobsrequirerepeatedexecutionofrule-basedtasksandarepurposelyde-signedtolimitworkerdiscretionandon-the-joblearning(LindbeckandSnower2000,Autoretal.2003,Bartlingetal.2012).Thusthesortingofeducatedworkersintononroutinejobssuggeststhatlifetimeearningsinequalitymaybelargerthancross-sectionalcomparisonssuggest(e.g.AabergeandMogstad2015,Hoffmannetal.2020,Guvenenetal.2022).Ace-mogluandRestrepo(2018)findthattheincreasingautomationofroutinetaskshasincreasedthecollegewagepremiumsinceeducatedworkersaremuchmorelikelytoholdnonroutinejobs.Thispapercomplementstheirfindingbyexploringoccupationalsortingdirectlyanddevelopingamodelthatexplainsitasanequilibriumphenomenon.Myfindingsarealsorelatedtotheliteratureonjobladdersandknowledgehierarchies,whereskilledworkersarepromotedtopositionsthatincreasetheirdecision-makingauthor-ityandspanofcontrolwithinthefirm(GibbonsandWaldman1999,GaricanoandRossi-Hansberg2006,GibbonsandWaldman2006).Byallowingfordifferencesintheproductivityofon-the-joblearning,mymodelissimilarinspirittoNelsonandPhelps(1966),whovieweducationasincreasingtheabilitytolearnandadapttochange.Finally,thispapercon-tributestothemacroeconomicliteratureonhumancapitalinvestment,learning-by-doing,andearningsdynamics(Rosen1972,Heckmanetal.1998,2002,Guvenen2006,BowlusandRobinson2012,ManuelliandSeshadri2014).Thepaperproceedsasfollows.Section2describesthedataandmeasurementofworkexperience.Section3establishestheempiricalpatternofgreaterlifecyclewagegrowthforeducatedworkers.Section4establishesoccupationalsortingasakeyexplanationforthelifecyclegrowthinthecollegewagepremium.Section5presentsthemodelanddevelopsitsimplications,andSection6concludes.62DataMymaindatasourceisthe1979NationalLongitudinalSurveyofYouth(NLSY79).TheNLSY79isanationallyrepresentativesampleof11,406youthages14to22in1979.5Thesurveywasconductedyearlyfrom1979to1993andthenbiannuallyfrom1994through2018,whenrespondentswereages53to61.Attritionisrelativelylow,withabout90percentofparticipantsretainedthrough1994and70percentretainedthrough2018.Isupplementthemainanalysiswithdatafromthe1997NLSY(NLSY97),anationallyrepresentativesampleof8,984youthage12to17in1997.6Thesurveywasconductedyearlyfrom1997to2011andthenbiannuallyfrom2011to2019,whenrespondentswereages34to39.AttritionisalsolowintheNLSY97,with77percentretainedthrough2019.BoththeNLSY79andtheNLSY97surveysincludeconsistentanddetailedmeasuresofeducationandpremarketskills,employment,wagesandearnings,andoccupationandemployerhistory.IharmonizeoccupationcodesacrossNLSYwavesandyearsusingthe“occ1990dd”cross-walkdevelopedbyAutorandDorn(2013)andextendedbyDeming(2017),andImatchtheseoccupationcodestomeasuresofthetaskcontentofworkfromO*NET.7Ifocusonroutineoccupations,definedhereasjobsthatscorehigheronthequestion“howimportantisrepeatingthesamephysicalactivities(e.g.keyentry)ormentalactivities(e.g.checkingentriesinaledger)overandover,withoutstopping,toperformingthisjob?”,followingDem-ing(2017).8Becauseresponsestothesequestionshavenocardinalmeaning,Iconvertthemtoa0-10scalewhereoccupationsaregiventhevaluethatreflectstheirpercentilerankinthelaborsupply-weighteddistributionofemploymentinthe2017-2019AmericanCommunity5TheNLSY79includesanationallyrepresentativesample(n=6,111),anoversampleofsomegeographiesinordertoobtainhighersharesofnonwhiteanddisadvantagedyouth(n=5,295),andanactivedutymilitarysamplewhichwasreduceddrasticallyin1985(n=1,280originally,n=186in1985).Idropthemilitarysamplefromallmyanalyses.6TheNLSY97includesanationallyrepresentativesample(n=6,748)andanoversampleofsomegeogra-phiesinordertoobtainhighersharesofnonwhiteanddisadvantagedyouth(n=2,236),7O*NETisasurveyadministeredbytheU.S.DepartmentofLabortoarandomsampleofworkersineachoccupation.Iusethe1998O*NETtomaximizeconsistencyacrosssamplewaves.8ThisdefinitiondiffersslightlyfromDeming(2017),whichusestheaverageofthisquestionandanother-“howautomatedisthejob?”.Idropthisquestionbecauseofconcernsabouttheconsistencyofresponsesacrosswaves,althoughmymainresultsarerobusttoincludingbothquestions.7Survey(ACS).9Mymainoutcomeistheinflation-adjustedloghourlywage(indexedto2018dollars),trimmedforvaluesbelow3andabove200followingAltonjietal.(2012).IuseArmedForcesQualifyingTest(AFQT)scorestomeasurecognitiveskill,followingmanyotherstudies,andIemploytheage-normedmappingofscoresacrosswavescreatedbyAltonjietal.(2012).BothwavesoftheNLSYincludemaskedemployerandjobidentifiers,enablingmetocalculateemployerandoccupationtenure.IsupplementtheNLSYanalyseswithcross-sectionaldatafromthe1980-2020MarchCPSAnnualSocialandEconomicSupplement(ASEC).IntheMarchCPSIcomputewagesbydividingannualwageandsalaryincomebyannualhoursworked,followingLemieux(2006a).2.1MeasurementofWorkExperienceandJobSpellsAkeyadvantageoftheNLSYdataisthedetailedcalculationofworkhistoriesandworkexperience.Theworkhistorydataincludeacomprehensiveweeklymeasurementoflaborforcestatus,totalnumberofhoursworked,andjobsheld(ifany)coveringeveryweeksinceJanuary1,1978,includingtheyearsinbetweensurveysorifrespondentsskipasurveywave.10Theseweeklymeasurementsarethensummeduptocalculatehoursandweeksworkedinthelastcalendaryearandsincethelastsurveyinterview.Toaccountforreportingerrorsandthedifferentagesatwhicheachsurveybegins,Idisregardworkexperiencepriortoage18inbothsurveys,althoughresultsareverysimilarifteenageworkexperienceisincluded.Iusethesevariablestocreate(fractional)measuresofyearlyworkexperience,withoneyearequalling2,080hours(52weekstimes40hoursperweek).Mybaselinemeasurementofworkexperienceusestheactualhoursreportedbyeachrespondentineachweek,evenifitexceeds40.Forrespondentsage25to54intheNLSY79,medianworkexperienceperyear9Forexample,ifweorderedallworkersintheU.S.economyin2017-2019accordingtotheiroccupation’sroutinenessscore,theoccupationatthe25thpercentilewouldreceiveascoreof2.5.10Allweeksarecounted,andrespondentswereunabletorecalltheiractivitiesinaweekinlessthan5percentofcases.Incasesofmissingweeks,Iinterpolatetheaveragefortherestoftheperiod(e.g.ifarespondentwasworkingfor40weeksoftheyear,notworkingfor10,andmissingfor2,Icodethe2missingweeksas0.8workingand0.2notworking).Theresultsarenotsensitivetootherreasonablechoices.8is0.65(about26hoursperweek),includingnon-employedrespondents.Amongrespondentsreportingatleastsomeworkinayear,39percentworkedlessthanfull-time,26percentworkedexactlyfull-time(2,080hours),22percentworkedbetween40and50hoursperweek,and13percentreportedworkingmorethan50hoursperweek.11Tostudylife-cyclepatterns,IconverttheNLSYdataintopanelformat,withoneob-servationperrespondent-year,andregressworkexperienceontwo-yearagebinscontrollingforindividualfixedeffects.Sincethereissomeattrition,thepanelisnotperfectlybalanced,buttheagecoe?icientsareidentifiedonlybywithin-workerdifferences.AppendixFigureA1presentshoursworked(infractionsofayear)byageandeducationlevel,forallNLSY79respondents.Four-yearcollegegraduatesworkfewerhoursinitiallybutquicklycatchup,accumulatingmoretotalhoursworkedbytheirlate20s.Hourspeakatage30forfour-yearcollegegraduatesandatage26fornon-collegegraduates,andthendeclinesteadilythroughage60.12TheNLSYmeasuresworkexperiencemoreaccuratelythanmostwidely-usedlabormar-ketdatasources.TheCurrentPopulationSurvey(CPS)andtheAmericanCommunitySurvey(ACS)askrespondentsabouttheir“usual”hoursworkedperweekandweeksworkedperyear,eitheringeneraloroveraspecificperiodoftimesuchasthelastyear(Rugglesetal2022).13Respondentsareinstructedtocountweeksinwhichtheyworkedevenafewhoursandtocountvacationandsickleaveaswork(Floodetal2022).Bothapproachesprobably11Toassesstherobustnessoftheresults,Icreateanalternativemeasurethatcapsworkexperienceat2,080hoursperyear.Ialsoexploreusingalldatareportedsincethelastinterviewevenifitisseveralyearsold,althoughmypreferredapproachusesdatafromonlythepastyeartominimizetheimpactofrecallbias.Finally,Icomputereturnstoworkexperienceintermsofannualearningsratherthanwages,toadd