University of Groningen
The rise of robots and the fall of routine jobs
de Vries, Gaaitzen J.; Gentile, Elisabetta; Miroudot, Sebastien; Wacker, Konstantin M.
Published in:
Labour Economics
DOI:
10.1016/j.labeco.2020.101885
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from
it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date:
2020
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
de Vries, G. J., Gentile, E., Miroudot, S., & Wacker, K. M. (2020). The rise of robots and the fall of routine
jobs. Labour Economics, 66, [101885]. https://doi.org/10.1016/j.labeco.2020.101885
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): http://www.rug.nl/research/portal. For technical reasons the number of authors shown on this cover page is limited to 10 maximum.
ContentslistsavailableatScienceDirect
Labour
Economics
journalhomepage:www.elsevier.com/locate/labeco
The
rise
of
robots
and
the
fall
of
routine
jobs
Gaaitzen
J.
de
Vries
a,∗,
Elisabetta
Gentile
b,
Sébastien
Miroudot
c,
Konstantin
M.
Wacker
a a Faculty of Economics and Business, University of Groningen, the Netherlandsb Asian Development Bank, Philippines
c Organisation for Economic Co-operation and Development (OECD), Trade and Agriculture Directorate, France
a r t i c l e
i n f o
JEL classification: E23 J23 O30 Keywords: Robots Tasks Occupations Eploymenta b s t r a c t
Thispaperexaminestheimpactofindustrialrobotsonjobs.Wecombinedataonrobotadoptionandoccupations byindustryinthirty-sevencountriesfortheperiodfrom2005to2015.Weexploitdifferencesacrossindustries intechnicalfeasibility– definedastheindustry’sshareoftasksreplaceablebyrobots– toidentifytheimpactof robotusageonemployment.Thedataallowustodifferentiateeffectsbytheroutine-intensityofemployment. Wefindthatariseinrobotadoptionrelatessignificantlytoafallintheemploymentshareofroutinemanual task-intensivejobs.Thisrelationisobservedinhigh-incomecountries,butnotinemergingmarketandtransition economies.
1. Introduction
Rapid improvements in robot capabilities have fuelled concerns abouttheimplicationsof robotadoptionforjobs.Whilethecreation ofautonomousrobotswithflexible3Dmovementcontinuestobea ma-jorchallengetoengineers,rapidprogressisbeing made.Robotscan nowperformavarietyoftasks,suchassealing,assembling,and han-dlingtools.Asrobotcapabilitiescontinuetoexpandandunitpricesfall, firmsareintensifying investmentinrobots(FreyandOsborne, 2017;
GraetzandMichaels,2018;AcemogluandRestrepo,2020).Whatisthe impactofrobotadoptiononlabourdemand?Dorobotssubstitutefor taskspreviouslyperformedbyworkers?
Themaincontributionofthispaperistoempiricallystudythe im-pactofindustrialrobotsontheoccupationalstructureoftheworkforce acrossindustriesinasetofhigh-incomeaswellasEmergingMarketand TransitionEconomies(EMTEs).Wecombinealargeanddetailed occu-pationsdatabasewithdataonindustrialrobotdeliveriesfromthe Inter-nationalFederationofRobotics.Thedatabaseonoccupational employ-mentfromReijndersanddeVries(2018)allowsustoexaminetheshare ofemploymentinoccupationswithahighcontentofroutinetasks– i.e. tasksthatcanbeperformedbyfollowingawell-definedsetof proce-dures.Wedelineateoccupationsalongtwodimensionsofthe character-isticsoftasksperformed,namely‘analytic’versus‘manual’,and‘routine’ versus‘non-routine’.1Wethusdistinguishfourkeyoccupational
group-∗Correspondingauthor.
E-mailaddress:g.j.de.vries@rug.nl(G.J.deVries).
1 Thedistinctionbetweenmanualandanalyticoccupationsisbasedon differ-encesintheextentofmentalversusphysicalactivity.
ings,namelyroutinemanual,routineanalytic,non-routinemanual,and non-routineanalytictask-intensiveoccupations(asinAutoretal.2003;
ReijndersanddeVries2018;Cortesetal.2020).WefollowGraetzand Michaels(2018)inconstructingmeasuresofrobotadoptionby country-industrypairsandrelatethesetochangesinoccupationalemployment shares.Oursamplecovers19industriesfor37countriesatvarying lev-elsofdevelopmentfrom2005to2015,andincludesmajorusersof in-dustrial robots,such asthePeoples Republicof China(PRC),Japan, SouthKorea,Germany,andtheUnitedStates.Ourmainfindingisthat country-industry pairsthatsawamorerapidincreaseinrobot adop-tionexperiencedlargerreductionsintheemploymentshareofroutine manualjobs.
Our approachis motivated bythefollowing economic considera-tions.Firmsproduceavarietyofproductsusingacontinuumoftasks (AcemogluandAutor,2011),andtheseproductsdifferinthenumber oftasksthatcanbeperformedbyrobots(GraetzandMichaels,2018). Forexample,theshareofreplaceabletasksbyrobotsdiffersbetween apparelandautomotiveandappearslargerinthelatter.2Thisgivesrise todifferencesacrossindustriesinthetechnicalfeasibilityofrobots substi-tutingtaskspreviouslyperformedbyhumans.Advancesinmachine ca-pabilitiesexpandthesetoftaskscarriedoutbymachines(Acemogluand Restrepo,2018).Firmswilladoptrobotsifitistechnicallyfeasibleand theprofit gainsexceed thecostsof purchasingandinstallingrobots. Givenhigherwagesinadvancedcountries,thetechnicalconstraintsto
2Seee.g.theEconomist,24August2017,“Sewingclothesstillneedshuman hands.Butforhowmuchlonger?”
https://doi.org/10.1016/j.labeco.2020.101885
Received31January2020;Receivedinrevisedform22June2020;Accepted10July2020 Availableonline11July2020
robotsreplacingtasksaremorelikelytobindforfirmsinthese coun-tries.Hence,improvementsinrobotcapabilitieswouldresultinalarger employmentresponseinadvancedcountriescomparedtodeveloping countries.
Weusethese economicinsightsinouranalysis.Inparticular,the technicalfeasibilityofadoptingrobotsguidesourinstrumentalvariables (IV)strategytoidentifythecausalrelationbetweenrobotsandlabour demand.Economicfeasibilitymotivatesourdistinctionoftheimpact ofrobotadoptionbetween advancedanddevelopingcountries.Using two-stageleastsquares(2SLS)estimation,wefindthatrobotadoption lowerstheemploymentshareofroutinemanualoccupations.This rela-tionisobservedinhigh-incomecountries,butnotinemergingmarket andtransitioneconomies.
Thispaperrelatestorecentstudiesthatexaminetheimpactofrobot adoptiononsocio-economicoutcomes.GraetzandMichaels(2018)find thatrobotadoptioncontributedtoanincreaseinproductivitygrowth across industries in high-income countriesbetween 1993 and 2007. Their findings suggest that robot adoption did not reduce employ-ment, which is corroborated in this paper. This is also observed by
Dauthetal.(2019),butnotbyAcemogluandRestrepo(2020),who ex-aminegeographicvariationinrobotadoptionacrosstheUnitedStates andfindthatrobotsarelabourreplacing.Dauthetal.(2019)use de-tailedlinkedemployer-employeedataforGermany toshow that dis-placementeffectsarecancelledoutbyreallocationeffects,suchthatin theaggregatenoemploymenteffectsfromrobotadoptionareobserved. Dataavailabilitydidnotallow GraetzandMichaels(2018)to exam-inetheimpactofrobotsonworkersthatperformdifferenttasks.Yet,
Autor(2015)emphasizesthatworkerswithroutinetask-intensive occu-pationsaremostlikelytobeaffectedbyautomation.Thispaperaimsto contributetoourunderstandingoftheimpactofrobotsonsuch occu-pationalshifts.
Theremainderofthispaperisorganizedasfollows. Section2 re-viewsthekeytheoreticalmechanismsbetweenautomationandlabour demand.Section3describesthemethodology andinstrumental vari-ables.Section4documentspatternsintheoccupationalstructureofthe workforceandrobotadoption.Section5empiricallystudiestheimpact ofrobotadoptiononthetaskcontentoflabourdemand.Section6 con-cludes.
2. Theoreticalframework
Thissectionstartswithadiscussionofrobotadoptioninthecontext ofatraditionalcapital-labourmodel.Inthismodel,technologyis factor-augmenting:itincreasestheefficiencyofoneoftheproductionfactors employed(AcemogluandAutor,2011).Themodelputsthefocuson thecomplementarityandsubstitutabilitybetweenrobotsandtasks per-formedbyworkers.Wethendescriberecentmodellingeffortsthat em-phasizetheabilityofmachinestoreplaceworkersinawideningrange oftasks(AcemogluandRestrepo,2018).Thesemodelshelptoclarify mechanismsbywhichrobotsmayimpactlabourdemandandmotivate ourempiricalanalysis.
Themodelswedescribeanalysetheimpactofautomation. Automa-tionreferstocomputer-assistedmachines,robotics,andartificial intel-ligence(AcemogluandRestrepo,2018).Thus,robotsareasubsetof automation.Robotsaredrivenbyalgorithms,whichhavebecome in-creasinglycomplex.Theycannowoperatewithoutrequiringanyoneto explicitlyprogramthemechanismsofthetasksperformed.Yet,notall algorithmsdriveaphysicalmachine.Infact,manyalgorithmsare em-bodiedindevicesorapplications.Oncethesealgorithmsaredesigned, theycanbeusedformanytasksanywhereandatanytime.Forrobots, thealgorithmsareembodiedinthemachines.Expandingtherangeof tasksperformedbyrobotsthusrequiresinvestinginrobots,i.e.robots arerival(MartensandTolan,2018).Thiscontraststoalgorithms,which arenon-rivalinnature.Robotsaremorefrequentlystudiedin empiri-calworkbecauseoftheavailabilityofstatisticsontheiruse.However,
giventhepropertiesofrobotics,studiesthatuserobotdatacaptureonly partoftheimpactofautomationonlabour.
In the traditional model, automation enhances the productiv-ity of workers by complementing the tasks they perform (see e.g.
Autoretal.1998; Feenstra,2008; VanReenen2011).Yet,for work-erswhoperformtasksthatcanbesubstitutedbyautomation, increas-ingavailabilityofmachineswilllowertheirlabourdemand.Scholars havearguedthatnewtechnologiestendtosubstitute foroccupations thatareintensiveinroutinetasks,suchasassemblers,andcomplement non-routinetask-intensiveoccupations,suchasmanagersand techni-calscientists(Autoretal.2003;VanReenen2011;Goosetal.2014;
Dauthetal.2019).Thisisbecauseforroutinetasks,suchasmonitoring, measuring,controlling,andcalculating,therearewell-specified proce-dureswhich allowthetask tobe automated.Yet,knowingtherules thatgoverntaskproceduresisnotatrivialrequirement.Formany non-routine tasks, suchasthoserequiringcreativity andproblem-solving skills,automationisdifficultandrathercomplementstheperformance ofthesetasksdonebyhumans.Inlinewiththisreasoning,ananalysis forWesternEuropeancountriesbyGoosetal.(2014)findsthatrecent technologicalprogresshasbeenreplacingworkersdoingroutinetasks. Thisisreferredtoas“routine-biasedtechnologicalchange” (RBTC).3
Predictionsinthetraditionalmodelarestraightforward.Firmsadopt robotsifitiseconomicallyfeasibletodoso,whichisthecasewhen prof-itsexceedpurchasingandinstallationcosts.Therefore,substitutionof robotsforroutinetasksismorelikelyincountrieswithhigherwage lev-els,andthereafallinthefixedcostsortherentalpricewillresultinan increaseinrobotadoption(GraetzandMichaels,2018).
RecentmodellingeffortsbyAcemogluandRestrepo(2018)adda distinctive featureof automation:thetechnical abilityof machinesto replaceworkersinawideningrangeoftasks.Theysplittheproduction processintotasksdonebyworkersandmachines.Advancesinmachine capabilitiesexpandthesetoftaskscarriedoutbymachinesandreplace labour,thusloweringlabourdemand.
However,roboticautomationtechnologiesalsoresultinthecreation ofnewtasksthatcannotbedonebymachines,suchasprogramming, design, andmaintenanceofhigh-techequipment (Acemogluand Re-strepo,2019).This‘re-instatementeffect’increaseslabourdemand.The combinationoftasksdisplacedbyrobotsandthere-instatementofnew tasksdeterminethereallocationoftasksbetweenworkersandmachines. ComplementaritybetweenmanandmachineintheAcemogluand Restrepo(2018)modeloriginatesfromtwoindirecteffectsthatcome ontopofcomplementarityeffectsinthetraditionalmodel(Martensand Tolan, 2018). The first is a price-productivity effect whereby robot adoptionlowerspricesofproducedgoods,leadingtheindustryto ex-pandsalesandincreaseitsdemandforlabour.Thesecondisa scale-productivityeffectwherebyloweraggregategoods’pricesenablethe (local)economytoexpandandthusalsoincreaselabourdemand.The overall impact of robotization on labour demand then depends on whetherthedisplacementorthecomplementaryeffectsdominate.So far,empiricalevidenceontheaggregateemploymenteffectsfrom robo-tizationareinconclusive.4
In line with Acemoglu and Restrepo (2018), Graetz and Michaels (2018) model the production process as a continuum of tasks. Yet, Graetz and Michaels (2018) assume that products differ
3Autoretal.(2003)examinetheimpactofcomputerizationonlabour de-mandinU.S.industriesfrom1960-1998.Theyfindapositiverelationbetween thedemandfornon-routinetasksandcomputerizingindustries.Ross(2017)and
DeLaRicaetal.(2020)studytheimpactofRBTConthewagepremiumforjob tasks.
4AcemogluandRestrepo(2020)findthatrobotadoptionlowerslabour
de-mand in US locallabour markets. Dauthet al.(2019) argue in an
analy-sisforGermanythatworkersdisplacedbyrobotsreallocatetoservicesand there is no decline in aggregate employment. In a cross-country analysis,
Ghodsietal.(2020)findthatrobotadoptiondoesnotsignificantlyaffect ag-gregateemployment,althoughtheimpactvariesattheindustrylevel.
intheshare oftasksthat canbe carriedout bymachines.Garments provide aclearexample:sewinggarmentsis acomplex processthat requireshumanintuitionanddexterity,whichisdifficulttoprogram. Incontrast,ithasproven easiertoprogramrobots toperform tasks inautomobileassemblylines.5 Automationofcarassembly lineshas helpedtoreduce error rates andenhancesthecontrol ofrepeatable tasks.Thetechnicalfeasibilityofmachinestakingovertasksthusdiffers byindustry.
Inthis expandedmodel,theimprovementofmachinecapabilities maydrive automation.6 That is,ifrobot adoption is constrained by theproductionnatureofcertainindustries,therentalpriceof robots doesnotmatter.Rather,itisanexpansioninmachinecapabilitiesthat willdriveautomation.Giventhatlabourcostsarehigherinadvanced economies,therelaxingoftechnologicalconstraintsbyexpandingrobot capabilitieswillleadtohighereconomicincentivesforrobotizationin advancedcountriesandhencestrongeremploymentresponses.
Thetraditionalandexpandedmodelcapturethekeyeconomic mech-anismsdrivingrobotadoptionandtheiremploymenteffects.ThePRCis aninterestingcasetoillustratehowadditionalfactorsdriverobot adop-tion.WagelevelsinChinaarebelowhigh-incomecountries,butitis theworld’slargestadopterofindustrialrobots(Chengetal.2019).This seemscounterintuitivetothemodellingofrobotadoption.Yet,robotuse inChinadoescoincidewithrisingwagesandaslowdowninthegrowth ofitsworking-agepopulation.Besideslabourcosts,concernsover prod-uctqualityandproductionexpansionarefoundtoinfluencedecisions byfirmsinadoptingrobots(Chengetal.2019).Inaddition,the Chi-nesegovernmenthasinitiatedvariousprogramsandprovidessubsidies thatencouragethedevelopmentoftheroboticsindustry(Yang,2017;
Lin,2018).
Robotsmayalsoreversethetrendtorelocatefabricationactivities fromadvanced towardslow-wagecountries.Inaninteresting contri-bution,Faber(2018)pointsoutthatadvancesinroboticswillreduce productioncosts,nomatterwheretheproductisproduced.That,he ar-gues,willincreasetheattractivenessofproducingdomesticallyrelative tooffshoring.Ineffect,workersinexportsectorsofdevelopingcountries canbedisplacedbytheadoptionofrobots,eitheronshoreoroffshore. Essentially,foreignrobotsactasaformofcompetitionontheexport market.UsingamethodologicalapproachsimilartoAcemogluand Re-strepo(2020),Faber(2018)findsthatUSrobotadoptionlowerslabour demandinMexicanexport-producingsectors.7
Thesemodelsinformtheempiricalanalysisinourpaper.Thenext sections describe the methodology and data to examine the aggre-gate(cross-country)implicationsofrobotization.Weviewthis analy-sisasacomplementaryapproachtothewithin-countrycomparisonsin
AcemogluandRestrepo(2020),Dauthetal.(2019),andFaber(2018).
3. Methodology
Toexaminetherelationbetweenrobotadoptionandchangesinthe structureoftheworkforce,weestimateregressionssimilartothosein
GraetzandMichaels(2018)thattaketheform
ΔLci=βΔRobotadoptionci+𝐗′ciγ +δc+εci, (1)
5 Clearly,sometextileproductioncannowalsobenearlyfullyautomated;an exampleistheAdidas’Speedfactory‘(Faber,2018).Yet,relativelyspeaking, theshareoftasksthatrobotscanperformvariesacrossindustries.
6 Wethankananonymousrefereeforpointingthisout.
7 Ifrobotsresultinreshoringofafactory,thiswillaffectallworkersatthe exportingplant inthedevelopingcountry.Faber(2018)findsthatMexican workersincommutingzonesmostaffectedbyU.S.robotsarelow-educated ma-chineoperatorsandtechniciansinmanufacturingandhigh-educatedworkersin managerialandprofessionaloccupations.UsingtheWorldInput-OutputTables,
Krenzetal.(2018)findevidenceforapositiverelationbetweenreshoringand thedegreeofautomation.
where∆Lciisthechangeintheemploymentoutcomeofinterestin
in-dustry iof countryc.8 ∆Robot adoption
ciis thechange oftherobot
stockrelativetolabourinputineachcountry-industrypair.9Most spec-ificationsincludecontrolvariableswhicharechangesin:investmentto valueaddedratios,and(thenaturallogarithmof)valueadded.Wealso examine resultscontrollingfortheadoptionofinformationand com-municationtechnologies(discussedbelow).𝛿crepresentscountryfixed
effects,whichinafirst-differenceequationareequivalentto country-specifictimetrendsinalevels’equation.Regressionsareestimatedin long-runchangesbetween2005and2015becauseweareinterestedin longer-termtrends.Theregressionsweightindustriesusingtheir2005 employmentshareswithineachcountry.Thisensuresthatestimates re-flecttheimportanceofindustrieswithincountries,butwegiveequal weighttocountriesintheanalysis(ase.g.inGraetzandMichaels,2018). Weuseheteroscedasticity-robuststandarderrorsthataretwo-way clus-teredbycountryandindustry.10Thisisaconservativeapproachbecause theresultingstandarderrorsaretypicallylargercomparedtoone-way clusteringbycountryorindustry.
3.1. Endogeneityconcernsand2SLSestimation
Estimating(1)usingOLSraisesseveralconcernsaboutendogeneity. First,onemightworryaboutreversecausalityandomittedvariablebias. Forinstance,industriesthatexperienceafastergrowthinproduct de-mandmayinvestmoreinrobots.Especiallyifthelabourmarketistight, apositivedemandshockismorelikelytoresultininvestmentinrobots ratherthananexpansionofemployment(Faber,2018).11Thisisacase ofreversecausality,becauseloweremploymentgrowthresultsinhigher robotadoption.Also,relevantvariablesmightbeomittedfromthe re-gressionanalysis.Forinstance,Harriganetal.(2016)findthatadoption ofnewtechnologiesismediatedbytechnicallyqualifiedworkers. Sec-ond,onemayworryaboutattenuationbiasof𝛽 in(1)dueto measure-menterrorinthevariablerobotadoption.Clearly,theavailabledata onrobotadoption,discussedinSection4.1,isimperfect,asitdoesnot informonthequalityandothercharacteristicsofrobotsinstalled. In addition,weestimateregressionspecificationsinchanges,whichmay worsenthesignal-to-noiseratiocomparedtoregressionsofvariablesin levels.Duetomeasurementerror,thevariablerobotadoptioncouldbe correlatedwiththeerrorterm𝜀ciandOLSestimationof𝛽 wouldbe
bi-aseddownwards.Finally,industriesthatadoptrobotsmaydifferfrom otherindustriesinnon-randomways,whichwouldalsobiasthe coeffi-cientifnotappropriatelycontrolledfor.Hence,thedirectionofbiasin
𝛽 isnotclearapriori,althoughthepreviousliteraturesuggeststhata downwardbiasinOLSismorelikely(e.g.GraetzandMichaels,2018). In an attempt to address these endogeneity concerns, we use two industry-specific instruments introduced by Graetz and
8Theemploymentoutcomeofinterestiseithertheaverageannual percent-agegrowthrateinemploymentbycountry-industrypair,whichisestimatedas ((ln(EMPci,2015/EMPci,2005))/10)∗100,oritisthechangeinthetask-specific employmentsharebycountry-industrypair,measuredasthesharein2015 mi-nusthesharein2005.
9Robotadoptionisdefinedasthenumberofrobotsinstalledperthousand personsemployed.WefollowGraetzandMichaels(2018)andusethepercentile rankofthechangeinrobotadoptionasourmainexplanatoryvariable.Thisis furtherelaborateduponinSection4.1.
10WeimplementStata’s‘ivreg2’commandforOLSand2SLSregressions. Two-wayclusteredstandarderrorsarerobust toarbitraryheteroscedasticity and intra-groupcorrelationwithineachofthetwo(non-nested)categories “coun-try” and“industry” (Cameronetal.2012).Thisallowsforrobustinference, forexample,iferrorsarecorrelatedwithincountries(e.g.duetounobserved country-specificpolicies)andhaveseparatecorrelationstructureswithin indus-tries(e.g.duetotechnologyshocks).
11InhisanalysisoftheMexicanlabourmarket,Faber(2018)pointsoutthata positivedemandshockduetotheNorthAmericanFreeTradeAgreementmay haveputupwardpressureonindustriesorlocallabourmarketstoadoptrobots iftheyhadlessroomtoexpandemployment.
Michaels(2018)andestimate (1)using 2SLS.12 Thefirstinstrument measurestheshareofeachindustry’s labourinputthatisreplaceable byrobots.Thisinstrumentisconstructedusinginformationonthetasks performedbyrobots(IFR,2012).Asdiscussedabove,theextentof robo-tizationforeachtaskcouldbeendogenoustoindustryconditions. There-fore,GraetzandMichaels(2018)useinformationonUSoccupationsin eachindustryfromthe1980census,whichdatesbackbeforetheriseof robots.Occupationsaredefinedas‘replaceable’if(partof)theirtasks couldhavebeenreplacedbyrobotsin2012.Theythencomputethe fractionofhoursworkedineachindustryin1980thatwasperformed byoccupationsthatsubsequentlybecamemorepronetoreplacement byrobots.Thisinstrumentisnotwithoutlimitations:itisbasedondata fromtheUSandlaboursharesmightthereforebedifferentifconstructed usingdatafromothercountries.13
Thesecondinstrumentismotivatedbyrapidimprovementsinthe abilityofroboticarmstoperform‘reachingandhandling’tasks.It mea-surestheprevalenceofoccupationsineachindustrythatrequire reach-ingandhandlingtaskscomparedtootherphysicaldemandsin1980,prior torobotadoption.Roboticarmsareasalientcharacteristicofrobots, andmuchtechnologicaladvancesarelinkedtothedevelopmentofthese roboticarms(GraetzandMichaels,2018).Itisthereforemorelikelythat roboticarmsareatechnologicalcharacteristicofrobots,lessdrivenby thedemandside(duetoindustries’taskrequirements),whichcould re-flectreversecausality.Thisinstrumentisconstructedusingtheextent towhichoccupationsineachUSindustryrequirereachingandhandling taskscomparedtootherphysicaltasksin1980.14Similarlimitationsas tothefirstinstrumentapplyhere,butonemayarguethatthis instru-mentislesslikelytoviolatetheexclusionrestriction.
Clearly,neitherinstrumentcanguaranteetoresolveall endogene-ityconcerns.Bothinstrumentsreflectvariationacrossindustriesinthe shareoftasksthatarepotentiallyreplaceablebyrobots,whichmay cor-relatewithotherchangesovertime.Nevertheless,theinstrumentsare helpfultocontrastOLSwith2SLSresults.
4. Dataanddescriptiveanalysis
WefirstdescribethedataonrobotsandoccupationsinSection4.1. DescriptivestatisticsarepresentedinSection4.2.
4.1. Occupationsandrobots
Wecombinetwodatasetswithinformationonoccupationsandrobot purchases.Thefirstdatasetwithoccupationalemploymentby country-industryoriginatesfromReijndersanddeVries(2018)andwasupdated byBuckleyetal.(2020).Thedataisconstructedusingdetailedsurvey andcensusdatafromstatisticalofficesfortheperiodfrom2000to2015. Thesourcesusedinconstructingthisdatasetcloselyalignwiththose fromotherstudies.15Thedatasetprovidesemploymentforthirteen oc-cupationalgroupingsbycountry-industrypairs.Itcovers40countries, namelythe27membersoftheEuropeanUnion(perJanuary2007), Aus-tralia,Brazil,Canada,India,Indonesia,Japan,Mexico,thePRC,Russia, SouthKorea,ChineseTaipei, TurkeyandtheUnitedStates.Foreach
12 Theinstrumentsarecomputedfor2-digitindustriesintheISICrevision3 classification,whichmatcheswiththeindustryinformationonrobotstocksand occupationalemploymentsharespresentedinSection4.1.Notethatthe instru-mentsdonotvaryacrosscountriesbutonlyacrossindustries.
13 Alsonotethereplacementvaluesareanupperboundbecauseoccupations areconsideredtobereplaceableevenifonlypartoftheirworkcanbereplaced byrobots.
14 InformationonthetaskcontentofoccupationsistakenfromtheDictionary ofOccupationalTitles.
15 Forexample,fortheU.S.,thesourcesarethe2000Censusandtheannual AmericanCommunitySurveys.ThesesourcesarealsousedinAutor(2015). DataforEuropeancountriesarefromtheharmonizedindividuallevelEuropean UnionLabourForceSurveys,whicharealsousedinGoosetal.(2014).
ofthesecountries,occupationalemploymentsharesby35ISICrevision 3.1industriesthatcovertheoveralleconomyaredistinguished.They include14 two-digitmanufacturingindustries(suchastextile manu-facturingandelectronicsmanufacturing),aswellasagriculture, min-ing,construction,utilities,finance,businessservices,personalservices, tradeandtransportservices,andpublicservicesindustries.Thedataset thushasdimensionsof13occupationalgroupings×35industries×40 countries×16years.Occupationdataisintrinsicallynotexactly compa-rableacrosscountries,andinpracticewillalsovaryduetodifferences inthetypeofsourcesandnationaldatacollectionpractices. Intertempo-ralchangeswithincountry-industriesarelikelymoreconsistentbecause
ReijndersanddeVries(2018)usedatafromthesamenationalsource for eachcountry. Ourempiricalanalysisexploits thiswithin-country variation.
Weexaminetheimpactofrobotadoptionontasks,whichwe dis-tinguish intoroutine versus non-routine andmanual versus analytic tasks.Ourmeasurementstrategyistoinfertheimpactofrobot adop-tionontasksfromdataontheoccupationalstructureoftheworkforce. Thedistinction betweenoccupationswithdifferent taskintensitiesis basedontheso-calledRoutineTaskIntensity(RTI)indexdevelopedby
Autoretal.(2003)andmappedintotheInternationalStandard Classifi-cationofOccupations(ISCO88)byGoosetal.(2014).Table1provides theallocationofoccupationalgroupingstotasks.
The second database includes deliveries of industrial robots by country-industryfromtheInternationalFederationofRobotics(IFR).16 TheIFRprovidescountrydataonthenumberofindustrialrobots de-liveredfrom 1993onwards. Yet coveragevaries andthebreakdown ofrobotinvestmentbycountry-industryisonlyconsistentlyavailable formostcountriesafter2004.Inaddition,robotinvestmentsincreased rapidlyduringthe2000s.Wethereforebuildthedatasetusing informa-tionforallavailableyearsbutfocusontheperiodfrom2005to2015 intheempiricalanalysis.17
Weusetheperpetualinventorymethodtobuildrobotstocks, assum-ingadepreciationrateof10%asinGraetzandMichaels(2018).18We thendefine‘robotdensification’orsimply‘robotadoption’astherobot stock perthousandpersonsemployed.Weexamine changesin robot adoptionovertime.Thedistributionofchangesinrobotadoptionfor thecountry-industriesincludedinouranalysishasmostlyeitherzeroor smallpositivevalues,withalongrighttail.Analysingrawchangesin robotdensityisthereforenotrecommendableandweusethepercentile ofchangesinrobotadoption(basedontheemployment-weighted dis-tributionofchanges)asinGraetzandMichaels(2018).19
16Purchasesofservicesrobotsareonlyavailableforrecentyearsandfew coun-tries,whichlimitsstudyingtheimpactontaskdemandofrobotadoptionin servicessectors.
17Programcodetoreplicatetheanalysisisavailablefromtheauthorsupon request.
18The perpetual inventory method to build robot stocks is: RS
ci,t = (1-d)∗RS
ci,t-1+RDci,t ,whereRSistherobotstockofindustryiincountrycat timet;RDarerobotdeliveries,anddisthedepreciationrate.Ourmainresults arerobusttobuildingtherobotstockusinga5anda15percentdepreciation rate.
19We follow Graetz and Michaels (2018) and calculate
within-country employment-weighted distributions of changes in robot
adoption between 2005 and 2015. We use the Stata code
that Graetz and Michaels (2018) made available at https:// dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/5JWBXU. Specifically,wedenoterobotadoptionbyRAci,t=RSci,t/EMPci,t,i.e.therobot stockperthousandpersonsemployedinindustryiofcountryc.Wedenotewsc theweightedchangeinrobotadoptionofcountryc,whichisthesummationof changesinrobotadoptionbyindustryiweightedbytheiremploymentshares. Thechangeinrobotadoptionnetoftheweightedchangeinrobotadoption isΔRAci= (RAci,t -RAci,t-1)-wsc.Wethencalculate thepercentilerankof thechangeinrobotadoption(ΔRAci)andusethisvariableintheregression analysis.Theuseof percentilesiscommon intheeconomicsliterature and helpfulwhenthedataisskewed,seeforexampleAutoretal.(2003).
Table1
Mappingoccupationstotasks.
Routine Non-routine
Manual Production workers (71-74, 81-82, 93) Agricultural
workers (61-62, 92) Others (01, 999)
Support-services workers (51, 910, 912-916) Drivers (83)
Analytic Administrative workers (41-42) Legislators (11) Managers (12-13) Engineers (21, 31) Health professionals
(22, 32) Teaching professionals (23, 33) Other professionals (24, 34) Sales workers (52, 911)
Notes:MappingofthirteenoccupationsfromReijndersanddeVries(2018)tofourdifferentgroupsbasedonAutoretal.(2003)andGoosetal.(2014).Numbers inbracketsrefertoInternationalStandardClassificationofOccupationscodes(ISCO88).
Table2
Descriptivestatistics.
Obs. Mean SD p5 p95
Dependent variables
Employment growth (average annual, in %) 700 -0.78 3.41 -6.0 3.9
Δ Routine employment share 700 -0.04 0.10 -0.2 0.1
Δ Routine manual employment share 700 -0.04 0.12 -0.2 0.1 Δ Routine analytic employment share 700 -0.00 0.05 -0.1 0.1 Δ Non-routine manual employment share 700 -0.00 0.06 -0.1 0.1 Δ Non-routine analytic employment share 700 0.04 0.10 -0.1 0.2
Independent variables
Percentile of changes in robot adoption 700 0.50 0.29 0.0 1.0
Robot adoption, 2005 700 2.23 10.17 0.0 10.5
Robot adoption, 2015 700 4.98 22.54 0.0 21.1
Δ Investment to value added ratio 700 0.02 0.69 -0.2 0.2 Δ (natural logarithm of) value added 700 0.21 0.60 -0.7 1.1 Percentile of changes in information technology adoption 277 0.51 0.29 0.0 1.0 Percentile of changes in communication technology adoption 277 0.50 0.30 0.0 1.0
IV: Reaching and handling tasks 700 0.45 0.05 0.3 0.5
IV: Replaceable tasks 700 0.25 0.12 0.0 0.4
Notes:A‘Δ’infrontofavariablereferstothechangebetween2005and2015.Forvariabledescriptions, seeSection4.1.Inthecolumns,‘obs’referstothenumberofobservations,SDthestandarddeviation,p5 the5thpercentile,andp95the95thpercentile.
Wematchthedataonrobot adoptionwithoccupational employ-ment.20Thenineteensectorsthatarematchedare14manufacturing in-dustries,agriculture,mining,utilities,construction,and‘educationand R&D’.The(unweighted)averageemploymentshareofthesesectorsin thetotaleconomyacrossthesampledcountriesis46%and39%in2000 and2015,respectively.Thesharevariesacrosslevelsofdevelopment. Itisabout aquarteroftheworkforcein advancedcountriessuchas Denmark,theNetherlands,andtheUnitedStatesthroughoutthe sam-pleperiod.Itisover50%oftotalpersonsemployedinindustrializers suchasthePRC,Turkey,andPoland.
Inmostregressionspecifications,wecontrolforchangesinthe in-vestmenttovalueaddedratios,and(thenaturallogarithmof)value added.21 Although robotsare a visible andmuch discussedform of
20 Aftermatchingthedatasets, wehavedatafor37countriesand19 sec-tors,withmissingdataforafewcountry-industrypairs.High-income coun-triesincludethe‘old’EU15countries,westernoffshoots,andhigh-incomeEast Asiancountries,namelyAustralia,Austria,Belgium,Canada,Germany, Den-mark,Spain,Finland,France,theUnitedKingdom,Greece,Ireland,Italy,Japan, SouthKorea,Malta,theNetherlands,Portugal,Sweden,ChineseTaipei,andthe UnitedStates.EMTEsaretheothers,namelyBrazil,thePRC,CzechRepublic, Estonia,Hungary,Indonesia,India,Lithuania,Latvia,Mexico,Poland,Romania, Russia,Slovakia,Slovenia,andTurkey.
21 ThisdataisobtainedfromtheWIOD2016release(Timmeretal.2015). Thefirstcontrolvariable,investmenttovalueaddedratiosmaybesubjectto concernsaboutmulti-collinearityasrobotsarepartofphysicalcapital invest-ment.Weexploredtheshareofrobotinvestmentinoverallinvestmentbyusing turnover-basedpricesofrobotsfortheUSprovidedinIFR(2012).Thenumber ofrobottimestheirunitpricegivesaroughapproximationofnominal invest-ment.Ourestimatessuggestthattheshareofrobotinvestmentintotal invest-mentissmall,typicallynotexceeding1percent.Thefirstdifferencesofour dataforrobotadoptionandinvestmenttovalueaddedratiosareonlyloosely correlated,withacorrelationcoefficientof-0.06.
automation, computersandotherdigitaltechnologiesimpact jobsas well.InformationandCommunicationTechnologies(ICTs)havebeen foundtobe skill-biased,raisingtheproductivityof high-skilled work-ers and lowering demand for low-skilled workers (Feenstra 2008;
Michaelsetal.2014).Incontrast,robotsarepartofrecentinnovations andconsideredroutine-biased,astheysubstituteforworkersperforming routine-manualtasks(Goosetal.2014).Theseroutinetasksareoften performedbyworkerswithamiddlinglevelofeducation,suchas fab-ricationjobs involvingrepetitiveproductiontasks(Autor,2015).We therefore expectadirecteffectofrobot adoptiononthedemandfor routine-manualtask-intensiveoccupationsindependentofICT invest-ment.
TocontrolforICTadoption,weusedatafromtheEUKLEMSRelease 2019forgrossfixedcapitalformationincomputingandcommunication equipment(Stehreretal.2019).TheseICTinvestmentsareexpressedas ashareintotalinvestment.ChangesintheICTinvestmentshareare in-cludedintheanalysis,alsointheformofthepercentileofchangesinICT adoption(basedontheemployment-weighteddistributionofchanges).
4.2. Descriptiveanalysis
Table 2showsdescriptivestatistics ofourkeydependentand ex-planatoryvariables.Thetoprowsshowchangesinemploymentshares for occupations by task intensity. Onaverage,the routine (manual) employmentsharedeclinedby4percentagepointsbetween2005and 2015.Thistrendisobservedin35outof37countries,butthedecline intheroutinesharediffersacrosscountriesandindustries.Thiscanbe seeninAppendixFigs.1and2,whichdepictthechangesinemployment sharesforourfouroccupationalgroupingsbycountryandindustry, re-spectively.Thedeclineinroutinemanualoccupationsis mirroredby theriseofnon-routineanalyticjobs,whichincreasedby4percentage
Fig.1. Robotsandtheroutineemploymentshare.
Notes:Observationsarecountry-industrycells.Thesizeofeachcirclecorrespondstoanindustry’s2005within-countryemploymentshare.Verticalaxisdisplays thechangeintheroutineemploymentsharebetween2005and2015.Horizontalaxisofpanel(a)showsthepercentileofchangesinrobotadoption(basedonthe employment-weighteddistributionofchanges),seeSection4.1.Panel(b)changesinrobotadoption(basedontheemployment-weighteddistributionofchanges). Fittedregressionlinesareshown.Coefficients(standarderrors)ofthelinearfitarerespectively-0.00033(0.00010)and-0.0013(0.0004).Sources:seeSection4.1.
pointsonaverage.22Thecomparabilityoftheshiftsinroutinemanual andnon-routineanalyticoccupationsacrossoursampleofhigh-income countriesandEMTEsmakesitlikelythatacommonsetofforces con-tributestoshareddevelopmentsinlabourmarkets.Theprimesuspect isautomation(Autor, 2015).Atthesametime,variationin country-specificexperiencesunderscoresthatnocommoncausewillexplainthe fulldiversityoflabourmarketdevelopmentsacrosstheseeconomies.
Theaveragerobotstockperthousandpersonsemployedmorethan doubledfrom2.23in2005to4.98in2015.Thestandarddeviationof robotizationrevealssubstantialvariationinrobotizationacross coun-triesandindustries.Mostofthisvariationstemsfromcross-industry dif-ferenceswithincountriesasopposedtovariationbetweencountries.23 Morerobotswereinstalledinallcountries,withthenumberofrobots perthousandpersonsemployedsurginginGermany,Japan,andSouth Korea(seeAppendixFig.A3).24Highrobotdensityisobservedin ma-chinery,electronics,andautomotive(seeAppendixFig.A4).For indus-triesthatproducechemicalsandmetalproductswealsoobservean in-creaseinrobotdensity,albeitstartingfromlowlevels.
AppendixFig.A5showsthenumber ofrobots per1,000 persons employedbyindustryinthePRCandGermanyfor2015.Thisfigure helpsclarifythelowerlevelofrobotsperthousandpersonsemployedin China.Forexample,in2015,thenumberofrobotsinstalledinChina’s automotiveindustrywasabout50,000,whichcomparestoaslightly lowernumberofaround48,500robotsinthatindustryforGermany. Yet,in2015thenumberofpersonsemployedinautomotiveisabout 6.8millioninChinacomparedto965thousandinGermany,soa fac-tor7differenceinthesizeoftheworkforceinthatindustry.Hencethe
22 Changesinthesharesofroutineanalyticandnon-routinemanualjobsare typicallysmallerandweobservesubstantialvariationacrosscountries(see Ap-pendixFig.A1).
23 Thestandarddeviationoftherobotstockperthousandemployedbetween countriesis8.06in2015.Incomparison,thestandarddeviationofrobot adop-tionwithincountriesis21.06in2015.Thosearecalculated,respectively,as thestandarddeviationsofcountrymeans𝑥𝑐andoftheirdeviations𝑥𝑐𝑖−𝑥𝑐+𝑥, wherexindicatesrobotadoptionand𝑥isitsglobalaverage.
24 ForJapan,reporteddeliveriesandstocksofrobotschangedovertimedue toa reclassificationof machinesasrobots(GraetzandMichaels, 2018).In
Section5.2weshowthatthemainresultsarerobusttodroppingJapanfrom thesample.
numberofrobotsinstalledperthousandpersonsemployedisabout7in Chinacomparedto50inGermany.
Table2alsoprovidesdescriptivestatisticsfortheinstrumentsand controlvariables.Theinstrumentsreplaceabletasksandreachingand han-dling tasksarepositivelycorrelated,butdifferent.25 Forexample,the highestshareofreplaceabletasksisobservedinautomotiveandmetal manufacturing,whereastheextentofreachingandhandlingstasksis highestintextileandfoodmanufacturing.
Fig.1plotsthechangeintheroutineemploymentshareagainst mea-suresofincreasedrobotuse.Insub-figure(a),weplotthepercentileof thechangeinrobotdensitynetofcountrytrendsonthehorizontalaxis, aswellasthefittedregressionline.Theslopeisnegativeandstatistically significant.Thedistributionofdatapointsaroundthefittedlinesuggest thattherelationshipbetweentheroutineshareandthepercentileof robotdensificationiswellapproximatedbyalinearfunctionalform.In subfigure(b),weinsteadplotchangesinrobotdensityonthehorizontal axis(againnetofcountrytrends),togetherwiththefittedline.Herea linearfunctionalform(thoughalsonegativeandsignificantat conven-tionallevels)seemsmuchlessadequate,andtheestimatedslopeappears sensitivetoseveraloutlyingobservationsnearthetopofthe distribu-tionofrobotdensification.Thus,followingGraetzandMichaels(2018), intheregressionanalysiswewillusethepercentileofchangesinrobot densification.
Panel(a)ofFig.2showsadescriptiverelationbetweenrobot adop-tionandindustryaveragechangesintheroutineemploymentshare be-tween2005and2015(seeTableA1fortheindustrydescriptions).We observea(slightly)strongerreductionintheroutineshareforindustries thatinvestedmoreinrobots.Sectorssuchaspaperandutilities experi-encedadeclineintheshareofroutinejobswithonlyarelativelysmall increaseinrobotization.Inmanufacturingindustriessuchasmachinery, electronics,andautomotive,weobserveadecreaseintheshareof rou-tinejobs.Theseindustriesarealsoamongtheoneswiththestrongest increaseinrobotadoption.Panels(b)and(c)suggestbothinstruments aregoodpredictors, asindustrieswitha highershareof replaceable tasksorthosemoreintensiveinreachingandhandlingtaskshave
in-25NotetheinstrumentsaremeasuredbyindustrybasedondatafortheUS(see Section4.1)andmatchedtothecountry-industrypairs.
Fig.2. Cross-industryvariationinIVsandchangesintheroutineemployment share.
Notes:Onthehorizontalaxisisthe(unweighted)averagepercentileofchanges inrobotadoptionbyindustry.Inpanel(a),theverticalaxisshowsthe
indus-try(unweighted)average changein theroutineemploymentshare between
2005and2015.Thecoefficient(standarderror)ofthelinearfitinpanel(a) is-0.013(0.007).Theverticalaxisofpanels(b)and(c)showthevaluesfor theinstruments,coefficients(standarderrors)ofthelinearfitarerespectively 0.59(0.11)and0.20(0.07).Sources:seeSection4.1.
stalledmorerobotscomparedtoothers.Thenextsectionformallytests theserelationships.
5. Econometricresults
We present our main results from OLS and 2SLS regressions in
Section 5.1. We find thatrobot adoption relatestoa decline in the employmentshareofoccupationswithahighcontentofmanual rou-tinetasks.InSection5.2wepresentseveralextensionsandrobustness checks.Wefirstdocumentthatresultsappearneitherdrivenbyspecific sectorsorcountriesnorspuriousindustrytrends.Wethenexploit het-erogeneityintaskintensityacross(blue-collar)productionworkersand findthatrobot adoptionrelatestodecliningdemandforoccupations thataremoreintensivein routinetasks.Finally,weexplorewhether globaldevelopmentsinrobotizationimpactlabourdemandinEMTEs.
5.1. MainOLSand2SLSresults
Ourmainregressionresults aresummarizedinTable3,withOLS results inpanelA and2SLSresults inpanelB.Westarttheanalysis byregressingtheaverageannualpercentagegrowthofemploymenton robotadoption.Countryfixedeffectsareincluded;thus,coefficientsare identifiedfromvariationacrossindustries.Weuseaconservative two-wayclusteringofstandarderrorsatthecountryandindustrylevel. Col-umn1ofTable3indicatesthatrobotadoptionisnegativelycorrelated withtheaveragegrowthrateofemploymentbetween2005and2015. However,thisrelationshipisnotstatisticallydifferentfromzero.It sug-gestsrobotadoptionisnotlabourreplacing,whichwasalsoobserved byGraetzandMichaels(2018).Ourfindingindicatesthisresultholds inalargercountrysample.
Incolumn(2)ofTable3,weexaminetherelationbetweenrobot adoptionandtheshareofroutinejobs.Wefindthatincreasedrobot usecontributestoadeclineintheroutineemploymentshare.Toassess theeconomicmagnitude,considerthedifferencebetweenanindustry withamediantrendinrobotadoptionandanindustrywithnorobot adoption,whichequals0.5x-0.047=-0.02intheOLSregression.This differenceamountstoabout59%oftheaveragechangeintheroutine employmentshare(whichis-0.04,seeTable2).Whilethisindicatesa sizeableimpactofrobotsonoccupationalshifts,theR-squaredof2% incolumn(2)wherecountryfixedeffectsarepartialledout,indicates thatmanyotherfactorsthanrobotadoptionaffectchangesintheshare ofroutinejobs.Thecoefficientmorethandoublesinthe2SLS regres-sion,whereweusetheshareof replaceabletasksinindustriesasan instrument(panel B,column2).Theinstrumentispositivelyand sta-tisticallysignificantlycorrelatedwithrobotadoptioninthefirststage, whichisreportedincolumn(4)ofpanelB.Identificationisstrong,with theCragg-DonaldWaldFstatistic(268.53,assumingi.i.d.errors)and theKleibergen-PaapF-statistic(23.42)surpassingthe10%criticalvalue (16.38).Under-identificationisrejectedatthe5%levelofstatistical sig-nificance.Theconsiderableincreaseintheestimatedsecondstage co-efficientforrobotadoption,whencomparedtoOLSresults,mayreflect measurementerrorinourmainexplanatoryvariable:anincreaseinthe noise-to-signalratioinrobotadoptionwillbiasOLSestimatestowards zero.Moreover,theincreaseinthecoefficientin2SLSestimatesmay re-flectthatourinstrumentforrobotadoptiononlyvariesacrossindustries andthatglobalindustrytrendsimpactchangesinroutineemployment shares(seeSection5.2below).Using‘reachingandhandling’tasksas aninstrumentgivessimilarresults,althoughmorepronetoweak iden-tificationconcerns(seeAppendixTableA2).
Anadvantageofourdatasetisthebroadcountrycoverage,including variousemergingmarketand(post-)transitioneconomies.Incolumn (3)ofTable3,wedifferentiatetherelationbetweenrobotadoptionand routinesharesacrosshigh-incomecountriesandEMTEs.26Wedosoby
26GiventhenumberofrobotsinstalledinthePRC,itmightbelessappropriate toclassifyitasanEMTE.Tocheckforrobustnessofreportedresults,weomitted
Table3
Baselineregressionresultsofemploymentgrowthandchangeinroutineemploymentshare.
Panel A: OLS
(1) (2) (3) (4)
ΔEmployment ΔRoutineemployment share ΔRoutineemployment share Percentile of changes in robot adoption Percentile of changes in robot
adoption -0.354 -0.047
∗∗∗ -0.055 ∗∗∗
(0.73) (0.02) (0.02)
Percentile of changes in robot adoption x dummy EMTE
0.040 ∗∗∗
(0.02)
R2 0.001 0.025 0.028
Observations 700 700 700
Number of countries 37 37 37 37
Panel B: 2SLS (IV: Replaceable tasks)
Percentile of changes in robot adoption
-2.714 -0.120 ∗∗ -0.156 ∗∗
(3.03) (0.05) (0.06)
Percentile of changes in robot
adoption x dummy EMTE 0.136
∗∗
(0.06)
Replaceable tasks 0.892 ∗∗∗
(0.18)
Cragg-Donald Wald F statistic 268.53
Kleibergen-Paap F-statistic 23.42
Kleibergen-Paap under identification test ( p -value)
0.013
R2 -0.052 -0.027 -0.053
Observations 700 700 700 700
Number of countries 37 37 37 37
Notes:Robuststandarderrorsinparentheses.Multi-wayclusteringbycountryandindustry.Thedependentvariableemploymentgrowthincolumn(1)isthe averageannualpercentagegrowthinemploymentfortheperiodfrom2005to2015.Thedependentvariableincolumns(2)-(3)isthechangeintheroutine employmentsharebetween2005and2015.Column(4)reportsthefirststagefor2SLSestimation.Theshareofreplaceabletasksinanindustryisusedasan instrumentforrobotadoption.Regressionsincludethechangeintheinvestmenttovalueaddedratioandthechangein(thelogof)valueaddedbetween2005 and2015ascontrolvariables.CountryfixedeffectsareincludedinallregressionsandpartialledoutinthereportedR2.
∗p<0.1. ∗∗∗p<0.01 ∗∗p<0.05
interactingadummyvariableforEMTEswithrobotadoption.27The re-lationshipbetweenrobotadoptionanddecliningroutinesharesappears tomainlyoccurinhigh-incomecountries:forboth,theOLSand2SLS regressions,thenegativeoverallcoefficientestimateforrobotadoption incolumn(3)isalmostequalinsizetothepositiveinteractionterm withtheEMTEdummy,indicatingthattheeffectofrobotadoptionis essentiallynullifiedinthosecountries.28Sincetechnicalconstraintsto robotsreplacingtasksaremorelikely tobindforfirmsinhigh-wage advancedcountries,improvementsinrobotcapabilitiesmightaccount forthelargeremploymentresponseinadvancedcountriescomparedto EMTEs.
Additionally,ourdatasetallowsustofurtherdisaggregateroutine and non-routine employment shares into manual and analytic task-intensiveoccupations.ResultsarereportedinTable4,againwithOLS resultsinpanelAand2SLSresultsinpanelB.29Wefindthatthe neg-ativerelationbetweenrobotadoptionandroutineemploymentshares isexclusivelydrivenbymanualroutinejobs:theestimatesincolumn
Chinafromthesampleandre-classifieditasanon-EMTE.Thisdidnotalterthe results(availableuponrequest).
27 Inthereported2SLSregressions,weonlyinstrumentrobotadoptionbutnot theinteraction.Weadditionallyestimated2SLSregressionswiththe interac-tioninstrumented,whichrequiredinteractionofourinstrumentwithanEMTE dummyinthefirststage.Results,whichareavailableuponrequest,were quan-titativelyandqualitativelysimilartothosereported,butmorepronetoweak identificationconcerns.
28 OLSand2SLSestimatesof𝛽 arenotstatisticallysignificantlydifferentfrom zerowhenestimatingequation(1)forEMTEsonly.Resultsareavailableupon request.
29 Notethatfirststageresultsforthe2SLScasearethesameasinTable3.
(1)ofTable4essentiallymimicthoseofcolumn(2)inTable3,while norelationshipcanbefoundbetweenrobotadoptionandanalytic rou-tineemploymentshares(Table4,column2).Itthusappearsrobotsare bettersuitedtosubstituteforroutine-manualtasksduetotheability of robotstomanipulateobjects.Conversely,theshareofnon-routine analytic occupationspositivelyrelatestorobot adoption(column4). Thisisconsistentwiththeintuitionthatnon-routineanalytictasksare complementedbyrobotsinproduction(Autor,2015).Norelevant rela-tionshipisobservedbetweenrobotadoptionandchangesinthemanual non-routineemploymentshare(column3).
5.2. Robustnessandextensions
We performed several robustness checks. These are summarized in Section5.2.1.TheotherSectionsfocusonaspectsconsidered rele-vant tobetterunderstandtherelationbetween robotizationand rou-tine employmentsharesandtomotivatefutureresearchinthisarea.
Section5.2.2examinestherelationbetweenrobotadoptionacross pro-ductionoccupationsthatdifferintaskintensity.Section5.2.3examines whethertheresultsaredrivenbylonger-termindustrytrends.Finally,
Section5.2.4explorestheroleofglobalindustrytrendsinrobot adop-tionfordrivingcountry-industrychangesinemploymentshares.
5.2.1. Robustnessandheterogeneity
WefirstexamineregressionresultswhenaddingICTinvestmentto theanalysis.Thisisbecausecomputersseemparticularlysuitedto sub-stitute foranalytictasksandthedevelopmentof computerand com-municationequipmentisnotindependentofrobotadoption,suchthat omittingICTmaybiasthecoefficientforrobotadoption.Including vari-ablesforcomputerandcommunicationinvestmentleadstoa
consider-Table4
Robotadoptionandchangesinemploymentsharesbytasktype.
Panel A: OLS (1) (2) (3) (4) ΔRoutinemanual employment share ΔRoutineanalytic employment share ΔNon-routinemanual employment share ΔNon-routineanalytic employment share Percentile of changes in robot
adoption
-0.049 ∗∗∗ 0.002 -0.008 0.055 ∗∗∗
(0.02) (0.00) (0.01) (0.02)
Δ Investment to value added ratio 0.003 ∗∗∗ 0.001 -0.001 -0.003 ∗∗∗
(0.00) (0.00) (0.00) (0.00)
Δ (natural logarithm of) value added 0.005 0.002 0.004 -0.009 (0.01) (0.00) (0.00) (0.01) R2 0.024 0.003 0.007 0.031 Observations 700 700 700 700 Number of countries 37 37 37 37
Panel B: 2SLS (IV: Replaceable tasks) Percentile of changes in robot
adoption
-0.119 ∗∗ -0.003 -0.032 0.152 ∗∗∗
(0.05) (0.01) (0.02) (0.05)
Δ Investment to value added ratio 0.004 ∗∗∗ 0.001 -0.001 -0.004 ∗∗∗
(0.00) (0.00) (0.00) (0.00)
Δ (natural logarithm of) value added 0.012 0.003 0.006 -0.019 ∗∗ (0.01) (0.00) (0.01) (0.01) R2 -0.020 0.001 -0.021 -0.059 Observations 700 700 700 700 Number of countries 37 37 37 37
Notes:Robuststandarderrorsinparentheses.Multi-wayclusteringbycountryandindustry.Thedependentvariableisthechangeintherespectiveemployment sharebetween2005and2015.Theshareofreplaceabletasksinanindustryisusedasaninstrumentforrobotadoption.Countryfixedeffectsareincludedinall regressionsandpartialledoutinthereportedR2.
∗p<0.1. ∗∗∗p<0.01 ∗∗p<0.05
abledeclineinthesampleto277observationsbecausetheEUKLEMS datasetdoesnotreportICTinvestmentbyindustryformanyEMTEs. Theestimatedcoefficientfortherelationbetweenrobotadoptionand routineemploymentsharesissmallerbutremainsnegativeand statisti-callysignificantintheOLSandIVregressions(seecolumn1ofAppendix
TableA3).30
Toavoidresultsbeingdrivenbycertaincountries,weinspectthe pat-ternofOLSresiduals(depictedinAppendixFig.A6).Furthermore,we lookatthedistributionofcountry-specificparameterestimates,which weobtainbyinteractingrobotadoptionwithamatrixofcountrydummy variablesinourmainOLSspecification(seeAppendixFig.A7).Thereis aclusterofhighfittedvaluesforIreland(AppendixFig.A6,panelA)and tworesidualsfromRomaniaandSwedenobtainarelativelyhigh lever-ageandarepotentialoutliers(AppendixFig.A6,panelB).Moreover,the country-specificestimationcoefficientsinAppendixFig.A7suggest co-efficientestimatesforIreland,Lithuania,andLatviadeviatefromother countries.Wehenceexcludethese5countriesaswellasPortugal,which sawsomewhatdifferentemploymentdynamicsthantherestofour sam-ple,accordingtoourdescriptiveanalysis(cf.AppendixFig.A1).Results arereportedincolumn(2)ofAppendixTableA3.Droppingthese coun-triesdoesnotqualitativelyaffectourmainresult.31
Similarly,wealsocomputeindustry-specificcoefficientsforthe re-lationshipbetweenrobot adoptionandtheshareof routinejobs. Ap-pendixFig.A8suggeststhattheelectricity,gas,andwatersupplysector
30 Moreover,thechangeintheparameterestimateappearstooriginatefroma samplecompositioneffectandnotfromomittedICTvariables:re-estimatingthe baselinemodelwiththe277observationsforwhichICTdataisavailable pro-ducesthesamecoefficientforrobotadoptionasinthepresenceofICTvariables: -0.033∗∗∗.
31 Wealsoexcludedseveralofthosecountries/countrygroupsseparately,with equallyrobustresults.ThisalsoappliestoexcludingJapanfromtheanalysis, whichwasdroppedfromthesamplebyGraetzandMichaels(2018).
couldbeanoutlierthatpotentiallydrivestheoverallresult,together withtheeducationandR&Dsector,whichsawdifferentroutine employ-menttrendsaccordingtoourdescriptiveanalysis.Wethusre-estimate ourbaselineregressionsandsequentiallyomitthesesectors.Columns (3)and(4)ofAppendixTableA3suggestourresultsarenotdrivenby thesesectors,althoughomittingtheeducationandR&Dsectorin2SLS estimationpushesstatisticalsignificanceoftherobotadoption parame-terslightlybeyondthecritical10%level(forthenullhypothesisofno relationship).Tocheckwhethercountriesthataccountforthemajority ofrobotsinstalledaredrivingourestimates,wealsoexcludedJapan, SouthKorea,Germany,thePRCandtheUSfromourestimates,leaving thebaselineestimateforrobotizationunaffected.Forthesamerationale, wealsoexcludedthehighrobot-adoptingautomotiveandelectronic in-dustries(columns(5)and(6)ofAppendixTableA3respectively).All parameterestimatesforrobotadoptionwherenegativeandstatistically differentfrom0andt-testsdonotallowrejectingthenullhypothesis ofequalityoftheseparameterestimateswiththebaselineresult(atthe 10%levelofstatisticalsignificance).
Wealsoinvestigatedwhetherasamplesplitatthemedian(0.5)of thepercentilechangeinrobotadoptionaffectsourresults.Theresults indicatethattheparameterestimateforthesloweradopters(<0.5)are considerablyhigherbutestimatedwithlowprecision,sothattheyare not statisticallydifferent from0.Neitheroftheestimated OLSorIV parametersforthesamplesplitarestatisticallyspeakingdifferentfrom thosein thebaselineresultof column(2)intable3, inlinewithan approximatelylinearrelationshipsuggestedbypanel(a)inFig.1.32
32Wealsoexaminedresultswhenclusteringstandarderrorsatthecountry levelandnotclusteringatall.Thealternativetreatmentofstandarderrorsdoes notaffectthestatisticalsignificanceoftherelationbetweenrobotadoptionand theshareofroutinejobsintheOLSregressionsandthecoefficient(𝛽)isdifferent fromzeroatthe1%levelofstatisticalsignificanceinthe2SLSregressions.
Table5
Robotadoptionandchangesintheemploymentshareofproductionworkers.
Panel A: OLS
(1) (2) (3) (4) (5)
No weight RII weight (global average) RII weight (U.S.) RII weight (Germany) RTI weight Percentile of changes in robot adoption -0.031 ∗ -0.066 ∗∗∗ -0.065 ∗∗∗ -0.058 ∗∗∗ -0.103
(0.02) (0.02) (0.02) (0.02) (0.08)
R2 0.016 0.036 0.054 0.035 0.019
Observations 450 450 450 450 450
Panel B: 2SLS (IV: Replaceable tasks)
Percentile of changes in robot adoption -0.083 ∗ -0.122 ∗∗ -0.143 ∗∗∗ -0.113 ∗∗ -0.318 ∗
(0.04) (0.06) (0.05) (0.06) (0.19)
R2 -0.018 0.013 -0.021 0.006 -0.033
Observations 450 450 450 450 450
Notes:Robuststandarderrorsinparentheses.Multi-wayclusteringbycountryandindustry.Dependentvariableisthechangeintheemploymentshareofproduction workersbetween2005and2015,withweightsindicatedinthecolumnheader.InPanelB,theshareofreplaceabletasksinanindustryisusedasaninstrument forrobotadoption.CountryfixedeffectsareincludedinallregressionsandpartialledoutinthereportedR2.
∗∗∗p<0.01 ∗∗p<0.05 ∗ p<0.1.
5.2.2. Robotadoptionandproductionworkers
InTable1,productionworkersarecategorizedashavingahigh con-tentofroutine-manualtasks.Yet,productionworkersaretypically la-belledblue-collarworkers.Hence, therelation betweenrobotsanda decliningemploymentshareofroutinemanualjobscouldreflecta sub-stitutionofrobotsforbluecollarproductionworkers,insteadofa sub-stitutionforroutinetasks.
Itishardtoruleoutsuchanalternativeinterpretation.Yet,for24 countriesinoursampleweareabletodistinguishseven2-digitISCO occupationsthattogethercomprisetheoccupationalgroupinglabelled ‘productionworkers’(cf.Table1).33Theroutinetask-intensityforeach ofthese2-digitoccupationsisprovidedbyAutoretal.(2003)and, us-inganalternativeapproach,byMarcolinetal.(2019).Weusetheseto createaweightedaverageofthechangeintheemploymentshareof productionworkers.Theweightsweusearetheroutineintensityindex (RII)fromMarcolinetal.(2019)andtheroutinetaskintensity(RTI) gaugedbyAutoretal.(2003).Thetask-intensitybyoccupationis re-portedinAppendixTableA4.Clearly,thesevenoccupationslabelled productionworkersareheterogeneousinthecontentofroutinetasks.
ThefirstcolumnofTable5regressesthechangeintheemployment shareofproductionworkersonrobotadoption.Resultsindicatea signif-icantnegativerelationbetweenrobotadoptionandchangesintheshare of(routinemanualtask-intensive)productionjobs.Subsequentcolumns examinethesamerelation,butherechangesintheshareofproduction jobsarecalculatedasaroutinetask-intensityweightedaveragechange. Occupationsthathaveahighercontentofroutinetasksreceiveagreater weightinthisapproach.34
Weightingbyroutineintensitystrengthensthenegativeassociation betweenrobotizationandchangesintheshareofproductionjobs:the resultingparameterestimatesincolumns(2)-(5)arelargercompared tocolumn(1).Thisresultisobservedifweuseasweightstheglobal averageroutineintensity(RII)reportedbyMarcolinetal.(2019),see column(2),ortheRIIfortheUSorGermany(columns(3)and(4), re-spectively).ItisalsoobservedifweweightoccupationsusingtheRTI
33 ThesevenISCO2-digitoccupationsthatcanbedistinguishedareISCO88 codes71,72,73,74,81,82,and93.Thecountriesforwhichweareableto makethissplitareAustria,Belgium,CzechRepublic,Denmark,Estonia, Fin-land,France,Germany,Greece,Hungary,Ireland,Italy,Lithuania,Latvia,Malta, theNetherlands,Poland,Portugal,Romania,Slovakia,Slovenia,Spain,Sweden, Turkey,andtheUnitedKingdom.
34 Thetask-intensitymeasuresarePearson-transformed,i.e.centredat0with astandarddeviationof1.Weadded+1tothemeasure.Hence,anoccupation withmeanroutineintensitygetsaweightof1,abelow-averageroutineintensity occupationalowerweight,andanabove-averageroutineintensityoccupation aweightabove1(seeAppendixTableA4).
fromAutoretal.(2003),seecolumn(5),althoughtheparameteris es-timatedwithlessstatisticalprecisionintheOLSand2SLSregressions. Overall,theseresultsprovideadditionalevidencethatrobotadoption is relatedtoadeclineintheshareofoccupationsthathaveahigher contentofroutinetasks.
5.2.3. Controllingforlong-termindustrytrends
Aremainingconcernisthattherecouldbealong-rundeclineinthe shareofroutinetasksdonebyworkers,whichismorepronouncedin industriesinvestingmoreinrobotsyetnotdrivenbyrobotizationper se.Acommonwaytoexaminethisconcernistoregressemployment outcomes fromapre-periodontheperiodduringwhichrobots were adopted.
Ideally,wethusrelatepre-periodemploymentoutcomesonthe cur-rentriseofrobots.However,weareconstrainedbycross-country occu-pationsdatawhichareavailablefrom2000onwards.By2000,robots werealreadybeinginstalled(GraetzandMichaels,2018).Still, descrip-tivestatisticsinTable2forthenumberofrobotsperthousandpersons employedin2005and2015suggesttheybecameubiquitousfromthe mid-2000sonwards.
Incolumn(1)ofTable6wethereforeregressthechangeinthe rou-tineemploymentsharebetween2000and2005onourpost-2005 mea-sureofrobotadoption.Weindeedfindarelationship,althoughthe coef-ficientissmallerandlesspreciselyestimatedcomparedtoourbaseline results(cf.column(2)ofTable3).35Pre-trendcorrelationisanecessary conditionforunobservedsectorheterogeneity,butitisnotasufficient conditiontorenderidentificationinvalid.Thisispartlybecausethe pre-trenddoesnotpre-datetheriseofrobots.Yet,tocontrolforlonger-term industrytrends,weprovidetwoadditionalestimationapproaches: ex-plicitlyaccountingforpre-trendsbyincludingthechangeintheroutine employmentsharebetween2000and2005asalaggeddependent vari-ableandincludingindustryfixedeffects.
Columns(2)and(3)ofTable6addpre-trendstotheregressionson changesintheroutineemploymentshareandtheroutinemanual em-ploymentshare,respectively(cf.column(2)ofTable3andcolumn(1) ofTable4).Weobserveapositiveautocorrelationinemployment dy-namics.Yet,robotadoptionaddsadditionalinformationbeyondthose pre-trendsasthecoefficientremainsstatisticallysignificant. The
esti-35Notethatthepre-trendsinemploymentsharechangescovera5year pe-riod.Estimatedcoefficientsandstandarderrorsthushavetobeapproximately multipliedbyafactor2tomakethemcomparablewithourmainresultsforthe 10yearperiodfrom2005to2015.Whenthepre-trendsareincludedaslagged dependentvariables(columns2and3ofTable6),theyaccordinglyhavetobe dividedby2.
Table6
Accountingforlong-termindustrytrends.
Panel A: OLS (1) (2) (3) (4) (5) ΔRoutineemployment share2000-2005 ΔRoutineemployment share ΔRoutinemanual employment share ΔRoutineemployment share ΔRoutinemanual employment share Percentile of changes in robot adoption -0.020 ∗∗ -0.044 ∗∗∗ -0.046 ∗∗∗ -0.016 ∗∗∗ -0.026 ∗∗∗ (0.01) (0.01) (0.01) (0.00) (0.01) Change in dependent variable, 2000-2005 0.174 ∗ 0.147 ∗ (0.10) (0.08)
Industry Fixed Effects No No No Yes Yes
R2 0.014 0.035 0.030 0.007 0.007 Observations 700 700 700 700 700 Panel B: 2SLS (IV: Replaceable tasks) Percentile of changes in robot adoption -0.053 ∗∗ -0.113 ∗∗ -0.114 ∗∗ (0.02) (0.05) (0.05) Change in dependent variable, 2000-2005 0.133 0.109 (0.09) (0.08)
Industry Fixed Effects No No No
R2 -0.018 -0.012 -0.010
Observations 700 700 700
Notes:Robuststandarderrorsinparentheses.Multi-wayclusteringbycountryandindustry.Thedependentvariableisthechangeintherespectiveemployment shareovertherespectiveperiod.Theshareofreplaceabletasksinanindustryisusedasaninstrumentforrobotadoption.Regressionsincludethechangeinthe investmenttovalueaddedratioandthechangein(thelogof)valueaddedbetween2005and2015ascontrolvariables.Countryfixedeffectsareincludedinall regressionsandpartialledoutinthereportedR2.
∗∗∗p<0.01 ∗∗p<0.05 ∗p<0.1.
matedcoefficientiscomparabletothebaselineresults.Perhapsthemost convincingevidencethatthenegativerelationshipbetweenroutine em-ploymentsharesandrobotadoptionisnotexclusivelydrivenby spuri-ousindustrydynamicscanbefoundincolumns(4)and(5)ofTable6, whereweaddindustryfixedeffectstoourOLSregressions.36Thisis arestrictivemodelthat assumesindustry-specific timetrendsin lev-elsandthusnotonlyaccountsforheterogeneousindustryemployment trendsbutalsoremovesaconsiderabledegreeofvariationinthedata thatmayberelevantforidentification.Yet,thenegativeassociation be-tweenrobotizationandroutineemploymenttrendsisstillobservedand statisticallysignificant.
5.2.4. Globaldevelopmentsinrobotadoption
AsdiscussedinSection2,advancesinthetechnicalabilityofrobots mightrelatetothe“reshoring” ofjobstoadvancedcountries.For ex-ample,Faber(2018)observesadecreaseinlabourdemandinMexico associatedwithrobotadoptionintheUnitedStates.Weexplorethis re-lationinacross-countrycontextusingtwomeasuresofrobotadoption thatvaryacrossindustriesbutnotacrosscountries.First,wetakeglobal averages,definedasthecross-countrymeanofthepercentilechangein robotadoptionbyindustry.Thisreflectstheideathatinan intercon-nectedworldthoseindustrieswithhigherrobotadoptionwillseefaster declinesinroutineemploymentsharesregardlessofthelocationof pro-duction.Second,weuserobotadoptionofU.S.industriestorepresent globalindustrytrends.
ResultsarereportedinTable7.Incolumns(1)and(2)theglobal av-eragesofindustry-specificrobotadoptionisused.Theregressions sug-gestastatisticallysignificantandnegativerelationbetweenchangesin theroutineemploymentshareandglobaltrendsinrobotadoption.37 In-terestingly,thepositiveinteractionbetweenrobotadoptionandEMTEs
36 Wecannotestimatethemodelwithindustryfixedeffectsusing2SLSbecause theinstrumentonlyvariesacrossindustries.
37 Usingmeasuresofrobotadoptionthatvaryacrossindustriesbutnotacross countries,wealsodonotfindastatisticalsignificantassociationbetweenrobot
shownincolumn(2)nolongermakesupforthenegativeoverallrobot adoptionparameter: thehypothesis thatthesumofbothparameters addsupto0canberejectedatthe5%levelofstatisticalsignificance. Thissuggeststhatglobaldevelopmentsinrobotadoptionimpactlabour marketsinEMTEs.Note,however,thisisnotobservedifweuserobot adoptionin U.S.industriestocharacterizeglobaltrends(seecolumn (4)).38 Nevertheless,these exploratoryregressionsprovide suggestive evidenceforthepotentialrelevanceofglobalproductionnetworksand associatedjobreshoringpatternsduetoautomation,whichremainsan interestingareaforfurtherresearch.
6. Concludingremarks
Westudytherelationbetweenindustrialrobotsandoccupational shiftsbytaskcontent.Usingapanelof19industriesin37high-income andEMTEsfrom2005-2015,wefindthatincreaseduseofrobotsis as-sociatedwithpositivechangesintheemploymentshareofnon-routine analyticjobsandnegativechangesintheshareofroutinemanualjobs. Thepatternsthatwedocumentarerobusttoinstrumentalvariable es-timationandtheinclusionofvariouscontrolvariables,buttheydiffer acrosslevelsofeconomicdevelopment:weobserveasignificant rela-tionforhigh-incomecountries,butnotinEMTEs.Finally,wedonot findasignificantrelationbetweenindustrialrobotadoptionand aggre-gateemploymentgrowth.Thissuggeststhatindustrialrobotsdidnot replacejobs,buttheydidimpacttaskdemandandthushaddisruptive effectsonemployment.
Our analysis covered industrial robots, but much of the recent roboticdevelopmentshavebeentakingplaceinservices,suchasthe emergenceofmedicalrobots,logisticshandlingrobots,anddeliveryby
adoptionandtheaverageannualpercentagegrowthinemploymentin specifi-cationswithandwithouttheinteractionwithadummyforEMTEs.
38ItisalsonotobservedifweuserobotadoptioninGermanindustriesto characterizeglobaltrends.