Employment protection and firm-provided training in dual labour markets

ContentslistsavailableatScienceDirect

Labour

Economics

journalhomepage:www.elsevier.com/locate/labeco

Employment

protection

and

firm-provided

training

in

dual

labour

markets

Massimiliano

Bratti

_,

_Maurizio

_Conti

_,

_Giovanni

_Sulis

a Department of Economics Management and Quantitative Methods (DEMM), Università degli Studi di Milano, via Conservatorio 7, 20122 Milan, Italy b Department of Economics and Business, Università degli Studi di Genova, via Vivaldi 5, 16126, Genoa, Italy

c Department of Economics and Business, Università degli Studi di Cagliari, viale S. Ignazio da Laconi 17, 09123 Cagliari, Italy d Centro Studi Luca D’Agliano, (LDA), Italy

e Institute of Labor Economics (IZA), Germany f Global Labor Organization (GLO)

g Centro Ricerche Economiche Nord Sud (CRENoS), Italy

a

r

t

i

c

l

e

i

n

f

o

Dual labour markets Permanent contracts

a

b

s

t

r

a

c

t

In this paper we leverage a labour market reform (Fornero Law) which reduced firing restrictions for open-ended contracts in the case of firms with more than 15 employees in Italy. The results from a Difference in Regression Discontinuities design demonstrate that after the reform, the number of trained workers increased in firms just above the threshold by approximately 1.5 additional workers. We show that this effect can be explained by the reduction in worker turnover and a higher use of permanent contracts. Our study highlights the potentially adverse effects of employment protection legislation on training in dual labour markets.

1. Introduction

Employmentprotectionlegislation(EPL) hasbeen attheheartof thepoliticalandpolicydebateinmanycountriesforalongtime.By imposingconstraintsonfirms’abilitytoadjusttheirworkforcein re-actionto theshocks associatedwith demand,costs andtechnology, EPLhasoftenbeenblamedfornegativeeffectsatboththemicroand macrolevel,e.g.,lowerproductivityandpossiblyloweraggregate em-ployment(HopenhaynandRogerson,1993).Therefore,overthepast 20years,variouscountrieshavetriedtoreduceEPLacrosstheboard. However,incountriesthathaveintroducedlabourmarketreforms,the EPLrulesforopen-endedcontractshavebeenbarelychanged,atleast untilveryrecently(BoeriandGaribaldi,2019;Dolado,2016),partly owingtopoliticaleconomyconsiderations(Saint-Paul,1997).A pro-cessof‘labourmarketflexibilizationatthemargin’(i.e.,fortemporary oratypicalworkers)hasbeenfrequentlyimplemented,whichhasledto duallabourmarkets.1

AbetterunderstandingoftheeffectofEPLinduallabourmarkets isimportantbecauseitcanprovideinsightsonsomeunsettleddebates

∗_{Corresponding author.}

E-mailaddresses:massimiliano.bratti@unimi.it (M. Bratti), mconti@economia.unige.it(M. Conti), gsulis@unica.it(G. Sulis).

1 _{Dolado(2016)provides an insightful discussion about the emergence of dual labour markets in some European countries. See also Bentolilaetal.(2019).} 2 _{Consistent with the negative effect on productivity, LeonardiandPica(2013)report negative effects on individual wages, which is mainly limited to job switchers}

and low earners, young white-collar workers and workers in low-employment regions.

intheempiricalliteratureontheimpactofEPLonfirmperformance. Indeed,severalscholarshavereportedrobustevidenceofthenegative impacts ofEPL onlabour andtotal factorproductivity(Autoretal., 2007; Bassaniniet al., 2009; Bentolila etal., 2019; Bjuggren,2018; Cinganoetal.,2010;2016;Hijzenetal.,2017).2_{However,other} schol-arshavereportedthatEPLmayincreaselabourproductivity,suchas throughapositiveimpactoninnovation(Acharyaetal.,2014;Griffith andMacartney,2014;Koeniger,2005)orinvestmentsinfirm-specific trainingbyemployees(Belotetal.,2007).Therefore,asforcefully ar-guedbyBoerietal.(2015),moreworkisneededbecausemostofthe empiricalliteraturetendstolookattheoverallneteffectsofEPLon productivityratherthanattheexactmechanisms.

Inthispaper,weprovidenewinsightsonthemechanismslinking labourproductivity,EPLandtemporarycontractusebyinvestigating theeffectofEPLonfirm-providedtraininginduallabourmarkets.To thisend,weprovidecleanevidenceonthecausaleffectsofEPLon firm-providedtrainingusingalabourmarketreform,theForneroLaw,that wasintroducedinItalyin2012.Usingtwowavesofarepresentative surveyofItalianfirms,weleveragequasi-experimentalvariationinEPL

https://doi.org/10.1016/j.labeco.2021.101972

Received 5 June 2020; Received in revised form 20 January 2021; Accepted 7 February 2021 Available online 11 February 2021

0927-5371/© 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/)

usingaDifferenceinRegressionDiscontinuities(diff-in-disc)design(see, forinstance,Cinganoetal.,2016;Grembietal.,2016).This identifica-tionstrategyallowsustoovercomelimitationsofpreviousstudieson theeffectsofEPLusingRegressionDiscontinuityDesign(RDD) strate-gies(forexample,BolliandKemper,2017),i.e.,theroleofconfounding policiesoperatingatthesamecut-off asEPLandpossibledynamic sort-ing.

Beforethereform,theItalianlegislationincludedsize-contingent fir-ingrestrictionsforopen-endedcontracts,andbasedonthese restric-tions,thefiringcostsincreasedsharplyabovethe15-employee thresh-old.3_{Theserestrictionsweregreatlyreducedin2012,butonlyforfirms} abovethe15-employeecut-off,byalabourmarketreformknownasthe ForneroLaw.CombiningthedifferentlevelsofEPLbelowandabove the15-employeecut-off withtheEPLchangesintroducedin2012gives usauniqueopportunitytoobtainclearcausalevidenceontheeffect ofEPLontrainingusingadiff-in-discdesign.Providingnewevidenceis importantandtimelygiventhepaucityofstudiesthathaveempirically investigatedtheinterplaybetweenEPLandfirm-providedtraining(see Section2)andinlightof severalreformsthathavereduced employ-mentprotection,especiallyatthemargin,i.e.,fortemporaryworkers, inmanycountries.4

Themainresultsofourpapercanbesummarisedasfollows.Our pre-ferredestimatessuggestthatbyreducingEPLforlargefirms(i.e.,firms abovethe15-employeecut-off),theForneroreformincreasedthe aver-agenumberoftrainedworkersbyapproximately1.5individuals,which correspondstoanapproximately50percentincreaseinthenumberof trainedworkersatthecut-off firmsize(3.1workersweretrainedon av-eragepriortothereform).Theresultsarenotsensitivetoanextensive setofrobustnesschecks,includingdonut-holeregressionstoaccount forpotentialmanipulationoffirmsizearoundthecut-off,changesin thebandwidth,changesintheorderofthepolynomialinfirmsize,data heapingonfirmsize,placeboregressions,possiblemisclassification as-sociatedtomeasurementerrorintheforcingvariableandrestrictingthe analysistothepanelcomponentofthedata(andincludingfirmfixed effects),amongothers.

Wefurtherexplorethemainmechanismsthatmightbebehindthe estimatedeffectofrelaxingEPLontraining.Thefirstisthesubstitution betweentemporaryandpermanentworkers(Cahucetal.,2016). Sev-eralscholarshavefoundempiricalevidenceconsistentwiththis predic-tion(Cahucetal.,2019;CentenoandNovo,2012;Hijzenetal.,2017; SchivardiandTorrini,2008).Thus,owingtothetraininggapexisting betweenpermanentandtemporaryworkers(Albertetal.,2005; Arulam-palamandBooth,1998;Arulampalametal.,2004;Boothetal.,2002; Cabralesetal.,2017;Ferreiraetal.,2018),thegreateruseofpermanent contractsshouldalsoentailhighertraining.Thesecondmechanismis proposedbyDoladoetal.(2016).Theauthorsarguethat,under plausi-bleconditions,firms’temporary-to-permanentconversionratesgodown whentheEPLgapbetweentemporaryandpermanentworkersincreases. Temporaryworkersrespondinturntolowerconversionratesby exert-inglesseffort,whilefirmsreactbyprovidinglesspaid-fortrainingto them.Thus,theForneroreformbyreducingtheEPLgapmighthave con-tributedtoanincreaseoffirms’trainingprovisiontotemporaryworkers. Weprovideevidenceconsistentwiththefirstexplanation.Indeed, after the Fornero Law, firms for which EPL was reduced increased the numberof permanentworkers, which enjoy higher training, by about1.7unitsatthethreshold. Undertheassumption thattraining isprovidedtonewpermanenthires,forinstance,temporary-permanent workersubstitutioninducedbythereformatthecut-off wouldbeable toexplainalonethewholeeffectwefindonfirm-providedtraining.

3 _{In Section3, we discuss why these policy-induced differences in employment}

protection substantially differentiate firing costs according to firm size.

4 _{On labour market liberalisation reforms at the margin, see, for example,}

BoeriandGaribaldi(2007) and BertonandGaribaldi(2012) and the survey paper by Dolado(2016).

Wemaketwomaincontributionstotheexistingliterature,whichis discussedinmoredetailinSection2.First,weproviderobustevidence ontheeffectsofEPLontraininginItalyusingadiff-in-discdesignin aquasi-experimentalsetting.Ourapproachrepresentsanimprovement overtheexistingliteraturesinceusingthediff-in-discdesignintheItalian contextallowsustoaddresssomeoftheweaknessesofRDDs,namely, theexistenceofotherlabourmarketinstitutionsalsochanging discon-tinuouslyatthesamemarginoffirmsizeastheEPL(i.e.confounding policies).Thesearetherighttocreateworkscouncilsinthefirmsandthe existenceofshort-workprogramsforemployeesinfirmsundersevere economicdifficulties,whichmightaffectafirm’sprovisionoftraining. Accountingfortheconfoundingpoliciesthrougha diff-in-discdesign, ouranalysisleadstodifferentconclusionscomparedtoBolliand Kem-per(2017).5_{Second,weinvestigatethemainmechanismsoftheeffect} ofEPLontraining,andprovideevidenceofsubstitutionoftemporary withpermanentworkersaroundthe15-employeecut-off.

All in all, our analysis shows that in countries characterised by verystringentEPLforpermanentworkersandpersistentdualisminthe labourmarket,theexcessiveuseoftemporarycontractsandtheshort durationofemploymentspellsmaybeonekeydeterminantofthe in-centivesforfirmsto(not)providetraining.

Theremainder ofthepaperis organisedasfollows. InSection2, weprovidethetheoreticalgroundsforourempiricalanalysisand re-viewtheempiricalliterature.InSection3,weintroducetheinstitutional frameworkandpresentouridentificationstrategy.Afterdiscussingthe datainSection4,wepresentourmainresultsinSection5.Inthe on-lineAppendixB,wediscusspotentialthreatstoidentificationandtest theassumptionsofthediff-in-discdesign.Somerobustnesschecksare conductedinSection6.Section7exploresthemechanismsbehindthe estimatedeffect.Finally,Section8summarisesthemainfindingsand drawsconclusions.

2. Theoreticalunderpinningsandempiricalevidence:EPL, temporarycontractsandtraining

TounderstandthepotentialmechanismsthatcouldlinkEPL, tempo-rarycontractsandfirms’trainingprovision,itisusefultobrieflyreview somerelatedliterature.

Usingthestandardhumancapitalmodel(Becker,1964),afirm’s in-centivestoinvestinitsworkforcedependonthetimetheemployer ex-pectstoreapthebenefitsofamoretrainedworkforce.Thus,astronger EPL,e.g.anincreaseinfiringcosts,shouldalsoleadtoreducedworker turnoverandhigherworkertenure(BoeriandJimeno,2005;Kanand Lin,2011) therebycreatinghigherincentivestotraintheemployees. Bythesametoken,workersintemporarycontractsareexpectedto re-ceivemuchlessfirm-providedtrainingthanworkershiredwith perma-nentcontracts.Anothermechanismthatcangenerateapositive associ-ationbetweenEPLandtrainingishighlightedbyAcemoglu(1997)and AcemogluandPischke(1999),whoshowthatwhenlabourmarket in-stitutions,suchasEPL,generatewagecompression,firmsmayhavea greaterincentivetopayfortrainingbecauselabourmarket imperfec-tions,suchassearchfrictions,informationasymmetriesandlabour mar-ketinstitutions,determinethegapbetweenaworker’smarginalproduct andherwage,thusgeneratingrentstobesharedbetweenworkersand firms.Anecessaryconditionforfirmstosponsor(general)trainingis thattheserentsareincreasingintraining(AcemogluandPischke,1999). Theabovepredictionschangewhenpermanentandtemporary posi-tionsinthelabourmarketcoexist.Severalpapershavealreadytestedthe relationbetweentemporarycontractsandtrainingandfoundatraining gapinfavourofpermanentworkers(Albertetal.,2005;Arulampalam andBooth,1998;Arulampalametal.,2004;Boothetal.,2002;Cabrales etal.,2017;Ferreiraetal.,2018).Inturn,muchlessempiricalevidence existsontherelationbetweenEPLandfirms’traininginvestments.

TheinterplaybetweenEPLandtemporarycontractsinduallabour markets,inwhichtemporaryandpermanentcontractsenjoydifferent levelsof EPL,explains whythe expectation of a positive association betweenstrongerEPLandfirm-providedtrainingmayactuallybreak down.Thetheoreticalmodel inCahuc etal.(2016) predictsthat in labourmarkets withsignificantasymmetry in thedegreeof employ-ment protection enjoyedby permanent and temporary workers,the firmshaveanincentive tosubstitutetemporaryforpermanent work-ersbyusing asequenceoftemporarycontracts, therebycreating ex-cessworkerturnover.Thispredictionisindeedsupportedbythedata. Hijzen etal. (2017) show that in thecase of Italy, the stricter EPL abovethe15-employeethresholdisassociatedwithhigherratesof ex-cessworkerturnover,whichisdefinedasworkerturnoveroverthe ab-solutevalueofnetemploymentchange,withthis valuemeasuredas thedifferencebetweenhiringandseparationrates.Interestingly,the authorsalsofindthatthiseffectisentirelyexplainedbythegreateruse oftemporaryworkersabovethethreshold.Similarevidenceisfoundby CentenoandNovo(2012),whoreportanincreaseintheproportionof fixed-termcontractsfollowingaPortuguesereformthattightenedEPL forregularworkersinthecaseoffirmswith11to20workers.Thus, giventhattemporaryworkersreceivelesstraining,induallabour mar-kets,stricterEPLmayinduceawidespreaduseofcontractsassociated withlesstraininginvestments,thusgeneratinganegativeassociation be-tweenEPLandtraining.

However, a change in the level of EPL may affect firms’ train-ing decisions independent of an increase in the use of temporary contracts, which is nicely illustrated in the theoretical model in Doladoetal.(2016).Intheirmodel,firmsdonotusetemporary con-tractsasapurescreeningdevicebecausesuchcontractsensurean ini-tiallyhighersurplusthanpermanentcontracts,duetotheirlowerfiring costs.6_{Temporaryworkerscannotberenewed,andattheircontract’s} termination,firmshavetodecidewhethertohirethempermanentlyor not.Workers’productivitydependsontheireffort,althoughinaddition towages,firmshaveanotherinstrumentavailabletoimprovetemporary workers’performancewhentheirincentivesarenotaligned: temporary-to-permanentconversionrates.Moreover,firmsprovidecostlyspecific humancapitaltotemporaryworkersthroughtraining,whichonly in-creasesworkers’productivitywhentheybecomepermanentworkers. Themodel’sbasicinsightisthatwhenEPLbecomesstricter,i.e.,the gapinfiringcostsbetweentemporaryandpermanentcontractswidens, firms reduce therate of temporary-to-permanent conversion and

re-ducetraininginvestmentsintemporaryworkers.Asaconsequence,

tem-poraryworkersreducetheireffortlevelbasedonthelikelyconditionin whichpermanentworkersreacttothechangeinEPLbyexertinglessor equaleffort(whichisconsistent,forinstance,withthepositiveeffectof stricterEPLonabsenteeismfoundbyIchinoandRiphahn,2005).

In brief, although simple economic reasoning indicates that we shouldexpectapositiveassociationbetweenstricterEPLandtraining, thepresenceofduallabourmarketscomplicatestheoverallpicture.In particular,stricterEPLmayinducealargeruseoftemporarycontracts orreducethetemporary-to-permanentconversionrates,thusreducing traininginvestmentsinbothcases.Inthiscontext,Choi(2019)proposes asearchandmatchingmodelinwhichtheexcessiveuseandpersistence oftemporarycontractsariseendogenouslyduetoselectioneffectsand optimaltrainingdecisions.Inhisframework,lessskilledworkers self-selectintotemporaryjobswhilelowlevelsoftrainingemergeasthe resultoftheshortdurationoftemporaryemploymentrelations. Con-sequently,hisframeworkpredictsthat,inresponsetoariseinfiring

6 _{Daruichetal.(2017)exploit an Italian reform that lifted constraints on the}

employment of temporary contracts while maintaining the level of EPL in per- manent contracts unaltered and demonstrate that firms increased the use of temporary contracts and experienced lower labour costs and higher profitabil- ity. The authors also report that workers on a temporary contract receive only 66% of the rents shared by firms with workers hired under a permanent contract.

costs,open-ended(permanent)contractsaresubstitutedbytemporary ones,whichresultsinalowincidenceoftraining.

ScholarshaveextensivelyinvestigatedtheeffectofEPLona num-ber of outcomes for both workers and firms; however, few papers have focusedon theeffect of EPLon firm-provided training.7 _Using a large firm-level datasetacross developing countries, Almeida and Aterido(2011)showthatstricterenforcementoflabourregulationsis significantlyassociatedwithhigherinvestmentsbyfirmsintheir em-ployees’ humancapitalbutthat themagnitude of theassociation is verysmall.Similarly,PierreandScarpetta(2013)usecross-country har-monisedsurveydataandfindthathigherEPLisassociatedwithhigher investmentintrainingandgreateruseoftemporarycontracts.Theyalso findthatEPLhaslargereffectsonsmallfirmsandin sectors charac-terisedbygreaterjobreallocation.

Furthermore,studiesexploitingwithin-countryvariationinlevelsof EPLdonotfindstrongpositiveeffectsofEPLontraining.Forinstance, PicchioandvanOurs(2011)useDutchdataformanufacturingfirmsand findthathigherlabourmarketflexibility(i.e.,lowerEPL)marginally re-ducesfirms’investmentintraining;however,thiseffectisrathersmall. AstudybyMesseandRouland(2014)exploitsareformofEPLinFrance toidentify,usingaDifferenceinDifferences(DID)approachcombined withpropensityscoremethods,theeffectofEPLontheincentivefor firmstopayfortraining.TheyfindthathigherEPL(intheformofa taxonfirings)hadnoeffectonthetrainingofoldereligibleworkers, whileithadapositiveeffectontrainingforworkersjustbelowthe el-igibility threshold.Theauthorsinterpret thisfindingasstressingthe complementaritybetweentrainingandfiringdecisions.

ThepapermostcloselyrelatedtooursisBolliandKemper(2017),in whichtheauthorsuseanRDDframeworkexploitingvariationinfiring regulationsacrosssizethresholdsinItalyandFinlandusingadifferent sourceofdata(from2005and2010)tostudytherelationshipbetween EPLandtrainingprovision.TheirRDDresultsdonotshowany statisti-callysignificanteffectofEPLonfirm-providedtraining(measuredasa dichotomousindicatorofafirm’strainingprovision,traininghoursand numberoftrainedemployees).Weaddtotheiranalysisbyleveraging quasi-experimentalvariationprovidedbytheForneroreformina

diff-in-discframework,whichenablesustocontrolforotherlabourmarket

institutionsthatinItalychangediscontinuouslyatthethreshold,such

astheCassaIntegrazioneGuadagniStraordinariascheme, ashort-term

workprogrammefeaturingaredundancyfundsystem,orthepresence ofworkscouncilsinthefirm.Weshowthataccountingforthese con-foundingpoliciesleadstoverydifferentconclusionsontheeffectofEPL.

3. Institutionalframeworkandidentification

3.1. Institutionalframework

Sincethe1960s,theregulationofunfairdismissalshaschanged sev-eraltimesinItaly.Themostsignificantreformoccurredin1970with Lawn. 300/70,alsoknownasthe‘StatutodeiLavoratori’(Workers’ Statute)and,in 1990, withLawn.108/90,which strengthened em-ployeeprotectionfromunfairdismissalonlyinthecaseofsmallfirms.8 Beforethelegislativechangesthatoccurredin2012(ForneroLaw), thedegreeofprotectionenjoyedbyunfairlydismissedworkerswas con-siderablygreaterinthecaseofemployeesworkinginfirmswithmore than15employees(Article18oftheWorkers’Statute).Indeed,ifa

dis-7_{The literature on the various effects of EPL is vast and cannot be compre-}

hensively reviewed here. See MessinaandVallanti(2007)for the effects of EPL on job flows, Bassaninietal.(2009) and Bjuggren(2018) for EPL and produc- tivity, Cinganoetal.(2010)and Cinganoetal.(2016)for EPL and investment in physical capital, SchivardiandTorrini(2008)for EPL and a firm’s’ propensity to grow, LeonardiandPica(2013)for EPL and wages, KuglerandPica(2008)for EPL and worker flows, and Bottassoetal.(2017)for EPL and firm dynamics.

8_{See Cinganoetal.(2016)and Hijzenetal.(2017)for a brief overview of the}

missalwasdeclaredunfairbyajudge,anemployeeunfairlydismissed fromafirmwithmorethan15employeescouldasktobereinstated andreceiveforgonewagesandthehealthandsocialsecurity contribu-tions(foraminimumof5months)relatedtotheperiodbetweenthe dismissalandthesentence.Althoughreinstatementwasthemostlikely occurrencein practice,the unfairlydismissedemployeeretainedthe righttoinsteadreceiveaseverancepaymentamountingto15months’ salary.Incontrast,inthecaseoffirmswithfewerthan15employees, itwasuptotheemployertochoosewhethertoreinstatetheunfairly dismissedworker(withoutpayinganyforgonewages)ormakea sever-ancepayment,whichrangedfrom2.5to14monthsinthecaseofvery seniorworkers(Hijzenetal.,2017).9

Thehigherdejurecostsforemployersinthecaseoffirmswithmore than15employeeswerefurtherincreasedwhenconsideringthedefacto costsassociatedwiththeverylongaveragedurationoflabourtrialsin Italy:GianfredaandVallanti(2017)report averagetrialdecisionsof approximately850daysovertheperiod2007–2010,withlarge vari-ationacross regions.10 _{Such adifference inthe lengthof labour} tri-alsescalatesthefiringcostsabovethethreshold.Indeed,usinga for-mulaproposedbyGaribaldi andViolante(2005) tocomputeex post firingcosts,GianfredaandVallanti(2017)reportfiringcostsequivalent toapproximately36monthsofwagesinTrentoversus160monthsin Salernoforablue-collarworkerwith8yearsoftenureinafirmabove the15-employeethreshold.11_{Becausenoforgonewageswereduefor} firmsbelowthethreshold,thelengthoflabourtrialsonlymattersfor firmsabovethethreshold,withfiringcostsrapidlyincreasingabovethe 15-employeethresholdifthelabourtriallastslongerthan5months.12 Moreover,thelackofacleardefinitionofunfairdismissalinItalian leg-islation(Hijzenetal.,2017)ledtosomeinconsistenciesinits implemen-tation,asnotedbyIchinoetal.(2003),whoshowthatinregionswith highunemploymentrates,judgestendedtoruleinfavourofemployees. Thevariabilityindecisionsthereforeledtouncertainty,whichfurther increasedthecostsassociatedwiththestricteremploymentprotection forfirmsabovethethreshold.

Thusfar,wehavediscussedonlyemploymentprotectionfor open-endedcontracts.However,asinothercountries,suchasSpainorFrance, theItalianlabourmarkethasinthepast15yearsbeencharacterisedby anotableincreaseintheuseoftemporaryandatypicallabourcontracts followingtheliberalisationthatstartedattheendofthe1980s(inthe caseoftemporarycontracts)andattheend ofthe1990sinthecase ofsemi-autonomousatypicalworkers.Itis,however,importanttonote thatthedegree of EPLfortemporary andatypicalworkersdoes not changediscontinuouslyatthe15-employeethreshold;indeed,itdoes notdependatallonfirmsize.

InJuly2012,areformknownastheForneroLawsignificantly re-ducedfiringcostsforpermanentworkersinthecaseoffirmswithmore than15employees.WerefertoBertonetal.(2017)foradetailed anal-ysisofthenoveltiesintroducedbythe2012reform,buthere,wenote thattheForneroLawlimitedthepossibilityforpermanentworkersin firmswithmorethan15employeestochoosebetweenreinstatement andamonetary compensationin caseof unfairdismissal toa setof

9 _{Above the 15-employee threshold, employment protection is also greater in}

the case of collective dismissals.

10 _{For instance, GianfredaandVallanti(2017) report an average length of}

labour trials of 313 days in Trento, in the north of Italy, versus 1397 days in Salerno, in the south of the country.

11 _{If one takes into account the expected probability of a settlement between}

the parties and the fact that some rulings are decided in favour of the firm, the exantefiring costs fall to approximately 15 months of wages in Trento (north) compared with 65 months in Salerno (south). The formula is based on the time it takes to reach a sentence, the forgone wage, the health and social security contributions, the penalty rate on forgone contributions, the legal fees and the severance payments (see GaribaldiandViolante,2005).

12 _{Indeed, 5 months is the minimum amount of forgone wages and contribu-}

tions that the unfairly dismissed worker has the right to receive for firms above the threshold.

well-definedcases.13_{Moreover,itsubstantiallyreducedtheamountof} monetarycompensationandeasedtheuncertaintysurroundingthe du-rationandcostsoflitigation,which,ashighlightedabove,wasfairly high,especiallyinsomeareasofthecountry.14

As explainedin following section,weuse the reductioninfiring costsbroughtaboutbytheForneroLawinfirmsabovethe15-employee thresholdtoidentifytheeffectofEPLonfirms’propensitytotrain

work-ersinadiff-in-discdesign.

3.2. Identificationstrategy

GiventhesharpdiscontinuouschangeinthelevelofEPLatthe 15-employeethreshold,awaytoestimatetheeffectofEPLontrainingis usingaRDDsuchas

𝑦𝑖=𝛽0+𝛽1𝑎𝑏𝑜𝑣𝑒𝑖+𝑓(𝑥𝑖−15)+𝑔(𝑥𝑖−15)×𝑎𝑏𝑜𝑣𝑒𝑖+𝜖𝑖, (1)

where𝑖isthefirmsubscriptand𝑦𝑖isthenumberoftrainedworkers;

𝑎𝑏𝑜𝑣𝑒𝑖isadichotomousindicatorthatequalsoneforthefirmsabove

the15-employeecut-off (i.e.thosesubjectedtostrongerEPLofArticle 18),andzerootherwise;𝑓(𝑥𝑖−15)and𝑔(𝑥𝑖−15)aretwodifferent poly-nomialsinfirmsizenormalizedwithrespecttothecut-off.15_Finally𝜖

𝑖 isanidiosyncraticerrorterm.Undertheusualcontinuityassumptions (Hahnetal.,2001),theparameter 𝛽1 identifiestheeffectof EPLon

training.

AlthoughRDD hasbeenalready usedintheliterature toidentify theeffectofEPLontraining(BolliandKemper,2017),intheItalian contextapotentialthreattoidentificationcomesfromtheexistenceof confoundingfactorsthatchangediscontinuouslyatthesamecut-off as EPL.Amongthelatter,thetwomostnotableconfoundersaretheCassa

IntegrazioneGuadagniStraordinaria(CIGS)scheme, ashort-termwork

programmefeaturingaredundancyfundsystem,andtherightto con-stituteworkscouncilsinthefirm.CIGSaimstohelpfirmsthatareeither intheprocessofreorganizationandrestructuring,facingasevere eco-nomiccrisisorunderaninsolvencyprocedure.TheItalianlegislation fortheperiodrelatedtothisstudymandatedthatonlyfirmsabovethe 15-employeethresholdcoulduseCIGS.16_{Ingeneral,firmswithahigh} share of workersunderCIGS schemesarealsolikely toprovideless trainingsincetheir levelof activityisdecreasing.Conversely, works councilscouldpositivelyaffectafirm’strainingprovision(Dustmann andSchönberg,2009;Stegmaier,2012),andItalianlegislationgrants workerstherighttoformworkscouncilsinfirmswithmorethan15 employees(TitoloIIIofWorkers’Statute).17

Inthiscase,usinganRDD,itwouldnotbepossibletoseparately identifytheeffectofEPLfromtheeffectof theconfoundingpolicies.

13_{For instance, the judge was granted the ability to order a reinstatement only}

if she believed that the just cause of justified subjective reason invoked by the firm simply did not exist or the collective agreement applied by the firm foresaw a different punishment. Similarly, in the case of an economic lay-off, reinstate- ment was allowed only as long as no justified objective reasons actually existed.

14_{Another important labour market reform, the so called Italian Jobs Act, was}

introduced in 2015. The reform further lowered EPL by introducing a new open ended contract: “The new open ended contract introduced in March 2015 al- lowed for employment protection increasing with tenure, confining the possi- bility of reinstatement of workers to discriminatory dismissals, hence exclud- ing this possibility for dismissals due to economic reasons ( motivooggettivo). ” ( BoeriandGaribaldi,2019, p. 2). However, the last period considered in the main analysis precedes the Jobs Act.

15_{Thus the polynomial is allowed to be different on each side of the cut-off.}

At the cut-off𝑔(.) is left continuous with 𝑔(0)= 0 .

16_{In turn, all firms, independently of firm size, may have access to Cassa}

Inte-grazioneGuadagniOrdinaria(CIGO) scheme, aimed to help firms in a temporary financial crisis.

17_{While there are firms with fewer than 15 employees with a worker council,}

our data show a clear increase at the threshold in the likelihood that there is a worker council within a firm, even controlling for the number of employees. See Cardulloetal.(2020) for more details on this aspect of the institutional background.

However,Grembietal.(2016)demonstratethat,eveninthepresence ofconfoundingpolicies,ifthereisachangeofthepolicyofinterest(EPL inourcase)overtime,theeffectcanbeestimatedusingadiff-in-disc de-sign.18_{Usingaparametricspecificationofdiff-in-disc(cf.Cinganoetal.,} 2016),theestimatedequationreadsasfollows

𝑦𝑖𝑡=𝛼0+𝛼1(𝑎𝑏𝑜𝑣𝑒𝑖𝑡×𝑝𝑜𝑠𝑡𝑡)+𝛼2𝑝𝑜𝑠𝑡𝑡+𝛼3𝑎𝑏𝑜𝑣𝑒𝑖𝑡+𝑓(𝑥𝑖𝑡−15)

+𝑔(𝑥_𝑖𝑡−15)×𝑎𝑏𝑜𝑣𝑒_𝑖𝑡+𝜖_𝑖𝑡, (2)

wherewehaveaddedatime𝑡subscripttoallvariables,with𝑡=0for thepre-ForneroReform(wave2010ofthedatadescribedinthenext section)and𝑡=1forthepost-Forneroreform(wave2015).Thevariable 𝑝𝑜𝑠𝑡𝑡isadichotomousindicatorthatequalsoneintheperiodafterthe reformandzerootherwise.19_{Alltheothervariablesandpolynomials} aredefinedasinequation(1).

Undersomeidentifyingassumptions, thecoefficient 𝛼1 ofthe

in-teraction𝑎𝑏𝑜𝑣𝑒𝑖𝑡×𝑝𝑜𝑠𝑡𝑡representsthecausaleffectof relaxingEPLon firm-providedtraininginthecaseoffirmsjustabovethethreshold.In particular,Grembietal.(2016)showthatinordertoidentifytheeffect ofthepolicyofinterest,inadditiontotheusualcontinuityassumption (Assumption1,intheirpaper),twootherassumptionsareneeded.The assumptionthattheeffectoftheconfoundingpoliciesinthecaseofno treatment(i.e.intheabsenceoftheForneroLaw)isconstantovertime (Assumption2inGrembietal.,2016),allowsustointerpret𝛼1asthe

lo-caltreatmenteffect(i.e.atthecut-off)ofrelaxingEPLonfirmssubjected

totheconfoundingpolicies.Ifweintroducethefurtherassumptionthat

theeffectofEPLatthethresholddoesnotdependontheconfounding policies(Assumption3inGrembietal.,2016),then𝛼1measuresthe

causaleffectofrelaxingEPLinaneighborhoodofthecut-off,i.e.the usualRDDestimand.

Totestthevalidityofthediff-in-discdesign,inonlineAppendixBwe implementsome tests suggested by Grembi etal. (2016). We check whetheranymanipulationoftherunningvariablechanges(orarises) overtimebytestingforthecontinuityofthedifferenceinthedensities beforeandaftertheFornero Law(Assumption1).20 _{Assumption2is} directlytestedbyrunningaplaceboanalysisfortheperiodbeforethe Forneroreform.Undertheassumptionthattheeffectoftheconfounding policiesistimeinvariant,intheabsenceoftreatmentthediff-in-disc co-efficientshouldbezero.However,withadiff-in-discdesignanddataon theconfoundingpolicies(i.e.useofCIGSandpresenceofworkscouncils inthefirm),onemightdirectlytestforbothAssumption2and

Assump-tion3ofthediff-in-discdesign.Inparticular,Assumption2canbetested

byshowingthattheeffectoftheconfoundingpoliciesonfirm-provided

18 _{The combination of an RDD and a DID identification framework has also}

been referred to in the statistical literature as Comparative Regression Disconti- nuity Design (see, for instance Hendrenetal.,2019;Kisbu-Sakaryaetal.,2018). Unlike in Grembietal.(2016) — who exploit the DID component in order to isolate the effect of other policies that change discontinuously at the cut-off— Kisbu-Sakaryaetal.(2018)argue that the use of pre-treatment data may lead to increased statistical power compared to conventional RDD.

19 _{The binary variable 𝑝𝑜𝑠𝑡}

𝑡captures, among the other things, a common shock that might have hit firms post-reform. Moreover, in some specifications (see equation(3)) the effect of 𝑝𝑜𝑠𝑡𝑡is allowed to change with the firm size variable, so as to capture the effect on training of any temporal shock that may have had a differential effect on firms of different sizes, such as changes in the availability of bank credit.

20 _{As a possible additional check for manipulation, one could report balanc-}

ing tests of some firm characteristics around the cut-off before and after the Fornero reform. Unfortunately, many of these covariates are not predetermined but may instead act as mediating factors for the effect of EPL. Thus, checking for balancing will not help judge the validity of our diff-in-disc framework. To take a few examples, firm characteristics affected by EPL that also interact with worker training may include investments in physical capital ( Cinganoetal., 2010;2016), access to credit ( Cinganoetal.,2016), innovation performance ( Acharyaetal.,2014;GriffithandMacartney,2014;Koeniger,2005), use of temporary contracts ( Hijzenetal.,2017), wages ( LeonardiandPica,2013) and workers’ mismatch ( Bertonetal.,2017).

trainingwasnotchangingovertimeatthecut-off beforethe implemen-tationoftheForneroLaw(i.e.localparalleltrends).Thiscanbedone byaugmentingequation(2)withaninteractiontermbetweenthe con-foundingpoliciesandthe𝑝𝑜𝑠𝑡𝑡indicator.21Assumption3canbetested runningasimilarregressionbutincludingbothperiodsbeforeand af-tertheForneroLaw(i.e.the2010and2015waves)andaninteraction betweentheconfoundingpoliciesandthe𝑎𝑏𝑜𝑣𝑒𝑖𝑡×𝑝𝑜𝑠𝑡_𝑡indicator.

Equation(2)isestimatedwithlocallinearregressiontechniques,i.e., weconsideralinearpolynomialandquiteanarrowbandwidtharound thethreshold,namely,6–25employees.However,thebaseline specifi-cationisalsoestimatedwithdifferentbandwidths,namely,11–20,6–30 and6–50,withbothalinearandaquadraticpolynomialspecification, andexcludingvs. includingavectorofcontrolvariables,comprising sector-by-yearandregion-by-yearfixedeffects.22_Moreover,asa robust-nesscheck,wefollowGrembietal.(2016)andweallowforthe poly-nomialtodiffernotonlyaboveandbelowthethresholdbutalsobefore andafterthereformandwealsoincludeatripleinteraction,whichis clearlyamoregeneralandconsiderablymoredemandingspecification thanthatinequation(2):

𝑦𝑖𝑡=𝛾0+𝛾1(𝑎𝑏𝑜𝑣𝑒𝑖𝑡×𝑝𝑜𝑠𝑡𝑡)+𝛾2𝑝𝑜𝑠𝑡𝑡

+𝛾3𝑎𝑏𝑜𝑣𝑒𝑖𝑡+𝑓(𝑥𝑖𝑡−15)+𝑔(𝑥𝑖𝑡−15)×𝑎𝑏𝑜𝑣𝑒𝑖𝑡

+ℎ(𝑥_𝑖𝑡−15)×𝑝𝑜𝑠𝑡_𝑡+𝑗(𝑥_𝑖𝑡−15)× (𝑎𝑏𝑜𝑣𝑒_𝑖𝑡×𝑝𝑜𝑠𝑡_𝑡)+𝜖_𝑖𝑡. (3) whereℎ(𝑥_𝑖𝑡−15)and𝑗(𝑥𝑖𝑡−15)aretwoadditionalpolynomialsin nor-malizedfirmsize.23_{Sincethediff-in-discdesigninthepresenceof} het-erogeneouseffectsprovidesalocalaveragetreatmenteffect(LATE)at thecut-off (Hahnetal.,2001),wearenotabletoestimatetheeffectsof theForneroLawonfirm-providedtrainingthataremediatedbychanges (e.g.increases)infirmsize,i.e.ourestimatesapplytofirmsaroundthe 15-employeecut-off.Forthesamereason,theestimatedeffectsonthe numberoftrainedworkersalsocorrespondtotheeffectsontheshare oftrainedworkers.Yet,inSection6,wereportpanelfixedeffects esti-matesthataresimilartoaDIDdesign,andthatalsoapplytofirmsizes farfromthecut-off.

Alast commentisinorder.Inourbaselineestimatesweuse two cross-sectionsthatarerepresentativeofItalianfirmsin2010and2015 (i.e.theyearsof thesurvey, butemploymentandtrainingdataused intheempiricalanalysisrefertotheyearbefore),whicharedescribed inthenextsection:itisimportanttonotethatwepoolthetwo cross-sectionsandthat,therefore,thefirmsinthetwowavesaregenerally notthesame.Inparticular,theidentityofthefirmsaboveandbelow thecut-off,andmoreimportantlyforidentificationtheircharacteristics maychange.Thismayaffectthepopulationofcompliersonwhichthe

diff-in-discLATEeffectisestimated.Forthisreason,firmdynamicsare

investigatedintheonlineAppendixBwherewealsocheckforchanges inthecompositionofthetwocrosssectionsaroundthecut-off imple-mentingthetestproposedbyCarrelletal.(2018).

4. Data

Weusetwowaves(collectedin2010and2015)oftheRILSurvey dataset(RilevazioneLongitudinalesuImpreseeLavoro)providedby IN-APP(NationalInstitutefortheEvaluationofPublicPolicies).The IN-APPinstitutehasbeenrecentlycreated(replacingISFOL,Istitutoperlo

sviluppodellaformazioneprofessionaledeilavoratori),anditsmain

activ-itiesareorientedtowardsresearch,monitoringandpublicpolicy eval-uation.Itconstitutesabuildingblockinsupportingpolicymakingby theMinistryofLabourandSocialPolicies.Usingtheuniverseofactive

21_{Although in this case 𝑝𝑜𝑠𝑡}

𝑡represents a placebo post-reform period.

22_{Industry-by-year fixed effects are included to capture any time-varying in-}

dustry specific differences in training provision. Similarly, by including region- by-year fixed effects, we allow for time-varying regional differences in training provision.

ItalianfirmsprovidedbyISTAT(theItalianNationalStatistical Insti-tute),calledASIA(ArchivioStatisticoImpreseAttive,StatisticalArchiveof ActiveEnterprises),theRILsampleisrepresentativeofthepopulation ofboththelimitedliabilitycompaniesandpartnershipsintheprivate (non-agricultural)sectors.Apanelversionofthedatasetisavailablefor asmallernumberoffirms.

Thedatasetcontainsindicatorsoffirmsize,performance,training andadditionalvariablesrelatedtothesystemofindustrialrelations.An importantfeatureofthedataisthattheycontaindetailedretrospective informationontrainingactivities(in2009and2014),whichisusually unavailableinadministrativedataonfirmsorworkers,andonfirmsize (alongwiththenumberofhiringsandfirings).Furtherinformationis availableonthepresenceofworkscouncilsintheworkplaceandthe levelofbargainingandcontractuallabouragreements.Thesurveyalso containsinformationonthecompositionoftheworkforceintermsof skillsandtypesofcontractsforworkers(butnotretrospectively,onlyfor thesurveyyears2010and2015).Onthefirmside,althoughthedataset isquiterichintermsofvariablesrelatedtofirmactivities,suchastheir export,innovationoroffshoringactivities,onlylimitedinformationis availableonbalancesheetdata.24

TwofeaturesoftheRILsurveyareworthmentioning.First,firmsize isprovidedindiscreteunits,i.e.,headcount(excludingcollaborators). Unlikeemploymentsize,whichisavailablefor2009,2010,2014and 2015,thecompositionof employment,in termsof part-timeand full-timeworkersandtypeofcontractsrefertotheyearofthesurveyonly (2010and2015)whiletrainingdata,aswesaid,referstotheyearbefore (2009and2014).Forthisreason,wecannotbuildacontinuousmeasure ofemployment in2009and2014usingproxy measuresof thelegal definitionoffirmsize,i.e.,theonestrictlyrelevantfortheapplication ofArticle18,butweareforcedtouseemployeeheadcountforthesame years.25_{Togetasenseofthepotentialimpactofmeasurementerror,we} checkthesensitivityofourestimatestoexcludingfirmswith16,16–17 and16to18employeesinadonut-holetypeofregression(seeSection6) because,giventheincidenceofpart-timeandtemporaryemployment aroundthecut-off,theyarethemostlikelytobespuriouslyconsidered asabovethethresholdwhentheyareinfactbelowit(e.g.,iftheyhave atleasttwopart-timeemployees,whicharecountedasafractionofa full-timeemployee).

Aseconddatafeaturedeservingattentionatthisstageisthatthe forcingvariable,i.e.,self-reportedfirmsize,ischaracterisedbyahigher frequencyoffirmsizesthataremultipleoffive.Ontheonehand,a pos-siblereasonisheaping,i.e.roundingbytheindividualthatwas inter-viewedinthefirm.Ontheotherhand,multiplesoffivearealso rela-tivelymorefrequentinthefirmsizedistributionreportedinpapers us-ingadministrativedataonNationalSocialSecurityarchives,suchasin Cinganoetal.(2016).26_{Barrecaetal.(2016)presentanddiscuss} simula-tionevidencesuggestingthatneglectingnon-randomheapingcanlead tobiasesandthatomittingobservationsatdataheapsshouldleadto unbiasedestimatesofthetreatmenteffectsforthe‘non-heapedtypes’.

24 _{Devicientietal.(2018)use the RIL data as a primary source of information}

to study the relationship between unions and temporary contracts.

25 _{According to Italian legislation, part-time workers count as less than one}

full-time employee and proportionally to the number of hours worked, when defining the firm size which is relevant for the application of EPL. By way of ex- ample, a firm with 16 employees, three of which have a 50% part-time contract, would be equivalent to a firm with 14.5 full-time employees and is therefore de facto below the 15-employee threshold. Similarly, only temporary employees with at least a 9-month contract should be counted as far as the definition of the threshold is concerned, while apprentices should be excluded. Some mea- surement error is also present in papers using National Social Security archives such as LeonardiandPica(2013)or Hijzenetal.(2017), which do not have full information on individual working hours and make some simplifying assump- tions to compute “legal ” firm size. Hijzenetal.(2017), for instance, assume that part-time workers work half time (50%).

26 _{Unfortunately, administrative data on firm size is not available in our sam-}

ple.

Table1

Descriptive statistics.

Variable Mean Std. Dev.

nr. of employees wave 2010 (year 2009) 10.98 4.50 wave 2015 (year 2014) 10.80 4.59 nr. of trained workers wave 2010 (year 2009) 2.26 4.28 wave 2015 (year 2014) 3.54 5.06 nr. of temporary workers wave 2010 (year 2010) 1.23 2.13 wave 2015 (year 2015) 1.12 2.79 nr. of permanent workers wave 2010 (year 2010) 9.70 4.64 wave 2015 (year 2015) 9.54 4.95 excess worker turnover

wave 2010 (year 2009) 0.51 1.22 wave 2015 (year 2014) 0.44 0.78 Note. Descriptive statistics use sample weights and are calculated for the sample used in the re- gression reported in column (1) of Table2. See Section4for the timing of all variables. Employ- ees represents the total number of employees. Trained workers represents the number of work- ers trained. We imputed trained workers equal to employees when the number of trained was greater than the number of employees; and we imputed 0 when this information was missing. Excess worker turnover is calculated at the firm level following Hijzenetal.(2017) as follows: 𝐸𝑊𝑇 = 2 ⋅ min ( 𝐻,𝑆)∕𝐸,where 𝐻 and 𝑆 arethe number of hiring and separations, respectively, and 𝐸 isaverage firm employment.

Notbeingsureontherealamountofmeasurementerror,inour pre-ferredempiricalspecifications weusetheentireavailabledata,since

diff-in-discdesigns,asRDDs,aredata-intensive,butinsomesensitivity

checkswealsoshow thatthebaseline resultsarerobusttodropping observationswithmultiplesoffiveinemploymentsize.

Inwhat follows,we describeour sampleselectionprocedure.We beginwith24,459observationsforthe2010waveand30,091forthe 2015wave.Wedropfirmsthathaveimplausibleorinconsistent(i.e. negative,ofdifferentorderofmagnitudesacrosswavesormissing val-uescodedasstrings)numberofemployeesin2010(196observations) andin2015(83observations).Theaboveselectionsresultin24,263and 30,008observationsforthetwowaves,andthewholesampleis54,271 observations.For10,214firms,wehavetwoobservations(panel),while theremainingfirms(14,049and19,794forthe2010and2015waves, respectively)representarepeatedcrosssection.Intheeconometric anal-ysis,werestrictthesampletofirmssizedinthe6–25employeerange (althoughwe presentrobustnesschecksusing different bandwidths); moreover,wetrimthedatabydroppingfromtheanalysisfirmsthat ex-periencedayear-on-yeargrowthrateinthenumberofemployeeslarger (smaller)thanthe95th(5th)percentile.27_{Finally,werestrictthesample} tostill-activefirms,resultinginafinalsampleof16,532observations (5,794forthepanelcomponent).InTable1,wereportdescriptive statis-ticsforthesampleusedinthebaselineregressionsreportedinTable2.

5. Mainresults

Thissectionreportsourpooledcross-sectionbaselineestimatesof theeffectofEPLonfirm-providedtrainingusingthenumberoftrained workersastheoutcomevariable.InthefirstfourcolumnsofTable2, wereportestimateswithapolynomialinfirmsizethatis allowedto

27_{This can be computed because each wave reports employment data for the}

0 2 4 6 8 10 N. of trained workers -10 -5 0 5 10

N. of normalized employees (firm size)

Pre-Fornero reform

N. of normalized employees (firm size)

Post-Fornero reform

Fig.1. Firm size and observed training provision before and after the Fornero reform Note. The figure presents a scatter plot for the average number of employed workers by one employee-bins of firm size (computed using survey weights) before and after the Fornero reform as well as the fitted (solid) line of a regression of the number of trained workers on normalized employment (see column (1) of Table2) and the 95% confidence interval (dashed lines). Normalized employment is reported on the horizontal axis, with “0 ” corresponding to the cut-off (i.e., firm’s employment level equal to 15). The scatter plot is presented for the bandwidth 6–25 employees of firm size (i.e., normalized size between - 10 and 10).

differoneachsideofthecut-off butthatisinsteadassumedtotakeon thesamecoefficientbeforeandafterthereform.Wealsoinclude (ex-clude)sectorandregionfixedeffects(whichsometimeswewillrefer toas‘firmcontrols’),whoseeffectisallowedtovarybeforeandafter theForneroreform.Theestimatesincolumn(1)showthatatthe 15-employeethresholdandfollowingtheForneroreform,therehasbeen anaverageincreaseof1.72trainedworkers,whichissignificantatthe 1percentlevel.28_{Themagnitudeofthediscontinuitycanalsobe} ap-preciatedfromFig.1,whichshowsnosignificantjumpinthenumber oftrainedworkersbeforethereform,althoughsmallerfirmsseemedto trainworkersslightlymore,andasignificantjumpinfavouroflarger firmsaftertheForneroLaw.29_{Theestimatesarenotsensitivetothe} in-clusionofregionandsectorfixedeffects,asshownincolumn(2),with aslightlysmallerestimatedeffectof1.5(ourpreferredestimate).30

28 _{All estimates and descriptive statistics are computed using RIL sampling}

weights.

29 _{Fig.1shows a slight increase in training provision for firms with 15 employ-}

ees, which, strictly speaking, should be not affected by the Fornero reform. This is however consistent with some heaping that we observe in the data, and the fact that some firms might have rounded their self-reported size from 16 to 15. This issue is discussed in the next Section. In particular, estimates sensitivity to dropping observations close to the cut-off (including exactly at the cut-off) is investigated through donut-hole regressions.

30 _{The number of observations between models omitting and including con-}

trol variables differs by few units owing to missing values in the controls. We

Intheremainingcolumns,weestimatethemodelinequation(2), allowingforadifferentbandwidtharoundthe15-employeethreshold. Theresults reportedacrosscolumnsconfirmthatthe𝑝𝑜𝑠𝑡×𝑎𝑏𝑜𝑣𝑒 co-efficientisalwayspositiveandstatisticallysignificantatconventional levels,withanorderofmagnitudethatvariesacrosscolumns,ranging from1.9incolumn(3)forthebandwidth11to20employeesto approx-imately3incolumn(7)forthelargestbandwidth(6to50employees). Again,wedetectveryminordifferencesdependingonwhetherornot firmcontrolsareincluded.Empiricalresultsarealsobroadlyconfirmed ifweconsideraquadraticpolynomialspecification,whichisreassuring, especiallyinthecaseofthelarger6–50bandwidthforwhichalinear terminfirmsizemaybeinsufficienttofitthedata(seeTableA1inthe onlineAppendixA).

Interestingly,the𝑎𝑏𝑜𝑣𝑒dummyisnegativeinallspecificationsand statistically significant in the 6–30 and 6–50 bandwidth cases; this meansthattherewerefewertrainedworkersabovethethreshold be-forethereform.Itispossiblethatinastronglyduallabourmarket,to escapethemorestringentfiringcostsonopen-endedcontractsabove thethreshold,firmswererelyingmoreontemporarycontracts.31

How-always report in the tables the estimates in the largest possible samples with non-missing values.

31_{Consistent with this prediction, BoeriandJimeno(2005)study the variable}

enforcement of EPL for permanent and temporary workers at the threshold to analyse the dynamics of hiring and firing in Italy. They find that firing decreases (increases) with size for permanent (temporary) workers; moreover, hiring is

Table2

Baseline pooled-cross section results.

(1) (2) (3) (4) (5) (6) (7) (8) post 1.084 ∗∗∗ _-2.416 ∗∗∗ _1.291 ∗∗∗ _-3.287 ∗∗∗ _1.084 ∗∗∗ _-2.611 ∗∗∗ _1.084 ∗∗∗ _-2.635 ∗∗∗ (0.137) (0.611) (0.303) (1.107) (0.137) (0.642) (0.137) (0.690) above -0.407 -0.487 -0.501 -0.718 -0.848 ∗∗ _-0.857 ∗∗ _-1.966 ∗∗∗ _-1.925 ∗∗∗ (0.382) (0.382) (0.575) (0.556) (0.358) (0.349) (0.412) (0.394) post ×above 1.722 ∗∗∗ _1.544 ∗∗∗ _1.946 ∗∗∗ _1.642 ∗∗∗ _2.049 ∗∗∗ _1.887 ∗∗∗ _3.075 ∗∗∗ _2.857 ∗∗∗ (0.422) (0.402) (0.594) (0.535) (0.383) (0.368) (0.532) (0.495) Bandwidth (6–25) (6–25) (11–20) (11–20) (6–30) (6–30) (6–50) (6–50)

Polynomial Linear Linear Linear Linear Linear Linear Linear Linear

Pol. inter. Above Above Above Above Above Above Above Above

Sec. ×year f.e. No Yes No Yes No Yes No Yes

Reg. ×year f.e. No Yes No Yes No Yes No Yes

Observations 16,486 16,462 7851 7836 17,826 17,797 21,266 21,229

R-squared 0.110 0.154 0.058 0.119 0.132 0.171 0.235 0.265

Note. Robust standard errors in parentheses, ∗∗∗_{p <0.01,} ∗∗_{p <0.05,} ∗_{p <0.1. The dependent variable is the number}

of trained workers. The analysis uses the 2010 and 2015 RIL waves. Polynomials in employment have been inter- acted with the dummy above (15-employee threshold). The number of observations between models omitting and including control variables may differ by few observations owing to missing values in the controls.

ever,temporaryworkerstendtoreceivelesstraining.TheForneroLaw, byreducingthewedgeinthedegreeofEPLenjoyedbypermanentand temporaryworkersinthecaseoffirmsabovethethreshold,mighthave inducedfirmstohiremorepermanentemployeesandthereforeto in-creasetrainingrelativetofirmswithfewerthan15employees. Alterna-tively,thereductioninthegapbetweentheEPLenjoyedbypermanent relativetotemporaryemployeesmighthaveincreasedtheincentivesfor firmstotrainmoretemporaryworkersasarguedbyDoladoetal.(2016). MechanismsareinvestigatedinSection7.

Returningtothemagnitudeofthe𝑝𝑜𝑠𝑡×𝑎𝑏𝑜𝑣𝑒coefficient,ifwefocus onourpreferredspecification,namely,thatwitha6–25bandwidth,a linearpolynomialandfirm-levelcontrols,ourresultssuggestthatfirms affectedbytheForneroLawmighthaveincreasedtrainingbya magni-tudeofapproximately1.5additionaltrainedworkers.Consideringthat beforethereform,theaveragenumberoftrainedworkersinfirmswith 15 employeeswas approximately 3.1,ourestimates suggestthatthe ForneroLawmighthaveincreasedthenumberof trainedworkersby approximately50%atthethreshold.

All potential threats toidentification arediscussed in the online AppendixB.

6. Robustnesschecks

WeconductseveralrobustnesschecksinTable3.First,becausethere isevidenceofheapingsintheforcingvariableatmultiplesoffive em-ployees,wefollowBarrecaetal.(2016)anddropfirmswith10,15,20 and25employeesfromtheestimationofequation(2).Reassuringly,the resultsreportedincolumns(1)and(2)ofTable3andthoseincolumns (1)and(2)ofTable2areverysimilar.32_{Second,incolumns(3)and} (4),werunaseriesofdonut-holeregressionstoaddresspossiblefirms’ self-sortingjustbelowthethresholdandtotakeintoaccountthe possi-bilitythatfirmswith16employeesare,infact,belowthethresholddue tothepresenceofpart-timeemployees:again,theresultsarebroadly

somewhat reduced at the threshold, with the emergence of an asymmetric U- shaped relationship between hiring and firm size.

32 _{Because the forcing variable is potentially continuous (i.e., the legal defi-}

nition of firm size, for which part-time workers count as fractions of full-time employees) but data limitations force us to treat the size as if it were discrete, we also re-estimate equation(2)by clustering standard errors by number of em- ployees as suggested by LeeandCard(2008). Reassuringly, we can reject the null hypothesis that the 𝑝𝑜𝑠𝑡×𝑎𝑏𝑜𝑣𝑒coefficient is equal to 0 at the 1% level of confidence.

unchanged.33_{Third,incolumns(5)to(8),werunaplaceboanalysisby} assumingthatthethresholdwasat10(20),ratherthanat15, employ-ees.Inthesecases,theestimatesofthediff-in-disctermarestillpositive butmuchsmaller(inthecaseof10employees)orlargelystatistically insignificant(inthecaseof20employees),asoneshouldexpectwith anincorrectlyspecifiedresearchdesign.

Intheremainingcolumns,werepeatthesameeconometricexercise butconsideramoregeneralspecification.Indeed,weallowthe poly-nomialinfirmsizetotakeondifferentcoefficientsbeforeversusafter thereformandnotjustaboveandbelowthethreshold,i.e.,weestimate differentversionsofequation(3).Theresultsincolumns(9)and(10) (linearpolynomial)confirmthemagnitudeoftheeffect,whichisequal to1.6and1.4,dependingontheexclusionornotofthefirmcontrols, re-spectively.Whenweconsiderapolynomialofsecondorder(columns11 and12),themagnitudeisslightlylargerthanthatreportedinprevious columns.

Aswehavealreadymentioned,theuseofrepeatedcross-sectionsin aDID-likeframeworkmightleadtoanestimationbiasifthe composi-tionofthecross-sectionschangessignificantlybeforeandafterthe re-form,possiblyastheresultoftheverysamereform.Indeed,theFornero Lawmighthavealteredtheincentivesforfirmstoself-selectbelowthe threshold.Although,aswereportintheonlineAppendixB,byrunning asetofSchivardiandTorrini(2008)testsandthetestinthechangein densitiesbeforevs.afterthereform,wedonotfindclearevidencethat thereformincreasedthepropensityforfirmstogrowinsizeorcrossthe 15-employeethreshold.Inwhatfollows,asafurtherrobustnesscheck (inadditiontothedonut-holeregressions),weinvestigatethispotential biasbyrestrictingtheestimationsampletothepanelcomponentofthe dataset,evenifthis reducesthesamplesizeandtheprecisionof the estimates.

InTable4,wereportestimatesofequations(2)and(3)witha poly-nomialof firstdegreewithandwithout firmcontrols;moreover,we includeasetoffirmfixedeffectstocapturepossibleunobserved firm-level heterogeneitypotentially correlatedwith treatmentstatus, and we clusterstandarderrors atthefirm level.Incolumns (1)and(2), whereweallowfordifferentpolynomialsonlybelowandabovethe 15-employeethreshold,wefindapositiveandstatisticallysignificanteffect

ofthe𝑝𝑜𝑠𝑡×𝑎𝑏𝑜𝑣𝑒interaction,butwithalowermagnitudecompared

tothepooledcross-sectionalestimates,ofapproximately1additional trainedworker.Incontrast,inthemoregeneralspecificationreported incolumns(7)and(8),whereweestimateequation(3),thecoefficient

33_{In regressions that are not reported but available from the authors upon}

request, we have re-estimated all models in Tables2,3 and 4, and the results are generally consistent.

Table3

Robustness: heaping, donut, fake thresholds, different interactions.

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

Heaping Donut Fake 10 Fake 20 Interaction post Interaction post

post 1.004 ∗∗∗ -2.646 ∗∗∗ _1.055 ∗∗∗ -1.843 ∗∗∗ _0.983 ∗∗∗ -2.657 ∗∗∗ _1.302 ∗∗∗ -2.316 ∗∗∗ _1.503 ∗∗∗ -1.886 ∗∗∗ _1.508 ∗∗∗ _-1.893 ∗∗ (0.139) (0.653) (0.133) (0.544) (0.134) (0.623) (0.135) (0.611) (0.390) (0.682) (0.547) (0.758) above 0.0336 -0.101 -0.240 -0.134 -0.702 -0.714 ∗ _{-0.867 -0.692 -0.356 -0.430 -0.359 -0.657} (0.421) (0.411) (0.529) (0.514) (0.493) (0.395) (1.722) (1.698) (0.478) (0.491) (0.732) (0.767) post ×above 1.384 ∗∗∗ 1.262 ∗∗∗ _1.566 ∗∗∗ 1.351 ∗∗∗ _0.810 ∗∗∗ 0.815 ∗∗∗ _{0.668 0.490 1.631} ∗∗ 1.437 ∗ _2.096 ∗ _2.064 ∗ (0.474) (0.450) (0.469) (0.446) (0.280) (0.248) (0.629) (0.611) (0.801) (0.764) (1.193) (1.143) Bandwidth (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25)

Polynomial Linear Linear Linear Linear Linear Linear Linear Linear Linear Linear Quadratic Quadratic

Pol. inter. Above Above Above Above Above Above Above Above All All All All

Sec. ×year f.e. No Yes No Yes No Yes No Yes No Yes No Yes

Reg. ×year f.e. No Yes No Yes No Yes No Yes No Yes No Yes

Observations 13,113 13,095 13,761 13,746 16,486 16,462 16,486 16,462 16,486 16,462 16,486 16,462

R-squared 0.109 0.151 0.116 0.159 0.108 0.153 0.106 0.151 0.111 0.155 0.111 0.155

Note. Robust standard errors in parentheses, ∗∗∗_{p <0.01,} ∗∗_{p <0.05,} ∗_{p <0.1. The dependent variable is the number of trained workers. The analysis uses the}

2010 and 2015 RIL waves. In columns (1) and (2) we drop multiples of 5 employees (heaping), in columns (3) and (4) we drop firms with 14, 15, 16 employees (donut); and the fake threshold in columns (5) and (6) is set at 10 employees while that in columns (7) and (8) is set at 20 employees. In columns (9) to (12) polynomials in employment have been interacted with the dummies above(15-employee threshold), post, and the above×post; in the table, these interactions are referred to as “All ”. The number of observations between models omitting and including control variables may differ by few observations owing to missing values in the controls.

Table4

Panel estimates.

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Baseline panel Heaping Donut Interaction post Non-switchers

post 1.360 ∗∗∗ _2.386 ∗ _1.217 ∗∗∗ _{2.130 1.231} ∗∗∗ _{2.060 2.250} ∗∗∗ _3.029 ∗∗ _1.376 ∗∗∗ _1.428 (0.125) (1.245) (0.135) (1.396) (0.126) (1.386) (0.363) (1.261) (0.126) (1.052) above -0.465 -0.443 -1.301 ∗ _{-1.131 -1.359 -1.134 -0.916 -0.997} (0.692) (0.688) (0.774) (0.757) (1.177) (1.174) (0.827) (0.823) post ×above 1.027 ∗∗ _0.838 ∗ _1.424 ∗∗ _1.222 ∗∗ _1.163 ∗ _{0.993 1.858} ∗ _1.869 ∗ _1.036 ∗ _0.832 (0.500) (0.495) (0.587) (0.579) (0.615) (0.610) (1.002) (0.988) (0.556) (0.550) Bandwidth (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25) (6–25)

Polynomial Linear Linear Linear Linear Linear Linear Linear Linear Linear Linear

Pol. inter. Above Above Above Above Above Above All All Above Above

Sec. ×year f.e. No Yes No Yes No Yes No Yes No Yes

Reg. ×year f.e. No Yes No Yes No Yes No Yes No Yes

Firm f.e. Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 5754 5732 3778 3766 4232 4220 5754 5732 4994 4976

R-squared 0.754 0.760 0.767 0.774 0.760 0.766 0.756 0.762 0.752 0.759

Note. Clustered standard errors at the firm level in parentheses, ∗∗∗_{p <0.01,} ∗∗_{p <0.05,} ∗_{p <0.1. The dependent variable is the number of trained}

workers. The analysis uses the 2010 and 2015 RIL waves. In columns (3) and (4), we drop multiples of 5 employees (heaping), and in columns (5) and (6), we drop firms with 14, 15, 16 employees (donut). In columns (7) and (8) polynomials in employment have been interacted with the dummies above(15-employee threshold), post, and the above×post; in the table, these interactions are referred to as “All ”. Finally, columns (9) and (10) estimate the baseline models of columns (1) and (2) but restricted to the sample of firms that remained above or below the cut-off in both waves ( “non-switchers ”). The number of observations between models omitting and including control variables may differ by few observations owing to missing values in the controls.

ofthe𝑝𝑜𝑠𝑡×𝑎𝑏𝑜𝑣𝑒interactionincreasestoapproximately1.9,whichis

statisticallysignificantatthe10percentlevel.Finally,incolumns(3)to (6),weconductsimilarrobustnesscheckstothoseconductedinTable3, i.e.,wetakeintoaccountpossibledataheapingsatmultiplesof5forthe forcingvariable,andwerundonut-holeregressions.Again,ourmain re-sultsareconfirmed.

Incolumns(9)and(10)wereportresultsforregressionsinwhichwe excludethosefirmsinthepanelthathavecrossedthethresholdbetween the2010and2015surveywavesin eitherdirection,sothatwe can keepthesamplesoftreatedanduntreatedfirmsunalteredbeforeand afterthereform.Thiscanalsobeimportantbecausethechangeinthe incentivestotrainworkersinducedbytheForneroLawcoulddepend onthenumberofyearsafirmisactuallyexposedtothesechanges,i.e. thenumberofperiodsthefirmisabovethecut-off.34_{Whenwedothat,} ourfixed-effectestimatessuggestthattheForneroreformmighthave

34 _{Specifically, firms above the cut-off in 2009 and that are above the cut-off in}

2014 are more likely to have been above the cut-off in all periods, and to have

determinedanincreaseofapproximatelyoneadditionaltrainedworker atthethreshold(about21%),aneffectthatisalmostidenticaltothe panel resultsreported incolumns (1)and(2)ofTable 4on thefull sample.35_{Itisworthnotingthatinthiscasethe𝑎𝑏𝑜𝑣𝑒dummydoesnot} changeovertime,thepolynomialsinfirmsizedropfromtheequation (unlesstheyareinteractedwith𝑝𝑜𝑠𝑡),andidentificationonlystemsfrom within-firmpre-vs.post-treatmentdifferenceslikeinaDIDdesign.

Apotentialthreattoourdiff-in-discestimatesmaycomefrom the misclassificationof the𝑎𝑏𝑜𝑣𝑒 indicator,sincewearenotableto pre-ciselymeasurethefirmlegalsize(seeSection4).Themeasurementerror islikelytodependonthenumberoftemporaryworkers(workingless than9months)andofpart-timeworkers.Computingtheaverageshare

been “fully ” exposed to the EPL reform. Conversely, in the cross-section sample we do not have any measure of the amount of firms’ “exposition ” to the reform.

35_{We drop 760 firm-year observations, which represents about 13 per cent of}

our panel sample. This indicates that most firms do not cross the cut-off in either direction between the two periods.

Table5

Measurement error.

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

Drop 16 Drop 16 and 17 Drop 16, 17 and 18

post 1.084 ∗∗∗ _-0.964 ∗ _1.084 ∗∗∗ _-1.039 ∗ _1.084 ∗∗∗ _-0.912 (0.137) (0.555) (0.137) (0.561) (0.137) (0.562) above -0.357 -0.344 -0.499 -0.485 -0.699 -0.706 (0.502) (0.486) (0.686) (0.650) (1.119) (1.042) post ×above 1.538 ∗∗∗ _1.407 ∗∗∗ _1.717 ∗∗∗ _1.584 ∗∗∗ _1.593 ∗∗∗ _1.454 ∗∗ (0.470) (0.447) (0.531) (0.503) (0.603) (0.569) Bandwidth (6–25) (6–25) (6–25) (6–25) (6–25) (6–25)

Polynomial Linear Linear Linear Linear Linear Linear

Pol. inter. Above Above Above Above Above Above

Sec. ×year f.e. No Yes No Yes No Yes

Reg. ×year f.e. No Yes No Yes No Yes

Observations 15,894 15,875 15,348 15,329 14,840 14,823

R-squared 0.108 0.145 0.107 0.145 0.105 0.143

Note. Robust standard errors in parentheses, ∗∗∗_{p <0.01,} ∗∗_{p <0.05,} ∗_{p <0.1. The dependent}

variable is the number of trained workers. The analysis uses the 2010 and 2015 RIL waves. Polynomials in employment have been interacted with the dummy above(15-employee thresh- old). The number of observations between models omitting and including control variables may differ by few observations owing to missing values in the controls.

ofworkersforeachfirmsizebinjustabovethecut-off,wefind percent-agesoftemporary(part-time)workersof12.4(11.5),12.4(14.3),13.1 (11.7),11.4(17.7)and12.0(14.5)forfirmswith16,17,18,19and20 employees,respectively.Onthebasisofthesefigures,weputforward thatfirmsaremorelikelytobemisclassifiedinthe16,17and18size bins.ThusTable5reportsdonut-hole estimatesdroppingthe16,16– 17and16to18firmsizegroups.Reassuringly,theestimatesarevery robustandsimilartotheonesobtainedinthefullsampleinTable2, suggestingthatfirmsizeisunlikelytobeaffectedbysubstantial mea-surementerror.

7. Potentialmechanisms

Aswehavementionedintheliteraturereviewsection,scholarshave suggestedthatinthecaseofduallabourmarkets,firmsmaytrytoavoid thecostsassociatedwithstricterEPLforregularworkersandincrease profitsbymakinggreateruseoftemporarycontracts.

Moreover,whenfiringcostsforregularworkersarehighandthere arerulesforbiddingtherenewaloftemporarycontracts,firmsmightbe reluctanttoconverttemporaryjobsintopermanentones.Thiscould,as aresult,increasetheincentivesforfirmstorelyonasequenceof tem-poraryjobs(CahucandPostel-Vinay,2002),therebyincreasing(excess) workerturnover. Cahucetal.(2016)presentasearchandmatching modelfeaturingregularjobs(withpossiblystricterEPL)andtemporary contracts(whichcanbeterminatedatzerocostwhentheyexpire,but whichcannotbeterminatedbeforetheirexpirydate):theyshowthat stricterEPLforregularworkersleadsfirmstoemploythelatteronly toexploitproductionopportunitiesthatareexpectedtolastforavery longtime.This,inturn,canleadtoanimportantsubstitutionof per-manentjobswithtemporaryones,leadingtoastrongexcessoflabour turnover:thispredictionfindsstrongsupportfromtheempirical stud-iesofHijzenetal.(2017)forItalyandCentenoandNovo(2012)for Portugal.

Ifthisisacorrectrepresentationofwhathappensinastronglydual labourmarket,theninlight ofthewidespread evidencethat tempo-raryworkersreceivelesstraining(Albertetal.,2005;Arulampalamand Booth,1998;Arulampalametal.,2004;Boothetal.,2002),onecould arguethatstricterEPLmightcauselowertrainingbyfirms,withthe me-diatingfactorsbeingtheexcessuseoftemporarycontractsandworker turnover.Clearly,onemightalsoexpectthattherelaxationofEPLfor permanentemployeesabovethethresholdbytheForneroLawshould beassociatedwithadecreaseinexcessworkerturnoverandintheuseof temporaryworkersatthethresholdbecauseofareductioninthewedge

Table6

Excess worker turnover and number of permanent workers. Dependent variable

(1) (2) (3) (4)

excess worker turnover number of permanent workers post 0.391 ∗∗∗ _0.486 ∗∗∗ _-3.013 ∗∗∗ _-3.557 ∗∗∗ (0.092) (0.092) (0.629) (0.725) above 0.098 ∗∗∗ _{0.025 -0.656} ∗∗ _-0.484 (0.032) (0.051) (0.265) (0.433) post ×above -0.104 ∗∗ _-0.135 ∗ _{0.504 1.735} ∗∗ (0.049) (0.075) (0.612) (0.738) Bandwidth (6–25) (6–25) (6–25) (6–25)

Polynomial Linear Quadratic Linear Quadratic

Pol. inter. All All All All

Sec. ×year f.e. Yes Yes Yes Yes

Reg. ×year f.e. Yes Yes Yes Yes

Observations 10,724 10,724 16,508 16,508

R-squared 0.197 0.205 0.737 0.738

Note. Robust standard errors in parentheses, ∗∗∗_{p <0.01,} ∗∗_{p <0.05,} ∗_{p <0.1.}

The analysis uses the 2010 and 2015 RIL waves. Excess worker turnover is calculated at the firm level following Hijzenetal.(2017), as 𝐸𝑊𝑇 = 2 ⋅ min ( 𝐻,𝑆)∕𝐸, where 𝐻 and𝑆 are the number of hiring and separations, respectively, and 𝐸 isthe average firm employment. Number of permanent workers represents the number of permanent contracts. Polynomials in em- ployment have been interacted with the dummies above(15-employee thresh- old), post, and above×post; in the table, these interactions are referred to as “All ”.

betweenfiringcostsforpermanentversustemporaryemployeesatthe cut-off.

To further explore the above conjectures, in Table 6, we report estimates of equation(3)using asdependent variableexcessworker turnover. FollowingHijzen etal.(2017), we measure excessworker turnoveras𝐸𝑊𝑇=2⋅ min(𝐻,𝑆)∕𝐸,where𝐻and𝑆 arethenumberof hiresandseparations(in2009and2014forthe2010and2015waves, providedbyRIL),respectively,and𝐸 istheaveragefirmemployment.36 Theresultsdisplayedincolumns(1)and(2)pointtowardsanegative

ef-36_{It can easily be shown that this formula is equivalent to the definition of}

excess worker reallocation as the difference between worker turnover and the absolute value of net employment change: it therefore represents worker flows in excess of job flows, and it is sometimes referred to as churning ( Burgessetal., 2000).