Can welfare conditionality combat high school dropout? Labour Economics

(1)

ContentslistsavailableatScienceDirect

Labour Economics

journalhomepage:www.elsevier.com/locate/labeco

Can welfare conditionality combat high school dropout? ^☆

Øystein Hernæs

^a^,^b^,^∗

, Simen Markussen

^a

, Knut Røed

^a

aThe Ragnar Frisch Centre for Economic Research, Norway

bInstitute for Social Research, P. box 3233 Elisenberg, 0208 Oslo, Norway

a r t i c le i n f o

JEL classiﬁcation:

H55 I29 I38 J18 Keywords:

Social assistance Activation Conditionality Welfare reform School dropout

a b s t r a ct

Basedonadministrativedata,weanalyzeempiricallytheeffectsofstricterconditionalityforsocialassistance receiptonwelfaredependencyandhighschoolcompletionratesamongNorwegianyouths.Ourevaluationstrat- egyexploitsageographicallydifferentiatedimplementationofconditionality.Thecausaleffectsareidentifiedon thebasisoflarger-than-expectedwithin-municipalitychangesinoutcomesthatnotonlycoincidewiththelocal timingofconditionalityimplementation,butdosoinawaythatcorrelateswithindividualexantepredicted probabilitiesofbecomingasocialassistanceclaimant.Wefindthatstricterconditionalitysignificantlyreduces welfareclaimsandincreaseshighschoolcompletionrates.

1. Introduction

Canaconditionalityregimedesignedtoactivate,counselandmoni- toryoungwelfarerecipientsplayaroleinreducingwelfaredependency andpromotinghighschoolcompletionamongvulnerableyouths?

Thelargeshareofyouthsthatdonotcompletehighschoolisacon- cerninmanydevelopedcountries;see,e.g.Lambetal.(2011)andOECD (2013).Secondaryeducationistoanincreasingextentconsideredthe basis,notonlyforfurtheruniversityorvocationaleducation,butalso forobtainingastablefootholdinthelabormarket.Dropoutratesare particularlyhighamongyouthswithsocio-economicallydisadvantaged backgrounds,andprobableconsequencesincludehighsubsequentun- employmentandlowearnings(RumbergerandLamb,2003;Campolieti etal.,2010).

Inthispaperweanalyzetheeﬀectsonyoungpeopleofbeingexposed toamorerestrictivepracticeregardingsocialassistanceclaims.There hasbeenanongoingdiscussioninNorwayofwhetherpartsofthewel- faresystemaretoolenient,andinthelate1990sandearly2000s,many localsocialinsuranceoﬃces– whichtraditionallyhavehadaconsider- ablediscretioninthedeterminationofpoliciesregardingmeans-tested socialassistance(welfare)– increasedtheiruseofsuchconditions.As weexplaininmoredetailbelow,thetypesofconditionsrangedfrom merely requiringclaimantstoattendcounselingmeetingswith case-

☆We gratefully acknowledge support from the Ministry of Labor and Social Aﬀairs and the Norwegian Research Council (grant No. 236992 ). We wish to thank the Telemark Research Institute for making their survey data available to us, Simen Gaure for programming assistance, and an anonymous referee for valuable comments and suggestions. Administrative register data from Statistics Norway have been essential for this project.

∗Corresponding author at: The Ragnar Frisch Centre for Economic Research, Gaustadalléen 21, 0349 Oslo, Norway.

E-mail addresses: [email protected] (Ø. Hernæs), [email protected] (S. Markussen), [email protected] (K. Røed).

workers todemand participation in fulltimeactivation programs. In somecases,theyalsorequiredwillingnesstoundertakeamedicalex- aminationand/ortodocument(orreduce)personalexpenses.Mostof theoﬃcesthatchangedpolicydidsoinaquitecomprehensiveway,in thesensethattheyincreasedtheiruseofseveralconditionssimultane- ously.

Conditionalitycanbeviewedasameanstooﬀsetmoralhazardprob- lemsembeddedinincomesupportprograms,aswellasatoolforimpos- ingamorestructureddailylifeoninactiveadolescents,andthuspre- ventmoreseriousmarginalization.Whenyouthsabouttodropoutfrom schoolshowupatthesocialinsuranceoﬃcetoseekincomesupport,a strictconditionalityregimemayinsomecasesbewhatisrequiredto convincethemtocompletetheirschoolingratherthantohavetopar- ticipateinstrenuoustrainingorcommunitywork.

Ourempiricalevaluationbuildsonadministrativedata,andinthe mainpartofouranalysis,westudytheincidencesofsocialassistance claims andhigh-school completion bythe age of 21 for Norwegian youthsbornbetween1972and1984.Theseoutcomevariablesarecou- pledwithsurvey-basedinformationfromlocalmunicipalitiesregarding changesinconditionality-practicesfrom1994through2004.Approxi- matelyhalfoftheNorwegianmunicipalitiesprovidedinformationabout theincidence,natureandtimingofsuchchanges.Identiﬁcationofthe

http://dx.doi.org/10.1016/j.labeco.2017.08.003

Received 12 September 2016; Received in revised form 2 August 2017; Accepted 7 August 2017 Available online 9 August 2017

(2)

causaleﬀectsofthechangesbuildsonabefore-after-comparisonofout- comes,whereweusepeopleinmunicipalitiesthatdidnotchangeprac- tice– orchangedpracticeatanotherpointintime– asimplicitcontrols.

Wedonotrelyonthestandardcommontrendassumption,though,as weidentifycausalitythroughtheinteractionbetweenaconditionality- indicator(treatment)andapre-determinedindividualsocialassistance propensityindicator.Theintuitionbehindthisstrategyisasfollows:If, say,theintroductionofconditionalityforsocialassistancepaymentsac- tuallyhadapositiveeﬀectonthelocalhigh-schoolcompletionrate,we shouldnotonlyobserveanincreaseinthelocalhigh-schoolcompletion rate,butweshouldseeanincreasethatisdisproportionallylargefor personswhohadahighexantelikelihoodofbecomingasocialassis- tanceclaimant.

Thereareclearlychallengesassociatedwiththisidentificationstrat- egyalso;themostimportantbeingthatlocalintroductionofcondition- alitymayhavebeentriggeredbyrisingsocialassistanceclaimsinthe past,whichevenintheabsenceofpolicyinterventionstendtobefol- lowedby“regressionstowardthemean”.Wereturntothis potential endogenous-policyproblemandotherthreatstowardouridentification approachafterhavingpresentedourmainempiricalstrategyandre- sults.Thebottomlineisthatwefindnoevidenceofpolicyendogeneity, andthatourresultsarehighlyrobustwithrespecttoboththechoiceof pre-treatment(comparison)period,thewayweallowforlocal(differ- entiated)trends,andanumberofothermodelingissues.

Ourpaperrelatestoalargeexistingliteraturedocumentingmoral hazardproblemsinsocialinsuranceprograms;seeKruegerandMeyer (2002)foranoverviewoftheliterature,andRøedandZhang(2003;

2005)andFevangetal.(2017)forrecentNorwegianevidence.Italso relatestoafast-growingliteratureontheimpactsofactivation,monitor- ing,andsanctionsinsocialinsuranceaswellaswelfareprograms;see, e.g.,Blank(2002),Moﬃtt(2007),andRøed(2012)forrecentreviews.

Aconsensusviewcomingoutofthisliteratureisthatactivation,aswell asmonitoringandsanctions,dotendtolowerthepubliccostsofpro- vidingtransferprograms,bothbyreducingthenumberofclaimsandby reducingtheiraverageduration.Mostofthepapersalsoidentifyfavor- ableeﬀectsonsubsequentemploymentandearnings.Apaperofpartic- ularrelevanceforourowncontributionisDahlbergetal.(2009)who investigatestheimpactsofmandatoryactivationprogramsforwelfare recipientsinSweden,takingadvantageofagradualintroductionofsuch programsinStockholm.Akeyﬁndingoftheirpaperisthatactivation requirementsimproveemploymentandearningsprospectsforyoung persons(aged18–25)considerably,buthaveno,orevennegative,ef- fectsonadults.Wearenotawareofexistingresearchlookingdirectly attheimpactsofsocialassistanceconditionalityonhighschoolcomple- tion.

Whyshouldsocialassistanceconditionalityaﬀecthighschoolcom- pletion?Asweexplaininmoredetailbelow,alladultsinNorway(i.e., personsaged18yearsormore)whoareunabletosupportthemselves, areentitledtomeanstestedsocialassistance.Yet,aslongastheappli- cantisenrolledinregularsecondaryeducation,socialassistanceclaims mayberejectedwithreferencetotheparents’ economicsituation,even whentheapplicantisabove18years.Hence,socialassistancetoyoung adultsprimarilyrepresentsaneconomicsafetynetforpupilsfromvery poorfamiliesandforadolescentswhoquitschool,butfailtoﬁnd– or evengenuinelysearchfor– gainfulemployment.Akeyroleofcondition- alityinthiscontextistoraisethepotentialcostofquittingschool,asthe alternativeoflivingonwelfaremaybecomeconsiderablylessattractive.

Inaddition,itispossiblethatsomeofthosewhoclaimwelfaredespite thestricteruseofconditionalityarepushed/coercedbacktoschoolby theactivitiesimpliedbytheconditions.Whiletheformermechanism impliesthatconditionalitycauseshighschoolcompletiontosubstitute forsocialassistanceclaims,thelatterimpliesthatitcomplementsthem.

Ourempiricalﬁndingsindicate thatwhenalocalinsuranceoﬃce increasestheiruseofconditionality,welfareclaimsamong21-yearolds inthatareadeclinesubstantially,whilehighschoolgraduationrates increase.Forexample,forthequarterofindividualsestimatedtohave

thehighestpropensitytoreceivewelfare,theincidenceofwelfarere- ception fallsbyaround3.1percentagepoints,while thehighschool graduationrateincreasesby2.2percentagepoints.Thefavorableeﬀects onhighschoolcompletionisfullyexplainedbyahigherprobabilityof completingwithoutclaimingsocialassistance;henceconditionalityin- duceshighschoolcompletiontosubstituteforsocialassistanceclaims.

Wealsoﬁndevidencethatthefavorableeﬀectsofconditionalitypersist andcontributetohighereducationalattainment,higherlaborearnings, andlowertransferdependencyatage25.

2. Institutionsanddata

AccordingtoNorwegianlegislation,adultpersons(aged18ormore) whoareneitherabletosupportthemselvesthroughworknorcovered bysocialinsuranceprograms,areentitledtomeans-testedsocialassis- tancefromtheirmunicipality.Thereisonepossibleexceptionfromthis rule,however,andthatisiftheyoungadultisstillinregularsecondary education,andtheparentsaredeemedtohavesufficienteconomicre- sourcestosupporttheiradultoffspring.Inthiscase,thecaseworkermay obligatetheparentstosupporttheiroffspringeconomically.Thisissub- jectedtoadiscretionarydecision,however,andsocialassistancecannot berejectedunlessitisclearthattheparentsactuallytakeontheireco- nomicresponsibility.Yet,itcannotberuledout thatthelegislation’s referencetocontinuedparentalresponsibilityduringhighschoolen- rolmentmayrepresentanincentivetoquitschoolforsome(potential) socialassistanceclaimants.

Theprobabilityofclaimingsocialassistanceduringacalendaryear peaks at a level close to 7% by age 20–21, after which it declines monotonouslywithage;seeFig.1.Thehighclaimratesatage20–21are drivenbyacombinationofrelativelyhighratesofunemploymentdur- ingtheschool-to-worktransitionphase,andlowlevelsofsocialinsur- ancecoverage;thelatterbecausesocialinsuranceentitlementsrequire pastworkexperienceandsocialsecuritycontributions.

Thelegislationimpliesthatlocalauthoritiescannotrefusetohelp personsintrueneed.Theycansetconditions,however,forexamplein theformof workrequirements,providedthattheconditionsarenot disproportionateorunreasonable.¹Intheperiodcoveredbyourdata, themunicipalitieshavehadampleroomfordiscretionregardingthe useofsuchconditions,andthepracticeshavevariedalotacrossthe country.²In2006,TelemarkResearchInstitute(TRI)publishedareport ontheNorwegiansystemofmeans-testedsocialassistance(Brandtzæg etal.,2006).Aspartofthiswork,theauthorsadministeredasurvey toalllocalsocialinsuranceofficesinNorway,asking,interalia,about changesduringthelast10years(1994–2004)intheoffices’ practices regardingtheuseofconditionsforreceivingsocialassistance.Itisthe answerstothesequestionsthatformthebasisforidentificationofthe treatmentsevaluatedinthispaper.Basedonthesocialinsurancedistrict ofresidenceatage21,wematchthetreatmentdatatopopulation-based administrativeregisterscontaininginformationaboutindividualsocial assistanceclaims,educationalandlabormarketoutcomes,aswellasa largerangeof(family)backgroundcharacteristicsforallpersonsborn between1972and1984.

Intotal247ofthe470localinsuranceoﬃces(locatedin433mu- nicipalities)existingin2005returnedtheTRI-survey.³Outofthese,46 oﬃcescouldnotbeusedbyusduetomissinginformationabouttiming, ambiguitywithrespecttothedirectionofchanges,inconsistentinfor-

1Lov om sosiale tjenester i arbeids- og velferdsforvaltningen (Sosialtjenesteloven), §§

18–20.

2New legislation implies that activation requirements now have become compulsory for social assistance claimants who are deemed able to work.

3With the exception of the largest cities (Oslo and Bergen) there is one single social insurance office in each municipality, ensuring that adolescents living in the same municipality at the same point in time have all been subjected to the same treatment status. Since we do not have sufficiently detailed information about address to link adolescents in Oslo and Bergen to the correct social insurance office, we have dropped these municipalities from the analysis.

(3)

0.02.04.06.08

20 30 40 50 60 70

Fig. 1. Fraction receiving social assistance (welfare) by age in 2011.

Table 1

Sample restrictions – social insurance districts.

Number of social insurance districts in Norway 470 - Non-responding districts − 223

= Oﬃces with returned surveys 247 - Missing time information − 32 - Cannot link oﬃce to individuals − 7 - Ambiguous policy change − 6 - Inconsistent information − 1

= Final sample 201

mation,orlackoflinktoindividuals;seeTable1.Hence,ouranalysis buildsoninformationfrom201socialinsurancedistricts(municipalities),coveringroughly60%oftheNorwegianpopulationintherelevant birthcohorts.Outofthese,43unambiguouslyincreasedtheiruseofcon- ditionsatsometime,and158maintainedstatusquo.Itisnotablethat noneofthesocialinsuranceoﬃcesunambiguouslyreducedtheiruseof conditionality.Toobtainabetterideaonthegeographicaldistribution ofthe43treatmentandthe158controlmunicipalities,Fig.2provides amapofNorwaywherethetreatmentandcontrolmunicipalitiesare highlighted.Asonecansee,bothtreatmentandcontrolmunicipalities arescatteredacrossthecountry.

Despitethelackofgeographicalconcentration,thefactthatwecan usedatafromlessthanhalfoftheNorwegianmunicipalitiesdoesraise questionsaboutgeneralizability.InTable2,weshowdescriptivestatis- ticsforthreegroupsofmunicipalities;thosewhodidnotreplyandfor thatreasonarekeptoutoftheanalysis,thosewhorepliedanddidnot change theirpolicies – whichwillserve as thecontrolgroupin the analysis– andthosewhorepliedandchanged theirpolicies– which constitutesourtreatmentgroup.Foreachgroupwepresentdescriptive statisticsfortwoyears,1993and2005,thatareonoppositesidesof anypolicychange.Thesocioeconomiccharacteristics,aswellastheir developments,aresimilarforthethreemunicipalitytypes.Itisnotable, however,thatthefractionreceivingwelfarebeneﬁtsdeclinedmostin thetreatedmunicipalitiesandleastinthecontrolmunicipalities.

Thepolicyshiftstowardstricterconditionalitywereconductedindif- ferentcalendaryearswithamajorityofthereformstakingplacetoward theendofthe1994–2004period(seeAppendix,TableA1fordetails).

Thisprobablyreﬂectsanincreasingconcernaboutrisingwelfareexpen-

dituresandageneralshifttowardmoreemphasisonactivationinsocial policies;see,e.g.,Gubriumetal.(2014).

TheTRI-surveydistinguished9diﬀerentcondition-types.Theseare describedinTable3,togetherwithanoverviewoftheirfrequenciesin the43socialinsuranceoﬃceswhichimplementedatleastoneofthem.

Onaverage,thereformingsocialinsuranceoﬃces(municipalities)re- portedtohavechanged4.14suchpoliciesatthesametime.Thefour mostcommonconditionsusedare(i)arequirementofdocumentingex- penses(29cases),(ii)requirementtoparticipateinaprogramtypically involvingworkortraining(26cases),(iii)requirementtoparticipatein generalcounselling(26cases),and(iv)arequirementtoregisterasan activejobseeker(25cases).

Youngwelfare recipients werebyfarthegroupforwhomcondi- tionalitywasappliedthemost– 97%ofrespondentsreportedthatthey

“often” usedconditionstowardsthisgroup.TheTRI-report(Brandtzæg etal.,2006)alsocontainstranscriptionsofinterviewswithcaseworkers, explaininginmoredetailwhyandhowconditionalityhasbeenusedin practice.Theinterviewsindicatethattheconditionshaveﬁrstandfore- mostbeenusedforyoungclients(below25yearsofage),withafocus on preventingthemfrom startingtheirpotentiallabormarket career aswelfareclients,andthatconditionalitythereforetypicallyinvolved somesortofactivationrequirement,eitherintheformofcommunity work,ortraining/education.Inmanycases,theconditionsaredesigned suchthattheyareeﬀectiveimmediately,e.g.,byrequiringapplicants toshowupatsomestructuredactivityalreadythefollowingmorning.

Thispotentiallyinducessome“secondthoughts” aboutalifeonwelfare andthusgeneratesa“threateﬀect” ofthetypereportedbyBlacketal.

(2003).Andforthosewhochoosetosatisfytheconditions,theactivities mayrepresentagreatlyneededelementofstructureinthedailylife,a pointalsoemphasizedbycase-workers(Brandtzægetal.,2006,p.115).

Notethatwedonotexploitinformationabouttheuseofconditions inthecontrolmunicipalities,exceptthattheydidnotchangepolicybe- tween1994and2004.Wearethusnotgoingtocomparemunicipalities withandwithoutconditionsinthispaper,butfocusexclusivelyonthe waychangesinoutcomescoincidewithchangesinconditionality.Note alsothatwedonothaveanyinformationaboutpolicychangesoccurring after2004.InamorerecentsurveyofNorwegianmunicipalities(Proba Research,2013),ithasbeenshownthatthetrendtowardmoreintensive useofconditionscontinuedafter2004.Hence,inordertoavoidcontam-

(4)

Fig. 2. Treatment and control municipalities.

inationofcontrolmunicipalitiesintheformofunobservedtreatments, welet2005beourlastobservationyear.

Giventheapparentlargedifferencesincontent,wewouldhaveliked eithertoevaluatetheimpactsofdifferentcondition-typesseparately,or toevaluatealternative“reformpackages”.However,duetothesimul- taneityintheimplementationofthevariousconditionsandthelarge number(37)ofcondition-combinationsactuallyobserved,thisissim- plynotdoable.Inthemainpartofouranalysis,wearethereforegoing tousetheimplementationofnewcondition(s)asasingledichotomous treatmentvariable.Thetreatmentindicatorthusreflectsthatthelocal socialinsuranceadministrationhastakendeliberate– andinmostcases several– stepstotightentheconditionsforpayingoutsocialassistance.

Inasmuchas89%ofthetreatmentcases,anactivation-conditionality wasincludedinthe“reformpackage”.Inasupplementaryanalysis,we alsoprovideseparatepartialeﬀectestimatesforeachofthethreemain typesofconditions;i.e.,activationrelated,healthrelated,orpersonal- economyrelated,respectively.

Apartfromthesurveydatacoveringthesocialinsuranceoﬃcepoli- cies,thedatausedinthispaperallstemfromadministrativeregisters coveringthecompleteNorwegianpopulation.Inourmainanalysis,we studyoutcomesfor21-yearoldswhoatthatageresidedineitheroneof thecontrol-ortreatedmunicipalities.Weincludeinthedatasettheco- hortsbornbetween1972and1984,whoturned21yearsintheyearsbe- tween1993and2005.Sincetheactualtimingofthepolicyshiftwithin

(5)

Table 2

Municipality characteristics in excluded, control and treated municipalities.

Excluded municipalities ( n = 178) Control municipalities ( n = 158) Treated municipalities ( n = 43)

1993 2005 1993 2005 1993 2005

Inhabitants 11,674 12,621 7,207 7,581 10,392 11,235

Employment rate 0.66 0.69 0.67 0.70 0.65 0.69

Mean income (1000 NOK, inﬂated to 2015 value; see note below) 361 399 343 374 334 369

Fraction with tertiary education 0.23 0.31 0.18 0.25 0.17 0.24

Fraction with at least secondary education 0.47 0.62 0.42 0.58 0.41 0.58

Fraction receiving welfare beneﬁts 0.027 0.020 0.021 0.017 0.027 0.019

….below age 30 0.039 0.028 0.033 0.027 0.041 0.030

Fraction receiving disability beneﬁts 0.085 0.094 0.087 0.102 0.090 0.104

Unemployment rate 0.044 0.030 0.040 0.027 0.046 0.028

Note: All variables refer to the age group 18–61 years, and reported means are weighted by population size. Income levels are measured in 1000 NOK, inﬂated to 2015-value with the adjustment factor used in the Norwegian pension system (approximately corresponding to the average wage growth).

Table 3

Policies and conditions changed, conditional on at least one policy change.

Activation and work requirements Number of

municipalities

Fraction of treated

Fraction of treated persons Participate in program: A requirement to take part in a work/training or educational program. 26 0.60 0.72 Work for welfare: Requirement to participate in a work program either organized by the municipality or

others.

15 0.35 0.21

Register as seeking work: A requirement to register as an active job-seeker, keeping an updated CV etc. 25 0.58 0.62 General counseling: Attend counseling meetings with caseworker or others to discuss the current situation. 26 0.60 0.59 Career counseling: Attend career counseling meeting(s) with caseworker or others to improve work

prospects.

10 0.23 0.18

At least one activation/work requirement 41 0.95 0.89

Health

Health examination: Willingness to undertake a health examination. 14 0.33 0.22

Economic

Document expenses: A requirement to show documentation for housing costs and other additional costs exceeding the welfare beneﬁt

29 0.67 0.65

How to use the beneﬁt: Restriction on how the recipient spend the beneﬁt 17 0.40 0.37 Move to cheaper housing: Refuse to cover housing costs exceeding the norm and require that one move to

cheaper housing for obtaining housing support.

16 0.37 0.48

At least one economic condition 34 0.79 0.79

Total number of conditions changed 175

Total number of municipalities changing policy 43

ayearisunknowntouswehavechosentoexcludethereform-yearco- hortinthetreatedmunicipalities.Thedataalsocontainlinksbetween childrenandparents,makingitpossibleforustoincludeinformation aboutthechildren’sparents,includingtheirearnings,countryororigin, ageandeducation.

Ourmainoutcomesofinterestaresocialassistance(welfare)recep- tionandhighschoolcompletion byindividuals’ 21styear(thestan- dard/normalageofcompletionis19).Somedescriptivestatisticsare showninTable4.Inafollow-upanalysistowardtheendofthepaper, wealsoexaminevariouslabormarketandeducationoutcomesatage 25.

3. Empiricalanalysis

Inthissection,wesetupandestimatestatisticalmodelsaimedat identifyingthecausaleﬀectofsocialassistanceconditionalityonthe probabilityofactuallyreceivingsocialassistanceduringthecalendar yearinwhichpersonsbecome21years,andontheprobabilityofhaving completedhighschoolbythatage.

Withinourdatawindowjustabout8%oftheadolescentsreceived socialassistanceduringthecalendaryeartheyturned21years.Stricter conditionsforwelfarebenefitsarethuslikelytohavenegligibleimpacts onthemajorityofyouths,andanycausaleffectscanbeexpectedtobe largerthemoreexposedapersonistotheriskofbecomingawelfare claimantinthefirstplace.Thisargumentisgoingtoplayakeyrole inouridentificationstrategy. Thefirststepofthisstrategyisthusto identifyindividual“exposurerisks”,basedonpre-determinedparental characteristicsonly.Inasecondstep,weinteractthepredictedpropen-

sitieswithtime-varyingindicatorsofconditionality-reform.Intuitively, foralocalshiftinindividualoutcomestobeinterpretedascausallyre- latedtotheintroductionofconditionality,itisnotsufficientthatthe shift islargerin reformingthanin non-reformingmunicipalities; the differencesalsoneedstobepositivelycorrelatedtoindividualpredicted exposurerisks.Ourempiricalstrategyissimilartotheapproachusedby MarkussenandRøed(2016)toevaluateanothersocialprogramwitha small,butimperfectlyidentifiable,targetgroup.

3.1. Auxiliaryregressionanalysis:thepropensityofwelfareuptakeatage 21

Westartoutbyestimatingthepropensityofwelfareuptakeatage21, basedonpre-determinedfamilybackgroundcharacteristicsonly.Todo thisweconstructasimilardatasetastheoneusedinthemainanalysis (anddescribedintheprevioussection),butcontainingonlythe1971 birth-cohortinthetreatmentandcontrolmunicipalities;i.e.,thelast birth-cohortnotusedinourcausalanalysis(23,852observations).We thensetupalogitregressionmodelwithanindicatormodelforwelfare receiptatage21(duringthecalendaryearofthe21stbirthday)asthe dependentvariableandavectoroffamilybackgroundcharacteristics𝐛_𝐢 asexplanatoryvariables.Thevectorofexplanatoryvariablesincludes bothparents’ educationattheoﬀspring’sage10(4categoriesforeach parent)andtheirrespectivecumulativeearningsbetweentheoﬀspring’s ages 0and10.Inaddition weincludedummy variablesfor parents’

countryoforigin(7categories).Theresultsfromthisregressionshow thatfamilybackgroundcharacteristicsarepowerfulpredictorsforlater social assistanceclaims; see theAppendix, TableA2, fordetails.We

(6)

Table 4

Descriptive statistics for estimation sample.

Mean SD

Outcomes

Welfare uptake at age 21 0.081

Completed high school by age 21 0.693

Background characteristics

Fraction female 0.484

Parental income, mean over child’s age 0–9, 1000 NOK (2015-value)

…Father 503 190

…Mother 133 130

Parental education, when child is 10 years

…Father has college degree 0.209

…Father has high-school 0.516

…Mother has college degree 0.171

…Mother has high-school 0.493

Nationality background

…Native 0.893

…Western Europe or North America 0.076

…Rest of the world 0.030

Calendar year turning 21 1998.7 3.777

Treated by age 21 0.051

Local unemployment rate at age 21 0.048 0.026

Number of observations 259,220

Number of municipalities 201

Table 5

Descriptive statistics by quartile in the predicted welfare propensity distribution. 1972–84 birth cohorts.

Q1 Q2 Q3 Q4

Mean predicted welfare propensity, based on 1971-cohort 0.031 0.058 0.086 0.158 Outcomes

Welfare uptake at age 21 0.030 0.052 0.084 0.159

Completed high school by age 21 0.850 0.751 0.662 0.510 Background characteristics

Mean parental income (1000 NOK, 2015-value)

..Father 678 517 454 363

…Mother 201 141 110 79

Parental education

…Father with college 0.657 0.109 0.051 0.018

…Father with high school 0.327 0.782 0.590 0.365

…Mother with college 0.509 0.124 0.039 0.014

…Mother with high school 0.476 0.813 0.593 0.090

Nationality

…Native 0.922 0.932 0.907 0.815

…Western Europe or North America 0.071 0.058 0.077 0.100

…Rest of the world 0.008 0.010 0.016 0.085

Number of observations 64,798 64,808 64,805 64,809

canthususetheseresultsobtainedforthe1971-cohorttomakeout-of- samplepredictionsforthe1972–84cohortsusedinourcausalanalysis. Thatis,wecomputeawelfarepropensityscore ̂𝑝𝑖as

̂𝑝𝑖= exp(𝐛^′_𝐢̂𝛑)

1+exp(𝐛^′_𝐢̂𝛑), (1)

wherê𝛑isthevectorofparameterestimates(includingaconstantterm) fromthe1971-cohortwelfareclaimregression.

Toillustrate the empiricalrelevance of these predictions for the 1972–84cohortsusedinthecausalanalysis,wehavedividedthemem- bersofthesecohortsintofourquartiles,basedontheirpositioninthe distributionof̂𝑝_𝑖,andpresentinTable5descriptivestatisticsseparately for eachquartile. Afirst pointtonote is that thepredicted welfare propensitiesquitenicelymatchestheactuallyrealizedclaims.Asecond pointtonoteishowstrikinglydifferentfamilybackgroundspersonsin thedifferentquartilestendtohave.Forexample,thelikelihoodofhav- ingafatherwithacollegedegreeis37timeshigherinthefirstthanin thefourthquartile,whereasthelikelihoodofhavingparentswhoim- migratedfromanon-westerncountryis11timeshigherinthefourth quartilethaninthefirst.

Thepredictedwelfarepropensities ̂𝑝_𝑖canbeusedtoillustratehow purechangesin thepopulationcompositionhave(orhavenot)con-

tributedtochangesinwelfareclaimsoverthecohortsusedinthecausal analysis.Fig.3showsthe ̂𝑝_𝑖-valuesforthe1st,5th,10th,25th,50th, 75th,90th,95thand99thpercentileinthedistributionofwelfareup- takepropensityforeachcohort/yearinthegroupoftreatmentandcon- trolmunicipalities.Withsomeexceptionsattheveryhighestpercentiles, theﬁgureindicatesparalleltrendsinthetreatmentandcontrolmunic- ipalities. Thepredictedclaimpropensityin thetenupperpercentiles increasedsomewhatmoreinthetreatmentmunicipalitiesthaninthe controlmunicipalitiestowardtheendoftheobservationperiod.The reasonforthisisthatthefractionof21yearoldswithnon-nativepar- entsincreasedmoreinthetreatedmunicipalities,andtheseyouthshave ahigherpredictedwelfareuptake.

Thefactthatthefraction ofimmigrant youthsincreased morein thetreatmentthaninthecontrolmunicipalitiesmayraiseconcernsre- gardingourabilitytodisentangletheimpactsofthisparticularchange fromthecausalimpactsofthereforms.Asimmigrantyouths tendto havehigherwelfareuptakepropensities,andalsosomewhatlowerhigh schoolcompletionrates,ifunaccountedfor,therisingimmigrantshare maymaskanyfavorabletreatmenteﬀects.Asweexplainbelow,wedo accountforimmigrantstatusinouranalysis,andwewillreturntoa morespeciﬁcrobustnessanalysiswithrespecttothistopicinSection 4.4.Anotherconcerncouldbethattherisingshareofimmigrantsinthe

(7)

0 .1 .2 .3 .4 .5

1993 1995 1997 1999 2001 2003 2005

Control group: Grey solid line, Treatment group: Black dashed line

Fig. 3. Predicted propensity of welfare uptake ( ̂𝑝 _𝑖) over time in treatment and control municipalities.

Note: The graph draws the 1st, 5th, 10th, 25th, 50th, 75th, 90th, 95th and 99th percentile in the distribution of predicted welfare propensities within each year and within treatment and control municipalities.

treatmentmunicipalitiesresultedindiﬀerenthighschoolpassingstan- dardsforall,includingnatives.Thisishighlyunlikely,however,asthe highschoolsinNorwayaretheresponsibilityofcountiesandnotmu- nicipalities(with23municipalitiesineachcountyonaverage),andas highschoolcompletionalsoinvolvesanumberofanonymousnational tests.

3.2. Causalregressionanalysis:theeﬀectsofwelfareconditionality

Inthissection,weexaminetheimpactsofwelfareconditionalityon individualindicatorvariablesforwelfareclaimsandhighschoolcom- pletion,respectively,bothmeasuredatage21.Thebasicideaofour empiricalstrategyistoassesswhetherthereisatendencyforoutcomes toshiftinresponsetotheintroductionofconditionalityinawaythat correlateswithpredictedpropensityofwelfareuptakê𝑝_𝑖.Beforeweturn totheformalregressionanalyses,weprovideasimplegraphicalexpo- sitionofhowthisidentiﬁcationstrategyplays outinthedata.Fig.4 presents(calendaryearadjusted)averageoutcomesforthe10deciles inthê𝑝_𝑖-distributionforthetreatmentgroupbeforeandafterthepolicy shift.

Startingoutwithpanel(a),showingwelfareuptakeatage21,we seethatthetwogroupsarealmostidenticalfortheﬁrstsixdecilesin thepredictedwelfarepropensitydistribution.However,forthefourup- permostdeciles, thetreated municipalitieshadasubstantiallyhigher welfareuptakebeforethanafterthepolicyshift.Asimilarpicturecan beseeninpanel(b)showinghighschoolcompletionbyage21.Inthe lowersevendecilesinthepredictedwelfarepropensitydistributionwe canhardlyseeanydiﬀerences,whereasfortheupperthreedecilesthere isaclearshifttowardshighercompletionratesafterthepolicyshift.

Wenowturntotheformalregressionanalysis.Lety_itdenotetheout- comeofinterestforpersonimeasuredincalendaryeartandletC_mtbe anindicatorvariableequalto1intreatmentmunicipalitiesinallyears strictlyaftertheintroductionofconditionalityandotherwisezero(we dropfromtheanalysisalloutcomesmeasuredinthesameyearasare- form,sinceinthesecaseswedonotknowwhetherclaimsweremade

beforeoraftertheintroductionofconditionality).Intreatmentmunici- palitiesitwillthusbethecasethatallpersonswithC_mt=1havebeen exposedtothenewconditionalityregimeatleastoneyearatage21, whereas personswithC_mt =0hadnot beenexposed toitatall.We alwaysclusterstandarderrorsatthe201municipalities.Furthermore, letx_ibeavectorofindividualcovariatesincludingfamilybackground characteristics(b_i)andgender(seeacompletelistinTableA3inthe Appendix),andletu_mtbethemunicipality-specificunemploymentrate inyeart.Westartoutwithasimpledifference-in-difference(DiD)de- sign,andestimatelinearprobabilitymodelswiththefollowingstructure

𝑦_𝑖𝑡=𝐱^′_𝐢𝛃+𝜆_𝑚+𝜎_𝑡+𝜌𝑢_𝑚𝑡+𝜃𝐶_𝑚𝑡+𝑣_𝑖𝑡, (2)

where(𝜆m,𝜎t)aremunicipalityandtimefixedeffects,respectively,and v_itisaresidual.Thecoefficientofinterestistheintentiontotreat(ITT) effect𝜃,whichcapturestheextrashift– overandabove thegeneral changescapturedbytheyearfixedeffects– occurringintreatmentmu- nicipalitiesaftertheintroductionofconditionality.Theresultantesti- matesof𝜃areprovidedinTable6,Column(1),andindicatethatthe introductionofconditionalityreducedtheprobabilityofwelfareuptake atage21by1.1percentagepointsandraisedtheprobabilityofhigh schoolcompletionby1.2percentagepoints.Theseaffectsappearsmall.

Model(2)isnotparticularlyinformative,however,sinceitexaminesan intentiontotreateffectonapopulationinwhichthemajorityisalmost certaintohavebeenunaffectedbythetreatment;i.e.,youthsforwhich socialassistanceisnotarelevantalternativeregardlessofconditionality regime.Asdiscussedabove,giventhattherearecausaleffectsofcon- ditionality,wewouldexpectthemtobelargerthelargeristheexante probabilityofbeingexposedtoit.

Toinvestigatethisfurther,weestimateEq.(2)separatelyforeachof thequartilesinthedistributionofpredictedwelfarepropensitieŝ𝑝_𝑖.The resultsfromthisexercisearedisplayedinTable6,Columns(2)–(5).As expected,wefindnoeffectsinthefirstquartile,andthenincreasingef- fectsaswemoveupinthewelfarepropensitydistribution.Intheupper quartile,weestimateanITTeffectofconditionalityequalto−3.1per-

(8)

0.05.1.15.2.25

0 .05 .1 .15 .2 .25

Predicted probability of welfare uptake

(a) Welfare uptake at age 21

.4.5.6.7.8.9

0 .05 .1 .15 .2 .25

Predicted probability of welfare uptake

(b) Completed high school by age 21

Pre treatment Post treatment

Fig. 4. High school completion and welfare uptake before and after treatment

Note: Outcomes have been calendar-year-adjusted by regressing them on calendar year dummy variables, obtaining the residuals, and then adding a constant term such that the outcomes are measured in 2000-levels.

Table 6

Main results. Estimated intention to treat (ITT) and average treatment eﬀects on the treated (ATET) of welfare conditionality (standard errors in parentheses).

( 1 ) ( 2 ) ( 3 ) ( 4 ) ( 5 ) (6) (7) A. Dependent variable: Welfare uptake at age 21

ITT all − 0.011 ^∗^∗

(0.006)

ITT quartile 1 − 0.001

(0.003) –

ITT quartile 2 − 0.007 − 0.006

(0.005) (0.006)

ITT quartile 3 − 0.009 − 0.009

(0.010) (0.010))

ITT quartile 4 − 0.031 ^∗^∗ − 0.025 ^∗^∗

(0.010) (0.010)

ATET − 0.196 ^∗^∗^∗

(0.074) Mean of dependent variable 0.08 0.03 0.05 0.08 0.16 0.08 0.08 B. Dependent variable: Completed high school by age 21

ITT all 0.012 ^∗

(0.007)

ITT quartile 1 − 0.005

(0.010) –

ITT quartile 2 0.004 0.009

(0.010) (0.014)

ITT quartile 3 0.028 ^∗^∗ 0.032 ^∗^∗

(0.011) (0.016)

ITT quartile 4 0.022 ^∗ 0.024 ^∗^∗

(0.013) (0.016)

ATET 0.170 ^∗

(0.094) Mean of dependent variable 0.69 0.85 0.75 0.66 0.51 0.69 0.69 Number of observations 259,220 64,798 64,808 64,805 64,809 259,220 259,220 Note: Quartiles relate to individual predicted welfare propensity as reﬂected in the ̂𝑝 distribution. Standard errors are clustered at the 201 municipalities.

∗( ^∗^∗)( ^∗^∗^∗) indicates statistical signiﬁcance at the 10(5)(1) percent level.

(9)

centagepointsforwelfareclaimsand+2.2percentagepointsforhigh schoolcompletion.

Giventhatconditionalitybecomesmorerelevant aswe moveup- wardsinthedistributionofpredictedwelfarepropensities ̂𝑝𝑖andthat itsimpactsarenegligibleinthefirstquartile,wecanexploitthisprop- ertydirectlyasafoundationforatripledifference(3D)identification strategy. Wecanthenalsoallow forothersourcesof geographically differentiatedcalendartimedevelopmentsbysubstitutingmunicipality- by-yearfixedeffectsfortheseparatemunicipalityandyearfixedeffects usedinEq.(2).Let𝑞=1,...,4denotequartileinthê𝑝_𝑖distribution.The 3D-estimatoristhenbasedon aregressionequationofthefollowing form:

𝑦_𝑖𝑡=𝐱^′_𝐢𝛃+𝜓_𝑚𝑡+

∑4

𝑞=2𝐼_𝑞(𝜆_𝑞𝑚+𝜎_𝑞𝑡+𝜌_𝑞𝑢_𝑚𝑡+𝜃_𝑞𝐶_𝑚𝑡)+𝑣_𝑖𝑡, (3) whereI_qisanindicatorforquartileq,𝜓mtaremunicipality-by-yearfixed effects,and(𝜆qm,𝜎qt)areadditionalmunicipalityandyearfixedeffects relevantforpersonsin ̂𝑝-quartileq(q=2,3,4).Withthisapproach,we estimatetheintentiontotreat(ITT)effectsinquartiles2–4asthe“extra” differenceindifferencethatarisesineachofthesequartilescomparedto thefirstquartile.TheresultsarereportedinTable6,Column(6).Given thattheunrestrictedestimatedeffectinthefirstquartilewascloseto zeroanyway(seecolumn(2)),itisnosurprisethattheyaresimilarto theseparatelyestimatedeffectsreportedinColumns(3)–(5).

Whatalltheestimatespresentedsofarhaveincommonisthatthey measuretheintentiontotreateffectonagroupofpersonsforwhich thetreatmentinquestionmayormaynotberelevant.Themagnitudes ofsucheffectsdependontwofactors:i)thefractionofpersonsactually exposedtothetreatment(inourcase,thefractionofpersonswhowould claimsocialassistanceinatleastoneoftheregimes)andii)theaverage sizeoftheeffectforthese persons.Itwouldclearlybeof interestto disentanglethesetwofactorsempirically,andthusarriveatestimates thatcanbeinterpretedassomethingakintotheaveragetreatmenteffect onthetreated(ATET).Sincewedonotobservetheindividualtreatments inourcase(theeffectoperatesthroughbothpersonsactuallyclaiming socialassistance,andpersonspotentiallyclaiming),wecannotuse a standardinstrumentalvariablesapproach, asin,e.g., Markussenand Røed(2016).Whatwecando,however,istomeasuretheestimated effectsrelativetotheindividualpredictedwelfarepropensityindicators

̂𝑝_𝑖.Thisway,wecanestimatetheaverageeffectofconditionalityrelative tothefractionofadolescentsactuallyexposedtothetreatment;i.e.,the effectobtainedperaffectedperson.Withsomeabuseoflanguage,we refertothisastheATET.

Theregressionmodelcanbewrittenas

𝑦_𝑖𝑡=𝐱^′_𝐢𝛃+𝜓_𝑚𝑡+𝜆_𝑝𝑚𝐼_𝑚̂𝑝_𝑖+𝜎_𝑝𝑡𝐼_𝑡̂𝑝_𝑖+𝜌_𝑝̂𝑝_𝑖𝑢_𝑚𝑡+𝜃_𝑝̂𝑝_𝑖𝐶_𝑚𝑡+𝑣_𝑖𝑡, (4) where(I_m,I_t)areindicatorvariablesformunicipalityandyear,respectively.Hence,inthismodel,wecontrolforcommonmunicipality-by- yearfixedeffectsaswellasseparateeffectsofsocialassistancepropen- sity ̂𝑝_𝑖foreachmunicipalityandforeachyear. Hence,itisonlythe

“extra” associationbetween̂𝑝_𝑖andoutcomesthatshowupintreatment municipalitiesaftertheintroductionofconditionalitythatidentiﬁesthe causalimpact(ATET)𝜃p.

TheresultanttreatmenteﬀectsareprovidedinTable6,Column(7).

Takenatfacevalue,theresultsindicatethatforayouthwhowouldhave receivedsocial assistancewithcertaintyintheabsenceoftreatment, theintroductionofconditionalityreducedtheclaimprobabilityby20 percentagepointsandincreasetheschoolcompletionprobabilityby17 percentagepoints.

3.3. Mechanisms

The literature on treatment eﬀects makes a distinction between

“regimeeffects” (exante)and“participationeffects” (expost);see,e.g., Arnietal.(2015).Inourcase,apotentialexanteeffectisthatcondi- tionalitymakesalifeonwelfarelessattractive,inwhichcaseweexpect

lowerclaimpropensities,possiblyincombinationwithhigherratesof highschoolcompletion.AsexplainedinSection2,thelegislationthat makesparents’ economicallyresponsiblefortheiradultoffspringaslong astheyareenrolledinhighschoolmayrepresentaperverseincentive forsomeadolescentstoquitschoolprematurely;andbymakingsocial assistancelessattractive,conditionalitymayservetooffsetthatincen- tive.Apotentialexposteffectisthattheconditionsimposedonactual welfare claimantscontributetoamorestructureddailylife, possibly includingvaluableworkexperienceortraining.Thiscouldinturnef- fecthighschoolcompletionpositivelyifitsubsequentlyinspiresare- turn tohighschool,ornegativelyiftherequiredactivitiessubstitute forregulareducation.Unfortunately,ourstatisticalapproachisnotde- signedtomakeacleandistinctionbetweenthesetwomechanisms,given thatanyregimeeffectswillendogenouslyalterthecompositionofac- tualclaimants,andhencemakeitdifficulttodisentangleselectionand causalityinthisgroup.Whatwecando,however,istoexaminewhether thefavorableschoolcompletioneffectsareassociatedwithmoreorless socialassistanceuptake.

InTable7wereportestimatesforoutcomesconstitutedbyallpossi- blecombinationsofwelfareuptakeandhighschoolcompletion,based onourbaselineATETmodelinEq.(4).Thecoefficientsreportedinthis tablerelatestotheATET-effectsreportedinTable6,suchthatthecoeffi- cientsinTable7’sColumns(1)and(2)adduptothecoefficientinTable 6’sColumn7,panelA,whereasthecoefficientsinTable7’sColumns (1)and(3) adduptothecoefficientinTable6’sColumn7,panelB.

What theseestimates essentiallyshow isthatthefavorableeﬀecton highschoolcompletionmaterializesincombinationwiththeabsence ofwelfareuptakeatage21;seeColumn(3).Thissuggeststhatexante regimeeﬀectsmusthavebeenimportant;i.e.,thatthehigherpropen- sitytocompletehighschoolinsomesensesubstitutedfor,ratherthan accompanied,welfareclaims.⁴Theprobabilityofcombiningwelfareup- takewithhighschoolcompletionactuallydeclined;seeColumn(1).This istheoppositeofwhatwewouldexpecttoseeifstricterconditionality increasedhighschoolcompletionprimarilybyrequiringwelfarerecip- ientstogobacktoschoolasapreconditionforcontinuedreceipt.

AsexplainedinSection2,theconditionalityreformsinvolvedanum- berofdifferentelements;conf.Table3.Whilealmostallofthereforms involvedsome formof activationrequirements,their actualcontents differedandtheyweretovaryingdegreescombinedwithrequirements regardingpersonaleconomyand/orhealthexamination.Whileitisnot sufficientvariationinthedataeithertoevaluateeachconditionsep- arately ortoevaluateall thedifferent combinations, we have made an attemptto evaluatethepartialimpacts of thethree maincondi- tioningtypes.Table8reportstheresultantaveragetreatmenteffects (ATET).Thepointestimatesbasedonthisspecificationsuggestthatthe activation-andwork-relatedrequirementsandtherequirementthatre- cipients undertakeahealth examinationarethemost importantand havecomparableeffects,whileconditionsregardingthepersonalecon- omyareofminorimportanceatthemargin.However,thehighcorre- lationbetweenthedifferentconditioningtypesimpliesthattheeffects areestimatedwithgreatstatisticaluncertainty.

Howdemandingwasthepolicyimplementationfortheoﬃces?Data on thebudgetsofthelocalinsuranceoﬃcesshowthatoperatingex- pensesrelatedtowelfaredecreasedbothinthetreatmentyearandlater.

Thissuggeststhatthetreatmenteﬀectofareducedcaseloadmorethan madeupforsome oftheconditionsrequiringhigherexpensesatthe oﬃce.Thefactthattherearealsosavingsrelatedtoareducednumber

4Given this finding, it would have been interesting to examine whether the effect on high school completion is larger for adolescents with parents who have sufficient economic resources to be made responsible for their adult kids as long as they are registered in school. However, it is difficult to disentangle this mechanism from the fact that the propensity to claim social assistance in the first place is an order of magnitude higher for offspring with poor parents. Hence, by, say, interacting the causal effects of interest with indicators for poor parents, this interaction would also capture a much higher exposure to treatment.