How Changes in Financial Incentive Affect the Duration of Unemployment

The Review of Economic Studies, Ltd.

How Changes in Financial Incentives Affect the Duration of Unemployment Author(s): Rafael Lalive, Jan van Ours, Josef Zweimüller Reviewed work(s): Source: The Review of Economic Studies, Vol. 73, No. 4 (Oct., 2006), pp. 1009-1038 Published by: Oxford University Press Stable URL: http://www.jstor.org/stable/4123257 . Accessed: 14/01/2012 00:12
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Oxford University Press and The Review of Economic Studies, Ltd. are collaborating with JSTOR to digitize, preserve and extend access to The Review of Economic Studies.

http://www.jstor.org

Review of Economic Studies (2006) 73, 1009-1038 @ 2006 The Review of Economic Studies Limited

0034-6527/06/00391009$02.00

How

Changes
Affect of

in

Financial

Incentives

the

Duration

Unemployment
RAFAELLALIVE
Universityof Zurich

JAN VAN OURS
TilburgUniversity

and JOSEFZWEIMULLER
Universityof Zurich
First version received October2004;final versionaccepted February2006 (Eds.) This paperstudieshow changes in the two key parameters unemploymentinsurance-the benefit of replacementrate (RR) and the potentialbenefit duration(PBD)-affect the durationof unemployment. To identify such an effect we exploit a policy change introducedin 1989 by the Austriangovernment, which affected various unemployed workers differently: a first group experienced an increase in RR; a second group experienced an extension of PBD; a third group experiencedboth a higher RR and a We longerPBD; and a fourthgroupexperiencedno change in the policy parameters. find thatunemployed workersreactto the disincentivesby an increasein unemployment duration,and our empiricalresults are consistent with the predictionsof job searchtheory.We use our parameter estimates to split up the total costs to unemploymentinsurancefunds into costs due to changes in the unemploymentinsurancesystem with unchangedbehaviourand costs due to behaviouralresponses of unemployedworkers.Our results indicatethatcosts due to behaviouralresponsesare substantial.

1. INTRODUCTION This paperstudies how changes that make unemploymentsystems more generousaffect the durationof unemployment. Most unemploymentinsurancesystems are characterized two major by the parameters: earningsreplacementrate (replacementrate RR), that is the level of unemployment benefits in relationto expected earnings;and the maximumdurationthat an unemployed workercan draw such benefits (potentialbenefit duration,PBD). A considerabletheoreticalliteraturehas shown that a more generous unemploymentinsurancesystem (througha higher RR and/or a longer PBD) will reduce the optimaljob search effort of an unemployedworker and hence resultin longer unemploymentduration.Moreover,these theories also offer sharppredictions on how changes in the key parametersin unemploymentinsurancewill affect the uneman ployment exit rate. In particular, increase in RR should have its strongesteffect early in the unemploymentspell, whereas an increase in PBD will have its strongesteffect aroundthe time when benefitsexpire, creating"spikes"in the unemploymentexit rate. The present paper identifies the causal effect of benefit duration on the willingness of individualsto acceptjobs using a policy change that took place in Austriain 1989. The policy
1009

1010

REVIEWOF ECONOMICSTUDIES

affected various unemployedworkersdifferently: a first group experiencedan increase in RR; a second group experiencedan extension of PBD; a thirdgroup experiencedboth a higher RR and a longerPBD; and a fourthgroupexperiencedno change in the policy parameters. More prebenefits(andhence RR) were increasedby about 15%for workersearning cisely, unemployment below a certainthresholdwhereasfor workersabove this thresholdthe RR remainedunchanged. The size of the increase in PBD depended on age and experience: For workersbelow age 40 and/orfor workerswith little previouswork experiencePBD remainedunchanged.For workers with high previous work experience PBD increased,respectively,from 30 to 39 weeks for the age group 40-49; and from 30 to 52 weeks for workersaged 50 and older. Hence, this policy change providesa nice empiricaldesign, because it has elements of a "natural" experiment. the Furthermore, policy change took place in 1989 which, in Austria, was quite a stable macroeconomic environment.This implies that our study is less subject to endogenous policy bias than other studies. Endogenouspolicy bias arises when more generous unemployment insurancerules are implementedin anticipationof a deteriorating labourmarket.Such a policy bias has been found importantin several recent studies (Card and Levine, 2000; Lalive and Zweimiiller,2004a). To assess the effect of these changes in unemploymentinsurancerules, we use a large and informativedata-setthatallows us to traceworkers'unemployment historiesover an extended period of time. We compare the entire unemploymentinflow that took place two years before the policy changeto the entireinflow two years afterthe change.This leaves us with a ratherlarge data-setand allows us to estimatethe interestingpolicy parameters quite precisely. We find thatboth the increasein RR as well as the extension of PBD significantlyincreases the durationof unemployment. line with theoreticalpredictions,we findthatmost of the effect In takes place early in the unemploymentspell in the case of the RR increaseand aroundthe dates when benefits expired in the case of the PBD extension. Furthermore, sensitivity analysis our shows that, despite the clear heterogeneityin treatedpopulations,results are ratherrobust. In particular,we find that the effect of the RR increase, the effect of PBD extensions, and the additiveeffect of simultaneousRR and PBD changes are independent variationsin the control of effect. While the size of the effects are quite robust,there groupsused to identify the treatment are two importantexceptions. First, older workersreact more strongly to PBD extensions than prime-age workers, and second, the additive effect of a simultaneouschange in RR and PBD is larger for older workers than for prime-age workers. In a theoreticalcontext, this stronger reaction of older workersto changes in unemploymentinsurancerules can be rationalizedby theirweakerlabour-market environment position and/orby incentivescreatedby the institutional (earlyretirement). Reliable empiricalevidence on how the two key unemployment insuranceparameters affect the searchbehaviourof unemployedworkersis crucial for designing appropriate policies. Most previousempiricalstudiesidentify such incentiveeffects from "exogenous" variation,separately, either in a change in benefit levels or a change in the PBD. While the ceteris paribus effects of are changes in such parameters clearly interesting se, it does not allow to addresspotentially per of unemploymentinsurance.To assess the relative important questionsof the appropriate design of we estimatesto split up the total costs importance the two policy parameters use ourparameter to unemploymentinsurancefunds into costs due to changes in the unemploymentinsurancesystem with unchangedbehaviourand costs due to behaviouralresponses of unemployedworkers. Ourresults indicate that costs due to behaviouralresponses are modest for increases in RR, but substantialfor increasesin PBD. From these results, a simple policy conclusion can be derived: If policy makerswant to influenceincentives, the potentialdurationof unemploymentis a more effective tool thanthe level of unemploymentbenefits. The paperis organizedas follows. The following section discusses the relevanttheoretical argumentsand also provides a discussion of previous empiricalevidence regardingthe effects
? 2006 The Review of Economic Studies Limited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES

1011

of the two key parameters unemploymentinsurance.Section 3 gives the relevantinstitutional of backgroundon the changes to unemploymentinsurancein Austriathat are used to identify their effects on unemployment durations.Section 4 providesa firstdescriptiveanalysisof the effects of insuranceon unemployment duration. Section 5 presentsthe econometricanalysis unemployment regardingthe effects on the unemploymentexit rate,and Section 6 concludes with a summaryof the resultsand drawssome policy conclusions.

2. THEORYAND PREVIOUSSTUDIES 2.1. Theory We assume thatan unemployedworkeris entitledto unemployment benefitsfor a fixed duration. After benefits expire he or she is entitled to unemploymentassistanceof infinitedurationthat is lower thanunemployment benefits.We are interestedto see how job searchbehaviouris affected the presence of a fixed benefit duration.1The argumentsin the optimizationproblem are by expected costs and benefits, that is, the value of being unemployed comparedto the value of having a job. The value of unemploymentis determinedby the level of the unemployment benefits, the searchcosts, the situationin the labourmarket(i.e. the way search intensity translates intojob offers), the expected gain from acceptinga job and the risk of not findinga job before unemploymentbenefits expire. While an increasein searchintensityincreases searchcosts it also increases the probabilityof finding a job and reduces the probabilityof runningout of benefits.Optimalsearchintensitybalancesmarginalcosts and benefits of search. unemployment At the start of the unemploymentspell search intensity is low because the probabilityof finding a job before unemploymentbenefits expire is large anyway.With increasingdurationof the unemployment riskof runningout of benefitsincreases,whichinducesan unemployedworker to increase search intensity. Once an unemployed workerruns out of unemploymentbenefits, searchintensityremainsconstantbecausethe workerfaces a stationary environment. exit rate The from unemploymentis then determinedby search costs, the level of unemploymentassistance, and the situationin the labourmarket.2 Let us now consider an increase in the RR. A higher RR implies that, at the start of the unemploymentspell, a worker will search less intensively than before, as the costs of being unemployed are lower. As benefit expirationcomes closer, the risk of runningout of the now higherbenefitsinducesthe unemployedworkerto increasesearchintensitymore strongly.Search effect".The value of intensityeventuallybecomes even largerthanbefore due to an "entitlement a new job increasesbecause losing this job is associatedwith less damagedue to the higherRR. These predictionsare summarized the upperpanelin Figure 1, which drawsthe unemployment in exit hazardagainstthe elapseddurationof unemployment differentRR (dashedfor the higher for RR).3The ratiobetweenthe exit ratein the new andthe old system is initiallyless thanunity,will
1. See, for example, Mortensen(1977, 1990), Burdett(1979), and Van den Berg (1990). The unemployedcould also choose both an optimal search intensity and an optimal reservationwage, but this would not essentially change behaviourand is thereforeignored here. Note that, with a fixed reservationwage, there is a one to one relationship between searchintensityandjob-findingrate. 2. Mortensen(1977) arguesthatsearchintensityshifts down afterbenefitexpirationif non-market time andmarket goods used in householdproductionare substitutes,or it goes up if time and marketgoods are complementsin household production. 3. This figure is based on a discrete time model in which job seekers receive unemploymentbenefits equal to 40% of their previous wage for six successive (five-week) periods and no benefits thereafter. Job seekers choose how intensivelyto searchfor jobs. The unemploymentexit hazardis the productof the searchintensityand the arrivalrate of job offers, takento be equal to 0.25 per period.The RR change increasesthe RR from 40% to 45%, whereasthe benefit durationchange increasesPBD from 6 to 8 periods.Note thatthe increasein PBD with 30% and the increasein RR with about 10%mimic the Austrianchanges in the UI system.

? 2006 The Review of Economic Studies Limited

1012
0 .20 S•-1 0 89 .1 IN 0.17 0 .16
0.15

REVIEWOF ECONOMICSTUDIES
Increasing the RR
O 1-04 1-02 1 .00 0 -:96 o 0.92 3 4 5 6 7 8 9 10 11 12
0.90

Increasing the RR

- - -- -- -

1 I_" 2

3

4

5

6

7

8

9

10 11

12

Ca N

a

0-20

Increasing the PBD

CD
1-02

Increasing the PBD

0.19102 0-18
0"6 -

1.04

100oo N 098 •
0-96 . 0-94 O 0.92 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12

0.15 __0-90
1 32

0-20
S019

Increasing both RR and PBD

C)
S1-04

Increasing both RR and PBD

--.•
c

01 0.18

aCIS 1.02 N 0.98 1.00_"
0C96

0-17-c

S0.160 0 5 .1

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

0-94 0 0.92 C 0$

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

1 FIGURE Changeto replacementrate (RR) and potentialbenefitsduration(PBD) and the unemploymentexit hazard

increasewith increasingunemploymentduration,and will eventuallyexceed unity when benefits expire.4 In contrast, an extension of PBD entails only small immediate disincentive effects. The reason is that, at the beginning of the spell, the extension of PBD does not strongly affect the risk of runningout of benefits. However,with increasingunemploymentdurationa longer PBD makes a difference. When unemploymentdurationis equal to PBD under the old system, the difference in search intensities of the two systems is largest. At that date, searchintensity is at its maximumunderthe old system, while it is still comparablylow underthe new system. As unemploymentdurationcomes closer to PBD underthe new system, search intensity increases strongly.Eventually,it becomes even largerthan underthe old system, again due to the entitlement effect. As illustratedin the middle panel of Figure 1, most of the action will take place just before benefit exhaustionin the old system until benefit exhaustionin the new system. The ratio between the exit rate in the new and the old system will not only be less thanunity,but also will be decreasinguntil benefitsexpire in the old system. Afterwards,the ratioincreasesstronglyand eventuallyexceeds unity when benefitsexpire underthe new system. Finally, let us consider a simultaneousincrease in RR and PBD. Now, an unemployed workeris facing not only a higher benefit over his initial PBD and a longer PBD with his initial benefitsbut he or she is also entitledto higher benefits over a longer PBD. Suppose PBD is
4. Note that the magnitudeof the entitlementeffect depends on the expected durationof employmentcompared to the expected durationof unemployment.If employmentdurationsare long comparedto unemploymentdurations,the entitlementeffect is boundto be small. @ 2006 The Review of Economic StudiesLimited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES

1013

increasedfrom Eo to El and the RR increase raises benefits from b0 to bl. To get the intuition of how these changes affect incentivesconsider a workerwho exhausts all his benefit claims. If this workeris eligible to the RR increase only, total benefit paymentsincreaseby AbEo. For a workereligible to the PBD increase only, total benefit payments increase by boAE. A worker who is eligible to both an increase in RR and an increase in PBD, total benefits increase by AbEo + boAE + AbAE. The latter interactioneffect causes an additionaldisincentive effect and cannot be ignored when changes in b and E are substantial.Therefore,we expect that the disincentiveeffects of a simultaneouschange in RR and PBD will cause an additionaldecline of the job-findingrate that is largerthan the sum of two separatechanges. Relativeexit rates from a joint change are lower until benefits have expired in the old system increase more strongly thereafter reach a higherlevel when benefitshave expiredin the new system. and The abovediscussionhas obvious empiricalimplicationsfor transitions fromunemployment to employment.First,providedthatentitlementeffects are negligible and/ormost unemployment exits take place before benefits expire, increases in RR and extensions in PBD should lead to a reductionin job search effort and hence to longer unemploymentdurations.Second, increases of RR should trigger strong behaviouralresponses early in the unemploymentspell whereas extensions of PBD should lead to strongresponsesaroundthe dates when benefitsexpire. Third, for individualsentitled to a simultaneousincrease in RR and PBD, we expect an increase in unemploymentdurationslargerthan the sum of the increases from two isolated changes in the policy parameters. Clearly,the exact size and natureof the relationshipbetween the unemploymentinsurance and of parameters the unemploymentdurationdependson the parameters the model. In particudifferencesin labour-market and the institutionalsettingmay explain why one group lar, position of workersreacts more stronglythan another.For instance, older workersreceive fewer job offers because employers often preferto hire young or prime-age workersto fill their vacancies. and/orhave a greaterchance to take up Moreover,older workersare closer to (regular)retirement retirementbenefits. This implies that older workersbenefit less from a given searcheffort early and may react strongerto disincentives.This may be particularly relevantfor large extensions of PBD. In that case, the consequences of not finding a job are less severe because the period between benefit expirationand the startof (early)retirement may be reducedsubstantially. 2.2. Previousempiricalstudies SeveralU.S. studiesestimatethe effects on the exit ratefromunemployment variationsin PBD of thattakeplace duringrecessions.5Earlystudies,includingMoffittand Nicholson (1982), Moffitt (1985), andGrossman(1989) find significantlynegativeincentiveeffects. KatzandMeyer (1990) andMeyer(1990) show thatthe exit ratefromunemployment rises sharplyjust beforebenefitsare exhausted.Such spikes are absentfor non-recipients.More recentworkby Addison andPortugal (2004) confirmsthese findings.6 A common objection againstthese studies is policy endogeneity.Benefits are typically extendedin anticipationof a worse labourmarketfor the eligible workers.Cardand Levine (2000) condition and exploit variationin benefit durationthat occurredindependentlyof labour-market show thatpolicy bias is substantial.Lalive and Zweimiuller (2004a,b) show similarevidence for
5. Fredrikssonand Holmlund(2003) give a recent overview of empiricalresearchrelatedto incentives in unemploymentinsurance.See Ham and Rea (1987) and Greenand Riddell (1993, 1997) for studies thatfocus on Canada. 6. Note thatthereis no theoreticalexplanationfor the existence of end-of-benefitspikes. It could be thatthe spikes have to do with strategictiming of the job startingdate, that is, workershave already found a job, but they postpone startingto work until theirbenefitsare close to expiration.Cardand Levine (2000) point at the possibility thatthereis an implicit contractbetween the unemployedworkerand his previousemployerto be rehiredjust before benefitsexpire.

? 2006 The Review of Economic StudiesLimited

1014

REVIEWOF ECONOMICSTUDIES

the Austrianlabourmarket.Evidence on the effect of PBD in Europeanstudies is mixed. Hunt disincentiveeffects of extendedbenefitentitlementperiodsfor Germany. (1995) finds substantial and Carling,Edin, Harkman Holmlund(1996) find a big increasein the outflowfrom unemployment to labour-market programmeswhereas the increase in the exit rate to employmentis sub(1998) uses Austriandata and finds significantbenefit duration stantiallysmaller.Winter-Ebmer effects for males, butnot for females. Rod andZhang(2003) findfor Norwegianunemployedthat the exit rateout of unemployment increasessharplyin the monthsjust priorto benefitexhaustion where the effect is largerfor females thanfor males. Puhani(2000) finds thatreductionsin PBD in Polanddid not have a significanteffect on the durationof unemploymentwhereasAdamchik (1999) finds a strongincreasein re-employment probabilitiesaroundbenefitexpiration.VanOurs and Vodopivec(2006) studyingPBD reductionsin Slovenia find both strongeffects on the exit rate out of unemploymentand substantialspikes aroundbenefitexhaustion. To estimate the effects of RR on unemploymentdurations,the early literaturefocuses on differencesin benefitreplacement ratiosbetweenindividuals.Estimatedelasticitiesof unemployment durationwith respect to benefit levels range from 0.1 to 1-0 (Atkinsonand Micklewright, is 1991). This literature problematicdue to the possibility of unobservedheterogeneitydistorting identificationin cross-sectionaldata.More recent Europeanstudies have focused on the impact of policy changes affecting benefit levels. Carling, Holmlund and Vejsiu (2001), studying the effects of a reductionin the RR from 80%to 75% in Sweden in 1995, find thatthis policy change increasedthejob-findingrateby roughly 10%,implying an elasticity of around1.7. Bennmarker, Carlingand Holmlund(2004), studyingchanges in the Swedish system in the early 1990's, finda smallerelasticityof around0.6. Rod andZhang(2003) estimateelasticitiesfor Norwayof around 0-95 for males and around0.35 for females. 3. UNEMPLOYMENT INSURANCEAND THE AUSTRIANLABOUR MARKET 3.1. Thesystembefore thepolicy change Like in a numberof other countries,the Austrianunemploymentinsurancesystem is characterized by a limited period over which unemployedindividualscan draw "regular" unemployment benefits (UB). UB depend on previousearningsand, comparedto other Europeancountries,the replacementratio (UB relative to gross monthly earnings) is ratherlow. Figure 2 shows that, before August 1989, the replacementratio declined strongly from a maximum of about 63% (monthly income is below 2210 ATS) to 41% in the income range of between 3000 and 5000 ATS previous monthly earnings.7The benefit replacementratio then stays just below 41% for incomes up to the cap of 27,430 ATS previousmonthlyearnings.Individualsearningmore than 27,430 ATS get UB of 11,233 ATS per monthirrespectiveof their income. Thus, the benefitRR decreasesmonotonicallyin previousmonthlyincome for the high-incomegroup. On top of UB, family allowancesarepaid. UB paymentsare not taxed andnot meanstested, Voluntary quittersand workersdischargedfor misconductcannotclaim benefits duringthe first four weeks of a new unemploymentspell. UB recipients are expected to search actively for a new job that should be within the scope of the claimant'squalifications,at least duringthe first monthsof the unemploymentspell. Non-compliancewith the eligibility rules is subjectto benefit sanctionsthatcan lead to the withdrawalof benefitsfor up to four weeks. Before August 1989, an unemployedpersoncould drawregularUB for a maximumperiod of 20 weeks providedthathe or she hadpaidunemployment for insurancecontributions at least 52 insurance(UI) weeks withinthe last two years.Individuals to who hadcontributed unemployment
7. The medianmonthlyincome was about 16,400 ATS in the unemployment inflow from regularjobs in 1988.

? 2006 The Review of Economic StudiesLimited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES
Gross replacementrate

1015

0.6

-

1989 Before August
After August 1989

05

04

0.3

0

5000

15,000 10,000 Monthly income (AS)

20,000

25,000

364/1989. no. Source: Austrian federal (Bundesgessetzblitter)594/1983, laws 2 FIGURE The change to the RR in August 1989

for at least 156 weeks in the last five years were eligible for 30 weeks of regularunemployment benefits.8 Once the period of regularunemploymentbenefits has expired, individualscan apply for As "transfer paymentsfor those in need"("Notstandshilfe").9 the name indicates,these transfers are means tested, and the job seeker is considered eligible only if she or he faces economic difficulties.These paymentsdependon the income and wealth situationof otherfamily members and close relatives and may, in principle, last for an indefinitetime period. These transfersare are grantedfor successive periods of 39 weeks after which eligibility requirements recurrently checked. These post-UB transfersare lower than UB and can be at most 92% of UB. In 1990, the median post-UB transferpayment was about 70% of the median UB. Note, however, that individualswho are eligible for such transfersmay not be comparableto individualswho collect UB because not all individualswho exhaustUB pass the means test. The system described above applies to all of Austria except for regions with a dominant steel industry.In these steel regions, a regional extended benefit programmewas introducedin June 1988, that entails a dramatically differentUI system for workersaged 50 or older.In order to focus attentionon the policy change in August 1989, we focus on the non-steel regions that were never entitled to the regional extended benefit programme.See Winter-Ebmer (1998) and Lalive and Zweimiiller(2004a,b) for analyses of this programme. 3.2. The 1989 changes in policy parameters As of 1 August, 1989 the Austriangovernmentenacted a series of importantchanges to the The unemploymentinsurancerules (Arbeitslosenversicherungsgesetz). first change was that the durationof UB payments became dependentnot only on previous contributions,but potential
In 8. UB durationwas 12 weeks for job seekers who did not meet eitherpreviouscontribution requirement. order for to guarantee minimumdegreeof homogeneity,this paperfocuses on workerswho have contributed at least 52 weeks a to UI in the two years priorto unemployment. Thus, all workersare eligible for at least 20 weeks of UB before the policy change. 9. This implies thatjob seekers who do not meet UB eligibility criteriacan apply at the beginningof their spell. @ 2006 The Review of Economic Studies Limited

1016

REVIEWOF ECONOMICSTUDIES

also on age at the beginningof the unemploymentspell. Benefitdurationfor the age group40-49 was increased to 39 weeks if the unemployed has been employed 312 weeks of employment within the last 10 years priorto the currentspell. For the age group 50 and older, UB duration was increasedto 52 weeks if the unemployedhas been employed for at least 468 weeks within the last 15 years. The second importantchange was that the RR was increasedfrom about 41% to roughly 47% in the previous income bracketbetween 5000 ATS and 10,000 ATS (Figure 2). After the income thresholdof 10,000 ATShas been crossed,the RR is fadedout quicklyto reachthe former level of 41% at 12,610 ATS previousmonthlyincome. These changes affect all job seekersas of 1 August, 1989. In the period that we study, there was a furtherslight increase to the benefit RR. As of 1 June, 1990, the Austriangovernmentenacted an increase to the RR for those with previous income exceeding 12,610 but below 27,000 ATS. This change essentially ensuresthat the RR is faded out from a level of 47% to a level of 41% over the income range 10,000 ATS to 27,000 ATS insteadof the rangefrom 10,000 ATS to 12,610 ATS. In this paper,we focus on the effect of the increasein PBD andthe increasein RR in August 1989. Clearly, the change in policy parametersis different for differentgroups related to the threeeligibility determinants: age, experience,and earnings.Earningsdeterminewhetherthe RR workerswith low earningsget an increase in RR. Age and experiencedetermine goes up-only whetherthe PBD goes up-only workersfrom age 40 onwardswith a high workingexperience the get an increasein PBD. Furthermore, size of the increasedependson the age of the individual. These policy changes createda nice empiricaldesign thatcan be exploitedfor empiricalresearch as it has elements of a "natural experiment". We do not estimate the effect of the increase in RR in July 1990, for two reasons. First, this policy change entails a much weaker change to the benefit RR than the policy change in August 1989.10Second, the July 1990 change to the RR occurs rathershortly after the August 1989 change. This implies that the unemploymentexit rate in the period before the July 1990 change, but afterthe August 1989 change is not identifiedin the period48 weeks afterentryand interestin the presentanalysis.The fact that beyond. This time periodis, however,of substantial we do not account for the July 1990 change implies that the estimated treatmenteffects refer of to treatmentsrelativeto a control treatmentthat does not leave unaffectedall parameters the unemploymentinsurance system. Thus, not accounting for the July 1990 policy changes the of interpretation the effects we reportbut it does not affect the internalvalidity of these effects. Moreover,since we evaluate the effects of these policy changes relative to a slight increase in benefits,our resultsgive a lower boundon the effects relativeto no change.Moreover,in Section 5.3 we will undertake sensitivitytests thataccountfor possible effects of the 1990 policy change. 3.3. Thesituationon the Austrianlabourmarket1987-1991 Before we go into the detailsof dataand statisticalanalysis, it is instructive look at the situation to on the Austrianlabour marketduring the period 1987-1991. This is the period on which the empiricalanalysis below will be concentrated. Table 1 shows thatin 1987 the economy was at the end of a recession and startedto improve. Real GDP growth was 1.7% in 1987 and then startedto grow to as much as 4.7% in 1990. The favourablesituationof the business cycle led to strong employmentgrowththroughoutthe period under consideration.However, it did not show up in decreasing unemploymentrates.
10. The averageincreasein the benefit RR is almost 6 percentagepoints for the August 1989 change. In contrast, the RR increasesby a mere 1.3 percentagepoints due to the second policy change (Table3).

? 2006 The Review of Economic Studies Limited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES
TABLE 1 TheAustrianlabour market1987-1991

1017

Real GDP growth 1987 1988 1989 1990 1991 1-7 3.2 4.2 4.7 3.3

Employmentgrowth 0-6 0.7 1.4 2.8 2.8

Unemploymentrate 5.6 5-3 5.0 5.4 5.8

Source: StatisticsAustria.

The reason was primarilya strong increase in labour supply (a strong increase in immigration and rising female labourforce participation). That is why unemploymentrose slightly despite a strongemploymentgrowth. It is worth noting that this situationis favourablefor our empirical strategy.Employment growth during the treatmentperiod was stronger than before. Hence it is unlikely that our resultsfrom a comparisonof the labour-market experiencesof older workersbetween the period labour prior to the policy change to the post-policy period is strongly drivenby a deteriorating market. 4. DESCRIPTIVE ANALYSIS 4.1. Data To assess the impactof changes to financialincentiveson transitionrates out of unemployment, we use longitudinalindividualdata from two differentsources: (i) the Austriansocial security database, which contains detailedinformationon the individuals'employment,unemployment and earningshistory since the year 1972 and some informationon the employerlike region and industryaffiliation;and (ii) the Austrianunemployment registerfrom which we get information on the relevantsocio-economic characteristics. Fromthese datawe extracta samplethatcontains all unemploymententrantsin the period between 1 August, 1987 (two years before the policy change) and 31 July, 1991 (two years afterthe policy change). We concentrateon job seekersin the age bracket35-54, who have at least 52 weeks withinthe last two yearsand with residencein Furthermore, regions thatwere nevereligible for a special regionalextendedbenefitprogramme. in orderto isolate the effects of changesin PBD, we concentrate workerswho eitherfulfil both on or previouscontribution requirements1" neither.We end up with 225,821 unemploymentspells. The mediandurationof unemploymentis 12 weeks. More than 85% of spells end in a job, 14% of spells in a non-job exit destination(long-termsickness, pension, unknown).Since spells are observeduntil May 1999, only 1%of spells are rightcensored. 4.2. Construction groups of Table 2 summarizesthe changes to unemploymentinsurancethat were enactedin August 1989. Eligibility depends on three criteria: previous gross monthly income, age, and previous work experience. Thus, in the data it is possible to distinguish four groups of job seekers. The first group comprisesjob seekers with monthly income exceeding 12,610 ATS who are aged 40 or older with much previouswork experience(6 out of previous 10 years and 9 out of previous 15
11. That is, at least 312 weeks contributionsin the previous 10 years and at least 512 weeks contributionsin the previous 15 years.

? 2006 The Review of Economic Studies Limited

1018

REVIEWOF ECONOMICSTUDIES
TABLE2 Eligibility(e) for change to the relativerate (RR)andfor change to potential benefitduration(PBD) Age Youngerthan40 years Workexperience Low Monthlyincome <12,610 AS > 12,610 AS eRR Control High eRR Control 40 years and older Workexperience Low eRR Control High ePBD-RR ePBD

Notes: Work experience "low" refers to less than 6 out of previous 10 years and less than 9 out of previous 15 yearsworkexperience.Workexperience"high"refersto workedmore than6 out of previous 10 and worked more than 9 out of previous 15 years. RR, replacementrate; PBD, potential benefits duration;ePBD, eligible for increasein potentialbenefit duration;eRR, eligible for increase in benefit RR; ePBD-RR, eligible for both.

years). This groupis eligible for the changeto PBD (the ePBD group).The second groupconsists of job seekerswith income lower than 12,610 ATS andage less than40 yearsor age exceeding 40 years but with little work experience.This groupbenefits from the increasein the RR (the eRR group). The thirdgroup containsjob seekers with income lower than 12,610 ATS and aged 40 or older with much previouswork experience.This groupis affected by both, the increasein the benefitRR andthe increasein PBD (the ePBD-RR group).The fourthgroupcontainsjob seekers with income exceeding 12,610 ATSwho areeitheryoung or havelittle workexperience.Forthose individualsthere was no change in the two centralparametersof the unemploymentinsurance system (control group). The heterogeneityin treatmentis obvious from the existence of four but groups.Thereis not only heterogeneityin termsof the natureof the treatment also in the size of the treatment.The RR increases with about 15%whereas PBD increases either with 30% or 75%. This additionalvariationexpandsthe range over which parameters estimated.Table 3 are reportsthe means of the RR and of PBD together with the numberof spells in the respective groups both before August 1989 and after August 1989. The first row shows that almost all job seekers entering unemploymentbefore August 1989 in the ePBD group were eligible for the maximumdurationof regularbenefits of 30 weeks.12 In contrast,the PBD was almost 43 weeks for spells in the ePBD group after the policy change. Thus, this group experiencedan increaseby 13.5 weeks in PBD. The second row in Table 3 shows that there is a slight increase in the RR. This increase is due to the fact that there is a second policy change in June 1990 affecting the high-income workersin the ePBD group. The thirdrow shows that this group is the largestin size and thatthe numberof spells in the ePBD groupincreases.However,note that this probablyreflectsthe fact thateligibility to the change in the RR dependson previousincome in nominal terms ratherthan the fact that more individualsregister to collect unemployment insurancebecause benefits are more generous. In line with economic growth over the period 1987-1991 (Table 1), the total numberof spells is almost identical before and after the policy change. The second set of rows in Table 3 refers to the eRR group. PBD is lower on average in the RR group than in the PBD group. This reflects the fact that many spells in the RR group do not fulfil the 30-week requirement that at least 60% out of the previousfive years have to be While thereis virtuallyno change to PBD, the RR increases strongly,from 41% spent working. to 47%. This group is smallerthan the first group,reflectingthe fact that median income in the
12. The empiricalanalysisbelow accountsfor the fact thatunemployment insuranceparameters varydue to the fact that the changes enactedin August 1989 pertainto all spells, not just to those startedafter August 1989. @ 2006 The Review of Economic Studies Limited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES
TABLE3

1019

The changes to unemployment insurancein August 1989 Before August 1989 ePBD group PBD (weeks) RR (%) N eRR group PBD (weeks) RR (%) N ePBD-RR group PBD (weeks) RR (%) N Controlgroup PBD (weeks) RR (%) N Total PBD (weeks) RR (%) N 29-5 40-0 48,294 25-1 41-4 17,160 29-0 41-3 11,992 27-4 40.2 33,815 28.1 40.4 111,261 After August 1989 42.5 40-9 51,110 24-6 47.3 15,310 42.6 47.0 9182 26-9 41.5 38,958 34-8 42-5 114,560 Change (after-before) 13.0 0-9 Diff-in-diff (change comparedto "control") 13.5 -0.3

-0.5 5.9

0.0 4.6

13-6 5.7

14.1 4.4

-0-5 1-3

6-7 2.0

Notes: Diff-in-diff, difference-in-difference; PBD, potential benefits duration;RR, replacementrate; ePBD, eligible for increase in potential benefit duration;eRR, eligible for increase in benefit RR; ePBD-RR, eligible for both. Source: Own calculations,based on AustrianSocial SecurityData.

unemploymentinflow is higher than the nominalincome thresholdof 12,610 AS. The decrease in the numberof spells in the group is due to the fact that previous income must be below a nominallyfixed income thresholdin orderto be a memberof the RR group. The third set of rows in Table 3 refers to the ePBD-RR group. This group very much resemblesa combinationof the ePBD and of the eRR groupexhibitinglong PBD as in the ePBD group, and a ratherhigh RR of 41% as in the eRR group. Interestingly,this group experiences an increase in PBD and in the RR in exactly the same magnitudeas both previously discussed groups.This is the smallest group. Again, the numberof spells allocated to this group declines since all individualsmust have earnedless than the nominal income thresholdof 12,610 AS in orderto be allocatedto the ePBD-RR group. The fourthset of rows in Table3 refers to the controlgroup. This second largestgrouphas rather long PBD andrelativelylow RR beforethe policy change.Thereis a slight decreasein PBD over time and a slight increasein the RR (reflectingthe policy change in June 1990) over time. The last column in Table 3 reports difference-in-difference (diff-in-diff) estimates of the effect of the policy change on both parameters the unemploymentcompensationsystem. This of column shows thata corresponding diff-in-diffcalculationfor some outcome,identifiesthe effect of extendingthe PBD by 14 weeks (startingfrom 30 weeks), the effect of increasingthe RR by 4-6 percentagepoints (startingfrom41%), andthe effect of increasingbothPBD by 14 weeks and increasingthe RR by 4.4 percentagepoints. Thus, this design allows for an exhaustiveanalysis of how financialincentivesaffect the durationof unemployment.
@ 2006 The Review of Economic Studies Limited

1020

REVIEWOF ECONOMICSTUDIES
TABLE4 durationinfirst 104 weeks (measuredin weeks) Averageunemployment Before August 1989 ePBD group 16-25 (0-08) 48,294 17-79 (0.12) 17,160 19-01 (0.17) 11,992 15-24 (0-08) 33,815 After August 1989 18-67 (0-09) 51,110 20-03 (0.16) 15,310 23-55 (0.24) 9182 16-52 (0-09) 38,958 Change (after-before) 2-42 (0-12) 2-24 (0.20) 4-53 (0.20) 1.29 (0-13) Diff-in-diff (changecomparedto "control") 1.13 (0-18) 0-96 (0.24) 3.25 (0.24)

eRR group

ePBD-RR group

Controlgroup

Notes: Standard errorsin parentheses. Diff-in-diff,difference-in-difference; replacement RR, rate;PBD, potentialbenefits duration;ePBD, eligible for increase in potentialbenefit duration;eRR, eligible for increasein benefitRR; ePBD-RR, eligible for both. Source: Own calculations,based on AustrianSocial SecurityData.

4.3. Unemployment duration Table4 reportsaverageunemployment durationin the first 104 weeks by programme implementation status and by group.13Let tu denote the realized durationof unemploymentmeasuredin weeks. Unemploymentdurationin the firsttwo years is t04 - min(tu, 104). The first column in Table 4 shows that average unemploymentdurationis longest in the ePBD-RR group and shortest in the control group in spells that startedbefore August 1989. The second column in Table 4 shows average unemploymentdurationafter August 1989. The thirdcolumn in Table4 shows thatunemploymentdurationincreasesin all groups. However,as column4 in Table4 shows, the change in unemployment strongerin groups,which areeligible is for either the change to PBD or RR or both. For instance, unemploymentdurationincreases by 1.13 weeks more strongly in the ePBD group that is eligible for the extension of PBD but not for the increasein RR. There is a slightly weakerincreaseby 0-96 weeks in unemployment durationin the eRR group, which is eligible for the increase in RR but not for the extension of PBD. The strongesteffect is in the group that is eligible for both RR and PBD. In the ePBDRR group, unemploymentdurationincreases by 3.25 weeks more strongly than in the control thereappearsto be an excess increaseof 1-16 weeks (= 3-25 - 1.13 - 0.96) group.Furthermore, in unemploymentdurationin the sense thatthis effect exceeds the sum of the effects reportedin the ePBD and in the eRR groups. Obviously,Table4 providesonly a first crude check on how the policy changes may have affected durations.Interestingly,the results are in line with theoreticalpredictionsdiscussed in Section 2. It is worth noting, however, that the diff-in-diff estimates of Table 4 are based on ratherdifferentgroups. Moreover,unbiased estimates are obtained only if there are no groupspecific trends in unemploymentdurations.To check for potential biases resulting from such group-specifictrends, Section 5.3 below will provide a variety of robustnesstests, including a focus on more narrowlydefined groups and on groups that are just below or just above the
13. Wereport in we the First, willcalculate contribution average unemploymentthefirst104weeksfortworeasons. to thetotalchange a function elapsed as of is duration. Second, unemployment rightcensoring not anissuein the first 104weeks.Moreover, thatresults note to duration verysimilar. are referring totalunemployment

The of Studies Limited @ 2006 Review Economic

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES

1021

trendsshouldbe of minor eligibility threshold.Withinsuch morehomogenousgroupsdifferential importance.14 4.4. Survivorfunctions Job search theory predictsthat financialincentives in UI affect the shape of the unemployment exit hazarddependingon the time to benefitexhaustion.In orderto studythis prediction,it is useful to decompose the effect on averageunemploymentdurationreportedin Table4 as follows. It is well known thatexpectedunemploymentdurationin the firsttwo years is given by E (Tu04)= where S(z) = exp (- f 0O(x)dx)is the survivorfunction,thatis, the probabilitythat fo4 S(z)dz unemploymentspells are longer than z weeks and where O(x) is the unemploymentexit hazard, that is, O(x) = f(x)/S(x) where f(x) is the density of unemploymentspells (Lancaster,1990). This says, for instance,thatthe increasein averageunemployment durationin the ePBD groupby 2.42 weeks is due to the fact thatthe survivorfunctionin the ePBD groupafterAugust 1989 was survivor functionin the ePBD groupbasedon spells thatstartedbehigherthanthe corresponding fore August 1989. Moreover,the differencein these survivorfunctionsintegratesto 2.42 weeks. Thus, analysingthe effect of the policy change on survivorfunctionsallows decomposingthe total changein unemployment duration into contributions this changeas a functionof duration.15 to 3 shows the Kaplan-Meiersurvivorfunctions for the four groups.The top left subFigure figurecontraststhe survivorfunction afterthe policy change with the survivorbefore the policy change in the ePBD group.Clearly,after 15 weeks of elapsed unemploymentduration,the survivor function after the policy change startsto diverge from the correspondingfunction before the policy change. The difference widens until about 40 weeks have elapsed. After this point, the differencebecomes smalleragain and stays constantat a low level after65 weeks of elapsed duration.Thus, extendingthe PBD appearsto create a "lens"that starts 15 weeks before benefit exhaustionin the old system, and that ends about 15 weeks after benefit exhaustiontakes place in the new system. The top right subfigurein Figure3 reportsthe differencein the survivorfunctionsin the eRR group.In contrastto the previousfindings,we note a slight increasein the survivorfunctionthat takesplace almostfrom the startof the unemploymentspell. The differencebetween the survivor functionsbecomes largerafter20 weeks of unemployment durationand again afterweek 30. The survivorfunctionremainsat a higherlevel thanbefore the policy change even in the periodwhen benefitshave been exhausted,afterweek 30. The bottom left subfigurein Figure 3 reportsthe survivorfunction analysis for the ePBDRR group. This figure very clearly combines aspects of both policy changes. On one hand, we note a relativelystrongincreasein the survivorfunctionrightfrom the startof the unemployment spell. Again, there a "lens" startsto appearafter 15 weeks of unemploymentduration,which disappearsonly after65 weeks of unemploymentdurationhave elapsed. The bottom right panel reportsthe survivorfunctions in the "control"group. There is no differenceat all in the survivorfunctionsup to 20 weeks of unemploymentduration.Thereafter, we note a slight upwardshift in the survivorfunction.Thus, in orderto isolate the effects of the changes to the unemployment compensationsystem in August 1989, it is necessaryto net out the over time from the raw effects on the survivorfunctions in the previousthree change occurring subfigures.
14. A furtherreason why results in Table4 could be biased are substitutioneffects. As individualsfrom the treatment group are less likely to acceptjobs, individualsfrom the control group may be able to leave unemploymentmore quickly. 15. Note that elapsed unemploymentdurationis not time to benefit exhaustion. However, recall that in the old system, only two levels of the PBD prevailed,that is, 20 and 30 weeks. Thus, elapsed durationis very closely relatedto time to benefitexhaustion.

? 2006 The Review of Economic Studies Limited

1022

REVIEWOF ECONOMICSTUDIES
PBD increase 6
a 00

RR increase

6adP

in

6

a
>

>

t--6

VBeforei-After

----- Before
6

After

0

- - - - -

- - -- -

0

10 20

30 40

50

60

70

80 90 100

- -- -

-- - --

--6-

- - - -

-

0

10 20

30 40
-----

50

60 70

-- -

80

90 100

Unemploymentduration(weeks)
-After ...Before

Unemploymentduration(weeks)
Before After

RR and PBD increase

No treatment

~l

t.,

-

-----Before -After-...Before

-After

1989. Notes: Before, before 1989; after, August spellstarts August spellstarts Data. Source: Owncalculations, on Austrian based SocialSecurity FIGURE 3 Kaplan-Meiersurvivorfunctions

4.5. Exit hazards Figure 4 reportsthe Kaplan-Meierestimates of the unemploymentexit hazardby period and group. The top left subfigurerefers to the ePBD group. The unemploymentexit rate before the policy change (dashed line) is very low at the startof the unemploymentspell, reaches a maximum of 0-1 per week after 20 weeks of unemploymenthave elapsed, and declines gradually to a very low level. Interestingly,there is an importantspike in the unemploymentexit rate in week 30-the week when regularunemploymentbenefits are exhaustedfor almost all individuals in this group.This replicatesthe important findingsin Meyer (1990). Thereare two important rate differencesbetween the unemploymentexit ratebefore August 1989 and the corresponding after August 1989. First, the spike that was observed in week 30 "moves"to weeks 39 and 52. exit Second, the unemployment rateis stronglydepressedin the periodfrom week 20 and ending in week 40. This is the periodjust before exhaustionin the old system and in between old and new exhaustionweeks. The exit rate in the eRR group is characterized two spikes in the old system in weeks by 20 and 30 (top rightsubfigure).In the new system, the exit rateis slightly depressedalreadyfrom the start of the unemploymentspell. Thus, an importantdifference between changes to PBD and to RR emerges. In line with theoreticalpredictions,the exit rate is depressedfrom the start
@ 2006 The Review of Economic StudiesLimited

LALIVEETAL.
a
PBD increase

CHANGESIN FINANCIALINCENTIVES
2
6

1023

RR increase

0

10 20

30 .....

40

50

60

70

80

90 100

0

10 20

30 .....

40

50

60

70

80

90 100

Unemploymentduration(weeks)
Before After

Unemploymentduration(weeks)
Before After

RR and PBD increase
0

.

No treatment

0

10 20 30 40 50 60 70 80 90 100 Unemploymentduration(weeks) ----- BeforeAfter

0

10 20 30 40 50 60 70 80 90 100 Unemploymentduration(weeks) ----- BeforeAfter

Notes: Data. Owncalculations, onAustrian based SocialSecurity Source: Before, spell startsbefore August 1989; after,spell startsAugust 1989.

Source: Owncalculations, based Austrian on SocialSecurity Data. FIGURE 4

Kaplan-Meierunemploymentexit rates

with changes to RR. In contrast,with changes to PBD this rate is initially unaffected,but varies stronglyarounddates of benefit expiration. The combinedeffects of RR andPBD can potentiallybe studiedin the bottomleft subfigure (ePBD-RR group).In the old system, the unemploymentexit rate is ratherlow until 30 weeks of regularbenefits have elapsed. In the new system, we observe a depressedhazardfrom the start of the spell. Moreover,whereasthe unemploymentexit rate shoots upwardbetween week 30 and week 40 in the old system, the exit rate is stronglydepressedin the corresponding period in the new system. Furthermore, there are two notable spikes centredafter39 and 52 weeks in the new system. The remainingsubfigureshows the unemploymentexit rate for the group that was not affected in August 1989. We note thateven thoughtherewas no changeto unemployment insurance in this group,the exit rate afterAugust 1989 appearsto be lower from 10 weeks until parameters 65 weeks of elapsed unemploymentduration.There are at least two reasonsfor this reductionin the exit rate. First, real GDP growthwas lower in 1991 than in 1990 (Table 1). Second, in June 1990 this groupwas affectedby a slight increasein the benefitlevel. Both factorsmay have contributedto a lower unemploymentexit hazardafterAugust 1989 comparedto the period before August 1989. This descriptiveanalysis alreadyprovidesimportantinsights into the mechanismby which financialincentivesin UI affectunemployment duration. However,a numberof important aspects were not accommodatedso far. First, job search theory models the unemploymentexit hazard for homogeneous workers. So far, however, we have discussed unemploymentexit rates that
@ 2006 The Review of Economic StudiesLimited

1024

REVIEWOF ECONOMICSTUDIES

refer to very heterogeneousgroups of job seekers. It has been shown that failure to accountfor heterogeneitybiases the durationdependenceof the unemploymentexit hazardand, potentially, the effects of financial incentives on unemployment.Second, the change to PBD is heterogeneous: PBD increasesby 9 weeks for workersaged 40-49 years and by 22 weeks for workers aged 50 years and older. These problems will be addressedin the context of an econometric model of the unemploymentexit ratein the following section. 5. RESULTS 5.1. Statisticalmodel To estimate how financialincentives affect the unemploymentexit hazard,we apply a proportional hazardmodel.16The proportional hazardmodel posits the following specificationfor the exit rate 0(tu I x) = A(tu)exp(xfl), where 2(tu) capturesthe baseline durationdependence of the hazard(in weeks), and x are the observed characteristics the individuals.17 The baseline of durationdependenceis of centralinterestin this paperbecause it refersto the exit ratefor a homogeneous group of workers.We specify the durationdependenceof the hazardas a piecewise constantfunctionof elapsed durationas follows: (tu)= exp ( i1(41 < tu < 4(1 + 1))+151(tu > 60) . (1)

(14=0

Thus, the hazardrate shifts in every four-week interval.Because there are very few transitions beyond week 60, the last time intervalcovers the entire remainingdurationof the spell as of week 60. The treatmenteffect can be identified in a (log) diff-in-diff setting. Denote eligibility for the extension of PBD from 30 to 39 weeks by eP39 = I (ePBD = 1, age < 50), eligibility for an extension to 52 weeks is denoted by eP52 = I (ePBD = 1, age > 50). Second, introducethe calendartime varyingfunction A89(tc) = I (tc > mdy(8, 1, 1989)) where tc measurescalendar time in days since 1 January,1960, and mdy(x, y, z) gives the numberof days since 1 January, 1960 of day y in monthx and year z. The functionA89(tc) recordsthe momenta spell entersthe termeP39*A89(tc) indicates periodafterthe policy change has takenplace. Thus, the interaction that an individualsatisfying all eligibility criteriafor the extension to 39 weeks has enteredthe period when this policy change has been enacted.The durationdependenceof the hazardrate is specified as follows:18
=

lfol + flleP39 + f21eP52 + fl31eRR + P41 (eP39 + eP52) * eRR + fi5sA89
* * + 61teP39* A89 + 621eP52 A89 + 631eRR A89

* + 641eP39 eRR* A89 + 65seP52* eRR* A89 l = 0,..., 15. (2)

16. See Vanden Berg (2001) for a recent survey of the propertiesof the mixed proportional hazardmodel and for applicationsof this model to durationdata. 17. These are age, maritalstatus,female, education,log(previousmonthlyincome), recall status, blue collar, seasonal industry, time spentnon-employed,tenure,andquarter inflow.Use of the exp(.) function of manufacturing industry, guaranteesthatthe hazardrate is non-negativein the entiredomainof xfi. 18. For ease of exposition, we suppressdependenceof A89(tc) on calendartime tc and write A89.

? 2006 The Review of Economic Studies Limited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES

1025

The set of fi parameters and capturesex ante differencesbetween groups (fli, ..., 41l) changes to durationdependenceoccurringover time (51st). Note that we assume thatthere are no ex ante differences between those individualswho are eligible for the change to RR and an extension of PBD from 30 to 39 weeks comparedto individuals eligible for the change to RR and the change to PBD from 30 to 52 weeks. The set of 6 parametersmeasurethe change in the duration dependenceof the hazardrate due to changes in financialincentives. There are five sets of 6 parameters because the policy change entails five interventions(P39, P52, RR, and combinations). 6(i and 621capturethe effect of extendingPBD, 631capturesthe effect of increasingthe gross RR. The parameters and651test whetherchanges to both dimensionsof unemployment 641 insuranceaffect the unemploymentexit rate in a way that would not be expected from two separatechanges to one dimensiononly. Thus, these parameters addressthe issue whetherincreasing the generosity of the unemploymentinsurancesystem due to simultaneouschanges in RR and PBD generatesdisincentive effects beyond the effects expected from two uni-dimensional changes. With this proportionalspecificationof the exit hazard,the individuallikelihood contribution is
Li = 0(tui I xi)1-ci I xi), S(tui (3)

= where ci = 1 if the spell is rightcensored,and ci = 0 otherwise;and S(tui Ixi) exp(- f'i O(z I is the survivorfunction. The likelihood function is obtained as the productof the indixi)dz) vidual likelihood contributions.The parameterestimates are obtained by maximizing the log likelihood.19 The conditionalhazardestimates addresstwo important issues. First, the model allows for the differencesacross treatments.Second, the hazardrate is identifiedconditionalon important differencesacross individualscapturedby x.

5.2. Main results hazto Figure5 reportsthe differencein the factualhazardratewith treatment the counterfactual ardratewithouttreatment.20 obtainthe factualhazardrate(solid line of Figure5), we calculate To the predictionfor the individualhazardrate with treatment 0l1(tuI xi), tu = 0, 4, ..., 56, 60, 100, andtakethe averagewithinthe groupof treatedindividuals,thatis, we averagewith respectto the distributionof individualcharacteristics in the populationreceiving the treatment.To obtain xi the counterfactual hazardrate (dashedline in Figure5) we impose all treatment effects 6 to be 0 to get the individualhazardrateOo(tu xi) and again averageacross treatedindividuals.The difI ference between the two hazardsis equal to the "averagetreatmenteffect on the treated".21 The top left subfigurereportsthe effect of extendingthe PBD from 30 to 39 weeks. IncreasingPBD does not affect the unemploymentexit rate until an elapsed durationof about 16 weeks. In week 30 (the old exhaustiondate), the differencesin hazardratesbetween the new and the old system become very large. Thereafter,the hazardrate in the new system increases strongly compared to the correspondingrate in the old system. In weeks 42-50-when benefits have expired also in the new system-the unemploymentexit hazardis higher in the new system. The higher exit rates after benefit expirationunderthe new system could be takenas evidence for a (transitory)
19. This specificationdoes not allow for unobservedheterogeneity.It turnsout that allowing for unobservedheterogeneitydoes not affect results.Results are availableupon requestfrom the authors. 20. See table Al on http: / /www. restud. org/supplements .htm for the coefficient estimates. 21. Note that the hazardrate is constant over successive four-week periods. Figure 5 plots the hazardrate with respectto the startof the four-weekperiod.

?

2006 The Review of Economic Studies Limited

1026

REVIEWOF ECONOMICSTUDIES
PBD 30-39 weeks PBD 30-52 weeks

\5

0

10

20

30

40

50

60

70

80

90

100

0

10

Unemployment duration (weeks)
Treated ..... Control

20

30

40

50

60

70

80

Unemployment duration (weeks)
Treated -.... Control

90

10

RR increase

/,

\B

\/

A03

ek

n

Rices

B

05

ek

n

Rices

0

10

20

30

40

50

60

70

80

90

100

Unemployment duration (weeks)
Treated Control

/.
6ore 0 10 20 30

Ir. .......... .
6w acltos 40 50 60 70 ae 80 nAsra 90 100 oilScrt 0 10 aa 20

...

/
40 50 60 70 80 90 100

30

Unemployment duration (weeks)
Treated----Control

Unemployment duration (weeks)
_

Treated-----

Control

of interval. Notes: x-axisgivesthebeginning theduration Data. SocialSecurity based Source: Owncalculations, on Austrian 5 FIGURE Estimatedaveragetreatedand controlhazardrates (basedon table Al)

entitlementeffect (discussed in Figure 1). Alternatively,it could simply mean that success from increasedsearcheffort takes some time to materialize.Fromweek 54 onwards,thereis no effect of extendingPBD on the unemploymentexit rate. In contrast,extending benefits from 30 to 52 weeks reduces the exit hazardin an earlier stage of the unemploymentspell (top right subfigure).While the unemploymentexit ratesunder the two systems are initially very similar,they startto divergeat week 14. The largestdifference occurs, again, in week 30-the period when benefits expire in the old system. Between weeks
@ 2006 The Review of Economic StudiesLimited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES

1027

30 and 42 the hazardunder the old system continues to be above the one of the new system. the Thereafter, benefit exhaustioneffect of the new system dominates,reachingits peak at week 54 (which covers the periodof benefitsunderthe new system) and stays higheruntil week 58. In week 62 (which covers the entire periodfrom week 60 to the end of the spell) there is no effect of extendingthe PBD. Increasingthe RR tends to depress the unemploymentexit rate much less strongly than benchangingPBD (middlesubfigure).Individualswith access to more generousunemployment efits tend to leave unemploymentless rapidly in the covered period (weeks 2-30, exceptions are weeks 22 and 26). After benefits have expired for all individualsthere is no effect on the unemploymentexit rate (with the exception of weeks 50 and 62). Does it matterwhetherindividualsare affected by a combinedchange insteadof two separate changes to financialincentives?The bottomleft subfigureanswersthis questionby comparing exit rates for the groupthat experiencedboth an increasein RR and a nine-week increasein PBD. The hazardrateis slightly (yet insignificantly)lower in the firstfour-weekinterval.Thereafter, the hazardrates are very similar-with the exception of the hazardat week 18. The main differencebetween the two graphsshows up at weeks 26, 30, and 42 (aroundold and new benefit expirationdates). The hazardunderthe new system is higher until week 60. Thereafter, no differencesremain. The bottom right subfigureshows the hazardrates for the group, which experiencedboth an increase in RR and a 22-week increase in PBD. The two graphs diverge already from the beginning of the unemploymentspell. The hazard rate under the old system lies above the one under the new system until week 48. Only around the date of benefit expiration (weeks 46, 50, and 58) hazard rates are higher under the new system. No difference remains after week 60. So far we have discussed hazardrate estimates. These are importantbecause job search on theory offers sharppredictionsregardingthe impactof unemploymentinsuranceparameters the exit hazard.However,policy makersare interestedin the implied effects on unemployment durationsor, equivalently,effects on survivorfunctions. Figure 6 reportsthe difference in the factual survivorfunction with treatmentto the counterfactualsurvivorfunction without treatment. Specifically, in the first step we estimate the survivorfunction with treatmentfor each individual as implied by the hazard rate estimates S1 I xi) = exp(fo (1 xi)dz) with (z (tu tu = 0, 4, ..., 56, 60, 100. The correspondingsurvivorfunction withouttreatmentis So(tuIxi) = hazardratewithouttreatment obtainedby imposing is exp(- fo Oo(zIxi)dz). The counterfactual all treatment effects 6 to be 0. This gives the change to the survivorfunctionin the treatedgroup at the individuallevel. In the second step, we averagethis change with respectto the distribution of individualcharacteristics in the populationreceiving the treatment.This gives the change xi in the populationsurvivorfunctionreportedin Figure6. Extending PBD by 9 weeks entails a positive contributionto the change in expected unemployment durationin the time period between 20 and 50 weeks (top left subfigure).Both, the periods before 20 weeks of elapsed durationand the period after 50 weeks of durationhave elapsed do not contributeto increasingexpected duration.The maximumcontributionarises in week 35, exactly in between the old benefitexhaustionweek (30) and the new benefitexhaustion week (39). Results are qualitativelysimilarbut quantitatively much strongerfor an increaseof PBD by 22 weeks (top right subfigure).Again, the unemploymentspell can be divided in three periods. From week 0 to week 12, the contribution expected unemploymentdurationis slightly negato is tive, from week 12 to week 60, the contribution stronglypositive, and from week 60 onwards, the contribution expected unemploymentdurationis positive but small. Again, the maximum to contributionoccurs at week 40, which is roughly in between the old exhaustionweek (30) and
? 2006 The Review of Economic Studies Limited

1028
" PBD 30-39 weeks

REVIEWOF ECONOMICSTUDIES
PBD 30-52 weeks

o

>

;

0

10 20 30 40 50 60 70 80 90 100 Unemploymentduration(weeks) RR increase

0:

10 20 30 40 50 60 70 80 90 100 Unemploymentduration(weeks)

O

-4

0

10 20 30 40 50 60 70 80 90 100 Unemploymentduration(weeks) PBD 30-39 weeks and RR increase PBD 30-52 weeks and RR increase

"5

0

10

20

30

40

50

60

70

80

90 100

0

10

20

30

40

50

60

70

80

90

100

Unemploymentduration(weeks)
Notes:

Unemploymentduration(weeks)

x-axis gives the end of the durationinterval.Differencerefersto the end of the duration interval.

Source: Own calculations,based on AustrianSocial SecurityData.

FIGURE 6
Simulations:differencebetween surviorfunctionwith treatment survivorfunctionwithouttreatment and (control)

the new exhaustionweek (52). However, the strongest difference lies in the magnitudeof the contribution.Whereas extending durationby 9 weeks generates a maximum contributionon the orderof 2-5 percentagepoints, the correspondingmaximumcontributiondue to a 22-week increaseexceeds 5 percentagepoints. In contrastto benefit durationextensions, an increase in the benefit RR generates a positive contributionto expected unemploymentdurationright from the startof the unemployment spell (middle subfigure).Most of the prolongingcontributionoccurs in the covered period of the unemploymentspell (weeks 0-30). There is also a positive, but much weakercontribution to
@ 2006 The Review of Economic Studies Limited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES
TABLE5

1029

Simulatedeffects on expecteddurationinfirst 104 weeks Treated No treatment Changeto one parameter PBD 30-39 weeks PBD 30-52 weeks RR increase Changeto two parameters PBD 30-39 and RR increase PBD 30-52 and RR increase 16-91 17-53 20-62 20-97 21-95 29.43 Control 16-91 17-08 18-35 20.60 21-09 23-70 Effect 0.00 0.45 2.27 0-38 0.86 5-72

in Notes: Based on populationreceivingthe treatment the periodafterthe policy change. RR, replacementrate;PBD, potentialbenefitsduration. Source: Own calculations,based on AustrianSocial SecurityData.

expected unemploymentdurationin the period that is no longer covered by regularunemployment benefits (week 30 onwards). The bottomtwo subfiguresreportresultsfor interventionsthatincreasePBD as well as RR. Two interestingresultsemergein comparisonwith isolated changes to PBD (top two subfigures). First, the contributionto expected unemploymentdurationis positive from the startof the unemploymentspell. This is clearly the impact of the RR on top of the PBD effect. Second, the maximumcontribution expectedunemploymentdurationincreasesstronglyfrom 2.5 percentto age points to more than4 percentagepoints (PBD 30-39 weeks) and from 6 percentagepointsto almost 14 percentagepoints (PBD 30-52 weeks). To indicatethe effects of the changesin financialincentivesTable5 shows the averageunemploymentdurationin the first 104 weeks of the unemploymentspell.22The firstcolumn in Table 5 gives the factual expected unemploymentdurationwith treatmentfor the five treatedgroups and the group that is not affected by an interventionin August 1989.23The second column in Table5 gives the counterfactual expected unemploymentdurationwithout treatmentfor the five on treatedgroups.The thirdcolumn gives the effect of the interventions expectedunemployment duration.24 Extending the PBD by nine weeks tends to increase expected unemploymentdurationby 0.45 weeks or by 0.05 weeks per additionalweek of PBD (second row). Increasing PBD by 22 weeks generatesabout 2-3 additionalweeks of unemployment(thirdrow). Thus, the second PBD extension produces twice as many weeks of unemploymentper additionalweek of PBD (0.10). This result lies within the range of previousfindingsregardingthe effect of PBD on un(see Section 2) andis also similarto the estimatesof LaliveandZweimiiller employmentduration (2004a) who find a disincentive effect of 0.05 weeks per additionalweek of PBD. In contrast,
22. We reportexpected unemploymentdurationin the first 104 weeks because, in orderto estimatetotal expected unemploymentdurationwe need to know the survivorfunction until infinity.Since inference on the survivorfunction limit our tends to become ever more unreliableas we extend the durationof the unemploymentspell, we arbitrarily discussion to the first 104 weeks, which are quite well identifiedin our large data-set.Expectedunemploymentduration is obtainedby integratingthe populationsurvivorfunctionwith respectto time up to 104 weeks. 23. Recall thataverageunemployment durationin the controlgroupis 16-5weeks in the periodafter 1989 (Table4). The corresponding numberimplied by the econometricmodel is 16-9 weeks (top, left cell). This is strongevidence that the econometricmodel fits the data well. The resultingdifferenceis due to the fact that averageunemploymentduration treatsspells, which are right censoredin the first 104 weeks as completed,whereasthe econometricmodel accountsfor rightcensoring. 24. Note that the simulationresults in Table 5 give the "effect of treatmenton the treated".A concern with these simulationsis that the treatedgroups differ from the control group. We deal with this concern below in the sensitivity analysis. @ 2006 The Review of Economic Studies Limited

1030

REVIEWOF ECONOMICSTUDIES

increasingRR by 6 percentagepoints tends to prolong unemploymentdurationby 0-38 weeks durationwith respectto the RR of about (fourthrow). This implies an elasticityof unemployment 0.15, which is small comparedto the resultsof otherstudies discussed in Section 2. Individualseligible to a combined nine-week increase to PBD and a 6 percentagepoint situationwithincreasein RR are unemployedfor 0-86 weeks longer than in the counterfactual out this combined intervention(fifth row). Interestingly,this change to unemploymentduration equals almostexactly to the predictionobtainedfrom two separatechanges,thatis, 0.38 + 0.45 = 0.83. Individualswho get both, a 22-week increasein PBD and a 6 percentagepoint increasein RR are unemployedmuch longer than in the counterfactualsituation of no intervention(5-72 weeks, sixth row). Note that the "addingup result"no longer obtains for this group of individuals, that is, 2.27 + 0-38 = 2.65 weeks instead of 5.72 weeks.25 This result is obvious already from Figure 6. Whereasthe survivorcurve difference in the case "PBD 30-39 weeks and RR Increase"is approximately sum of "PBD 30-39 weeks" and "RR Increase",this is not true the for the case "PBD 30-52 weeks and RR Increase". 5.3. Sensitivity analysis A questionthatarises from the previoussubsectionis to what extent the resultsare drivenby the heterogeneityof treatment groupsandcontrolgroup.Forexample,if the exit ratesof these groups effects will be biased. To investigatehow are subjectto differenttime trends,estimatedtreatment robustour results are, we performtwo types of sensitivity analyses focusing on the main issue of heterogeneityin treatmentand control groups.First, we redo the analysis for several specific subgroupsthat are more similarthan treatedand controls in the baseline model in terms of age, monthlyincome, and previouswage. Second, we redo the analysis for a subsamplethatis closer in to the "policyinterventionthreshold" termsof calendartime and age. is effects when identification basedon smaller, Columns2-6 of Table6 presentthe treatment but more similar groups. Column 1 of Table 6 reproducesour baseline estimates for ease of comparison. In column 2 we identify the RR-effectfrom restrictingthe sample to young, low-wage individuals. Restrictingthe sample to individualsbelow age 40 means that includedworkersare not eligible to extendedPBD. Restrictingthe sample to individualswith income below ATS 17,610 lets us comparelow-wage individualsbelow the RR-eligibility threshold(ATS 12,610) to lowwage individualsnot widely above the threshold.This is in the spiritof regressiondiscontinuity analysis (see Hahn, Todd and Van der Klaauw,2001), which exploits situationswhere assignment to treatmentis a discontinuousfunction of some given variable(in the presentcontext, an individual'spre-unemployment income). Column 2 shows that expected durationis 0.31 weeks for treatedindividuals, an effect, which is very close to our baseline estimate of 0-38 longer weeks. Columns 3 and 4 of Table 6 identify the PBD effect from comparingunemploymentduraincome (to ensurethatRR remains tions of individualswith a sufficientlyhigh pre-unemployment or have high work experience (column 4). unchanged)who are either older than 40 (column 3) In otherwords, in column 3 treatedand controlsare more similaralong the dimensionof age and income, whereasin column 4 the groupsare more similaralong the dimensionof experienceand

25. We have also investigatedthe addingup result by comparingexpected durationwhen we ignore the parameter estimates that measurethe additionaleffect of the joint change to expected durationthat allow for the effect due to the joint change. This analysis indicates that the effect of changing RR and PBD for the prime-ageworkersis the same as the effect predictedfrom two separatechanges. In contrast,two separatechanges in the PBD and in the RR for the older workersare predictedto increaseunemploymentdurationby 2.98 weeks whereasthe total change is 5.72 weeks.

? 2006 The Review of Economic Studies Limited

TABLE6

Sensitivityanalyses: making the control group more similar.Expectedduration(we Baseline result Identify RR Identify PBD Identify PBD+R Identify PBD (5) (4) (3) (2) (1) Sample restrictions Age Monthly income Workexperience Change to one parameter PBD 30-39 weeks PBD 30-52 weeks RR increase -

<40 <17, 610
-

>40 > 12, 610
-

-

>40

> 12, 610 high

<12, 610 and hig > 12, 610 and lo

0.45 2.27 0.38 0.86 5.72 225,821 -792,903

0.31
-

0.36 1.99 -

0.63 2.29 -

-

-

O

Change to two parameters PBD 30-39 and RR increase PBD 30-52 and RR increase N InL

39,685 -134,303

117,208 -415,646

142,523 -489,566

CD
0 0 r 0 0 o

0.76 5.97 38,978 -145,098

Notes: "High"refers to individualswith 6 years out of previous 10 and 9 years out of previous 15 work experience; "low criterion.RR, replacementrate;PBD, potentialbenefits duration. Source: Own calculations,based on AustrianSocial Security Data.

CD

CD C:L.

1032

REVIEWOF ECONOMICSTUDIES

income. The resulting PBD-treatmenteffects indicate that the increase in PBD from 30 to 39 weeks increases average unemploymentdurationby 0.36 weeks (column 3) and 0.63 weeks (column 4), respectively.Similarly,the increasein PBD from 30 to 52 weeks increases average duration 1-99 and2.29 weeks. Notice thatourbaseline estimatesof 0.45 weeks unemployment by (PBD 30-39 weeks) and 2-27 weeks (PBD 30-52 weeks) are withinthe rangeof these estimates. Finally, columns 5 and 6 identify the joint effect of RR and PBD increases. In column 5 we confine the sample to workers older than 40 and in column 6 we concentrateonly on in high-experienceworkers.Variation RR plus PBD eligibility relies, respectively,on differences across groupsin experience and income (column 5), and on differencesin age and income (column 6). Dependingon the includedgroups, an increasein RR togetherwith an increasein PBD from 30 to 39 weeks raises the averagedurationof unemploymentby 0.76 and by 0.65 weeks, respectively.Again, this is very close to our baseline estimateof 0.86 weeks. The same holds for the joint RR and PBD increase from 30 to 52 weeks. Confiningsamples to older workersyields an estimatedincreasein unemploymentdurationof 5-97 weeks. Alternatively, when we confine to high-experienceworkersthis estimateis 6-25 weeks. Both estimatedeffects are very samples close the 5.72 weeks baseline estimate.In sum, the sensitivityanalysis of Table6 makes us conin fidentthat our baseline results are ratherrobustand not particularly sensitive to permutations the samplesused to identify the treatment effects. A final sensitivityanalysis conductedin Table6 concernsthe 1990 policy change in RR that has so far been disregarded.Recall that in June 1990 policy change led to an increase in RR, which was roughly linear in income (RR = 0.47 with income 12,610 AS and RR = 0-41 with income 27,000 AS). On average, this led to an increase in the gross RR of about 2 percentage points (Table3). Column7 shows thatcontrollingexplicitly for the 1990 policy change leads to but slightly higher effects of the estimatedparameters, does not affect the generalpicture.26 The resultsof the second type of sensitivityanalysis are shown in Table7. Here, in the spirit of regressiondiscontinuityanalysis, we adjustgroupsin such a way thatthey are close to the eligibility threshold.More precisely,we only include workersaged 38-41 and48-51, respectively (as opposed to workersaged 35-54 in the baseline model); and we confine our analysis to the inflow to one year before and one year after the policy change (as opposed to two years before and after in the baseline model). This reduces the size of the sample strongly,from more than 225,000 workersin the baseline case to less than 37,000 workersin the thresholdsample. The formersample restrictionshould substantiallyreduce heterogeneitybetween groups as age is a of very importantdeterminant unemploymentexit rates. The latterrestrictionis a simple check whether the results in the baseline model are driven by a trend that adversely affects treated workers.If this is indeed the case we would expect smaller treatmenteffects in the "threshold" analysis, which is based on a much more narrowtime window.For ease of comparisoncolumn 1 in Table7 reproducesour baseline results,column 2 shows the resultsfrom the thresholdsample. The results from the thresholdsample indicate that increasingPBD from 30 to 39 weeks increasesthe expected durationof unemploymentby 1-38 weeks, whereasincreasingPBD from 30 to 52 weeks increasesexpected durationby 2.56 weeks. While the lattereffect is of the same orderof magnitudeas in the baseline model (2-27 weeks) the formereffect is now largerthanin the baseline model. Similarly,we find that the increasein RR has a strongereffect on expected durationin the thresholdsample where it raises expected durationby 1.50 weeks as opposed to 0.45 weeks in the baseline model. The joint PBD and RR changes lead to respective increases in unemploymentdurationby weeks (PBD 30-52). The weeks (PBD 30-39) and by 6.23 1.40
26. We control for the 1990 policy change by allowing for a change in the baseline hazardrate affecting all spells as of 1 July, 1990. A second change is modelled only for the high-income group that is affected by the 1990 change. This implies thatthe 1989 policy parameters identifiedusing informationfrom the two pre-programme are years and the post-programme periodlasting from August 1989 to July 1990.

? 2006 The Review of Economic Studies Limited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES
TABLE7

1033

Sensitivityanalysis: getting close to age and calendar time threshold.Expectedduration(weeks) Baseline result Samplerestrictions Age Calendartime Changeto one parameter PBD 30-39 weeks PBD 30-52 weeks RR increase Changeto two parameters PBD 30-39 and RR increase PBD 30-52 and RR increase N InL 35-54 1 August, 1987 to 31 July, 1991 0.45 2.27 0.38 0.86 5-72 225,821 -792,903 Threshold 38-41 or 48-51 1 August, 1988 to 31 July, 1990 1.38 2-56 1-50 1.40 6-23 36,587 -130,269

RR, replacementrate;PBD, potentialbenefitsduration. Source: Own calculations,based on AustrianSocial SecurityData.

formeris somewhat largerthan, the latter similar to the baseline results. However, the general pictureis similarthan in the baseline model. ChangingRR and PBD simultaneouslyleads to an increasein expected unemploymentdurationlargerthan the effects of two isolated increasesfor the older workers(PBD 30-52) whereasthis is not the case for younger workers(PBD 30-39). The results in Table7 also indicatethat groupheterogeneityin the baseline model does not bias our results upwards.In contrast,we find somewhat strongertreatmenteffects in the threshold analysis. In additionto these two types of sensitivity analysis we performeda numberof additional analyses. We investigatedto what extent unobservedheterogeneitymatters,what happensif we exclude seasonal workers and whetherthe results change if we focus on exits to regularjobs insteadof exits out of unemployment.Basically,our main resultsdo not change.27 5.4. Empiricalestimatesand theoreticalpredictions The empiricalanalysisin Sections 5.2 and 5.3 has revealedseveralinterestingpieces of evidence. First, we have seen thatan isolatedincreasein RR by about 15%(6 percentagepoints) leads to an increasein unemploymentdurationof roughlyequal size as the 30% increasein PBD (9 weeks). Second, the increasein unemploymentdurationsis relativelysmall for the PBD extension from 30 to 39 weeks, but is much largerfor the increasefrom 30 to 52 weeks. Whereasfor the PBD extensionfrom 30 to 39 weeks the extendedunemployment durationis 0.35 days for every week of extraPBD, this is 0.70 days for every week of extraPBD for the increasefrom 30 to 52 weeks. Third,the effect of a joint increase in RR and PBD from 30 to 39 weeks is only slightly larger thanthe sum of the effects from two separatechanges whereasthe effect of joint increasein RR and PBD from 30 to 52 weeks is much largerthanthe effects of two separatechanges. In Section 2 we have reviewedthe main predictionsfromjob searchtheory.Let us now discuss to which extent our empiricalresultsare consistentwith these predictionsand which pieces of evidence are not. A first predictionfromjob searchtheory was that if entitlementeffects are
27. We dealt with unobservedheterogeneityby allowing for a discrete distributionof unobservedheterogeneity with two mass points (Heckmanand Singer, 1984). Althoughwe find evidence of the presenceof these two mass points the estimatedeffects of a change in RR and PBD are not affected by taking unobservedheterogeneityinto account.We did not find evidence of the presenceof more thantwo mass points. All resultsof the additionalanalyses are availableon request.

? 2006 The Review of Economic Studies Limited

1034

REVIEWOF ECONOMICSTUDIES

negligible and/or most unemploymentexits take place before benefits expire, increases in RR and extensions in PBD should lead to a reductionin job searcheffort and hence to longer unemploymentdurations.This predictionis clearly consistentwith the evidence. In all our estimations we found thatincreasingRR and/orextendingPBD resultedin an increasein unemploymentduration.A second predictionwas thatincreasesof RR shouldtriggerstrongbehavioural responses early in the unemploymentspell, whereas extensions of PBD should lead to strong responses aroundthe dates when benefits expire. Also, this predictionis clearly consistent with data. We find that increases in PBD are associated with strongchanges in exit rates from unemployment around(old and new) dates of benefit expiration,whereas behaviouralresponses due to an increase in RR are distributedmore uniformly over the unemploymentspell. This patternalso helps to rationalizethe result that a 15%increase in RR can lead to roughly the same effect on unemploymentdurationthan a 30% extension of PBD. This is because the change in RR affects behaviourstrongestfrom the startof the unemploymentspell (hence affecting all unemployed) while the extension of PBD has the largest effects arounddates of expiration(when many have alreadyleft unemployment).A thirdpredictionwas thata simultaneousincreasein RR and PBD should lead to an increasein unemploymentdurationslargerthanthe sum of the increasesfrom two isolated changesin these policy parameters. estimatesclearly confirmthis predictionfor Our older workerswhereasfor prime-ageworkersonly small (and statisticallyinsignificant)additive effects were found. Hence, while there are several theoreticalpredictionsthat are clearly consistent with the data, we are left with two empiricalresults that areprimafacie difficultto reconcile with theoretical predictions.The first concerns the heterogeneityof PBD effects across age groups. The reasonwhy older workersreact much more stronglythanprime-ageworkershas to do with conditions that are specific to this group.Olderworkershave a weakerlabour-market position.They get fewerjob offers because employerspreferablyfill theirvacanciesby hiring young or primeinstitutionalenvironment effect on age workers.Moreover,the particular may have an important older workers'searchefforts. Olderworkersare close to retirement(or face a higherprobability of becoming eligible to early retirementbenefits than prime-age workers).Hence, they have a lower incentiveto searchbecause the value of findinga job is small. For this reason,a more generous unemploymentinsurancesystem may induce older workersto reduce their search efforts more stronglythanprime-ageworkers. The second result that is difficult to reconcile with theoreticalpredictionconcerns the age differences in reactions to a joint variationin policy parameters. the negligible interaction For effect for prime-ageworkerswe lack a theoreticalexplanation.One reasoncould be thatthe size of the two policy changesis simply too small meaningthatsuch an interaction effect is difficultto measureanyway.In contrast,the large interactioneffect for older workerscan be rationalizedin much the same way as the large isolated PBD effects discussed above: labour-market conditions togetherwith the institutionalenvironmentmost likely explain the strongbehaviouralresponses of the older workersto a joint increasein RR and PBD.28 estimates Finally, note that our sensitivity analysis of Table 6 above shows that parameter are otherwise quite robust. The effect of the RR increase, the effect of PBD extensions, and the additive effect of simultaneousRR and PBD changes are quite robust to variationsin the control groups used to identify the treatmenteffect. In particular, estimatedtreatmenteffect the of an RR remains unchangedwhen the control group is confined to workersyounger than 40 who are slightly above the earningsthreshold.The effects of a given PBD extension and of a simultaneousRR and PBD change remain unchangedwhen the control group either consists
28. Note thatthe earlyretirement explanationis supported the facts. Given an elapseddurationof unemployment by of at least 30 weeks, the probabilityto exit to an early retirementscheme (some form of disability insurance)is about 20% among individualsaged 40-49 and is about60% for individualsaged 50-54.

? 2006 The Review of Economic StudiesLimited

LALIVEET AL.

CHANGESIN FINANCIALINCENTIVES

1035

only of too unexperiencedworkers(of the same age) or of too young workers(with the same work experience). This suggests that, apartfrom the much largerreactions to PBD extensions and joint RR and PBD changes for older workers mentioned above, the estimated effects are quite homogenousacross the variousgroupsand largely in line with theoreticalpredictions. 5.5. Disincentive effects This final subsection comparesthe disincentiveeffects of changes in PBD and changes in RR. Since such a comparisonneeds a commondenominator, studythe effects in termsof the total we increase in benefit payments. Suppose, for instance, that policy makers increase the RR. This raises the total amount of benefit payments for two reasons. First, the RR increase will raise benefitpaymentseven if individualsdo not change theirbehaviour,simply because higherbenefits have to be paid for the same numberof days individualsspend in unemployment.Second, the RR increasewill induce individualsto stay longer in unemployment, thus raisingbenefitpayments further. intuitivelyappealingmeasuresplits up the total costs into the fractionof direct An costs (withoutbehaviouralchanges) and the fractionof indirectcosts resultingfrom individuals' change in behaviour. Specifically, we estimate direct cost due to the change in the UI system from an increase in the benefits level from boi to bli and/oran increase in PBD from Eoi to Eli as follows. The is expecteddurationof unemploymentwithouttreatment fEoi Soi(t)dt, where Soi (t) So(t Ixi) is the counterfactual survivorfunction without the policy change.29The direct costs due to the change in the UI system are given by bli foE" Soi (t)dt - boi fEOi Soi (t)dt. The indirect costs due to the change in the behaviourof job seekers are given by bli OEli (t)dt - bli Sli Soi(t)dt OEli where Sli (t) - S1 t Ixi) is the factualsurvivorfunctionfor treatedindividualsand liE"Sli (t)dt is the expected durationof unemploymentin the new system that takes behaviourchanges into account.Denote by Ds the fractionin total additionalcosts due to the directeffect of the change in the system and by DB the fraction due to the indirect effect from changes in individuals' behaviour,with Ds + DB = 1. Formally,Ds and DB are given by
Ex, [bli Ds
1OEi

S Oi SEoi Ex, [bli 0ESli(t)dt - bo f
-

Soi(t)dt -boi fEoi Soi (t)dt] Soi (t)dt

(t)dtt

and DB =
(Sli Ex, [bli oiE" li(4) Soi)(t)dt]

Ex, [bli oE Sli(t)dt -boi fEoi Soi(t)dt]

Table 8 shows the estimated split of the total increase in benefit payments into Ds (first column)and DB (second column)of the variouschanges for the respectivetreatedgroups.Forthe PBD change from 30 to 39 weeks, more than 80%of additionaltotal costs are direct.Obviously, these costs result only from spells thatlasted longer than 30 weeks alreadyunderthe old system andthe behavioural effects, less than20%of totaladditionalcosts, reflectan increasein the length of all spells. This comparablysmall numberreflectsthe relativelysmall behaviouralchanges we have estimatedabove.Forthe RR change,the directcost componentis even largerand amountsto about90%. This is not surprising, because with an increasein RR all eligible spells are affected.
29. Note thatb = B w, where B is the gross RR and w is gross weekly income. Moreover,this ex ante expectedcost measuredoes not accountfor unemploymentassistancepayments,which are availableafterunemployment benefitshave run out. The datado not containinformationon unemploymentassistance.Also, results are qualitativelynot sensitive to using imputedunemploymentassistancepayments. @ 2006 The Review of Economic Studies Limited

1036

REVIEWOF ECONOMICSTUDIES
TABLE8 Simulatedeffects on benefitpayments Percentageof total change in cost due to change UI system Changeto one parameter PBD 30-39 weeks PBD 30-52 weeks RR increase Change to two parameters PBD 30-39 and RR increase PBD 30-52 and RR increase 82-0 54.8 89.9 84.3 51.1 Behaviourof job seekers 18.0 45.2 10-1 15-7 48-9 Total 100.0 100.0 100.0 100.0 100.0

Notes: Based on populationreceiving the treatmentin the period after the policy change. RR, relatedrate;PBD, potentialbenefitsduration. Source: Own calculations,based on AustrianSocial SecurityData.

Results aredifferentfor the groupseligible to the strongincreasein PBD from 30 to 52 weeks. In this groupbehavioural effects aremuch strongerand of roughlyequal size as the directeffects. In otherwords, individualsreact stronglyto the increasein benefit duration,and these behavioural changes are the main factordrivingthe total additionalcosts of the policy change. Differences in replacementratiosare here of lesser importance.While a somewhathigherfractionof additional total costs has to be attributed behaviouralchanges, the split is of a similarorderof magnitude to for changes in two parameters for the corresponding as change in only one parameter. 6. CONCLUSIONS This paper addressesthe issue of how financialincentives embeddedin the unemploymentinsurance system affect the durationof unemployment.This issue is importantfor a numberof reasons. On the one hand, the years since the turnof the centuryhave witnessed importantreforms to unemployment insurancein many (particularly European)countries.On the otherhand, studies of how simultaneouschanges of UI parametersaffect the unemploymentexit rate are lacking. Hence, it is difficultto comparethe relativeeffects of changes in the policy instruments. This paper relies on a change to unemploymentinsurancein the late 1980's in Austria. This reformleads both to extensions of the potentialduration(PBD) of regularunemployment benefits for a first group of individuals;to an increase in the earnings replacementrate (RR) for a second group of individuals;to both extended PBD and higher RR for a third group of individuals;and to no changes for a final group of individuals.This means that it is not only possible to study the relative magnitudesof two key parametersdeterminingthe generosity of unemployment compensation,but also to analysewhetherthereareexcess effects from combined in PBD and RR. changes Age, experience, and earningsdeterminethe treatmentgroup. Earningsdeterminewhether the RR goes up-only workers with low earnings get an increase in RR. Age and experience determinewhetherthe PBD goes up. Only workersfrom age 40 onwardswith a high working experienceget an increasein PBD, where the size of the increasedependson the age of the individual. Individualsbetween 40 and 50 get an increasein PBD from 30 to 39 weeks; individuals from age 50 onwardsget an increase from 30 to 52 weeks. Furthermore, changes in the UI the are introducedin a periodof employmentgrowthso that any increasesin unemployment system durationwere not relatedto labour-market conditions.The Austriancase has clear elements of a "natural experiment".
@ 2006 The Review of Economic Studies Limited

LALIVEETAL.

CHANGESIN FINANCIALINCENTIVES

1037

Although it is clear that there is heterogeneityin the treatedpopulationswe find that the effects are quite robust. In particular,we find that both the estimated effects are invariantto variationsin the control groups used to identify the treatmenteffect. This suggests that the various groups do not react systematicallydifferentto financialincentives. However,there are two exceptions. First, the relative effect of a PBD extension is smaller for prime-age workers.For durationincreasesby 0.35 days per week of extraPBD, while prime-ageworkersunemployment for older workersit increases by 0-70 days per week of extra PBD. Second, the additiveeffect of a simultaneouschange in PBD and RR is larger for older workers. The reason why older workersreact differentlyis because they face differentlabour-market conditions and a different institutionalsetting. They get fewer job offers because employerspreferablyfill their vacancies by hiringyoung or prime-ageworkers.Furthermore, they are close to retirement,and the closer to retirement less incentivethey have to searchfor a job because the value of finding the they get a job is reduced.Both labour-market conditionsand institutionalsettingmay cause older workers to reactstrongerto disincentives.In short,despitethe apparent heterogeneityin responseworkers react to incentivesin line with severalpredictionsfrom theory. The changes in incentives are interestingfrom a scientific point of view to investigateto what extent results are in line with theoreticalpredictions.Theory provides predictionsabout the patternof changes in searchintensityover the durationof unemploymentand aboutthe way separatechanges in incentivesinteract.Nevertheless,theoryonly providesa predictionaboutthe sign of the effects and does not give an indicationabouthow large these effects should be. With respectto the strongerreactionof olderworkersthereis not muchthatcan be derivedfrom theory to explainthe exact differencein response.However,it does give some idea aboutpotentialcauses of the differences:labour-market conditionsand institutionalsettings. From a policy point of view our study is interestingas well. An innovativepartof our analestimatesto split up the total costs to unemployment ysis concernsthe way we use ourparameter insurancefundsinto costs due to changes in the unemployment insurancesystem with unchanged behaviourand costs due to behavioural responsesof unemployedworkers.Forthe increasein RR we findthatadditionalcosts due to behavioural responsesaremodest, of the orderof 10%of total costs. This is differentfor increasesin PBD where the behaviouralcomponentamountsto about 20% in case of the PBD increase from 30 to 39 weeks and to close to 50% in case of the PBD increasefrom 30 to 52 weeks. Takingthese resultsat face value, we concludethatPBD is a more effective policy parameterthan RR to affect individual'sjob search behaviourand unemployment durations.From this we derive two simple policy recommendations. First, if governments change variousincentivesfor unemployedworkersthey should be awareof behaviouraleffects, and above all they should take behaviouraleffects relatedto interactionsbetween the changes in incentives and interactionbetween changes in incentives and institutionalsettings into account. Second, if they want to influence incentives the potentialdurationof unemploymentis a more effective policy tool thanthe level of the unemployment benefits.
REFERENCES in of Benefitson the Probability Re-employment Poland",Oxford ADAMCHIK,V. (1999), "TheEffect of Unemployment Bulletin of Economicsand Statistics,61, 95-108. ADDISON, J. T. and PORTUGAL,P. (2004), "How Does the UnemploymentInsuranceSystem Shape the Time Profile of Jobless Duration?" (Mimeo, IZA WorkingPaper976, Bonn). J. ATKINSON, A. and MICKLEWRIGHT, (1991), "UnemploymentCompensationand Labor Market Transitions: A CriticalReview",Journalof EconomicLiterature, 1697-1727. 29, BENNMARKER,H., CARLING,K. and HOLMLUND,B. (2004), "Do Benefit Hikes Damage Job Finding?"(Mimeo, IFAU,Uppsala). Insuranceas a SearchSubsidy: A TheoreticalAnalysis",EconomicInquiry,17, BURDETT,K. (1979), "Unemployment 333-343. CARD, D. E. and LEVINE, P. B. (2000), "ExtendedBenefits and the Durationof UI Spells: Evidence from the New Journalof Public Economics,78, 107-138. JerseyExtendedBenefit Program",

? 2006 The Review of Economic StudiesLimited

1038

REVIEWOF ECONOMICSTUDIES

Duration,Unemployment CARLING,K., EDIN, P.-A., HARKMAN,A. and HOLMLUND,B. (1996), "Unemployment Benefits, and LaborMarketProgramsin Sweden",Journal of Public Economics,59, 313-334. CARLING,K., HOLMLUND,B. and VEJSIU, A. (2001), "Do Benefit Cuts Boost Job Findings? Swedish Evidence from the 1990s",EconomicJournal, 111, 766-790. FREDRIKSSON,P. and HOLMLUND,B. (2003), "ImprovingIncentivesin UnemploymentInsurance:A Review of Recent Research"(Mimeo, IFAU,Uppsala). GREEN, D. A. and RIDDELL, W. C. (1993), "The Economic Effects of UnemploymentInsurancein Canada: An Journalof LaborEconomics, 11, 96-147. EmpiricalAnalysis of UI Disentitlement", GREEN, D. A. and RIDDELL, W. C. (1997), "Qualifying for UnemploymentInsurance: An Empirical Analysis", EconomicJournal, 107, 67-84. GROSSMAN, J. B. (1989), "The Work Disincentive Effect of Extended Unemployment Compensation: Recent Evidence",Reviewof Economicsand Statistics,71, 159-164. Effects with a and HAHN, J., TODD, P. and VAN DERKLAAUW,W. (2001), "Identification Estimationof Treatment RegressionDiscontinuityDesign",Econometrica,69, 201-209. HAM, J. and REA, S. (1987), "UnemploymentInsuranceand Male UnemploymentDurationin Canada",Journal of LaborEconomics,5, 325-353. HECKMAN,J. J. and SINGER, B. (1984), "A Method for Minimizing the Impact of DistributionalAssumptions in EconometricModels for DurationData",Econometrica,52, 271-320. HUNT, J. (1995), "The Effect of UnemploymentCompensationon UnemploymentDurationin Germany",Journal of LaborEconomics, 13, 88-120. KATZ,L. and MEYER,B. (1990), "TheImpactof the PotentialDurationof UnemploymentBenefits on the Durationof Journalof Public Economics,41, 45-72. Unemployment", J. LALIVE,R. and ZWEIMOLLER, (2004a), "BenefitEntitlementand UnemploymentDuration:Accountingfor Policy Journalof Public Economics,88, 2587-2616. Endogeneity", J. LALIVE,R. and ZWEIMOLLER, (2004b), "BenefitEntitlementand the LaborMarket:Evidence from a Large-Scale (eds.) Labor MarketInstitutionsand Public Policy Policy Change",in J. Agell, M. Keene and A. Weichenrieder (Cambridge,MA: MIT Press) 63-100. Data (Cambridge:CambridgeUniversityPress). LANCASTER,T. (1990) TheEconometricAnalysis of Transition Insuranceand UnemploymentSpells",Econometrica,58, 757-782. MEYER,B. (1990), "Unemployment MOFFITT, R. (1985), "Unemployment Insurance and the Distribution of Unemployment Spells", Journal of Econometrics,28, 85-101. R. MOFFITT, and NICHOLSON,W. (1982), "The Effect of UnemploymentInsuranceon Unemployment:The Case of FederalSupplemental Benefits",Reviewof Economicsand Statistics,64, 1-11. MORTENSEN,D. (1977), "UnemploymentInsuranceand Job Search Decisions", Industrial and Labor Relations Review,30, 505-517. MORTENSEN,D. (1990), "A StructuralModel of Unemployment InsuranceBenefit Effects on the Incidence and Duration of Unemployment",in Y. Weiss and G. Fishelson (eds.) Advances in the Theory and Measurement (New York:St Martin'sPress) 57-81. of Unemployment PUHANI, P. A. (2000), "Polandon the Dole: The Effect of Reducing the UnemploymentBenefit EntitlementPeriod Journalof PopulationEconomics, 13, 35-44. DuringTransition", Economic ROD, K. and ZHANG, T. (2003), "Does UnemploymentCompensationAffect UnemploymentDuration?", Journal, 113, 190-206. VAN DEN BERG, G. J. (1990), "Nonstationarity Job SearchTheory",Reviewof EconomicStudies,57, 255-277. in VAN DEN BERG, G. J. (2001), "Duration Models: Specification, Identification, and Multiple Durations", in 3381-3460. J. J. Heckmanand E. Leamer(eds.) Handbookof Econometrics,Vol. V (Amsterdam: North-Holland) VAN OURS, J. C. and VODOPIVEC,M. (2006), "How Shorteningthe PotentialDurationof UnemploymentBenefits EntitlementAffects the Durationof Unemployment: Evidence from a NaturalExperiment",Journal of Labor Economics,351-378. WINTER-EBMER,R. (1998), "PotentialUnemploymentBenefit Durationand Spell Length: Lessons from a QuasiExperimentin Austria",OxfordBulletin of Economicsand Statistics,60, 33-45.

? 2006 The Review of Economic StudiesLimited