Labour Economics 13 (2006) 773 798 www.elsevier.

com/locate/econbase

Unemployment duration and the interactions between unemployment insurance and social assistance
Michele Pellizzari
IGIER-Bocconi University, Italy IZA, Italy fRDB, Italy Received 17 April 2005; received in revised form 6 October 2005; accepted 6 October 2005 Available online 28 November 2005

Abstract Existing studies suggest that reforms that reduce the generosity of the unemployment benefits should lower unemployment. Despite the large number of such reforms implemented in Europe in the past decades, evidence from various data sources shows very little correlation with the evolution of unemployment. This paper suggests that the scant success of these labour market reforms can be explained by the interactions between unemployment insurance and other social assistance programmes. Evidence from the European Community Household Panel shows that recipients of unemployment insurance who are also eligible for other welfare schemes are indeed less sensitive to changes in the level and the duration of their benefits. D 2005 Elsevier B.V. All rights reserved.
JEL classification: J64; J65 Keywords: Unemployment duration; Unemployment insurance; Social assistance; Duration models

1. Introduction Many studies have already explored the effects of unemployment insurance (UI) on the duration of unemployment (Atkinson and Micklewright, 1991). Two empirical findings are now widely accepted. First, as initially showed by Nickell (1979) and Lancaster (1979), higher benefits are associated with longer unemployment spells. Later, Moffitt (1985) and Meyer (1990), having access to information about both the level and the duration of benefit entitlement at the individual level, showed that the probability of exiting unemployment increases around the time of benefit exhaustion.
E-mail address: michele.pellizzari@unibocconi.it. 0927-5371/$ - see front matter D 2005 Elsevier B.V. All rights reserved. doi:10.1016/j.labeco.2005.10.003

774

M. Pellizzari / Labour Economics 13 (2006) 773798

The literature also provides a comprehensive theoretical framework for interpreting these results. Mortensen (1977) develops a simple search model that easily delivers negative correlation between exit rates and unemployment benefits via search effort and reservation wages being respectively negatively and positively affected by income out of work. When UI entitlement expires, income out of work drops, inducing an increase in search effort and a decrease in the reservation wage, thus providing an explanation for the observed higher exits rates around the time of exhaustion. Cahuc and Lehmann (2000) recently extended the model by endogenising wages and allowing the unemployment benefit to gradually decline over time. On the basis of these findings, one would expect reforms that reduce either the level or the duration of unemployment benefits to have a positive impact on the unemployment rate. However, this prediction does not seem to conform with the recent experience of many European countries. Fig. 1 plots the time series of the unemployment rate for selected European countries. The vertical lines indicate the implementation years of reforms that have modified either the level or the duration of the unemployment benefits. The solid bars refer to changes in the amount of the benefits and the dashed lines to changes in their duration. The thickness of the lines indicates the direction of the reform: thick for reductions (either in the amount or duration of benefits) and thin for increases. A first simple fact is evident from Fig. 1: the past 20 years have been constellated by labour market reforms in virtually all European countries (also the ones not shown in the figure). Moreover, despite the coexistence in many countries of reforms of opposite sign, often implemented close to each other (in Finland and France, for example), most of the changes (19 out of 29 in all the 15 pre-enlargement EU countries) modified the system towards less generous benefits paid for a shorter time. However, already a simple visual inspection of Fig. 1 suggests that the correlation between these reforms and the evolution of unemployment is rather weak and it does not improve even when controlling for the business cycle. More convincing evidence can be produced for those countries where reforms took place during the years covered by the European Community Household Panel (ECHP)1, namely between 1994 and 2001. This is possible for six countries: Austria, Belgium, Finland, Germany, Ireland and the United Kingdom. Table 1 shows the conditional difference in the probability of finding employment between individuals who entered unemployment before and after the reform. The estimates are produced with a standard hazard model controlling for gender, age, health status, education, marital status, family size, presence of children, household income, regional unemployment and year dummies.2 Results confirm the visual impression from Fig. 1: reforms that reduced the generosity of the UI system did not have a significant effect on the probability of finding a job. There are marginally significant effects only in Austria (where the estimate is actually negative) and Ireland. The scant success of labour market reforms in Europe has been acknowledged by other authors as well, for example Coe and Snower (1997), Elmeskov et al. (1998), Orszag and Snower (1999), Nickell and van Ours (2000). This paper suggests an explanation for the failure of so many reforms of the unemployment compensation system. All European countries have complex welfare states and unemployment compensation is only one element of the system that necessarily interacts in various ways with all the other programmes. In particular, many unemployed persons receive other social assistance benefits together with their unemployment insurance.
1 2

The data are described in details in Section 4. A fuller description of the econometric model is provided in Section 6.

Belgium
12 20

Finland
12

France

unemployment rate

unemployment rate

unemployment rate

10

15

10

10

M. Pellizzari / Labour Economics 13 (2006) 773798

1980

1984

1988

1992

1996

2000

200 4

1980

198 4

1988

1992

1996

2000

2004

6 1980

1984

1988

1992

1996

2000

2004

year

year

year

Germany
10 20

Ireland
12

United Kingdom

unemployment rate

unemployment rate

unemployment rate

15

10

10

1980

1984

1988

1992

1996

2000

200 4

1980

1984

1988

1992

1996

2000

2004

1980

1984

1988

1992

1996

2000

2004

year

year
Type of reform: UI benefit UI duration

year
Direction of the reform: thick: less generous (i.e. lower and/or shorter payments) thin: more generous

Fig. 1. Unemployment and social reforms in Europe. 775

776

M. Pellizzari / Labour Economics 13 (2006) 773798

Table 1 The effect of UI reforms on the probability of leaving unemployment Country Austria Belgium Finland Germany Ireland United Kingdom Reform reduced reduced reduced reduced reduced reduced reduced duration-2000 benefit-1999 benefit-1997 benefit-2000 duration-1995 benefit-1995 duration and benefit-1995 Effecta 0.459* 0.292 0.156 0.258 0.031 0.486* 0.291 (0.252) (0.361) (0.142) (0.253) (0.169) (0.279) (0.334)

Source: ECHP 19942001. a Coefficient of a dummy for unemployment spells that started after the implementation of the reform. The estimates are obtained from a proportional hazard model with a fully non-parametric baseline and the following explanatory variables: sex, age (linear and squared), health status, education, marital status, family size, dummy for children in the household, household income, regional unemployment rate and rate of long term unemployment, year dummies.

For instance, in the sample used in this paper, more than 40% of UI recipients receive other welfare benefits at some point during their unemployment spell. Most of these other benefits are means-tested, therefore a reduction or an anticipated withdrawal of unemployment insurance is often compensated, at least partly, by higher transfers from other programmes. Moreover, even those unemployed who only receive unemployment insurance may still expect to become eligible for some social assistance programmes when their benefits expire. For these workers, too, reductions or exhaustion of unemployment insurance are less of a concern and might not affect much their search effort nor their reservation wages. These arguments will be tested empirically using data from the European Community Household Panel, which allow to reconstruct monthly labour market histories for samples of individuals from all the 15 pre-enlargement EU countries. Contrary to most studies in this literature that use data from administrative sources, this paper exploits survey data which have two main advantages. First, by relying on self-reported information about ones labour market status, it is possible to separate unemployment spells that end into employment from those that end into inactivity. Hence, I will be able to estimate the probability of finding a job rather than that of leaving unemployment or the UI registry. This is an important innovation since a non negligible fraction of unemployment spells usually terminate into other types of non employment. For example, in the largest sample used in this study exits into inactivity represent about 15% of total exits from unemployment. A second advantage of the ECHP is the high level of cross-country comparability, which allows to use variation in the administrative rules of the welfare systems to identify causal effects. In particular, the institutional features of family benefits offer an interesting source of variation. In 10 out of the 15 pre-enlargement EU countries eligibility for family benefits is determined only by the number and the age of the children in the household, regardless of income. Only in the remaining 5 countries Germany (but only up to 1997), Greece, Italy, Portugal and Spain family benefits are subject to means testing. This paper exploits this variation to identify the differential effect of unemployment benefits on the job finding probabilities of potential recipients of family benefits. Intuitively, in countries where family benefits are means tested UI recipients who are potentially eligible for child assistance will be less worried about reductions or withdrawal of their benefits, as these could be compensated by higher social assistance payments. In countries

M. Pellizzari / Labour Economics 13 (2006) 773798

777

were family benefits are universal, i.e. not means tested, there should be no such interaction effect. Results indicate significant differences in the elasticity of the job finding probability to the unemployment benefit and to the time of benefit exhaustion between male heads of households with children who are potentially eligible for means-tested family benefits as opposed to those who have access to universal benefits. The importance of interactions between welfare programmes has lately been recognised by both academicians and policy makers. Belot and van Ours (2000) provide evidence from macro data showing that countries where unemployment has fallen often owe their success to comprehensive rather than piecewise reforms of labour market policies. A theoretical justification for the importance of these interactions is discussed in Coe and Snower (1997) within a search and matching framework. Despite the acknowledged importance of potential overlappings between welfare programmes, specific evidence from micro data is still lacking. This paper aims at filling this gap by providing detailed evidence on one specific interaction. The paper is organized as follows. Section 2 briefly describes the institutional details of unemployment benefits and other social assistance programmes in Europe. Section 3 shows how interactions between welfare programmes can be analysed in a simple search model. Section 4 describes the data while Section 5 discusses the empirical strategy. The results of the empirical analysis are presented in Section 6 and Section 7 concludes. 2. Unemployment compensation and social assistance in Europe All European countries have large and well developed welfare states, nevertheless there still exist important differences in the institutional details across countries. The comparison of different types of welfare states in Europe has been the object of a large number of studies in both the economic and the political literature.3 This section simply provides a very general overview of the welfare programmes available in Europe during the period covered by the empirical analysis that follows. Excluding old age pensions (which are still the major component of social expenditure in many countries), welfare benefits are generally grouped into 5 large categories: unemployment related benefits, family cash benefits, disability benefits, housing benefits and general social assistance. In kind benefits are not considered here, even if they might play an important role for some groups of beneficiaries (disabled persons, large families, et.). Unemployment benefits are generally distinguished into unemployment insurance and unemployment assistance. The first is a standard insurance scheme by which workers who have paid sufficient contributions out of their salaries are allowed to receive a compensation if they become unemployed. Unemployment insurance is usually rather generous but benefits are typically paid only for a limited period of time (with Belgium, where the duration of payments is unlimited, being a notable exception) and various re-eligibility conditions apply for repeated spells of unemployment. Clearly then, unemployment insurance does not cover the whole population of job seekers: young workers and anybody who has not paid enough contributions and the long term unemployed who have exhausted their benefits cannot claim compensation. For this reason many countries have introduced a parallel unemployment assistance programme that pays a (typically lower) benefit to those who are not entitled to unemployment insurance and is financed through general taxation.
3

Bertola et al. (2000), Esping-Andersen (1990), Ferrera (1998).

778

M. Pellizzari / Labour Economics 13 (2006) 773798

Disability benefits are another important component of the welfare system. The typical scheme pays a benefit to individuals whose capacity to work and earn is substantially reduced by some sort of invalidity. The level of the benefit is usually determined on the basis of a measure of bnormalisedQ earnings, i.e. earnings of a similar person who does not suffer from the invalidity. Payments are usually carried over to retirement, when a disability benefit is typically converted into an old age pension. Family cash benefits are the most important welfare programme that is not related to employment. In the majority of countries (10 out of 15) family benefits are actually paid to any household with children, regardless of their income. Germany is the only country that, during the period covered by the ECHP, changed from means-tested to universal family benefits in 1997. These benefits are paid until the child reaches a certain age and the amount varies according to the childs age and to the number of children in the household. Some countries also offer supplements for single parents. Children who undertake higher education or training are often allowed to receive benefits for some time beyond the age limit. The provision of housing benefits is more varied. Some countries offer a generalised housing benefit available to everyone whose income is sufficiently low (Germany, France, Netherlands, Finland) while others simply provide specific housing supplements for those on low-income benefits (Ireland, Luxembourg, Austria and Portugal). Denmark and Sweden have both. Some countries do not offer any housing benefit but often public housing is available for low income families. Finally, all countries, with the exception of Greece and Italy,4 also provide a general social assistance scheme that aims at preventing poverty for those families that do not qualify for any of the other bcategoricalQ benefits or who still remain under a variously defined income threshold. Qualifying conditions for this type of assistance are usually related to nationality, residence and age. All countries also require those who are able to work to prove that they are actually willing to take up job offers and beneficiaries are often required to participate in training or other active labour market programmes. During the past two decades most European countries have undertaken important reforms of their welfare systems. Most reforms have actually changed the unemployment compensation system in the direction of lower benefits paid for shorter periods while much less effort has been put in reforming other welfare programmes. In fact, one of the main claims of this paper will be that the scant success of UI reforms is partly due to the lack of coordination with other parts of the welfare system. 3. A simple theoretical framework The interaction between unemployment benefits and social assistance can be analysed in the framework of a very standard search model. The theory in this section is a mere revised and simplified version of that in Mortensen (1977). In each period t , an unemployed worker needs to invest leisure time (s t ) to search for job offers. These arrive with per-period probability as t , from an exogenous wage distribution F (d ) with support (0,w ]. In each period utility is a non-decreasing function of income ( y t ) and leisure (l t ), u ( y t ,l t ).5
4 A minimum income scheme was experimented in Italy between 1998 and 2002 (Reddito Minimo dInserimento ) but was never introduced on a universal basis. y;l Bu y;l 5 The per-period utility satisfies the standard assumptions, i.e. twice differentiable with: BuB y N0; Bl N0 and Bu y;l Bu y;l V 0 ; V 0. By By Bl Bl

M. Pellizzari / Labour Economics 13 (2006) 773798

779

For simplicity, assume that working time is constant and equal to h in all jobs and that the total endowment of leisure is normalised to 1. Then, per-period utility while unemployed with benefit b t and searching for a job is u (b t ,1s t ), while a job that pays w t generates a utility flow equal to u (w t ,1h ). Unemployed workers also need to choose a reservation wage w* t such that only offers above w* are accepted, all the others are turned down. t Under these assumptions the value of unemployment U at time t can be written as: 1 rUt ubt ; 1 st ast Prfwbw4 t gUt 1 Prfwzw4 t gE fV wjwzw4 t g 1

where V (w ) is the value of employment at wage w. For simplicity assume that there is no job destruction: once workers enter employment they stay in the job forever at constant wage.6 Eq. (1) can be rewritten as: Z w 1 rUt ubt ; 1 st 1 ast 1 F w4 V wdF w 2 t Ut 1 ast
w4 t

The optimal levels of w* t and s t are then chosen in order to maximise Eq. (2), according to the following first order conditions: V w4 t Ut 1 Bubt ; 1 st a Bst Z
w

3  V wdF w 1 F w4 t Ut 1 4

w4 t

The interpretation of these two conditions is very intuitive. Eq. (3) shows that the optimal reservation wage is set at a level that equalises the value of employment and unemployment. An unemployed worker can allocate time to two different activities, search and leisure, hence optimal search time equalises the marginal utilities of search and leisure, as shown in Eq. (4). Note that both s t and w* t are time-varying: Eqs. (3) and (4) hold for all t and, for any known l sequence of benefits, {b t }l 0 , identify a series of reservation wages, {w*} t 0 , and optimal search l times, {s t }0 . The per-period probability of exiting unemployment the hazard rate is then calculated as: qt ast d 1 F w4 t and is higher when job search is more intense and when the reservation wage is lower. These results are useful to analyse the implications of different assumptions about the sequence of benefits. Consider, for example, the case of an unemployed worker who receives a constant unemployment benefit (b ) for a given number of periods, T, and nothing after that.7 For such worker the value of unemployment decreases over time as periods of positive benefit payments run out and expected future income out of work decreases. Eq. (3) implies that the

In this case: V w; h
l X uw; h j jt 1 r

Introducing exogenous job destruction does not modify the empirical implications of the model. 7 This is the specific case discussed in Mortensen (1977).

780

M. Pellizzari / Labour Economics 13 (2006) 773798

reservation wage also decreases over time. Similarly, the right hand side of Eq. (4) increases with time (the value of unemployment enters with a negative sign and the derivative with respect to w*, t given Eq. (3), is zero), therefore in order for the equality to hold the left hand side must increase as well and this can only be achieved with higher s t , i.e. search time also increases with time. At time T, when unemployment benefit entitlement expires, b t discontinuously drops to zero. This requires the reservation wage to jump down and optimal search time to go up. Note incidentally that for these effects to be non-ambiguous leisure and income must be complements (i.e. u 21(b t , 1s t ) V 0). These results are represented in Fig. 2. The previous analysis can be easily extended to a worker who receives social assistance together with his/her unemployment benefit, or, similarly, to somebody who expects to become eligible for some social assistance programmes once his/her unemployment insurance expires. Eventually all changes from one scheme to another simply generate jumps in the time profile of the benefit and can be analysed within the same framework. For example, it might happen that, given the particular rules and household composition of applicants, family cash benefits top up household income once unemployment insurance expires, leaving the time profile of benefit payments flat. In this case the model predicts no discontinuous jumps in reservation wage, search effort and exit rate, which will all remain constant throughout the entire unemployment spell. In other instances, payments under social assistance may actually be higher than under unemployment insurance. In such an extreme case the value of unemployment increases with time and all the effects derived previously are reversed.
benefit reservation wage wt*

bt

unemployment duration hazard rate

unemployment duration

search time

st T unemployment duration

qt T unemployment duration

Fig. 2. Time profiles of benefit, reservation wage, search time and hazard rate-Case 1.

M. Pellizzari / Labour Economics 13 (2006) 773798

781

4. The data: the European Community Household Panel Most of the existing studies of unemployment insurance and unemployment duration make use of administrative data.8 The advantage of these data usually consists in having detailed information about the amount and the sequence of payments as well as about individual eligibility and entitlement conditions. However, for the purpose of this paper the use of administrative data would be problematic for at least two reasons. First, in many countries unemployment insurance and social assistance programmes are administered by different governmental bodies and data available from one institution rarely include information about benefits paid by others. Therefore, using administrative data would make it very difficult to look at interactions between different programmes. Secondly and more importantly, even if comprehensive administrative data were available, in order to explore the interactions between different programmes one would need to compare similar individuals facing different unemployment benefits and social transfers: in other words one would need enough variation in the rules and regulations of both unemployment insurance and social assistance. However, there is typically little variation in such rules within one country and for identification purposes it would be helpful to use some cross-country variation as well. Unfortunately, cross-country comparable administrative micro data are simply not available. Alternatively one could use some exogenous time variation induced, for example, by a reform but, as already mentioned in Section 2, there has not been much reforming in social assistance programmes over the past years. In order to overcome these problems, comparable cross-country survey data from the European Community Household Panel (ECHP) are utilized in this paper. The ECHP is a panel of households and individuals from all the 15 pre-enlargement EU countries produced by Eurostat in cooperation with the member states statistical offices. The main advantage of this data source is the high level of cross-country comparability which is guaranteed by a common questionnaire and uniform sampling procedures.9 The ECHP started in 1994 and 8 waves of data have been released so far, covering the period from 1994 to 2001. Not all countries entered the survey at the same time and for three of them Germany, Luxembourg and the United Kingdom the original sample has been replaced after the first three waves with harmonised versions of household panels already been produced nationally: the German Socio-Economic Panel (GSOEP), the Luxembourgs Socio-Economic Panel (PSELL) and the British Household Panel Survey (BHPS). When possible data from the existing panels have been provided for the first three years too. For the purpose of this paper it is important to note that the ECHP includes information about unemployment benefit payments and social assistance transfers received both at the individual and at the household level. Moreover, it also contains retrospective information which allow to reconstruct employment/unemployment/inactivity monthly spells. 5. Empirical strategy and sample selection The theory of Section 3 suggests that unemployment insurance recipients who also receive some social assistance should be less sensitive to both the level and the duration of their benefits.
8 Boeri and Steiner (1998), Katz and Meyer (1990), Lancaster (1979), Meyer (1990), Moffitt (1985), Narendranathan and Stewart (1993b). Hunt (1995) is an exception and uses survey data for Germany. 9 See Peracchi (2002) for a detailed description of the ECHP.

782

M. Pellizzari / Labour Economics 13 (2006) 773798

However, simply comparing UI recipients who receive some social assistance to those who do not receive any additional benefits might not identify any causal effect. The selection into social assistance is, in fact, likely to be highly non-random. Ideally, one would like to rely on a natural experiment where UI recipients are randomly assigned to a treatment group of persons eligible for means-tested social assistance and a control group of persons who are either not eligible or eligible for universal, non means tested, benefits. UI recipients in the treatment group will be less concerned about their benefits being reduced or withdrawn, as they could be compensated by higher social assistance payments. Such interaction effect should not affect the control group. The ECHP allows to exploit variation in the institutional details of family benefits across European countries that mimic this ideal experiment. Specifically, in 5 countries Germany (until 1997), Greece, Italy, Portugal and Spain family cash benefits are means tested while in all the other EU-15 countries family assistance is universally assigned according to the number and the age of the children but regardless of household income. This paper focuses on the subsample of UI recipients who are male heads of households with children and assigns them to a treatment group of those who are eligible for means-tested family benefits and a control group of those who are either eligible for universal family assistance or not eligible altogether because the children are above the age limit. Note that here the terms treatment and control are used more loosely than in the standard evaluation literature as the variation used for identification only mimics an ideal experiment and the individuals are not really randomly assigned to the two groups. Also eligibility here has a particular meaning because some persons in the treatment group may actually fail the means-test. For the purpose of this study, however, it would be inappropriate to condition eligibility on income. In fact, what we are interested in is the behaviour of persons who are either receiving means-tested benefits or expect to receive them as soon as their income falls below a certain threshold. Thus, the treatment group includes all persons who live in countries that offer means-tested family benefits and have children below the age limit. As long as the generosity of family benefits does not affect the decision about ones country of residence and about the timing of childbirth, the selection into these treatment and control groups can be considered fairly exogenous. Note also that, since family benefits are assigned to the household and not to the individual, focusing on household heads avoids spurious correlation between adults in the same household who would be sharing the same benefit. Female headed households are also excluded since women, especially when heading families with children, have a very particular labour supply behaviour and often rely heavily on social assistance programmes such as minimum income schemes, for which the ECHP does not provide enough information. Table 2 reports the eligibility rules of family cash benefits in the countries included in this study and the distribution of the sample by treatment and control groups. Luxembourg, the Netherlands and Sweden have been excluded because information on retrospective employment status is lacking for these countries, while in Greece some important indicators of the labour market institutions are not available (see below). Eventually, the sample includes 2,703 unemployment spells experienced between January 1994 and December 2001 by male heads of households who received unemployment benefits at some point during their unemployment spells. The sample is divided into a treatment group of 944 persons who are potentially entitled to means-tested family assistance and 1,759 persons who are either entitled to universal assistance (because they live in countries where family benefits are not conditional on income) or are not eligible at all (because the children are above the age limit).

M. Pellizzari / Labour Economics 13 (2006) 773798 Table 2 Family cash benefits in Europe and composition of the sample Country Characteristics of Family cash Benefit Programmes Income Test Austria Belgium Denmark Finland France Germany (up to 1997) Ireland Italy Portugal Spain United Kingdom Total No No No No No Yes No Yes Yes Yes No Age Limita 19 to 26 18 to 25 18 16 18 to 20 16 16 to 19 18 15 to 24 18 16 to 19 Sample sizes (% in parentheses) All male heads of households with kids 291 111 149 275 187 199 277 149 80 890 95 2703 (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) Control 291 111 149 275 187 134 277 35 5 200 95 1759 (100.0) (100.0) (100.0) (100.0) (100.0) (67.3) (100.0) (23.5) (6.3) (22.5) (100.0) (65.1) Treatment 0 0 0 0 0 65 0 114 75 690 0 944 (0.0) (0.0) (0.0) (0.0) (0.0) (32.7) (0.0) (76.5) (93.8) (77.5) (0.0) (34.9)

783

Average family benefit (% of average UI benefit )b 0.85 0.56 0.40 0.59 0.49 0.53 0.25 0.76 0.15 0.31 0.82

Source: European Commission Missoc and OECD Tax and Benefit Database. a Benefits are paid until the child reaches this age limit, which is extended for children in training or higher education. b Ratio between average family benefits and average UI benefits for unemployed persons with 2 children between 1996 and 2000. The averages are taken considering only positive values. Source: ECHP.

The last column of Table 2 gives an indication of the generosity of family benefits relative to unemployment compensation. The numbers are computed as the ratio between the average family benefit for families with 2 children over the average UI compensation of unemployed persons over the period 1996 to 2000. According to these numbers child benefits often amount to a large fraction of unemployment insurance and their generosity varies largely across countries with average family benefits ranging from 15% to over 85% of average unemployment compensation. In order to avoid further issues related to recipiency of other benefits, the sample is also restricted according to age, health status and homeownership. In particular, the persons in the sample are aged between 25 and 50 and are thus very unlikely to receive any type of old age benefits. They are also in good health conditions and they report owning their houses for the entire duration of their unemployment spell, thus making them ineligible for both sickness and housing benefits. Some of the households in the sample might still be entitled to some minimum income schemes but unfortunately the ECHP does not contain enough information to identify them. Furthermore, left-censored spells have been dropped from the sample to avoid stock-sampling bias. Unemployment spells, thus, end into either employment or inactivity or are right-censored. Spells experienced by new entrants in the labour market are also excluded because these workers are rarely entitled to unemployment insurance. In the ECHP the sample-to-population ratios differ from country to country and the observations need to be weighted accordingly when pooled together across countries. The 2,703 observations represent 685,888 unemployment spells which are distributed across countries as shown in Fig. 3. This distribution is influenced by both the relative size of each country and the level of the unemployment rate experienced in each area (plotted in Fig. 3 along the dashed line and scaled on the right hand side vertical axis). This explains why Spain accounts for such a large fraction of the entire sample (approximately 30%). Unemployment in Spain was particularly high during the sample period,

784

M. Pellizzari / Labour Economics 13 (2006) 773798

300000

282122

200000

12

16 100000 95612 84908 46058 24877 15460 14847 9023 0 Austria Belgium Denmark Finland France Germany Ireland Italy Portugal Spain United Kingdom 50496 42879 19606 0

20

Fig. 3. Distribution of spells by country.

averaging around 19%, almost twice as much as the average for the other EU-15 countries. Despite this particular distribution of observations across countries, the empirical results of the next section are robust to the exclusion of Spain from the sample as well as to the introduction of country dummies (see Table 6). The identification strategy proposed here, that uses variation in institutions across countries, requires controlling for all other possible institutional differences across Europe. A lot of effort has been put in recent years, especially by the OECD, in producing quantitative measures of labour market institutions and a rather complete set of indicators is now available for statistical analyses. This paper fully exploits these indicators and the set of explanatory variables used for the estimation always includes the following yearly measures: the regional rate of unemployment, the regional rate of long-term unemployment, expenditure in active labour market programmes (as a % of GDP), the share of part-time dependent employment, the share of temporary dependent employment, the legal minimum wage (normalised by the median wage), the average effective tax rate, union density and the index of the strictness of employment protection legislation.10 Before proceeding with the empirical analysis two issues need to be briefly discussed. First, although the ECHP allows the identification of monthly unemployment spells, it reports all other variables, included the amount received in bunemployment related benefitsQ, only annually. A monthly measure of UI has been constructed by imputation. In fact, in all the countries the amounts and the duration of unemployment benefits are calculated on the basis of individual characteristics, most of which are easily available from the ECHP: previous employment records, age, family composition, et.. Combining these data with the rules of each countrys unemployment benefit system, a rather precise imputation of both the levels and the duration of
10

See OECD (2004a,b) for more details about these indicators.

Average un. rate

Frequency

M. Pellizzari / Labour Economics 13 (2006) 773798

785

payments can be obtained. Once controlling for individual characteristics, which are always included in the set of controls, variation in these imputed values is only generated by differences across countries (and to a smaller extent by reforms) and is therefore robust to potential endogeneity that might arise from unobservable characteristics such as preferences for work or the feeling of stigmatisation due to welfare use. A more detailed discussion of the imputation routine is in the Appendix. Second, the pre-unemployment wage is a crucial variable for the analysis of the duration of unemployment but in the ECHP only workers who are employed at the time of the interview report their current wages. Thus, this variable is missing for many observations and it is replaced by the average wage of full-time workers with the same level of education, age, experience, gender and region of residence. 6. The empirical analysis Table 3 shows the descriptive statistics for the entire sample and for the treatment and the control group separately. Encouragingly, the differences between the two groups are not particularly marked. Education and income are significantly higher in the control group, while age and experience are slightly lower. Interestingly, the composition of the households (size and number of children) is not particularly different between the two groups, suggesting that the selection into treatment and control is mostly due to cross-country differences in the meanstesting procedures. The fraction of families that report receiving family assistance is very similar in the two groups and so are UI replacement rates-the ratio between the UI benefit and the preunemployment wage-and durations. Important differences emerge for unemployment assistance. Most of the countries that offer unlimited unemployment assistance happen to also have universal family benefits, hence, 70% of the persons in the control group are potentially entitled to unlimited unemployment assistance versus only 8% among the treated. Finally, the average duration of unemployment is only marginally different, with the treated experiencing on average spells longer by less then 1 month. The differences in the regional and country variables are more marked, but this is not surprising, being the treatment and the control group selected mostly on the basis of the features of the countries welfare systems. This brief inspection of the descriptive statistics already suggests that there should be no systematic differences in the individual characteristics of the treated and the controls. In order to clear any further doubt, Table 4 reports the results of a simple probit regression where the dependent variable is equal to 1 if the individual is in the treatment group and zero otherwise. Most of the individual characteristics turn out to be not significant, apart from the number of children. A test of the joint significance of all the personal and family variables does not reject the null hypothesis of the coefficients being all zero. On the other hand, all the regional and country characteristics (apart from the regional long term unemployment rate and the share of temporary employment) are strongly significant. Once again this result is not surprising. The crucial identification assumption here is that, after controlling for a large set of institutional differences, cross-country variation in the administrative rules of the welfare system can safely be used to identify causal effects. Obviously, there still remains a possibility that the treated and the controls might significantly differ according to some unobserved characteristics (culture, attitudes towards family and work, etc.). However, since the variables of interest UI replacement rate and duration of UI entitlement have been imputed only on the basis of the countries regulations, endogeneity

786 Table 3 Descriptive statistics Variable

M. Pellizzari / Labour Economics 13 (2006) 773798

All male heads of households with kids Mean Std. Dev. (7.41) (8.93) (0.50) (0.47) (0.36) (1.32) (1.08) (3.02) (0.41) (0.23) (18.18) (0.50) (11.53) (9.65)

Control group Mean 35.88 17.63 0.38 0.42 0.20 4.24 1.03 9.17 0.21 0.60 20.62 0.70 10.67 10.65 79.93 93.63 6.37 1759 Std. Dev. (7.89) (9.24) (0.49) (0.49) (0.40) (1.30) (1.11) (1.79) (0.41) (0.24) (13.98) (0.46) (11.73) (8.69)

Treatment group Mean 38.61 22.47 0.72 0.20 0.08 4.46 1.57 6.47 0.22 0.61 21.05 0.08 11.85 11.43 85.81 96.77 3.23 944 Std. Dev. (6.72) (8.46) (0.45) (0.40) (0.27) (1.21) (0.88) (3.98) (0.42) (0.22) (17.54) (0.27) (11.17) (10.88)

Age Experiencea 1=primary education 1=secondary education 1=tertiary education Household size Number of chilren (b 15yo) Log Household income (PPP)b 1=Family cash benefit recipients UB replacement ratec Maximum entitlement of UB (in months)d 1=unlimited UAd Duration of all unemployment spells Duration of completed unemployment spells Fraction of completed spells ending into employment ending into inactivity Number of spells Countries and regions: Regional unemployment ratee Regional long-term unemployment ratee Expenditure in ALMP (%GDP)f Share of part-time employmentf Share of temporary employmentf Share of public employmentf Minimum wage/median wageg Average effective tax rateh Uniuon densityf Employment protection indexf
a b

37.22 20.08 0.52 0.32 0.15 4.35 1.31 8.15 0.22 0.64 20.89 0.44 11.13 10.61 81.98 95.27 4.73 2703

14.09 0.39 0.95 11.86 18.19 0.21 0.19 38.00 34.88 2.45

(7.49) (0.14) (0.47) (4.58) (11.13) (0.06) (0.21) (5.73) (22.22) (0.81)

11.68 0.41 1.01 14.43 14.14 0.21 0.23 39.47 30.64 2.37

(6.06) (0.16) (0.45) (4.62) (8.65) (0.06) (0.26) (6.09) (20.65) (0.82)

20.39 0.35 0.66 8.63 27.48 0.18 0.25 37.43 19.88 3.04

(7.53) (0.10) (0.26) (2.56) (10.38) (0.02) (0.14) (4.12) (7.80) (0.26)

Years since first entrance in the labour market. Income of all other household members. c Monthly UB amount/previous wage. UB amounts imputed on the basis of country regulations and personal characteristics (MISSOC, 19932001). d Imputed on the basis of country regulations and personal characteristics (MISSOC, 19942001). e Source: ECHP. f Source OECD. g Zero for the countries with no minimum wage. Source: OECD. h Income taxes plus social security contributions (employer and employee) over total gross labour costs. Source: OECD.

could only arise in the unlikely situation in which these unobserved characteristics influence the design of the welfare programmes considered here. Let us now move on to the analysis of the exit rates from unemployment for the two groups. Figs. 4 and 5 show the empirical Kaplan-Meier hazards11 for the entire sample and for the
The Kaplan-Meier hazard at time t is computed as the ratio of individuals who actually exit unemployment at time t over the number of all individuals who have been unemployed at least until t , i.e. all individuals who could have exited at time t .
11

M. Pellizzari / Labour Economics 13 (2006) 773798 Table 4 Selection into treatment and control group Dependent variable=1 if treated Personal and family characteristics: Age Experiencea 1 = secondary education 1 = tertiary education Household size Number of kids (b15yo) (log) hh incomeb 1 = zero hh income Country/region characteristics: Regional unemployment ratec Regional long-term unemployment ratec Expenditure in ALMP (%GDP)d Share of part-time employmentd Share of temporary employmentd Share of public employmentd Minimum wage (over the median wage)d Average tax ratee Union densityd Employment protection indexd Year dummies: 1995 1996 1997 1998 1999 2000 2001 Constant Observations Log-likelihood

787

UB recipients, male heads of households with kids 0.027 0.008 0.197 0.093 0.145 2.030** 0.052 0.272 (0.025) (0.022) (0.207) (0.244) (0.086) (0.234) (0.079) (0.843)

0.021* 0.076 4.460** 0.789** 0.074 0.101** 49.174** 0.212** 0.555** 3.541**

(0.011) (1.061) (1.132) (0.239) (0.049) (20.381) (8.191) (0.067) (0.110) (0.941)

0.229 (0.479) 0.601 (0.495) 0.478 (0.660) 0.453 (0.628) 1.619 (0.828) 1.689* (0.750) 0.956 (0.820) 3.560 (5.503) 2703 338.91

Robust standard errors (clustered by country and year) in parentheses. a Years since first entrance in the labour market. b Income of all other household members. c Source: ECHP. d Source: OECD. e Income taxes plus social security contributions (employer and employee) over total gross labour costs. Source: OECD. * Significant at 5%. ** Significant at 1%.

treatment and control group separately. The bars in these figures represent the distribution of imputed duration of unemployment benefit, scaled on the right hand side vertical axis. The figure for the entire sample (Fig. 4) shows the expected peak in the hazard around the time of UI exhaustion. In fact, the almost monotonic declining trend in the hazard rate is interrupted between the 12th and the 18th month, in correspondence with a series of important mass points in the distribution of UI duration at 12, 15, 16 and 18 months. The heavy concentration of UI durations at 24 months is due to Spain, where the majority of workers are entitled to two years of unemployment compensation. At this duration, however, the

788
0.08 0.07

M. Pellizzari / Labour Economics 13 (2006) 773798

30

25 0.06 20 0.05

0.04 0.03

15

10 0.02 5 0.01 0.00 2.00 4.00 6.00 8.00 10.00 12.00 14.00 16.00 18.00 20.00 22.00 24.00 26.00 28.00 30.00 32.00 34.00 36.00 0

month of unemployment
Males heads of households with children - - - - - - [95% Conf. Int.]

Fig. 4. Empirical hazard for unemployment-to-job transitions.

size of the sample is already dramatically reduced (more than 90% of exits occur within the first 2 years). Thus, the lack of a peak in the hazard around the 24th month is due to the small size of the sample at this duration and especially of the control group, which, as we shall see later on, is mainly responsible for the spikes. By looking at the same picture for the two groups (Fig. 5), it is already evident that these peaks are mostly due to the control group. The hazard rate for the treated, in fact, is a lot flatter, especially between the 12th and the 18th month. Moreover, the control group shows an earlier peak at the 6th month of unemployment, in correspondence with mass points in the distribution of UI duration at 6 and 7 months. The hazard rate of the treated is much more fluctuating at the same duration. These empirical hazards do not condition on any observable characteristics of the individuals, the household or the country, nonetheless, they already suggest that the persons in the control group respond more to the duration of their UI as their exits rates show more evident spikes around the mass points of the distribution of UI duration.
0.10 0.09 30

0.07 0.06

20

0.05 0.04 0.03 0.02 0.01 0.00 2.00 4.00 6.00 8.00

15

10

0 10.00 12.00 14.00 16.00 18.00 20.00 22.00 24.00 26.00 28.00 30.00 32.00 34.00 36.00

month of unemployment
CONTROL GROUP TREATMENT GROUP

Fig. 5. Empirical hazard for unemployment-to-job transtions by groups.

frequency of UB duration (%)

0.08

25

hazard

frequency of UB duration (%)

hazard

M. Pellizzari / Labour Economics 13 (2006) 773798 Table 5 Proportional hazard model for unemployment duration-exits into jobs All UB recipients UB recipients, male heads of households with kids [2] 0.318 (0.283) 0.150 (0.291) 0.987* (0.467) 0.409 (0.411) Control Treatment

789

Difference

[1] Replacement rates: UBUad (timeb=2) UBUd (2Ntimeb=6) UBUd (timeN6) UAUa 0.035 (0.111) 0.066 (0.065) 0.468** (0.132) 0.014 (0.156)

[3] 0.217 (0.307) 0.272 (0.302) 1.427** (0.526) 0.777 (0.922)

[4] 2.548** (0.930) 2.295** (0.848) 0.965 (0.991) 0.775 (0.596)

[3][4] * ** *

bMonths to UB exhaustionQ dummies: 1=less than 2 months 0.235 (0.136) 1=2 to 4 months 0.072 (0.146) 1=4 to 6 months 0.302** (0.121) bMonths to UA exhaustionQ dummies: 1=less than 2 months 0.112 (0.250) 1=2 to 4 months 0.159 (0.225) 1=4 to 6 months 1.235** (0.275) Personal and family characteristicsb yes Country/region characteristicsc yes Year dummies yes Baseline hazard (time, timeb, timec) Observations 106,268 Subjects 16,813 Log-likelihood 36,832.70

0.533 (0.448) 0.256 (0.253) 0.308 (0.223)

0.077 (0.319) 0.630* (0.254) 0.535* (0.238)

1.448 (0.867) 1.457 (0.794) 0.692 (0.399)

* ** **

0.892 (0.625) 0.745* (0.376) 1.049* (0.410) yes yes yes (time, timeb, timec) 24,080 2703 9146.90

0.442 (0.483) 0.876 (0.733) 1.010* (0.457) yes yes yes (time, timeb, timec) 14,372 1759 4813.56

2.451** (0.802) 1.160 (0.927) 0.002 (0.585) yes yes yes (time, timeb, timec) 9708 944 4177.92

**

Robust standard errors (clustered by country and year) in parentheses. a Monthly UB or UA amount/previous wage. UB and UA amounts imputed on the basis of country regulations and personal characteristics (MISSOC, 19932001). b Age, experience, education, health status, house ownership, household size, number of kids (b 15yo), (log) household income. See Table A1. c Regional unemployment rate, regional long-term unemployment rate, expenditure in ALMP (%GDP), share of parttime employment, share of temporary employment, share of public employment, minimum wage (/median wage), average tax rate. See Table A1. * Significant at 5%. ** Significant at 1%.

More convincing estimates that control for all observable personal, household and country characteristics are reported in Table 5, which shows the results from a proportional hazard model estimated for various samples. The econometric model is rather standard, with one minor

790

M. Pellizzari / Labour Economics 13 (2006) 773798

difficulty arising because in the ECHP unemployment durations are recorded in months i.e. in discrete intervals of time whereas the underlying process of job search occurs essentially in continuous time (workers can find a job at any moment within a month). Following the custom in the literature, let us assume that the hazard rate q(t |X i ,b ) of the underlying continuous process for individual i , i.e. the instant exit rate at time t conditional on duration until time t , can be written as the product of two parts: a baseline hazard that depends on duration only, h 0(t ), and a bproportional shifterQ, e Xib , that, according to each individuals characteristics X i , shifts the baseline up or down: #t jXi ; b h0 t d eXi b The discrete time analog of q(t |X i ,b ) for spell i that ends between month Ti and Ti +1, is usually written as: hTi jXi ; b PrfTi bt bTi 1jt NTi ; Xi ; bg 1 exp eXi ;b Hi Hi1 S Ti jXi ; b S Ti 1jXi ; b S Ti jXi ; b 5

where Hi 0 h0 udu. It is useful to apply to Eq. (5) the following transformation: log log1 hTi jXi ; b Xi b si 6

R Ti

Eq. (6) allows to recover for (a transformation of) the discrete time hazard h (Ti |X i ,b ) the separability property of its continuos-time analog. In fact, this transformation of h (Ti |X i ,b ), just like q(t |X i ,b ), can now be separated into two parts: one, H i =log[H (Ti )H (Ti +1)], that depends on the shape of the baseline hazard only, and another one, X i b , which depends only on individuals characteristics (possibly time-varying). According to Eq. (6), the discrete time hazard can then be rewritten as: hTi jXi ; b 1 exp eXi bsi 7 Then, it is possible to express the likelihood contributions of completed and uncompleted spells in terms of the discrete-time hazard and apply the transformation of Eq. (6) for the estimation:
i 1 Y T completed spells : PrfTi bt bTi 1jXi ; bg h Ti jXj ; b 1 hk jXi ; b

k 0 Ti Y k 0

uncompleted spells : Prft NTi jXi ; bg

1 hk jXi ; b

In our data, a spell can end either into employment or into inactivity. Assuming that the probabilities of ending in any of these two states are independent, Narendranathan and Stewart (1993a) showed that, by making the additional, harmless but greatly simplifying assumption, that exits can only occur at the boundaries of the interval (i.e. either at the beginning or at the end of each month), the correct hazard for exits into employment can be estimated by considering as censored all those spells that end into inactivity. This is also the approach taken here and the investigation of the determinants of exits into inactivity is left for future research. The set of controls included in the model contains all the relevant observable personal characteristics, country and region controls and year dummies. Table 5, however, only reports

M. Pellizzari / Labour Economics 13 (2006) 773798

791

the estimated coefficients on the variables of main interest the UI replacement rate and the btime to exhaustionQ dummies while the full set of results can be found in Table A1 in the Appendix. The estimates in Table 5 are obtained by imposing a third order polynomial form for 3 the baseline hazard, H i = Ti + T 2 i + T i , but the results are robust to different specifications. The first column in Table 5 shows the estimates obtained for the larger sample of all UI recipients, i.e. including males and females, all persons aged between 16 and 64 and all households with and without children. The purpose of this first set of results is to show that, using cross-country variation in the administrative rules of the unemployment insurance system, it is possible to replicate the usual results in the literature. In fact, higher UI replacement rates (UI U ) reduce the probability of finding a job, however, contrary to the findings of most other authors (see for example Narendranathan and Stewart (1993b)), this effect is significant only after the sixth month of unemployment.12 According to these estimates a 10 percentage point increase in the replacement rate leads to a reduction of about 4.7% in the hazard rate after the sixth month of unemployment. Once UI expires, some of the individuals in the sample are still eligible for unemployment assistance but its generosity (UA U) does not have a significant impact on the hazard rate. To capture the presence of peaks in the hazard rate around the time of benefit exhaustion,a set of bmonths to exhaustionQ dummies has been included in the model for both unemployment insurance and unemployment assistance. The coefficients on these dummies indicate a significant increase in the hazard rate only between 4 to 6 months from exhaustion of UI. Here the hazard rate is about 35% higher than at previous times. A similar but opposite effect ( 29%) is found for UA. The results for UA, however, should be interpreted with care. In fact, only in a few countries this scheme is limited and, even when it is limited, its duration is quite long and only a few persons in the sample reach the time of UA exhaustion. The following column in Table 5 repeats the same estimation for the sample of 2,703 UI recipients who are male heads of households with children. Results are generally confirmed, apart form the peak in the hazard at 4 to 6 months from UI exhaustion, which is no longer significant. The most interesting part of Table 5, however, is the comparison between column 3 and 4, where the estimation is repeated for the control and the treatment group separately. Following the previous discussion, one would expect the negative coefficient on the UI replacement rate to be significantly closer to zero for the treated as opposed to the controls. Moreover, the treatment group should show less pronounced spikes around the time of benefit exhaustion and we should then expect the coefficients on these dummies to be smaller. The estimates for the control group confirm and reinforce the results obtained for the entire sample, both in columns 1 and 2. A detectable effect of UI compensation appears only after the sixth month of unemployment, when a 10 percentage point increase in the replacement rate reduces the hazard by about 14%. The spike around the time of benefit exhaustion is also present: with 4 to 6 months left of benefit entitlement the hazard is about 1.7 times higher than during previous months and it goes to 1.9 times higher between 2 to 4 months. The negative impact of high UI replacement rates disappears for the treated, who instead show significantly higher exit rates associated with more generous UI in the first 6 months of unemployment and no effect afterwards. For this group there is no significant peak in the hazard as benefit exhaustion approaches. A jump in the hazard appears towards UA exhaustion but,
12 Most other papers actually find a negative effect of UI benefits on the exit rate at the beginning of the spell, which declines with duration and becomes insignificant after the first 56 months. However, this result is well documented only for the US and the UK and there still is very limited evidence for the other European countries.

792 Table 6 Robustness checks

M. Pellizzari / Labour Economics 13 (2006) 773798

Excluding Spain Control [1] Replacement rates: UBUbd (timeb=2) UBUd (2Ntimeb=6) UBUd (timeN6) UAUb 0.058 (0.355) 0.503 (0.323) 1.576** (0.530) 0.826 (0.976) Treatment [2] 3.574** (1.327) 3.487** (0.963) 1.344 (1.059) 0.950 (0.919) Diff. [1][2] ** ** *

Including country dummiesa Control [3] 0.083 (1.744) 0.501 (1.456) 1.836 (1.346) 0.884 (1.194) Treatment [4] 2.510* (1.030) 2.284* (0.959) 0.962 (1.094) 0.862 (0.622) Diff. [3][4]

bMonths to UB exhaustionQ dummies: 1=less than 2 months 0.105 (0.340) 1=2 to 4 months 0.650* (0.266) 1=4 to 6 months 0.500* (0.242) bMonths to UA exhaustionQ dummies: 1=less than 2 months 0.484 (0.517) 1=2 to 4 months 0.911 (0.709) 1=4 to 6 months 0.947* (0.457) Personal and family characteristicsc yes Country/region characteristicsd yes Country dummies no Year dummies yes Baseline hazard 3rd order polynomial of time Observations 12726 Subjects 1559 Log-likelihood 4043.03

0.860 (1.106) 1.399 (1.108) 0.495 (0.364)

0.028 (0.493) 0.212 (0.717) 1.278** (0.436)

1.026 (0.741) 1.291 (0.662) 0.590 (0.393)

**

1.395 (0.975) 0.924 (0.998) 2.309** (0.754) yes yes no yes 3rd order polynomial of time 2563 254 865.32

0.576 (1.107) 0.778 (0.999) 5.194** (1.558) yes yes yes yes 3rd order polynomial of time 2860 374 1326.63

2.088** (0.714) 1.015 (0.763) 0.063 (0.571) yes yes yes yes 3rd order polynomial of time 9699 944 4180.48

**

Robust standard errors (clustered by individual and year) in parentheses. a Only the countries with both treatment and control observations are included: Germany, Italy, Portugal, Spain. b Monthly UB or UA amount/previous wage. UB and UA amounts imputed on the basis of country regulations and personal characteristics (MISSOC, 19932001). c Age, experience, education, health status, house ownership, household size, number of kids (b 15yo), (log) household income. d Regional unemployment rate, regional long-term unemployment rate, expenditure in ALMP (%GDP), share of parttime employment, share of temporary employment, share of public employment, minimum wage (/median wage), average tax rate. * Significant at 5%. ** Significant at 1%.

M. Pellizzari / Labour Economics 13 (2006) 773798

793

once again, these estimates are likely to be driven by the few observations reaching the limit of UA entitlement. The positive effect of UI benefits for the treatment group goes beyond what we expected and is admittedly a rather puzzling result. One possible explanation is related to the availability of means-tested family benefits for this group. When receiving generous UI benefits, these families probably receive very little, possibly nothing, from family assistance. The higher exit rates associated with higher replacement rates could then be due to the overall benefits (UI plus family assistance) being lower, thus improving the incentives to work. This is only one possible explanation and a better understanding of this finding certainly requires further investigation. The last column of Table 5 reports the significance level of the difference between the estimated coefficients for the treated and the controls. Equality of the coefficients in the two groups is rejected in all interesting cases. As already noted in the previous section, the heavy concentration of observations in Spain might affect these results. Table 6 reports some robustness checks that show how the results are not fully dependent on Spain. The first two columns of Table 6, in fact, contain estimates from the same model of Table 5 produced excluding Spain from the sample. The results are qualitatively and quantitatively very similar. The negative effect of the UI replacement rate, which is present for the control group after the sixth month of unemployment, disappears for the treated and so do the peaks between 2 and 6 months from benefit exhaustion. The second two columns of Table 6 report the results of an additional robustness check. Some readers might not be fully convinced by the identification strategy adopted in Table 5, which is based on cross-country comparisons controlling for a large set of observable institutional characteristics of the countries. There might be other unobservable institutional and/or cultural differences across Europe that are correlated with some features of the unemployment insurance system. It is, then, interesting to repeat the estimation with the more standard approach of exploiting only within-country variation. This is obviously possible only for those countries with observations in both the treatment and the control groups, i.e. Germany, Italy, Portugal and Spain. Columns 3 and 4 of Table 6 reports the estimates from the same model of Table 5 where the sample has been restricted to these 4 countries and country dummies have been included in the set of explanatory variables. The identification in this case comes from comparing households that are entitled to meanstested family benefits with similar households that are not entitled because their children are above the age limit or because the rules have changed over time (like in Germany where family benefits have been made universal after 1997). Obviously, the variation available for identification is now more limited and some of the coefficients of interest are no longer significant. However, the negative coefficient on UB U for the controls turns positive for the treated, although none of them are significant. The controls still show a significant peak in the probability of finding a job between 4 to 6 months from UI exhaustion which disappears for the treated. Several other robustness checks have been performed which are not reported here. Neither the inclusion of renters nor the introduction of a parametric distribution of unobserved heterogeneity affect the results significantly. 7. Conclusions This paper investigates how interactions between unemployment insurance and social assistance affect the search behaviour of the unemployed. The theoretical framework presented

794

M. Pellizzari / Labour Economics 13 (2006) 773798

in Section 3 formalizes the idea that people are eventually interested in total payments (i.e. unemployment benefit and social assistance) and their time profile. Unemployed workers will react differently to changes in the rules of the UI system depending on what alternative or complementary welfare programmes are available. These theoretical predictions are tested using monthly labour market histories of European unemployed persons and exploiting variation in a particular feature of family benefits across Europe. While family benefits are universal in most countries, i.e. allocated according only to the number and the age of children but regardless of income, in 5 countries they are actually means tested. The comparison of similar unemployed persons who are subject to different family assistance schemes indicate significant differences in the elasticity of the probability of finding employment to both the generosity and the duration of the unemployment benefit. These results suggests that interactions between welfare programmes are, indeed, important. From a more general policy perspective, this paper highlights the need to design welfare reforms with very careful consideration for the overlappings between different programmes in the system: reducing the level or the duration of unemployment benefit may not be very effective in inducing the unemployed to search harder if they can easily shift into other social assistance programmes. This result is consistent with some recent studies that show how comprehensive reforms of labour market policies, even if politically harder to implement, are often more effective than reforming one programme at a time. Acknowledgements I would like to thank Tito Boeri, Marco Leonardi, Alan Manning, Steve Pischke, Federico Perali, Barbara Petrongolo, Joachim Wol .and an anonymous referee for useful comments and suggestions. This paper has been presented at the conference QChanging Social Policies for LowIncome Families and Less-Skilled Workers in the EU and the USQ at the National Poverty Centre, University of Michigan, where I received many useful comments, especially from my discussant, Jenny Hunt, and the conference organizer, Rebecca Blank. All errors are my own responsibility. Appendix A. Imputation of monthly unemployment benefit payments The imputation of monthly unemployment benefits requires two basic ingredients: a detailed description of the unemployment compensation system in all countries and years, and all relevant personal characteristics of the beneficiaries. The institutional features of all welfare programmes in the member countries of the European Union are systematically collected in the bMutual Information System on Social Protection in the Member States of the European UnionQ (MISSOC), a yearly publication of the European Commission that contains comparative descriptions of the rules and regulations of welfare programmes in the EU member states. Additional information can be extracted from institutional databases created by other research institutions, like the Fondazione Rodolfo Debenedetti (2004) and the CESIfo (2004) centre. As for the relevant personal characteristics, most of them are directly available from the ECHP. Combining these two sources institutional details from the MISSOC and other sources and personal characteristics from the ECHP it is possible to write imputation procedures for each country and year. The outcome of these procedures is a vector of imputed variables including the duration of benefit entitlement in months and the monthly sequence of payments for every individual in the sample.

M. Pellizzari / Labour Economics 13 (2006) 773798

Fig. A1. Comparison between imputed and reported annual unemployment benefits (national currencies). 795

796

M. Pellizzari / Labour Economics 13 (2006) 773798

Table A1 Proportional hazard model for unemployment duration-exits into jobs All UB recipients UB recipients, male Control heads of households with kids [2] (0.111) (0.065) (0.132) (0.156) 0.318 (0.283) 0.150 (0.291) 0.987* (0.467) 0.409 (0.411) [3] 0.217 0.272 1.427** 0.777 (0.307) (0.302) (0.526) (0.922) Treatment

[1] Replacement rates: UBUad (timeb=2) UBUd (2Ntimeb=6) UBUd (timeN6) UAUa 0.035 0.066 0.468** 0.014

[4] 2.548** (0.930) 2.295** (0.848) 0.965 (0.991) 0.775 (0.596)

bMonths to UB exhaustionQ dummies: 1=less than 2 months 0.235 (0.136) 1=2 to 4 months 0.072 (0.146) 1=4 to 6 months 0.302* (0.121) bMonths to UA exhaustionQ dummies: 1=less than 2 months 0.112 (0.250) 1=2 to 4 months 0.159 (0.225) 1=4 to 6 months 1.235** (0.275)

0.533 (0.448) 0.256 (0.253) 0.308 (0.223)

0.077 (0.319) 0.630* (0.254) 0.535* (0.238)

1.448 (0.867) 1.457 (0.794) 0.692 (0.399)

0.892 (0.625) 0.745* (0.376) 1.049* (0.410)

0.442 (0.483) 0.876 (0.733) 1.010* (0.457)

2.451** (0.802) 1.160 (0.927) 0.002 (0.585)

Personal and Family characteristics: 1=female 0.290** (0.054) Age 0.002 (0.008) 0.014 (0.013) Experienceb 0.002 (0.007) 0.019 (0.010) 1=secondary education 0.074 (0.061) 0.066 (0.107) 1=tertiary education 0.286** (0.081) 0.197 (0.182) 1=bad health 0.391** (0.126) 1=house owner 0.245** (0.057) Household size 0.083** (0.032) 0.016 (0.039) Number of kids (b15yo) 0.012 (0.039) 0.100 (0.066) (log) hh incomec 0.067* (0.033) 0.036 (0.053) 1=zero hh income 0.667* (0.298) 0.169 (0.445) Country/region characteristics: Regional unemployment rated Regional long-term unemployment rated Expenditure in ALMP (%GDP)e Share of part-time employmente Share of temporary employmente Share of public employmente Minimum wage (/median wage)e Average tax ratef Union densitye Employment protection indexe Year dummies Baseline hazard Observations Subjects Log-likelihood

0.011 0.017 0.061 0.247 0.077 0.213* 0.008 0.184 (0.056) (0.088) (0.089) (0.806) (0.013) (0.012) (0.130) (0.218)

0.002 0.007 0.038 0.106 0.102 0.106 0.122* 0.717 (0.055) (0.072) (0.053) (0.432) (0.029) (0.022) (0.186) (0.262)

0.011* (0.005) 0.201 (0.236) 0.175 (0.101) 0.054** (0.019) 0.014* (0.007) 3.468** (0.833) 0.996** (0.214) 0.022** (0.008) 0.014** (0.003) 0.158 (0.102) yes (time, timeb, timec) 106,268 16,813 36,832.70

0.022** (0.008) 0.072 (0.446) 0.219 (0.157) 0.480 (0.537) 0.072* (0.010) 0.364 (1.669) 0.005 (0.031) 0.039* (0.016) 0.011* (0.006) 0.452* (0.189) yes (time, timeb, timec) 24,080 2703 9146.90

0.005 (0.014) 0.255 (0.486) 0.068 (0.208) 0.063 (0.667) 0.000 (0.016) 1.763 (1.990) 0.822 (0.038) 0.014 (0.026) 0.011 (0.006) 0.181 (0.324) yes (time, timeb, timec) 14,372 1759 4813.56

0.029* (0.012) 0.781 (0.972) 0.049 (0.687) 0.204 (0.046) 2.986 (5.312) 32.830* (15.079) 0.031 (0.182) 0.202** (0.063) 0.134 (0.093) 1.856** (0.364) yes (time, timeb, timec) 9708 944 4177.92

M. Pellizzari / Labour Economics 13 (2006) 773798

797

Evaluating the goodness of this imputation is not easy. One possibility is to compare the cumulated annual amount of imputed benefits with annual income from bunemployment related benefitsQ, as recorded in the ECHP. This is done in Fig. A1 where imputed annual benefits are shown on the vertical axis against reported annual benefits on the horizontal axis. Although the two measures appear to be strongly correlated, there still is a large dispersion around the 45degree line. Italy is a particularly anomalous case: in this country there are many individuals who appear to be entitled but receive no benefits. This result is likely to be due to the high degree of discretionality of the Italian system. In fact, in this country unemployment benefits are highly differentiated by sector of industrial activity and firm size. Moreover, access to the most generous programmes (Cassa Integrazione Guadagni Ordinaria and Straordinaria ) is often subject to governmental approval. The high degree of discretionality necessarily leads to a poor imputation. It is unclear, however, whether one would actually expect the points in Fig. A1 to lie on the 45-degree line. The imputed measure of benefits is more an indicator of entitlement than receipt. It is a known fact that the degree of benefit take-up (i.e. the fraction of persons entitled to a benefit who actually claim it) varies largely across countries and is often very low (Hernanz et al., 2004). Moreover, even if imputed and reported benefits were actually measures of the same underlying quantities, they are probably both affected by measurement error. Survey respondents will tend to underreport the amounts they receive from public transfers, typically because of social stigma, one of the most popular explanations put forward for the observed low take-up rates. Moreover, the ECHP only asks the annual amount received in Qunemployment related benefitsQ, thus, it is potentially subject to recall bias and, possibly, misunderstanding about what should fall into this group of transfers. The imputation, on the other hand, does not take into account various elements of the system for which no information is easily available. For example, in many countries benefits might be reduced or withdrawn in case the recipient does not fulfill various job search requirements. These sanctions are normally applied with some discretion and it is extremely hard to identify the persons in our sample who might be subject to them. Another potential problem concerns the timing of the reforms. Changes in the unemployment compensation system could in principle be applied to the newly unemployed only, i.e. those who lost their jobs after the reform, or to all recipients. Information about these details of the reforms is very difficult to obtain. For simplicity, the imputation procedures used here assume that all changes always affect all recipients, regardless of when they entered unemployment.

Notes to Table A1: Robust standard errors (clustered by country and year) in parentheses. a Monthly UB or UA amount/previous wage. UB and UA amounts imputed on the basis of country regulations and personal characteristics (MISSOC, 19932001). b Years since first entrance in the labour market. c Income of all other household members. d Source: ECHP. e Source: OECD. f Income taxes plus social security contributions (employer and employee) over total gross labour costs. Source: OECD, 19952004. * Significant at 5%. ** Significant at 1%.

798

M. Pellizzari / Labour Economics 13 (2006) 773798

Finally, in the ECHP employment histories are perfectly known since the time the individual joined the survey but little is known about previous periods and some assumptions need to be made. Specifically, it has been assumed that individuals have always worked and paid contributions since the start of their first job.13 References
Atkinson, A., Micklewright, J., 1991. Unemployment compensation and labour market transitions: a critical review. Journal of Economic Literature 29, 1679 1727. Belot, M., van Ours, J.C., 2000. Does the recent success of some OECD countries in lowering their unemployment rates lie in the clever design of their labour market reforms? CEPR Discussion Paper: 2429. Bertola, G., Jimeno, J.F., Marimon, R., Pissarides, C., 2000. EU welfare systems and labour markets: diverse in the past, integrated in the future? In: Bertola, G., Boeri, T., Nicoletti, G. (Eds.), Welfare and Employment in a United Europe. MIT Press. Boeri, T., Steiner, V., 1998. bWait unemploymentQ in economies in transition: the case of Poland. Konjunkturpolitik 44 (3), 287 311. Cahuc, P., Lehmann, E., 2000. Should unemployment benefits decrease with the unemployment spell? Journal of Public Economics 77, 135 153. CESIfo, 2004. Ifos Database for Institutional Comparisons in Europe (DICE). www.cesifo.de. Coe, D.T., Snower, D.J., 1997. Policy complementarities: the case for fundamental labour market reform. CEPR Discussion Paper: 1585. Elmeskov, J., Martin, J.P., Scarpetta, S., 1998. Key lessons for labour market reforms: evidence from OECD countries experiences. Swedish Economic Policy Review 5, 205 252. Esping-Andersen, G., 1990. The Three Worlds of Welfare Capitalism. Policy Press, Cambridge. Ferrera, M., 1998. The four social europes: between universalism and selectivity. In: Rhodes, M., Meny, Y. (Eds.), The Future of European Welfare: A New Social Contract? Macmillan, Basingstoke. Fondazione Rodolfo Debenedetti, 2004. fRDB Social Reforms Database. www.frdb.org. Hernanz, V., Malherbet F., Pellizzari, M., 2004. Take-up of welfare benefits in OECD countries: a review of the evidence. OECD Social, Employment and Migration Working Papers no. 17. Hunt, J., 1995. The effect of unemployment compensation on unemployment duration in Germany. Journal of Labor Economics 13 (1), 88 120. Katz, L.F., Meyer, B.D., 1990. Unemployment insurance, recall expectations and unemployment outcomes. Quarterly Journal of Economics 105 (4), 973 1002. Lancaster, T., 1979. Econometric models for the duration of unemployment. Econometrica 47, 939 956. Meyer, B.D., 1990. Unemployment insurance and unemployment spells. Econometrica 58 (4), 757 782. MISSOC, 19942001. Social Protection in the Member States of the European Union. European Commission, Bruxelles. Moffitt, R., 1985. Unemployment insurance and the distribution of unemployment spells. Journal of Econometrics 28, 85 101. Mortensen, D., 1977. Unemployment insurance and job search decisions. Industrial and Labor Relations Review 30, 505 517. Narendranathan, W., Stewart, M.B., 1993a. Modelling the probability of leaving unemployment: competing risks models with flexible base-line hazards. Applied Statistics 42 (1), 63 83. Narendranathan, W., Stewart, M.B., 1993b. How does the benefit effect vary as unemployment spells lengthen? Journal of Applied Econometrics 8 (4), 361 381. Nickell, S., 1979. Estimating the probability of leaving unemployment. Econometrica 47 (5), 1249 1266. Nickell, S., van Ours, J.C., 2000. The Netherlands and the United Kingdom: a European unemployment miracle? Economic Policy 15 (30), 135 175. OECD, 19952004. Benefit Systems and Work Incentives. OECD, Paris. OECD, 2004a. OECD Employment Outlook. OECD, Paris. OECD, 2004b. Taxing Wages 2003/2004. OECD, Paris. Orszag, J.M., Snower, D.J., 1999. Anatomy of policy complementarities. IZA Discussion Paper No. 41. Peracchi, F., 2002. The European community household panel: a review. Empirical Economics 27, 63 90. This assumption is particularly problematic for the United Kingdom, where contribution conditions are quite complex.
13