an evaluation bart cockx and geert ridder march, 13 2000 · to stimulate t heir re -employment. ......
TRANSCRIPT
SOCIAL EMPLOYMENT OF WELFARE RECIPIENTS IN BELGIUM:
AN EVALUATION1
Bart Cockx 2and Geert Ridder 3
March, 13 2000
Abstract:
In Belgium, welfare agencies receive a subsidy to employ welfare recipients for a
period sufficiently long to entitle them to unemployment benefits. This work
experience program is called Social Employment (SE). We investigate the effect of SE
on the exit rate from welfare. We propose a grouping/IV estimator of the SE effect
that eliminates selection bias. The estimator is consistent, even if the selection into SE
depends on the average unobserved characteristics of welfare recipients in a region and
in a welfare duration interval. The empirical analysis suggests that there is creaming in
the selection process. Without correction for selectivity we find that SE reduces
welfare dependence, but after correction this conclusion is reversed. These results are
consistent with the adverse incentives faced by the welfare agencies.
Keywords: Welfare, selection bias, duration data, grouping, work experience
JEL classification numbers: C41, I38.
1We are grateful for the research assistance provided by Annick Persoon. This project has been financed by several grants: first, by a Human Capital and Mobility Fellowship which enabled the first author to visit the Tinbergen Institute in Amsterdam during the first semester of 1994; second, by a grant (n° 30.1.92.237) of the Belgian Fund of Collective and Fundamental Research (F.K.F.O); thirdly, by a grant "Actions de Recherches Concertées" n° 93/98-162 of the Ministry of Scientific Research of the Belgian French Speaking Community; finally, by a program, Poles d'Attraction inter-universitaires PAI P4/01, of the Belgian government. We also thank the Centre of Economic Studies of the Katholieke Universiteit Leuven, the Tinbergen Institute, and IRES of the Université Catholique de Louvain for the accommodation and the facilities provided. We thank the Minister of Health, Environment and Social Emancipation for access to the data and Frans Spinnewyn for his support in setting up this research project. We are grateful to Kenneth Y. Chay, Geert Dhaene, Thierry Magnac, Michel Mouchart, Bruno Van der Linden, and Marno Verbeek for their valuable comments. We are also indebted to, the managing editor, and especially to an anonymous referee whose detailed comments led to significant improvements of this article. The usual disclaimer applies. 2 IRES and Department of Economics, Université Catholique de Louvain, Place Montesquieu 3, B-1348 Louvain-la-Neuve, Belgium, phone: +32.10.47.34.39. E-mail: [email protected].
2
3 Department of Economics, The Johns Hopkins University, Baltimore, MD 21218, USA, phone +1.410.515.7614. E-mail: [email protected].
1
1. Introduction
In this study, we evaluate the effectiveness of a public sector work experience program,
the Belgian Social Employment (SE) program. Economists have argued that work
experience programs can reduce long-term unemployment by countering the
discouragement, the loss of work habits, and the skill deterioration induced by a long spell
of inactivity (cf. e.g. Layard, Nickell and Jackman 1991). A prerequisite for the
effectiveness of a work experience program is that the responsible agencies foster the re-
employment of inactive workers. In Belgium, the local welfare agencies (WA) that are in
charge of SE, not only run the program, but they also provide community services, as
services to nursing homes, hospitals and homes for the elderly, and meals and cleaning
services for households in need of support. We argue below that, as employers, the WA
have an incentive to select the most productive welfare recipients for SE and no incentive
to stimulate their re-employment.
Not only the program administrators but also the participants in SE face adverse
incentives that may offset the beneficial effect of improved work habits. First, participants
in SE earn the minimum wage, and thus have a significantly higher income than non-
participants. This increases the wage at which they are willing to accept a regular job.
Second, participants who remain in SE until the statutorily defined end-date are
automatically entitled to unemployment benefits that are higher than welfare allowances
and not means-tested. Finally, job search effort could be directly affected, if it is harder to
search for a regular job while participating in SE.
The fact that the WA are likely to assign more employable welfare recipients to
the program complicates the estimation of the SE effect on the exit rate from welfare,
because even without the program SE participants would have a better position on the
labor market. Moreover, the effect of a work experience program may be smaller for
more employable recipients (see Gueron and Pauly 1991). To evaluate the program we
need a procedure to correct for selection bias. Research by Ashenfelter and Card (1985),
LaLonde (1986), and Fraker and Maynard (1987) has cast doubt on the ability of non-
experimental methods to correct for selection bias. Estimates are sensitive to the model
specification and the estimation method. Recently, Dehejia and Wahba (1999) have
shown that in the National Supported Work (NSW) training program, studied before by
LaLonde (1986), the selection bias is mainly determined by observable variables.
2
Conditioning on these variables (or the probability of selection) removes the selection
bias1. This requires the availability of high quality and comparable data on both
participants and non-participants, in particular data on the pre-program labor market
histories of both groups. Participants and non-participants should also be recruited from
the same local labor markets (Heckman et al. 1998).
Unfortunately, the data used in our evaluation do not contain the variables that
make the participants and non-participants comparable. We therefore follow another
strand in the evaluation literature that identifies the program effect from exogenous
variation in program participation (see Meyer 1995 or Angrist and Krueger 1999 for a
survey). In particular, we use Angrist’s (1991) observation that if participation is selective
at the individual level, aggregation may reduce or even eliminate selectivity, if variation in
the participation rate at the aggregate level is exogenous2. The grouping estimator was
first proposed by Wald (1940) to deal with measurement error. Durbin (1954) showed
that the Wald estimator is an Instrumental Variable (IV) estimator.
In our study, we group the data by duration class and region. This yields a
consistent estimator of the effect of SE if the participation rate is not correlated with the
average characteristics of the welfare recipients grouped by duration interval and region3.
This assumption may not hold in our application. Because the WA may be aware of the
variation of the average unobserved characteristics with the elapsed duration of welfare,
there may be a correlation between the average unobserved characteristics by duration
interval and the participation fraction. However, we include indicators of the duration
intervals (and of the region4), and this eliminates the correlation, at the expense of a
reduced variation in the participation fraction that is now in deviation of the mean by
duration interval (and by region).
We estimate the effect of SE on the rate of leaving welfare. In other words, the
response variable is the welfare duration. Unless we impose restrictive distributional
1 Heckman et al. (1998) find that conditioning on a similar set of variables does not remove all of the bias in an evaluation of the JTPA program and propose a conditional (on the propensity score) differences-in-differences method to remove any remaining selection on bias. 2 Angrist (1991) uses this approach to deal with the endogenous wage in a labor supply equation. He aggregates waves in the PSID to the national level and this aggregation replaces individual wages by average wages. 3 Note that this is perfectly consistent with selectivity at the individual level. To see this, assume that there are two types of welfare recipients and that the type with favorable characteristics is more likely to be selected for SE. Such a composition of welfare recipients within a group is irrelevant at the group level. 4 This eliminates the potential correlation between the participation fraction and the average unobserved "quality" of welfare recipients in a region.
3
assumptions and censoring is not important, an IV estimator is not consistent in models
for duration data5. Grouping solves this problem, since we can then base estimation on the
Minimum Chi-Square (MCS) method. The MCS method justifies estimation on the
linearized duration model (Cockx, 1997). Without further assumptions on the variation of
the program effect in the population, the probability limit of an IV estimator is not equal
to the average effect of the program. In general, an IV estimator gives the effect of SE for
marginal participants6 (Imbens and Angrist, 1994). The IV estimator gives the usual
average program effect for participants, if the program effect does not vary among
participants or if the variation in the program effect does not influence the decision to
participate in the program (Heckman and Robb 1985, p.196, Heckman and Smith 1996,
p.59-68, Heckman 1997). We test the latter assumption and show that it cannot be
rejected for our data.
Krueger (1990), Meyer (1989), and Meyer, Viscusi and Durbin (1990) analyzed
duration data that were generated by a natural experiment. Their approach cannot be used
in the evaluation of SE. First, because of censoring, restrictive distributional assumptions
must be imposed (see Krueger 1990, p.14). Second, their approach cannot account for
time-varying explanatory variables. They assume that individuals are selected into the
program prior to entry into the state of interest. WA, however, employ and therefore
select welfare recipients at some instant during the welfare spell. The participation
indicator is therefore time varying.
In the following section we discuss the main features of the institutional setting.
We compare it to the institutional setting of work experience programs in other countries
and argue that adverse incentives for administrators of the kind discussed in this paper can
explain the ineffectiveness of these programs in other countries as well. In Section 3 we
describe the data and provide a benchmark of the program effect on the basis of a
matching estimator. Section 4 presents the statistical model, justifies the identifying
assumption and presents the baseline results. Section 5 discusses the robustness of the
results. The final section contains the conclusions.
5 Only if one restricts the duration distribution to be Weibull and if there is no right censoring, duration data can be analyzed in the linear regression framework (cf. Lancaster 1990, p. 220-221). 6 This is the Local Average Treatment Effect (LATE).
4
2. Social Employment and its Relevance for other Work-Experience Programs
2.1. Social Employment
In Belgium the welfare system is a safety net for those who are not covered by social
insurance, because of insufficient work experience, delay in administrative procedures, or
because they have been punished by the social insurance administration. Individuals, who
pass a means test, can claim a supplement to their income up to the legally determined
Minimum Income Guarantee (MIG) at the welfare agency (WA) of their municipality.
During the 1987-90 period on average 52,659 adults and 31,672 children received the
MIG at any given time (Cockx 1992, p.36). This is roughly 0.8% of the Belgian
population. To compare, in January 1989, 389,672 individuals received Unemployment
Insurance benefits (Cockx 1992, p.39). The number of welfare recipients is relatively
small, because unemployment insurance benefits have an indefinite duration7. After a
waiting period of six months, even school-leavers are entitled to unemployment insurance
benefits.
Legislation stipulates that WA may employ welfare recipients for a period that is
sufficiently long8 to entitle them to unemployment benefits. This type of employment is
called Social Employment (SE). During the 1987-90 period on average 1282 individuals
received SE at each instant of time. This is 2.4% of the adult welfare population.
The possibility to employ welfare recipients for this purpose brought legislation
into agreement with the principle that social assistance offers relief only if the main social
insurance system fails to do so. The national and regional9 authorities have gradually
widened the scope of SE as a means of enhancing the integration of the poor. This is
reflected in increased financial support to WA offering SE. The regional authorities
subsidize SE in Flanders since 1983 and in the Walloon provinces since 1989. The central
7 There is one exception to this rule. Cohabitants who are not the head of a family can lose entitlement to unemployment benefits on grounds of "excessive" duration. 8 Unemployment benefits are only paid to workers who, within a specified period prior to their claim, have been employed for at least 75 days if younger than 18, and up to 600 days if older than 50 (Van Langendonck 1991, p. 450). 9 Readers who are familiar with the Belgian institutional setting will notice that we use regional in a loose sense. As such, we can divide Belgium into 4 regions: Flanders, referring to the Flemish (Dutch) speaking community in the North; the Walloon provinces, referring to the French speaking community in the South; Brussels, referring to the region of the bilingual community of the capital city; the German speaking community in the East. Given the marginal importance of the latter region, we ignore it in the sequel.
5
authorities finance 50% of the MIG paid by the local WA. Since 1985 the WA remain
entitled to this subsidy if it socially employs the welfare recipient, and since January 1993
the subsidy is up to 100% of the MIG.
It is important to realize that the WA receive this financial support and not the
employed welfare recipients. SE participants are paid the minimum wage, so that they
earn at most as much as in an alternative job. Incentives for a welfare recipient to
participate in SE are therefore, either the (possibly higher) unemployment benefits to
which he/she is entitled after SE, or the (subjective) benefits of the work experience.
These incentives are not the sole determinants of participation, for the WA can mandate
participation, as the receipt of welfare benefits is conditional on the willingness to work.
Van de Velde (1989) demonstrates that SE is concentrated in the community
services that the WA offer to the general public. These community services consist of
domestic services to households, such as care, meal provision, cleaning, and to
institutions, i.e. nursing homes, hospitals, homes for the elderly, etc. These jobs require
few qualifications. Women are typically (95%) employed in the domestic services
provided by the WA. This work involves cleaning, cooking, washing and ironing. Men are
required to do all kinds of odd jobs (37%), to maintain roads or to plant vegetation
(13%), to help in the kitchen (10%), to do administrative work (10%) (see Van de Velde,
p.55).
WA can save on outlays by employing welfare recipients. An example clarifies this
point10. Consider a welfare recipient living alone in Flanders in January 1989. The net
monthly cost of SE for this individual is the cost of the minimum wage for the employer
(=49,354 Belgian Francs) minus the central authorities’ subsidy of 50% of the MIG
(=8,152 B.F.) and the regional authorities’ subsidy (=30,500 B.F.): 10,702 B.F.
Moreover, during the initial period of SE, in which the participant would otherwise
depend on welfare, the WA no longer needs to pay welfare benefits and therefore saves at
least11 50% of the MIG (=8,152 B.F.) minus the regional subsidy for welfare (=1,141
B.F.), that is 7,011 B.F. This implies that employing a worker, who initially produces
goods and services with a value of more than 3,691 B.F. per month and later on of more
than 10,702 B.F. per month, is financially attractive to the WA. The WA needs not worry
10 The figures in the example are taken from an internal document of a WA to which we had access. 11 WA's can offer a supplement to the MIG, but they do not receive a subsidy for this supplement.
6
about costs after SE, because after SE welfare recipients are entitled to unemployment
benefits, and hence will not depend financially on the WA. We conclude that SE can result
in savings for WA. These savings are larger if the WA select the most productive
recipients for participation in the program.
This observation is not only valid for the specific example. However, the size of
the savings differs with the level of the MIG12, the time period and the region. For
instance, Brussels does not provide a specific subsidy for SE and due to the growing
popularity of the program in Flanders, the subsidy per participant has declined. In the case
of a lower subsidy the WA can still save on expenses by being more selective in the choice
of participants. The specific subsidy rules provide no incentive to the WA to use SE to
integrate welfare recipients in the labor market. Efforts to this effect will only increase its
expenditures. Consequently, the findings of Van de Velde 1990, p. iv, that SE was hardly
accompanied by training and assistance and that the WA encouraged only 6% of the
participants to apply for another job, are not surprising.
2.2. Comparison with Other Work Experience Programs
In other European countries, program administrators face similar adverse incentives. For
instance, in Germany the dramatic rise in unemployment since the early nineties has
increased the number of welfare recipients substantially. In response, several
municipalities, like Leipzig, Frankfurt and Lübeck, now offer temporary employment to
welfare recipients in municipal job-creation companies, the so-called
"Beschäftigungsgesellschaften". Feist and Schöb (1999) argue that municipalities can
benefit from these initiatives for reasons similar to those in Belgium. First, the job-
creation companies provide local public goods. Second, after being employed for a year
the participants are entitled to unemployment insurance benefits of the "Bundesanstalt für
Arbeit", a federal authority. Consequently, the German municipalities have similar adverse
incentives as the WA in Belgium13.
12 The level of the MIG differs by household type. Legislation distinguishes between singles living alone, singles living with dependent children, cohabiting individuals and cohabiting married couples. 13 We are not aware of any study evaluating this program. Eichler and Lechner (1998) find that public employment programs reduce the unemployment risk for participants in the East German State of Sachsen-Anhalt. However, this study could not distinguish employment in the municipal Beschäfti-gungsgesellschaften from other public employment programs.
7
The British Youth Training Scheme (YTS) of the second half of the eighties
provides an example in which private employers face adverse incentives. Participants in
the YTS are placed in a job where they can acquire work experience. Trainees were paid
an allowance (set slightly above the unemployment benefits level) and firms could employ
trainees without incurring any further wage costs. As the WA in Belgium, firms could
therefore gain from cheap labor. Moreover, firms did not have explicit incentives to
increase the employability of trainees. The work experience was supplemented by
classroom training, but one can have serious doubts on the quality of this training
component, as this was a constant theme in discussions of reform. Dolton et al. (1994)
find that ex-trainees obtain jobs at a slower rate than non-trainees even when the time
spent in YTS is excluded14.
Work experience programs that place participants in jobs that do not require much
training induce (public or private) employers to hang on to the participants for the subsidy
period. After this period, employers, who have not invested in the participants, have little
incentive to keep them. For example, Edin and Holmlund (1990) report that participation
in temporary jobs in the public sector in Sweden decreases the re-employment rate of
young and displaced workers15. Bonnal et al. (1997) find that in France lower educated
young workers participating in public employment programs do not have a higher
transition rate from unemployment to employment after program participation. Moreover,
participation decreases this transition rate for workers with a professional or technical
diploma16.
The SE program is, at first sight, comparable to the workfare programs in the US,
e.g. aimed at women who receive Aid to Families with Dependent Children (AFDC) (see
Gueron 1990, Gueron and Pauly 1991, Moffitt 1992 and LaLonde 1995). Typically,
welfare applicants or recipients are required to look for a job for two to four weeks. If
they do not find a job, participants may be required to work for up to three months
without pay (workfare). Usually, these are entry-level jobs in public or nonprofit
14 Female (but not male) ex-trainees obtain ‘good’ jobs at a faster rate than non-trainees when time spent on YTS is excluded. 15 The authors do find that program participation reduces the length of subsequent unemployment spells. However, the study imposes strong parametric (Weibull) restrictions on the baseline hazard and does not explicitly account for selection on unobservables, which may bias the results. 16 According to the authors, the latter finding suggests that participation in such programs may signal low job performance
8
agencies, e.g. maintenance, clerical, park upkeep, or human services functions. Monthly
working hours are equal to the welfare grant divided by the hourly minimum wage rate.
The US workfare programs, however, differ from SE in important respects. First,
both participants and program administrators face different incentives in the two
programs. Second, most US programs are not pure work experience programs, but also
include job-search assistance, education or training components. A welfare recipient on
workfare earns no more than the welfare benefits. The participant is therefore worse off
by participating, because his/her time costs are not compensated. Contrary to SE,
workfare therefore makes it more attractive to leave the welfare rolls. The US welfare
and workfare programs are administered by the states, but subsidized by the Federal
Government. In contrast to the SE-program, the funding rules of the US Federal
Government, as determined by the Family Support Act (FSA) of 1988, provide incentives
for enhancing the employability of program participants (see Gueron and Pauly 1991, pp.
55-59 for a more extensive discussion). The rules for matching grants contain provisions
that prevent states from using workfare as cheap labor. Finally, contrary to SE in
Belgium, welfare recipients who do not exit to a regular job, return to welfare, and
increase the state's welfare expenditures.
Despite these differences, the finding of this paper that SE increases rather than
reduces welfare dependence is not in conflict with the evaluations of US workfare
programs. First, there is limited and inconsistent evidence on the effect of pure unpaid
work experience: Studies have problems in disentangling the independent effects of each
of a sequence of program components. Nevertheless, one finding is robust. No effect is
found, when resources are so limited that staff can provide almost no direct assistance. As
participants in SE hardly receive any assistance, this is in line with our results.
A noteworthy example of a successful work experience program is the National
Supported Work (NSW) demonstration targeted at extremely disadvantaged welfare
recipients17. In contrast with other work experience programs, NSW offered highly
structured, full-time work experience positions for up to 18 months with close
supervision and peer group support (Gueron and Pauly 1991, p.101). Hollister, Kemper
and Maynard (1984) and Grossman, Maynard and Roberts (1985) give evidence that
participation in NSW led to a sizable earnings increase for former AFDC recipients, and
17 NSW was not only offered to AFDC women, but also to disadvantaged youth. Evaluation studies do not find
9
that it significantly reduced the probability of being on welfare. Couch (1992) shows that
the initial positive earnings effect is even maintained 8 years after the program.
The experience with NSW is similar to that of SE pilot projects in Belgium
evaluated by Wouters, Van Meensel and Nicaise (1994). In these pilot projects work
experience is complemented by training and intensive assistance. Wouters, Van Meensel
and Nicaise find that these projects increase the employment rates, even in the long run.
3. The Data We analyze administrative data on recipients of the MIG that have been collected since
June 1987 by the Ministry of Social Integration. From these data we calculate the length
of welfare and SE spells (in days18) for all MIG recipients, who claim benefits after June
1st 1987. In the analysis we consider only spells that started in the period June 1987 -
July 1990. On July 1st 1990 additional measures to stimulate the re-employment of
welfare recipients were introduced. We only consider welfare recipients who were
younger than fifty at the start of the welfare spell.
Our data do not allow us to distinguish between exits to regular employment and
other states. Hence, a positive effect of SE on the total exit rate from welfare need not
correspond to a larger transition rate to regular employment. For example, Dehaes (1994,
p.119 and 128) found that 26% of the exits involve a change in labor market status19,
35% involve entitlement to other social benefits, 19% a move to another municipality,
11% a change in the family composition20, 4% a withdrawal of benefits by the WA, and
finally 4% other reasons.
Table 1 summarizes some characteristics of the population under consideration.
As the data were collected for administrative purposes and not for analysis, the
information is limited. The first column refers to all welfare spells, the second to those
spells in which some time was spent in SE.
INSERT TABLE 1 APPROXIMATELY HERE
positive short- or long-term effects for the latter group (cf. Hollister et al. 1984 and Couch 1992). 18 In Table 2 we express statistics in months by dividing the number of days by 30.4375. 19 This refers to events such as re-employment, a transition from a part-time to a full-time job, an increase of self-employment earnings, etc. 20 This refers to events such as marriage to or cohabitation with a partner with a sufficiently high income, moving into the house of parents or children with sufficient means, etc.
10
In contrast to Garcia (1990) we do not find an age difference between SE
participants and non-participants. This is a consequence of the restriction to welfare
recipients younger than fifty. On the other hand, the over-representation of women
among the socially employed is confirmed. In addition, individuals who are legally
married but living alone are over-represented among the participants in SE. Furthermore,
participants are more likely to have dependent children. The rate of participation in SE is
larger in small municipalities and in Flanders.
INSERT FIGURE 1 APPROXIMATELY HERE
The median welfare spell is 5 months. This is somewhat smaller than the number
reported in Cockx (1997). The difference can be attributed to the exclusion of recipients
older than fifty. Welfare spells of participants in SE are much longer. The median
duration for this group is 13 months. Moreover, as can be seen from Figure 1, the
survival fraction of participants in SE is uniformly higher than that of non-participants.
From this we cannot conclude, that participation in SE increases welfare dependence.
First, as reported in Table 2, the median duration until selection into SE is 6 months, and
this increases the survival fraction of participants. However, even if we subtract the time
until selection, the median duration of participants (=13-6=7 months) exceeds the median
duration of all welfare spells. Second, if transitions from welfare exhibit negative duration
dependence, the difference of the medians understates the effect, because the exit rates of
participants are lower due to delayed entry in SE. Third, even at an equal elapsed
duration, participants and non-participants differ in both observed and unobserved
characteristics. This induces the well-known selection bias. Note that, if the WA indeed
cream, i.e. select the best welfare recipients, the difference between the two survivor
functions in Figure 1 under- rather than overestimates the impact of SE.
INSERT TABLE 2 APPROXIMATELY HERE The finding that the median duration until selection into SE is larger than the
median welfare duration seems to be evidence against the creaming hypothesis of Section
2. However, WA will not offer SE to welfare recipients who are likely to leave welfare
shortly, because of the fixed costs associated with SE. As a consequence, SE is only
11
offered to welfare recipients who have been on welfare for some period. The creaming
hypothesis applies to this subgroup of the welfare recipients, which is a negative selection
from all welfare recipients.
Some of the biases associated with a simple comparison of participants and non-
participants in SE are reduced if we consider a matched comparison. Despite recent
pessimistic evidence of the validity of matching methods (cf. Friedlander and Robins
1995), the interest in their application has recently revived. Dehejia and Wahba (1995a,b),
for instance, reproduce, by matching on the probability of selection or propensity score,
the treatment effect estimated by a randomized experimental design. This requires that the
treatment assignment is independent of the outcome conditionally on observed
variables21. The administrative data used in our study are clearly not rich enough. For this
reason, the matching estimator gives only a benchmark estimate of the effect of SE on
welfare duration. Rather surprisingly, however, the estimate found by this method are
consistent with those found by the grouping/IV estimator proposed in Section 5 below.
Participants in SE are matched with non-participants on the basis of the
characteristics reported in Table 1. We only retain exact matches22. This explains the
significantly lower number of individuals in the matched sample. In the matched sample,
welfare recipients living in small municipalities are underrepresented. This should not
come as a surprise, as the number of welfare recipients is much smaller in these
municipalities, and consequently the probability of finding an exact match. For the same
reason foreigners, widow(er)s, women, families with dependent children, married
persons, and persons living in Flanders are underrepresented This feature can bias the
matching estimator of SE to the extent that the treatment effect varies systematically with
these characteristics. In Section 5 we argue that this is unlikely to be the case.
INSERT FIGURE 2 APPROXIMATELY HERE
In order to account for the bias induced by the elapsed duration until selection, we
match only to non-participants who are still on the rolls after the selection time of the SE
participant. Moreover, as SE can only produce effects from the moment of enrollment,
21 Heckman et al. (1998), using a rich dataset, reject the hypothesis of conditional independence given only observables in favor of the hypothesis that unobservables are important for treatment assignment. 22 Hence, we do not match on the propensity score. The matching procedure is nonparametric.
12
we subtract this elapsed duration from the welfare duration in both groups. In Figure 2
the survival functions of the welfare durations are plotted for participants and non-
participants23. This shows that participants in SE are more likely to be on welfare for up
to one year. Afterwards, they seem to be more likely to have left. This pattern is
consistent with (i) a lower transition rate out of welfare of SE participants (in comparison
to individuals who have been on welfare for the same time period) (ii) creaming of SE
participants. Initially the lower transition rate dominates, but over time the positive
selection into SE diminishes this effect among the survivors, and even reverses the
effect24.
4. Dealing with Bias: Discrete Durations, Competing Risks and Unobservables Nonrandom selection of participants leads to differences in the (average) characteristics
of participants and non-participants. Part of this is due to the delay in selection into SE.
The matching procedure of the previous section makes the two groups comparable and
eliminates the bias due to selection on observables and the delay in participation. It does
not eliminate bias due to selection on unobservable characteristics, which as argued is
potentially important in SE.
In this section we introduce a model for discrete duration data with competing
risks that deals with both types of bias simultaneously. The competing risks specification
allows for an arbitrary starting time in SE. In a first step, we control for even fewer
observable characteristics then in the matching procedure. If our hypothesis of creaming
in the selection process holds, this benchmark estimator of the effect of SE on the
transition rate hazard out of welfare will be even more biased upwards. This is confirmed
below. In a second step, we present the grouping/IV estimator that controls for selection
on both observables and unobservables.
We group the observed durations by duration interval and explanatory variables.
The advantage of grouping is that we do not need parametric assumptions on the joint
distribution of unobservables that affect both the participation decision and the transition
rate out of welfare. Grouped durations can be modeled in continuous or in discrete time.
23 As a consequence of right censoring, we cannot simply take differences of durations between two matched individuals as a measure of the treatment effect. 24 We also matched the SE participants to non-participants in WA’s that do not have welfare recipient in SE. This resulted in a somewhat larger matched sample. The results are the same as in Figure 2.
13
Following Prentice and Gloeckler (1978) and Cockx (1997) we use time-aggregation in a
continuous time proportional hazards model. In the resulting model the parameters are
invariant to the grouping. After a simple transformation of the dependent variable, the
model can be written as a (heteroskedastic) linear regression model.
4.1. Time-Varying Covariates and Competing Risks with Grouped Duration Data
Let the duration data be grouped into K+1 intervals: [ t
0, t
1), ...., [ t
k −1, t
k ), ..., [ tK −1,tK),
[ tK , tK+1 ) of (possibly unequal) length ∆ k = tk − tk −1 with t0 = 0 , tK +1 = ∞ . In the simplest
grouped duration model it is assumed (Prentice and Gloeckler 1979) that the (base-line)
hazard λ is constant on these intervals
(1) kk ttt <≤= 1-k , (t) λλ Let d (t ) be the indicator of participation in SE at t. In this section, we assume that the
process of participation is exogenous, an assumption that will be relaxed later. Note that
the time-scale is time spent on welfare. This is the time-scale used in the rest of this
paper.
If we assume that the effect of SE on the exit rate is independent of t, then the exit
rate h can be written as
(2) )())(exp()( ttdth λα= This means that participation in an individual's SE status at time t leads to a proportional
change in the baseline hazard that is equal to α . Let Pk be the conditional probability of
exit in the k-th duration interval [ t k −1 , tk ) given that the individual is still on welfare at t k −1 .
With (1) and (2) we have
(3) ( )
−−= ∫
−
k
k
t
t
kk dssdP1
)(expexp1 αλ
This probability depends on the evolution of d in the k-th duration interval. With grouped
duration data this evolution is unknown. We could assume that exit from SE is only
possible by leaving welfare and that all selection into SE occurs at time t k −1 . Then we
14
only need to know the SE-status at the start of the k-th interval to determine Pk , because
the SE-indicator is constant on that interval. However, these assumptions are
counterfactual, because individuals may leave SE and remain on welfare, and they may be
selected into SE at any point during the welfare spell. For that reason we make the more
reasonable assumption that there is at most one transition in a time interval.
Let θ0 (t) , θ1 (t ) be the transition intensities into and out of SE at welfare duration
t. As in (1) we assume that these intensities are piece-wise constant, and θ0k , θ1k are the
values on the k-th interval. We condition on the SE-status at the beginning of the k-th
interval. Let PkWSE ,OW be the conditional probability of leaving welfare in the k-th duration
interval while being in SE until exit from welfare, given that the individual is on welfare
and in SE at t k −125. Analogously, let Pk
WNSE ,OW be the conditional probability of leaving
welfare in the k-th duration interval while not being in SE until exit from welfare, given
that the individual is on welfare and not in SE at t k −1 . These probabilities are transition
probabilities in the k-th interval from the state welfare with SE (WSE) to the state out of
welfare (OW) and from the state welfare without SE (WNSE) to the state out-of-welfare
(OW). With (1) and (2) we have
(4)
( ) ( )
( )[ ]kkkkkk
k
kkkOWWSE
k dsssPk
∆−∆−−+
=
−−= ∫∆
11
1
0
,
)exp(exp1)exp(
)exp(
exp)exp(exp)exp(
θαλθαλ
αλ
θαλαλ
In a similar way we obtain
(5) ( )[ ]kkkkkk
kWNSEWSEkP ∆−∆−−
+= 1
1
1, )exp(exp1)exp(
θαλθαλ
θ
The conditional probability of all other events is the complement of the sum of these
probabilities.
For welfare recipients who are not in SE at the start of the k-th interval we have
25 Combining the event of leaving welfare with the event of being in SE until the transition, implies that we only consider the first transition within the interval. With the states defined below, we assume that there is only one transition in a time interval.
15
(6) ( )[ ]kkkkkk
kOWWNSEkP ∆−∆−−
+= 0
0
, exp1 θλθλ
λ
and
(7) ( )[ ]kkkkkk
kWSEWNSEkP ∆−∆−−
+= 0
0
0, exp1 θλθλ
θ
The expressions (4)-(7) show that our model for selection into SE is equivalent to
two competing risks models with piece-wise constant transition intensities: one for the
origin state WSE and destination states OW and WNSE, and one for the origin state
WNSE and destination states OW and WSE.
4.2. Modeling Grouped Durations with Competing Risks
It will be convenient to attach numbers to the three states introduced in Section 4.1
1. Welfare without SE
2. Welfare with SE
3. Out of welfare
In the sequel we identify the states by these numbers. We are interested in comparing the
transition intensity from state 1 to state 3 with that from state 2 to state 3. This identifies
the effect of SE on the exit rate from welfare. State 3 is assumed to be an absorbing state.
Our data do not contain multiple welfare spells.
Origin/destination states are indicated by superscripts u and v, and duration
intervals and types of individuals by subscripts k and m, respectively. We make the
following assumptions on the transition intensities between the states
(8) ( ) uvuh uv
kmuvv
mvk
uvkm ≠==+++= 3,2,1 ,2,1 ,log εαβγ
There are K+1 duration intervals, where the last interval is open, and M types. The
duration effect γ kv and the type effect v
mβ are assumed to depend only on the destination
state, i.e. they are the same for SE-participants and non-participants. We normalize the
type effects and set β1
v = 0 . The parameter α uv gives the effect of SE on the exit rate
from welfare. The effect is assumed to be independent of the type m. By definition
16
α12 = α 21 = 0 , and by normalization α13 = 0 . Note that α 23 is the parameter of interest.
Finally, the ε kmuv are unobserved variables, which are similar to the specification errors
introduced by Amemiya and Nold (1975) in their grouped logit model (see also Parks
1980). If we consider these unobserved variables as parameters, we obtain a saturated
model. In the sequel we impose restrictions on the unobserved variables, by assuming that
they are draws from some distribution. The specification in (8) is a proportional hazard
model for the continuous-time transition intensities.
The specification allows for unobserved heterogeneity in the transition intensities.
Unobservables may be important determinants of the transition intensities, because the
number of included observed characteristics is small. Because unobservables may also be
important for the selection into SE and correlation between unobservables in the
transition intensities and in the selection process biases the estimate of the SE-effect, we
must assess their contribution to the explanation of the transition intensities. Because the
data are grouped, the unobservables vary between, but not within groups. Hence, the
analogy between the group-specific unobserved heterogeneity in our model and the
individual unobserved heterogeneity in continuously observed duration data is imperfect.
An important advantage of our approach is that it is not necessary to specify the
distribution of the unobservables. An attempt to distinguish between the distribution of
unobservables and duration effects may bias the estimate of the SE effect (see e.g. Baker
and Melino 1999).
Under these assumptions the transition probability Pkmuv , i.e. the probability of a
transition of an individual of type m from state u to state v in duration interval k, is equal
to
(9) uvuh
h
hP k
uww
uwkm
uww
uwkm
uvkmuv
km ≠==
∆−−= ∑
∑ ≠=
≠=
3,2,1 ,2,1 ,exp13
,13
,1
This expression maps transition intensities to transition probabilities. The inverse mapping
from transition probabilities to transition intensities is
(10) ( )( ) ( ) uvuh
P
PPz uv
kmuukmk
uukm
uvkmuv
km ≠===
−∆−
= 3,2,1 ,2,1 , log1
loglog
17
with
(11) ∑≠=
−=3
,1
1uww
uwkm
uukm PP
the probability of staying in u in the k-th duration interval.
Let rkmu be the number of individuals of type m in state u at the start of duration
interval k, i.e. at time t k −1 , and let q kmuv of these individuals make a transition from state u
to state v in the k-th duration interval. Then we estimate Pkmuv by
(12) uvur
qP
ukm
uvkmuv
km ≠=== 3,2,1 ,2,1 ,ˆ
Upon substitution of these estimates in the middle expression of (10), we obtain uv
kmz . If in
(10) we replace the transition probabilities by their estimates, then the second equality
does not hold exactly. However, a Taylor series expansion around Pkmuv yields
(13) u32,,1 ,2,1 ,)log(ˆ ≠==+= vuhz uvkm
uvkm
uvkm ω
with
(14) ∑≠
−=3
w1,=w
)ˆ(u
uwkm
uwkm
uvwkm
uvkm PPbω ,
where uvw
kmb is defined in the Appendix. We may ignore the remainder of the Taylor
expansion as it can be shown that its probability limit for rkmu tending to infinity converges
to zero at a faster rate than uvkmω (Amemiya 1985, p. 276-77). Hence, omission of the
remainder does not affect the consistency of the estimator nor its asymptotic distribution.
Conditional on rkmu the random vector of the number of transitions of individuals
of type m from u to states v ≠ u in the k-th interval has a multinomial distribution with
parameters rkmu and Pkm
u , which is the vector of transition probabilities from state u to
states v ≠ u in interval k for type m. The errors of the system of regression equations
(13) for v ≠ u have mean 0, but they are heteroskedastic and correlated. The variance
matrix is given in the Appendix.
18
Upon substitution of (8) into (13) we obtain a linear regression model. More
precisely, we obtain for each origin state u = 1, 2 two regression equations that
correspond to the two destination states uv ≠= 3,2,1
(15) 2,1,3,2,1 ,ˆ =≠=++++= uuvz uvkm
uvkm
uvvm
vk
uvkm ωεαβγ
In the sequel we use the notation uv
kmuvkm
uvkm ωεν += .
We assume that the 4xMxK unobserved variables uvkmε are random variables26 with
a joint distribution such that
(16) ( ) 0=uv
kmE ε
(17) ( ) uvw
uummkkwumk
uvkmE σδδδεε ′′′
′′′ =
The unobserved variables corresponding to different duration classes, types, and origin
states are assumed to be uncorrelated. The unobserved variables corresponding to
different destination states may be correlated. Note that the disturbances uvkmω have the
same pattern of zero correlations. Hence, for origin state u and for duration interval k and
type m, the distribution of the 2-vector of disturbances of the two regression equations
(15) with v = 1, 2, 3 ≠ u , vkmu ≡ vkm
uvvkm
uw[ ] , which is the sum of the unobserved group
effect and the approximation error has a variance-covariance matrix with typical element
σ uvw + skmuvw , with v, w = 1, 2, 3 ≠ u . We denote this 2x2 matrix by Vkm
u . This completes the
specification of the regression equations.
The regression equations (15) are heteroskedastic and have correlated
disturbances. In the Appendix we discuss the estimation of these equations. Because we
have grouped data, we can use a χ 2 -goodness-of-fit test to evaluate the specification of
the model. This statistic can be interpreted as a GMM test for the null hypothesis of
regressor-error orthogonality. Since we do not control for selection on unobservables, we
anticipate a rejection of the null hypothesis on the basis of this test statistic. The details
can be found in the Appendix.
26 The assumption that the omitted variables are random is analogous to the treatment of unobserved heterogeneity in duration (and other) models. The expectation is with respect to the distribution of the omitted variables.
19
The estimation results for model (15) are reported in Table 3. Durations are
measured in months. We have chosen K=4 duration intervals: {[0,3), [3,6), [6,12),
[12,24)}. There are three types (M=3) that correspond to the three regions: Flanders,
Walloon Provinces and Brussels. Hence, 4 duration intervals, three regions, two origin
states and two destination states make up 48 empirical hazard estimates and 19 regression
parameters.
INSERT TABLE 3 APPROXIMATELY HERE
We obtain negative estimates for the variances of the unobserved group effects for the
transitions from state 1 to state 2 and from state 2 to state 1. Following Parks (1980,
p.299, footnote 5) we set these variances to 0. Hence, the unobserved group effects are
set to 0. We only find an unobserved group effect in the transitions from welfare with SE
and welfare without SE to out-of welfare.
The matching estimators of Section 4 suggested that SE decreases the exit rate
from welfare. Table 3, which corrects for the timing of selection and the resultant lower
exit rate due to the duration dependence of the exit rate, leads to the opposite conclusion:
participation in SE increases the exit rate by a factor 1.25 (= exp(.224)). This difference
can be explained by the fact that the matching estimator controls for more observable
characteristics and hence reduces the selection bias induced by creaming. Nevertheless,
both estimators are biased if selection in SE is based on unobservable characteristics. If
this is the case, then the distribution of the unobserved determinants of the exit rates ε kmuv
depends on the type of transition. In particular, creaming implies that >),( 23 mkE kmε
),( 13 mkE kmε , and as a result the effect of participation in SE is overestimated. The
specification is rejected by the chi-square goodness-of-fit test.
4.3. Correcting for Selection on Unobservables
In this subsection we derive a grouping/IV estimator that corrects for selection on
unobservables. We show that aggregation over the program participation indicator
identifies the effect of SE under a weaker assumption than the one invoked in the
previous section.
20
Because we observe all welfare spells that start between June 1 1987 and July 1
1990, the population fraction of spells that start out in the two origin states fkmu is known
exactly27. The fraction given k and m is kmu
km ff , where ∑ =≡
2
1u
ukmkm ff . Hence, if we
average regression equation (15) over the origin state u, for the destination state of
interest, i.e. for v=3, and if we recall that 013 =α , we obtain:
(18) 3323333ˆ kmkmkmmkkm pz ωεαβγ ++++=
where kmkmkm ffp 2≡ is the fraction of welfare recipients in k, m that participate in SE.
The new unobserved group effect, 3kmε , is the average of the unobserved
characteristics of the total welfare population in a particular region and duration class,
irrespective of participation in SE. Selective participation in SE in a region and duration
class implies that, in the case of creaming, welfare recipients with above average
unobserved characteristics are more likely to be selected for SE, so that ),( 13 mkE kmε
),( 23 mkE kmε> , i.e. the average unobserved characteristics of participants are better than
those of non-participants. This correlation between the participation indicator and the
average unobserved characteristics of participants and non-participants is absent in (18):
even if SE divides the welfare population into two subgroups on the basis of
(unobserved) characteristics, the fraction in SE can be unrelated to the average
characteristics of the welfare population. Hence (18) gives an unbiased estimate of the SE
effect, if the participation fraction is uncorrelated with 3kmε , a much weaker assumption
than the uncorrelatedness of characteristics and participation invoked in (15).
If 3kmε and kmp are uncorrelated, then the SE effect can be estimated from (18).
The resulting estimator is a grouping estimator, similar to the estimator first proposed by
Wald (1940) to deal with measurement error. Durbin (1954) showed that the Wald
estimator is an IV estimator. Angrist (1991) used this estimator to estimate a labor supply
equation with an endogenous wage rate, and in his application the instruments are year
dummies. In our application the instruments are the region and duration class indicators.
27 Hence, the averaging does not result in an errors-in-variables problem as in Deaton (1985).
21
The regression equation (18) includes region and duration class dummies, because
we want to allow for differences in the exit rate from welfare by region and by welfare
duration. As a consequence, the regressor in (18) is the residual kmp~ after regressing the
participation rate in SE on region and duration class dummies. The residual participation
rate is uncorrelated with the average participation rate by region and the average
participation rate by duration interval. Hence, if the SE participation rate is correlated
with 3kmε through the average unobserved characteristics of the welfare recipients in the
regions or through the average unobserved characteristics of welfare recipients in the
duration intervals, the residual rate kmp~ is uncorrelated with 3kmε . The inclusion of
regional and duration interval indicators makes the SE effect in (18) thus unbiased under
even weaker assumptions.
The price for these weaker assumptions is that the variation in the residual
participation rate kmp~ is less than the variation in kmp , and this increases the variance of
the estimator. If we include interactions between region and duration class in (18), there
is no variation, and (18) cannot be used for the estimation of the SE effect.
Although we can not estimate (18) if we include all interactions between region
and duration class indicators, there is indirect evidence that these are not important. First,
the chi-square goodness-of-fit test does not reject model (18). If the interactions were
important the model should be rejected. Second, in the next section we experiment with
the inclusion of some interactions and we find that the estimates do not change. The
regression equation (18) can be estimated by OLS, or, because of the heteroskedastic
errors, preferably by GLS. The details of the estimation procedure can be found in the
Appendix28.
The GLS estimates for regression equation (18) and the regression equations for
the other transitions are reported in Table 4. A comparison with the results in Table 3
shows that the parameter estimates for the destination states welfare without SE and
welfare with SE are almost identical in the two tables. Also the duration effects in the
transition intensity to the state out-of-welfare do not differ by much between these tables.
The only significant differences are in the regional effects and the effect of SE.
28 An alternative is to use difference-in-difference estimators where the differences are between duration intervals and regions. With our data these estimators are imprecise. Following Angrist (1991) we combine the difference-in-difference estimates to increase the precision. Angrist’s Efficient Wald estimator is our GLS estimator.
22
Eliminating unobserved differences between participants and non-participants makes the
transition intensity to the state out-of-welfare significantly smaller in the Walloon
provinces as compared to Flanders. This is reassuring given the structurally weaker
economy of the Walloon Provinces. More importantly, the effect of SE is negative in
Table 4, while it was significantly positive in Table 3. Note that the GMM test statistic
only rejects at a significance level of more than 27%.
INSERT TABLE 4 APPROXIMATELY HERE
5. Sensitivity Analysis
The standard error of the (efficient) GLS estimator of the SE-effect is relatively large.
Hence, it is important to investigate the sensitivity of this estimate to changes in the
specification. First, we consider the use of (stochastic) prior information. Cockx (1992,
1997) gives estimates of the regional effects on the hazard out of welfare. On the basis of
his estimates, taking Brussels as the region of reference, we calculate29 the following
point estimates (and standard errors) for Flanders and the Walloon Provinces: .0541
(.0165) and -.0797 (.0170). Following Theil (1963) or Judge et al. (1985, p.58), we
impose these stochastic linear restrictions on the parameters in the estimation. The results
are reported in Table 5.
INSERT TABLE 5 APPROXIMATELY HERE
The standard error of the effect of SE is smaller, and the point estimate becomes even
more negative.
In the estimation we group the durations in a relatively small number of intervals,
and we assume that the transition intensities are constant during these intervals. To
investigate the sensitivity to the specification of the duration dependence we estimate (18)
29 Cockx (1992,1997) reports separate estimates for men and women and considers other regressors besides region. The above point estimates were calculated by taking the appropriate transforms of the mixture of the duration distribution of men and women. As information on the covariances is not available, the standard errors are calculated by assuming a zero covariance between the estimated coefficients.
23
and the regressions for the other two transition intensities with 8 instead of 4 duration
intervals. The results are reported in Table 6.
INSERT TABLE 6 APPROXIMATELY HERE
Upon comparison with the estimates of Table 4 we conclude, that only the estimates of
the duration effects differ between these two tables. In particular, the estimate of the SE-
effect is similar to that in Table 4. Note that the model now is rejected against the
saturated model.
The estimates of Table 6 indicate that there is a problem with the specification.
We already noted that Brussels is rather different from Flanders and the Walloon
provinces. We re-estimate the model with a separate duration dependence for Brussels.
The estimate of the SE-effect for this specification is reported in the column (1) of Table
7.
INSERT TABLE 7 APPROXIMATELY HERE
The goodness-of-fit statistic improves considerably. The point estimate of the SE-effect is
hardly affected and its standard error is significantly reduced.
During the 1987-90 period the demand for labor was rising, peaking in 1990. One
may question to what extent the effect of SE is influenced by labor market conditions. In
particular, the review of the literature suggests that work experience programs are more
likely to be effective in periods of high labor demand (see e.g. Ridder 1986). To test this
hypothesis we estimate a model that allows for a different the SE-effect for 1987-88 and
for 1988-90. Column (2) of Table 7 reports the estimated SE-effects. The point estimates
are both negative, and in line with expectations, less negative in the second sub-period.
Throughout this paper we have made the assumption of effect homogeneity. If the
SE-effect varies among participants, then the grouping estimator does not estimate the
effect of SE on the average participant, but rather the effect on the marginal participant.
There are two cases, in which the grouping estimator does estimate the usual effect on
the average participant. Either the SE effect does not vary between participants or
selection into the program is unrelated to the individual specific variation in the benefit
difference between participation and non-participation, but is related to the average
24
difference in benefits between these two options (Heckman and Robb 1985, p. 196,
Heckman and Smith 1996, p.59-68, Heckman 1997). If the SE effect varies between
participants, then this induces within group unobserved heterogeneity in the exit rate. We
perform two tests for the presence of unobservables in the exit rate.
It is well known (see e.g. Lancaster (1990), p.64-65), that unobserved individual
heterogeneity induces a trend towards 0 in the SE effect. Hence, the absence of such a
trend, is an indication that within-group heterogeneity is not important. If we introduce a
time-varying SE effect in model (15), correcting for selection on observables only, we
find that the effect increases with duration.
Second, if within-group heterogeneity is important the hazard rate at a particular
duration is known to depend on the hazard rates of the past. This suggests testing for
within-group heterogeneity by checking whether the (weighted) residuals of our
regression equation are autocorrelated. For this purpose we calculated the Durbin-
Watson and the (modified) Lagrange multiplier (LM) test statistic.
Denote the weighted residuals corresponding to the estimated regression model
reported in Table 4 by ekmv . We adapt the Durbin-Watson statistic (d) to account for the
panel structure of the data in the following way:
(21) ( )
( )∑ ∑ ∑∑ ∑ ∑
= = =
= = = −−=
3
1 1 1
2
3
1 1 2
2
1 v
M
m
K
k
vkm
v
M
m
K
k
vmk
vkm
e
eed
The finite sample properties of this statistic are not obvious. However, asymptotically we
have d ~ N(2, 4/(3MK )) . The Durbin-Watson statistic corresponding to the model of
Table 4 is 2.4027. Consequently, we cannot reject the null hypothesis of zero
autocorrelation of the residuals at a significance level of 11%.
The Breusch (1978)-Godfrey (1978) LM test of autocorrelation provides an
alternative procedure. Operationally, the test is carried out by regressing the (weighted)
least squares residuals, ekmv , on the contemporaneous regressors and ek−1m
v . 3M (K − 1)R2
is then distributed as a chi-square random variable with one degree of freedom. Kiviet
(1986) finds that such a test can be poorly behaved when the regression model is
overspecified. He therefore suggests that it is safer to employ a modified LM procedure,
by testing the significance of the estimated autocorrelation coefficient in the regression
25
(see Godfrey 1988, p.116-117 for a discussion). Given the flexible parameterization of
our model we follow Kiviet's suggestion. The autocorrelation coefficient is estimated to
be equal to -.192 with an asymptotic T-ratio of -1.014. The hypothesis of zero
autocorrelation can therefore not be rejected at a significance level of 33%. The test
statistics for the model in Table 6 lead to similar conclusions.
6. Conclusion In this paper we estimate the effect of participation in social employment (SE) on welfare
duration in Belgium. For this purpose we extend the Minimum Chi-Square approach that
Cockx (1997) proposed for the analysis of grouped duration data. First, we show that the
time-varying participation in SE can be modeled in a competing risks framework.
Secondly, by allowing for a specification error, that captures the effect of unobservables
on the transition rates, we can study selection on these unobservables. The proposed
estimator deals with the selection bias by aggregation to a level where the participation
rate is exogenous. Inclusion of specific indicators eliminates the correlation between the
participation rate and the average unobserved differences between regions and duration
intervals. The resulting GLS estimator is efficient under these assumptions. The
estimation results can be summarized as follows. If we correct for selection on
observables only (Section 5.3, Table 3) participation in SE reduces the median welfare
spell from 13.1 to 11.6 months30. The corresponding coefficient is significantly different
from 0. If we correct for selection on unobservables, the change in the estimate of the SE
effect suggests that there is substantial creaming in the selection process. Participation in
SE is now reduces the exit rate from welfare, be it that the effect is not significantly
different from 0. The median duration of a welfare spell is 12.5 months for non-
participants and 14.8 months for participants in SE.
These results are in line with expectations. Welfare agencies (WA) face adverse
incentives. Since the welfare recipient who is in SE will eventually become entitled to
unemployment benefits, and will therefore no longer depend on the WA, the WA has no
incentive to enhance the professional integration of the participants in SE. Moreover,
since the SE is heavily subsidized, the WA, as an employer and provider of community
services, has an incentive to keep welfare recipients in the program. With the current
30 These median durations refer to a welfare recipient in Flanders, who is selected into SE after 6 months.
26
incentives to the WA, welfare recipient do not benefit and are likely to be hurt by
participation in SE. Moreover, international comparison of work experience programs
reveals that adverse incentives for administrators of the kind discussed in this paper can
explain the ineffectiveness of these programs in other European countries.
Appendix: The system of SUR regression equations (13) In (14)
(A.1) uv
kmvwuu
kmuu
kmuu
km
uvwkm
PPPPb
1
)log(
1
1
1δ+−
−=
and δvw the Kronecker delta. If we order the errors in (13) for v ≠ u in a vector, we have
for all k, m, and u = 1, 2
(A.2) 0)( =ukmE ω
so that the disturbance in (13) has expectation 0. The variance-covariance matrix of the
vector ukmω has typical element
(A.3) [ ] == uw
kmuvkm
uvwkm Es ωω
−−−++
−−∑ ∑ ∑
≠= ≠= ≠=
ukmvw
uvkm
uvkmvw
uukm
ukm
uxx uxx xyy
uykm
uxkm
uxkm
uxkm
ukm rPPPbPPPPbE )1()1(2)1()(
3
,1
3
,1
3
,1
2 δδ
with v, w = 1, 2, 3 ≠ u , )log(
1
1
1uu
kmuu
kmuu
km
ukm
PPPb −
−= , and where the expectation is taken
with respect to the distribution of rkmu . A consistent estimate is obtained by omitting the
expectation and replacing the transition probabilities by their estimates. Estimation of SUR system (15) Under assumption (16) which in particular implies that there are no unobserved
differences in the exit rates of SE-participants ( 3,2 == vu ) and non-participants
( 3,1 == vu ), we can estimate the parameters of these equations by OLS, or more
efficiently by GLS. Because we do not observe the transitions that occur in the open
27
K+1-th duration interval, we omit the regression equations that correspond to that
interval. As a result, we have MxKx2x2 equations. If we order the regressions by m, k, u,
v, in this order, we obtain a block-diagonal variance-covariance matrix of the MxKx2x2
vector of disturbances with the diagonal blocks being 2x2 matrices Vkmu .
The feasible GLS procedure consists of two steps. In the first step we use OLS to
estimate the 3xK+3x(M-1)+1 parameters γ kv , k=1,...,K, v=1,2,3, βm
v m=2,...,M,
v=1,2,3, and 23α . Next, we estimate the 2x2 variance-covariance matrix of the
unobserved group effects by
(A.4) [ ]
−= ∑∑
= =
M
m
K
k
uvwkm
uwkm
uvkm
uvw svvMK 1 1
ˆˆˆ1
σ
with ˆ v km
uv the residuals of the regression (15), and ˆ s kmuvw the estimate of (A.3). Substitution
of this consistent estimate gives a consistent estimate of Vkmu , which is used in the second
step of the feasible GLS procedure.
Let ˆ v be the 4KM-vector of residuals of the regression equation (15) ordered by
m, k, u, v, in that order. Let the estimated variance-covariance matrix of the disturbances
in (15) be denoted by ˆ V . This is a block diagonal matrix, which simplifies the
computation of its inverse. Under the assumption that the model is correctly specified we
have that ˆ ′ v ˆ V −1 ˆ v follows a χ 2 -distribution with 4KM-3K-3(M-1)-1 degrees of freedom.
As the regression equations can be seen as a set of 4KM moment conditions, this statistic
can be interpreted as a goodness-of-fit test statistic.
GLS estimation of the grouped regression equation (18)
In (15) we had four regression equations for each k, m. In (18) two of those are
combined, so that we have in total 3xKxM equations. If we order these equations by m, k
and for each such pair in the order: transition from 1 to 2, from 2 to 1, combined 1 and 2
to 3, then the variance matrix of the 3xKxM vector of disturbances is block-diagonal with
blocks
28
(A.5)
( )( )( )
+
++++
∑=
2
1
33233
2132213123211
1231123211122
......
0
0
u
ukmkm
ukm
kmkmkmkm
kmkmkmkm
sff
sffs
sffs
σ
σσσσ
In the two-step feasible GLS procedure we first use OLS to estimate the residuals. Next,
σ 122 and σ 211 are estimated as in (A.4). If we denote the OLS residuals of (23) by ˆ v km3 ,
then we have
(A.6) [ ]
−= ∑∑
= =
M
m
K
kkmkmkmkmkm sffvv
MK 1 1
1231312123 ˆ)(ˆˆ1
σ
with an analogous estimator for σ 213 , and
(A.7)
−= ∑∑ ∑
= = =
M
m
K
k u
ukmkm
ukmkm sffv
MK 1 1
2
1
3322333 ˆ)()ˆ(1
σ
In the second step we estimate the variance matrix (A.5) and estimate the parameters by
GLS.
29
References Amemiya, Takeshi (1985), Advanced Econometrics, Basil Blackwell, Oxford. Amemiya, Takeshi and Nold, Frederick (1975), "A Modified Logit Model", Review of Economics and Statistics, 57, 255-257. Angrist, Joshua D. (1991), "Grouped-data estimation and testing simple labor-supply models", Journal of Econometrics, 47, 243-266. Angrist J.D. and A.B. Krueger (1999), "Empirical strategies in labor economics", in O.C. Ashenfelter and D. Card (eds.), Handbook of Labor Economics, Volume 3A, Amsterdam: Noth-Holland, 1278-1329 and 1339-1366. Ashenfelter, Orley and Card, David (1985), "Using the longitudinal structure of earnings to estimate the effect of training programs", The Review of Economics and Statistics, 67, 648-60. Baker, Michael, and Melino, Angelo (1999), “Duration Dependence and Nonparametric Heterogeneity: A Monte Carlo Study”, Working Paper, University of Toronto. Berger, Jean-Marie (1987-92), Mémento CPAS 87-93, Kluwer, Editions Scientifiques, Zaventem. Bonnal, Liliane, Fougère, Denis and Sérandon, Anne (1997), "Evaluating the Impact of French Employment Policies on Individual Labour Market Histories", The Review of Economic Studies, 64, 683-713. Breusch, T. (1978), "Testing for Autocorrelation in Dynamic Linear Models," Australian Economic Papers, 17, 334-355. Cabinet du Secrétaire d'Etat à l'Emancipation (1990), Les Bénificiaires du Minimex en Belgique 1988 - 1989, Brussels. Cockx, Bart (1992), The Minimum Income Guarantee. Some Views from a Dynamic Perspective, Ph.D.-Thesis, Faculteit der Economische en Toegepaste Economische Wetenschappen, Katholieke Universiteit Leuven. Cockx, Bart (1997), "Analysis of Transition Data by the Minimum Chi-Square Method. An Application to Welfare Spells in Belgium", The Review of Economics and Statistics, 79(3), 392-405. Couch, Kenneth A. (1992), "New Evidence on the Long-Term Effects of Employment Training Programs", Journal of Labor Economics, 10(4), 380-88. Cox, D.R. (1972), "Regression Models and Life Tables," Journal of the Royal Statistical Society B, 34, 187-220.
30
Deaton, A. (1985), “Panel Data from Time Series of Cross-sections”, Journal of Econometrics, 30, 109-126. Dehaes, Viviane (1994), Leven van de bijstand. Een onderzoek naar de persistentie van armoede in Vlaanderen, CBGS-monografie, 1, Centrum voor Bevolkings en Gezinsstudiën, Brussel. Dehejia, Rajeev H. and Wahba, Sadek (1995a), "A Matching Approach for Estimating Causal Effects in Non-Experimental Studies", mimeo, Department of Economics, Harvard University, Cambridge, MA. Dehejia, Rajeev H. and Wahba, Sadek (1995b), "Causal Effects in Non-Experimental Studies", mimeo, Department of Economics, Harvard University, Cambridge, MA. Dolton, Peter J., Makepeace, Gerald H., Treble, John G. (1994), "The Youth Training Scheme and the School-to-work Transition", Oxford Economic Papers, 46(4), 629-57. Durbin, J. (1954), "Errors in variables," Review of the International Statistical Institute, 22, 23-32. Edin, Pers-Anders and Holmlund, Bertil (1990), "Unemployment, Vacancies and Labour Market Programmes: Swedish Evidence", in Padoa Schioppa, Fiorella (ed.), Mismatch and Labour Mobility, Cambridge University Press, Cambridge. Eichler, Martin and Michael Lechner (1998), "An Evaluation of Public Employment Programmes in the East German State of Sachsen-Anhalt", mimeo, Universität Mannheim/Universität St. Gallen. Feist, Holger and Ronnie Schöb (1999), "Workfare in Germany and the Problem of
Vertical Externalities", paper presented at the 11th EALE-conference, Regensburg, 23-26 September, CES, Ludwig-Maximilians-Universität München. Fraker, T. and Maynard, R. (1987), "The Adequacy of Comparison Group Designs for Evaluations of Employment-Related Programs", Journal of Human Resources, 22, 194-227. Friedlander, Daniel and Robins, Philip K. (1995), "Evaluating Program Evaluations: New Evidence on Commonly Used Nonexperimental Methods", The American Economic Review, 85(4), 923-37. Garcia, A. (1990), La remise au travail des bénéficiaires du Minimex dans le cadre de l'article 60 §7, Département des Sciences Politiques et Sociales, Université Catholique de Louvain, Louvain-la-Neuve. Godfrey, L. G. (1978), "Testing Against General Autoregressive and Moving Average Error Models When the Regressors Include Lagged Dependent Variables," Econometrica, 46, 1293-1302.
31
Godfrey, L.G. (1988), Misspecification tests in econometrics, Cambridge University Press, Cambridge. Grossman, Jean Baldwin, Maynard, Rebecca and Roberts, Judith (1985), Reanalysis of the Effects of Selected Employment and Training Programs for Welfare Recipients, Mathematica Policy Research, Inc., Princeton. Gueron, Judith M. (1990), "Work and Welfare: Lessons on Employment Programs", Journal of Economic Perspectives, 4(1), Winter, 79-98. Gueron Judith M. and Edward Pauly (1991), From Welfare to Work, Russell Sage Foundation, New York. Heckman, James J. (1997), "Instrumental Variables. A Study of Implicit Behavioral Assumptions Used in Making Program Evaluations", Journal of Human Resources, 32(3), 441-62. Heckman, James J., Heinrich, Carolyn and Smith, Jeffrey (1997), "Assessing the Performance of Performance Standards in Public Bureacracies", The American Economic Review, Papers and Proceedings, 87(2), 389-95. Heckman, James J., Ichimura, Hidehiko, Smith, Jeffrey, and Todd, Petra (1998), "Characterizing Selection Bias Using Experimental Data", Econometrica, 66(5), September, 1017-98. Heckman, James J. and Robb, Richard, Jr. (1985), "Alternative Methods for Evaluating the Impact of Interventions", in Heckman, James J. and Singer, B. (eds.), Longitudinal Analysis of Labor Market Data, Econometric Society Monograph series, Cambridge University Press, New York. Heckman, James J. and Smith, Jeffrey A. (1995), "The Effects of JTPA Performance Standards in Efficiency, Equity and Participant Outcomes", mimeo, University of Chicago, Chicago. Heckman, James J. and Smith, Jeffrey A. (1996), "Experimental and Nonexperimental Evaluation", in Schmid, G., O'Reilly, J. and Schöman K. (eds.), International Handbook of Labour Market Policy and Evaluation, Cambridge University Press, Cambridge, 37-88. Hollister, Robison G., Jr., Kemper, Peter, Maynard, Rebecca A. (eds.) (1984), The National Supported Work Demonstration, University of Wisconsin Press, Madison, Wis. Imbens, Guido W. and Angrist, Joshua D. (1994), "Identification and Estimation of Local Average Treatment Effects", Econometrica, 62, no. 2, 467-475. Judge, George G., W.E. Griffiths, R. Carter Hill, H. Lütkepohl and T.-C. Lee (1985), The Theory and Practice of Econometrics, (2nd ed.), John Wiley and Sons, New York.
32
Kiviet, J. (1986), "On the Rigour of Some Specification Tests for Modeling Dynamic Relationships," The Review of Economic Studies, 53, 241-262. Krueger, Alan B. (1990), "Workers' Compensation Insurance and the Duration of Workplace Injuries", NBER Working Paper Series, no.3253. Lalonde, R. (1986), "Evaluating the Econometric Evaluations of Training Programs with Experimental Data", American Economic Review, 79, 604-620. Lalonde, R. (1995), "The Promise of Public Sector-Sponsored Training Programs", Journal of Economic Perspectives, 9(2), Spring, 149-68. Lancaster, Tony (1990), The Econometric Analysis of Transition Data, Cambridge University Press, Cambridge. Layard, R., Nickell, S. and Jackman, R. (1991), Unemployment, Oxford University Press, Oxford. Meyer, Bruce D. (1989), "A Quasi-Experimental Approach to the Effects of Unemployment Insurance", NBER Working Paper Series, no. 3159. Meyer, Bruce D., Viscusi, W. Kip and Durbin, David L. (1990), "Workers' Compensation and Injury Duration: Evidence from a Natural Experiment", NBER Working Paper Series, no. 3494. Meyer, Bruce D. (1995), “Natural and Quasi-Experiments in Economics”, Journal of Business and Economic Statistics, 13(2), April, 151-161. Moffitt, Robert (1992), "Incentive Effects of the US Welfare system: A Review", Journal of Economic Literature, 30 (March), 1-61. O.C.M.W.-Visies (1993), "Het Bijzonder Fonds voor Maatschappelijk Welzijn: een overzicht van de gehanteerde criteria", O.C.M.W.-Visies, 1, 14-22. Parks, Richard W. (1980), "On the Estimation of Multinomial Logit Models from Relative Frequency Data", Journal of Econometrics, 13, 293-303. Prentice, R. L. and L.A. Gloeckler (1978), " Regression Analysis of Grouped Survival Data with Application to Breast Cancer Data," Biometrics, 34, 57-67. Ridder, Geert (1986), "An Event History Approach to the Evaluation of Training, Recruitment and Employment Programmes", Journal of Applied Econometrics, 1(2), 109-126. Theil, H. (1963), "On the Use of Incomplete Prior Information in Regression Analysis", Journal of the American Statistical Association, 58, 401-414.
33
Van de Velde, Veerle (1990), Wat na de sociale tewerkstelling? Evaluatie van de toepassing van art. 60 par.7 van de organieke wet op het OCMW, Hoger Instituut voor de Arbeid, Katolieke Universiteit Leuven, Leuven. Van Langendonck, J. (1991), Handboek Sociale Zekerheid, Acco, Leuven. Wald, Abraham (1940), "The fitting of straight lines if both variables are subject to error", Annals of Mathematical Statistics, 11, 284-300. Wouters, Martine, Rien Van Meensel and Ides Nicaise (1994), De TOK-projecten en hun cursisten, drie jaar later. Follow-up onderzoek van de projecten van 1989, Hoger Instituut voor de Arbeid, KU Leuven, Leuven.
34
Table 1: Population characteristics
all welfare spells (including SE)
spells with some time in SE
matched sample: Controls in same
WA Number of spells 80,621 2,813 681* Average age in years 30.0 30.9 29.3 Male 39.1% Belgian 92.2% 94.3% 97.8% Children dependent children present
39.2% 51.7% 39.8%
Number of children 1.8 1.8 1.7 Marital status unmarried 45.3% 37.5% 46.8% married 30.6% 38.5% 34.6% widow(er) 1.2% 0.6% 0.3% divorced 22.8% 23.4% 23.4% Household type Living alone 69.5% 73.1% ** Married / cohabiting. 10.1% 9.8% ** Unmarried / cohabiting 20.4% 17.2% ** Region Flanders 39.4% 59.4% 52.0% Walloon Provinces 47.0% 25.7% 29.5% Brussels 13.6% 14.9% 18.5% Size of municipality > 100,000 36.0% 28.4% 42.4% 50-100,000 18.3% 18.1% 27.5% < 50,000 45.7% 53.5% 30.1% * 2 groups of 681 individuals ** This information was not available at the moment that the groups were matched.
35
Table 2: Quintiles of Survivor Functions by Region and Participation Indicator (95% confidence intervals between brackets)
1. Duration in Months until the Beginning of SE
Quintiles Flanders Walloon P. Brussels Belgium 75% 1.4
[1.1,1.8] 1.0
[0.6,1.2] 1.0
[1.0,1.2] 1.1
[1.0,1.3] 50% 6.8
[6.0,7.5] 5.3
[4.5,6.5] 4.5
[3.7,5.9] 6.0
[5.5,6.5] 25% 16.5
[15.3,17.9] 14.1
[12.8,15.5] 12.0
[10.4,14.0] 15.0
[14.1,15.9] 2. Duration in Months of SE spells
75% 4.0 [3.9,4.8]
2.5 [2.0,3.0]
2.4 [1.9,3.0]
3.1 [3.0,3.6]
50% 10.5 [9.0,11.5]
6.0 [6.0,6.8]
6.5 [6.0,7.0]
7.6 [7.0,8.2]
25% 18.0 [17.9,18.2]
12.5 [12.0,14.0]
13.0 [12.0,17.0]
15.9 [14.9,17.9]
3. Welfare Durations in Months for Non-Matched Groups
75% Non-participants
1.9 [1.9,1.9]
2.0 [1.9,2.0]
2.0 [2.0,2.0]
1.9 [1.9,2.0]
Participants in SE
7.6 [7.0,8.0]
5.0 [4.4,6.0]
5.6 [4.0,6.7]
6.7 [6.0,7.0]
50% Non-participants
4.0 [4.0,4.2]
5.0 [4.9,5.0]
4.8 [4.5,4.9]
4.8 [4.6,4.9]
Participants in SE
14.8 [14.0,15.5]
11.8 [10.5,12.8]
11.4 [10.3,12.8]
13.1 [12.8,14.0]
25% Non-participants
11.0 [10.9,11.5]
13.0 [12.9,13.3]
11.4 [11.0,12.0]
12.0 [11.9,12.0]
Participants in SE
23.7 [22.4,25.0]
20.0 [18.6,22.0]
20.7 [19.0,23.4]
22.3 [21.3,23.4]
4. Welfare Durations in Months for Matched(i) Groups - Controls in Participating WA
75% Non-participants
2.0 [1.8,2.4]
1.8 [1.4,2.4]
1.5 [1.0,2.3]
1.9 [1.7,2.0]
Participants in SE
3.7 [3.0,4.2]
2.8 [1.9,3.5]
1.7 [1.0,2.0]
2.8 [2.2,3.1]
50% Non-participants
4.4 [3.9,5.9]
4.0 [3.0,5.1]
3.3 [3.0,4.0]
4.0 [3.8,4.5]
Participants in SE
7.0 [6.0,9.0]
5.5 [4.4,6.0]
5.2 [3.0,6.9]
6.0 [5.9,6.9]
25% Non-participants
14.5 [10.6,19.9]
8.8 [8.0,12.3]
7.0 [5.1,13.6]
11.0 [9.2,13.6]
Participants in SE
13.0 [12.0,15.0]
9.3 [9.0,11.9]
10.0 [7.2,12.0]
12.0 [11.9,12.5]
(i) The individuals are matched on the year of birth, gender, nationality, number of children, marital status (see Table 1 for a definition), and WA (panel 4). The welfare duration of non-participants is required to be greater than the elapsed duration of participants at the time of selection in SE. We redefine welfare duration by subtracting the time elapsed until selection in SE for each pair of matched individuals.
36
Table 3: GLS estimates of the parameters of the transition intensities (standard error in parentheses)
Destination state (v) 1: Welfare
without SE 2: Welfare with
SE 3: Out of welfare
Duration interval (months)
0-3 (γ1v) -3.084***
(.188) -4.977***
(.064) -1.838***
(.095) 3-6 (γ 2
v − γ 1v) -.224
(.236) -.011
(.095) -.341***
(.108) 6-12 (γ 3
v − γ 1v) -.610**
(.236) -.035
(.095) -.465***
(.107)
12-24 (γ 4v − γ 1
v ) -.693** (.265)
.089 (.107)
-.579*** (.109)
Region
Walloon Prov. (β2v) .448**
(.184) -1.190***
(.086) -.016
(.092) Brussels (β3
v) -.364
(.318) -.359***
(.104) .086
(.095) SE-effect
α 23 - - .224**(i)
(.093) Unobserved vars.
ˆ σ 1vw 0 - 0.00868
ˆ σ 2vw
- 0 0.01986
Weighted sum of squared residuals (degrees of freedom)
72.54 (29)
P-value .00 (i) This is the parameter of interest, signifying that SE increases the transition rate out of welfare by 25% (= *100
( )[ ]1224.exp − ).
* significant at the 10% level ** significant at the 5% level *** significant at the 1% level
37
Table 4: GLS estimates of the parameters of the transition intensities with participation rate in SE instead of participation indicator
(standard errors in parentheses) Destination state v
1: Welfare without SE
2: Welfare with SE
3: Out of welfare
Duration interval (months)
0-3 (γ1v) -3.095***
(.127) -4.977***
(.043) -1.716 ***
(.029) 3-6 (γ 2
v − γ 1v ) -.222
(.159) -.011
(.064) -.321***
(.025)
6-12 (γ 3v − γ 1
v) -.608***
(.159) -.035
(.064) -.509***
(.033)
12-24 (γ 4v − γ 1
v) -.614***
(.180) .088
(.072) -.709***
(.052) Region
Walloon Prov. (β2v ) .469***
(.124) -1.191***
(.058) -.116***
(.033) Brussels (β3
v ) -.349 (.215)
-.360*** (.070)
-.007 (.026)
SE-effect
α 23 - - -.258(i)
(.814) Unobserved vars
ˆ σ 1vw 0 - -
ˆ σ 2vw
- 0 -
ˆ σ 33
- - .000464
Weighted sum of squared residuals (degrees of freedom)
19.36 (17)
P-value .31 (i) This is the parameter of interest, signifying that SE decreases the transition rate out of welfare by 23% (= *100
( )[ ]258.exp1 −− ).
* significant at the 10% level ** significant at the 5% level *** significant at the 1% level
38
Table 5: GLS estimates of the parameters of the transition intensities with participation rate in SE instead of participation indicator and
imposing the stochastic linear restrictions β13 = .0541 (.0165) and
β 23 = -.0797 (.0170).
(standard errors in parentheses) Destination state v
1: Welfare without SE
2: Welfare with SE
3: Out of welfare
Duration interval (months)
0-3 (γ1v) -3.446***
(.238) -5.337***
(.071) -1.742 ***
(.018) 3-6 (γ 2
v − γ 1v ) -.222
(.162) -.011
(.065) -.318***
(.025)
6-12 (γ 3v − γ 1
v) -.607***
(.162) -.035
(.065) -.503***
(.029)
12-24 (γ 4v − γ 1
v) -.613***
(.183) .088
(.074) -.698***
(.043) Region
Flanders (β1v ) .351
(.219) .360***
(.071) .042***
(.014) Walloon Prov. (β2
v ) .821*** (.228)
-.830*** (.081)
-.086 (.014)
SE-effect
α 23 - - -.499(i)
(.621) Unobserved vars
ˆ σ 1vw 0 - -
ˆ σ 2vw
- 0 -
ˆ σ 33
- - .000464
Weighted sum of squared residuals
21.79
(i) This is the parameter of interest, signifying that SE decreases the transition rate out of welfare by 39% (= *100
( )[ ]499.exp1 −− ).
* significant at the 10% level ** significant at the 5% level *** significant at the 1% level
39
Table 6: GLS estimates of the parameters of the transition intensities with participation rate in SE instead of participation indicator: 8
duration intervals (standard errors in parentheses)
Destination state v
1: Welfare without SE
2: Welfare with SE 3: Out of welfare
Duration interval (months)
0-3 (γ1v) -3.103***
(.148) -4.977***
(0.052) -1.706***
(.015)
3-6 (γ 2v − γ 1
v) -.234
(.191) -.011
(.077) -.335***
(.015) 6-9 (γ 3
v − γ 1v) -.211
(.190) -.087
(.092) -.619***
(.023) 9-12 (γ 4
v − γ 1v) -.118
(.195) -.115
(.109) -.532***
(.027)
12-15 (γ 5v − γ 1
v ) -.205 (.210)
-.047 (.124)
-.868*** (.036)
15-18 (γ 6v −γ 1
v ) -.176 (.226)
-.070 (.145)
-.975*** (.042)
18-21 (γ 7v − γ 1
v) -.214
(.263) -.097
(.175) -.977***
(.047) 21-24 (γ 8
v − γ 1v) -.207
(.304) -.225
(.222) -.936***
(.053) Region
Walloon Prov. (β2v) .472***
(.119) -1.180***
(.069) -.112***
(.017) Brussels (β3
v) -.128
(.191) -.374***
(.085) -.020
(.018) SE-effect α 23
- - -.171(i)
(.502)
Weighted sum of squared residuals (degrees of freedom)
67.463 (41)
P-value .0057 (i) This is the parameter of interest, signifying that SE decreases the transition rate out of welfare by 16% (= *100
( )[ ]171.exp1 −− ).
* significant at the 10% level ** significant at the 5% level *** significant at the 1% level
40
Table 7: Efficient Wald Estimates of the SE-effect for various models*
(standard errors in parentheses)
(1) (2) SE-effect 87-88 89-90
23α -.248 (.316)
-1.663 (1.860)
-0.458 (1.400)
Weighted sum of squared residuals (degrees of freedom)
7.730 (14)
47.873 (34)
P-value .90 .058 * The complete results can be obtained from the authors on request. (1) GLS estimates of the parameters of the transition intensities with participation rate in SE instead of participation indicator, allowing the duration effects to be different in Brussels (2) GLS estimates of the parameters of the transition intensities with participation rate in SE instead of participation indicator: proportionality over calendar time and SE-effect varies over calendar time.
41
Figure 1: Survival Functions of Welfare Durations
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 3 6 9 12 15 18 21 24 27 30 33 36
months
non-participants participants in SE
42
Figure 2: Survival Functions of the Welfare Durations of the Matched Groups - Contols in Participating WA's
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 3 6 9 12 15 18 21
months
non-participants participants in SE
The individuals are matched on year of birth, gender, nationality, number of children, marital status (see Table 1 for a definition), and WA. The welfare duration of non-participants is required to be greater than the elapsed duration of participants at the time of selection in SE. We redefine welfare duration by subtracting the time elapsed until selection in SE for each pair of matched individuals