This article was downloaded by: [Florida Atlantic University]

On: 10 November 2014, At: 13:41

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

Applied Economics
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/raec20

Do the skilled and prime-aged unemployed benefit

more from training? Effect heterogeneity of public
training programmes in Germany
a a b
Ulf Rinne , Marc Schneider & Arne Uhlendorff
a
IZA , PO Box 7240, D-53072 Bonn, Germany
b
University of Mannheim, IZA , DIW Berlin, L 7, 3-5, D-68161 Mannheim
Published online: 04 Nov 2010.

To cite this article: Ulf Rinne , Marc Schneider & Arne Uhlendorff (2011) Do the skilled and prime-aged unemployed benefit
more from training? Effect heterogeneity of public training programmes in Germany, Applied Economics, 43:25, 3465-3494,
DOI: 10.1080/00036841003670697

To link to this article: http://dx.doi.org/10.1080/00036841003670697

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
 Applied Economics, 2011, 43, 3465–3494

Do the skilled and prime-aged

unemployed benefit more from
training? Effect heterogeneity
of public training programmes
in Germany
Ulf Rinnea,*, Marc Schneidera and Arne Uhlendorff b
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

a
IZA, PO Box 7240, D-53072 Bonn, Germany
b
University of Mannheim, IZA, DIW Berlin, L 7, 3-5, D-68161 Mannheim

This study analyses the treatment effects of public training programmes

for the unemployed in Germany. Based on propensity score matching
methods, we extend the picture that has been sketched in previous studies
by estimating the treatment effects of medium-term programmes for
different skill and age groups. Our results indicate that programme
participation has a positive impact on employment probabilities and
earnings for almost all sub-groups. We find little evidence for the presence
of heterogeneous treatment effects, and the magnitude of the differences is
quite small. Our results thus – at least in part – conflict with the strategy to
provide training increasingly to individuals with better employment
prospects.

I. Introduction expenditure. However, the number of participants

has decreased over the past few years (Fig. 1). While
One central aim of Active Labour Market Policy more than 500 000 unemployed individuals entered a
(ALMP) is to increase the employment prospects of training programme in 2000, this number was only
unemployed individuals. For this purpose, the around 130 000 in 2005 but increased to nearly
Federal Employment Agency (FEA) in Germany 250 000 in 2006. In contrast, the unemployment rate
spends a substantial amount of money on measures remained rather constant during this period.
such as job creation schemes, public training pro- A number of studies have evaluated the effective-
grammes or employment subsidies. For instance, ness of public training programmes in Germany. For
about E20.5 billion were spent on ALMP measures in a recent review of the results see Caliendo and Steiner
2002 (Eichhorst and Zimmermann, 2007). The most (2005).1 The results are rather heterogeneous –
important measures in Germany are public training depending on the method, the investigation period
programmes. Totalling almost E7 billion, these and the underlying data set. While earlier studies
programmes account for more than 32% of the often find insignificant or even negative effects

*Corresponding author. E-mail: [email protected]

1
LaLonde (2003) and Kluve (2010) summarize the international literature on the evaluation of ALMP. Recent examples for
evaluation studies of other ALMP measures in Germany include Jaenichen and Stephan (2009) for wage subsidies,
Baumgartner and Caliendo (2008) for start-up programmes and Caliendo et al. (2008b) for job creation schemes.
Applied Economics ISSN 0003–6846 print/ISSN 1466–4283 online ß 2011 Taylor & Francis 3465
http://www.informaworld.com
DOI: 10.1080/00036841003670697
 3466 U. Rinne et al.
of public training programmes, partly differentiated
by gender, programme type and region. The contri-
bution of this article is to extend the picture outlined
so far by investigating potential effect heterogeneity
across two dimensions which have a large impact on
the reemployment probabilities: vocational education
(as a proxy for the skill level) and age.2 Individuals
without a vocational degree as well as older workers
have a relatively low probability of leaving unem-
ployment for a job. However, this does not necessar-
ily imply that participation in training programmes
is not effective for these groups.3
Based on whether or not an individual has a formal
vocational degree, we distinguish between skilled and
Fig. 1. Entrants into public training programmes, unem- unskilled individuals. There is no unambiguous
ployment rate (2000–2006) hypothesis for the direction of potential effect heter-
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Source: Federal Employment Agency (FEA).

Note: Bars show annual number of entrants into public ogeneity with respect to the level of vocational
training programmes (left axis). The dashed line displays education, and hence across skill groups. On the
the average unemployment rate (right axis, in %). one hand, public training programmes for unem-
ployed individuals may involve diminishing marginal
(Lechner, 1999, 2000; Hujer and Wellner, 2000), most returns. The more human capital a given individual
recent studies, which are based on rich administrative has already accumulated, the less the training pro-
data sets, find positive treatment effects for at least gramme enhances his or her human capital. On the
some sub-groups (Biewen et al., 2007; Lechner et al., other hand, the effect of medium-term training
2007, 2010; Fitzenberger et al., 2010). Hujer et al. programmes, which is the focus of our study, could
(2006) is an example of a recent study which finds be positively related to the human capital the
negative effects. However, the authors concentrate on individual has already accumulated, and it could be
the duration of the initial unemployment spell, and complementary to the existing skills. In contrast to
the negative impact of programme participation long-term programmes, which in general aim to
probably reflects the locking-in effect of training provide a vocational degree and thus enhance
programmes. Another example for a recent study human capital, shorter programmes can only activate
reporting negative effects is Lechner and Wunsch human capital already accumulated, at least accord-
(2008). Despite differences in the data and the ing to this line of argumentation. These two opposing
investigation period, the mixed results may also effects leave the direction and the extent of potential
reflect different methodological approaches. For effect heterogeneity across skill groups ambiguous.
instance, Stephan (2008) finds that estimated treat- Similarly, one can state expectations about the
ment effects differ considerably across different direction and the extent of potential effect heteroge-
definitions of nonparticipation. neity across age groups: for instance, young indivi-
In Germany, a stricter selection of participants for duals have only recently finished their education and
training programmes by caseworkers based on the initial training. It is reasonable to assume that within
expected reemployment probability has been intro- this age group, the choice which has been made with
duced as a part of the reform of ALMP in 2003. This respect to the optimal level of human capital still
implies that after the reform, caseworkers now tend represents the optimum or is at least reasonably close
to select individuals with higher reemployment to it. Additional training would have only small
probabilities for participation – independent of the returns. On the other hand, the temporal distance to
individual gain resulting from participation. The schooling and initial education is considerably larger
above-mentioned studies focus on the average effects for prime-aged and older individuals. For these age
2
Caliendo et al. (2008a) investigate a similar question for job creation schemes in Germany and present evidence for the
presence of effect heterogeneity. Although previous results of negative average effects are confirmed in their study, some strata
of the population benefit from participation in job creation schemes.
3
One could think of additional interesting dimensions of effect heterogeneity like nationality, health restrictions and family
status. In this article, we focus on age and vocational education. The reasons for this are (a) that we have for many variables,
like health restrictions and nationality, relatively small sample sizes, and (b) that age and vocational education are highly
correlated with other dimensions like family status and previous earnings. However, the analysis of additional dimensions of
effect heterogeneity could be an interesting topic for further research.
 Public training programmes in Germany 3467
groups, it seems likely that either the once accumu- training and employment programmes from 2000 to
lated human capital has depreciated or the skill 2002. Treatment effects are investigated at a fairly
requirements have changed over time. In any case, disaggregated level, and compared to our study, a
updating existing skills or acquiring new skills would relatively small inflow sample into unemployment is
then exhibit sizeable returns, which could be increas- used. The authors find evidence for effect heterogene-
ing in age. However, Büsch et al. (2009) present ity and show that job seekers with relatively good
empirical evidence supporting the presence of age a priori employment prospects are worse off because of
discrimination in the hiring process in Germany. It is large locking-in effects from which they recover only
thus likely that older workers have difficulties in very slowly, while job seekers with disadvantageous
finding a job. Furthermore, this age group will be a priori employment prospects show below average
facing retirement in the near future. In other words, locking-in effects and positive employment effects for
the length of the remaining working life clearly differs some of the shorter training programmes, including
across age groups and thus the time to recoup an job related training. Biewen et al. (2007) use similar
investment in human capital. This may negatively data and analyse effect heterogeneity by regressing
affect the motivation and incentives for training for outcome variables after matching on different socio-
older workers. If these factors are correlated with the economic covariates. They find little heterogeneity
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

training’s effectiveness, returns will be lower. In sum, along observed characteristics, although in some cases
these considerations lead to the hypothesis of lower older and less educated participants seem to benefit
returns to training, and hence a lower effectiveness of less or not at all from programme participation.
training programmes, for younger as well as for older In comparison to Biewen et al. (2007) and Lechner
individuals. Under this hypothesis, returns will be the and Wunsch (2008) we have access to a much larger
largest for prime-aged workers, and the overall sample of participants in training programmes. This
pattern of treatment effects across age groups will allows us to apply matching methods within several
be inversely U-shaped. sub-groups – e.g. within the sample of unskilled
We estimate effects for three different programme women – and to investigate the effect heterogeneity in
types: (a) programmes with a focus on class-room greater detail. Moreover, we analyse the effects on
training, (b) programmes with a focus on practical monthly earnings by comparing the shares of indivi-
experience and (c) training within practice firms, duals with and without training in different quartiles
which simulate a real working environment. These of the earnings distribution. This approach provides
three types of programmes are shorter programmes in insights into the effect of programme participation on
comparison to other ALMP measures in Germany the probability of finding higher or lower paid jobs.
and have a median duration between 6 and 8 months. Our analysis is based on an inflow sample into
We expect that the effect heterogeneity with respect training programmes for the year 2002. We ensure that
to age may differ between programmes with a more the control group consists of individuals who have
theoretical focus (type 1) and programmes with a been unemployed as long as the participants by exactly
more practical focus (types 2 and 3). Young matching on the previous unemployment duration.
individuals have only recently finished their educa- Furthermore, a propensity score matching aims to
tion and initial training, and it seems reasonable that balance differences in a wide range of observable
practical experience is more relevant to them than characteristics, including socio-demographic vari-
additional classroom training. In contrast to that, the ables, information about the labour market history
temporal distance to schooling and initial education of individuals and the regional labour market envi-
is considerably larger for older individuals, and we ronment. It is quite plausible that once we condition
thus expect theoretical training to be more relevant on this rich set of variables the Conditional
for them than practically oriented courses. With Independence Assumption (CIA) holds, and that
respect to the skill level, we do not have unambiguous there are no additional variables which jointly influ-
expectations about the effect heterogeneity across ence the participation decision and the outcome
programme types. Potential differences depend on variable which are not reflected by our rich set of
differences in the extent of diminishing marginal variables.
returns and potential complementarities to existing Our results indicate that on average, after an initial
skills between more practical and more theoretical reduction due to locking-in effects all programme
training, and we do not have a theoretical model in types have a significantly positive impact on the
mind which would give us corresponding predictions. employment prospects of the participants, both in
Two recent contributions point to a similar direc- terms of employment probabilities and earnings. This
tion as this article. Lechner and Wunsch (2008) result holds if we divide our sample into skilled and
analyse the effectiveness of several West German unskilled individuals. Moreover, if we compare
 3468 U. Rinne et al.
employment probabilities 24 months after pro-

0 10 20 30 40 50 60 70 80
percentage share among all program types
gramme entry, we find no significantly different
treatment effects for skilled and unskilled participants
within the same gender and programme type.
Similarly, the extent to which the average treatment
effects vary between age groups is relatively small.
However, while we observe significantly positive
effects across gender and programme types for all
prime-aged individuals, we observe only insignif-
icantly positive effects at the end of our observation 2000 2001 2002 2003 2004
period for some of the younger and older sub-groups. Type 1: occupation−related or general training
Nonetheless, we find for each age group at least one Type 2: practice training in key qualifications
Type 3: practice firms
programme with a positive impact on subsequent
employment prospects. Hence, our results conflict at
least in part with the strategy of providing training Fig. 2. Share of entrants by programmmes type
increasingly to individuals with better employment (2000–2004)
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Source: Federal Employment Agency (FEA).

prospects. Although our results seem to suggest
that better targeting of training programmes in
Germany – especially with respect to age – may be qualifications, and (c) type 3: practice firms.
promising, we do not find evidence that skilled and Participants in type 1 learn specific skills required
prime-aged individuals in general benefit more from for a certain vocation (e.g. computer-aided design for
training than other groups. a technician/tracer) or receive qualifications that are
This article is structured as follows. Section II of general vocational use (e.g. MS Office, computer
describes the data and the programme types being skills). Type 2 is a predominantly practically oriented
analysed. Section III presents the econometric meth- programme with only few theoretical parts. It follows
ods and Section IV discusses the results. Section V the principle ‘learning by doing’ and is often
concludes. combined with internships. Within type 3, the simu-
lation of real operations is conducted, and most of
the time technical training is provided. For example,
participants are taught practical skills of wood
II. Data working and how to operate work benches and
machines, under the supervision of instructors.
We use a sample of a particularly rich administrative Figure 2 shows that type 1 is by far the most
data set, the Integrated Employment Biographies important programme type. By 2003, around 60% of
(IEB) of the FEA.4 It contains detailed daily infor- all participants in public training programmes were
mation on employment subject to social security assigned to this particular type. It became even more
contributions including occupational and sectoral important after the labour market reforms were
information, receipt of transfer payments during introduced in 2003, as this share increased to more
periods of unemployment, job search and participa- than 70%. The three types together accounted for
tion in different programmes of ALMP. roughly 85% of all participants in public training
Furthermore, the IEB comprises a large variety of programmes over the period 2000 to 2004.
covariates – e.g. age, marital status, number of Figure 3 indicates that the three programme types
dependent children, disability, nationality and are rather short in comparison to other ALMP
education. measures in Germany: after a year, more than 90% of
As the public training programmes currently in the participants have left. The median programme
place in Germany are quite heterogenous, we con- duration is about 8 months for type 1 and roughly
centrate on and differentiate between three particular 6 months for types 2 and 3. While a comparatively
types: (a) type 1: occupation-related or general large fraction of participants finishes type 1 exactly
training, (b) type 2: practice training in key after 12 months, an even larger share finishes type 3
4
The IEB is in general not publicly available. Only a 2.2% random sample (the Integrated Employment Biographies Sample,
IEBS) can be obtained for research purposes. See e.g. Jacobebbinghaus and Seth (2007) for details on the IEBS. The IEB
consists of four different administrative data sources: the employees’ history (BeH), the benefit recipients’ history (LeH), the
job seekers’ database (ASU/BewA), and the programme participants’ master data set (MTH). For a more detailed description
see e.g. Schneider et al. (2007).
 Public training programmes in Germany 3469

Type 1 Type 2

1
.2 .3 .4 .5 .6 .7 .8 .9

.2 .3 .4 .5 .6 .7 .8 .9
Kaplan−Meier Survivor Function

Kaplan−Meier Survivor Function

.1

.1
0

0
0 6 12 18 24 0 6 12 18 24
Months since Program Entry Months since Program Entry

Type 3
1
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

.2 .3 .4 .5 .6 .7 .8 .9
Kaplan−Meier Survivor Function
.1
0

0 6 12 18 24
Months since Program Entry

Fig. 3. Actual programme durations (participants 2002)

Source: IEB, own calculations.

exactly after 6 months. For type 2, we observe a individuals participating in programme type 2 is
sizeable share who ends the measure exactly after 6 or particularly low. The share of individuals who did not
12 months. graduate from school is at 5% the lowest among the
Our sample of participants consists of almost participants in type 1, compared to 12% and 9% for
60 000 unemployed persons entering the three pro- type 2 and type 3, respectively. This share is around
gramme types in 2002. More precisely, we observe 11% in the comparison group. Furthermore, 13% of
22 843 participants in type 1, 13 744 participants in participants in type 1 have the highest educational
type 2, and 20 042 participants in type 3. This sample attainment, while this share is relatively low for type 2
allows us to draw conclusions on the average partic- (4%) and type 3 (5%) and for nonparticipants (7%).
ipant starting a given programme in 2002.5 In order We obtain a similar picture with respect to vocational
to apply the matching approach described in Section attainment: the share of individuals without a
III, around 520 000 nonparticipants were drawn. vocational degree is lowest for type 1 (18%) and
Both participants and nonparticipants are aged highest for type 2 (36%). In general, the majority of
between 17 and 65 years.6 individuals have completed an apprenticeship and
Table 1 displays descriptive statistics of our sample, only a minority have a university degree. However,
differentiated by participation status and programme these shares differ across participation status and
type.7 Participants in all programme types appear to across programme types. For example, around 10%
be on average younger and more likely to be German of participants in type 1 hold a university degree;
than nonparticipants. The share of East German whereas this share is relatively low for type 2 (2%)
5
The number of participants entering a programme differs between the quarters analysed. We take this into account by
applying corresponding weights when calculating the average treatment effects on the treated.
6
One could argue for stricter age restrictions, for example, because of early retirement regulations in Germany. However, if
one is interested in the average effects of treatment on the treated and there are participants older than 55 or 60, there is no
reason to exclude these individuals.
7
A detailed description of selected variables can be found in Table A1 of the Appendix.
 Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

3470

Table 1. Descriptive statistics (selected variables)

Participants type 1 Participants type 2 Participants type 3 Nonparticipants

Mean SD Mean SD Mean SD Mean SD

Socio-demographic characteristics
Female 0.4156 (0.4928) 0.4347 (0.4957) 0.4265 (0.4946) 0.4271 (0.4947)
East Germany 0.4298 (0.4951) 0.1391 (0.3461) 0.3797 (0.4853) 0.3029 (0.4595)
Age 37.248 (9.5597) 35.797 (11.106) 37.788 (9.9481) 39.975 (11.439)
German citizenship 0.9179 (0.2745) 0.8959 (0.3054) 0.9116 (0.2838) 0.8693 (0.3370)
Educational attainment
No formal degree 0.0545 (0.2270) 0.1198 (0.3247) 0.0949 (0.2930) 0.1147 (0.3186)
Secondary school (9 years) 0.3263 (0.4689) 0.5438 (0.4981) 0.4351 (0.4958) 0.4922 (0.4999)
Secondary school (10 years) 0.4207 (0.4937) 0.2606 (0.4390) 0.3913 (0.4880) 0.2879 (0.4528)
Advanced technical college entrance 0.0664 (0.2490) 0.0335 (0.1799) 0.0320 (0.1760) 0.0365 (0.1875)
qualification (11–12 years)
General qualification for university 0.1321 (0.3386) 0.0424 (0.2015) 0.0468 (0.2111) 0.0687 (0.2530)
entrance (12–13 years)
Vocational attainment
No formal degree 0.1808 (0.3848) 0.3608 (0.4803) 0.2581 (0.4376) 0.3389 (0.4733)
Apprenticeship (dual system) 0.6253 (0.4841) 0.5507 (0.4974) 0.6406 (0.4798) 0.5519 (0.4973)
Specialized vocational school 0.0953 (0.2936) 0.0671 (0.2502) 0.0742 (0.2621) 0.0659 (0.2481)
University, technical college 0.0986 (0.2982) 0.0214 (0.1447) 0.0271 (0.1625) 0.0433 (0.2034)
Unemployment history
Share of unemployment in 1st year before programme entry 0.5942 (0.3080) 0.6566 (0.3057) 0.6040 (0.3062) 0.6225 (0.3192)
Share of unemployment in 2nd year before programme entry 0.2479 (0.3293) 0.3425 (0.3776) 0.2873 (0.3555) 0.3384 (0.3688)
Share of unemployment in 3rd year before programme entry 0.2065 (0.3164) 0.2853 (0.3682) 0.2454 (0.3432) 0.3016 (0.3743)
Share of unemployment in 4th year before programme entry 0.1851 (0.3032) 0.2524 (0.3567) 0.2215 (0.3274) 0.2715 (0.3643)
Employment history
Share of employment in 1st year before programme entry 0.2687 (0.3057) 0.2115 (0.2831) 0.2626 (0.3030) 0.2260 (0.2984)
Share of employment in 2nd year before programme entry 0.5176 (0.4276) 0.4364 (0.4162) 0.4993 (0.4284) 0.4181 (0.4186)
Share of employment in 3rd year before programme entry 0.5287 (0.4380) 0.4551 (0.4310) 0.5098 (0.4383) 0.4438 (0.4329)
Share of employment in 4th year before programme entry 0.5264 (0.4379) 0.4387 (0.4317) 0.5073 (0.4371) 0.4568 (0.4345)
ln(last daily income from employment) 3.8728 (0.6220) 3.7037 (0.6234) 3.7830 (0.5753) 3.7950 (0.6383)
Number of observations 22 843 13 744 20 042 520 488

Source: IEB, own calculations.

U. Rinne et al.
 Public training programmes in Germany 3471
Table 2. Sub-sample sizes by vocational education

Male Female

West Germany East Germany West Germany East Germany

Participant type 1 No degree 1746 413 919 288
Vocational degree 4429 4433 4240 2997
Participant type 2 No degree 2118 126 1457 133
Vocational degree 2699 503 2224 672
Participant type 3 No degree 2473 391 1251 248
Vocational degree 3090 3390 3873 2253
Nonparticipants No degree 72 866 13 449 52 164 9989
Vocational degree 104 923 63 125 85 006 50 384

Notes: Completed apprenticeships and degrees from a specialized vocational school, a university or a technical college are
considered as vocational degrees. Individuals who are 25 years or younger are excluded.
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

and type 3 (3%). Overall, the participants in type 1, contrast to the ‘dual system’, participants do not
which focus on classroom training, have on average have a contract with an employer; the theoretical
higher skills than participants in the alternative training is combined with internships and training
programmes with a more practical focus. within practice firms. University training is provided
Furthermore, they have a lower share of time spent by two types of institutions: universities and technical
in unemployment, a higher share of time spent in colleges. The university graduates, independent of the
regular employment and higher previous earnings. subject and the type of institution, are coded having a
As we focus on the effect heterogeneity of pro- vocational degree.9 Our classification of skill groups
gramme participation across skill and age groups, we closely follows Dustmann and Meghir (2005). The
divide our sample into sub-samples for each pro- importance of this distinction is emphasized, as the
gramme type. With respect to the skills, we stratify our authors find substantial differences between the two
sample into sub-samples for each programme type groups in terms of job mobility, wage growth and
according to the skill level (skilled and unskilled ), returns to experience. In their view, these differences
gender (male and female) and region (East and West have important implications for the design of ALMP.
Germany). Our classification of skilled and unskilled Furthermore, we exclude individuals who are 25 years
individuals is based on whether an individual had old or younger from this part of the analysis. Those
received a formal vocational degree before entering individuals, who constitute roughly 18% of the
the programme. We consider completed apprentice- original sample of participants, are just beginning
ships and degrees from a specialized vocational their working career and have a much lower proba-
school, university or technical college as vocational bility of already having obtained a vocational degree.
degrees. Most individuals obtain a vocational degree Table 2 shows that the resulting sample sizes are in
through the German apprenticeship system (cf. general reasonably large. However, we observe rela-
Table 1). The distinctive feature of this system is tively few East German participants without a
that it involves training both at the workplace and at vocational degree and for programme type 2 also
vocational school (‘dual system’).8 Firms that train few East German participants with a vocational
apprentices have to follow a prescribed curriculum, degree.
and apprentices take a rigorous external exam in their With respect to age, we stratify our sample into
trade at the end of the apprenticeship (Acemoglu and sub-samples for each programme type according to
Pischke, 1998). The apprenticeship usually lasts about age (three age groups), gender (male and female) and
3 years, and there are more than 340 trades which region (East and West Germany). The construction of
qualify: examples are doctor’s assistant, car mechanic, age groups is based on the arguments about the
plumber, etc. Other institutions providing vocational direction and extent of potential effect heterogeneity
training are specialized vocational schools. In discussed in Section I. Therefore, we stratify our
8
The apprentices have previously finished secondary education, which ranges from 9 to 13 years of schooling.
9
See e.g. Winkelmann (1996) for a broader and more general overview about the three pillars of the German educational
system: general schooling, vocational training and university education.
 3472 U. Rinne et al.
Table 3. Sub-sample sizes by age groups

Male Female

West Germany East Germany West Germany East Germany

Participant type 1 25 years 1176 1152 514 536
26–49 years 5459 3980 4582 2812
50 years 716 866 577 473
Participant type 2 25 years 2064 260 1270 218
26–49 years 3792 496 3064 599
50 years 1025 133 617 206
Participant type 3 55 years 1181 970 565 357
26–49 years 4853 3049 4383 1903
50 years 710 732 741 598
Nonparticipants 525 years 29 903 13 899 17 990 6790
26–49 years 127 600 55 121 98 034 42 465
50 years 50 189 21 453 39 136 17 908
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

sample into young individuals who are beginning their programmes (Y1) with the outcomes for the same
working career (25 years or younger), prime-aged individuals if they had not participated (Y 0). If D
workers (26–49 years old ) and older workers denotes participation, with D ¼ 1 if a person partici-
approaching retirement (50 years or older). The pates in the programme and D ¼ 0 otherwise, the
resulting sample sizes are depicted in Table 3. While actual outcome for individual i can be written as
the number of prime-aged participants is relatively
large across all programme types, the resulting sample Yi ¼ Y1i  Di þ Yi0  ð1  Di Þ ð1Þ
sizes for younger individuals and older workers are
The individual treatment effect would then be given
smaller. In the case of programme type 2, only few
by the difference Di ¼ Y1i  Yi0 . However, it is
individuals from East Germany participate.
impossible to calculate this difference because one
The success of programme participation is evalu-
of the outcomes is counterfactual. Instead, the
ated by looking at the probability of being employed
evaluation literature concentrates on the population
for a period of 24 months after programme entry. This
average gains from treatment – usually on the average
period is based on the fact that we focus on
treatment effect on the treated (ATT or DATT), which
programme participants in 2002, and we can observe
is formally given by
reliable data for all employment states until 31
December 2004. Individuals are regarded as DATT ¼ EðDjD ¼ 1Þ ¼ EðY1 jD ¼ 1Þ  EðY 0 jD ¼ 1Þ
employed if they hold a job in the primary labour
ð2Þ
market. For instance, participation in job creation
schemes is not included in this outcome measure. It is the principle task of any evaluation study to
Moreover, the administrative data set only includes find a credible estimate for the second term on
employment subject to social security contributions.10 the right hand side of Equation 2, which is
Hence, self-employment cannot be observed in our unobservable.
data. In addition, we evaluate the effect of programme One possible solution could be to simply compare
participation on monthly earnings in the primary the mean outcomes of participants and nonpartici-
labour market. We apply the described definition of pants. However, if EðY 0 jD ¼ 1Þ 6¼ EðY 0 jD ¼ 0Þ, esti-
employment and consider remunerations in terms of mating the ATT from the difference between the
monthly earnings associated with these spells. sub-population means of these two groups would
yield a selection bias. On the other hand, if treatment
assignment is strongly ignorable, i.e. if selection is on
III. Evaluation Approach observable characteristics X (unconfoundedness or
CIA) and if observable characteristics of participants
Ideally, one would like to compare the outcomes for and nonparticipants overlap (common support), the
the individuals participating in public training matching estimator is an appealing choice to estimate
10
This means that we do not observe self-employment earnings, and remunerations are only reported up to the social security
contribution ceiling.
 Public training programmes in Germany 3473
the desired counterfactual (Rosenbaum and Rubin, calculating the variance of the estimated treatment
1983). Under these conditions, the distribution of the effects based on bootstrapping procedures. Although
counterfactual outcome Y 0 for the participants is nearest neighbour matching does not satisfy the basic
the same as the observed distribution of Y 0 for the conditions for the bootstrap and the bootstrap
comparison group conditional on the vector of variance diverges from the actual variance (Abadie
covariates X. Formally, and Imbens, 2008), this alternative method leads to
similar variances of the estimated treatment effects
EðY 0 jX, D ¼ 1Þ ¼ EðY 0 jX, D ¼ 0Þ ð3Þ and does not change the implications presented
Applying this relation into Equation 2 allows the below.
ATT to be estimated by comparing mean outcomes The focus of the subsequent analysis lies on the
of matched participants and nonparticipants. differences in treatment effects between separate
Rosenbaum and Rubin (1983) show that if treatment sub-groups. To assess whether these differences are
assignment is strongly ignorable given X, it is also significantly different from zero, we assume that the
strongly ignorable given any balancing score that is a treatment effects follow a normal distribution and are
function of X.11 One possible balancing score is the independent from each other.12
The probability of participation in the three
propensity score P(X), which is the probability of
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

programme types under consideration is estimated

participating in a given programme. Mueser et al.
conditional on a number of observable characteristics
(2007) present evidence that if administrative data is
using binary probit models with participation as the
used to measure the performance of training pro-
dependent variable. These characteristics include
grammes, propensity score matching is generally
socio-demographic characteristics (e.g. age, national-
most effective.
ity, marital status), regional information (regional
There are several propensity score matching
type, unemployment rate), educational and voca-
methods suggested in the literature, see, e.g.
tional attainment, the employment history (4 years
Caliendo and Kopeinig (2008) for an overview.
prior to programme entry) and information on the
Based on the characteristics of our data, we apply
last employment spell (duration, income, business
nearest neighbour matching without replacement. sector).13 We run these regressions separately for the
This matching method has the advantage of being different sub-samples of participants and nonpartici-
the most straightforward matching estimator: a given pants according to programme type, gender, region
participant is matched with a nonparticipant closest (East and West Germany) and skill or age groups,
in terms of the estimated propensity score. We avoid respectively.
an increased variance of the estimator, as we match The matching method we apply is based on the
without replacement (Smith and Todd, 2005), which CIA. This is, in general, a very strong assumption and
is justified as the ratio between participants and hence its plausibility is crucial. Caliendo et al. (2008b)
nonparticipants – i.e. potential matching partners – is provide a good example of a careful discussion of this
comparatively high in our data. Hence, the con- issue. The implementation of matching estimators
structed counterfactual outcome is based only on requires choosing a set of variables simultaneously
distinct nonparticipants. To check the sensitivity of influencing the participation decision or, more gen-
our results with respect to the matching algorithm, erally, the selection process into the programme and
we apply additional methods to our data and find the outcome variable. For example, information
evidence for robust estimates (Section, ‘Sensitivity about socio-demographic characteristics and the
analysis’ for details). educational background is important in this regard
For the variance of the estimated treatment effects, to capture important factors that determine individ-
we base our inference on the assumption that the ual employment prospects, on which caseworkers
estimators are asymptotically normally distributed. then base their assignment decisions. Our data
This distribution is derived from the difference of two contain a variety of such information, as described
weighted means of two independent observations. above. Moreover, our data allow us to construct
Lechner (2002) employs a similar approach. We detailed employment history for each individual. It
checked the accuracy of this approximation also by seems reasonable to assume that the employment
11
When there are many covariates, it is impractical to match directly on covariates because of the curse of dimensionality.
See e.g. Zhao (2008) for some comments on this problem.
12
If we drop the assumption of independence and allow for nonzero correlation between treatment effects, implications only
change marginally.
13
The exact specifications are not reported here, but available upon request. Table A1 provides additional information about
the variables used in the regressions.
 3474 U. Rinne et al.
history also contains information about unobserved matched participants and nonparticipants, the former
variables not included in our data, e.g. motivation, condition builds on similar arguments Sianesi (2004)
attitudes and aptitudes. It is likely that such put forward. She argues that participation decisions
characteristics are persistent over time and have in ALMP are to be viewed sequentially over time in
already impacted on an individual’s labour market unemployment, as choices faced by unemployed
history before programme entry and are thus individuals are not whether to participate or not to
reflected therein. For this reason, conditioning on participate at all but rather whether to join a
past labour market outcomes is crucial for the CIA to programme now or not to participate for now.
hold (Heckman et al., 1998). Furthermore, regional According to this line of argumentation, it is funda-
labour market characteristics play an important role mental to ensure elapsed unemployment duration for
for employment prospects. Such information is matched treated and controls is the same.
included in our data (e.g. unemployment rate). However, we use programme entry as our point of
Altogether, it is quite plausible that once we reference rather than following entrants into unem-
condition on this rich set of variables, the conditional ployment over time. The estimates we present below
independence assumption will hold, and that there can thus be viewed as the outcome of the joining/
will not be any additional variables which jointly waiting-decision after the same elapsed duration of
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

influence the participation decision and the outcome unemployment for given individuals. Our approach
variable which are not reflected by our rich set of allows us to estimate the ATT for average partici-
variables. pants in given programme types in 2002, as opposed
After estimating the propensity score we match to the ATT for participants in given programme types
each participant with a distinct nonparticipant within of a specific entry cohort into unemployment. Exact
the different sub-samples by exact covariate matching matching on the previous unemployment duration
plus propensity score matching.14 The nonpartici- only considers the past up to the entry into the given
pants are required to have participated neither in the programme. Future outcomes are not considered in
respective type of public training programme before this context. In particular, nonparticipants can
nor in the quarter of the participant’s programme potentially participate in the given programme type
entry. The variables used for exact matching are after programme entry. Sianesi (2004) employs a
region, previous unemployment duration (in months) similar definition of nonparticipation. She argues for
and quarter of programme entry. Therefore, we the case of Sweden that, in principle, any unemployed
stratify the sub-samples by these variables first and individual will join a programme at some time,
then implement propensity score matching within provided he remains unemployed long enough. We
each cell without replacing the matched nonpartici- think Sweden is similar to Germany in this respect.
pant. In general, the overlap between the group of Hence, a restriction on future outcomes – i.e. to
participants and nonparticipants is sufficient in all require nonparticipation in the follow-up period after
sub-samples. Nonetheless, in some cases there are programme entry – affects estimated treatment effects
parts of the propensity score distribution where negatively (Stephan, 2008), as a substantial fraction
participants lack comparable nonparticipants. These of the ‘never treated’-individuals de facto leave the
individuals are dropped from our pool of partici- unemployment register.15
pants. We find comparable nonparticipants for more After forming the matched pairs, a suitable way to
than 95% of the sample of participants. This means assess the matching quality is comparing the stan-
that 2651 participants (1789 participants) are dardized bias before matching, SBb, to the standar-
dropped from the original sample when we match dized bias after matching, SBa. The standardized
within age groups (skill groups). biases are defined as
Our matching procedure thus ensures that matched
pairs are (a) previously unemployed for the same ðX 1  X 0 Þ
SBb ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ;
duration and (b) enter the programme in the same 0:5  ðV1 ðXÞ þ V0 ðXÞÞ
region and quarter. While the latter condition makes
ðX 1M  X 0M Þ
sure that regional and seasonal influences are con- SBa ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð4Þ
stant and that the observation period is the same for 0:5  ðV1M ðXÞ þ V0M ðXÞÞ

14
The matching algorithm is implemented using the PSMATCH2 Stata ado-package by Leuven and Sianesi (2003).
15
For instance, Lechner and Wunsch (2008) require nonparticipation in the follow-up period after programme entry for
comparison individuals. Applying the same definition of nonparticipation to our data lowers the estimated treatment effects
(Section ‘Sensitivity analysis’ for details). Although we opted for the above-stated definition of nonparticipation and do not
exclude future participants, the alternative approach clearly has the advantage of employing a very straightforward definition
of nonparticipation.
 Public training programmes in Germany 3475
Table 4. Matching quality within sub-samples by vocational education

Before matching After matching

Type Vocational education Sex Region Mean bias R2 Mean bias R2

1 No degree Male West Germany 10.411 0.049 1.205 0.004
1 No degree Male East Germany 9.043 0.038 2.886 0.012
1 No degree Female West Germany 12.059 0.060 1.414 0.004
1 No degree Female East Germany 9.118 0.036 3.714 0.011
1 Vocational degree Male West Germany 12.879 0.072 0.958 0.002
1 Vocational degree Male East Germany 11.939 0.057 1.083 0.001
1 Vocational degree Female West Germany 9.795 0.052 0.696 0.001
1 Vocational degree Female East Germany 12.054 0.052 1.116 0.002
2 No degree Male West Germany 9.327 0.076 2.191 0.008
2 No degree Male East Germany 12.762 0.078 5.040 0.042
2 No degree Female West Germany 10.251 0.062 2.081 0.006
2 No degree Female East Germany 13.599 0.086 6.991 0.072
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

2 Vocational degree Male West Germany 9.114 0.076 1.733 0.005

2 Vocational degree Male East Germany 10.016 0.070 3.530 0.011
2 Vocational degree Female West Germany 10.077 0.051 1.583 0.003
2 Vocational degree Female East Germany 9.430 0.055 2.672 0.011
3 No degree Male West Germany 9.595 0.040 1.065 0.002
3 No degree Male East Germany 12.099 0.081 2.951 0.015
3 No degree Female West Germany 9.123 0.036 1.439 0.004
3 No degree Female East Germany 13.218 0.078 4.340 0.023
3 Vocational degree Male West Germany 8.241 0.039 0.983 0.001
3 Vocational degree Male East Germany 9.932 0.045 1.034 0.002
3 Vocational degree Female West Germany 9.135 0.044 0.818 0.001
3 Vocational degree Female East Germany 7.915 0.031 1.144 0.002

Notes: Reported indicators refer to 75 variables that are at least included in the specification. Mean bias: mean standardized
bias; R2: pseudo-R2 of propensity score estimation. Completed apprenticeships and degrees from a specialized vocational
school, a university or a technical college are considered as vocational degrees. Individuals who are 25 years or younger are
excluded.

where X1(V1) is the mean (variance) in the treated and decrease substantially compared to before
group before matching and X0(V0) the analogue for matching.
the comparison group. X1M(V1M) and X0M(V0M) are Training programmes may have an influence on
the corresponding values after matching (Rosenbaum the employment probability as well as on the earnings
and Rubin, 1985). Following the example of Sianesi of the participants. Realized earnings are the product
(2004), we also re-estimate the propensity score on of the employment probability and the observed
the matched sample to compute the pseudo-R2 before individual earnings.17 In order to gain additional
and after matching. insight into the effects of participation on the
Tables 4 and 5 suggest that the quality of our (observed ) monthly earnings we compare the earn-
matching procedures is satisfactory: the mean stan- ings distributions next to the mean of the earnings
dardized biases of the matched samples are noticeably over time between treated and controls. From a
smaller than those of the unmatched sample. policy point of view, it is interesting to know to which
They are in general below 4% after matching.16 extent the share of individuals ending up in higher
Similarly, the pseudo-R2 after matching are fairly low paid jobs is increased by participating in

16
Exceptions are the sub-samples of East German participants in type 2 without a vocational degree or 50 years and older,
as well as female East German participants in type 3 without a vocational degree.
17
Lechner and Melly (2007) argue that realized earnings are only a ‘crude’ measure of the causal effect of training on
productivity and propose to estimate bounds for the earnings effects. However, this approach goes beyond the scope of this
article.
 3476 U. Rinne et al.
Table 5. Matching quality within sub-samples by age groups

Before matching After matching

Type Age Group Sex Region Mean bias R2 Mean bias R2

1 25 years Male West Germany 11.134 0.042 1.585 0.005
1 25 years Male East Germany 13.385 0.062 1.841 0.006
1 25 years Female West Germany 14.905 0.062 2.213 0.010
1 25 years Female East Germany 14.477 0.067 2.510 0.011
1 26–49 years Male West Germany 11.939 0.060 0.808 0.001
1 26–49 years Male East Germany 12.377 0.053 1.341 0.002
1 26–49 years Female West Germany 9.902 0.057 0.958 0.001
1 26–49 years Female East Germany 10.981 0.042 1.328 0.003
1 50 years Male West Germany 14.901 0.112 2.590 0.012
1 50 years Male East Germany 14.641 0.109 2.874 0.008
1 50 years Female West Germany 19.936 0.148 2.133 0.009
1 50 years Female East Germany 13.771 0.100 3.385 0.020
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

2 25 years Male West Germany 12.289 0.077 1.891 0.005

2 25 years Male East Germany 18.000 0.105 3.740 0.020
2 25 years Female West Germany 12.836 0.073 1.994 0.006
2 25 years Female East Germany 20.683 0.117 3.609 0.021
2 26–49 years Male West Germany 9.501 0.078 1.448 0.003
2 26–49 years Male East Germany 10.223 0.072 3.019 0.012
2 26–49 years Female West Germany 9.022 0.048 1.388 0.003
2 26–49 years Female East Germany 10.936 0.064 2.978 0.013
2 50 years Male West Germany 10.466 0.131 2.485 0.012
2 50 years Male East Germany 11.119 0.100 6.145 0.053
2 50 years Female West Germany 10.825 0.116 3.196 0.018
2 50 years Female East Germany 11.966 0.090 5.090 0.030
3 25 years Male West Germany 10.957 0.042 2.297 0.006
3 25 years Male East Germany 16.557 0.094 1.735 0.004
3 25 years Female West Germany 17.185 0.083 2.556 0.012
3 25 years Female East Germany 19.259 0.109 2.521 0.019
3 26–49 years Male West Germany 7.062 0.027 0.864 0.001
3 26–49 years Male East Germany 11.820 0.053 0.980 0.001
3 26–49 years Female West Germany 8.964 0.039 0.783 0.001
3 26–49 years Female East Germany 10.729 0.042 1.192 0.003
3 50 years Male West Germany 10.194 0.091 1.916 0.007
3 50 years Male East Germany 12.269 0.072 2.258 0.008
3 50 years Female West Germany 15.605 0.117 2.207 0.009
3 50 years Female East Germany 10.497 0.075 2.433 0.008

Notes: Reported indicators refer to 75 variables that are at least included in the specification. Mean bias: mean standardized
bias; R2: pseudo-R2 of propensity score estimation.

training programmes. This effect is given by compar- across programme types given that positive monthly
ing the shares of individuals entering a job above earnings are observed and compare the fraction of
certain thresholds or within a given strata. In contrast treated and controls within these thresholds for the
to a comparison of mean earnings, we can gather sub-groups under consideration.
information on whether new jobs are mainly lower or
higher paid jobs, given participation or nonparticipa-
tion. The mentioned thresholds (or strata) are in our
case based on the overall distribution of monthly IV. Results
earnings 2 years after programme entry. In other
words, we calculate quartiles of the earnings distri- After applying the matching approach as described
bution for participants and matched nonparticipants above, the ATT can be calculated as the difference in
 Public training programmes in Germany 3477
mean outcomes between the groups of matched when the median duration of the programme is
participants and nonparticipants. We present esti- reached, we observe increasing employment prob-
mates of differences in employment probabilities and abilities. In these graphs, the vertical line indicates the
monthly earnings from employment in the primary median duration of the programme for the corre-
labour market for a period of 2 years after programme sponding sub-group. 2 years after programme entry,
entry.18 While average treatment effects for the whole the point estimates differ slightly across gender: we
sample are discussed in Section ‘Average treatment observe a point estimate of about 8.0 percentage
effects’, the effect heterogeneity of these effects with points for male and 6.7 percentage points for female
respect to vocational education (i.e. across skill individuals.
groups) is considered in Section ‘Treatment effects: Our findings on the general effectiveness of type 2
vocational education’ and across age groups in are also rather positive. Although the effect of being
Section ‘Treatment effect: age’. We briefly describe locked-in in the programme is again apparent, we
differences in our results between East and West find that participation significantly increases the
Germany in Section ‘Treatment effect: East versus probability of being employed and begins about 8–9
West Germany’. The sensitivity of our results is then months after programme entry for both male and
assessed in Section ‘Sensitivity analysis’. female individuals. We observe a sharp increase in
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

employment probabilities 6 months after programme

Average treatment effects entry, i.e. exactly after the median duration of the
programme (indicated by the vertical line). 2 years
To obtain a general impression of the ATT on after programme entry, the point estimates are
employment probabilities and monthly earnings, we slightly lower than for type 1 but similar across
aggregate the matched sub-groups for each pro- gender. They amount to 5.2 percentage points for
gramme type separately by gender. We calculate men and 5.8 percentage points for women.
treatment effects as the difference in mean outcomes A positive impact of participation on employment
between participants and nonparticipants in the probabilities is also found for programme type 3.
resulting samples. Although this procedure was
We compute a point estimate of 5.6 percentage points
implemented both for the matched sub-samples
for men and 6.9 percentage points for women 2 years
previously stratified according to the level of voca-
after programme entry. The treatment effects become
tional education and with respect to age groups, the
significantly positive about 12 months after entering
former results are not reported in this section as they
the programme for both male and female individuals.
are virtually the same.19
We display ATT on employment probabilities and
monthly earnings effects for a period of 24 months Earnings effects. For all programme types and over
after programme entry. These effects consist of the whole 2-year period after programme entry, the
locking-in effects for the group of participants due ATT on monthly earnings (Fig. 5) do not exhibit
to reduced search activities while participating in a major differences compared to the ATT on employ-
programme (van Ours, 2004). As an opposing effect, ment probabilities described above. To give an idea of
we expect an increase in employment probabilities the magnitude of the monthly earnings effects 2 years
during and after completing the programme. after entering the programme, male (female) partici-
pants in type 1 earn about E125 (E100) more per
Employment probabilities. For programme type 1, month than comparable nonparticipants. These
we find that participation has a significantly positive estimates are lower for programme types 2 and 3,
impact on the probability of being employed for male but virtually the same across gender: participants in
individuals and begins about 12 months after pro- type 2 earn roughly E60 per month more than
gramme entry (Fig. 4). For female individuals, matched nonparticipants; and it is about E80 for
significantly positive treatment effects realize about participants in type 3.
14 months after programme entry. However, in Figure 6 displays the monthly earnings distribution
previous months the impact of being locked-in in along with the employment effects 2 years after
the programme leads to significantly negative point programme entry. Again, the above described posi-
estimates for both men and women. Around the time tive employment effects for each programme type can
18
We thus follow the prevailing approach in the recent evaluation literature. A different approach concentrates on treatment
effects only after the end of the programme. For advantages and disadvantages of both approaches see e.g. Caliendo and
Kopeinig (2008).
19
However, these results are available from the authors upon request. Minor differences are very likely due to the exclusion of
individuals who are 25 years or younger in our analysis of treatment effects across skill groups.
 3478 U. Rinne et al.
(a) Type 1

Men Women

.2
.2

ATT Employment Probability

ATT Employment Probability

.1
.1

0
0

−.1
−.1

−.2
−.2

0 6 12 18 24 0 6 12 18 24
Months since Program Entry Months since Program Entry

(b) Type 2
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Men Women

.2
.2

ATT Employment Probability

ATT Employment Probability

.1
.1

0
0

−.1
−.1

−.2
−.2

0 6 12 18 24 0 6 12 18 24
Months since Program Entry Months since Program Entry

(c) Type 3

Men Women
.2
.2

ATT Employment Probability

ATT Employment Probability

.1
.1

0
0

−.1
−.1

−.2
−.2

0 6 12 18 24 0 6 12 18 24
Months since Program Entry Months since Program Entry

Fig. 4. ATT employment probabilities

Notes: Thick lines are point estimates of the ATT based on aggregated matched sub-samples across age groups, while thin
lines represent 95% confidence intervals. The vertical line indicates the median duration of the programme.

be observed. Moreover, it is possible to assess to E1080, the third quartile up to E1499, the second up
which type of job (in terms of monthly earnings) the to E1980, and the first quartile includes monthly
positive employment effects lead. Therefore, indivi- earnings above the latter threshold.
duals with earnings from regular employment in the The graphs show that participation in type 1
primary labour market are divided into quartiles. The increases the share of both male and female
fourth quartile includes gross monthly earnings up to individuals in better paid jobs, while the fraction of
 Public training programmes in Germany 3479
(a) Type 1

Men Women

300

300
200

200
100

100
ATT Earnings

ATT Earnings
0

0
−100

−100
−200

−200
−300

−300
0 6 12 18 24 0 6 12 18 24
Months since Program Entry Months since Program Entry

(b) Type 2
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Men Women
300

300
200

200
100

100
ATT Earnings

ATT Earnings
0

0
−100

−100
−200

−200
−300

−300

0 6 12 18 24 0 6 12 18 24
Months since Program Entry Months since Program Entry

(c) Type 3

Men Women
300

300
200

200
100

100
ATT Earnings

ATT Earnings
0

0
−100

−100
−200

−200
−300

−300

0 6 12 18 24 0 6 12 18 24
Months since Program Entry Months since Program Entry

Fig. 5. ATT monthly earnings

Notes: Gross monthly earnings from employment (in E). Thick lines are point estimates of the ATT based on the aggregated
matched sub-samples across age groups, while thin lines represent 95% confidence intervals. The vertical line indicates the
median duration of the programme.

participants in the bottom quartile is about the same nonparticipants, while the share of both groups is
as for comparable nonparticipants. For male partici- virtually the same in the top quartile. Finally,
pants in type 2, we observe increased shares in all participation in type 3 increases the share of both
quartiles of the earnings distribution. Female parti- male and female individuals in jobs with higher
cipants in this type exhibit larger shares in the bottom earnings. While the fraction of female participants in
quartiles of the distribution than matched this type also increases in the bottom quartile
 3480 U. Rinne et al.
(a) Type 1

Men Women

.8

.8
Non−Participants Non−Participants
.7 Participants Participants

.7
.6

.6
.5

.5
Shares

Shares
.4

.4
.3

.3
.2

.2
.1

.1
0

0
Not Employed 4th Quartile 3rd Quartile 2nd Quartile 1st Quartile Not Employed 4th Quartile 3rd Quartile 2nd Quartile 1st Quartile

(b) Type 2
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Men Women
.8

.8
Non−Participants Non−Participants
Participants Participants
.7

.7
.6

.6
.5

.5
Shares

Shares
.4

.4
.3

.3
.2

.2
.1

.1
0

Not Employed 4th Quartile 3rd Quartile 2nd Quartile 1st Quartile Not Employed 4th Quartile 3rd Quartile 2nd Quartile 1st Quartile

(c) Type 3
Men Women
.8

.8

Non−Participants Non−Participants
Participants Participants
.7

.7
.6

.6
.5

.5
Shares

Shares
.4

.4
.3

.3
.2

.2
.1

.1
0

Not Employed 4th Quartile 3rd Quartile 2nd Quartile 1st Quartile Not Employed 4th Quartile 3rd Quartile 2nd Quartile 1st Quartile

Fig. 6. Monthly earnings distribution and employment effects 24 months after programme entry
Notes: Quartiles are based on the distribution of monthly earnings in the matched samples, aggregated across age groups and
programme types. 4th quartile: gross monthly earnings 5E1080; 3rd quartile: gross monthly earnings E1080–1499; 2nd
quartile: gross monthly earnings E1500–1980; 1st quartile: gross monthly earnings 4E1980.

compared to nonparticipants, it is about the same for individuals with and without a vocational degree.
male participants and nonparticipants. Below, we present treatment effects for skilled and
unskilled participants in the three programme types
separately by gender.
Treatment effects: vocational education
To analyse the effect heterogeneity of treatment Employment probabilities. For all three types, the
effects across skill groups, we distinguish between resulting treatment effects across all sub-groups are
 Public training programmes in Germany 3481
quite similar within each type. Nonetheless, some skilled individuals (Fig. 7c). In line with this pattern,
differences appear between the three types, and some we observe the smallest treatment effect 2 years after
small differences occur between sub-groups within programme entry for unskilled men (4.0 percentage
each type. points) and the largest effect for skilled women (7.0
For type 1, significantly positive treatment effects percentage points) (cf. Table 6). However, only these
can be observed about 13 months after programme two effects can be regarded as significantly different
entry for individuals with a vocational degree (Fig. from each other. All other differences between
7a). For male individuals without a vocational treatment effects across sub-groups are insignificant,
degree, we find significantly positive treatment effects even if we relax the assumption of independence and
9 months after programme entry. On the other hand, allow for nonzero correlation.
significantly positive ATT can only be observed 15 In summary, it is important to note that we find
months after programme entry for women without a only two significant differences between sub-groups
vocational degree. The point estimates of the treat- with respect to the ATT 2 years after programme
ment effects 2 years after entering the programme are entry. For programme type 2, men without a degree
rather similar across the different sub-groups, with gain significantly less by participating in training than
the exception of male individuals without a voca- women without a degree; and for programme type 3,
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

tional degree: while the estimated ATT for skilled the treatment effect for unskilled men appears to be
individuals and unskilled women lie between 8.1 and significantly lower than for skilled women. No other
8.7 percentage points, the corresponding effect significant differences are observed. Therefore, treat-
amounts to 10.6 percentage points for unskilled ment effects across skill groups within the same
men (cf. Table 6). However, as the latter treatment gender cannot be regarded as significantly different
effect is rather imprecisely estimated, it cannot be from each other. This observation is independent of
regarded as significantly different from the estimated the type of content: whether the programmes have a
effects for the other sub-groups. more theoretical focus (type 1) or are more practically
For type 2, treatment effects become significantly oriented (types 2 and 3).
positive around 9 months after programme entry for
all sub-groups except male individuals with a voca- Earnings effects. There are also some differences in
tional degree (Fig. 7b). It takes about 12 months to monthly earnings effects across sub-groups within the
observe significantly positive treatment effects in this three programme types, as in the case for the ATT on
sub-group. We generally observe a sharp increase in employment probabilities previously discussed. Once
employment probabilities exactly after the median again, the treatment effects on monthly earnings for
duration of the programme (indicated by the vertical all three programme types do not exhibit any major
line). The magnitude of the estimated treatment differences compared to the ATT on employment
effects 2 years after programme entry varies slightly probabilities, see Fig. A1 and Table 6.20
across skill groups (cf. Table 6). While the point For type 1, Fig. A2a shows the monthly earnings
estimates for skilled participants are positive but distribution along with the employment effects 2
relatively moderate (3.8 percentage points for men years after programme entry. Across all sub-groups,
and 4.6 percentage points for women), treatment an additional fraction of participants enter jobs in the
effects are larger for unskilled participants in type 2 top quartiles of the earnings distribution. The
(5.2 percentage points for men and 7.2 percentage increase in the top quartile is significantly positive
points for women). Hence, unskilled women in for all sub-groups except unskilled women.
particular seem to outperform the other sub-groups. Furthermore, the share of participants in the
However, the treatment effect for this sub-group can bottom quartile of the earnings distribution is virtu-
only be regarded as significantly different from the ally the same as the share of matched nonparticipants
effect for one sub-group (men without a degree) if we for men and skilled women. It is slightly higher for
allow for nonzero correlation between treatment unskilled women.
effects; and it cannot be regarded as significantly When looking at the monthly earnings distribution
different from the estimated effects for the other for type 2, two years after programme entry (Fig.
sub-groups at all. A2b), we can distinguish a slightly different impact of
For type 3, the estimated ATT become significantly programme participation for men and women: while
positive about 12 months after programme entry in we observe higher shares of male participants in all
all sub-groups. The overall pattern indicates slightly quartiles of the earnings distribution, we find that
larger treatment effects for women as well as for additional jobs of women appear to be mainly located
20
Figures A1–A4 are included in the Appendix of this article.
 Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

3482

Fig. 7. ATT employment probabilities by vocational education

Notes: Thick lines are point estimates of the ATT for the respective sub-group, while thin lines represent 95% confidence intervals. Completed apprenticeships and degrees
from a specialized vocational school, a university or a technical college are considered as vocational degrees. Individuals who are 25 years or younger are excluded. The
vertical line indicates the median duration of the programme.
U. Rinne et al.
 Public training programmes in Germany 3483
Table 6. ATT 24 months after programme entry by vocational education

Employment probability Monthly earnings

Type Sex Vocational education NP P DATT (SE) NP P DATT (SE)

1 Male No degree 0.2086 0.3147 0.1061 (0.0151) 289.45 460.81 171.36 (26.03)
1 Male Vocational degree 0.3065 0.3895 0.0830 (0.0082) 503.94 632.29 128.35 (17.42)
1 Female No degree 0.1684 0.2551 0.0868 (0.0183) 170.60 299.25 128.64 (25.16)
1 Female Vocational degree 0.2716 0.3530 0.0814 (0.0088) 352.65 453.67 101.02 (15.17)
2 Male No degree 0.1890 0.2414 0.0523 (0.0130) 242.31 315.32 73.01 (20.21)
2 Male Vocational degree 0.2509 0.2889 0.0380 (0.0122) 394.10 443.82 49.72 (23.84)
2 Female No degree 0.1800 0.2518 0.0718 (0.0167) 165.61 247.83 82.23 (19.41)
2 Female Vocational degree 0.2389 0.2851 0.0461 (0.0140) 266.45 304.50 38.05 (18.90)
3 Male No degree 0.2226 0.2626 0.0400 (0.0122) 314.43 380.15 65.72 (20.48)
3 Male Vocational degree 0.2895 0.3448 0.0552 (0.0088) 430.18 516.76 86.57 (16.58)
3 Female No degree 0.1820 0.2419 0.0599 (0.0152) 186.13 236.72 50.58 (18.56)
3 Female Vocational degree 0.2516 0.3218 0.0702 (0.0090) 279.97 366.63 86.67 (12.82)
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Notes: NP: nonparticipants; P: participants; DATT: average treatment effect on the treated. Completed apprenticeships and
degrees from a specialized vocational school, a university or a technical college are considered as vocational degrees.
Individuals who are 25 years or younger are excluded.

in the bottom quartiles. For both skilled and unskilled their working career (25 years or younger),
women, the share of participants in the fourth prime-aged workers (26–49 years old ) and older
quartile of the earnings distribution is significantly workers approaching retirement (50 years or older).
higher than for comparable nonparticipants.
For type 3, we find once more only minor Employment probabilities. The extent to which the
differences between skilled and unskilled individuals ATT on employment probabilities vary between age
within the same gender, but a different pattern for men groups is relatively small. In most cases, differences
and women (Fig. A2c). Additional jobs of male between sub-groups are not significant – and if so,
participants, irrespective of their skill level, appear to primarily when comparing male and female partici-
be located in the top three quartiles of the earnings pants. However, while we observe significant positive
distribution; and the share of male individuals in the effects across gender and programme types for all
bottom quartile is virtually the same for participants prime-aged individuals, for some of the younger and
and matched nonparticipants. On the other hand, we older subgroups we observe only insignificant posi-
observe significantly higher shares of both skilled and tive effects at the end of our observation period.
unskilled female participants in the bottom quartile of For type 1, we observe significant effects for the
the distribution. There are also higher shares of skilled three male age groups (Fig. 8a; Table 7). The point
female participants in the top quartiles (a notably estimates after 24 months are the largest for
significantly higher share in the top quartile), but the prime-aged individuals (8.5 percentage points),
shares in the top quartiles of unskilled women are which is in line with our hypothesis we derived in
virtually the same for participants and nonparticipants. Section I. However, they are not statistically different
In sum, our analysis of the treatment effects on from the ATT for younger (6.4 percentage points)
earnings across skill groups reveals some differences and older (6.3 percentage points) unemployed. For
between sub-groups. These differences, however, are women we see a similar picture: the point estimate of
almost exclusively between men and women, whereas the ATT is the highest for the prime-aged individuals
differences between skill groups of the same gender (7.6 percentage points), but this effect is not
are basically negligible. This result holds both for significantly different for younger and older females.
programmes with a more theoretical focus and for The treatment effect for the younger group does not
programmes with a more practical focus. become significantly positive during our observation
period, the one for the older group is significantly
different from zero only at some points in time, but
Treatment effects: age
for both groups the point estimates are positive at the
For the analysis of the employment effects of training end of our observation period (3.1 and 3.6 percentage
programmes with respect to age, we distinguish three points, respectively). If we compare male and female
age groups: young individuals who are beginning individuals, we find that prime-aged males show
 3484 U. Rinne et al.
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Fig. 8. ATT employment probabilities by age groups

Notes: Thick lines are point estimates of the ATT for the respective sub-group, while thin lines represent 95% confidence
intervals. The vertical line indicates the median duration of the programme.
 Public training programmes in Germany 3485
Table 7. ATT 24 months after programme entry by age groups

Employment probability Monthly earnings

Type Sex Age group NP P DATT (SE) NP P DATT (SE)

1 Male 25 years 0.3616 0.4260 0.0644 (0.0165) 466.18 568.96 102.78 (27.73)
1 Male 26–49 years 0.3110 0.3962 0.0852 (0.0079) 507.15 643.53 136.38 (16.47)
1 Male 50 years 0.1728 0.2362 0.0634 (0.0166) 225.37 301.21 75.84 (29.84)
1 Female 25 years 0.3807 0.4112 0.0306 (0.0256) 432.09 451.38 19.30 (35.46)
1 Female 26–49 years 0.2828 0.3583 0.0755 (0.0087) 349.48 459.43 109.95 (14.55)
1 Female 50 years 0.1740 0.2096 0.0356 (0.0204) 143.16 225.19 82.03 (27.94)
2 Male 25 years 0.3542 0.4015 0.0473 (0.0152) 472.28 527.94 55.66 (25.09)
2 Male 26–49 years 0.2357 0.2817 0.0460 (0.0101) 361.34 409.33 47.99 (19.00)
2 Male 50 years 0.1454 0.2174 0.0720 (0.0176) 194.98 313.00 118.01 (33.94)
2 Female 25 years 0.3159 0.3872 0.0713 (0.0230) 325.44 403.49 78.05 (28.95)
2 Female 26–49 years 0.2343 0.2911 0.0568 (0.0118) 259.27 306.82 47.56 (16.14)
2 Female 50 years 0.1488 0.1804 0.0317 (0.0226) 139.62 167.54 27.93 (25.90)
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

3 Male 25 years 0.3531 0.4217 0.0686 (0.0163) 445.27 555.73 110.46 (26.37)
3 Male 26–49 years 0.2830 0.3368 0.0539 (0.0079) 423.20 508.85 85.65 (14.59)
3 Male 50 years 0.1636 0.2081 0.0445 (0.0157) 217.66 256.59 38.93 (28.34)
3 Female 25 years 0.3601 0.4196 0.0595 (0.0264) 342.27 453.41 111.14 (33.43)
3 Female 26–49 years 0.2515 0.3302 0.0787 (0.0088) 282.45 374.35 91.90 (12.25)
3 Female 50 years 0.1799 0.1998 0.0199 (0.0172) 168.60 178.81 10.21 (21.19)

Note: NP: nonparticipants; P: participants; DATT: average treatment effect on the treated.

significantly larger treatment effects than young and more relevant for younger than older cohorts.
older women. However, none of these effects are significantly
The treatment effects for type 2 exhibit no signif- different from each other. The direction of effect
icant differences at all across subgroups 2 years heterogeneity for women across age groups, however,
after programme entry (Fig. 8b; Table 7). Interesting is in line with our initial argumentation. We estimate
to note is that we observe similar treatment effects the largest ATT at the end of our observation period
2 years after programme entry for young and for prime-aged female individuals (7.9 percentage
prime-aged male individuals (4.7 and 4.6 percentage points), and obtain lower estimates for both younger
points) and obtain a larger point estimate for older and older women (6.0 and 2.0 percentage points,
men (7.2 percentage points). For females, we estimate respectively). In particular, the estimated ATT for
the largest ATT for young women (7.1 percentage older women appear to be small and not significantly
points) and the smallest for older women (3.2 different from zero during most of the observation
percentage points), which is not significantly different period – and it is statistically different from the point
from zero. Although the differences between the age estimate for prime-aged individuals at the end of our
groups are not statistically different, the point observation period.
estimates suggest that for males the argument that Overall, we find some weak evidence that partic-
practical training may be more important for younger ipation in training programmes impacts differently on
than for older individuals is questionable. In contrast different age groups for females, and that this may
to that, the point estimates for females are in line with vary with the type of programme. However, we find a
this argument. positive ATT on employment probabilities in at least
For male individuals participating in type 3, we do one training programme for all age groups.
not find a pattern which is consistent with our initial
hypothesis about the direction of effect heterogeneity Earnings effects. The few differences in monthly
across age groups (Fig. 8c; Table 7). On the contrary, earnings effects across sub-groups which we observe
the estimated treatment effects for this sub-groups are similar to the ones just discussed, see Fig. A3 and
decline in age: 2 years after programme entry, the Table 7. Especially for women not in the prime-aged
largest effect is obtained for young men (6.7 percent- group, we find small point estimates for some groups,
age points) and the smallest is found for older men which are not different from zero in some cases.
(4.5 percentage points). However, this decline is in These are similar to the employment probabilities for
line with the argument that practical training may be younger women in type 1 and older women in types 2
 3486 U. Rinne et al.
and 3. The impact of participation in public training Germany. While the pattern of treatment effects in
programmes on the monthly earnings distribution West Germany is similar to the results previously
2 years after programme entry is depicted in Fig. A4. discussed, it appears that in East Germany only some
For type 1, the shares of participants located in the sub-groups who participate in types 1 or 3 exhibit
top quartile of the monthly earnings distribution significantly positive treatment effects. This is the
2 years after programme entry are significantly higher case for skilled individuals, male individuals across all
in most sub-groups than the respective shares of age groups and prime-aged women. However, treat-
comparison individuals (Fig. A4a). The only excep- ment effects are rather imprecisely estimated in the
tions are older male and younger female individuals, sub-groups for which we do not find significantly
who do not exhibit any significant differences positive effects. An analysis which is based on a larger
between the shares of participants and nonpartici- number of observations in East Germany is therefore
pants. On the other hand, the differences between the desirable to deliver deeper insight into the direction
shares in the bottom quartile of the earnings distri- and extent of effect heterogeneity in this region.
bution are in general not significant. Only a fraction
of prime-aged male participants is significantly lower Sensitivity analysis
than that of control observations.
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

For types 2 and 3, the overall picture is less To assess the sensitivity of our results with respect to
consistent than for type 1. For most sub-groups, we the matching method, we additionally employ some
do not find any significant differences between the alternative algorithms. Besides nearest neighbour
share of participants and nonparticipants – neither in matching without replacement, on which the above
the top quartile, nor in the bottom quartile. Although described results are based on, we calculate treatment
the share of participants in the top quartile of the effects based on (a) nearest neighbour matching with
monthly earnings distribution is in most sub-groups replacement, (b) caliper matching without replace-
higher than the share of nonparticipants, we find ment (with a maximum tolerance level of 0.001) and
significantly increased fractions only for male partici- (c) radius matching (with a maximum tolerance level
pants in type 2 above 49 years as well as for male and of 0.001). The results based on these three procedures
female participants in type 3 who are between 26 and very closely reflect those presented above. This is in
49 years at programme entry. On the other hand, we line with Mueser et al. (2007), who also report quite
find significantly higher shares of female participants similar results across a variety of matching methods
in the bottom quartile of the earnings distribution in when these methods are based on the same set of
some age groups: female participants in type 2 below control variables.
26 years, female participants in type 2 between 26 and As mentioned earlier, one could choose in principle
49 years and prime-aged female participants in type 3. a stricter definition of nonparticipation. Lechner and
Wunsch (2008), for instance, require future nonparti-
cipation for potential members of the control group.
Treatment effects: East versus West Germany
In contrast to our results, they find a negative impact
Heterogeneous treatment effects could also be mate- of participation on employment probabilities 2 years
rialized by region. In particular, there could be after programme entry for most analysed programme
differences in the programmes’ effectiveness between types. If we also use a similar – and thus stricter –
East and West Germany. Although our participants definition of nonparticipation and require that
enter the programme more than 10 years after nonparticipants do not participate in any public
reunification, economic conditions in both regions training programme for a period of 2 years after
still differ considerably. entering the programme, we exclude almost 30% of
We differentiate between East and West Germany, our nonparticipants. Moreover, we find that exclud-
gender and programme type, and our results are in ing these individuals affects our results: in general, the
line with previous findings on programme effective- ATT on the probability of being employed 2 years
ness in East Germany (see, e.g. Biewen et al., 2007). after programme entry are substantially lower. This
In particular for women in East Germany, we change is mainly driven by increasing employment
observe – at best – only slightly significantly positive probabilities among matched nonparticipants.
treatment effects. But in general, and particularly for Moreover, the point in time at which the ATT
type 2, the number of observations for East Germany becomes positive occurs in most sub-groups around
is considerably low compared to the overall sample 1–4 months later with this alternative definition of
(cf. Tables 1–3). Therefore, we are able to draw only nonparticipation. However, we still find positive ATT
tentative conclusions with respect to the effect heter- for all sub-groups considered and the overall picture
ogeneity across skill and age groups in East of relatively homogeneous treatment effects remains
 Public training programmes in Germany 3487
unchanged. More specifically, the analysis of treat- between age groups is relatively small. However,
ment effects across skill and age groups leads to the while we observe significantly positive effects across
same conclusion: in general, we do not observe a gender and programme types for all prime-aged
pattern which is consistent with larger treatment individuals, we observe only insignificantly positive
effects for prime-aged or skilled individuals. effects for some of the younger and older sub-groups
Another source of bias in our primary approach at the end of our observation period. Nonetheless, we
may stem from the fact that nonparticipants enter find at least one programme type with a positive
training at a later stage; and by doing so, they may impact on subsequent employment prospects for each
experience the locking-in effect of participation at the age group.
time when the outcome variable is measured. To get Our results thus – at least in part – conflict with the
a rough estimate of the likely importance of such strategy to provide training increasingly to indivi-
effects, we first calculate the average locking-in effect duals with better employment prospects. This strat-
of participation, i.e. the mean (negative) treatment egy has been implemented in Germany as a part of
effect until the median programme duration ends. the reform of ALMP in 2003. Following the reform,
Subsequently, we subtract 30% of this number from caseworkers are now asked to evaluate the employ-
the point estimates of ATT, as this is the share of ment prospects of the unemployed in advance and
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

participants who enter training at a later stage. As a provide training only to individuals with a relatively
result, we see that later involvement in programmes high probability of entering employment after train-
does not cause a large bias. The estimated treatment ing participation. Age and skill level are important
effects are only slightly lower (between 0.6 and 3.0 determinants of future employment prospects. Our
percentage points). The bias, however, varies across findings suggest that an exclusion of unskilled unem-
skill and age groups because average locking-in effect ployed from training programmes would not increase
also vary in this regard. More specifically, the bias is the effectiveness of this policy measure, but instead
less important for unskilled and older individuals.21 would prevent potential gains for this group of
disadvantaged individuals. On the other hand, our
results suggest that better targeting of training
programmes in Germany might be promising, espe-
V. Conclusion cially with respect to age.
The results derived in this study are specific for the
This article studies the effects of participation in German context and its institutional setting, includ-
public training programmes for the unemployed in ing its educational and retirement system. In addi-
Germany. We apply propensity score matching tion, we show that the programme content may
methods and estimate the treatment effects for matter for the extent of effect heterogeneity.
participants entering training in 2002 using a rich Therefore, it is not easy to draw general conclusions
administrative data set. We focus on treatment effects for the effect heterogeneity of training programmes in
for different skill and age groups, next to average other countries. However, one general conclusion
treatment effects on the treated. which can be drawn from our results is that an
Considering three medium-term programme types, exclusion of individuals with disadvantageous
with a median duration between 6 and 8 months and employment prospects from participation in training
together accounting for roughly 85% of all partici- programmes does not generally increase the effec-
pants in public training programmes, our results tiveness of training programmes. Instead, this study
indicate that programme participation has on average underlines that better targeting is only possible on the
a positive impact on employment prospects for all basis of thoroughly measured treatment effects.
programme types. We present little evidence for the This study finds weak but interesting evidence that
presence of heterogeneous treatment effects, and the the programme content – i.e. the focus on practical
magnitude of these differences is quite small. With or theoretical training – might interact with the age of
respect to skill level, we find significantly positive the participants. This indicates that the effectiveness
impacts on the employment prospects of participants of an investment in human capital for unemployed
for all sub-groups, both in terms of employment may depend on (a) the type of an investment and (b)
probabilities and earnings. Although we find a the timing within the employment biography. It
slightly different picture with respect to age, the seems to be an interesting question for further
extent to which the average treatment effects vary research to analyse these interdependencies and the
21
The detailed results of Sections ‘Treatment effect: East versus West Germany’ and ‘Sensitivity analysis’ are not reported
here but available from the authors upon request.
 3488 U. Rinne et al.
underlying production function of human capital in (Eds) D. L. Millimet, J. A. Smith and E. J. Vytlacil,
more detail, including potential differences across the Emerald Group Publishing, Bingley, UK, pp. 383–430.
Caliendo, M. and Kopeinig, S. (2008) Some practical
skill distribution and between occupations. guidance for the implementation of propensity score
matching, Journal of Economic Surveys, 22, 31–72.
Caliendo, M. and Steiner, V. (2005) Aktive
Arbeitsmarktpolitik in Deutschland: Bestands-
aufnahme und Bewertung der mikroökonomischen
Acknowledgements Evaluationsergebnisse, Journal for Labour Market
The data used in this article originate from the Research, 38, 396–418.
Dustmann, C. and Meghir, C. (2005) Wages, experience
evaluation of public training programmes as part of and seniority, The Review of Economic Studies, 72,
the evaluation of the proposals of the Hartz 77–108.
Commission. Schneider et al. (2007) contains details. Eichhorst, W. and Zimmermann, K. F. (2007) And then
We would like to thank Marco Caliendo, Marton there were four. . . how many (and which) measures of
Csillag, Michael Lechner, Peter Mueser, Hilmar active labor market policy do we still need?, Applied
Economics Quarterly, 53, 243–72.
Schneider, Zhong Zhao, two anonymous referees as Fitzenberger, B., Osikominu, A. and Völter, R. (2010) Get
well as participants at the ESPE Conference in training or wait? Long-run employment effects of
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Chicago, the IZA Conference on the Evaluation of training programmes for the unemployed in West
Labor Market Programmes in Bonn, the EALE Germany, Annales d’Economie et de Statistique,
forthcoming.
Conference in Oslo, the EEA Meeting in Budapest,
Heckman, J. J., Ichimura, H., Smith, J. A. and Todd, P.
the CAPE Conference in Nuremberg, the Conference (1998) Characterizing selection bias using experimental
of the German Statistical Society in Kiel, the VfS data, Econometrica, 66, 1017–98.
Conference in Munich and the Seminar at the George Hujer, R., Thomsen, S. L. and Zeiss, C. (2006) The effects
Washington University for the valuable discussions of vocational training programmes on the duration of
unemployment in Eastern Germany, Allgemeines
and helpful comments. Arne Uhlendorff also thanks Statistisches Archiv, 90, 299–322.
DIW DC, where part of this research was pursued Hujer, R. and Wellner, M. (2000) Berufliche Weiterbildung
during his stay in fall 2007. All remaining errors are und individuelle Arbeitslosigkeitsdauer in West- und
our own. Ostdeutschland: eine mikroökonometrische Analyse,
Mitteilungen aus der Arbeitsmarkt- und
Berufsforschung, 33, 405–19.
Jacobebbinghaus, P. and Seth, S. (2007) The German
integrated employment biographies sample IEBS,
References Schmollers Jahrbuch, 127, 335–42.
Jaenichen, U. and Stephan, G. (2009) The effectiveness of
Abadie, A. and Imbens, G. W. (2008) On the failure of the targeted wage subsidies for hard-to-place workers,
bootstrap for matching estimators, Econometrica, 76, Applied Economics. DOI: 10.1080/00036840802600426.
1537–57. Kluve, J. (2010). The effectiveness of European active labor
Acemoglu, D. and Pischke, J.-S. (1998) Why do firms train? market policy, Labour Economics. DOI: 10.1016/
Theory and evidence, The Quarterly Journal of j.labeco.2010.02.004.
Economics, 113, 79–119. LaLonde, R. J. (2003) Employment and training pro-
Baumgartner, H. J. and Caliendo, M. (2008) Turning grammes, in Means Tested Transfer Programmes in the
unemployment into self-employment: effectiveness of US (Eds) M. Feldstien and R. Moffitt, University of
two start-up programmes, Oxford Bulletin of Chicago Press for the National Bureau of Economic
Economics and Statistics, 70, 347–73. Research, Chicago, USA, pp. 517–86.
Biewen, M., Fitzenberger, B., Osikominu, A. and Lechner, M. (1999) Earnings and employment effects of
Waller, M. (2007) Which programme for whom? continuous off-the-job training in East-Germany after
Evidence on the comparative effectiveness of public unification, Journal of Business and Economic
sponsored training programmes in Germany, IZA Statistics, 17, 74–90.
Discussion Paper No. 2885, Institute for the Study of Lechner, M. (2000) An evaluation of public-
Labor (IZA), Bonn. sector-sponsored continuous vocational training pro-
Büsch, V., Dahl, S.-A. and Dittrich, D. A. V. (2009) An grammes in East Germany, Journal of Human
empirical study of age discrimination in Norway and Resources, 35, 347–75.
Germany, Applied Economics, 41, 633–51. Lechner, M. (2002) Some practical issues in the evaluation
Caliendo, M., Hujer, R. and Thomsen, S. L. (2008a) of heterogeneous labour market programmes by
Identifying effect heterogeneity to improve the effi- matching methods, Journal of the Royal Statistical
ciency of job creation schemes in Germany, Applied Society, Series A, Statistics in Society, 165, 59–82.
Economics, 40, 1101–22. Lechner, M. and Melly, B. (2007) Earnings effects of
Caliendo, M., Hujer, R. and Thomsen, S. L. (2008b) The training programmes, IZA Discussion Paper No. 2926,
employment effects of job creation schemes in Institute for the Study of Labor (IZA), Bonn,
Germany: a microeconometric evaluation, Germany.
in Estimating and Evaluating Treatment Effects in Lechner, M., Miquel, R. and Wunsch, C. (2007) The curse
Econometrics, Volume 21 of Advances in Econometrics and blessing of training the unemployed in a changing
 Public training programmes in Germany 3489
economy: the case of East Germany after unification, methods that incorporate the propensity score, The
German Economic Review, 8, 468–509. American Statistican, 39, 33–8.
Lechner, M., Miquel, R. and Wunsch, C. (2010) Long-run Schneider, H., Brenke, K., Jesske, B., Kaiser, L., Rinne, U.,
effects of public sector sponsored training in West Schneider, M., Steinwede, J. and Uhlendorff, A. (2007)
Germany, Journal of the European Economic Evaluation der Maßnahmen zur Umsetzung der
Association, forthcoming. Vorschläge der Hartz-Kommission – Bericht 2006,
Lechner, M. and Wunsch, C. (2008) What did all the money IZA Research Report No. 10, Institute for the Study of
do? On the general ineffectiveness of recent west Labor (IZA), Bonn, Germany.
German labour market programmes, Kyklos, 61, Sianesi, B. (2004) An evaluation of the active labor market
134–74. programmes in Sweden, The Review of Economics and
Leuven, E. and Sianesi, B. (2003) PSMATCH2: Stata Statistics, 86, 133–55.
module to perform full Mahalanobis and propensity Smith, J. A. and Todd, P. E. (2005) Does matching
score matching, common support graphing, and overcome LaLonde’s critique of non-experimental
covariate imbalance testing. Statistical Software estimators?, Journal of Econometrics, 125, 305–53.
Components, Boston College Department of Stephan, G. (2008) The effects of active labor market
Economics. Available at http://ideas.repec.org/c/boc/ programmes in Germany: an investigation using
bocode/s432001.html (accessed 26 February 2010). different definitions of non-treatment, Jahrbücher für
Mueser, P. R., Troske, K. R. and Gorislavsky, A. (2007) Nationalökonomie und Statistik, 228, 586–611.
Using state administrative data to measure programme van Ours, J. C. (2004) The locking-in effect of
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

performance, The Review of Economics and Statistics, subsidized jobs, Journal of Comparative Economics,
89, 761–83. 32, 37–55.
Rosenbaum, P. R. and Rubin, D. B. (1983) The central role Winkelmann, R. (1996) Employment prospects and skill
of the propensity score in observational studies for acquisition of apprenticeship-trained workers in
Germany, Industrial and Labor Relations Review, 49,
causal effects, Biometrika, 70, 41–55.
658–72.
Rosenbaum, P. R. and Rubin, D. B. (1985) Constructing a
Zhao, Z. (2008) Sensitivity of propensity score methods to
control group using multivariate matched sampling
the specifications, Economics Letters, 98, 309–19.
 Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Appendix
3490

Fig. A1. ATT monthly earnings by vocational education

Notes: Gross monthly earnings from employment (in E). Thick lines are point estimates of the ATT for the respective sub-group, while thin lines represent 95% confidence
intervals. Completed apprenticeships and degrees from a specialized vocational school, a university or a technical college are considered as vocational degrees. Individuals
who are 25 years or younger are excluded. The vertical line indicates the median duration of the programme.
U. Rinne et al.
 Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Public training programmes in Germany

Fig. A2. Monthly earnings distribution and employment effects 24 months after programme entry, by vocational education
Notes: Quartiles are based on the distribution of monthly earnings in the matched samples, aggregated across programme types. 4th quartile: gross monthly earnings
5E1110; 3rd quartile: gross monthly earnings E1110–1529; 2nd quartile: gross monthly earnings E1530–2040; 1st quartile: gross monthly earnings 4E2040. Completed
apprenticeships and degrees from a specialized vocational school, a university or a technical college are considered as vocational degrees. Individuals who are 25 years or
younger are excluded.
3491
 3492 U. Rinne et al.
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Fig. A3. ATT monthly earnings by age groups

Notes: Gross monthly earnings from employment (in E). Thick lines are point estimates of the ATT for the respective
sub-group, while thin lines represent 95% confidence intervals. The vertical line indicates the median duration of the
programme.
 Public training programmes in Germany 3493
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

Fig. A4. Monthly earnings distribution and employment effects 24 months after programme entry, by age groups
Notes: Quartiles are based on the distribution of monthly earnings in the matched samples, aggregated across programme
types. 4th quartile: gross monthly earnings 5E1080; 3rd quartile: gross monthly earnings E1080–1499; 2nd quartile: gross
monthly earnings E1500–1980; 1st quartile: gross monthly earnings 4E1980.
 3494 U. Rinne et al.
Table A1. Description of variables

Variable Description
Socio-demographic characteristics
Age Age at programme entry
Nation Nationality: three categories (German, EU, non-EU)
disabled Disability: four categories (not disabled, three degrees of disability)
marstat Marital status: four categories (single, single parent, single in cohabitation, married )
kid03 Youngest child 0–3 years in household (dummy variable)
kid414 Youngest child 4–14 years in household (dummy variable)
nkids Number of dependent children in household
Regional information
region Place of residence at programme entry: 12 regions reflecting clusters of similar macroeconomic
conditions (e.g. GDP growth, unemployment rate)
unemp_rate Unemployment rate in local FEA district at programme start
Educational attainment
school1 No formal degree
Downloaded by [Florida Atlantic University] at 13:41 10 November 2014

school2 Secondary school (9 years)

school3 Secondary school (10 years)
school4 Advanced technical college entrance qualification (11–12 years)
school5 General qualification for university entrance (12–13 years)
Vocational attainment
voceduc1 No formal degree
voceduc2 Apprenticeship (dual system)
voceduc3 Specialized vocational school
voceduc4 University, technical college
Previous employment
business Business sector of last employment: five categories (agriculture, manufacturing, construction,
services, other/missing)
duration Duration of last employment in days
income Daily income in last employment
Employment history
p-emp-y1 Share of time spent in regular employment in 1st year before programme entry
p-emp-y2 Share of time spent in regular employment in 2nd year before programme entry
p-emp-y3 Share of time spent in regular employment in 3rd year before programme entry
p-emp-y4 Share of time spent in regular employment in 4th year before programme entry
p-unemp-y1 Share of time spent in unemployment in 1st year before programme entry
p-unemp-y2 Share of time spent in unemployment in 2nd year before programme entry
p-unemp-y3 Share of time spent in unemployment in 3rd year before programme entry
p-unemp-y4 Share of time spent in unemployment in 4th year before programme entry
Income history
trinc-y1 Average daily transfer income in 1st year before programme entry
trinc-y2 Average daily transfer income in 2nd year before programme entry
trinc-y3 Average daily transfer income in 3rd year before programme entry
trinc-y4 Average daily transfer income in 4th year before programme entry
labinc-y1 Average daily labour income in 1st year before programme entry
labinc-y2 Average daily labour income in 2nd year before programme entry
labinc-y3 Average daily labour income in 3rd year before programme entry
labinc-y4 Average daily labour income in 4th year before programme entry

Notes: Variables used in the matching specifications (at least). The regressions were run separately for the sub-samples of
participants and nonparticipants previously stratified with respect to programme type, gender, region (East and West
Germany) and level of vocational education or age group.