Should Welfare Administration be Centralized or Decentralized? Evidence from a Policy Experiment



The 2005 reform of the German welfare system introduced two competing organizational models for welfare administration. In most districts, a centralized organization was established where local welfare agencies are bound to central directives. At the same time, 69 districts were allowed to opt for a decentralized organization. We evaluate the relative success of both types in terms of integrating welfare recipients into employment. Compared to centralized organization, decentralized organization has a negative effect on employment chances of males. For women, no significant effect is found. These findings are robust to the inclusion of aspects of internal organization common to both types of agencies.

1. Introduction

Studies from economics, management and organization theory suggest that the form of organization of an institution, particularly the centralization or decentralization of responsibilities, may have far-reaching implications for their outcomes (e.g., Besley and Coate, 2003; Richardson et al., 2002; Hutchcroft, 2001). Also, in the case of welfare administration, different organizational systems are likely to result in different incentives and strategies and can influence the success of integrating unemployed welfare recipients into employment. Given that public welfare spending accounts for a significant portion of total government expenditure and given that labor market integration of welfare recipients is the principal task of the public welfare administration for the unemployed, the improvement of organizational effectiveness is a question of foremost economic importance.

One key component in the organization of welfare administration is the degree of local autonomy. In a decentralized setting, local authorities are responsible for the activation of welfare recipients and act independently from central directives and guidelines. Conversely, in a centralized structure, welfare administration is organized by a countrywide government agency that issues directives on how the activation of welfare recipients should be implemented at the local level. Theoretical arguments in favor of a decentralized organization are based on the idea that local authorities are better informed about the characteristics of the local labor market. They are assumed to have detailed knowledge about the specific regional attributes relevant for a successful activation process, and, therefore, they are effective in providing services that are tailored to local conditions. Centralized organizations, on the other hand, are often considered to have an advantage in bundling resources, collecting information from various sources and imposing best-practice strategies for its local offices (e.g., Finn, 2000).

The degree of local autonomy of welfare administration varies considerably across countries. In the Netherlands, local authorities form the basis of the public welfare system. In the UK, in contrast, public welfare administration is part of the central government structure. In other countries, welfare reform has changed the degree of centralization of welfare administration. The 1996 US welfare reform, for instance, devolved greater program authority from the federal level to the states, and the Canadian reform that same year gave greater discretion to the provinces (Blank, 2002).

Even though there is an increasing evaluation literature concerning the effectiveness of active labor market programs (ALMP) and certain elements of welfare reform (most of them from the US, Germany or other European countries), evidence of the effects of the welfare system organization is scarce.1 One reason for this is that centralization or decentralization applies to countries as a whole, which makes it difficult to disentangle the effects of a particular organizational setting from other aspects of the welfare system or its reform. So far, conclusions are derived from case studies only (e.g., Lindsay and McQuaid, 2008; Tergeist and Grubb, 2006). To the best of our knowledge, this is therefore the first study to provide a quantitative assessment of the relative performance of a centralized and a decentralized organization of welfare administration.

We exploit the 2005 reform of the German welfare system that introduced two competing types of organization – a centralized and a decentralized one – in an otherwise homogenous institutional framework. Both approaches were pursued in parallel for a fixed period of time, after which the more successful model should be determined.2 In most of the 439 German districts, a centralized organization was established, in which the welfare agencies are subject to the directives and guidelines of the Federal Employment Agency. However, 69 districts were initially3 allowed to opt-out in favor of a decentralized organization legally and organizationally independent from central directives and guidelines. All other components of public welfare and labor market policy – such as benefit entitlements, the tax-benefit system in general and labor market institutions such as minimum wages and employment protection – apply equally to the centralized and decentralized systems of welfare administration.

On the basis of a unique data set compiled from surveys of welfare administration, Federal Employment Agency (FEA) register data, comprehensive surveys of welfare recipients and extensive regional information, we evaluate the relative performance of the two organizational systems in terms of successful integration of welfare recipients into the labor market. For our purpose, successful integration means that an unemployed welfare recipient takes up employment without receiving public welfare transfers any longer.4 Since the decentralized organization was adopted voluntarily, we need to worry about systematic selectivity. To estimate the effect, we apply a propensity score matching estimator to make sure that regional and individual characteristics are balanced between both types of organization. In addition, we address the possibility of further confounding factors at regional level in a number of robustness checks. A limitation of our study is due to the fact that our data are restricted to the year 2007; thus, we are not able to estimate long-run effects.

The estimated effects show that decentralized welfare agencies are less successful than centralized welfare agencies in placing male welfare recipients in employment; for female welfare recipients, the point estimates are also negative, but mostly not statistically significant. We also estimate the effects for persons living as singles and persons in non-single households separately. The results tend to be more pronounced for singles. Finally, we investigate whether the effects hinge on centralization itself or are due to internal organization features, exploiting data on the organizational strategies applied in the welfare agencies. We find that the significant negative effect of decentralized welfare agencies on employment for men is largely robust to the inclusion of further organizational details. The superior performance of centralized as compared to decentralized welfare agencies, therefore, is due to inherent differences between the two types of organization, and not to the adoption of particular forms of internal organization.

2. The German Welfare Reform of 2005

Before 2005, the same organization of welfare administration applied to all 439 districts (in German Kreise and kreisfreie Städte) in Germany. There were two different types of welfare benefits: unemployment assistance and social assistance, which were administered by two different authorities. The centrally organized FEA, represented by the local employment offices, was in charge of unemployment assistance, a means-tested benefit for long-term unemployed individuals whose claims to unemployment insurance benefits had expired. In contrast, local authorities were responsible for social assistance, a benefit for individuals who were not eligible for unemployment assistance or unemployment insurance benefits. This organization of the welfare system, with its two distinct administrative bodies, was often judged as overly fragmented (Tergeist and Grubb, 2006; Eichhorst et al., 2010) and resulted in disincentives with respect to integration into the labor market.

In January 2005, the welfare system reform merged unemployment and social assistance into a single benefit, the so-called unemployment benefit II (UBII). In contrast to unemployment assistance, and similar to the former social assistance, UBII is not conditional on former earnings. To be eligible for UBII, persons must be aged between 15 and 64 and must be able to work for at least 15 hours per week. Means-testing takes into account the wealth and income of all individuals living in the household. Individuals who are employed but have insufficient household income are also eligible for the benefit. Recipients of UBII are obliged to actively look for work and to participate in the welfare-to-work programs that are assigned to them. An important part of the reform was the reorganization of the welfare agencies: After the reform, for each district, all welfare services (benefit payments, counseling, labor market activation, etc.) were provided by one welfare agency, as opposed to the previous division of tasks and responsibilities between two administrative entities.

However, there was no political consensus on where the new welfare agencies should be established: within the system of the centralized FEA or decentralized at the level of local authorities. Ultimately, the legislator mandated a policy experiment and the evaluation of the relative performance of the two competing models. In the majority of the 439 German districts, local employment offices and local authorities formed a joint venture that is subject to the central controlling standards of the FEA [centralized agencies; in German: Arbeitsgemeinschaft (ARGE)]. Within the joint venture, the FEA is in charge of the administration of benefits, job placement and the application of the main instruments of ALMP. In particular, guidelines for these aspects and technical standards as computer software of the FEA are binding for centralized districts. Local authorities are in charge of administering payments for housing costs and for additional needs. Moreover, they provide counseling in specific contexts such as lone parent families, home care for elderly or disabled relatives, or alcohol and drug addiction.5

Of the 439 German districts, 69 were allowed to opt for a decentralized organization of welfare administration [decentralized agencies, in German: zugelassener kommunaler Träger (zkT)]. Under this system, local authorities autonomously operate the entire activation process including counseling, benefits disbursement, job placement, and the allocation of benefit recipients to ALMP. In particular, local welfare agencies are legally and organizationally independent from central directives and guidelines in the decentralized system.

Table 1 summarizes the key characteristics of centralized and decentralized welfare agencies. Decentralized welfare agencies can adopt their own organization and integration strategies, and are not bound to central directives. In contrast, centralized agencies are subject to central directives and codes of best-practice.6 In both the centralized and the decentralized systems, the largest share of welfare payment is financed by the federal government; only a small fraction of overall expenditure – identical in all districts – is taken from local tax budgets. This is different from reforms in other countries, where budgets have been shifted to local authorities as part of the decentralization process. We are therefore able to investigate the effect of decentralized organization independently of budgetary matters.

Table 1. Organizational features of decentralized and centralized welfare agencies
 Decentralized agenciesCentralized agencies
  1. Notes: The numbers of decentralized and centralized welfare agencies presented here refer to October 2006 and are based on the 439 German districts at this time.

Number of Entities69370
Legal FormPart of local administrationPart of FEA, but is a separate legal entity
Organizational AffiliationLocal authoritiesJoint venture between local employment office of the FEA and local authorities
Main Source of FinancingFederal governmentFederal government
Centralized Standards of FEANot binding, although legal restrictions existBinding for job placement, provision of ALMP, monitoring of efforts
SoftwareSpecific solutions for each local authorityStandard system of FEA

To evaluate the relative performance of both regimes, it is important to understand selection of districts into the two types. The overall number of decentralized districts (69) is equal to the number of deputies in the Bundesrat, the second chamber of the German Parliament. The maximum number of centralized districts in each federal state ranges between three and six, corresponding to the state's number of deputies in the Bundesrat. Individual districts could apply to the federal states to opt-out of the centralized system. In cases of excess demand, the state government made a selection from the applying districts. In several federal states, the maximum number of decentralized districts was not exhausted. The vacant places could then be filled by the districts not selected from other states in the first round. With respect to the regional distribution of applications, it appears that the selection process was strongly influenced by political affiliations of the state governments. In two states, Lower Saxony and Hesse, where the conservative governments were strongly in favor of the decentralized system, 13 districts were allowed to opt-out, even though these states only had six and five seats in the Bundesrat, respectively. In contrast, hardly any districts from Mecklenburg-Western Pomerania or Rhineland-Palatinate applied, both of which were run at that time by social democrats. Thus, the rules for selection resulted in a concentration of decentralized agencies in certain states (WZB, infas and FHS Frankfurt/M, 2008).

On 1 January 2012, a further 41 welfare agencies were admitted to choose the decentralized organization. Given the decline of long-term unemployment in Germany, many more counties and municipalities found it attractive to engage in counseling welfare recipients at this time. The move to the decentralized model by some municipalities was further pushed by the uncertainty whether the centralized model would be constitutional7 and reports of differences in approach between the public employment service and local municipalities.

3. Description of the Data

3.1. The estimation sample

To investigate whether centralized or decentralized welfare agencies are more successful in integrating welfare recipients into employment, we use a unique data set that was specifically created for our research question.8 For a random sample of 51 of 69 decentralized agencies, the aim was to identify regional units that were comparable in terms of labor market characteristics prior to the reform, but chose the centralized organization after the reform. The motivation for this is that although evidence suggests that the adoption of a decentralized system was driven by the political affiliation of the state governments (WZB, infas and FHS Frankfurt/M, 2008), some association could remain between local labor market characteristics and the opt-out from centralized welfare administration. Therefore, the distribution of regional characteristics is accounted for in the sampling procedure, leading to a data set of 154 pre-selected districts (out of a total of 439 German districts): the 51 decentralized welfare agencies sampled and 103 centralized welfare agencies selected on the basis of comparability.

The selection of comparable districts is explained in detail in Arntz et al. (2006).9 On the basis of a comprehensive description of the regional labor market situation until 2004 (before the reform took place), the authors chose variables that are relevant (i.e., significant at the 2.5% level) to the transition of the long-term unemployed into the labor market Among others, these include the local unemployment rate, the share of long-term unemployed, the creation of job vacancies, the share of individuals on welfare, local GDP, population size and urbanization, local transfer payments and active labor market policies. In a second step, the authors used this reduced set of relevant regional variables to identify districts that are (apart from the different forms of organization) comparable in terms of labor market outcomes of long-term unemployed. To this purpose, they applied the matching algorithm suggested by Zhao (2004). The latter defines the matching distance between two districts as the sum of a weighted difference in the districts’ regional variables, where the weights are a function of each variable's predictive power with respect to the individual labor market transition, such that more relevant characteristics obtain a higher weight.

Appendix 1 illustrates the regional location of the districts in our sample. Appendix 2 shows that the matching of regions equalizes the (unweighted) means and distributions of the relevant regional variables over the 154 agencies of our sample. The table reveals that equality of means and distributions cannot be rejected for the majority of the variables. The only exceptions are some variables that depend on the degree of urbanization of the district such as, for example, the share of commuters, the rate of social assistance recipients and the ratio of working population to resident population. Here, the mean in centralized districts is slightly higher than it is in decentralized districts. Individuals tend to have longer durations in welfare in urban areas than in rural areas (Bundesagentur für Arbeit, 2010). We will account for the possible effects of these differences in the estimation below.

3.2. Available information

To obtain data on the organizational structure of the welfare agencies, repeated interviews were conducted with the agencies’ management and staff in the 154 sample units. These surveys have been used to build aggregate measures of the type of case management, the activation concept, the placement strategies and the mix of ALMP. In addition, a wide range of regional variables (e.g., unemployment rates, welfare ratios, GDP, population density, share of foreigners) were collected on district-level for several months before and after the 2005 reform.

The individual-level data consist of a survey of welfare recipients who were registered at the 154 agencies. Between January and April 2007, 100–300 telephone interviews were conducted within each agency; the number of interviews depended on the size of the welfare agency. In total, nearly 20,300 individuals were interviewed who were drawn from the stock of UBII recipients in October 2006. This sampling scheme could impose a difficulty for the estimation of the relative effects of decentralized and centralized welfare agencies since the sample was not drawn in January 2005 (when the reform was introduced) but in 2006, that is, more than one year after the implementation of the reform.10 The reason for this delay is that the disruptions caused by the reform created considerable problems for the quality of administrative data during several months after the introduction of the reform.11 Because a large share of UBII recipients depend on welfare benefits for an extended period of time, the stock sample covers those individuals for whom the organization of welfare administration matters the most.

The survey data include individual characteristics (gender, age, marital and parental status, education, health and disability status, migration background, etc.), information on members of the household (number and age of household members and respondent's relation to them), and details concerning the labor market status and labor market history (current labor market state, former spells of insured and minor employment, former spells of unemployment, receipt of welfare benefits, participation in activation programs). Moreover, information is available about basic skills (e.g., reading, writing, math and computer skills), further qualifications (e.g., driver's license), job search activities and the concessions that respondents would be willing to make to obtain a new job.

The survey data were linked with administrative data from the FEA at the individual level. The administrative data include daily information about periods of employment and unemployment, job seeking, participation in ALMP and benefit receipt. This information allows for the construction of comprehensive labor market histories of the sampled individuals. An overview of the available information is given in Table 2. However, no information on social assistance is available for the time before 2005. Due to the decentralized administration of social assistance (see above), there were no uniform standards of data collection and storage. This also fostered the need to conduct the large-scale survey.

Table 2. Overview on characteristics included in the analysis
Basic sociodemographic informationAge (18–24, 25–34, 35–44 and 45–57 years), schooling (secondary general school, intermediate secondary school, university entrance diploma, other), migration background, household size (one person, two persons, three or more persons), no. of children (no children, one child, two or more children)
Obstacles to employmentDisability, care obligation
Labor market and employment historyStatus before welfare receipt ((minor) employment), no. of half-months unemployed in 2004, no. of half-months unemployed in 2003, no. of half-months unemployed in 2002, no. of half-months unemployed in 2001, no. of half-months out of labor force from 2001 to 2004, mean duration out of labor force from 2003 to 2004 in half-months, no. of programs from 2003 to 2004, mean duration of programs from 2003 to 2004 in half-months
Current welfare spellMonths in welfare before 10/2006, start after 10/2006 or missing
Regional informationUnemployment ratio (binary), urban district, GDP per employed person (binary), population density (binary), labor market conditions (above average, on average, below average), East Germany
Further sociodemographic variablesAt least one child aged below three in the household, lone parent status, vocational qualification (none, in-firm training, off-the-job training, university degree, other), self-assessment of overall state of health (good, satisfactory, poor), impairments to health (gastro-intestinal diseases, cardiovascular diseases, rheumatism and other articular trouble, sleep disorders, nervous disorders, allergies, back complaint, other complaints, no health problems), self-assessment of daily working capacity (less than three hours, three to six hours, six to eight hours, eight or more hours), self-assessment of basic skills (reading and writing in mother tongue, mathematics, emails and internet), driver's license
Further information on the labor market history from 2001 to 2004No. of half-months employed in 2004, no. of half-months employed in 2003, no. of half-months employed in 2002, no. of half-months employed in 2001, no. of half-months seeking for a job while employed in 2004, no. of half-months seeking for a job while employed in 2003, no. of half-months seeking for a job while employed in 2002, no. of half-months seeking for a job while employed in 2001, no. of half-months in a program in 2004, no. of half-months in a program in 2003, no. of half-months in a program in 2002, no. of half-months in a program in 2001, no. of employment spells in 2004, no. of employment spells in 2003, no. of employment spells in 2002, no. of employment spells in 2001
No. of unemployment spells in 2004, no. of unemployment spells in 2003, no. of unemployment spells in 2002, no. of unemployment spells in 2001, no. of spells of job seeking while employed in 2004, no. of spells of job seeking while employed in 2003, no. of spells of job seeking while employed in 2002, no. of spells of job seeking while employed in 2001, no. of programs in 2002, no. of programs in 2001, no. of spells out of labor force in 2004, no. of spells out of labor force in 2003, no. of spells out of labor force in 2002, no. of spells out of labor force in 2001

The information used for the outcome variable is also provided by the FEA and indicates for each month between January and December 2007 the employment status of individuals.12 We define employment without welfare receipt as the outcome of interest. In this case, gross labor earnings (plus any income from other sources such as capital earnings) exceed the income threshold which limits eligibility for welfare benefits.13 Because our analysis focuses on integration into employment, we restrict the sample to individuals who were unemployed at the time they entered the welfare system and at the time of sampling. Furthermore, we restrict the data to persons aged between 18 and 57 years. Persons aged 58 or older are no longer required to actively search for employment but may remain on welfare benefits until they reach the official retirement age of 65. Individuals aged 15–17 years are subject to compulsory schooling and cannot be expected to take up employment. Due to these restrictions, we have 13,286 observations in the estimation sample (4,489 persons from districts with decentralized welfare organization and 8,797 from districts with centralized organization).

4. Estimation Approach

To evaluate the relative performance of decentralized vs. centralized organization on the individual level, we estimate the Average Treatment Effect on the Treated (ATT) by propensity score matching, where the organization of the local welfare administration is used as the treatment variable. For identification of the ATT, the Conditional Independence Assumption (CIA) must hold (Lechner, 2001) so that conditional on observable covariates, the potential outcome is independent of the organizational model. Since the decision on welfare administration was made non-randomly by municipalities but on the basis of characteristics such as the regional labor market situation, we include regional information in our matching approach and make various checks to ensure that regional heterogeneity is properly accounted for.

In addition, we take into consideration that the composition of welfare recipients may differ between – and possibly due to – centralized and decentralized welfare districts. A first potential reason for this is selectivity into UBII receipt. In determining whether claimants are able to work and, thus, eligible to UBII, welfare agencies possess a considerable degree of leeway. If ability criteria differ systematically between centralized and decentralized welfare agencies, this may result in a different composition of welfare recipients with regard to characteristics such as illness or disability.14 A second potential reason is that our sample is drawn with a time lag after the start of the treatment. Changes in the composition of welfare recipients after the reform are likely to be influenced by the treatment. Thus, even if treatment were initially as good as randomly assigned at the regional level, differences in behavior of agencies with respect to inflow and activation require that we perform matching at the individual level.

We have access to sociodemographic characteristics beyond the standard set of controls such as migration background, basic mathematics, literacy and computer skills, self-assessed working capacity (measured in hours per day) and obstacles to employment such as provision of long-term care of relatives. In addition, detailed information on the labor market history of each individual, including frequency and duration of employment, unemployment, job seeking activity, ALMP participation and benefit receipt between 2001 and 2004 as well as on the recent labor market state is available. Thus, we are confident that our rich data include the factors importantly affecting both the treatment and the outcome.

We choose different specifications to check the robustness of the estimated treatment effects. The first specification contains the most important individual characteristics (age, schooling, migration background, household size, number of children, obstacles to employment and several indicators for labor market history) as well as the duration of the current welfare spell and limited regional information.15 Based on the results of balancing tests, this parsimonious specification is our preferred choice. To the latter, we add further regional information in the second specification. The third specification contains the full set of covariates.16

Since many evaluation studies have found the effectiveness of labor market activation to differ between genders (e.g., Bergemann and van den Berg, 2008; for a survey on recent findings for Europe), all estimations are done separately for men and women. Furthermore, note that activation by welfare agencies targets households as a whole. Only for single households, this is the unit which may also be integrated in employment. For this reason, we look also at single and multiperson households separately. Descriptive statistics for all variables included in the different propensity score specifications and the results for the propensity score models with the preferred specification (baseline specification) are provided in an online Appendix.17

For estimation, we use kernel density matching on the (estimated) treatment propensity score with bootstrapped standard errors and 250 replications (see Heckman et al., 1999 for an overview on ATT estimation with matching).18 Individuals residing in the same district are affected by common shocks, which may affect the efficiency of the estimates (e.g., Moulton, 1986, 1990). We account for this by estimating clustered standard errors at the agency level using the non-overlapping block bootstrap.

To assess the quality of matching, we apply four balancing tests: (1) the t-test for mean differences in each of the covariates included in the propensity score between matched individuals in centralized and decentralized agencies, (2) the standardized difference test of Rosenbaum and Rubin (1985), (3) re-estimation of the propensity score in the matched sample and checking whether the explained treatment variation is close to zero as measured by the McFadden-R2 (see Sianesi, 2004), and regressing each covariate on a 4th order polynomial of the propensity score, the treatment indicator, and the interaction between both and testing whether the coefficients on the interaction are jointly zero (see Smith and Todd, 2005).

As can be seen from the results of the balancing tests depicted in Table 3, matching quality is very satisfactory.19 According to Table 3, the mean standardized difference in per cent is strongly reduced after matching. The McFadden-R2 estimates of the third test are almost zero after matching and almost all of the variables included in the propensity score model pass the Smith and Todd (2005) test. In addition, the Online Appendix shows that the equality in means of variables in the propensity score specification between individuals in centralized and decentralized agencies cannot be rejected in just about any case.

Table 3. Indicators for matching quality
  1. Notes: McFadden-R2 derives from a probit estimation of the propensity score on all covariates considered. The LR-statistic and the corresponding p-value derive from a likelihood-ratio test of the joint insignificance of all covariates. The mean standardized difference in per cent has been calculated as an unweighted average of all covariates. The Smith–Todd test displays the number of covariates passing the test at the indicated significance level. There are 26 covariates included in the preferred specification.

Before Matching
Mean standardized difference in per cent6.3096.648
After Matching
Mean standardized difference in per cent1.0031.271
Smith and Todd (2005) balancing test
p-values >0.052118
p-values >0.012320

5. Empirical Results

Before presenting the estimation results, we briefly describe the development of our outcome variable, employment without welfare receipt, where we distinguish between individuals who are registered at centralized and decentralized welfare agencies (see Figure 1). For men, employment rates in districts with centralized welfare agencies are larger than they are in districts with decentralized organization. By December 2007, we observe a mean difference of about one and a half percentage points between decentralized and centralized welfare agencies (16.8% for centralized and 15.2% for decentralized welfare agencies). There is no difference between the two organizational models for women.

Figure 1.

Means of the outcome variable ‘employment without welfare receipt’

Note: Displayed are results for 2007 from raw data; sampling date: October 2006. All sampled persons are receiving welfare benefits at sampling date.

Our econometric analysis is consistent with these descriptive findings. As discussed in section 'Estimation Approach', we use three different specifications of the propensity score (with baseline denoting the preferred specification). The estimated treatment effects of decentralized welfare agencies are presented in Figures 2 and 3 for both men and women. To provide a benchmark and to indicate the changes due to the matching step, we also display the unmatched difference between centralized and decentralized welfare agencies. Rather than showing treatment effects at a single observation date, we display their evolution over the course of 2007, the year after sampling.

Figure 2.

Estimated treatment effects on employment, men

Notes: ♦ indicates significance at the 5% level, ◊ significance at the 10% level; displayed results for 2007; sampling date: October 2006.

Figure 3.

Estimated treatment effects on employment, women

Notes: ♦ indicates significance at the 5% level, ◊ significance at the 10% level; displayed results for 2007; sampling date: October 2006.

For men, we observe a negative treatment effect, that is, decentralized welfare agencies are less successful than centralized agencies in placing welfare recipients in jobs that provide a sufficient living income. The absolute effect rises from one to over three percentage points from January to August 2007, and declines moderately thereafter. These magnitudes are slightly larger than the unmatched differences. We interpret this in the sense that the composition effects mentioned in section 'Estimation Approach' introduce a positive bias in the treatment effect of decentralized agencies. The effects for May to November are significant at the 5% level, with t-statistics ranging from 1.96 to 2.91. With the exception of April, the effects for the other months are significant at the 10% level. The inclusion of further covariates leaves the estimated effects virtually unaffected (see Figure 2).

Given the relatively small fraction of people taking up employment (Figure 1), the effects for men are substantial. The largest estimated effect of nearly −3.5 percentage points, estimated for August 2007, implies that decentralized agencies have an about 24% lower integration quota than centralized agencies. For women, we also find negative treatment effects, which are however smaller in magnitude and not statistically significant (see Figure 3). Again, the results are insensitive to the specification of the propensity score and similar in magnitude to the unmatched difference between centralized and decentralized welfare agencies.

Gender differences are also present when we split the sample into single and non-single households (see Figures 4 and 5 providing estimated treatment effects based on the baseline specification of the propensity score models). For single men, we estimate a substantially negative employment effect of decentralized welfare agencies. The negative effect amounts up to 4.5 percentage points in absolute terms. For single women, we observe a negative treatment effect, too. This effect, however, is only slightly significant at the beginning of our observation period. Thereafter, it is insignificant and of smaller magnitude than the effect found for single men. In case of non-single men, we estimate a negative treatment effect of decentralized welfare agencies which has an absolute value of up to 2.9 percentage points. This effect is of smaller magnitude than the effect found for single men, but it is larger than the effect for non-single women. For the latter subgroup, we cannot establish a significant treatment effect.

Figure 4.

Treatment effects on employment, singles and non-singles, men

Notes: ♦ indicates significance at the 5% level, ◊ significance at the 10% level; displayed results for 2007; sampling date: October 2006.

Figure 5.

Treatment effects on employment, singles and non-singles, women

Notes: ♦ indicates significance at the 5% level, ◊ significance at the 10% level; displayed results for 2007; sampling date: October 2006.

To discuss potential reasons for the gender differences in our results, we refer to the study of IAQ, FIA and GendA (2009). This study combines analyses of survey and administrative data with case studies within welfare agencies. It shows that women are less intensively activated than men, irrespectively of the agency type. In particular, women are less frequently assigned to ALMP programs than men (see also Thomsen and Walter, 2010b; and Boockmann et al., 2011).

In addition, according to Chapter 10 of Book II of the German Social Code, parents of small children under the age of three years may not be activated at all. According to the results of IAQ, FIA and GendA (2009), many more mothers than fathers make use of the option to withdraw from active job search. Case study evidence also suggests that activation efforts of welfare agencies further differ between genders for efficiency reasons (IAQ, FIA and GendA, 2009). Due to limited time resources of the caseworkers and the overall goal to realize as many transitions to employment as possible, activation is mainly targeted to the most easy-to-place individuals. In most cases, welfare agencies assume that men are the easy-to-place individuals. If women are much less intensively activated than men or even not activated at all, we would not expect any significant difference in the success of decentralized and centralized welfare agencies to integrate female welfare recipients into employment without welfare receipt. Differences can only be present for individuals who are subject to activation like men. Thus, the findings of IAQ, FIA and GendA (2009) and other studies might explain why we observe gender differences in our results.

6. A Glance into the Black Box of Welfare Administration

The significant treatment effect for men raises the question of why centralized organization performs better in placing welfare recipients into jobs. Is the relative success of centralized agencies due to their use of more successful organizational approaches and strategies that could also be adopted by decentralized agencies as well? All centralized welfare agencies are subject to central FEA guidelines, central controlling, and certain directives regarding the use of activation strategies. Nevertheless, welfare agencies have leeway in the way they internally organize their services for welfare recipients. The implementation of organizational approaches is not specific to either administrative model, and we observe variations within both agencies with different organizational features. In the following, we analyze the effect of the adopted approaches and strategies and check if they are able to explain the positive effect of centralized organization.

To do so, we exploit data on the organizational strategies applied in the welfare agencies. According to studies conducted to evaluate the implementation of Germany's 2005 welfare reform (IAW and ZEW, 2008; WZB, infas and FHS Frankfurt/M, 2008), the following features are the most important elements in the internal organization of tasks and the cooperation with external partners:

  1. Generalized case management for all clients as opposed to case management by specialized staff for clients with multiple obstacles to employment,
  2. Integration of activation and placement as opposed to the separation of these functions,
  3. Use of customer segmentation procedures,
  4. Establishment of an employer service, that is, specialized staff maintaining contact with employers,
  5. Subcontracting of placement services to private providers.

Table 4 provides a more detailed description of the organizational features (measured in 2006) and outlines some arguments as to why they could affect the integration success of welfare recipients. Customer segmentation and, in particular, generalized case management tend to be used much more frequently by decentralized agencies, integration of activation and placement is slightly more common among centralized agencies, while the other two strategies are not related to agency type.

Table 4. Definition of organizational variables
DefinitionPossible impact on integrationFrequency in sample
  1. Note: The organizational variables were obtained from surveys conducted in 2006.

Generalized case management
Case managers counsel all types of clients. There is no assignment of welfare recipients with multiple obstacles to employment to specialist caseworkers.Better placement under specialized case management if clients with specific problems require specialized expertise. Generalized case management facilitates individual counseling as clients have fewer contact persons.0.69 (decentralized agencies)
0.24 (centralized agencies)
Integration of activation and placement
Clients are counseled (activated) and placed into employment by the same staff members. There is no assignment of specialized staff to the two tasks.Integration reduces the number of contact persons for each welfare recipient, and facilitates a holistic approach. In contrast, separation leads to gains from specialization but may create coordination problems at the interface of both tasks.0.51 (decentralized)
0.59 (centralized)
Customer Segmentation
Classification of clients into different groups receiving different treatment during activation.Segmentation may increase employment rates among groups that are activated more intensely but reduces integration into employment in other groups.0.84 (decentralized)
0.66 (centralized)
Employer service
A team of agency staff members maintains a network with employers and serves as contact persons for them.Networking may result in better placement. However, internal coordination problems between the employer service and caseworkers may arise.0.86 (decentralized)
0.83 (centralized)
Subcontracting of placement services
The welfare agency uses private employment services to place some of their clients into employment.Specialization gains may occur. However, private agencies may work more or less effectively compared to the public employment service. Requires proper assignment of welfare recipients to service providers.0.41 (decentralized)
0.40 (centralized)

To check whether the effect of decentralized agencies can be attributed to one of these strategies, we require a multivariate framework. For this purpose, we use binary probit models. The probit estimations contain all covariates used in the preferred specification of the propensity score (see above). In addition, dummy variables for decentralized welfare agencies and for each of the organizational features are included. Furthermore, we include the interaction of the organizational variables with the type of agency. We then test whether a significant effect of decentralized agencies on employment without welfare receipt remains despite controlling for organization.

Therefore, the estimated model is

display math

where yijt is the dependent variable for individual i in agency j at time t, Di is a dummy indicating whether agency j is decentralized, Bjk is the k-th organizational variable (k = 1, …, 5), Xi is a vector of individual characteristics (including a constant), Ri is a vector of regional characteristics and εijt denotes the error term, which is assumed to be normally distributed. The function 1() indicates that the dependent variable is binary.

Table 5 displays our estimation results for April, August and December, 2007. The standard errors account for potential clustering of error terms at agency level (e.g., Moulton, 1986, 1990). The entries in the table are marginal effects of the dummy variables on the outcome variable, and their magnitudes and treatment effects from matching are, therefore, comparable. Since results did not differ much between randomly chosen individuals and single or non-single households, we rely on the overall samples of men and women.

Table 5. Probit estimations for the effects of organizational features
  1. Notes: The table shows the results of five separate probit estimations. In the models, we include an interaction term between the respective organizational variable and decentralization. The table shows marginal effects and standard errors (in brackets). The dependent variable in each model and for each month is 1 if an individual is employed and does not receive welfare benefits and 0 otherwise. Number of observations for men (women): 6,217 (6,992). One centralized welfare agency had to be dropped from the analysis due to missing information. Standard errors take into account clustering at agency level. All models include the covariates used in the preferred propensity score specification of the matching analysis; detailed results are available from the authors on request. All results refer to the year 2007.

  2. ***p < 0.01; **p < 0.05; *p < 0.1.

Decentralized welfare agency−0.0188** (0.0089)−0.0259 (0.0162)−0.0286 (0.0180)−0.0122 (0.0081)−0.0098 (0.0091)−0.0037 (0.0160)
Generalized case management−0.0021 (0.0085)−0.0016 (0.0128)−0.0205 (0.0135)−0.0097 (0.0068)0.0014 (0.0090)0.0054 (0.0119)
Interaction0.0169 (0.0156)0.0016 (0.0234)0.0302 (0.0273)0.0122 (0.0131)0.0050 (0.0141)−0.0047 (0.0205)
Pseudo R20.0690.0670.0690.0800.0770.070
Decentralized welfare agency−0.0092 (0.0095)−0.0116 (0.0148)−0.0100 (0.0177)−0.0131* (0.0072)−0.0056 (0.0096)−0.0023 (0.0126)
Integration of activation and placement0.0010 (0.0072)0.0062 (0.0104)0.0024 (0.0124)0.0003 (0.0069)0.0100 (0.0086)0.0036 (0.0102)
Interaction−0.0010 (0.0135)−0.0333* (0.0184)−0.0214 (0.0230)0.0094 (0.0134)0.0017 (0.0149)−0.0043 (0.0190)
Pseudo R20.0690.0670.0690.0790.0780.070
Decentralized welfare agency−0.0281** (0.0132)−0.0408** (0.0197)−0.0253 (0.0191)−0.0055 (0.0120)0.0020 (0.0141)0.0028 (0.0172)
Customer segmentation−0.0041 (0.0075)−0.0126 (0.0104)0.0037 (0.0126)0.0012 (0.0071)−0.0034 (0.0088)0.0043 (0.0098)
Interaction0.0258 (0.0192)0.0224 (0.0259)0.0075 (0.0249)−0.0050 (0.0137)−0.0087 (0.0154)−0.0093 (0.0195)
Pseudo R20.0690.0670.0680.0790.0770.070
Decentralized welfare agency−0.0139 (0.0136)−0.0652*** (0.0235)−0.0778*** (0.0289)−0.0320*** (0.0107)−0.0536*** (0.0140)−0.0569** (0.0223)
Employer service−0.0099 (0.0096)−0.0094 (0.0142)−0.0374** (0.0186)−0.0216** (0.0110)−0.0313*** (0.0104)−0.0431*** (0.0135)
Interaction0.0059 (0.0172)0.0540 (0.0332)0.0831** (0.0416)0.0349* (0.0181)0.0730*** (0.0247)0.0779** (0.0353)
Pseudo R20.0690.0670.0700.0820.0800.074
Decentralized welfare agency−0.0150** (0.0074)−0.0271** (0.0110)−0.0252* (0.0137)−0.0038 (0.0065)−0.0023 (0.0078)0.0066 (0.0106)
Subcontracting of placement services−0.00560.0008−0.0065−0.00030.00140.0029
Interaction0.0156 (0.0160)0.0040 (0.0245)0.0181 (0.0293)−0.0146 (0.0098)−0.0096 (0.0138)−0.0273* (0.0154)
Pseudo R20.0690.0670.0680.0800.0770.071

Similar to the matching results, we tend to find a negative effect of decentralization for men. The main effect is significant for at least one period for four out of five specifications. The main effects of the organizational variables are, with the exception of those of the employer service, never significant. There is a (weakly significant) negative interaction effect for integration of activation and placement with decentralization, suggesting that decentralized agencies which integrated these functions perform worse than centralized agencies. Another significant interaction is with the employer service. The positive sign implies that decentralized agencies with an employer service performed better than those without. Importantly, in both cases it is not the typical organizational feature of decentralized agencies (i.e., not integrating activation and placement, as well as having an employer service) that explain their inferior performance.

Among women, the effects of decentralization, organizational strategies and interaction terms are insignificant for four out of five organizational variables. Only if controlling for employer service, all three factors are significant: decentralization and employer service have a negative impact on employment, but the interaction between employer service and decentralization is strongly positive. This indicates that, as in the case of men, the effect of an employer service offsets the negative effect of decentralization. A possible reason why an employer service is valuable for decentralized agencies is that this division strengthens their competences in the area of job placement where they have less previous experience than the centralized agencies.20

As a further robustness analysis, we included all organizational variables jointly into the specification (without interaction terms).21 Again, the effect of decentralization for men remained significantly negative at least over some part of the observation period, while the effect for women was insignificant. None of the organizational variables had a significant impact, with the exception of the employer service in case of women. All in all, we conclude that the effect of organization of welfare agencies is not due to the adoption of particular forms of internal organization. A more likely explanation of the difference in effects relates to the theoretical argumentation. The advantages of centralized organization in bundling resources, collecting information from various sources, and imposing best-practice strategies for the local offices tend to outperform the favorable properties of decentralized organization.22

7. Conclusions

The German welfare reform of 2005 introduced two competing organizational systems for the labor market activation of welfare recipients in an otherwise homogenous institutional setting: decentralized and centralized welfare agencies. To evaluate their relative performance, we have estimated their effect on the integration of welfare recipients into employment without welfare receipt, regarding regional differences as well as individual selection. The estimation is based on exceptionally rich data from various sources. We have combined a detailed survey of welfare recipients with administrative records from the Federal Employment Agency (FEA). In addition, we have used a large set of variables that describe the local labor market. Finally, we have considered unique information on the internal organization of the welfare agencies in our sample.

We find that decentralized welfare agencies have a negative effect on male welfare recipients with respect to integration into employment. Given the low transition intensity from welfare receipt into employment in general, the magnitudes of the effects for men are substantial. The integration quota of decentralized welfare agencies is up to 24% lower than the quota of centralized agencies. For women, we also find negative treatment effects, which are, however, smaller in magnitude than for men and which are not statistically significant. Gender differences are found within all subgroups considered (randomly chosen individuals, singles and non-singles). These might result from different activation intensity between men and women. Evidence suggests that, irrespective of agency type, the activation intensity of women is far lower compared to men. If welfare agencies concentrate their activation efforts predominantly on men rather than on women, it is harder to uncover significant differences in the relative performance of decentralized and centralized agencies for the latter subgroup.

We have further explored channels through which our results may have emerged. Because welfare agencies have significant discretionary power with respect to internal organization, we have checked whether the organization of tasks at individual welfare agencies is responsible for the result of decentralization. Although the effects are slightly weakened by the inclusion of the additional organizational strategies, the overall result is not affected. We conclude that the negative effect of decentralization is not due to differences in the adoption of strategies between centralized and decentralized welfare agencies and is not subject to their choices regarding the internal organization of tasks. The remaining differences are related to the very nature of (de)centralized organization. Examples are the application of central best practice guidelines of the FEA concerning the use of instruments of activation, as well as the (de)centralized controlling system.

Our findings point to the importance of the organizational aspects of welfare administration to the integration of welfare recipients into employment. Despite their importance, this topic has been largely neglected by existing literature on employment policy. Identifying successful and less successful strategies for the organization of welfare administration poses a number of challenges, and several open questions remain. A first aspect refers to potential changes over time: due to the limitations of the data, we only provide a snapshot relating to one particular year. Second, to study particular aspects of welfare administration, treatment variation at the individual rather than regional level could be exploited. Finally, the use of other econometric methods, such as experimental methods, may relax some of the assumptions required for propensity score matching, in particular the conditional independence assumption.

Appendix A

Figure A.1.

Map of the 154 welfare agencies in the sample

Appendix B

Table B.1. Probit estimations for the effects of organizational features
 Centralized agenciesDecentralized agenciesp-value (equality-of-means test)p-value (Kolmogorov–Smirnov test)


  1. Notes: All variables are measured for December 2003. The depicted numbers refer to the 154 sampled welfare agencies. The p-values in the third column derive from equality-of-means tests of the displayed variables for centralized and decentralized agencies. The p-values in the rightmost column derive from Kolmogorov–Smirnov tests of the equality of distributions. FF denotes the number of participants in activation programs designed on the discretion of the local employment offices (Freie Förderung). ABM stands for the number of participants in job creation schemes (Arbeitsbeschaffungsmaßnahmen). FbW denotes the number of persons participating in long-term training (Förderung der beruflichen Weiterbildung), TM the number of persons participating in short-term training (Trainingsmaßnahmen) and JUMP the number of participants in a program for the activation of young unemployed persons (Sofortprogramm der Bundesregierung zum Abbau der Jugendarbeitslosigkeit). ssc, social security contributions.

Unemployment rate (Source: FEA)11.30911.4120.9060.868
Unemployment rate of the young (age <25) (Source: FEA)10.62810.5050.8600.999
Unemployment rate of foreigners (Source: FEA)23.28524.3400.5670.959
Ratio of caseworkers to unemployed (classified)0.0160.0160.8370.574
Ratio of placement officers with fixed-term contract to unemployed0.0020.0020.8950.844
Ratio of young (<25) to old (>50) unemployed (in per cent)49.47850.9660.3390.538
Ratio of severely disabled unemployed to all unemployed0.0400.0390.8090.979
Ratio of long-term unemployed to all unemployed0.3320.3330.8930.872
Rate of social assistance recipients0.0360.0280.0040.013
Unemployment-Vacancy (UV) relation in textile industry73.59284.2130.3010.712
UV relation in construction sector37.12435.6400.7020.960
UV relation in engineering16.26717.8570.5670.395
UV relation in commerce sector24.82027.3320.4620.626
UV relation in service sector20.75324.2320.2120.720
UV relation in metal industry15.26114.6100.6610.998
UV relation in healthcare6.3466.3560.9830.572
UV relation in social sector11.43311.1210.7280.600
UV relation overall30.20832.3860.4710.770
FF per unemployed0.0070.0090.4080.939
FF per male unemployed0.0080.0100.4790.947
FF per female unemployed0.0060.0080.3370.361
FF per unemployed over age 500.0040.0050.4050.855
FF per unemployed under age 250.0140.0190.2530.511
Employer wage subsidies per unemployed0.0320.0330.7530.076
Employer wage subsidies per unemployed over age 500.0620.0650.7630.591
Employer wage subsidies for long-term unemployed per unemployed0.0020.0030.1680.039
Employer wage subsidies for long-term unemployed per male unemployed0.0020.0030.1490.172
Employer wage subsidies for long-term unemployed per female unemployed0.0020.0030.1310.021
Start-up grants per unemployed over age 500.0080.0090.6380.509
Start-up grants per unemployed under age 250.0080.0070.7350.896
ABM/unemployed + ABM0.0170.0190.4300.890
ABM/unemployed + ABM (women)0.0160.0180.4880.412
ABM/unemployed + ABM (men)0.0170.0200.3890.812
FbW/(unemployed + FbW)0.0580.0600.2050.593
FbW/(unemployed + FbW) (men)0.0490.0520.3100.268
FbW/(unemployed + FbW) (women)0.0690.0710.2640.386
FbW/(unemployed + FbW) (age >50)0.0140.0150.3600.093
FbW/(unemployed + FbW) (age <25)0.0540.0550.7410.945
TM/(unemployed + TM)0.0220.0220.6370.419
TM/(unemployed + TM) (women)0.0230.0230.7630.610
TM/(unemployed + TM) (men)0.0220.0210.5390.341
TM/(unemployed + TM) (age >50)0.0100.0100.8830.223
TM/(unemployed + TM) (age <25)0.0360.0350.8280.813
JUMP per unemployed (age <25)0.1210.1360.2090.565
Ratio of working population to resident population0.4650.4240.0750.098
Ratio of persons employed (subject to ssc) to resident population0.3200.3220.5350.855
Ratio of persons employed (subject to ssc) to resident population (men)0.3570.3610.4500.490
Ratio of persons employed (subject to ssc) to resident population (women)0.2840.2850.8230.884
Commuter balance per 1,000 employees−64.233−172.4310.0340.051
Business foundations per 10,000 inhabitants aged 15–64149.643146.7000.5170.228
GDP per economically active person (in 1,000 Euro)51.65751.3430.8260.602
Ratio of foreigners to resident population0.0840.0650.0320.158
Available infant care places per infant0.6370.6550.3390.518
Available child care places per child0.2810.2850.7770.802


Boockmann is also affiliated with the University of Tübingen and Institute for the Study of Labor (IZA). Thomsen is affilated with University of Hannover and ZEW. Göbel is also affiliated with ZEW. This study is based on results from the evaluation mandated by the ‘experimentation clause’ of Chapter 6 of Book II of the German Social Code (Sozialgesetzbuch Zweites Buch, SGB II) and was commissioned by the German Parliament. All opinions expressed are our own. We thank the Institute for Employment Research (IAB), Nürnberg, for providing the administrative data for this study. We gratefully acknowledge helpful comments by the editor, Christoph M. Schmidt, two anonymous referees, Martin Brussig, Bernd Fitzenberger, Gerd Heyer, Matthias Knuth, Michael Lechner and Conny Wunsch. We thank Markus Clauss and Martina Hartig for helping us to obtain the database. Moritz Hennig, Stefan Langer, Verena Niepel and Hans Verbeek provided excellent research assistance. Stephan L. Thomsen thanks the Stifterverband für die deutsche Wissenschaft for financial support.


  1. 1

    For a review of US welfare reforms and the related empirical literature, we refer to Blank (2002), Grogger and Karoly (2005) and Moffitt (2002). Bloom and Michalopoulos (2001) synthesize the results of 29 studies investigating the effects of various US welfare-to-work programs. German welfare-to-work programs that were introduced after 2005 have been analyzed by Aldashev et al. (2010), Bernhard et al. (2008), Boockmann et al. (2009), Hohmeyer and Wolff (2007), Huber et al. (2010), Thomsen and Walter (2010a) and Wolff and Jozwiak (2007). Surveys on welfare reforms in Europe are provided by Halvorsen and Jensen (2004), Kildal (2001) and Torfing (1999) for the Nordic countries, Beaudry (2002), Dostal (2008) and Finn (2000) for the UK, and Knijn and van Wel (2001) and Finn (2000) for the Netherlands. See also Kluve (2010) and Martin and Grubb (2001) for comprehensive overviews.

  2. 2

    This setting was introduced in the so-called experimentation clause in Chapter 6 of Book II of the German Social Code (Sozialgesetzbuch Zweites Buch, SGB II). A description of the experimentation clause with details of implementation, context and policy results is provided by Deutscher Bundestag (2008).

  3. 3

    The number was later extended, as described below.

  4. 4

    This definition does not preclude that the employer receives public employment subsidies for hiring. Since German hiring subsidies were found to give rise to huge deadweight effects (Boockmann et al., 2012), we feel justified to neglect the distinction between subsidized and non-subsidized hiring.

  5. 5

    A variant of this model arose where the local employment office and local authorities could not agree on forming a joint venture. In 19 of 439 cases, both institutions continued to work separately in the district. However, because tasks are shared in a similar way as in the case of the centralized system, we do not differentiate between these two types in the empirical analysis.

  6. 6

    As will be shown in section 'A Glance into the Black Box of Welfare Administration', the organizational independence of decentralized agencies leads to considerable variance in the implemented integration strategies.

  7. 7

    The German Constitutional Court had ruled in December 2007 that the form of cooperation between the public employment service and local authorities that had been chosen by most centralized agencies violated the constitution (2 BvR 2,433/04; 2 BvR 2,434/04).

  8. 8

    Parts of this data set are publicly available as a scientific use file at the Federal Employment Agency. See Oertel et al. (2009) for details on data access.

  9. 9

    The study by Arntz et al. (2006) was conducted to prepare the evaluation of the welfare reform.

  10. 10

    The composition of welfare recipients in the districts could, to some extent, itself be an outcome of decentralized or centralized organization at this point of time. If, for example, the centralized system were faster in integrating welfare recipients with good employment prospects in the early periods after the reform, the stock of welfare recipients in 2006 may contain fewer welfare recipients with favorable characteristics than in decentralized districts. Potential compositional differences of welfare recipients are considered in the estimation, by taking selection at the individual level into account.

  11. 11

    This particularly applied to decentralized welfare agencies, which continued to use their local computer systems. In principle, an interface for data collection was provided by the FEA, allowing these welfare agencies to interact directly with the FEA's mainframe computers. In practice, however, the adoption of the interface was incomplete until the second half of 2006. Centralized agencies, on the other hand, had issues with a newly introduced software system. For these reasons, the quality of the data during the early periods after the reform is insufficient for empirical analysis. Therefore, we rely on data from 2006 and 2007.

  12. 12

    Due to the proprietary nature of the data, the time horizon of December 2007 could not be extended.

  13. 13

    The administrative data only contain information regarding employment that is subject to social insurance contributions. Therefore, our outcome variable does not include spells of minor employment or self-employment. The outcome variable is measured as a binary dummy variable.

  14. 14

    Other forms of potential selectivity are unlikely in our case. In particular, welfare recipients cannot self-select into treatment. From the point of view of a welfare recipient or the caseworker, the 2005 reform of welfare administration and organization is an exogenous event that cannot be easily influenced or avoided. The only way to select into treatment would be to move to another district. However, welfare recipients usually cannot afford to relocate and are not encouraged to move as long as they remain on welfare.

  15. 15

    The duration of the welfare spell is measured as the number of months on welfare benefits before the sampling date. Due to the time span between sampling and interview date, not all individuals report a starting date of welfare receipt before the sampling date. Some left and re-entered the welfare system during fall and winter 2006/2007 and thus report a starting date after the sampling date. For these individuals, the duration variable is set to zero. An additional dummy variable takes these late starting dates into account.

  16. 16

    To further whether regional heterogeneity might bias our results, we limited the sample to 35 decentralized agencies directly bordering a centralized agency. For each of these 35 areas, we performed the same matching analysis as for the whole sample and chose the average of the regional treatment effects as our estimator. Furthermore, we used the regional matching approach described in Arntz et al. (2006) to form groups of comparable welfare agencies and repeated our matching analysis. In both cases, the averages of the treatment effects estimated are very similar to the effects presented below. Results for this robustness analysis are available on request from the authors.

  17. 17

    The online appendix is available at

  18. 18

    We use the matching algorithm provided by Leuven and Sianesi (2003).

  19. 19

    The corresponding results for the samples of singles and non-singles are given in the Online Appendix.

  20. 20

    More discussion on the role of the employer service in centralized and decentralized agencies can be found in WZB, infas and FHS Frankfurt/M (2008), p. 214ff.

  21. 21

    Results are available on request from the authors.

  22. 22

    The use of ALMP measures (public employment schemes, short-term training and qualification measures) by centralized and decentralized agencies has been investigated (ZEW, IAQ and TNS Emnid, 2008). The probability of participation in these measures did not differ much with centralization, although centralized agencies used slightly more public employment schemes and decentralized agencies tended to give programs more frequently to women, lone mothers and young individuals than centralized agencies. The estimated treatment effects did not differ substantially between centralized and decentralized agencies. Therefore, we have little indication that the different intensity or effectiveness of ALMP programs is behind the effect of centralization.