Is Local Better? Job Finding and the Decentralization of Public Employment Services

Transcription

1 Is Local Better? Job Finding and the Decentralization of Public Employment Services Lukas Mergele Michael Weber This version: August 22, 2016 Abstract This paper examines whether decentralization of public employment services (PES) improves their efficacy. We exploit a rare within-country variation in decentralization given by the partial devolution of German PES offices to the communal level in Using a difference-in-differences strategy, we find that decentralization decreased the job finding of unemployed by up to 20% in the first year and by about 10% throughout the following three years. While we attribute the short-term loss to transition and learning effects, the medium-term loss is associated with the increased use of inefficient public employment programs. These findings are consistent with local authorities not internalizing the full costs of increasing unemployment which are mainly born by the federal government. JEL-Classification: H11, H75, I38, J48 Keywords: Decentralization, public employment services, job-finding We thank participants of the ESPE 2014 Annual Conference, the SES 2014 Annual Conference, the 7th IAB GradAB Workshop, the 2015 Potsdam Workshop in Empirical Economics, the 2015 Warwick Economics PhD Conference as well as seminar audiences in Berlin, Dresden, Leipzig and Nuremberg for helpful comments and suggestions. In particular, we are grateful to Bernd Fitzenberger, Erik Grönqvist, Rafael Lalive, Ben Lockwood, Rajshri Jayaraman, Martin Lundin, Eva Mörk, Steve Pischke, Björn Öckert, Alexandra Spitz-Oener, Marcel Thum, Fabian Waldinger and Felix Weinhardt for their advice. We also thank the German Ministry of Labor and Social Affairs as well as our contact persons at eleven state ministries for providing invaluable guidance on the institutional setting. Mergele gratefully acknowledges financial support through the DFG grant RTG 1659 as well as excellent research assistance by Alex Graf and Alex Rebmann. Corresponding author. Humboldt University, Spandauer Strasse 1, Berlin. lukas.mergele@hu-berlin.de Ifo Institute, Dresden Branch, Einsteinstrasse 3, Dresden. Phone: +49(0)351/ weber.m@ifo.de 1

2 1. Introduction Many countries maintain public employment services (PES) to reduce coordination frictions in the labor market and thereby unemployment. These services advice job seekers on job offers, monitor their search efforts, assign active labor market programs (ALMPs) and administer welfare payments. Making PES more efficient has been found to be more effective at reducing unemployment than cutting unemployment benefits (Launov and Wälde, 2016). To increase the efficiency of public welfare institutions, international organizations including the World Bank and OECD (LEED program) have promoted their decentralization, that is to give local PES units more operative flexibility. During the 1990s and 2000s, Canada, Denmark, Germany, Italy and Sweden, among others, initiated reforms in this spirit. However, there is surprisingly little evidence on the success of these policies. 1 In Germany, an experimental devolution of PES offices in 2005 decreased the employment probabilities for men but not for women (Boockmann et al., 2015). The Danish experience indicates that this might reflect the localized PES offices emphasis on social integration rather than on re-employment of hard-to-place job seekers (Bredgaard and Larsen, 2009). In Sweden, localized PES offices tended to target training programs more towards hard-to-place job seekers (Lundin and Skedinger, 2006), which would temporarily reduce their job-finding rates. Yet, beyond these few studies, the impact of decentralizing PES offices on job finding is unknown. In particular, the mechanisms driving these effects have rarely been explored. This paper exploits a rare policy experiment in Germany to examine the effect of PES decentralization on job finding as well as the underlying mechanisms. In 2012, a partial devolution policy induced within-country variation in the degree of centralization of PES offices on a large scale. PES offices ( job centers ) in 41 out of 402 German districts 2 were decentralized so that local authorities could manage the placement strategy, the monitoring and sanctions policy, and the allocation of ALMP measures on their own. Before, these strategies were determined under the leadership of the Federal Employment Agency (FEA) although in cooperation with local authorities. In particular, the job placement strategy was strongly influenced by the FEA via guidelines and directives. This centralized steering of PES offices prevailed in 294 districts even after Our analysis employs an administrative dataset comprising the monthly stocks and gross flows of unemployed, welfare recipients and vacancies from 2009 to 2015 in all German districts. The data further includes the number of benefit sanctions imposed on job seekers and ALMP assignments per district. We use a difference-in-differences (DiD) framework to identify the causal effect of decentralizing PES offices. We proceed in two steps. First, we determine the overall 1 There is evidence on the effect of contracting out of placement services to private providers, see, for instance, Heinze et al. (2006) and Bennmarker et al. (2013). 2 German: Kreise und kreisfreie Städte. The German districts are an administrative subdivision similar to counties in the US. 3 PES offices in 67 districts had been decentralized in 2005 already, see Boockmann et al. (2015). 1

3 effect of decentralization on the PES efficacy in job finding of unemployed. Second, we open the black box of PES mechanisms affecting the job finding. In the first step, we estimate an aggregate stock-flow matching function as proposed by Coles and Smith (1998). We link a district s monthly number of unemployed finding a job to the district s total stocks and inflows of unemployed job seekers and vacancies. The German job centers likely have a significant impact on the number of job matches as they constitute the major job search channels for their clients, because these are either long-term unemployed, low productive, or both (Fougère et al., 2009; Gregg and Wadsworth, 1996; Pissarides, 1979). In addition, job centers may directly impact the number of matches by placing job seekers into specific vacancies. We find that decentralization decreased the conditional number of matches by 20% in the first year and by about 10% during the second to fourth post-reform year. The latter result is equivalent to an increased average unemployment duration of three months. While the short-term loss is likely related to administrative changes during a transition period, the medium-term loss points to severe negative effects of decentralization on a PES office s placement efficiency. In the second step, we analyze mechanisms that may explain the observed lower job finding after decentralization. Theoretical and empirical evidence suggests at least five potential channels. First, decentralized job centers might offer a different quantity, or secondly, a different mix of ALMP measures (Lundin and Skedinger, 2006) which would affect their clients re-employment prospects depending on the effectiveness of these programs (see Card et al., 2010, for a meta analysis). Third, decentralized job centers may have eased the monitoring and sanction policy, which would reduce the job seekers search efforts and thereby the aggregate job-finding rate (Arni et al., 2013; Lalive et al., 2005; van den Berg and van der Klaauw, 2015). Fourth, decentralized job centers might place more emphasis on stable rather than on quick placements of potentially lower quality, which leads to longer search durations and lower job-finding rates as well (Caliendo et al., 2013). Fifth, decentralized job centers might have increased their caseworkers caseloads or employ caseworkers with different backgrounds and skills, which could also negatively affect the number of job matches (Behncke et al., 2010; Hainmueller et al., 2015). Our results indicate that decentralized job centers shifted ALMP assignments towards more inefficient ALMP measures, namely public sector employment programs, while leaving the overall number of participants in training measures unchanged. Furthermore, sanction regimes became less strict during the transition period following the reform. We do not find evidence for temporary or permanent changes in the placement policy and we can rule out changes in the caseworker quality and quantity due to the reform s specific regulations. Multiple selection analyses, placebo tests, and the assessment of geographical spillovers confirm that our results are driven by decentralization rather than selection problems or other confounding effects. We conclude that local labor markets did not benefit from decentralization, and that the lower conditional number of matches in districts with decentralized job centers largely reflects the increased use of inefficient ALMP programs. By increasing the average duration in unemployment, the 2012 decentralization caused Germany fiscal costs of more than 1 bn. EUR. These findings 2

4 are consistent with the idea that decentralization is unfavorable if local agents fail to coordinate spillovers. In the German case, the fiscal costs of unemployment are mainly borne by the federal government whereas PES policies in reformed districts are defined by local authorities. Our research connects to two strands of the literature. First, we expand the literature dealing with the optimal design of labor market institutions (Boeri, 2011; Holmlund, 2014). In particular, we extend previous work by Boockmann et al. (2015) and Lundin and Skedinger (2006). Boockmann et al. (2015) examine the 2005 experimental decentralization of PES offices in 69 out of then 432 German districts and find a negative effect of decentralization on the job-finding rate of men (see also Holzner and Munz, 2013). Yet, the authors have to deal with an empirically difficult reform setting which forbade pre-post comparisons, coinciding with a major unemployment benefit reform. Our empirical setting is more favorable in these regards. Lundin and Skedinger (2006) study a Swedish pilot reform that granted municipal authorities a voting majority in the local employment committees of 25 municipalities. Lundin and Skedinger (2006) find that due to this reform, more ALMP projects were organized by the municipalities and that hard-to-place job seekers were more likely to be enrolled in municipal projects. However, this pilot lasted for three months only, limiting the scope for potential effects to unfold. In contrast, the German decentralization allows us to trace the evolution of the decentralization effect over a period of four years. Second, we contribute to the broader debate on whether governments should follow a centralized or decentralized provision of public goods and services (see Oates, 2005, for a review). So far, economists have investigated fiscal federalism mostly with respect to public finances (Baicker et al., 2012), education policies (Barankay and Lockwood, 2007), environmental policies (Banzhaf and Chupp, 2012) or political institutions (Enikolopov and Zhuravskaya, 2007). Little attention has been paid to labor market institutions, although they are of particular interest for decentralization due to their context-specificity. We provide evidence that decentralizing labor market institutions may cause enormous fiscal net costs, emphasizing the need to decentralize with caution and taking potential spillovers into account. Moreover, this literature has discussed whether decentralization actually alters the policies in use (Strumpf and Oberholzer Gee, 2002). Our results concerning ALMPs clearly support this hypothesis albeit these policies do not work in favor of the job seekers directly concerned. The use of inefficient ALMP programs that are useful for the local budget but detrimental to the job seeker s reemployment prospects may reflect clientelism or local capture (Faguet, 2004). The paper proceeds as follows. Section 2 provides details on the German PES offices and their reform in Section 3 describes the data and our empirical strategy. Section 4 presents the estimation results on the effects of decentralization and explores underlying mechanisms. Section 5 concludes. 3

5 2. Policy Background 2.1. The German Job Centers Job centers play a central role in the German welfare system. They were introduced as one-stop PES offices in 2005.Job centers are set up in each German district, serving welfare recipients living in the respective district. 4 These PES offices engage in job matching and job counseling, and assign clients into training or other ALMP measures. They also monitor their clients job search efforts and may temporarily reduce their unemployment benefits if the unemployed do not comply to their jobseeker duties. 5 The job centers have some discretionary power over the placement strategy, the monitoring and sanction regime and the assignment of clients into ALMP measures. The job centers clients are welfare recipients able to work whose household income falls below a subsistence threshold, irrespective of their employment status. They are roughly summarized as long-term unemployed job seekers and employed workers with very low labor incomes. As of January 2012, 2 million people across Germany were job center clients. These clients have comparably poor labor market prospects and little access to alternative job search channels. For them, welfare-to-work transitions likely result from placements by the PES (see, for instance, Fougère et al., 2009; Pissarides, 1979) Centralized vs. Decentralized Job Centers A political compromise led to the creation of two job center types, varying in their degree of centralization (see Table 1). 6 Centralized job centers are effectively governed by the Federal Employment Agency (FEA) by means of target agreements, directives and technical supervision, such that PES provision is effectively standardized. In particular, the placement, caseload, and sanction policy follows nationwide guidelines with little strategic leeway remaining for the local PES offices. Centralized offices cooperate with the district administration (gemeinsame Einrichtungen) but local authorities responsibilities are mainly non-pes related tasks such as the administration of additional welfare payments and specific counseling, for instance in case of drug addiction. The second type of job centers are decentralized and run by the district administrations alone (zugelassene kommunale Träger). They constitute a regular part of the local administration. They are independent of the FEA but receive the same financial support from the federal government as their centralized counterparts. Districts only have to sign target agreements with their respective state governments, their sole de-jure supervisors. This setup results in a decentralized PES provision. A 4 In very few cases, job centers serve more than one district. 5 Unemployed job seekers have to search actively for a new job, frequently meet their caseworkers, participate in assigned ALMP measures, and accept appropriate job offers. 6 See e.g. Feldkirchen and Sauga, Das Reform-Monster. Der Spiegel, 1/2005, p

6 Table 1: Comparing decentralized and centralized German job centers Commonalities Legal basis Tasks Financing Centralized Decentralized Job center Job center Rights and oblications according to the same law (SGB II) Job counseling, assignments into ALMPs, monitoring and sanctioning, benefit admistration Equal financial support from the federal government Differences De Jure Affiliation FEA Part of district administration Supervision FEA District authority, State FEA Process Standards Binding standards for counselling, ALMP provision and monitoring Not binding, independent solutions Provider of placement services FEA District Software FEA Standard System District-specific solutions Controlling system FEA Standards District-specific solutions NOTES. FEA: Federal employment agency. ALMP: Active labor market programs. Sources: Boockmann et al. (2015); Bundesagentur für Arbeit (2013); Deutscher Bundestag (2012) survey conducted in 2007 reveals that decentralized and centralized PES offices adopt different structures and strategies indeed (Boockmann et al., 2015; Deutscher Bundestag, 2008). For instance, centralized PES offices were more likely to counsel job seekers depending on their employment prospects while decentralized providers rather counseled each job seeker with equal intensity. The differences in the internal and external organization of the local PES offices may affect the job finding of their clients for two reasons. On the one hand, decentralized job centers have more leeway to adjust their placement, sanction, and ALMP strategies to the specific context of their local labor market which should improve the aggregate job-finding rate. On the other hand, they are part of the district administration and, therefore, their set of strategies might be subject to the local budget or local election cycles which could hamper the job finding of the unemployed The 2012 Decentralization Decentralized PES offices were established in two waves. On 1 January 2005, decentralized offices were created in 67 districts. On 1 January 2012, 41 other districts followed suit. The 2005 experimental decentralization coincided with 5

7 another major labor market reform which included the effective reduction of longterm unemployment benefits (Dustmann et al., 2014). Due to the comprehensive changes in the German unemployment system, unemployment statistics produced during the first one and half years after the decentralization are either missing or lacking reliability. An official evaluation of this reform led to inconclusive results (Deutscher Bundestag, 2008), such that no political consensus could be reached about the preferred regime. As a compromise, job centers in up to one forth of all German districts could be decentralized. Hence, 41 additional districts decentralized their job centers on 1 January This time, no other relevant welfare or labor market reforms took place. Although the decentralizing offices had to migrate to new software systems, statistics problems were largely absent for this wave (Deutscher Bundestag, 2012). For these reasons, we focus our analysis entirely on the second wave of reform. The districts whose PES offices were decentralized in 2012 were determined within a two-stage process. First, districts willing to decentralize had to apply to their respective state governments. For the 2012 reform, the application period started on 3 August 2010 and ended on 31 December 2010 (Deutscher Bundestag, 2010). Local councils were required to back the application with a voting majority of two thirds. Hence, districts that could gain the most should have been more likely to support the regime change. Second, the state governments nominated from all applications those districts that were allowed to decentralize. The number of nominations were subject to a quota specific to each state, roughly proportional to the state s size. 7 The list of districts that were allowed to decentralized their job centers was officially announced on 14 April Despite this process, the groups of decentralizing and non-decentralizing districts are observationally equivalent (see Section 3). Job centers were decentralized in districts all over Germany (see Figure 1). In particular, treated districts do not cluster in regions with particularly poor or strong labor market conditions. Decentralized districts also include major German cities such as Stuttgart or Essen. Decentralization came into force on 1 January Data and Empirical Strategy 3.1. Data We utilize a rich administrative dataset with monthly observations measured at the district level to examine the effect of decentralization on job finding and other labor market outcomes. The data are provided by the FEA, covering the full universe of districts and PES offices. Variables covered are aggregate unemployment transitions, vacancies, welfare recipients, benefit sanctions and ALMP participants. 8 Moreover, virtually all unemployed job seekers are considered in the official statistics because unemployment registration is mandatory for taking up unemployment benefits. 7 The total quota for Germany as a whole was 41 districts. If the number of applying districts fell short of the cap in one state, remaining places were filled by districts from other states. 8 We readily observe gross worker flows and thus have not to deal with time aggregation issues. 6

8 Geodata: GeoBasis-DE / BKG Figure 1: Spatial distribution of districts by their type of job center In our analysis, we omit 11 districts in which centralized and decentralized PES offices co-exist due to some administrative reforms. Moreover, districts that decentralized PES offices in 2005 are not part of our sample, although their inclusion as an additional control group does not alter our results. Our sample period ranges from January 2009 to December 2015, that is from three years before to four years after the decentralization. Due to incomplete data, we drop the first three post-reform months from our sample. We retrieve additional information on a district s application and decentralization status from the Federal Ministry of Labor and Social Affairs. The data allows us to study the effect of decentralization on various labor market outcomes. As the main task of PES is to increase the job finding of the unemployed, we consider the monthly outflow from unemployment into employment as our main outcome indicator. We also elaborate on potential channels that might explain reform-induces changes in the job-finding rate. To assess whether decentralized job centers adopt different ALMP, sanction or placement strategies, we also analyze alternative outcomes such as the flows of unemployed into different types of training measures, the number of sanctions issued in a month on job seekers who did not 7

9 comply to their job-seeker duties, and the outflow out of welfare. 9 To explore the quality of the placements, we additionally consider transitions out of welfare without subsequent returns into welfare within three months ( permanent outflows out of welfare ) Econometric Model To identify the effect of decentralization, we employ a difference-in-differences framework (DiD). Our treatment group comprises 40 districts whose PES offices were decentralized in 2012, while our control group contains 294 districts whose offices remained centralized throughout the sample period. The functional form of the econometric model is equivalent to a stock-flow matching model with Cobb-Douglas technology (Coles and Smith, 1998; Ebrahimy and Shimer, 2010). Analogous to a production function, the stock-flow matching function models the gross flow from unemployment into jobs ( matches ) M as an output produced by the stocks of vacancies and unemployed (V,U) and their respective inflows (Ṽ, Ũ). We interpret the total factor productivity of the matching function as an indicator for the efficiency of the local PES office in bringing unemployed back to work. The decentralization status of a PES office then is one component of this indicator. The stock-flow matching function has received empirical support both at the micro and the macro level (Andrews et al., 2013; Gregg and Petrongolo, 2005) with strong evidence for a Cobb-Douglas functional form (see Petrongolo and Pissarides, 2001, for a survey). Our estimation equation therefore reads ln M it = α i + τ t + β 1 ln U it + β 2 ln V it + β 3 ln Ũit + β 4 ln Ṽit +δ D it + γ 1 Z it + ε it. (1) where i indexes districts (which is equivalent to PES offices) and t indexes months. We include a district-specific effect α i to account for unobserved time-constant local labor market conditions and a time fixed effect τ t to account for general business cycle and seasonal effects. The dummy variable D it indicates whether a PES office is decentralized or not. Our parameter of interest is δ which provides the treatment effect of decentralization on the conditional outflow from unemployment to employment. For robustness checks, we include further variables Z it that account for the demographic composition of the unemployed. Standard errors are two-wayclustered to account for unobserved correlation within PES offices and time periods (Bertrand et al., 2004). In order to trace the evolution of the treatment effect over time, we substitute the uniform coefficient δ D it by a series of year-specific effects, i. e. by y δ y D iy. As treatment effects might occur since the announcement of the decentralizing districts 9 Unfortunately, we do not observe to what extent outflows out of welfare reflect welfare-to-work transitions. For instance, the observed number of total outflows may also include transitions into old-age pensions. Yet, balance checks do not indicate that the composition of the outflows varies between the treatment and the control group. 8

10 in April 2011, all treatment effects are estimated relative to the year Estimating quarterly-specific effects leads to qualitatively similar but less precise results Identifying Assumptions Our empirical approach relies on two main assumptions. First, fundamental labor market trends should be parallel in both types of districts in the absence of the policy change. We discuss this major assumption in the following paragraphs. Second, the decentralization had no effect on the job-finding rate in districts with still centralized job centers (stable unit value treatment assumption, SUTVA). Section 4.6 presents evidence in favor of this assumption. Moreover, there were no other simultaneous reforms that affected the two groups of districts systematically different, as argued in section 2. Table 2: Major district characteristics by type of job center in 2010 group means decentralized remained Variable in 2012 centralized difference Total population at working age (15 64 years, in 1000) (96.6) (179.3) (28.9) Employment rate (.170) (.216) (.036) Job-center unemployment rate (.025) (.028) (.005) Welfare dependency rate (.041) (.046) (.008) Districts Notes: * p < 0.1; ** p < 0.05; *** p < Standard deviations of group means and standard errors of differences in parantheses. Comprehensive balance checks indicate that districts with decentralized job centers plausibly constitute a random sample from the population of German districts with regard to observational characteristics (for details, see Table 11 in the Appendix). First, Table 2 shows that both district groups exhibited on average the same population size, employment, unemployment, and welfare dependency in the application year Besides, both groups of districts do not differ with regard to the sectoral distribution of employees, including self-employed. There are also no statistically significant differences in the socio-demographic structures of the population at working age, of the welfare recipients, or of the unemployed with the share of long-term unemployed 10 being the single exception. 10 Job seekers are long-term unemployed if they have been unemployed for at least one year without interruption. Interruptions are, for instance, employment, participiation in certain ALMP measures, or illness of at least six weeks. 9

11 Table 3: Changes in major district characteristics by type of job center in group means decentralized remained Variable in 2012 centralized difference Total population at working age (15 64 years, in 1000) (.063) (.068) (.011) Employment (.069) (.069) (.012) Job center unemployment a) (.126) (.113) (.019) Number of welfare-recipients a) (.098) (.080) (.014) Districts Notes: * p < 0.1; ** p < 0.05; *** p < Standard deviations of group means and standard errors of differences in parentheses. a) Change from 2007 to Second, districts of both groups also experienced remarkably similar fundamental labor market trends from 2000 to 2010, on average: This includes a shrinking working population, growing employment with sectoral employment shifts towards services, and a reduction in overall unemployment, especially in long-term unemployment (see Table 3). Districts that successfully applied for decentralization had performed poorer only with regard to some very narrow demographic groups, such as a smaller reduction of the number of foreigners on welfare. Third, the graphical inspection of our outcome of interest affirms that the jobfinding rates 11 in both district types have evolved very similarly in the three years before the reform (see Figure 2). During the first four years after the reform, the difference between the average job-finding rates increased again and did not reach its pre-reform level. This descriptive result points to a potentially negative effect of decentralization on the job-finding rate. Yet, the analysis could suffer from potential self-selection based on unobserved characteristic. To counter these concerns, section 4.6 presents additional placebo tests and triple-differences specifications. 10

12 Figure 2: Seasonally adjusted average monthly job-finding rates by PES office type 4. The Effects of Decentralization 4.1. The Effect on Job Finding Table 4 reports the DiD estimation results for our main outcome indicator, the outflow from unemployment into employment. Each column represents a regression of logged transitions into jobs on a decentralization indicator, district and month fixed effects, and subsequently introduced covariates. The first column gives the average treatment effect of decentralizing the PES office on unemployment outflows while controlling only for fixed effects. The point estimate implies a reduction of the monthly flows into jobs by 11%, on average. This estimate is quite robust when we additionally control for the monthly stocks of vacancies and unemployed in column 2, their inflows in column 3, or both stocks and inflows in column 4. The job-finding of the unemployed is more elastic with respect to the inflow of new vacancies than with respect to the stock of vacancies, which is in line with the concept of stock-flow matching. In column 5, we additionally control for shares of demographic groups that are typically hard to place into jobs, that is the share of youth unemployed below the age of 25, the share of elderly unemployed above the age of 50, the share of foreign unemployed and the share of long-term unemployed. As expected, higher shares of these hard-to-place job seekers in the group of unemployed ceteris paribus reduce the unemployment outflow into employment. Yet, controlling for these shares does not affect our estimate of the decentralization effect. Hence, we conclude that 11 Definition: Outflow out of unemployment during a month divided by the unemployment stock at begin of that month. We use job-finding rates as a standardized measure of job-finding. 11

13 Table 4: Difference-in-differences: Flows into jobs, Fixed Effects Stocks Flows Stock-Flow Controls Decentralized *** *** *** *** *** (0.025) (0.024) (0.022) (0.022) (0.022) Vacancies, stock 0.040** (0.015) (0.013) (0.013) Unemployed, stock 0.805*** 0.692*** 0.700*** (0.051) (0.046) (0.047) Vacancies, inflow 0.113*** 0.097*** 0.098*** (0.017) (0.012) (0.012) Unemployed, inflow 0.378*** 0.280*** 0.271*** (0.031) (0.025) (0.025) Unemployed <25 ys, share * (0.002) Unemployed >50 ys, share *** (0.002) Foreign unemployed, share *** (0.002) Long-term unempl., share (0.001) R-squared Districts Observations 26,998 26,998 26,998 26,998 26,998 Notes: * p < 0.1; ** p < 0.05; *** p < Each column presents a different estimation of equation 1. Decentralized is a dummy equaling 1 for districts whose PES office is decentralized and 0 otherwise. All continuous variables in logs. Standard errors given in parenthesis are two-way clustered on the PES office and the month level. Regressions include a full set of dummies for districts and months. decentralization has reduced the monthly flow into jobs on average by about 10% within four years after the reform. Given that the average monthly job-finding rate of centralized districts amounts to about 3.8%, our estimates imply an increase of the expected duration in (long-term) unemployment from about 26 months to about 29 months, that is by about 3 months. Next, we investigate whether the negative effect of decentralization is accelerating, stabilizing or regressing over time. For that reason, we add a full set of annual leads and lags of the reform to our model. Our reference is the year 2010, when districts had to apply for decentralization. 12 To avoid potential endogeneity concerns regarding the targeting of certain demographic groups, we continue with the stock-flow matching model from column (4) in Table 4 which does not include unemployment composition covariates. Figure 3 depicts the resulting evolution of the treatment effect of decentralization on monthly unemployment outflows from three years before 12 Estimates relative to the year 2011 might be flawed by announcement effects which potentially emerged after the announcement of successful applicants in April

14 to four years after the reform. During the pre-reform period, treatment effects are statistically insignificant which is in favor of the common trends assumption underlying our identification strategy. In the first year after decentralization, monthly unemployment outflows were strongly reduced by about 20%. During the following three years, this negative effect weakens over time, but still amounts to almost 10% in the fourth year after decentralization. We thus conclude that there is a permanent negative effect of decentralization on job-finding which is particularly pronounced in the transition phase in the first year. NOTE. The figure depicts coefficients and their 95%-confidence intervals of yearly leads and lags of the decentralization indicator from a stock-flow regression on the monthly outflow out of unemployment into jobs. The regression includes a full set of dummies for PES offices and months. Standard errors are two-way clustered on the PES office and the month level. Figure 3: Dynamic difference-in-differences: Flows into jobs, The Role of ALMPs We now explore potential mechanisms that might explain the negative effect of PES decentralization on unemployment outflows. In this subsection, we examine whether decentralized PES offices change the assignment of job seekers into ALMP measures, as pointed out by Lundin and Skedinger (2006). In the subsequent subsections, we investigate potential effects of decentralization on the sanction strategy, on the placement strategy, and on other features of the local PES offices that affect the job-finding rate of their clients. 13

15 Active labor market policies are an effective tool to increase the employment probabilities of welfare recipients. In Germany, the four most common ALMP types are activation and integration schemes (e.g. classroom and on-the-job trainings of up to 3 months), job-creation schemes (subsidized public sector employment), job integration subsidies and professional development programs (medium-term (re)trainings). Yet, some measures are more effective than others (see Card et al., 2010; Heckman et al., 1999; Kluve, 2010, for reviews). Moreover, potential lock-in effects temporarily reduce the job-finding rates of training participants (Lechner et al., 2011). The negative effect of decentralization on unemployment outflows may therefore reflect (i) an increased overall use of ALMP measures which would raise the importance of lock-in effects; (ii) a reduced overall use of ALMP measures which could hamper the employability of job-center clients; or (iii) a shift of ALMP assignments from more effective ALMP measures (such as classroom training) towards less effective ones such as public employment programs. Localized providers may find these programs attractive because they offer a cost-saving way to increase the communes capacity in delivering public services. For example, program participants may support municipal maintenance yards while their welfare benefit is still covered by the federal government. Table 5: Difference-in-differences: Active labor market policies, Total Activation Job creation Job integration Professional schemes schemes subsidies development Decentralized *** (0.068) (0.096) (0.079) (0.069) (0.076) R-squared Districts Observations 23,606 23,605 22,775 23,173 22,326 NOTE. * p < 0.1; ** p < 0.05; *** p < Each column presents a different estimation of equation 1. The dependent variables are inflows of welfare recipients into the respective ALMP categories. Decentralized is a dummy equaling 1 for districts whose PES office is decentralized and 0 otherwise. Standard errors given in parenthesis are two-way clustered on the PES office and the month level. Regressions include the stocks and flows of unemployed and vacancies as well as a full set of dummies for PES offices and months. To assess whether the two provider types use ALMPs differently, we employ our stock-flow matching model from equation (1) and include entries into ALMP participation as the outcome variable. Table 5 presents the results. The first column shows that decentralized PES offices overall do not assign their clients more or less often to ALMP measures than centralized PES offices. Yet, they tend to assign them to different measures, as subsequent columns indicate. Particularly, decentralized offices send 30% more welfare recipients into job-creation schemes than centralized PES offices (see column 3). Figure 4 which graphs the respective dynamic estimation shows that this difference is also stable over time. Thus, we can rule out that explanations related to changes in the overall use of ALMPs account for the jobfinding reductions. Instead, as job-creation schemes are generally a less effective 14

16 ALMP type, a change in the ALMP mix likely contributes to the negative effect of decentralization on the unemployment outflow into employment. NOTE. The figure depicts coefficients and their 95%-confidence intervals of yearly leads and lags of the decentralization indicator from a stock-flow regression on the monthly inflow into job-creation schemes. The regression includes a full set of dummies for PES offices and months. Standard errors are two-way clustered on the PES office and the month level. Figure 4: Dynamic difference-in-differences: New participants in job-creation schemes, The Role of Sanctions Changes in the sanction strategy of local PES offices constitute a second potential mechanism that might explain the negative effect of decentralization on job finding. Sanctions are temporary reductions of unemployment benefits when job seekers do not comply to their search and meeting duties. There is ample empirical evidence that stricter sanction regimes and even the credible threat of being sanctioned increase the job-finding rate (see, for instance, Abbring et al., 2005; Lalive et al., 2005; van den Berg et al., 2016; van den Berg et al., 2004; van der Klaauw and van Ours, 2013) A strict monitoring regime can also induce adverse effects if informal search channels are more effective than formal ones. This is because monitored job seekers tend to substitute informal channels by formal search channels to provide their caseworkers with observable signals of their efforts (see van Den Berg and van der Klaauw, 2006; van den Berg and van der Klaauw, 2015). 15

17 Sanction strategies may differ across PES office types: Local PES providers, could prefer to sanction welfare recipients less intensively, for instance, to increase the re-election prospects of the local politicians in charge. Centralized PES offices do not have this incentive as the FEA is a federal institution that does not rely on local constituents. NOTE. The figure depicts coefficients and their 95%-confidence intervals of yearly leads and lags of the decentralization indicator from a stock-flow regression on the log monthly number of new sanctions. The regression includes a full set of dummies for PES offices and months. Standard errors are two-way clustered on the PES office and the month level. Figure 5: Dynamic difference-in-differences: Sanctions, We estimate equation (1) for the imposition of sanctions, substituting the outflow out of unemployment by the number of new sanctions issued in the given month. The results given in Figure 5 indicate a strong reduction of sanctions in the first year after the reform, but not in the following years. This points to short-run transition effects that may have contributed to the drop in job-finding during However, these results do not indicate that decentralized PES offices treat their clients more generously overall. Therefore, it is unlikely that the permanently lower job-finding is due to a laxer sanctioning regime The Role of Placement Strategies A third potential mechanism is a shift in the placement strategy. A PES office may accept a lower placement rate if emphasizes the quality rather than the quantity 16

18 of placements, even though the search for jobs with a better match quality usually requires more personnel resources and more time (Caliendo et al., 2013). The FEA, and therefore centralized job centers, have an incentive to increase the number of placements as this reduces the number of unemployed and therefore the expenditures of the FEA. Decentralized job centers, in contrast, might be more interested in stable, higher-paying placements because the districts bear the housing costs for households in need, irrespective whether they are unemployed or not. To assess whether decentralized offices changed their placement strategy, we analyze permanent placements, that is outflows out of welfare that did non resulted into returns into welfare within three months. Using this additional outcome, we estimate equation (1) with respective stocks, flows and shares as covariates. 14 The resulting estimates in Table 6, first column, reject the idea that decentralization had a differential impact on more permanent placements relative to non-permanent placements. Subsequent columns present models to test whether the pool of unemployed changed in localized districts. 15 A change could be the result if localized PES offices particularly target demographic groups that have particularly difficult labor market prospects. However, decentralization did not significantly change the shares of unemployed younger than 25 years, older than 55 years of foreign unemployed. Hence, we do not find evidence that decentralized PES offices use differing placement strategies. Table 6: Difference-in-differences: Permanent outflows out of welfare, Outflow Stock Stock Stock Permanent share Young share Old share Foreign share Decentralized (0.002) (0.186) (0.285) (0.277) R-squared Districts Observations 26,869 27,054 27,054 27,054 NOTE. * p < 0.1; ** p < 0.05; *** p < The column presents an estimation of equation (1) with the shares of different welfare or unemployment characteristics as an outcome variable. Decentralized is a dummy equaling 1 for 2012-decentralizing PES offices after the reform and 0 otherwise. Standard errors given in parenthesis are two-way clustered on the PES office and the month level. Regressions include the stocks and flows of unemployed and vacancies as well as a full set of dummies for PES offices and months. 14 Due to data availability, this analysis is only possible for welfare recipients in general. The group of all welfare recipients is about twice as large as those recipients considered unemployed. The differential consists of people currently in ALMPs, on childcare leave or permanently ill, among others. Hence, the decentralization effects for this group will be less pronounced. 15 A demographic composition of unemployment outflows is not available. 17

19 4.5. Other Considerations The lower job-finding rate in decentralizing districts might also reflect a lower number of and less qualified caseworkers (Hainmueller et al., 2015) or differences in institutional monitoring. For the caseworker argument to be true, the decentralized PES offices should have replaced the caseworkers of the former centralized PES office by new employees. Yet, this is not the case. Partially due to legal regulations, about 95% of the administrative and caseworker staff in the decentralized job centers before their decentralization, a total of about 4,000 employees, continued to work for the communal PES offices after their reform (Deutscher Bundestag, 2012). However, these employees subsequently had to adapt to new IT systems and practices. This kind of learning effects likely contributed to the particularly pronounced reduction of the placement efficiency in the first months after decentralization occurred. Five expert interviews we conducted with division heads of state and federal ministries responsible for PES supervision confirm this view. They also point out that some functions previously provided by the FEA had first to be set up in decentralized PES offices, such as a controlling system. Those transition effects should be completed by about 12 months after the reform. Consequently, we cannot attribute the reduced job finding to changes of the PES staff. Yet, the transfer of caseworkers into new organizational structures likely explains some of the transition effects visible in the first post-reform year. Finally, our ministerial interview partners consistently suggest that controlling systems are another decisive factor differing between the two provider types. The FEA imposes a very rigorous target control system on centralized PES offices that include target agreements, performance dialogues, ranking comparisons and strict monitoring by a federal institution (Vorstand der Bundesagentur für Arbeit, 2014). Decentralized PES offices have to report to state ministries but are independent otherwise. Besides, decentralized PES offices are voluntary members of a benchlearning program organized by the Federation of German Cities and Communes. This comparatively loose initiative organizes semi-annual symposiums for exchange and discussion to promote a continuous improvement process. Although quantifications are unavailable, our interview partners suggest that the FEA controlling system has more stringent requirements. Consequently, differences in the institutional performance monitoring may contribute to the reduced performance of decentralized PES offices Threats to Validity We now assess the robustness of our main finding and then determine the validity of our identifying assumptions. There are three potential concerns. First, applicants and non-applicants for decentralization might differ with respect to unobserved characteristics. Second, districts of the treatment and the control group might have experienced different labor market trends or shocks. Third, our estimates might be flawed by spillover effects from treated on non-treated districts. The first two 18

20 concerns question the validity of the common trends assumption, the third one the validity of the SUTVA. Table 7 presents additional checks to verify the main model. While the first column repeats the main model, column two shows that the finding is robust if we allow for different time patterns between East and West Germany. The third column shows that the basic finding still holds even if the first post-reform year 2012 is omitted from the sample. The reduction in job finding is then closer to 8%. To address potential selection based on a district s intention to apply, we first repeat our analysis using only non-successful applicants as control observations. If applicants for decentralization differ from non-applicants, we should also obtain effect estimates different from our main model. Column four shows the results based on the restricted control group. We still observe significant reductions of job finding following decentralization which are quantitatively very close to our initial estimates. This similarity also holds for other main outcomes from our channel analysis (see appendix, table 12). We consider this as first evidence against selection into the program. Table 7: Difference-in-differences: Specification checks, Denied applicants Baseline East x Month FE W/o 2012 as control group Decentralized *** *** *** *** (0.022) (0.022) (0.022) (0.026) R-squared Districts Observations 26,998 26,998 24,318 6,100 NOTE. * p < 0.1; ** p < 0.05; *** p < Each column presents a different estimation of equation 1. The outcome variable is the monthly log outflow out of unemployment into jobs. Decentralized is a dummy equaling 1 for districts whose PES office is decentralized and 0 otherwise. East is a dummy equaling 1 for districts in East Germany and 0 otherwise. Standard errors given in parenthesis are two-way clustered on the PES office and the month level. Regressions include the stocks and flows of unemployed and vacancies as well as a full set of dummies for PES offices and months. To further assess this issue, we apply a placebo reform to non-successful applicants and compare their outcomes to the districts that did not apply for decentralization. The treatment effect thus now captures the effect of being interested in decentralization rather than being actually decentralized. If this treatment effect is statistically significant, applicants differ from non-applicants and our preferred control group is inappropriate to measure the effect of decentralization. In contrast to the decentralization status, as given by Table 8, the applicant status has no statistically significant effect on job finding (first column) or other main outcomes (see appendix table 13). We conclude that applicants and non-applicants do not differ with respect to unobserved characteristics, supporting the common trends assumption underlying our main analysis. 19