The outcome of coaching and training for self-employment. A statistical evaluation of outside assistance support programs for unemployed business founders in Germany

Abstract

This paper focuses on the question of whether improving the competence of new business founders through programs that offer external expertise enhances the duration of self-employment. In our analysis, we focus on three different programs that are provided along with a financial subsidy and that focus on founders who started a business while they were unemployed. We found that participation was strongly determined by regional patterns and time, and that individual characteristics were less important. These results reflect a particular regional specialization in promoting self-employment. A statistical matching approach was used to control for selectivity and was performed in a way that explicitly considered differences across regions and over time. The results show that the treatment effects tended to be low. However, we found evidence that external expertise increased passive learning.

Zusammenfassung

Die vorliegende Arbeit geht der Frage nach, ob sich Förderprogramme, die helfen externen Sachverstand bei einer Gründung einzubinden, positiv auf die Verbleibsdauern in Selbständigkeiten auswirken. Hierzu werden drei unterschiedliche Programme betrachtet, die zusätzlich zu einer finanziellen Basissicherung Gründungsvorhaben aus der Arbeitslosigkeit fördern. Wir finden, dass die Selektion in die Förderprogramme stark durch regionale Merkmale determiniert wird und dass individuelle Charakteristika bei der Inanspruchnahme der Förderleistungen wenig relevant sind. Dieses verweist auf eine regionale Spezialisierung in der Ausrichtung der Förderung bei der Aufnahme einer selbständigen Tätigkeit durch die aktive Arbeitsmarktpolitik. Das angewandte Selektionskorrekturverfahren (statistisches Matching) berücksichtigt diese Besonderheiten, so dass neben individuellen Merkmalen explizit auch regionale und zeitliche Aspekte kontrolliert werden. Die Analysen zeigen, dass die Wirkung der zusätzlichen Förderung für die Verbleibsdauer in Selbständigkeit eher gering ausfallen und dass Selbständigkeitsperioden bei Inanspruchnahme externer Expertise schneller beendet werden als ohne. Dieses deutet darauf hin, dass externe Expertise bei Gründungen aus der Arbeitslosigkeit tendenziell passives Lernen fördert.

1 Kurzfassung

In der vorliegenden Studie untersuchen wir die Teilnahmeeffekte zusätzlicher Fördermaßnahmen gemessen an der Stabilität der Selbständigkeit über einen Zeitraum von fünf Jahren. Wir evaluieren den Effekt, der von der Einbindung externer Expertise ausgehen dürfte. Dabei konzentrieren wir uns auf Personen, die mit Überbrückungsgeld gefördert wurden und hierzu parallel oder im Vorfeld eine weitere Förderung durch Gründertraining-, -coachings oder Freie Gründungsförderung (Maßnahmen der Freien Förderung nach § 10 SGB III mit zu erwartendem Schwerpunkt im Bereich Einbindung externer Expertise) erhalten. Damit untersuchen wir Fördermaßnahmen, die im Zusammenhang mit der aktiven Arbeitsmarktpolitik in Deutschland umgesetzt werden und einen nationalen Bezug aufweisen. Darüber hinaus können wir von weitestgehend homogenen Förderbedingungen ausgehen. Inhaltlich gehen wir mit der Evaluationsanalyse der Frage nach, ob das Einbinden externer Expertise eine positive Wirkung auf die Verbleibsdauer in Selbständigkeit hat.

Für unsere Analyse verwenden wir Daten der Integrierten Erwerbsbiografien (IEB). Dieser Datensatz enthält Informationen zu Perioden abhängiger sozialversicherungspflichtiger Beschäftigung und umfasst Teilnahmezeiten an Programmen der aktiven Arbeitsmarktpolitik, Lohnersatzleistungen und zu Meldungen zur Arbeitsuche. In unserer Analyse finden wir, dass die Wahrscheinlichkeit zur Programmteilnahme vorwiegend durch regionale Merkmale erklärt wird. Dieses Ergebnis weicht von bisherigen Erkenntnissen aus der Evaluationsforschung in Deutschland ab und verweist auf eine ausgeprägte regionale Spezialisierung im Bereich der Gründungsförderung durch die Bundesagentur für Arbeit. Wir berücksichtigen diese Besonderheit indem wir einen Matching-Algorithmus zur Selektionskorrektur verwenden, der auf Zeit-Regionen Stratas basiert.

Die Evaluationsergebnisse zeigen, dass die Effekte der zusätzlichen Förderung eher gering ausfallen und dass die Einbindung externer Expertise mit einer Verkürzung von Selbständigkeitsperioden verbunden ist. Zudem finden wir schwache Hinweise auf zeitlich abhängige Effekte (vgl. u.a. Rotger et al 2012). Insgesamt sind die gefunden Effekte aber von geringer statistischer Signifikanz. Die relativen Effektgrößen sind jedoch nicht zu vernachlässigen. So finden wir z.B. für die Teilnahme an Coaching-Programmen insgesamt 17 % weniger Austritte in eine abhängige Beschäftigung (verglichen zu Gründern die nur mit Überbrückungsgeld gefördert wurden). Signifikante Effekte bei Gründungstrainings konzentrieren sich auf erhöhte Austritte in Arbeitslosigkeit (7 %). Für Teilnahmen an der der freien Gründungsförderung finden wir 6,7 % mehr Austritte in Arbeitslosigkeit und 10 % weniger Austritte in eine abhängige Beschäftigung.

Unsere Ergebnisse zeigen auf, dass Gründertrainings, Gründercoachings sowie die Freie Gründungsförderung nicht die von der Politik erwartenden Effekte zur Stabilisierung von Selbständigkeitsperioden für Gründer aus der Arbeitslosigkeit hervorrufen. Mit diesem Aussage unterstützen wir die Ergebnisse von Karla und Valdvia (2011), die ebenfalls auf geringe und überwiegend insignifikante Fördereffekte ähnlicher Programme in Peru hinweisen (für ähnliche Ergebnisse siehe auch Shutt und Sutherland 2003 sowie Eckl et al. 2009). Dieses würde zusammenfassend bedeuten, dass eine zusätzliche Gründungsförderung, die auf das Einbinden externer Expertise konzentriert ist, keine kluge politische Entscheidung ist. Allerdings widerspräche diese Bewertung anderen Studien, die sehr deutlich machen, dass maßgeblich die Qualifikation des Gründers eine zentrale Erfolgskomponente darstellt (Chandler und Hanks 1998, Cressy 1996). Ebenso: basierend auf einer Studie zu ähnlichen Förderprogrammen wie wir sie untersucht haben, finden Michaelides und Benus (2012) positive Wirkungseffekte bei Gründern aus der Arbeitslosigkeit für die USA. Allerdings ist der Förderrahmen bei den untersuchten Programmen deutlich selektiver ausgerichtet. Letztendlich bleiben die Untersuchungsergebnisse zudem in mancherlei Hinsicht vorläufig, da die Datengrundlage der Analysen mit nicht unerheblichen Limitationen behaftet ist.

2 Introduction

From a practical perspective, policy and research have long focused on capital endowments as major constraints to entrepreneurship, small business development and self-employment (Almus 2004). Since the 1990s, it has been recognized that capital and qualifications interact and that deficits in expertise constitute further constraints to self-employment (e.g., Cressy 1996; Chandler and Hanks 1998; Shutt and Sutherland 2003). One political consequence was to begin initiatives to combine financial support and qualification in promoting entrepreneurship (e.g., Chrisman et al. 2005; Michaelides and Benus 2012). For European countries, the European Employment Strategy (EES) offered a master framework for implementing experimental settings in the late 1990s and the early 2000s to develop new promotional programs that focused on including external expertise to enhance the qualifications for starting a new business.

However, with respect to the inclusion of external expertise, evidence on the outcome of related programs for self-employment is mixed. For example, experience reported from the Small Business Development Center (SBDC) program in the U.S. indicates that the intensity and quantity of advisory services had a positive but inversely u-shaped effect on firm growth and sales development (Chrisman et al. 2005; Chrisman and McMullan 2004). Shutt and Sutherland (2003) did not find a significant effect of local advisory support programs on the chances for survival of newly founded businesses in England. Similar findings were also reported for the FINCA-Peru program (Karlan and Valdivia 2011). Additionally, Eckl et al. (2009) did not find that advisory support improved firm growth when they focused on ESF (European Social Funds) co-funded start-ups in Germany. Similar evidence has also been reported for the English Business-Link-Network program (Mole et al. 2008). However, research has provided evidence for the complexity of the mechanisms of training and advisory support programs. For example, Parker and Belghitar (2006) discuss potential quality effects of corresponding programs, just as Wren and Storey (2002) indicate that assistance programs may be most effective for medium-sized business start-ups. Chrisman and Leslie (1989) focus on the potential variation of treatment effects depending on the start-up period in which the support program actually begins and discuss whether the program focuses on strategic or operating assistance. Rotger et al. (2012), in addition, report that the effect of outside advice diminishes over time.

In this study, we investigated the outcomes of three different support programs (training courses, support for using business coaching and a flexible promotion program) that promoted the implementation of external expertise to enhance self-employment. These programs complemented a financial promotion program (bridging allowance) and were part of the German active labor market policy framework in the early 2000s.¹ At the heart of our study, we focused on the effect of the additional support on the sustainability of self-employment.² In contrast to earlier research, we evaluated programs that allowed for the study of heterogeneity in terms of flexibility (standardized topics, problem-oriented counseling and flexible promotion) and in terms of the timing of using external expertise (before and after start-up). Furthermore, we focused on promotion programs that were part of a nationwide policy program. Research that allows for insight into greater nationwide policies on promoting self-employment is scarce.

Assessing the net outcome of using external expertise to improve self-employment sustainability faces at least two important challenges that require extra attention. First, clarification is needed on how external expertise may operate in affecting particular outcome measures. This is not trivial because from a theoretical perspective, one must be aware that the inclusion of external expertise may be ambiguous with regard to its effects on the chances for survival. In fact, external expertise may improve productivity (Ericson and Pakes 1995), but it may also enhance passive learning (Jovanovic 1982). This may, in turn, also foster retiring from self-employment and can thus have effects opposite to those expected by politicians (see LeBrasseur et al. 2003; Castrogiovanni 1996). Note that the view of the potentially ambiguous outcomes of external expertise on self-employment sustainability differs from the view that was typically emphasized in earlier research (e.g., Rotger et al. 2012; Michaelides and Benus 2012). Second, selection effects are an important issue. Unobserved characteristics may govern the choice to take advantage of external expertise. With regard to our evaluation approach, we followed a broad strand of recent evaluation studies of labor market interventions, and we controlled for endogeneity and selectivity using a statistical matching approach (e.g., Hujer et al. 2004; Almus and Czarnitzki 2003; Baumgaertner and Caliendo 2008). In particular, to capture regionally embedded differences in the quality of the interventions and the treatment assignment, we extended the general framework by giving extra weight to regional characteristics in the matching approach.

The data we used were the Integrated Employment Biographies (IEB), which are compiled by the Institute for Employment Research of the German Federal Employment Agency. This data set consists of information from four distinct administrative registers and combines employment biographies and detailed information on participation in employment and training programs. We saw five advantages in using these data: (1) we were able to observe a five-year period to assess program outcomes; (2) the data rarely suffered from the types of participation or attrition bias that are usually found in survey data; (3) the data permitted valid identification during periods of self-employment of the types of nonfinancial support received and detailed information on the individual employment histories; (4) all of the individuals in our study received a bridging allowance to start their ventures, which ensured a relatively homogenous study population in terms of entrepreneurial intention and (5) the data allowed us to identify the regional context in which the intervention took place.

Section two describes the institutional and conceptual setting of German self-employment promotion as it is implemented in active labor market policy. Section three presents the dataset and describes the construction of the sample for analysis. Section four focuses on the analysis strategy. This includes a brief theoretical foundation of the research, descriptive information and a short discussion of the selection process plus the implementation of the statistical matching procedure. Section five presents and discusses the empirical results. Finally, section six summarizes the study, makes concluding policy-related remarks, and offers suggestions for future research.

3 Promotion of self-employment as part of active labor market policy

3.1 The basic framework of the promotion of self-employment since the late 1990s

The field of self-employment promotion in German active labor market policy was first addressed in 1986 in the form of a financial subsidy aimed at supporting the transition from unemployment to self-employment (known as ‘Überbrückungsgeld’: bridging allowance). During the mid-1990s, self-employment was promoted through a more generous bridging allowance. In the late 1990s, the promotion of self-employment had been expanded in general. As for example, the implementation of Social Code Book III (SGB III) in 1998 led to a greater degree of managerial responsibility for local employment offices, as based on § 10 SGB III (discretionary measures of regional active labor market policy administered by local employment agencies; ‘Freie Förderung’). To a large extent, this greater degree of freedom was used to increase the promotion of self-employment at local levels. Second, in 1998, the active labor market policy of the Federal Employment Agency implemented a nationwide program as part of the national ESF funding.³ Initially, this funding framework only focused on promoting general training programs, but it largely shifted toward promoting support for self-employment between 1998 and 2008.⁴

3.2 Characteristics of the programs that promoted self-employment⁵

During the early 2000s, the bridging allowance was the most important program in the field of self-employment promotion in Germany.⁶ Access to this program was limited to the unemployed or to individuals who were threatened by unemployment and sought to avoid unemployment by becoming self-employed. Furthermore, this program was only open to those who were entitled to unemployment benefits and only in cases in which the new venture would enable the individual to leave unemployment. Support was only granted for applications with a positive assessment of the business concept (e.g., a local chamber of commerce). Finally, the bridging allowance offered a subsidy comparable with the sum of the unemployment benefits and covered social security contributions for the first six months of the new business activity.

Building on the ESF funding framework, ‘training’ and ‘coaching’ programs were implemented to ensure qualified outside assistance during the preparation and early stage business development of new businesses founded by the previously unemployed. In accordance with the implementing regulations, training courses were focused on seminars that lasted between 4 and 12 weeks and were supposed to cover topics such as bookkeeping, business plan development, finance, sales and legal issues to ensure sufficient business preparation. In contrast, the coaching program was designed to cover expenditures for business consultancy, such as might be related to tax issues, sales development, marketing or accounting support, to improve early stage business development. Initially, there was no detailed official regulation concerning the form and content of the coaching.⁷ Both programs were legislated to cover all direct expenses (course fees or payments for the coach) as well as indirect costs for child care, accommodation, and travel (up to a maximum of 4,600 €).

Finally, the discretionary measures of the regional active labor market policy offered a more flexible promotion framework. Based on Social Code Book III, § 10 (‘Freie Förderung’; hereafter ‘discretionary measures’), local employment agencies were allowed to administer locally specialized programs. This strategy partly deviated from the generally centralized German labor market policy. In general, the discretionary measures offered a framework that allowed employment agencies to concentrate on special industries or target groups and permitted them to address specific regional problems. However, over time, this source of funding has increasingly been used to promote self-employment. For example, local entrepreneurship centers and financial subsidies or training programs for nascent entrepreneurs were funded within the discretionary measures (so-called ‘discretionary start-up support’ (DSUS)). Despite its heterogeneous setting, reports from the Federal Employment Agency indicate that DSUS—if used as additional support—mainly comprised types of support that we would typically define as qualification-oriented.⁸

3.3 Conceptual objectives

Note that our study addresses the combination of different support programs that promoted self-employment activities while concentrating on the outcome of programs that allowed for the inclusion of external expertise. We used the bridging allowance funding as a basic program to identify entries in self-employment. It is vital to note that this program ensures subsistence for everyday life during the first months of business activity. With respect to the total arrangement of self-employment promotion, this implies that the overall risk of financial distress during a new business’s start-up period is already relaxed. In contrast, the other programs can be characterized as different approaches that promoted the inclusion of external expertise for a new self-employment activity. Figure 1 summarizes the range of promotion activities in our research context and emphasizes the conceptual rationales of the related programs.

Fig. 1

Conceptual objectives related to policy interventions

The major issue with the training program is that it focuses on skill enhancement before the start-up begins to ensure sustainable and assessable business planning and market preparation. Hence, a substantial aim of training is to improve the ability of the business founder to assess the new business option and to evaluate ongoing business development. Furthermore, training courses also intend to improve specific knowledge on topics related to business activities such as technical skills in accounting, finance and marketing. For this, the implementing regulation outlines rather homogenous requirements for the form and content related to the training seminars. In contrast, the use of the coaching program focuses on the period after start-up and allows for an inflow of external knowledge during the early stage business development. The major aim of this program is to allow for individual and context-specific support to overcome technical problems and to compensate specific technical and personal deficits. Furthermore, skill enhancement may also be relevant in this context since learning is allowed to be problem- and context-specific. Therefore, coaching not only improves the business development of the newly founded business but also fosters the individual’s ability to assess and manage the economic potential of the business concept in general. Coaching in this sense should be understood as a highly flexible tool that allows for the inclusion of external expertise in the post entry period.⁹ Note that this implies a rather wide definition of coaching (e.g., Chrisman and McMullan 2004). Finally, discretionary start-up support can be characterized as having the highest degree of freedom in terms of the timing of the intervention, related topics and concerning the way this support is used. Nevertheless, it is plausible to assume that DSUS mainly focuses on incorporating external expertise for a new business venture (see above). Hence, DSUS may provide a mixture of the characteristics related to what is covered by the training and the coaching programs.

4 Data and sampling

The data used for the analysis were a sample from the Integrated Employment Biographies (IEB). These data were compiled from four administrative sources that originated from the registers of the Federal Employment Service.¹⁰ The data comprise employment and benefit histories till 1990 and official registrations for job search, periods of unemployment, and participation in active labor market programs till 2000. By combining these sources, the IEB provides a detailed historical perspective of periods of employment and unemployment.

The information provides the exact start and end dates of each period. Source-specific information adds data on individual schooling, type of employment, job characteristics, income and detailed information on qualifications. In addition, we added data from the Establishment History Panel (Betriebs-Historik-Panel—BHP); see Spengler 2008)) to include the characteristics of corresponding employment episodes and regional labor market information.¹¹ In the latter case, we added data from the official statistics of the Federal Employment Agency (for a detailed overview of the attributes, see Table 4 in the Appendix).

To control the context of the additional support programs, we restricted the analysis to individuals who received a bridging allowance starting not before the year 2000. This restriction ensured a valid identification of self-employment periods. Start-ups after the first quarter of 2003 (hereafter, 2003(I)) were excluded because they could have been affected by institutional changes that were introduced in 2003 (see Caliendo and Kritikos 2009).¹² Dropouts and episodes for which it was difficult to identify valid start or end dates were also removed from the analysis.¹³ To focus on valid additional support, including external expertise, we also restricted the analysis to cases of support that occurred within a certain time before and after receipt of the bridging allowance.¹⁴

The treatment was defined as participating in an additional self-employment support program. More precisely, the following three treatments were distinguished: (1) participating in short-term self-employment training, (2) being assigned to a coaching program, and (3) receiving discretionary start-up support. Unfortunately, there was no further information related to the quality of these interventions in the data source. Multiple treatments, such as combining training and coaching or discretionary start-up support for self-employment, were not studied. For the comparison group, we used individuals who received a bridging allowance but no additional support. This group included all individuals who did not receive extra support (external expertise support programs) during the time period under observation and those who did not receive valid additional support.¹⁵

The outcome measure of our evaluation concentrated on capturing the general objective of the active labor market policy. Here, we focused on the stability of an individual’s period of self-employment.¹⁶ However, it should be noted that the data did not provide a direct measure of this variable. Therefore, we used an inverse definition in which instability was approximated by any record that was not related to self-employment promotion after entry into self-employment. These records included any observation of unemployment benefits, employment, job-search promotion or non-self-employment promotion after the individual began receiving the bridging allowance.¹⁷ These observations were used to measure employment stability as associated with self-employment activity. In addition, we concentrated on the duration of self-employment and allowed for different types of exits from self-employment (i.e., departing for unemployment or employment), which allowed us to further capture aspects of the economic reasons for exiting self-employment. However, it is important to note that because of data limitations, we did not observe any non-labor-market-related positions (e.g., retirees) and that the observation of promotion activities may have depended on the fact that individuals had to be entitled to receive unemployment benefits. Both of these facts may have caused the underestimation of the true rate of exiting self-employment.¹⁸

The first outcome measure (Y ¹) was defined as the likelihood of exiting self-employment during the first 36 months after accepting the bridging allowance (T ≤ 36). The second measure (Y ²) provided information on the time-dependent survival probability. The second measure was the inverse of the probability of exiting self-employment during or before a time interval (t; t’) assuming that the individual had entered that time interval. Both measures were calculated for k types of exits (all types of exits, exits into unemployment, and exits into employment):

$${{Y}^{1}}:Pr({{T}^{k}}\le 36)$$

(a)

$${{Y}^{2}}:1-\Pr (t\le {{T}^{k}}<t'|{{T}^{k}}\ge t)$$

(b)

5 Analysis

5.1 Theoretical evaluation framework

To assess the theoretical outcome of expertise related to supporting programs, we assumed that the programs directly or indirectly affected the founder’s capability to evaluate and improve business prosperity. Imagine the following theoretical foundation: Any business idea embraces a specific distribution of potential outcomes (e.g., reward, income, utility)¹⁹ while the exact position on this potential outcome distribution is unknown to the individual unless the person begins to exploit the business idea. If the expected position on this distribution exceeds a certain threshold (the opportunity costs), the individual becomes self-employed (Gimeno et al. 1997). As new information comes in (along with business activity), the founder becomes more capable of assessing the true position of the potential outcome distribution. We characterize this as passive learning (Jovanovic 1982). Hence, increased information may have two effects: (1) the founder realizes that the initial assessment overrated the true business potential; or (2) the initial assessment of the business potential was correct or even underrated. In the first scenario, we expected that the founder would quit his or her new business. In the second scenario, we expected that the business would continue. Note that in this context, an active intervention (e.g., lowering production costs because of counseling) can be interpreted as improving information that helps uncover the true market potential of the business idea. We may call this active learning (Ericson and Pakes 1995).

For example, because training is conducted before a business is launched, better and faster assessments are possible for two reasons. First, start-up training enhances convergence toward the true option value of the business. This should lower the risk of initial overestimation and improve the chance for additional active improvements after the start-up. Second, training may allow for better assessments to show that the true option value is lower than was initially expected, which would increase the risk of exiting self-employment. In contrast, coaching is only focused on transferring expertise in the post-entry period. Nevertheless, the rationale related to coaching remains the same as the one related to training because it improves the realization of the true outcome potential of the business idea (active and passive learning). Similarly, we should expect a similar mechanism for the DSUS. However, we should be aware of a higher degree of freedom, which may allow a more accurate mix of active and passive assessments (learning). Nevertheless, the net outcome related to the additional supporting programs on including external expertise remains an empirical question.

5.2 The evaluation strategy

To evaluate the promotion outcome, we used a comparison framework in which the populations of individuals with and without policy interventions were used to identify counterfactual observations to estimate average treatment effects (Rosenbaum and Rubin 1983)_. ²⁰ Compared with other methods, the advantage of matching is that the set of necessary restrictions is highly limited (e.g., it does not need the exogeneity of conditioning variables and exclusion restrictions or the separability of outcome and choice equations). In particular, matching techniques do not require a parametric specification of the outcome function or the selection process. However, they emphasize the existence of a common support that makes it possible to study heterogeneous treatment effects. Because of matching, the bias reduction fundamentally depends on the availability of rich information that allows for including attributes that simultaneously determine the treatment assignment and the potential outcome of the comparisons (the conditional independence assumption; CIA).

Furthermore, the identification of net effects fundamentally relies on the absence of general equilibrium effects (the stable unit treatment value assumption, SUTVA; see Holland 1986). Participants had to be stochastically independent across all observations, and the outcome had to be independent of the mechanism by which participants received the treatment. In more practical terms, SUTVA implies that an individual’s potential outcome and his or her likelihood of receiving a treatment should not interfere with those of others. These conditions may be violated in a regional evaluation context. To clarify, consider an intervention that is small at the national level but may be highly relevant to a particular region. In such a case, we would need a better understanding of the regional level for the selection process. If regional characteristics are important, the validity of SUTVA will require a more local perspective, which must result in implementing a matching approach that considers the regional support context.

5.3 Distribution of participation

Observations entered the risk pool in 2000 and were right-censored to December 2005. We found that inflows into the bridging allowance increased from over 85,000 in 2000 to 140,671 in 2003. In total, and considering the sample restrictions presented above, 418,856 cases of bridging allowance were included in this study. Discretionary start-up support (DSUS) showed the largest number of participants (N = 30,481), followed by cases of coaching (N = 13,737). The number of participants in training courses remained relatively small (N = 2,131).

Following the discussion above and the outline of self-employment promotion in Sect. 2, we first took a closer look at the regional variation in the relative relevance of the individual programs. Figure 2 illustrates the ratio between the number of participants in an additional support program (training, coaching or DSUS) and the total number of participants who received the bridging allowance for each of the 176 local labor market districts (note: the x-axis is based on the official identifiers of the local districts).

Fig. 2

The relative importance of different self-employment promotion programs across regions

As observed in Fig. 2, most labor market districts had low ratios of additional support policies, which indicate the limited importance of external expertise as an instrument for promoting self-employment. However, in some regions, these extra support activities were close to or exceeded 40 % (e.g., DSUS and coaching). In contrast, training for self-employment remained relatively unimportant in most labor market districts (close to zero). Clearly, there were strong local differences in the costs of managing programs or in the expected gains that could have driven this regional heterogeneity.

This finding is important for the evaluation because it indicates the high relevance of a local policy implementation (Hirschenauer 2001). This observation also supports the hypothesis that there is a particular regional specialization in the strategy of self-employment promotion. Furthermore, this finding directly emphasizes the concern regarding general equilibrium effects (in regions with exposed promotion activities) and the problem of limited joint support (in regions with minimal additional activities). To overcome this potential source of bias, we excluded regions that had more than 40 % of additional promotion in one of the studied promotion programs.²¹ As a result, 17 local labor market districts were excluded from the study. This corresponded to a loss of nearly 29,700 observations (12,500 from the bridging allowance; 3,400 coaching observations; 12,200 from DSUS). Furthermore, for matching, we excluded all regions that did not support the statistical matching approach.²²

5.4 The selection process, potential outcomes, and the validity of the CIA

5.4.1 Treatment selection

Prior to the evaluation, we examined the selection process to gain further insight into the treatment assignment. Note that the selection processes associated with the interventions was complex in nature. First, interventions were implemented within highly regionalized policy frameworks. This induced a much higher supply-side effect on the selection process than is typically found in active labor market policy. Based on the local labor market structures and local policy strategies, there may have been varying motives to focus on self-employment as a promising way to improve local employment (e.g., offering more or less support). As such, particular regional policy strategies had to be considered in the matching frameworks. Second, selection only took place if the individual had evaluated the training, coaching or DSUS to be of advantage to the founder.

To simplify the selection process, imagine that the selection into an additional support program is a result of the negotiation between the local employment office case manager and the applicant. Considered from the agency perspective, regional differences in the negotiation may result from different local policy strategies, different cost and benefit structures, and the perceived success of the intervention. In contrast, the single case manager may be less important. Generally, case managers are not trained to evaluate the extra promotion needs of business founders. Instead, they mainly follow general routines and strategies that are developed at the agency level. Therefore, the supply-side selection will be driven largely by specific local labor market conditions. Note that time may be a crucial factor in this context because it captures variations in learning about efficient policies and allows the agency to establish quality benchmarks.

With regard to the selection process at the individual level (the demand-side), we expected that negotiation would be mainly driven by the individual’s cost-benefit functions. The quality of the business concept that may govern the need to improve business preparation is an important factor in this context. However, because the business concept itself was not observable, we assumed that the driving force that determined the quality of the business concept as well as the overall cost-benefit ratio in assessing the expected returns on additional support were related to individual expertise. As a result, the negotiation position of a founder’s selecting training, coaching or DSUS could be formulated as a function of the founder’s experience, formal qualifications, and employment biography. With this framework, we directly linked the founder’s ability to develop a certain quality level of the business concept and his or her competence to assess the quality of the business and to outline the potential return on the external expertise. Furthermore, this statement implies that better-qualified business founders may interact more efficiently with external experts, ask better and more precise questions, and be more likely to capitalize on previous knowledge to improve their businesses.

In sum, we expected that the quality of external expertise—and therefore the selection process—would strongly depend on a mixture of regional policy strategy and individual qualifications. Qualitative interviews support this perspective of an interrelated selection (Oberschachtsiek 2007). However, exact information on each individual selection process was missing. For general evidence on the role of individual and regional factors in the selection process, see Table 1, which reports the related statistics separately for each support program. These results are based on logit models and cover different sets of attributes. Because we were only interested in general information on the selection process, Table 1 only focuses on model fit statistics.²³ The reported statistics (Bayesian information criteria [BIC] and the likelihood ratio [LR]) provide information about the entropy of the statistical modeling that can be used to describe the general pattern of the selection process (Burnham and Anderson 2004). For further details on the selection process see also Table 5 in the appendix.

Table 1 Factors affecting treatment selection. (Source: IEB, own calculations)

As observed in Table 1, the greatest model improvement was gained by introducing regional characteristics (especially by introducing an indicator for the local labor market district). This finding directly supports the hypothesis that the local agent’s cost or utility function (policy strategy) is of high importance in the overall selection process. We found that using external expertise support programs had a higher ratio in eastern Germany (less pronounced for the DSUS) and that time and the local composition of the external expertise support programs strongly affected the selection process. In contrast, individual characteristics were of little informational value in explaining program participation. Nevertheless, we found that the probability of receiving support for additional external expertise was higher for males and that it increased with the age (inversely U-shaped) qualification level.

5.4.2 The validity of the CIA

Concerning the validity of the statistical matching approach, it was critical that we pay sufficient attention to information on both (individual and regional) selection levels, which correlated with the potential outcome. Particular concerns in our context may have been the role of the business concept in treatment selection, the role of the quality of the intervention and the effect of the local context on the treatment assignment (regional policy).

First, it is important to note that our study focused on self-employment creation that originated from an unemployment position. Note that this group is rather homogeneous in terms of capital endowment, business motivation and growth intentions. For example, Hinz and Jungbauer-Gans (1999) and Oberschachtsiek (2012) reported that previously unemployed founders typically began businesses that needed less financial capital endowment and relied relatively more on a founder’s human capital structure. Associated with this overall difference, we believed that it was acceptable to assume that the complexity of businesses founded to induce primary self-employment was lower than it was for those that were founded in general context of entrepreneurship (e.g., lower growth intentions). In addition, it seemed plausible to us that among the unemployed, the variation in business complexity would be rather low. As a result, missing information on business quality should—on average—have been less problematic in the population of the previously unemployed.

Second, we posited that the quality of the business concept and potential outcome would be strongly interrelated with the founder’s qualifications. Previous research indicates that experience, schooling, gender, and motivation are highly correlated with the quality of the business idea and the assessment of business prosperity as well as with the duration of self-employment.²⁴ As a result, the need to select into training and coaching strongly depends on the founder’s qualifications because it determines the need to improve the business concept and potential outcome. Therefore, having information on the individual’s previous job experience, schooling, and professional training—as we had in our data—is highly important.

Third, we expected that the quality of the interventions would strongly depend on regional policy strategy. On average, the local agency may control quality by managing the total participation rate. A high participation rate may signal market demand, which may cause a downshift in marginal quality. In addition, local economic conditions have a direct effect on the willingness to set up new ventures because economic conditions are correlated with the level of competition and overall demand (e.g., Falck 2007). As a result, the local level is highly important in supporting the conditional independence assumption. In this study, we controlled for the regional implementation of promotion programs by modifying the matching algorithm.

Finally, an additional important factor supporting the argument that sufficient information was included in our data is that the major issue of selection was already absorbed by the decision to apply for the bridging allowance (i.e., the motivation to start a business was already captured) and because specific data restrictions applied (see above). Because both groups—the treated and potential comparisons—entered self-employment in reality, most of the unobserved factors that govern an individual’s intention to start a venture should have been equally distributed in our study population. Therefore, we saw no fundamental aspect that would lead us to think this motivation would be different between the treated and untreated populations.

Overall, our opinion is that we were able to control for sufficient information to balance treated and untreated individuals. In particular, we relied on the assumption that motivation was controlled for by concentrating on bridging allowance as the basic selector. Furthermore, for the quality of the business, we assumed that the founder’s human capital was a good approximation, and we were confident that the policy strategies delineated by the treatment assignments were sufficiently approximated by local labor market information and local active labor market policy. Nevertheless, in our setting, we were not able to control for the use of similar programs that were provided by other authorities (e.g., local chambers of commerce). Therefore, we had to assume that in general, additional support for the inclusion of external expertise (focusing on business founders who exited unemployment) was mainly offered by the Federal Employment Agency. Furthermore, we also believe that for a substantial bias related to this deficit, it would have been necessary for a greater share of founders who had exited unemployment to have received support from other authorities. We state that this was unlikely the case.²⁵

5.5 Implementation of the matching procedure

In our evaluation, we concentrated on the average treatment effect on the treated (ATT) as the most interesting parameter. This estimator is defined as the difference between the mean outcome of the treated $Y_{i}^{D=1}$and the estimated counterfactual outcome $\hat{Y}_{j}^{D=1}$, which provides information about the net outcome of a treatment for those who were treated:

$$ATT=\sum\nolimits_{i}{\left[ Y_{i}^{D=1}-\hat{Y}_{j}^{D=1} \right]}\,given\,that\,\hat{Y}_{j}^{D=1}=\sum\nolimits_{j}{{{W}_{i,j}}Y_{j}^{D=0}}$$

(1)

i characterizes the treated and j the untreated individuals. In our analysis, individuals who only received the bridging allowance were defined as untreated, and individuals who had received additional support to include external expertise were defined as treated. As the right-hand side of Formula (1) shows, the estimated counterfactual outcome for those who received additional support (training, coaching or DSUS) was taken from the mean outcome of the bridging allowance population with no support ($\hat{Y}_{j}^{D=1}$). We calculated this counterfactual outcome as the weighted mean outcome of the non-treated, in which the individual weights ${{W}_{i,j}}$ referred to the distance between comparisons j and i. To ensure the equal importance of the treated and untreated observations, weights were restricted to the following conditions:

$$\sum\nolimits_{j}{{{W}_{i,j}}=1},{{W}_{i,j}}\in \left[ 0,1 \right]$$

(2)

The distance between the treated and the untreated was used to define the comparability of the comparisons. For technical reasons, we used the Mahalanobis distance, which allowed us to set a distance measure and was used as a measure of equality.²⁶ To stress the importance of specific characteristics, we used a more complex procedure to define the distance measure and apply it to the matching approach. For example, to permit a more detailed representation of the selection process, we carried out a direct matching procedure for the type of region and calendar time and then calculated three propensity scores (see the full matching approach on the next page) that were entered into the distance measurement.

Finally, the weighting program W was implemented using a kernel function K (Epanechnikov kernel) based on the bandwidth h and distance function u, where u was defined based on the distance between the balancing scores (B(x))—that is, the dissimilarity between the treated and untreated observations—and bandwidth h:²⁷

$${{W}_{i,j}}=\frac{{{K}_{i,j}}}{\sum\nolimits_{j}{{{K}_{i,j}}}}with\,{{K}_{i,j}}=\frac{3}{4}(1-{{u}^{2}})\,\,{{1}_{\left\{ \left| u\left| \left. \le 1 \right\} \right. \right.\right.}}and\,u=({{B}_{i}}(x)-{{B}_{j}}(x))/h$$

(3)

Our Mahalanobis distance kernel matching proceeded as follows:²⁸

• 1. Identify j and i.

• 2. Skip regions with no support (zero participants between 2000 and 2003).

• 3. Estimate three propensity scores Ps(x): Pr(D = 1|X ⁱ ), Pr(D = 1|X ^rc ) and Pr(D = 1|X ^rd );²⁹ where Pr(D = 1|X = x) =  1/ (1 + e ^X’b ).

• 4. Stratify the matching procedure into matching clusters (by annual quarter and type of region³⁰).

• 5. Calculate the Mahalanobis distance based on Ps ^{i, rc,rd} (x) and the selected X as the B(x).

• 6. Set a multiplier$m\in \left| 0,1 \right|$.

• 7. Run a pre-matching process to identify h based on the distance distribution of the nearest neighbors in each matching cluster: a) Select a treated observation i. b) Use the nearest neighbor in terms of Mahalanobis distance given that j lies within the cluster cl, except for the distances between the comparisons. c) Extract the 75th percentile of all distance values within cluster cl. d) Use the 90th percentile across all ‘cl p75-distance values’ as the bandwidth h.

• 8. Run the clustered matching algorithm based on h taken from (7), which is multiplied by m.

• → if the balancing property is not sufficient, re-run from (7) based on additional attributes that are added to the calculation of the Mahalanobis distance.

• → if balancing is not sufficient based on the addition of attributes, rerun from (6) with a smaller multiplier.

Note that we calculated the standard errors (SE) of the estimator in (4) following Lechner (2001):

$$SE(ATT)=\frac{1}{{{N}_{i}}}Var(Y_{i}^{D=1})+\frac{\sum\limits_{n=1}^{{{N}_{j}}}{W_{n}^{2}}}{(N_{i}^{2})}Var(Y_{j}^{D=0})$$

(4)

In the calculation based on Formula (4), we implicitly assumed that individuals (treated and matched untreated) were independent, thereby emphasizing the issue of regional clustering (the non-independence of observations within a regional entity), as reported in Section 4. We also calculated two measures that provided information about the potential misspecification of the standard error. The first measure was a design-effect indicator (denoted by ‘se ^r /se, I’) that focused on the ratio of the two standard errors taken from the non-weighted and unrestricted sample of the treatment effect estimation and based on a simple logit model with (se^r) and without robust standard errors (se). High values indicated a strong correlation between observations and, therefore, a high risk of misspecification of the common variance estimation. The second measure followed the same logic and was calculated as a ratio (denoted by ‘se ^r /se, II’). However, the ratio focused on the weighted and restricted population (matched sample). Nevertheless, using such indicators is not common in the context of evaluation, and they can only be considered rough indications of the potential effect of regional clustering.

6 Results

6.1 Results for the main groups

Table 2 reports statistics related to the treatment effect.³¹ In particular, we focused on the average treatment effect on the treated (ATT) measured in accordance with Formula (a) (Y¹: exits within the first three years) and the subsequent inference statistics. Note that for the interpretation of the ATT (Y¹), a positive sign is associated with a higher failure rate of the treated compared with those who only received the bridging allowance, thus indicating a negative effect of the treatment on the likelihood of remaining self-employed. As column five shows, the ATTs (Y¹) were relatively low and, in most cases, remained statistically non-significant. In empirical terms, this finding indicates that (on average) the additional support did not contribute to increasing the duration of self-employment. However, our findings indicate that it is important to disentangle the different reasons for exiting self-employment because the treatment effects significantly differed for exits into employment positions vs. exits into unemployment.

Table 2 Treatment effects (Mahalanobis-Distance-Kernel-Matching). (Source: IEB, own calculations)

For instance, in the case of training for self-employment, statistically significant effects could be identified only for exits into unemployment, which indicates that this form of support is associated with an increase in exiting self-employment if one focuses on exits into unemployment. In contrast, coaching significantly reduced exits into dependent employment (level of statistical significance: 95 %), meaning that business founders who received coaching were less likely to enter dependent employment when they exited self-employment. Furthermore, focusing on the DSUS, we found that exits into employment were less likely, whereas exits into unemployment increased.

With respect to regional clustering, the indicator for the design effect (‘se ^r /se, I’) showed a potentially high correlation in observations within regions. However, focusing on the ‘se ^r /se, the II’ ratio suggests that the matching procedure solved the problem to some extent. Furthermore, despite some statistically significant treatment effects, Table 2 shows that the magnitude of the identified treatment effects remained rather small. For example, a statistically significant difference of 0.021 in exit rates between the treated and matched untreated (see DSUS; exits into unemployment) means that including external expertise increased exit probabilities by no more than 2.1 percentage points over a period of three years. This result seems less likely to be of economic importance. However, compared with the baseline exit rate (see the ‘EXIT-column’), this shift indicated an increase of 6.7 %, which should at least be considered serious as seen from a relative perspective. Similar results were found for coaching (− 1.3 % points but − 17 % compared with the baseline exit probability) and DSUS (− 1.1 % points but − 10 % compared with baseline) when we focused on exits into an employment position. Similarly, trainings increased exits into unemployment by nearly − 7 % compared with the baseline exit rate.

However, when assessing this finding, we must be aware that there could have been different reasons for the low treatment effects, namely, time-variant effects, heterogeneous treatment effects, and methodical misspecifications. Considering these patterns suggests that existing effects may otherwise be averaged out. We focus on these issues below.

6.2 Time-variant treatment effects

To reveal time-dependent differences in survival (Rotger et al. 2012), Fig. 3 displays the treatment effect as the difference in the non-parametric survival functions between the treated and (weighted) untreated comparison groups. This finding provides information on the net outcome of promoting external expertise in terms of better survival chances over time. Note that the ATT at this point focused on Y² (time-dependent survival rate; see Formula (b)), and thus, a negative value reflected a lower survival chance (higher likelihood of exiting) in the treated population compared with those with no treatment. Again, the results are reported for different types of exits (all types, only exits into unemployment, and only exits into employment). Considering right-censoring, the survival functions were calculated as the proportion of observations (self-employed at time t) in relation to the pool of individuals who were still at risk. Confidence intervals (dashed lines) of 95 % were calculated using Greenwood’s (1926) approximation of standard errors (without controlling for clustering).

Fig. 3

Time-dependent treatment effects.

As Fig. 3 shows, time-dependent effects existed for all types of additional support. However, the extent to which variation occurred over time differed across the type of support and for the type of exit considered. For example—putting statistical significance aside—with regard to training, the findings revealed that the use of external expertise was associated in comparative terms with lower survival chances when we focused on all types of exits during the first 24 months. However, after 24 months, we observed that the survival difference between the treated and matched non-treated groups was almost zero. Additionally, there were relatively constant differences in survival for exits into unemployment and those into dependent employment. In general, there was little evidence that any benefit resulting from additional support increased with time when focusing on exits in general or exits into unemployment in particular (the opposite applies for exits into employment). Nevertheless, it is worth noting that those who experienced ‘trained’ periods of self-employment, in particular, had lower business survival rates immediately after the end of the bridging allowance and that entrants with coaching tended to have higher survival rates at this point in time. In particular, we found a lower survival rate in the ‘trained’ population, which indicates strong post-entry selection. A similar pattern was also found for the coaching population. However, this finding was less pronounced, which may indicate that the additional support increased the perception of self-employment as being an inferior option to employment.

6.3 Heterogeneous treatment effects

For plausible reasons, the effects caused by the treatment may have differed for specific subpopulations. As research on self-employment shows, outcome differences are likely to emerge between genders because they are associated with differences in risk attributes, investment behaviors, income, and growth intentions (Williams 2000; Georgellis and Wall 2005; Wagner 2007). Following this idea, we controlled for gender differences and differences between eastern as well as western Germany and stratified the population based on a generalized propensity score (five groups according to the 20th percentile). However, the findings did not differ substantially from the results for the whole population (see Tables 6, 7 and 8 in the Appendix). In most cases, we were unable to identify significant effects, except for the DSUS, for which we found the highest treatment effect for the subgroup with low treatment dispositions and increased exits into unemployment for those who had received additional support (ATT = 0.065; se = 0.027).

6.4 Common support and matching quality

To assess the quality of the matching procedures, we examined the joint distributions of the propensity scores for the treated and non-treated groups. According to the graphical assessment in Fig. 5 (in the Appendix), the included matched comparisons were sufficiently balanced. Furthermore, in accordance with Rosenbaum and Rubin (1983 and 1985), we used the mean standardized bias (MSB) as an indicator for the overall balance of the matched comparisons.³² As reported in Table 2, the average MSB decreased strongly after the matching procedure. This is a fairly good indication of a sufficiently good balance. In fact, this was a better balance than that found in other related studies (e.g., Baumgaertner and Caliendo 2008). Finally, the F-test statistic revealed the joint insignificance of the covariates in a logistic regression in the matched sample.³³ Similarly, t-tests provided support for rejecting the mean differences between the matched treated and non-treated on the observed attributes.

6.5 Additional findings and robustness checks

The most critical objection in this evaluation might refer to the point that individuals with unpromising business projects may expect greater returns form using additional support programs and are therefore more likely to take advantage of these promotions. Because this might go unobserved, matching may fail to estimate unbiased treatment effects. However, various checks were conducted to assess the robustness of the estimates. First, we performed different matching procedures, including single nearest neighbors, caliper matching, and propensity kernel matching, to check methodical issues. On the whole, the results of the procedures supported the reported findings. We also tested the potential effect of unobserved heterogeneity by explicitly excluding information and calculating post-estimation Rosenbaum bounds.³⁴ In particular, neither of the sensitivity tests supported the hypothesis that unobserved heterogeneity had affected the reported estimates. In addition, we re-ran the estimates and included only regions with high ratios of external expertise support programs to consider potential interference from ‘negative creaming’ (assuming that negative selection would be relatively higher in regions with only a small number of participants). Finally, we replicated estimations while focusing on regions with low levels of comparable self-employment promotion activities that were covered by state-specific ESF-funding to test for the effects of potential substitutes and to address the potential conflict with the SUTVA (see above).³⁵ Overall, none of the robustness checks revealed substantial differences from the findings reported above.

7 Summary and conclusions

In this study, we examined the treatment effects of additional start-up support in terms of individual self-employment stability over a period of five years. We evaluated the outcome of including external expertise by studying the nationwide promotion of programs that complement a financial support program. The programs we studied are part of the nationwide active labor market policy in Germany. From a conceptual perspective, we studied three different programs that promote the inclusion of external expertise in the venture process and that complement a basic financial support program. Here, we question whether the individual self-employment position is more sustainable when external expertise is involved.

In our analysis, we used data from the Integrated Employment Biographies (IEB), which is an integrated German database that makes it possible to examine all cases of participation in employment and training programs that are offered by the Federal Employment Agency. The data allowed us to control for detailed information about employment history and qualification levels as well as socio-demographic information. In addition, rich regional data on local labor market conditions and on local labor market policies could be controlled for in the evaluation context, which made the statistical matching approach a valid evaluation technique. The information collected at the individual level allowed us to capture qualifications and biographical information, which we related to the quality of the business project. Furthermore, the population itself and our data restrictions allowed us to focus on relatively homogeneously treated and untreated individuals, which improved our matching the Conditional Indepence Assumption (CIA). An important issue in this context is that the treated and control participants received bridging allowances, which ensured the homogeneity of their intentions to start businesses.

With respect to treatment selection, we found that the likelihood of participating in an additional support program was mainly the result of differences in local active labor market policy strategies across Germany. Note that this finding extends earlier research that emphasized the dominant role of individual factors in activities related to creating a venture (e.g., Santarelli and Vivarelli 2007). However, it provides support for the finding of low individual heterogeneity among the treated and the controls. In particular, the results show that few regions had large shares of additional support and that in most regions, using the external expertise promotion for self-employment was less attractive for previously unemployed business founders. This finding indicates a particular regional specialization in the promotion of self-employment that has not been addressed in previous evaluation studies. Nevertheless, this finding gives extra support to implement a weighting program in the matching procedure that includes specific regional context information.

The evaluation showed that support for including external expertise for self-employment tended to increase exits over a period of three years. We also found support for the time-dependent effects of the promotion (e.g., Rotger et al. 2012). However, statistical significance was limited for all programs and outcome measures. For example, coaching only showed relevance for (fewer) exits into dependent employment (17 %), and significant effects for training were limited to (increased) exits into unemployment (7 %). Statistically significant treatment effects concentrated on DSUS, following which exits into unemployment increased by 6.7 % and exits into dependent employment decreased by 10 %.

Our findings indicate that on average, training and coaching did not correspond to the intentions of the relevant policies. If individual ‘learning’ were improved because of the additional support, we would have expected the sustainability of the self-employment period to be higher or for exits into wage-earning positions to be accelerated. However, we found significantly lower exit rates into employment positions related to training and DSUS and significantly higher exit rates into unemployment related to DSUS. These findings are interesting for at least two reasons. First, they indicate that the promotion with the largest degrees of freedom was associated with a number of non-ignorable treatment effects. And, second, because the treatments were related to increased exits into unemployment (DSUS) and reduced exits into wage work (training and DSUS), ‘passive learning’ (Jovanovic 1982) may have been dominant in our study context. The only interpretation we see for this effect is that external expertise may have tended to improve the ability to critically assess the business’s future economic prospects (e.g., to avoid running into debt). This would provide support for our theoretical framework and for our determination that it is highly important to disentangle active and passive business assessments related to the inclusion of external expertise.

Compared with earlier research, our findings related to insignificant treatment effects support the results reported by Karlan and Valdvia (2011), who evaluated a support program in Peru in an experimental setting (see also Shutt and Sutherland 2003 or Eckl et al. 2009 for similar evidence in other countries). This would imply that providing extra support that includes external expertise is not a good idea. However, this general statement would be in conflict with the findings of Chandler and Hanks (1998) and of Cressy (1996), who reported that human capital is even more important for self-employment than is financial capital. Similarly, Michaelides and Benus (2012) reported positive outcomes for previously unemployed business founders in the U.S. Nevertheless, our research is the first study that concentrates on a nationwide promotion setting. A major difference in this context may have been that the governing of sufficient quality standards—such as those of the GATE program or those in the SBDC context—was not appropriately met in our study population. In addition, additional differences may exist because the policy context that our research related to consisted of few access barriers.

Finally, it should be noted that our findings should be treated as preliminary because our evaluation must address a number of data limitations. First, one should be aware that relevant regional characteristics as well as individual attributes may have existed that we were not able to control for. For example, in the regional context, we only focused on labor market characteristics. Hence, the level of market competition was not observed. Instead, we assumed that all relevant information was fully approximated by controlling for labor market issues. Furthermore, at the individual level, our findings were based on the assumption that the untreated business founders did not use similar promotion programs, which we were not able to observe. We only used rough approximations to control for this issue. Furthermore, it should be noted that we were not able to study the quality of the interventions or the quality of the business projects. Again, in both cases, we assumed that the quality was correlated with our observed attributes. Second, our findings were only based on a simple outcome measure that focused on the sustainability of self-employment. Although we were able to disentangle different post-exit positions, it should be noted that our definition only allowed for identifying post-entry positions that were observable in the data. Nevertheless, we think that the study of alternative outcome measures is necessary to uncover the mechanisms related to the inclusion of outside assistance.

Appendices

Appendix

Table 3 Self-employment promotion schemes of active labor market policy

Fig. 4

Definition of valid additional support

Table 4 Definition of the variables

To focus only on valid additional support the study uses time corridors as displayed in Fig. 4. Following this figure training, coaching and discretionary start-up support are included only if they lie within a certain time corridor in relation to the start date of the bridging allowance. For simplicity the figure displays the number of the episode (I to IV) and the type of period (b = before the start of bridging allowance; a = after the start) as well as the time corridor to define the cut-off-point for including or excluding observations. For the training population this includes observations where the end date of the training lies within a corridor of one year before the start date of the bridging allowance (entry date) and end dates that are no later than 3 weeks after the entry date (training courses must have started before the entry date). Hence only episodes I and II are included (see the left part of the figure). For coaching we set a time window between − 21 days and + 365 days related to the start of the coaching and coaching must begin no later than one year after the bridging allowance. Therefore, only episode III is included (see right part of the graph). For discretionary start-up support the time corridor is + /- one year which means that episodes I, II and III are included (see right part of the figure).

Listing A.1: Description of the matching algorithm

Labor market districts (regions) that have no participants in the support schemes evaluated do not enter the study, because the joint support in these regions is zero (step 2). To include statistical information about the assignment process, linear predictions of logit estimates are used. To emphasize the distinct levels of selection three separate scores are applied. The first one includes only individual characteristics, the second is based on regional attributes and time intervals, the third is based on dummy variables of the labor market district and time (step 3).

Before calculating M(x), the sample of i and j is stratified according to regional clusters and time intervals (step 4). Region (type of region) and time (13 annual quarters) define distinct matching clusters to ensure that comparisons are only taken from observations with the most similar external economic market conditions and to take into account the dominating pattern of the treatment assignment. The exact date of entry, the scores and the interaction between the scores are added to the list of attributes that enter the calculation of the Mahalanobis distance (step 5). The inclusion of further variables depends on step 8.

A pre-within sample matching procedure is then performed to identify a bandwidth parameter (step 7). This step ensures that h is drawn directly from the clustered sample. Nearest neighbor matching guarantees that only the closest j are used. Based on the realized n ^{i, j} distances the 75th percentile is taken as the cluster k specific bandwidth (h ^k). This avoids potential high distance matches within a cluster. Next, the 90th percentile of all h ^k is used as the overall bandwidth parameter h. This procedure is implemented to weight matches in clusters in which only high distance matches exist.

The final matching procedure is performed based on h and proceeds in step 8. The multiplier defined in step 6 ensures a flexible adjustment of the bandwidth parameter and is only rescaled if the inclusion of additional variables (or interaction terms) does not improve the balance between i and j after matching.

Table 5 Treatment selection (estimated effects). (Source: IEB, own calculations)

Table 6 Treatment effects for ‘training’ (sub groups and robustness checks)

Table 7 Treatment effects for ‘coaching’ (sub groups and robustness checks)

Table 8 Treatment effects for ‘discretionary start-up support’ (sub groups and robustness checks)

Fig. 5

Support-Overlap. (Source: IEB, own calculations)