A replication of ‘Entry regulation and entrepreneurship: a natural experiment in German craftsmanship’

Abstract

Rostam-Afschar (Empir Econ 47:1067–1101, 2014) analyzes the impact of the deregulation of the German Trade and Crafts Code of 2004 on entrepreneurial activity, using German microcensus (MC) data. He finds a uniform positive effect on market entry in partially and fully deregulated trades and no change in exit probabilities. We replicate and extend this study. Most importantly, we generate a novel classification scheme that aims to achieve an improved identification of crafts trades in the microcensus. It is necessary to remove non-craftsmen from the analysis as the policy change exclusively pertains to the crafts sector. In contrast to Rostam-Afschar’s findings, the increase in self-employment and entry is more pronounced in the completely deregulated B1-trades rather than the partially deregulated A-trades. In addition, exit probabilities in fully deregulated trades do not remain constant but rather increase.

Appendices

Appendix A: Classification of crafts trades

The following procedure was used to identify individuals working in the crafts sector by using the microcensus occupation codes (KldB1992). In a first step, we gathered information on all training occupations and their classification codes (KldB 1992). Training occupations are different from occupations but are nevertheless associated with a particular crafts trade. This was achieved by consulting the official classifications of the ZDH and the Federal Institute for Vocational Education and Training and including present as well as predecessor occupations (Bundesinstitut für Berufsbildung, BiBB 2012).

In a second step, we used data provided online by BiBB concerning the information about how many apprentices within one occupational field are trained within either crafts or non-crafts (mainly industrial) companies.⁸ We subsequently computed a proportion of crafts apprentices within each occupational code. To exclude occupation codes with a high proportion of non-crafts workers, we used the information on the proportion of crafts trainees and dropped codes if this proportion was less than 60%. Lowering or increasing this cut-off point by up to 20% hardly affects the classification as most occupations contain either a very low or a high proportion of craftsmen. We also removed observations if occupations could not be clearly marked as either an A or B occupation.

This method is not error-proof as it assumes that the proportion of crafts trainees strongly correlates with the proportion of crafts employees. However, this method allows us to remove some of the occupation codes from the analysis that most probably contain very low proportions of crafts workers. For example, while the KldB code 141 (“Chemiebetriebswerker”, chemical plant employee) may seem a good proxy for the B-trade of “Wachszieher” (candle maker, see RA classification), according to our results less than 1% of individuals in the occupation of chemical plant employee are actually craftsmen. Our classification scheme implies that most of the individuals in that occupation are industrial workers such as chemical production specialists, chemical technicians or pharmaceutical technicians.

In a final step, we scrutinized the occupation of building cleaners (KldB code: 934). The occupation comprises about 45% of all individuals in the deregulated B-trades in the microcensus dataset. Owing to its large size, it potentially biases any general conclusions about B-trades.

After a thorough inspection, we are doubtful that the occupational group of cleaners in the microcensus data reasonably captures the TCC trade of cleaners. While official company registration data points to a sharp increase in market entry in that trade after 2004, no such trend can be established in the microcensus data. The proportion of self-employed cleaners in the microcensus only increases from 1.6% (2004) to 2.3% (2011). Upon request, employees of the Research Data Centers of the German States confirmed our suspicion and suggested several other classification codes under which cleaners might be found, none of which can be identified as crafts trades based upon our classification scheme.

According to the documentation for an older occupation classification system (KldB1975), there are about seven activity profiles coded as 933 or 934 (cleaners). The classification scheme in the microcensus (KldB1992) merges these codes into one (934). According to the crafts classification scheme recently developed by the Federal Employment Agency (BAA, 2014⁹), only three of these seven occupations belong to the crafts sector.

Table 9 presents a comparison of the RA and the Runst et al. samples. According to our classification scheme, about 97,000 observations in the RA sample are in fact not crafts occupations or cannot be clearly identified as group A or B and thus must be dropped from the analysis. Furthermore, there are about 45,000 observations that we included but are not part of the RA sample.

Table 9 Comparison of samples based on different crafts classifications

Appendix B: Construction of entry and exit variables

There are two possible approaches to constructing dummies that indicate market entry. RA uses a 45% subsample of the microcensus, whereby he compares the employment status of the previous year and the year of the survey. Alternatively, one may rely on information about the start of current employment, which is part of the mandatory section of the questionnaire. If the starting date for self-employment coincides with the year of the survey, it is coded as a market entry. As the questionnaire is completed around March of each year, market entry during the summer, fall and winter is not recorded in this way. We report the regression results for both of these variables.

The exit variable is constructed as described by RA (2014), based on the non-mandatory question about the employment status in the previous year.

Appendix C: Mediating variables

Mediation refers to a causal chain when a variable A affects the mediating variable B, which in turn affects variable C. At the same time, A can also cause C directly. The concept was developed in psychology (Baron and Kenny 1986; Judd and Kenny 1981; MacKinnon et al. 2007) but has recently also been applied to econometric analyses (see Heckman and Pinto 2015).

In the context of our paper, the reform (A) lowers the educational credentials of market entrants (B), which in turn increases the exit probabilities in the market (C). At the same time, the reform is hypothesized to directly increase exit as the level of competition is higher than prior to 2004.

In order to explore mediation pathways, Baron and Kenny (1986) suggest performing four regressions in which each component of the causal chain is examined separately. The first regression does not include mediation variables (education), i.e., the direct channel from reform to exit probability. The second and third regressions follow the mediation channel (reform to education, education to exit probability). Finally, if the first three regressions have established significant relationships, the fourth model uses all variables.

The regression results without education controls can be found in Table 4, while the regression results for steps two and four can be found in Table 10. There is evidence of the existence of a mediation channel. The relationship between the reform and education is negative. Education and exit probabilities are also negatively related. The negative reform effect on exit probabilities holds regardless of whether we control for education or not.

Table 10 Testing for Mediation

Appendix D: Sensitivity analysis

In specification (1), German Microcensus data for the years 2002–2009 has been used.

The variable ‘exit’ is constructed as described by RA.

The following control variables are used: Secondary education (POS, Realschule, Fach-Abitur, Abitur), tertiary education (dual vocational training, school-based vocational training, master craftsmen, university for applied sciences, university degree, Ph.D.), age, age squared and cubed, gender, citizenship, state dummies, city size dummies, marital status, no. of children, year dummies, branch and occupation controls (Table 6).

In specification (2), ZDH panel-data for the years 1998–2015 has been used. The total number of firms within each trade is used as frequency weights. P values are displayed in parentheses.