Naturalisation and on-the-job training participation. of first-generation immigrants in Germany

Transcription

1 Naturalisation and on-the-job training participation of first-generation immigrants in Germany Friederike von Haaren * NIW Hannover and Leibniz Universität Hannover This version: January 31 st, Preliminary - Abstract This paper examines the relation between naturalisation and on-the-job training (OJT) participation among first-generation immigrants in Germany. Since OJT is associated with improved labour market outcomes it is an important measure for labour market integration. Naturalisation may act as a signal, because it exhibits the employee s commitment to the host country and this may increase employers likelihood to invest in employee s human capital. Results of pooled linear probability models estimated on the basis of the German Socio Economic Panel show a positive and significant correlation between citizenship acquisition and OJT participation. However, estimations based on propensity score matching indicate that this correlation is due to self-selection. Results obtained by individual fixed-effects and instrumental variable models support this conclusion. JEL: classification: J15; J24 Keywords: citizenship; integration; human capital * Niedersächsisches Institut für Wirtschaftsforschung (NIW), Königstr. 53, D Hannover, vonhaaren@niw.de, telephone: , fax:

2 1. Introduction This paper examines the relation between on-the-job training (OJT) participation and naturalisation among first-generation immigrants in Germany. The aim is to assess whether the event of naturalisation increases the participation of immigrants in OJT. OJT is characterised as job-related training that takes place during working hours and is financed by the employer. Since OJT is essential for post-school and firm-specific human capital formation it is an important aspect of labour market integration. Naturalisation, which means gaining the citizenship of the new home country, is closely connected to integration and might even be regarded as a proxy for unobservable variables related to integration like identification with the host society or the probability to stay in the country. There are two alternative hypotheses that might explain why naturalised immigrants participate more often in OJT measures than non-naturalised immigrants. First, naturalisation may act as a signal which exhibits the workers commitment to the host country. Hence, the expected length of employment might be higher for naturalised than for non-naturalised immigrants. Consequently, employers might be more willing to invest in human capital of naturalised workers. Second, differences in OJT participation might be due to systematic differences in other characteristics between naturalised and non-naturalised immigrants. Thus, the question is whether naturalisation has a causal effect on OJT participation or is a confounder. This is important for choosing adequate policy measures that enhance labour market integration and ensure the supply of skilled labour. If naturalisation has a causal effect on OJT, adjusting naturalisation laws could contribute to improve immigrants labour market outcomes. Furthermore, this paper adds to the literature on the economic consequences of naturalisation. While it is known that gaining citizenship is associated to improved employment probabilities and higher wages (e.g. Bevelander and Pendakur 2012, Bratsberg et al. 2002), the reasons why naturalised immigrants achieve better labour market outcomes are not well understood, yet. In exploring a possible channel that may explain why naturalisation relates to labour market success, namely the effect of naturalisation on the probability to participate in OJT, this paper sheds more 2

3 light on this question. Since OJT is strongly correlated with firm-specific human capital and associated with higher wages (e.g. Parent 1999) and other favourable labour market outcomes like promotions (Pfeifer et al. 2013), higher participation rates of naturalised immigrants in OJT may be one of the reasons why naturalised individuals have more favourable labour market outcomes compared to non-naturalised immigrants. Up to this point, the literature has not addressed the relation between naturalisation and OJT explicitly. Based on data of the German Socio Economic Panel (GSOEP), I estimate pooled linear probability models for a sample of first-generation immigrants aged between 25 and 55 years. Only 0.9% of nonnaturalised immigrants participate in OJT. Results indicate that naturalisation is associated with an increase of more than 150% in the OJT participation probability, controlling for a range of personal, migration and job characteristics. In order to reduce selection bias on observables, I apply propensity score matching and find that naturalisation has no significant effect on OJT participation. This indicates that the passport does not facilitate access to training. The observed higher share of OJT participation among naturalised immigrants rather seems to be due to positive self-selection. Results obtained by individual fixed-effects and instrumental variable models support this conclusion. The paper is organised as follows. The next section explains the relation between naturalisation and OJT. Section two describes the data and defines OJT and other important variables. Furthermore, basic descriptive statistics are illustrated. Section four specifies the empirical strategy. Results and robustness checks are discussed in section five. The last section concludes. 2. The relation between naturalisation and OJT participation So far, the effect of naturalisation on OJT participation has not been analysed explicitly before. However, there is a small strand of literature, examining the differences in participation rates of immigrants and natives. Results show that immigrants are less likely to participate in training than natives (e.g. Lochhead 2002, Hum and Simpson 2003, VandenHeuvel and Wooden 1997). Possible 3

4 reasons for this finding are discrimination (Oosterbeek 1998, VandenHeuvel and Wooden 1997) and concerns about the profitability of training. On the one hand, OJT might be less efficient for immigrants than for natives due to language problems (VandenHeuvel and Wooden 1997). On the other hand, immigrants are presumed to have a weaker attachment to the labour market (Oosterbeek 1998). The expected returns on investments in the employees human capital depend on the expected length of employment. This may explain, why individuals who are assumed to be less attached to the labour market and thus have shorter (expected) employment spells, like parttime employees, individuals with a temporary work contract or women, are found to receive less training (e.g. Arulampalam et al. 2004, Lynch 1992 and Pischke 2001). The naturalisation status, however, is not taken into account in these studies. Only Park (2011) distinguishes between naturalised and non-naturalised immigrants when comparing predicted probabilities of OJT participation between immigrants and natives in Canada. To account for diversity of immigrants, the author controls for various personal and job characteristics. He exhibits that the difference in the training probabilities is larger between Canadians and non-citizens than between Canadians and naturalised citizens. In particular, the predicted odds ratio of participation in employer-supported job training between native and non-naturalised immigrant men is On the contrary, between natives and naturalised immigrants it is 0.87 and the difference is not significant. According to Park (2011), naturalised immigrants may be better prepared to participate in OJT, because naturalisation ensures a particular duration of residence and a certain level of language proficiency. Since Park (2011) does, however, not account for self-selection, the causal effect of gaining citizenship on OJT participation remains presently unclear. In contrast to Park (2011), this paper does not compare OJT participation rates between natives and immigrants but focuses explicitly on the effect of naturalisation on OJT participation among firstgeneration immigrants and tries to attenuate the selection bias. Moreover, the aforementioned studies are mostly based on data for North America (mainly Canada) and Australia. However, results might be different for European countries, because the structure of immigrants to European 4

5 countries and in particular to Germany differs from Canada. While a large part of immigrants to Canada is highly skilled (50%), only 20% of immigrants to Germany are (OECD 2011). Thus, OJT participation of immigrants in Germany needs further examination. Figure 1 shows the basic descriptive relation between OJT participation and naturalisation for Germany on basis of data from the GSOEP. While 3.3% of naturalised immigrants participate in OJT, only 0.9% of non-naturalised do. In addition, Figure 1 indicates the share of OJT participation among the entire sample of first-generation immigrants and natives. Evidently, the share of participants is higher among natives (5.9%) compared to first-generation immigrants (1.2%). Thus, the results are in line with existing descriptive findings from Canada (Lochhead 2002, Hum and Simpson 2003, Park 2011). Figure 1: Share of OJT participation among natives and first-generation immigrants 7% 6% 5.9% 5% 4% 3% 3.3% 2% 1% 1.2% 0.9% 0% Natives First-generation immigrants (total) Non-naturalised first-generation immigrants Naturalised (t-1) first-generation immigrants Source: own calculations based on data from the GSOEP. As stated above, two alternative hypotheses might explain the difference between naturalised and non-naturalised immigrants. Hypothesis 1: Naturalised immigrants participate more often in OJT because naturalisation may act as a signal which exhibits the workers commitment to the host country (signalling effect). A possible explanation why naturalised immigrants tend to participate more often in OJT than their non-naturalised counterparts could be a signalling effect. The acquisition of citizenship may increase 5

6 employers investment in workers human capital, because naturalisation may act as a signal by exhibiting the workers commitment to the host country. Employers might perceive this also as commitment to the firm and, thus, expect that naturalised employees will be longer employed in the firm than non-naturalised. Since empirical findings of the literature support the human capital theory, which assumes that the decision to invest in employees human capital depends amongst others on the expected length of time in which the revenues can be obtained (e.g. Arulampalam et al and Lochhead 2002), this could be the mechanism behind the correlation between naturalisation and OJT participation. On the other hand, observed differences in OJT participation might be due to systematic differences in other observable or unobservable characteristics between naturalised and non-naturalised immigrants. If unobservable factors influence naturalisation and OJT participation at the same time, the naturalisation variable is endogenous. Hypothesis 2: OJT participation differs between naturalised and non-naturalised immigrants due to systematic differences in other characteristics between the two groups (self-selection). Aim of the paper is to find out which hypothesis can be corroborated. 3. Data and descriptive statistics 3.1. Data and sample restrictions The relation between naturalisation and OJT participation among first-generation immigrants is examined on the basis of the German Socio-Economic Panel (GSOEP). Since 1984 nearly 12,000 households are interviewed each year and asked a variety of questions. Amongst others, the data set contains detailed information on training attendance and migration characteristics, the key variables for the analyses. An advantage of the GSOEP is the overrepresentation of immigrants that increases the sample size (Wagner et al. 2007). 6

7 The definition of OJT is crucial, because the literature shows that the effects of OJT and also the influence of different determinants on training participation depend on this definition. For example, Park (2011) exhibits that participation differences between natives and immigrants are larger in employer-supported training. In accordance with the literature, I define OJT as participation in an occupational oriented course, that takes place during working hours, is organised and financed by the employer and lasts between one day and three months. 1 Information on training measures comes from retrospective questions referring to the past three years and is available for the years and Thus, the sample is restricted to these years. The explanatory variable of interest is naturalisation, which is approximated by using information on place of birth and nationality. Accordingly, foreign-born individuals with German citizenship are defined as naturalised. However, foreign-born individuals, who stated to have German citizenship since birth are not considered as naturalised. 3 Furthermore German citizens who lived abroad are excluded. Naturalisation of first-generation immigrants is associated with a certain level of integration, language proficiency and an increased probability to stay in the country. 4 It is hypothesised, that the acquisition of citizenship serves as a signal for these characteristics. This is why a positive influence of naturalisation on training participation is expected for first-generation immigrants. However, German-born children of immigrants (the second-generation) have better language proficiency (Haug 1 Theory often distinguishes between general and specific training. While general training enhances employees productivity in all firms, specific training is not portable. According to the theory, the firm would pay for specific training only when it is sure that the employee will not leave the firm after participating in training (Borjas 2008). However, empirically it has been shown that OJT is often a combination of general and specific training (e.g. Borjas 2008 or Parent 1999). In the GSOEP data it is only partly possible to assess whether the acquired skills in training measures would be useful in another job. One third of the acquired skills are not at all or only in a limited way portable. Due to a small sample size and data limitations it is not possible to apply this distinction in the analyses. 2 Questions on training are part of a special module of the questionnaire that was included in 1989, 1993, 2000, 2004 and Since the module was not part of the questionnaire in 1997, information on training is not available for the years This information is, however, only available since In the analysed sample 83% of naturalised immigrant judge their language proficiency as good, while only 55% of non-naturalised immigrants do. Constant and Massey (2002) show empirically that naturalised immigrants from the former recruitment countries are less likely to return to their home countries than non-naturalised immigrants in Germany. 7

8 2005) and a higher probability to stay in Germany than first-generation immigrants already without being naturalised (Tucci 2011), because they grew-up and were educated in Germany. Thus, naturalisation is a different signal for second-generation immigrants than for first-generation immigrants. Therefore, the sample does not include second-generation immigrants. Moreover, the so-called ethnic Germans (mostly repatriates from Eastern Europe and the former Soviet Union) differ in certain characteristics 5 and legal status from other first-generation immigrants. For example ethnic Germans do not have to meet standard naturalisation conditions and are naturalised shortly after arrival in Germany. Hence, in contrast to other immigrant groups, acquiring German citizenship is not an explicit decision for ethnic Germans, rather they are naturalised by definition (Worbs et al. 2013). Therefore, the effect of naturalisation is assumed to be different for them. For this reason, ethnic Germans are excluded from the analyses, as well. 6 Furthermore, only employed individuals aged between 25 and 55 years are considered, because training incidence is higher in the prime age group, than at the margins. Thus, there are 13,852 observations in the sample, 11.3% of all observations are naturalised Descriptive statistics Table A 1 shows means of important variables for the total sample of first-generation immigrants and separately by naturalisation status. While only 1% of non-naturalised immigrants participate in OJT, the share of naturalised immigrants participating in OJT is 3%. The means of almost all other observed variables differ significantly by citizenships as well (differences are significant to the 1% level according to the t-test and the Mann-Whitney-Test, Pearson s chi-squared test indicates that the distribution of the categorical variables origin, position in the job and firm size are not 5 Worbs et al. (2013) report for example lower shares of individuals without any educational degree and higher employment rates among ethnic Germans compared to all migrants living in Germany. Moreover, ethnic Germans have better language proficiency (Haug 2008) and a higher intention to stay in Germany compared to other immigrants (Tucci 2011). 6 The definition of ethnic Germans is described in the Methodological Appendix. 8

9 independent from naturalisation). Comparing for example the position in the job exposes that 67% of non-naturalised immigrants have a low position, while only 39% of naturalised immigrants are in this position. In addition, the share of people in a high position is substantially larger among naturalised immigrants compared to non-naturalised (13% versus 3%). At the same time, the position in the job is strongly correlated with OJT participation. The higher the position in the job, the more likely employees participate in OJT. Furthermore, non-naturalised immigrants are more often blue-collar workers (84%) than naturalised immigrants (53%). Although discrepancies by citizenship are smaller regarding the firm size, Pearson s chi-squared test indicates that the distribution of naturalisation status and firm size is not independent. Moreover, Table A 1 reveals that men are overrepresented in the sample: 64% of all observations are male. However, this does not imply that there are more male immigrants in Germany. Rather this disproportion is due to the sample construction and in particular to fact that less female immigrants are regularly employed. In order to account for cultural differences, immigrants are categorized into different origin groups according to their country of birth. Table A 1 shows that in the estimation sample, the largest group of immigrants comes from Western European countries (39%). 27% migrated from Eastern European countries to Germany and 30% from Turkey. 4% of the immigrants were born in other countries. 7, 8 Furthermore, Table A 2 describes the mean number of years since migration by origin and naturalisation status. The average duration of residence in Germany is quite similar for all groups. It varies between 22.8 years among immigrants from Western Europe and 19.0 years among immigrants from other countries. On average, the observed immigrants are living since 21.1 years in Germany. As expected, naturalised immigrants reside longer in Germany than non-naturalised. While 7 Although the group of immigrants from other countries is quite heterogeneous, the small sample size of this group (607 observations) does not allow a further distinction. 8 The share of European first-generation immigrants in the estimation sample is similar to the share in the overall population, the share of immigrants from Turkey in the overall population is only 18% (Statistisches Bundesamt 2013), thus, they are oversampled in the estimation sample. If Turkish immigrants naturalise less frequently, the true effect is underestimated in the estimation sample. 9

10 the average number of years since migration among non-naturalised immigrants is 20.6 years, it is 24.9 years among naturalised immigrants (Table A 1). The distinction into origin groups is not only important in order to control for cultural differences but also because immigrants source countries are closely related to the rights foreigners have in Germany. Due to differences in the legal status by immigrants origin, the incentives and thus the motives to naturalise differ by origin as well. On the one hand, immigrants from Western European countries are citizens of the European Union (EU) 9 and have thus, almost the same rights as German citizens. 10 That means that these immigrants have only small additional benefits from naturalisation. Therefore, it is not surprising, that only 7.7% of immigrants from Western European countries are naturalised (Table A 2). On the other hand, the naturalisation rate is highest for immigrants from Eastern Europe (51.8%). 11 Most of these immigrants come from non-eu member states or from countries that became member states only recently, like Poland or the Czech Republic in Thus, these immigrants have higher benefits from naturalisation than immigrants from Western European countries. The naturalisation rate for immigrants from Turkey is 17.8% in the estimation sample. 22.7% of immigrants from other countries gained German citizenship (Table A 2). Motives for naturalisation can be divided into emotional and instrumental reasons (Wunderlich 2005). Emotional or identificatory reasons are for example the sense of belonging to Germany, identification with Germany and the desire for political participation. Instrumental reasons include economic and pragmatic reasons that facilitate everyday life. Although I cannot observe the naturalisation motives in the data, it is known that instrumental reasons are particularly interesting for immigrants from non-eu member states, while immigrants from (the old ) EU member states 9 Most of the countries were already EU member states before Austria, Sweden and Finland joined the EU in Immigrants from countries of the European Free Trade Association (EFTA), like Switzerland, enjoy similar rights as immigrants from EU countries. 10 The major difference between German citizens and EU citizens is the lack of voting rights. Though, they are allowed to participate in local government elections. 11 Note that ethnic Germans are excluded from the sample. Although immigrants from Eastern Europe might partly be relatives to ethnic Germans, naturalisation is an explicit decision for them. 12 The Czech Republic, Estonia, Hungary, Latvia, Lithuania, Poland, Slovakia and Slovenia joined the EU in 2004, Bulgaria and Rumania followed in

11 tend to naturalise because of emotional reasons (Worbs 2008). Although immigrants from Turkey are not EU citizens, Sauer (2012) found that emotional ties to Germany are important motives to acquire German citizenship for them as well. Examining the duration between migration and naturalisation by origin supports this pattern. The duration between migration and naturalisation might serve as indicator for the relevance of emotional or identificatory motives. I assume that, on the one hand, immigrants who acquire German citizenship mainly because of instrumental reasons naturalise relatively shortly after arrival in Germany because they want to enjoy the rights of German citizenship as soon as possible. On the other hand, I expect that immigrants who naturalise mainly due to emotional reasons naturalise later, because identification and integration needs time. Furthermore, for immigrants who decide more emotionally, giving up the original citizenship might be more difficult. Table A 2 shows that the duration of residence before naturalisation differs clearly by origin, although the mean number of years since migration is quite homogeneous. Naturalised immigrants from Western European countries, who tend to acquire citizenship because of emotional reasons, have lived on average a very long time in Germany before they decided to acquire German citizenship (32.6 years). Immigrants from Turkey and immigrants from Eastern Europe live on average 23.1 years, respectively 21.5 years, in Germany before naturalisation, while immigrants from other countries gained citizenship after 15.2 years of residence in Germany. 13 In order to find out more about the factors that influence naturalisation, determinants of naturalisation are estimated for the total sample and for different origin groups (Table A 3). One of the relevant factors is the duration of residence in Germany. Living 10 years longer in Germany is associated with an increase of the naturalisation probability by 9 percentage points in the total estimation sample holding constant other characteristics. The relation seems to be weaker for immigrants from Western European countries and immigrants from Turkey compared to immigrants from Eastern European and other countries. This is in accordance with the previous discussion about instrumental and emotional naturalisation motives. The different naturalisation rates by origin are 13 This measure can only be calculated for those migrants, who changed citizenship status during the observation period. 11

12 reflected by the dummy variables for the immigrants origin. Immigrants from Eastern European countries are 20 percentage points more likely to acquire German citizenship than immigrants from Western European countries. Being born in Turkey is associated with an 11 percentage points higher naturalisation probability. Immigrants from other countries are 47 percentage points more likely to naturalise than immigrants from Western Europe. The most important determinant is being married to a German citizen, which can be regarded as proxy for integration. Immigrants who are married to a German have an almost 30 percentage points higher naturalisation probability than those who are married to a foreigner or not married. This is in line with other results indicating that having close German friends is strongly correlated with naturalisation (Zimmermann et al. 2009). For immigrants from Non-Western European countries this determinant seems to be even more important than for immigrants from Western European countries. Table A 1 illustrates that the share of women is larger among naturalised than among nonnaturalised immigrants. Estimation results confirm this observation: women are overall 2 percentage points more likely to naturalise than men when other characteristics are held constant. This is in accordance with other studies finding that naturalisation rates of women are larger than those of men in OECD countries (e.g. Liebig et al. 2010, Zimmermann et al. 2009, Yang 1994). According to Alvarez (1987), females are more vulnerable and may have an incentive to acquire citizenship in order to escape from disadvantageous situations like repressive marriages or jobs. Economic factors seem to have a stronger influence on naturalisation of immigrants from Non-Western countries. For them the position in the job is positively and blue-collar employment is negatively correlated with the naturalisation probability. These determinants are not significant in the sample of Western European immigrants. 12

13 4. Estimation strategy In order to analyse the relation between naturalisation and OJT participation, the following pooled linear probability model (LPM) is estimated. 14 ojt it 0 1natu it 1 2 persit 3 jobit 3 year u it it, (1) where ojt i denotes OJT participation of individual i in year t. The variable of interest is naturalisation. To ensure the correct order of cause and effect, the naturalisation status in the previous year is used (natu t-1 ) as explanatory variable. In addition, further control variables for personal and job characteristics are included. From the literature on OJT participation, it is known, that women tend to receive less training than men (e.g. Pischke 2001). Therefore the model controls for gender and is, furthermore, estimated separately by gender. Since Pfeifer et al. (2012) show that the training probability depends on age, variables for age and age squared are included, as well. In order to account for differences between immigrants, the model contains dummy variables for immigrant s origin (Eastern Europe, Turkey and other countries, Western Europe serves as reference category). Years since migration are included as an indicator for assimilation. Besides, the data allows controlling for various job characteristics. I include the position in the job 15, which is expected to be positively correlated with OJT participation, as already shown descriptively in section 3.2 and illustrated by Pischke (2001). Moreover, a common finding is that larger enterprises are more likely to provide training (e.g. Frazis et al. 2000, Lynch and Black 1998, Lochhead 2002), therefore, the model contains dummy variables for the firm size. According to the human capital theory, part-time employees would receive less training, because as they work fewer hours, returns are lower. Human capital theory also predicts that employees with a temporary work contract receive less training, because of a shorter amortisation duration. In line with the literature, I add tenure and tenure squared as well as a dummy variable for being a blue-collar worker. Furthermore, year fixed-effects 14 Logit models yield similar results. They are available upon request. 15 I decided to include the position in the job as explanatory variable instead of education because the position in the job explains a larger part of the variation in OJT than education. Including both characteristics would lead to problems due to multicollinearity. 13

14 are included. Since panel data is pooled across all years to estimate equation (1), the errors of individuals may be correlated over time. Therefore, standard errors are estimated heteroskedasticity-robust and clustered by individual. The significant differences in observed characteristics by naturalisation status illustrated in Table A 1 and the analysis of naturalisation determinants (Table A 3), suggest that naturalised immigrants are a selected group. Therefore, results of the pooled linear probability models might be biased. In order to reduce selection bias on observables, I apply propensity score matching interpreting naturalisation as treatment. The idea of matching is estimating the average treatment effect on the treated (ATT) by comparing the outcomes of treated and non-treated, who are similar to treated in their characteristics X (Caliendo and Kopeinig 2005). In this case, I compare OJT participation of naturalised immigrants to non-naturalised with similar characteristics. Identifying an adequate control group is crucial to obtain meaningful results. For this purpose so-called balancing scores (b(x)) can be used. These are functions of the characteristics X that balance treatment and control group so that both groups have the same conditional distribution of X given b(x) (Rosenbaum and Rubin 1983). The coarsest balancing score is the propensity score, which is in this case the probability to be naturalised given the observed characteristics X. That means the differences in the distribution of characteristics between naturalised and non-naturalised are balanced by the propensity score and differences in the outcome (OJT participation) can be attributed to treatment (naturalisation). However, to identify an unbiased treatment effect, two assumptions are required. First, selection has to rely only on observable characteristics X (conditional independence assumption). In that case OJT participation of naturalised and non-naturalised immigrants is independent of naturalisation conditional on observable characteristics X. (Rosenbaum and Rubin 1983). Second, the common support or overlap condition has to be applied. This ensures that individuals with the same characteristics X have a positive probability of being in both the treatment and control group (Heckman et al. 1999). The combination of these two assumptions guarantees that treatment is strongly ignorable. If both assumptions hold, the expected difference in the outcome between 14

15 treatment and control group is the unbiased average treatment effect (Rosenbaum and Rubin 1983), which can be obtained by the following three steps: Step one: Estimate the propensity score The propensity score (probability to be naturalised) is estimated by a probit model. Control variables that capture observable differences between treatment and control group are chosen based on results of the literature and descriptive statistics discussed in section 3.2. Included are personal characteristics (gender, age, age squared), migration specific characteristics (origin and duration of residence), socio-economic factors (position in the job and being a blue-collar employee), as well as dummy variables for time periods (reference: ). Moreover, since section 3.2 reveales that being married to a German citizen is strongly associated with the naturalisation probability, this variable is also included as proxy for integration. Due to missing values in this variable, the sample size is reduced to 13,204 observations. 13,144 observations are within the common support and 1,427 are treated. Step two: Match treatment and control group and estimate the treatment effect Based on the estimated propensity scores treatment and control group are matched. The average treatment effect on the treated is calculated by differencing the outcome of matched treatment and control group. Different matching algorithms can be applied (for an overview see for example Caliendo and Kopeinig 2005). I apply kernel matching using the Epanechnikov kernel in order to obtain matched pairs. The non-parametric kernel matching estimator creates the counterfactual outcome by using weighted averages of all individuals of the control group. Since kernel matching uses more information than other matching algorithms, lower variance is achieved (Caliendo and Kopeinig 2005). 15

16 Step three: Assess the matching quality The matching procedure aims at balancing the distribution of the relevant variables in treatment and control group. The matching quality can be assessed by comparing the situation before and after matching. Matching was successful when there are no differences between the two groups conditional on the propensity score (Claiendo and Kopeinig 2005). In particular, I test the matching quality by three indicators: the standardised bias, the two-sample t-test and the Pseudo-R squared. Although the matching analysis accounts for self-selection into naturalisation, this selection process is only conditioned on observable characteristics. If, however, the naturalisation decision depends also on unobservable characteristics, like motivation, which are not correlated with observed characteristics, the naturalisation status can be different for individuals with the same observed but different unobserved characteristics. Thus, the selection problem would not be solved by propensity score matching. Therefore, I estimate in addition individual fixed-effects and instrumental variable models to check the robustness of the results. 5. Results 5.1. Pooled linear probability model Table 1 shows the coefficients of the naturalisation status in the previous year for different specifications obtained from a pooled linear probability model estimated for the total sample and for different subsamples. The dependent variable is participation in OJT. In the model, referring to the total sample (first row), the coefficient of naturalisation (t-1) is significant in every specification. Naturalisation is associated with an increase in the probability to participate in OJT of 2.3 percentage points when personal characteristics (gender, age, age squared and origin) are held constant. After controlling additionally for years since migration (specification three), the coefficient decreases slightly to 2.1 percentage points. When dummy variables for the position in the job are added, the 16

17 coefficient decreases further to 1.5 percentage points. Given that only 0.9% of non-naturalised firstgeneration immigrants participate in OJT, an increase of 1.5 percentage points is equivalent to an increase of more than 150%, thus the probability to participate in OJT measures is two-and-a-half times higher for naturalised immigrants compared to non-naturalised immigrants. The naturalisation coefficient stays nearly constant when further job specific control variables like tenure, part-time employment and firm size are included (specification five). In specification six, which contains an additional dummy variable for being a blue-collar worker, the naturalisation coefficient equals 1.3 percentage points and is still significant at the 5% level. Results remain robust adding a dummy variable for having a temporary work contract (specification seven). The remaining control variables have the expected sign and are in line with the literature (they are displayed in Table A 5). In general, job characteristics, like the position in the job, seem to be more important than personal characteristics, like immigrants origin, for the probability to participate in OJT. In order to shed more light on the different role of naturalisation for men and women, the linear probability models are also estimated separately by gender. The results (second and third row in Table 1) are very similar to the results described above. However, as the share of OJT participation among non-naturalised women is even lower compared to non-naturalised men (0.7% vs. 1.0%, see also Table A 4), the percentage increase associated with naturalisation is higher for women (+214% vs %). This suggests that naturalised women are either strongly self-selected or that women benefit more from acquiring German citizenship than men, which is in line with the findings of Liebig et al. (2010). 16 Looking at the other determinants of OJT by gender (Table A 6), it becomes evident that the coefficient of part-time employment is very small (0.3 percentage points) and not statistically significant in the female sample, while it is highly significant in the male sample. For men, part-time employment is associated with a 1.5 percentage points lower OJT participation probability. This negative relation corresponds to the hypothesis that individuals with presumed weaker labour 16 In order to check whether the naturalisation premium differs by gender, I also estimate a model with an interaction term between naturalisation and gender. However, the coefficient is not significant. 17

18 market attachment (like immigrants) tend to receive less training. Since part-time employment is more common among women, the signal of working fewer hours is maybe different for men. Table 1: Pooled linear probability model, dependent variable: participation in OJT (1) (2) (3) (4) (5) (6) (7) No. of clusters Total 0.022*** 0.023*** 0.021*** 0.015** 0.016** 0.013** 0.019** (0.006) (0.007) (0.007) (0.006) (0.006) (0.006) (0.009) 2,068 Men 0.022** 0.023** 0.019** 0.015* 0.015* (0.009) (0.009) (0.009) (0.009) (0.009) (0.009) (0.012) 1,276 Women 0.023** 0.023** 0.024** * * (0.009) (0.010) (0.010) (0.009) (0.009) (0.009) (0.013) 792 Western Europe 0.042* 0.043* 0.047* * (0.024) (0.024) (0.025) (0.023) (0.023) (0.023) (0.031) 777 Non-Western Europe 0.021*** 0.022*** 0.018*** 0.012* 0.012** * (0.007) (0.007) (0.006) (0.006) (0.006) (0.006) (0.009) 1,291 Eastern Europe 0.018** 0.019** 0.015* (0.009) (0.009) (0.008) (0.008) (0.007) (0.008) (0.012) 551 Turkey (0.018) (0.018) (0.017) (0.017) (0.017) (0.017) (0.017) 626 Blue-collar 0.014** 0.016** 0.014* 0.014* 0.020* 1,752 (0.007) (0.008) (0.007) (0.007) (0.010) White-collar (0.011) (0.012) (0.013) (0.013) (0.016) Notes: Displayed are coefficients of the naturalisation status in the previous year obtained from linear probability models. The sample is restricted to first-generation immigrants aged 25 to 55. The outcome variable is "participation in on-the-job training". Reported standard errors in parentheses are robust and clustered by individual. * denotes statistical significance at the 10% level, ** at the 5% level and *** at the 1% level. Specification (1) controls only for naturalisation (t-1) and year effects. Specification (2) adds gender, age, age squared and dummy variables for origin. Specification (3) controls additionally for years since migration (measured in ten years). Specification (4) includes dummy variables for the position in the job. Further job characteristics are added in specification (5): tenure, tenure squared, dummy variables for part-time employment and firm size. A dummy variable for blue-collar employment is included in specification (6) and specification (7) controls for having a temporary work contract. Due to missings in that variable, the numbers of observations are slightly lower in specification (7). Table A 5 and Table A 6 in the appendix show details on coefficients of all explanatory variables. Source: German Socio-Economic Panel Dividing the sample into immigrants from Western European countries and Non-Western European countries shows that the naturalisation coefficient is significant in most specifications (Table 1, row four and five). Although the naturalisation coefficient is larger in the Western European sample, it is less significant compared to the sample of Non-Western European immigrants. Differences in legal status and naturalisation motives, described in section 3.2, are possible explanations for these 18

19 observations. On the one hand, immigrants from Non-Western countries are mainly Non-EU citizens who face stronger labour market barriers than immigrants from Western European countries (who are mainly EU citizens). 17 Thus, naturalisation might be more beneficial for Non-Western European immigrants. On the other hand, immigrants from Western European countries naturalise mainly because of emotional or identificatory reasons, therefore, the acquisition of citizenship might be an even stronger signal of commitment for them than for immigrants from Non-Western European countries, who are assumed to naturalise mainly due to instrumental reasons. Another reason for this hypothesis is the supposition that immigrants from Western European immigrants are more mobile than immigrants from Non-Western European countries. Due to the freedom of movement within the EU migrating is easier for them compared to immigrants from Non-EU countries, who have to undergo the visa process in order to migrate to Germany. This is in line with the results on return migration, indicating that immigrants from Non-Western European countries have a lower return migration probability than immigrants from Western European countries (Table A 11). Also Constant and Massey (2002) find in their analysis on return migration of immigrants from the former recruitment countries in Germany that those from EU countries are more likely to return to their home country than immigrants from Turkey or former Yugoslavia. This supports the hypothesis, that naturalisation is a stronger signal of commitment for immigrants from Western European countries. In addition, the group of immigrants from Non-Western European countries is subdivided into immigrants from Eastern European countries and immigrants from Turkey. The coefficients are similar, but mainly not significant, probably due to the smaller sample size. 18 Since OJT behaviour is more prevalent among white-collar employees compared to blue-collar employees (Table A 4), the models are estimated for these two subsamples as well. Naturalisation has a significant influence on OJT participation in each specification for blue-collar workers while in the white-collar sample the coefficient is not significant (Table 1, row six and seven). A reason might 17 Ethnic Germans are excluded. 18 The group of immigrants from Non-Western European countries contains also immigrants from other countries (Non-Eastern Europe and Non-Turkey). Since the number of clusters is even smaller in this subgroup (114), separate regression results are not shown. 19

20 be the small sample size of white-collar employees or the fact, that OJT is in general more prevalent in white-collar occupations. Besides, Table A 6 reveals, that the coefficient of years since residence in Germany is small and not significant for immigrants from Western European countries and white-collar employees, though it is larger and significant for immigrants from Non-Western European countries and blue-collar employees. This might suggest that these individuals need more time to adapt or to get along on the German labour market. Several sensitivity checks support the robustness of the correlation between naturalisation and OJT participation. Results of four alternative specifications are shown in Table A 7. In the original model, explanatory variables for the position in the job are included. However, not only the position in the job can influence the probability to participate in OJT, but OJT could also influence the position in the job (reverse causality). For example participating in a specific qualification measure might be a precondition for promotion. Thus, the position in the job might be endogenous. To overcome this endogeneity problem, column one controls for the position in the job in the previous year. The coefficients of these lagged explanatory variables are very similar to the non-lagged coefficients in the original specification and the naturalisation coefficient stays robust, as well. Therefore, reverse causality does not seem to be problematic here. Columns two additionally controls for (self-assessed) good language proficiency. 19 If language proficiency and naturalisation are correlated the naturalisation coefficient might incorporate this influence if the model does not control for language proficiency. Thus, results would suffer from omitted variable bias. In this case, the coefficient would decrease or loose significance when the omitted variables is included. However, results indicate that good language proficiency is only weakly associated with OJT participation (the point estimate is close to zero and statistically significant to 19 Information on language proficiency is not available in all years, therefore, it is not included in the baseline model. For the estimation in Table A 7 information on language proficiency is transferred to the following year. 20

21 the 10% level) and the coefficient of naturalisation stays significant. 20 The remaining coefficients and the R squared adjusted stay similar as well. In the baseline model the sample is restricted to immigrants aged 25 to 55 years. In order to check, whether the results are sensitive to this age restriction, the estimates of column three are based on a broader age restriction (20 to 65) and those of column four on a more restrictive boundary (30 to 50). However, in both specifications the naturalisation coefficients remain significantly positive and of similar size like in the original model. Furthermore, I estimated weighted linear probability models; the results are very similar to those of Table Since the dependent variable is binary, I also tested whether the results of non-linear models differ. However, average marginal effects obtained from logit models are very similar to the coefficients of the linear models Matching Average treatment effects on the treated obtained from kernel matching using the Epanechnikov kernel with a bandwidth of 0.06, are displayed in Table Treatment effects are calculated for the total sample and for subsamples of different origin groups. Results are not significant, indicating that the positive correlation between naturalisation and OJT participation shown in section 5.1 can be attributed to positive self-selection. 20 The magnitude of the point estimate even increases which is probably due to the change of the baseline category. 21 Results are available upon request. 22 Results are available upon request. 23 Estimation results of the propensity are available upon request. 21

22 Table 2: Matching results Total Western Europe Non-Western Europe Eastern Europe Turkey ATT SE (0.008) (0.033) (0.008) (0.014) (0.012) No. of observations 13,204 5,185 8,019 3,606 3,843 Notes: Displayed are average treatment effects on the treated (ATT) after propensity score matching with naturalisation as treatment. Results have been obtained by STATA procedure psmatch2 by Leuven and Sianesi (2013) (matching algorithm: Epanechnikov kernel with bandwidth 0.06). Standard errors in parenthesis are bootstrapped with 200 replications. Source: German Socio-Economic Panel Indicators of the matching quality are reported in Table A 8 and Table A 9. The pseudo-r squared indicates the percentage of the variance which is explained by the estimation model. Since there should be no systematic differences in the characteristics of naturalised and non-naturalised after matching, the pseudo-r squared should be low after matching (Caliendo and Kopeinig 2005). For the total sample the pseudo-r squared is below 0.04, for the subsamples it is of similar magnitude. Although the mean of standardised bias is still 8.2 after matching for the total sample, it was reduced by 78.1% after matching. Thus, the matching quality of the total sample can be assessed as moderate. Furthermore, the matching quality is worse for immigrants from Western European countries and best for immigrants from Eastern European countries (Table A 8). In addition, Table A 9 shows further matching quality indicators for the total sample. For most covariates the percentage reduction in standardised differences is greater than 60%. Since the covariates of treatment and control group should be similar after matching, the two-sample t-test should be insignificant after matching. However, the test suggests significant differences between naturalised and nonnaturalised for several covariates. However, treatment and control group are better balanced in the subsamples by origin. Overall, matching quality can be assessed as modest. 22