• Keine Ergebnisse gefunden

58

To show statistically that the post-matching difference between the means is not too large, following Rosenbaum and Rubin (1985), the table includes the measure of standardized percentage bias (%SB). %SB is calculated twice (before and after the matching procedure) to show the improvement in the comparability between sample means achieved by matching. Caliendo and Kopeinig (2008) report that an after-matching

%SB of less than 3% or 5% is often considered a sufficient indicator of good matching quality. For most of the conditioning variables, Table 3.3 shows that the achieved post-matching %SB is significantly lower than 5%.59 Another indicator of matching quality is the post-matching reduction in mean and median %SB. The mean %SB for the selected variables is 2.6, which is a substantial reduction of 86% from the unmatched sample. The median %SB of 1.7 is also well within the acceptable level of 5%.

The matching quality of the sample consisting of assimilation outcomes (#3-5) is briefly discussed. Variable balance is achieved without conditioning for state dummy variables, survey year dummy variables, and work experience. The means of the conditioning variables for the treatment and the control groups are shown in the appendix E. The matching quality is greatly affected due to the low sample size, as denoted by the substantial increases in %SB.

59

Figures are suggestive that Turkish immigrants were significantly more fearful of xenophobic hostility directed at them rather than the general crime level in their surroundings after the 2011 revelations.

To take one step further, Table 3.4 presents the DiD estimates of the "treatment" effect on worries about xenophobic hostility. 61 The first column of the Table suggests that in the aftermath of the 2011 revelations, worries about xenophobic hostility were increased among Turkish immigrants by 0.152, which is about 7.3 percent of the mean and 20.8 percent of one within-individual standard deviation in the estimated sample.62 After it was uncovered, the NSU episode was featured extensively in German newspapers. Building on SOEP question about newspaper reading, column 2 and 3 of Table 3.4 look at whether the impact of the 2011 revelations was different among on newspapers' readers and non-readers. Consistent with the literature about the role of newspapers (Heath, 1984), the regression results suggest that newspapers readership was actually associated with stronger effects of 2011 revelations. 63

Columns 4-6 present the baseline results for respondents who report reading only foreign newspapers, respondents who read only German newspapers, and respondents who read both. Although the coefficients are estimated with less precision due to low sample sizes, the magnitudes increase from left to right.64 The main result of column 4 suggests a rather limited treatment effect among the readers of foreign newspapers, which understates the role of Turkish newspapers’ coverage of the NSU episode in intensifying the treatment. Column 6 shows that respondents who read both German and foreign newspapers (a proxy for higher consumption of news) report statistically significant treatment effects with larger magnitude than the average effect.

The geographical span of NSU crimes (figure 3.2) also offers another opportunity for a treatment intensity check. As mentioned, five of the 10 murders were committed in the state of Bavaria alone. Additionally, the seven-year-long NSU trial was held in Munich, Bavaria. Thus, it may be expected that the NSU episode received more extensive and relatively frequent news coverage in Bavaria than other states in Germany and that the treatment could have been particularly intense among the respondents from this area. Panel B of

61 To show that the baseline estimates do not depend on the inclusion of control variables, appendix H, presents the estimates without them. The results do not change. Additionally, Table 3.A.5 in the supplementary appendix presents the simple means of the outcomes variables across survey years.

62 The estimated treatment effect and the CTA are not conditional on the implementation of the matching procedure.

Appendix C presents the lead and lag effects of the 2011 revelations for the unmatched sample. The main message of the paper holds.

63 The role of other media platforms with which respondents acquire information could also be interesting, e.g., internet use and time spent watching TV. Unfortunately, however, the information on respondents’ private use of the internet was asked in only the year 2013, i.e., post-treatment year, and the TV watching habits have not been recorded since 1989.

64 Although the sample sizes for the sub-group analysis presented in columns 3-5 are small, the coefficients presented in column 5 are statistically significant. Unfortunately, less precise estimation due to lower sample sizes makes it difficult to establish whether the coefficients in columns 3-5 are indeed statistically different.

60

Table 3.4 shows the re-estimated results that were obtained by restricting the sample to respondents from Bavaria. The estimates show that this subpopulation reported a much larger increase in their worries towards xenophobic hostility than the average effect, hence supporting the hypothesis.

Finally, we also checked whether a number of individual-level characteristics are important considerations for the heterogeneity of the treatment effect. In particular, respondent’s immigration status (FGI vs. SGI), education (high-educated vs. low educated) and religiosity (religious vs. non-religious) are considered. 65 Although the results are imprecisely estimated and have large standard errors, Table 3.A.14 in the supplementary appendix shows that we cannot reject the null hypothesis that the treatment effect was homogeneous across the respondent’s immigration status, education and religiousness.

3.5.2 2011 revelations and social assimilation of Turkish immigrants in Germany

Panel A of Table 3.5 presents the results for assimilation outcomes (responses range from 1 to 5). The results in column 1 suggest that Turkish immigrants were less likely to self-identify as Germans in the aftermath of the 2011 revelations. In terms of magnitude, the decrease in self-identification as German for Turkish immigrants is substantial at about 17 percent of the mean and 41 percent of one within-individual standard deviation in the estimated sample. The results in columns 2 and 3 show that Turkish immigrants substantially increased their self-identification as foreigners living in Germany and their connection with their country of origin. The event-analysis graphs in figures 3.5-3.7 present additional supporting evidence. Thus, these results conclusively highlight the impact of dissimilation by the 2011 revelations on Turkish immigrants in Germany.

Panel B of Table 3.5 shows the results regarding whether the revelations had a negative impact on respondents’ health satisfaction and life satisfaction. The results show that the revelations negatively impacted these factors. In terms of magnitudes, the decreases in health and life satisfaction were about 10 and 6 percent of one within-individual standard deviation, respectively.66 Although not very large, these decreases highlight the relevance of the revelations for the wellbeing of the Turkish Diaspora in Germany.

65 Information on the country of birth of SGI respondents' parents is not available for all respondents. In case where the country of origin was missing for both parents, observations were assumed to belong to the control group. We checked that we obtain similar results when we omit SGI respondents for which the country of origin is missing or when we focus on the FGI alone (see Table 3.A.15).

66 Steinhardt (2018) finds that the rise in xenophobic violence in the 1990s reduced the subjective wellbeing of Turkish immigrants by approximately 0.36 points, which is about 5% of the mean and 19% of the standard deviation in the present estimation sample. In contrast, the results presented here are smaller in magnitude. This is not surprising, however, as Steinhardt considers the case of actual violent incidents, whereas the present study focuses on the impact of news revelations about the crimes committed in the past.

61

Particularly, the results related to health satisfaction provide evidence of increased stress levels and negative health consequences of the 2011 revelations.

3.5.3 Additional robustness checks

To address the concern of ethnicity-specific shocks, the baseline results were re-estimated after controlling for ethnicity-specific linear time trends. In addition to controlling for Turkish immigrants’ specific time trends, the linear time trends were controlled for the following major ethnic groups in Germany: 1) Central and Eastern European immigrants, 2) European migrants, and 3) other ethnic groups. Panel A of Table 3.A.15 in the supplementary appendix presents the main results, which are robust to the inclusion of ethnicity-specific time trends.

Eventually, we checked that the 2011 revelations did not coincide with any significant change in immigrants' labor market outcomes, as measured by hourly wages or unemployment (see Table 3.A.16 in the supplementary appendix). Our baseline results cannot be interpreted as reflecting immigrants' increased economic problems.