Immigration and Crime: The 2015 Refugee Crisis in Germany Yue Huang Otto von Guericke University Magdeburg Michael Kvasnicka Otto von Guericke University Magdeburg, RWI, IZA February 1, 2018 Abstract. In the heyday of the European refugee crisis in 2015, nearly one million refugees came to Germany, causing widespread concern that crime would rise as a result. Using unique self-compiled data, we investigate this question empirically. We find evidence for a hump-shaped relation at county level between the scale of the refugee inflow and the crime rate and that the extent of decentralized accommodation of refugees exerts a negative effect on crime. Importantly, however, this systematic relation disappears if we restrict the analysis to crimes involving at least one German victim and one refugee suspect. Results of 2SLS regressions that control for potential endogeneity in regional refugee inflows corroborate this finding. Our result casts doubt on the explanatory power of most studies on immigration and crime, as these consider but the overall crime rate, not the rate of crimes committed by immigrants against natives. Keywords: Immigration, Refugees, Crime JEL Classification: F22, J15, K42. Acknowledgement: This paper has benefited from comments by participants of seminars at RWI - Leibniz Institute for Economic Research, Universität Hamburg and Otto-von-Guericke-Universität Magdeburg. All remaining errors are our own. Corresponding author. Otto von Guericke University Magdeburg, Universitätsplatz 2, 39106 Magdeburg, Germany. Email: yue.huang@ovgu.de. Phone: +49-391-67-58736. Fax: +49-391-67-41136. Otto von Guericke University Magdeburg, Germany. Email: michael.kvasnicka@ovgu.de.
1 Introduction In the heyday of the 2015 European refugee crisis, which was fueled by the conflict in the Syrian Arab Republic (Tumen, 2016), close to one million refugees arrived in Germany. This largest inflow of refugees to Germany since the early 1990s (Bundesamt für Migration und Flüchtlinge (BAMF), 2016a) posed, and continues to pose, great challenges for state authorities, society and the economy. A major concern among the German public, and one particularly exploited by populist right parties, has been that this mass immigration would lead to more crime. However, little is still known, whether such fears are indeed justified. Empirical studies on the 2015 refugee crisis and crime are still rare; and the broader literature on immigration and crime provides little guidance, given its mixed and inconclusive findings (Bell et al., 2013; Bianchi et al., 2012; Butcher and Piehl, 1998; Gehrsitz and Ungerer, 2017; Mastrobuoni and Pinotti, 2011; Piopiunik and Ruhose, 2017; Spenkuch, 2014). Given this paucity of evidence on the matter, we draw on two novel and unique data sets on regional refugee populations and crime, suspect and victim statistics to study the impact that the 2015 refugee crisis had on the scale and type of crime committed in Germany. The first data set includes detailed statistics on the regional distribution of refugees, their gender and age structure, as well as their type of accommodation (decentralized or centralized, i.e. group quarters) 1, prior to and after the mass inflow of refugees in 2015, drawn from special data extracts from the Statistic on Asylum Seekers Benefits ( Asylbewerberleistungsstatistik ), an administrative public registry with full coverage of the refugee population in Germany. The second data set comprises special data extracts from the police crime statistics of the Federal Criminal Police Office ( Bundeskriminalamt ), which allow us to consider not only the number of crimes of different kinds, as is most common in the literature, but also associated suspects, victims, and, what is unique, crimes differentiated by the nationality (German/foreign) and refugee status of their respective victims and suspects involved. In combination, these rich data make it possible to study in unprecedented detail potential links between the size of immigration, its gender and age composition, and its housing structure on crimes, suspects and victims, differentiated by the nationality and refugee status of individuals and combinations thereof (e.g. involving a suspect that is foreign or a refugee). crimes committed against Germans Exploiting for identification across county variation in the change across time (pre to post the mass inflow of refugees in 2015) of refugee-to-population shares and structures of refugee populations (by gender, age, and type of housing) in first-differenced regressions, we 1 Refugees in centralized accommodation are refugees who reside in a reception center for asylum seekers ( Aufnahmeeinrichtung ) or in a collective accommodation center ( Gemeinschaftsunterkunft ). 1
find evidence for a non-linear and hump-shaped relation between the scale of the refugee inflow and the overall crime rate (the same pattern emerges for the overall suspect rate). Furthermore, and of particular interest from a policy perspective, we also find evidence that decentralized accommodation of refugees tends to exert a negative effect on the crime rate. However, and importantly, the systematic relation between the scale of the refugee inflow and the crime rate disappears if we restrict the analysis to crimes involving at least one German victim and one refugee suspect. Using pre-treatment measures of regional centralized refugee accommodation as instruments in 2SLS regressions for observed county-level changes in refugee populations between 2014 and 2015, produces results that corroborate this finding. 2 This result casts doubt on whether public concerns about refugees and crimes are justified, and it underscores more generally the importance of detailed crime, suspect and victim statistics for analyses of immigration and crime against natives. It also casts doubt on the explanatory power of most studies on immigration and crime, which because of lack of data consider but the overall crime rate, not the rate of crimes committed by foreigners against natives. Our study relates and contributes to a growing body of literature on immigration (of different kinds) and crime. Bell et al. (2013), who investigate refugee migrant inflows to the UK in the late 1990s and early 2000s, exploit official allocation rules for asylum seekers of the National Asylum Support Service. They find evidence for a slight increase in property crimes, but no evidence for an effect on other types of crimes. Bianchi et al. (2012), who study immigration to Italy during the period 1990-2003, use instrumental variables that exploit information of migrant flows to other destination countries, as these should reflect supply-push factors operating in origin countries that do also affect the inflow to Italy. They find a statistical positive effect of immigration on the incidence of robberies at province level. Mastrobuoni and Pinotti (2011), who focus on the legal status of immigrants in Italy, exploit a natural experiment (European Union Enlargement) for identification. They show that possession of legal residence status decreased the propensity of immigrants to commit a crime in areas that provide better labor market opportunities. Butcher and Piehl (1998), who control for characteristics of metropolitan areas in the U.S., find no evidence for an effect of immigration into those areas on crime rates during the 1980s. They also produce evidence which suggests that young immigrants are actually less likely to commit a crime than young natives. Spenkuch (2014), using panel data on U.S. counties from 1980 to 2000 in a first-difference regression model, finds a positive but small impact of immigration 2 These measures include the minimum distance to a county that had a reception center for asylum seekers in 2014 (i.e., prior to the refugee crisis), the number of municipalities in a county which provided centralized accommodation in 2014, and a binary indicator for the existence of a reception center for asylum seekers in a county in 2014. 2
on property crimes but no evidence for an effect on violent crimes. Piopiunik and Ruhose (2017), who investigate the relationship between ethnic German immigrants and crimes, find that immigration increases crime rates in regions with high unemployment, high pre-existing crime levels and large foreigner shares. Finally, Gehrsitz and Ungerer (2017), who focus on the refugee inflow to Germany between 2014 and 2015, document a small positive effect of the refugee inflow on crimes, in particular for drug offenses and fare-dodging. None of these studies, however, makes use of detailed statistics on crimes by foreigners against natives, which are at the heart of public concerns about immigration and crime. Our results suggests that this is a major shortcoming, as such data appears crucial for causal inference concerning the relation between immigration and the likelihood of natives to become the victim of a crime committed by foreigners. The paper proceeds as follows. Section 2 provides background information on the 2015 refugee crisis and the scale and development of criminal activity in Germany over time. Section 3 describes the data sources we employ, defines all variables we use in the empirical analysis, and outlines our estimation strategy. Section 4 presents the main regression results, robustness checks, as well as analyses exploring potential effect heterogeneity. Section 5 discusses the basis for a causal interpretation of our findings and presents additional results from IV regressions. Finally, Section 6 summarizes our main findings and concludes. 2 Background From 1991 to 2015, 24.9 million immigrants came to Germany and they can be classified into different groups according to their migration purpose, such as repatriates, asylum seekers, migrant workers and EU citizens (BAMF, 2016b). However, the year 2015 was particularly marked by the mass inflow of asylum seekers. Around 890,000 asylum seekers arrived in Germany (BAMF, 2016b). The number of asylum applications increased steadily from January 2014 to middle of year 2015, but it rose suddenly dramatically in the second half, or to be precise in the last quarter, of 2015. Bavaria, North Rhine-Westphalia and Baden- Württemberg, the first three states that received the most asylum applications, got about 43.4% of all applications (BAMF, 2015). Around 62.3% of asylum applicants were aged between 16 to 39 and the sex ratio of these applicants was about 3.17 (BAMF, 2015). 3 Crimes by immigrants saw a significant increase in 2015 (BKA, 2016). In 2015, in terms of total crimes (excluding violations of residence act), 206,201 offences were cleared, in which at least one immigrant was identified as a suspect, representing an increase by 79% compared 3 The percentage number and the sex ratio are calculated by authors with the data provided by BAMF (2015). 3
to 2014 (115,011 offences) (BKA, 2016). 4 59,912 immigrants were recorded as suspects of total crimes (excluding violations of residence act) in 2014 while in 2015 this number rose to 114,238, an increase by about 91% (BKA, 2016). Media reports of increasing numbers of burglaries, and of sexual assaults in particular, such as those occurring on New Year s Eve 2015/16 in Cologne, Hamburg, Stuttgart and other cities, have fuelled concerns about public safety and spurred demands for state intervention, including the introduction of an upper cap on refugee inflows to Germany. Figure 1 shows the annual number of asylum seekers receiving standard benefits and the crime rate in Germany from 2009 to 2015. 5 Panel (a) displays a continuing increase in asylum seekers per 100,000 population with a kink from 2014 to 2015. Panel (b) illustrates the total number of crimes as well as the number of crimes excluding violations of residence act (crime key: 890000) 6 per 100,000 population. quite obvious that total crimes increased sharply from 2014 to 2015 while crimes excluding violations of residence act decreased, which means due to the mass inflow of refugees in 2015 the number of cases violating residence act (crime key: 725000) rose extraordinarily. However, violating residence act cannot directly result in damages on person or property in Germany. Therefore, in the following analysis we focus on total crimes excluding violations of residence act (crime key: 890000). Fig. 1: Annual number of asylum seekers and crime rates in Germany, 2009-15 It is refugees per 100k population 200 400 600 800 1000 1200 crimes per 100k population 7200 7400 7600 7800 total crimes excluding violations of residence act (890000) 2009 2010 2011 2012 year 2013 2014 2015 2009 2010 2011 2012 year 2013 2014 2015 (a) asylum seekers per 100,000 population (b) crimes per 100,000 population Increases over time in the number of crimes committed by immigrants, however, were far from uniform across Germany. The observed variation of immigration-related crimes 4 An act is considered cleared if the legitimate personal data of at least one suspect is known. 5 Data on asylum seekers are provided by Statistisches Bundesamt, GENESIS-Online Datenbank (code: 22221-0001). Data on crime rates are provided by Polizeiliche Kriminalstatistik (PKS) 2015, Grundtabelle ab 1987 https://www.bka.de/de/aktuelleinformationen/statistikenlagebilder/polizeilichekriminalstatistik/ PKS2015/Zeitreihen/zeitreihenFaelle.html 6 The full list of crime categories with identity numbers and explanations can be found in the catalogue of criminal offences 2015. 4
across German federal states and across counties within states raises the question, which regional factors contributed to or helped to reduce crimes committed by immigrants. Figure 2 illustrates county-level changes in crime rates (panel (a)), refugee population shares (panel (b)), refugee sex ratios (panel (c)) and shares of refugees in decentralized accommodation (panel (d)). 7 For each subfigure we classify the data into 5 quantiles. The darker the color is, the larger is the value of numbers. From eyeballing it is hard to tell how these 4 factors are correlated with each other. Crime rates seem to increase more in some counties of Rhineland- Palatinate, Baden-Wüttemberg and counties in the Northeast Germany. Counties in the Northern and Western Germany show a relative larger increase in the population share of refugees. Compared to 2014 much more refugees in counties of Bavaria lived in decentralized accommodation. In terms of sex ratios of refugees aged 15-39, it is not easy to say which counties attracted more young male refugees. Hence, in the next section we introduce our data, explain variables and discuss the empirical strategy with which we want to investigate whether the inflow of refugees, the gender structure of refugees as well as the type of refugee accommodation affect criminal activities. 3 Data and Empirical Strategy We use three main data sources to analyze the effect of the 2015 refugee crisis on crimes in Germany. The first data source is the information on the regional distribution of refugees, their composition (gender and age) and housing (in centralized or decentralized accommodation) for two points in time, one prior to the mass inflow of refugees (i.e. pre treament) and one after (i.e. post treatment). The second data source is statistics on annual county-level criminal activities in 2014 and 2015 in Germany, containing various special data extracts in addition to publicly available statistics on crimes, with which we can not only observe all crimes of different categories, but also crimes committed by refugees involving at least one German victim, with which we can answer the research question more accurately whether refugees do more crimes against German natives. The third data source is the demographic and economic characteristic of German counties in the pre-treatment period. Using all these three sources, we are able to set up different measures of criminal activities, generate the inflow and the structure of refugees and produce variables on the pre-treatment demographic and economic condition of German counties. In the following, we firstly describe in detail the data source and variables in Section 3.1. Secondly, we present and discuss our empirical strategy in Section 3.2. 7 Data source: crime data are from PKS 2014/15 and refugee data are from the Statistic on Asylum Seekers that will be discussed in detail in Section 3. 5
Fig. 2: County-level changes in crime rates, refugee population shares, refugee sex ratios and shares of refugees in decentralized accommodation, 2014-15 (a) crime rate (b) population share (c) sex ratio (15-39) (d) decentralized accommodation 3.1 Data and Variables The first data source is the information on refugee populations, their composition and housing, provided by the Statistic on Asylum Seekers Benefits (Asylbewerberleistungsstatistik). 6
These data are administrative records for individuals who seek refugee status in Germany and receive some kind of regular financial or other support from public authorities under the Act on Benefit for Asylum Seekers (Asylbewerberleistungsgesetz) (Statistisches Bundesamt, 2015). We use stock data for the sampling dates 31 December 2014 (pre treatment) and 31 December 2015 (post treatment) for German counties and municipalities. At county level, data extracts contain information on the total number of refugees, the number of male refugees, the number of total and male refugees aged 15-39 and the number of refugees in decentralized accommodation. In addition to the total number of refugees, the standard measure, we are also interested in the structure of refugees, gender, age and their accommodation. Since the mass inflow of refugees brought a large number of young males to Germany and on average males, especially young males, are more likely to commit crimes, we hence want to investigate whether the sex ratio of young refugees influences criminal activities in Germany. Also, how to properly accommodate these refugees, in centralized or decentralized accommodation, has become another question. In terms of crimes we want to analyze whether decentralized accommodation for refugees would increase or decrease crime rates. At municipality level, the data provide an indicator on whether or not there was a reception center or collective accommodation center for asylum seekers at the two sampling dates. We can then transfer these municipality-level data to county-level data, showing whether a county had centralized accommodation for asylum seekers on the last day of 2014/15. With this information we can discuss whether the effect of refugee inflows and their structure on crimes differs in counties where there was any or no reception centers in 2014. We can also use the pre-treatment refugee reception capacities, i.e. whether there was centralized accommodation, as an exogenous source of variations in refugee inflow. The second data source we use is the annual county-level official police crime statistics, provided by the Federal Criminal Police Office (Bundeskriminalamt, BKA) in 2014 and 2015. These data contain two types of data extracts. One type is the publicly available BKA statistics, including the number of crimes, suspects, foreign suspects, victims and German victims at county level in 2014 and 2015. The other type, the special data extract, provides us county-level data in 2014 and 2015 on the annual number of: (A) suspects (1) who are refugees (2) who are refugees and have committed a crime involving at least one German victim; (B) victims who are German and have been involved in a crime committed by at least one refugee. As far as we know, we are the first to exploit these special data. Because of the mass inflow of refugees, if more crimes could be observed, those crimes are not necessarily committed by refugees and the victim of crimes are neither necessarily Germans. But the heart of the public concern is whether refugees commit more crimes against Germans after the large number of refugees arrived. Using our special data we are able to differentiate 7
crimes in total, crimes committed by refugees and crimes committed by refugees against Germans. Therefore, we can discuss whether the public concern about refugees and crimes is justified, especially whether refugees do more crimes against Germans. Such analysis, to our knowledge, has never been done before. In addition to the 2014/15 data, we also have county-level information on criminal activities in 2013, both number of crimes and suspects. These data can be used to control for the pre-treatment crime developing trend. The last data source is the information on demographic and economic conditions of German counties in pre-treatment period, published by the Federal Statistical Office of Germany (Destatis). They are measures of the county-level population, GDP per capita and unemployment rate. Using our fruitful and special data, we are able to construct dependent and main independent variables and generate other control variables. Now, we introduce all variables in turn. Dependent variables: crime rate, suspect rate and victimization rate. A very common used measure of criminal activities is the number of crimes standardized by populations. Using this method, we define the crime rate as the ratio of numbers of crimes over total population. In stead of this standard measure, we also use suspect rates and victimization rates alternatively. The suspect rate is calculated as the population share of suspects. In terms of crimes committed by refugees, the overall refugee suspect rate is then defined as the ratio of refugee suspects over total population. When we observe crimes committed by at least one refugee and involving at least one German victim, the refugee suspect rate is the number of refugee suspects divided by the German population. The victimization rate is a ratio of total victims over total population. Since we are interested in whether Germans are more involved in crimes committed by refugees, we also define a German victimization rate as a German population share of German victims in those crimes. Main independent variables: population share of refugees, sex ratio of refugees aged 15-39, share of refugees in decentralized accommodation. The main independent variable in our analysis are three indicators for refugee structures, i.e. population share of refugees, sex ratio of refugees aged 15-39 and share of refugees in decentralized accommodation. Using the Statistic on Asylum Seekers Benefits, we standardize refugee populations by county-level total populations. Since we want to investigate whether young and male refugees commit more crimes, we use the sex ratio of refugees aged 15-39 as another main independent variable. How to accommodate refugees is a big concern for policy makers. In order to investigate the effect of the type of refugee accommodation on crimes, we generate the third variable indicating refugee structures, the share of refugees in decentralized accommodation. It is calculated as the number of refugees living in decentralized accommodation 8
over the total number of refugees. Other covariates. In addition to our main explanatory variables, we will in part of our analysis control for other potential determinants that can influence both regional crimes and refugee populations. There are two kinds of control variables, controls for pre-treatment changes and pre-treatment levels. For the pre-treatment change we include the change in county-level crime rates, suspect rates and the logarithm of population between 2013 and 2014. For the pre-treatment level we control for annual GDP per capita and unemployment rate in 2014. Using variables mentioned above, we will do a first-difference analysis on the effect of refugee inflows on crimes in Germany. We discuss our empirical strategy in the next section in detail. 3.2 Empirical Strategy In order to avoid bias due to time invariant omitted variables, i.e. time-fixed factors, we exploit changes over time in the county-level characteristics of refugees (the total number, sex ratios and types of accommodation) and correlate these with changes in measures of criminal activities. The first-difference model reads as follows: Y i,t = α + Ref i,t β + X i,t 1 γ + ε i,t, where Y i,t is the change in a particular outcome for county i between year t (post-treatment period 2015) and year t 1 (pre-treatment period 2014), i.e. the change in crime rates, suspects rates and victimization rates. Ref i,t is a vector of variables capturing changes in characteristics of refugees in county i between year t and year t 1, i.e. the change in the population share of refugees ( Ref), sex ratios of refugees aged 15-39 ( SR 1539 ) and the share of refugees living in decentralized accommodation ( Ref dec ). X i,t 1 is a vector of control variables for period t 1 characteristics of county i, including the pre-treatment crime development ( Crime 1314, Suspect 1314 ), pre-treatment population growth ( ln(p op) 1314 ) and GDP per capita (ln(gdp 14 )) and unemployment rate (Unemployment 14 ) in 2014. α, the constant term, captures the common difference across counties in the pre- and posttreatment period that may have an effect on outcomes. ε i,t is an error term. By correlating outcomes and refugee population characteristics in a first-difference framework rather than in levels, we can control for time-invariant (un)observable factors that may be correlated with both Y i,t and Ref i,t, e.g. persistent level differences between counties in demographic, economic, geographic factors or preferences. We also include various county characteristics in 9
t 1, i.e. X i,t 1, because the trend or the level of county characteristics may be correlated with both the inflow and structure of refugees and the change in outcomes over time. Unlike other international migrations that are more gradual, in peace time or with labor market correlated, the fuse of the 2015 refugee crisis in Germany was war, ethnic and religious conflict and it was sudden and massive, especially in the first several months. Therefore, the feature of the refugee crisis makes the inflow of refugees to Germany less self-selected. However, the regional distribution of refugees may not be completely exogenous, which might potentially confound the relationship with refugee inflows and criminal activities. From the country level to state level, refugees are allocated according to the Königsteiner Schlüssel that takes the economic and population situation in each state into consideration. Once we control for the state fixed effect, the difference of counties between states are netted out. Allocating refugees within a state is complicated. States have different regulations. For instance, Schleswig-Holstein, Lower Saxony, Hesse, Rhineland-Palatinate, Baden-Württemberg, Mecklenburg-Vorpommern, Saxony, Saxony-Anhalt and Thuringia allocate refugees across counties using the share of county population over the state population, defining that counties with more population should have more refugees, while North Rhine- Westphalia and Brandenburg use a quote that takes county population and county area into account. However, even with a fixed quote defining the number of refugees a county should accept, a slight deviation is but possible and allowed. In our empirical analysis we control for factors that may influence the regional refugee settlement patterns and the regional criminal activities, for example pre-treatment regional changes in criminal activities and population growth and pre-treatment level differences in GDP per capita and unemployment rates. Using our empirical strategy the effect of refugee inflows on criminal activities should be properly and correctly identified. Since all data we have are county-level data, the observations in the analysis are the 402 counties in Germany. However, 17 counties in the state Rhineland-Palatinate and 2 counties in the state Baden-Württemberg have missing data, so they are dropped from the estimation sample. Therefore, altogether we have 383 counties in our final estimation sample. Table 1 provides summary statistics for the estimation sample. In the next section we will present and interpret our main results, show a robustness check and discuss the effect heterogeneity. 10
Table 1: Summary Statistics for Estimation Sample Observation Mean Std. Dev. Min Max crime rate 383 0.0013 0.0046 0.0201 0.0152 suspect rate 383 0.0005 0.0016 0.0070 0.0060 Ref 383 0.0074 0.0064 0.0044 0.0716 SR 1539 383 0.2172 2.2109 28.2691 10.6769 Ref dec 383 0.0431 0.1669 0.9555 0.5775 Crime 1314 383 0.0002 0.0040 0.0188 0.0160 Suspect 1314 383 0.0002 0.0015 0.0091 0.0064 ln(p op) 1314 383 0.0037 0.0061 0.0185 0.0240 ln(gdp 14 ) 383 10.3537 0.3409 9.6121 11.8243 Unemployment 14 383 6.3352 2.9242 1.4000 15.4000 4 Results 4.1 Main Results In this section we present and interpret our main results by observing crime rate, suspect rate and victimization rate. We do not only investigate the effect of refugee population size on criminal activities, but also the composition, i.e. the sex ratio of refugees aged 15-39 and the share of refugees in decentralized accommodation. As the heart of the public concern about refugees is whether they commit more crimes against Germans, using our special and unique data we observe crimes in which there is at least one refugee suspect and at least one German victim when we analyze the refugee inflow effect on suspect rates and victimization rates. As mentioned in Section 2, crimes we observe are total crimes excluding those against the Residence Act, the Asylum Procedures Act, and the Freedom of Movement Act/E.U. (key: 890000). We now discuss our results in turn. 4.1.1 Crime Rate The first measure of criminal activities is the crime rate that has been often used in the literature (Bell et al., 2013; Bianchi et al., 2012; Butcher and Piehl, 1998; Gehrsitz and Ungerer, 2017; Piopiunik and Ruhose, 2017). We firstly include the three main independent variables separately into regressions and then put them together into one regression at the same time. The estimation result is shown in Table 2. In column (1) we regress the change in crime rates on the change in population share of refugees ( Ref), controlling for the differential state trend, but no significant effect of the change in refugee populations is found. It is possible that the effect of refugee inflows on crime rates is not linear. Therefore, 11
in column (2) we add in the square term of the change in population share of refugees ( Ref 2 ). Now the linear coefficient is statistically positive significant and the quadratic coefficient is statistically negative significant, which means the change in the population share of refugees is positively associated with the change in the crime rate, but the increase in the crime rate becomes slower and slower as the increase in the population share of refugees gets larger. In column (3) we include all other control variables, i.e. the change in the crime rate between 2013 and 2014 ( Crime 1314 ), the change in the logarithm of population between 2013 and 2014 ( ln(p op) 1314 ), the logarithm of GDP per capita in 2014 (ln(gdp 14 )) and the unemployment rate in 2014 (Unemployment 14 ). Compared to the result in column (2) the effect of refugee inflows in terms of magnitude and significance level only changes slightly. The estimated coefficient of Crime 1314 shows a significant negative effect of the pre-treatment crime trend, which indicates that counties experiencing a decrease in the crime rate from 2013 to 2014 had an increase in this factor from 2014 to 2015. It could be a mean reversion effect as the crime rate could fluctuate around its mean value. Hence, when it increases in one period, it may decrease in the next period. 8 ln(p op) 1314 has a significant positive effect on the change in the crime rate, indicating that crimes increase more in counties with a larger population growth. Both GDP per capita and the unemployment rate in 2014 describe whether counties with different economic conditions have different changes in the crime rate. The result shows that counties having a higher unemployment rate in 2014 experienced a larger increase in the crime rate from 2014 to 2015. In column (4) and (5), instead of the size of the refugee inflow, we investigate respectively the effect of the change in the refugee structure on the change in the crime rate, i.e. the change in sex ratios of refugees aged 15-39 ( SR 1539 ) in column (4) and the change in the share of refugees living in decentralized accommodation ( Ref dec ) in column (5). The refugee sex ratio displays no significant effect on the crime rate while the change of refugees in decentralized accommodation has a statistically significant negative effect on the change in the crime rate. Such result shows that more refugees in decentralized accommodation seems to be able to reduce crimes. In column (6) we include all three factors of refugees into the regression at the same time. Therefore, when we investigate the effect of one refugee factor, we can also control for the other two factors. Column (6) show that the effect of the change in the population share of refugees still has a hump-shaped effect on the change in the crime rate and the share of refugees living in decentralized accommodation is negatively associated with the crime rate. Using the result in this specification, we first calculate the marginal effect of Ref at the mean of all independent variables that is equal to 0.1602. 8 Due to data limitation we cannot control for the average crime rate change over several years before the refugee crisis. 12
Then, 1 standard deviation in Ref increases the crime rate by 0.001 (= 0.0064 0.1602) that is about 1.64% of the crime rate in 2014. With the result in column (6), panel (a) and (b) of Figure A-1 in Appendix show that the predicted crime rate reaches its peak and the marginal effect of Ref becomes zero when Ref approximately equals to 0.034 and most of the counties have a refugee change lower than this level. Whether those counties with Ref larger than 0.034 (outliers) drive our results will be discussed in Section 4.2. In case we drop the outliers from the estimation sample, our findings are still robust. Table 2: The effect of refugee inflows on crime rates mean of crime rate in 2014: 0.0627 (1) (2) (3) (4) (5) (6) Ref 0.038 0.205 0.224 0.205 (0.035) (0.088) (0.084) (0.085) Ref 2 2.947 3.215 3.008 (1.427) (1.362) (1.366) SR 1539 1.134e-05 1.84e-05 (9.83e-05) (9.74e-05) Ref dec 0.003 0.003 (0.001) (0.001) Crime 1314 0.333 0.329 0.322 0.326 (0.056) (0.056) (0.056) (0.056) ln(p op) 1314 0.085 0.087 0.081 0.080 (0.044) (0.044) (0.044) (0.044) ln(gdp 14 ) 0.001 0.001 0.001 0.001 (0.001) (0.001) (0.001) (0.001) Unemployment 14 2.39e-04 2.03e-04 1.79e-04 2.14e-04 (1.22e-04) (1.23e-04) (1.22e-04) (1.23e-04) observations 383 383 383 383 383 383 Notes: We focus on total crimes, excluding those against the Residence Act, the Asylum Procedures Act, and the Freedom of Movement Act/E.U., with the key number 890000. The dependent variable is the change in the population share of crimes, i.e. crime rate, between 2014 and 2015. The main independent variables are the change between 2014 and 2015 in the population share of refugees ( Ref), in the sex ratio of refugees aged 15-39 ( SR 1539), and in the share of decentralized refugees ( Ref dec ). As control variables we include the change in the crime rate between 2013 and 2014, the change in the logarithm of population between 2013 and 2014, the logarithm of GDP per capita in 2014 and the unemployment rate in 2014. Differential state trends are controlled for. *, **, *** denote statistical significance at 10%, 5% and 1% level. Using the same strategy as in this section, we will discuss our results when we use the suspect rate, instead of the crime rate, to investigate the effect of the refugee crisis on criminal activities in Germany in Section 4.1.2. Of most importance, we can figure out the refugee status of suspects and the nationality of victims, with which we can better answer the research question, also the concern of the public, whether refugees do more crimes against Germans. 4.1.2 Suspect Rate Following the previous analysis we change the outcome variable to the change in suspect rates and re-estimate the model. Table 3 shows the estimation result. Controlling for all 13
covariates, 9 in column (1)-(3) we estimate the effect of the three refugee characteristics separately and in column (4) we include all of them. Similar to findings in Section 4.1.1, the change in suspect rates increases when Ref increases, but the marginal increase becomes smaller as Ref gets larger (column (1)). Column (2) shows that the sex ratio of refugees aged 15-39 has no significant effect, but in column (3) the share of refugees in decentralized accommodation is estimated to have a statistical significant negative effect on the suspect rate. In column (4), when all refugee variables are included in one regression, the marginal effect of Ref at the mean of all explanatory variables is equal to 0.0787 and 1 standard deviation increase of Ref increases suspect rate by 0.0005 (= 0.0064 0.0787) that is about 1.93% of the mean value of suspect rates in 2014. Panel (c) and (d) of Figure A-1 in Appendix illustrate the predicted value of suspect rate and the marginal effect of Ref at different levels of the refugee change using the result in column (4) of Table 3. The humpshaped curve in panel (c) indicates a positive effect of the change in the population share of refugees on the change in suspect rates for most observations and the large confidence interval on the decreasing part of this curve is derived by the very few counties having extraordinary increase in the number of refugees. The predicted suspect rate reaches its peak when Ref approximately equals to 0.034 and most of the counties have a refugee change lower than this level. Till now, we observe all crimes that can be committed by Germans or foreigners, including refugees, and can involve German or foreign victims. As we have mentioned before, the public is caring more about whether refugees commit more crimes when the mass inflow of refugees arrived in Germany, especially crimes against Germans. Therefore, now, we restrict crimes firstly to those committed by refugees and secondly to those committed by at least one refugee and involving at least one German victim. Table 4 shows the estimation result and we only show the coefficient and its significance level of main independent variables. 10 In column (1)-(4) of Table 4 we observe crimes committed by refugees with no restrictions on victims, i.e. victims being Germans or foreigners including refugees. The outcome is measured by the change in the ratio of refugee suspects over total population. In column (1) the linear coefficient of Ref is statistically significant positive. Although the coefficient of the quadratic term is negative, it is not statistically significant. However, the significance level of the linear and quadratic term cannot be discussed separately; they are statistically jointly significant. The sex ratio of young refugees still shows no significant effect (column 9 Instead of the crime rate change from 2013 to 2014, we use the suspect rate change from 2013 to 2014 to control for the pre-treatment trend difference between counties and it also shows a significant negative effect on the outcome. 10 Results of the whole regression model as well as complete regression results in the following analysis can be given by authors on request. 14
Table 3: The effect of refugee inflows on suspect rates mean of suspect rate in 2014: 0.0261 (1) (2) (3) (4) Ref 0.105 0.098 (0.029) (0.029) Ref 2 1.382 1.310 (0.469) (0.470) SR 1539 3.07e-05 2.69e-05 (3.41e-05) (3.35e-05) Ref dec 0.001 7.46e-04 (0.001) (5.03e-04) Suspect 1314 0.227 0.210 0.210 0.225 (0.053) (0.054) (0.054) (0.053) ln(p op) 1314 0.029 0.029 0.027 0.027 (0.015) (0.015) (0.015) (0.015) ln(gdp 14 ) 4.49e-04 4.10e-04 4.52e-04 4.99e-04 (2.66e-04) (2.69e-04) (2.69e-04) (2.68e-04) Unemployment 14 6.51e-05 5.02e-05 4.10e-05 5.98e-05 (4.20e-05) (4.24e-05) (4.24e-05) (4.22e-05) observations 383 383 383 383 Notes: Crime key: 890000. The dependent variable is the change in the population share of suspects between 2015 and 2014. The main independent variables are the change between 2015 and 2014 in the population share of refugees ( Ref), in the sex ratio of refugees aged 15-39 ( SR 1539), and in the share of decentralized refugees ( Ref dec ). As control variables we include the change in the population share of suspects between 2014 and 2013, the change in the logarithm of population between 2014 and 2013, the logarithm of GDP per capita in 2014 and the unemployment rate in 2014. Differential state trends are controlled for. *, **, *** denote statistical significance at 10%, 5% and 1% level. (2)). The share of refugees living in decentralized accommodation ( Ref dec ) has a negative effect on the refugee suspect rate in column (3), but it becomes insignificant in column (4) as it is significantly correlated with the population share of refugees ( Ref) with a correlation coefficient equal to 0.1489. In the next four specifications (column (5)-(8)) we restrict crimes further to those committed by at least one refugee and involving at least one German victim. With such crimes, the outcome variable changes to the ratio of refugee suspects over German population, indicating the frequency of Germans to meet refugees who commit a crime against natives. In these four specifications in which the three characteristics of refugees are firstly separately and then together treated, the change in the population share of refugees and in the share of refugees in decentralized accommodation show no significant effect on refugee suspect rates and no positive effect of the change in refugee sex ratios is found. Obviously, comparing the results in Table 4, we can find that when we observe crimes committed by refugees with all possible victims (of domestic or foreign nationality), more refugees in the population tends to increase the refugee suspect rate. However, if we add a restriction on victims, i.e. crimes committed by refugees against Germans, the significant effect of the population share of refugees disappears, which means the significant positive 15
Table 4: The effect of refugee inflows on suspect rates of crimes committed by refugees crimes committed by refugees crimes committed by refugees against Germans ref. suspects ref. suspects pop., mean of suspect rate in 2014: 0.0007 G. pop., mean of suspect rate in 2014: 0.0001 (1) (2) (3) (4) (5) (6) (7) (8) Ref 0.027 0.025 0.002 0.002 (0.012) (0.012) (0.001) (0.001) Ref 2 0.228 0.202 0.023 0.021 (0.187) (0.188) (0.024) (0.024) SR 1539 8.57e-06 7.39e-06 2.97e-06 2.88e-06 (1.36e-05) (1.34e-05) (1.69e-06) (1.69e-06) Ref dec 3.74e-04 2.79e-04 2.10e-05 1.38e-05 (2.01e-04) (2.01e-04) (2.51e-05) (2.53e-05) observations 383 383 383 383 383 383 383 383 Notes: Crime key: 890000. The dependent variable of the first four specifications is the change in the population share of refugee suspects between 2014 and 2015 and of the next four specifications the change in the share of refugee suspects against Germans over German population between 2014 and 2015. The main independent variables are the change between 2014 and 2015 in the population share of refugees ( Ref), in the sex ratio of refugees aged 15-39 ( SR 1539), and in the share of decentralized refugees ( Ref dec ). As control variables we include the change in the population share of suspects between 2013 and 2014, the change in the logarithm of population between 2013 and 2014, the logarithm of GDP per capita in 2014 and the unemployment rate in 2014. Differential state trends are controlled for. *, **, *** denote statistical significance at 10%, 5% and 1% level. effect found in column (1) and (4) comes from crimes in which refugees attack on foreigners including refugees, e.g. fighting events occurred among refugees in the reception center for asylum seekers. Therefore, the public concern that refugees commit more crimes against Germans is not confirmed. 4.1.3 Victimization Rate To answer the question whether Germans are affected by the refugee crisis, it may be better to do an analysis from the perspective of German populations. Hence, the third measure of criminal activities in this study is the German victimization rate of crimes committed by refugees. The outcome variable is the change in the share of Germans who were as victims involved in a crime committed by at least one refugee. Therefore, this analysis tells what is the probability of Germans to be attacked by refugees, which may be the closest to the public concern from the view of natives whether refugees do more crimes against Germans and whether domestic citizens are more likely to be attacked by refugees. As the previous analysis, we firstly investigate the effect of the size and the structure of the refugee inflow one by one and then exploit the specification with all three refugee factors together. The estimated coefficient of main independent variables and their significance level are presented in Table 5. Very similar to findings in the analysis on the refugee suspect rate of crimes against Germans (Section 4.1.2), we find no statistical significant effect of Ref, SR 1539 and Ref dec on the change in the German victimization rate. Hence, it proves again that the public concern that refugees commit more crimes against Germans is not justified. 16
Table 5: The effect of refugee inflows on German victimization rates crimes committed by refugees against Germans G. victims G. pop., mean of victimization rate in 2014: 0.0001 (1) (2) (3) (4) Ref 0.003 0.003 (0.002) (0.002) Ref 2 0.035 0.035 (0.037) (0.037) SR 1539 3.70e-06 3.54e-06 (2.64e-06) (2.65e-06) Ref dec 4.49e-06 1.48e-05 (3.92e-05) (3.96e-05) observations 383 383 383 383 Notes: Crime key: 890000. The dependent variable is the change in the ratio of German victims attacked by refugees over German population between 2014 and 2015. The main independent variables are the change between 2014 and 2015 in the population share of refugees ( Ref), in the sex ratio of refugees aged 15-39 ( SR 1539), and in the share of decentralized refugees ( Ref dec ). As control variables we include the change in the population share of suspects between 2013 and 2014, the change in the logarithm of population between 2013 and 2014, the logarithm of GDP per capita in 2014 and the unemployment rate in 2014. Differential state trends are controlled for. *, **, *** denote statistical significance at 10%, 5% and 1% level. Concluding the findings in Section 4.1.1, 4.1.2 and 4.1.3, we can tell that the change in the population share of refugees is statistically significantly positively associated with the change in the overall crime rate, overall suspect rate and even the overall refugee suspect rate, and more refugees living in the decentralized accommodation tends to decrease the overall crime rate, overall suspect rate and overall refugee suspect. However, when we restrict crimes to those committed by at least one refugee and involving at least one German victim, the effect of these two refugee characteristics, Ref and Ref dec, on refugee suspect rates and German victimization rates are no longer statistically significant. 4.2 Robustness Check - Result Sensitivity In the main analysis (Section 4.1) we exploited a first-difference model to control for observed and unobserved factors that can be correlated with both characteristics of refugees and criminal activities. We used a non-linear model to investigate the effect of the refugee inflow on outcomes measured in different ways, which is rarely done in the literature. Also, we analysed whether the structure of the refugee inflow, i.e. sex ratios of refugees aged 15-39 and the share of refugees living in decentralized accommodation, has significant effect on criminal activities. In this section we discuss whether the hump-shaped effect we found with the non-linear model is driven by the outliers in the sample, i.e. counties with extraordinary large increase in the population share of refugees. In the main analysis we include 383 counties with exact data for all necessary variables 17
to the estimation sample and we find a hump-shaped effect of the change in the population share of refugees ( Ref) on the overall crime rate and overall suspect rate. Figure A-1 in Appendix shows that a very few counties have an extreme increase in the population share of refugees from 2014 to 2015 and the marginal effect of Ref on the corresponding outcome becomes negative in these counties. Do these counties drive down the marginal effect of Ref? If yes, what happens if we only observe the other counties that are still most of the observations in the original estimation sample? Is a linear or still a non-linear model suitable for the restricted sample? In order to answer these questions, we drop counties being the top 1% of Ref, which in practice drops three counties, i.e. Gießen, Schweinfurt and Saale-Holzland-Kreis, that have a Ref equal to 0.0448, 0.0714 and 0.0716 respectively, quite far from the mean value. Firstly, we still estimate a non-linear model for the specification only including Ref as the main explanatory variable and also for the one exploiting all three refugee characteristics at the same time, controlling for all other covariates. Secondly, in stead of the non-linear model, we exploit a linear model. For both non-linear and linear model, we investigate the effect of Ref on overall (refugee) suspect rates and on refugee suspect rates and German victimization rates of crimes committed by at least one refugee and involving at least one German victim. The estimation result is presented in Table A-1 in Appendix. After dropping the outliers we have 380 observations. In panel (a) we still exploit a non-linear model, which is the same as the main analysis. However, with this slightly restricted estimation sample we find no statistical significant hump-shaped effect of Ref on the overall (refugee) suspect rate (column (1)-(4) of panel (a)). When we exploit a linear model in panel (b), we find a statistically significant positive effect of Ref on the overall (refugee) suspect rate (column (1)-(4) panel of (B)). But then, if we restrict crimes to those committed by at least one refugee against at least one German, the setting that can answer the research question the best, the positive effect of Ref is no long statistical significant (column (5)-(8) of panel (b)). Therefore, on one hand, the hump-shaped effect found in Section 4.1 seems to be driven by the outliers. On the other hand, both linear and non-linear model show identical findings that Ref is positively associated with the change in the overall (refugee) suspect rate if we observe all crimes (committed by refugees); nevertheless, no significant effect of the size of refugees is found once we only analyse crimes by refugees against Germans. 18