Refugees welcome, but not in my backyard? The impact of immigration on right-wing voting: evidence from Germany

In recent years, the world has faced the second largest refugee movement since World War II. The increase in immigration is largely due to the high number of refugees who escape from the Syrian civil war and an increasingly unstable Middle East.

For the sake of simplicity, we will not distinguish whether an asylum seeker is already a registered refugee, is still in the registration process, or has not been registered yet but claims to receive refugee status at some time. Since the registration of asylum seekers is an ongoing process, we will refer to all three groups as refugees.

The election outcome of 2016 evokes the research question for this article. In general, we are interested in what degree the election outcome was caused by local refugee presence. Did the presence of local refugees induce a rise in AFD votes and a contraction in CDU votes? To be precise, we study the micro-level association between the density of refugee accommodations and the local election outcomes for the AFD and CDU parties at the municipality level. Our research question contributes to the rising interest in the influence of immigration on election outcomes. The refugee influx in 2015 acts as an ideal natural experiment to study the effect of a sudden immigration shock on short-run voting behavior. The event is similar to other famous immigration shocks, such as the 1980 Marial Boatlift in the sense that a large number of immigrants from one predominant ethnic group migrate to a certain location in a short period of time. In the present example, the predominant ethnic group, which accounts for 35.9% of the total immigration in 2015, were Syrians (BAMF, 2016). The refugee crisis is a unique natural experiment for two reasons. First, within a year, more than one million immigrants arrived in Germany, which is the highest immigration number to a single European country in recent history. The sheer mass of people required quick allocation and housing solutions. The housing market did not provide the necessary volume of accommodations at a sufficiently low price. Since vacancies were scarce and urgently needed, new refugee shelters were built, and existing vacancies were converted to refugee accommodations. The refugee allocation was managed by the German Government, which followed an allocation scheme embedded in the German legislature.

The distribution among German counties follows a distribution scheme, namely, the Koenigssteiner Schluessel, which directs more refugees to higher populated federal states with higher tax incomes. However, the scheme does not apply to municipalities within the states, which results in an increasingly uneven distribution.

Figure 1 illustrates the uneven distribution of refugee accommodations and the AFD vote shares in the 2016 state election for 1,101 municipalities in the German state of Baden-Wuerttemberg (BW). Sixty-one percent of the municipalities accommodate refugees, while 39% do not accommodate refugees. Our identification strategy exploits the regional variation of refugee settlement, voting behavior, and socioeconomic differences among the municipalities in the state of BW. We use a spatial econometric framework, which allows modeling spillover effects among the regions and reduces estimation bias. The framework is able to model two channels through which the presence of refugees can influence election outcomes directly and indirectly. We distinguish between the direct presence of refugees, that is, the backyard effect, and the indirect presence of refugee, that is, the neighborhood effect. The backyard effect analyzes how the number of refugee accommodations (per 1,000 residents) within a municipality affects the election outcome, while the neighborhood effect describes how the number of refugee accommodations in neighboring municipalities affects the election outcome.

Figure 1

Both effects are measures for the degree of refugee presence through physical distance. While the first effect reflects direct proximity, the second effect measures the influence of refugees in neighboring municipalities, which are further away but still close enough to be perceivable for voters. Although there is a rising number of literature that studies the influence of immigration on voting behavior, to the best of our knowledge, there is no paper that investigates how the physical presence of immigrants affects voting behavior directly and indirectly.

For a better understanding, it is useful to shed some light on the political landscape in Germany. In 2016, five state elections took place in Germany, particularly in BW, Berlin, Mecklenburg-Western Pomerania, Rhineland-Palatinate, and Saxony-Anhalt. Due to incomplete or insufficient data regarding the latter four states, we restrict the analysis to the largest of the five states, namely, BW. The state has a population of 10.9 million which corresponds to 13% of the total German population. BW is divided into 1,101 municipalities with an average of 9,900 residents per municipality. For the analysis, we chose the election year 2016 due to its temporal closeness to the refugee crisis.

The state election in Baden-Wuerttemberg was held on 13 March 2016.

For our analysis, we require three categories of data: election data, socioeconomic data, and refugee settlement data at the municipality level. Table 1 gives a variable overview. First, voting data was made available by the Federal Statistical Office of Baden-Wuerttemberg (FSOBW, 2016). We measure the election outcome for each party in percentage points per municipality. Furthermore, we include the vote share of the 2013 election to control for previous election outcomes. This is important because refugees were not a political topic during the 2013 election. We also account for the election turnout rate from the previous state election in 2013 because the refugee crisis might have mobilized nonvoters from 2013 which participated in the 2016 election. Second, the FSOBW and the Federal Institute for Employment Research provided socioeconomic data. We include a set of socioeconomic control variables to account for voting differences among social groups. The main variables are GDP per municipality, the unemployment rate, mean age, gender, marital status, the share of residents with immigration background, and the share of residents with basic education. We also include the population density per hectare to distinguish between rural and urban voting preferences.

Third, the Ministry of Social Policy and Integration provided refugee settlement data. We measure the impact of refugee presence on the election outcome in terms of refugee accommodations per 1,000 residents. Because this is a crucial variable for our analysis, we describe the identification strategy in detail. The record date for the inventory of refugee accommodations is March 2016, shortly before the state election. We do not distinguish between newly built refugee shelters and existing buildings that have been turned into refugee accommodations. A potentially more precise measure for refugee presence could be the absolute number of refugees per accommodation. This type of data is not at our disposal, so we cannot distinguish the capacity among the accommodations. This might induce some inaccuracy caused by an underrepresentation of large accommodations which host many refugees and an overrepresentation of regular accommodations which host fewer refugees. Nevertheless, the number of refugee accommodations is a fairly good measure for refugee presence. This is due to the fact that refugees are allocated to municipalities based on an allocation scheme which defines the population of the municipalities as the driving allocation factor. The detailed allocation scheme is defined in the Law on the Admission of Refugees (FlüAG).

In German: Flüchtlingsaufnahmegesetz.

Although panel data is a more elegant way to deal with endogeneity, there are two reasons why we choose a cross-section approach. First, the AFD party became an anti-immigration party in recent years. It evolved from a Euro-sceptic party, which focused on European economic policy, to a right-wing party with a restrictive stance toward immigration. In other words, voters who oppose immigration did not necessarily vote for the AFD party in 2013. Against this backdrop, it is unreasonable to compare election outcomes from 2013 and 2016. Second, the party was established in 2013 and participated only in two federal elections, that is, 2013 and 2017. As for the state elections, the party participated in only one election cycle in 2016. Due to the recent transition in AFD’s political agenda and the lack of election periods, we choose a cross-section approach.

Using municipality level data, we build a simple model where the vote share for a political party depends on the number of refugee accommodations per 1,000 residents (refhome_i) and a set of policy and socioeconomic control variables (X_i). We exploit the variation among i = 1, 2, …, N municipalities. Parameter a₁ represents the marginal effect of direct refugee presence, that is, the backyard effect. For a particular draw from the population, we formulate the following model.

For the sake of simplicity, we summarize the policy variables and the socioeconomic control variables in matrix X_i which includes k independent variables. The corresponding coefficient vector is b. For the sake of clarity, and to highlight the backyard effect, we separate the refugee presence variable (refhome_i) from the sum of independent variables.

Since all the data have a spatial context, that is, all variables are attributes of spatial units, we must account for spatial spillovers among the variables. Ignoring a spatial structure by assuming randomness will lead to biased and inefficient coefficients (Anselin, 1988). This makes spatial econometric techniques necessary for estimation. Nonrandomness in space can occur when, for instance, right-wing voting concentrates in a certain neighborhood resulting in right-wing voting clusters. Another advantage of spatial models is the analysis of the neighborhood effect itself. One motivation for the neighborhood effect stems from the literature review, since immigration to certain regions can improve election outcomes for right-wing parties. However, none of the literature has measured how immigration to neighboring regions affects the voting behavior. Spatial econometric models allow to distinguish between the number of refugees per 1,000 residents in a municipality, that is, the backyard effect, and the number of refugees per 1,000 residents in neighboring municipalities, that is, the neighborhood effect. We are not only interested in the backyard effect on the voting outcome but also in the neighborhood effect. Before we define the latter effect, we require a plausible definition of neighborhood. The most common way to implement a neighborhood structure in the sense of spatial econometrics is the spatial weight matrix (Drukker et al. 2013a). The spatial weight matrix defines which spatial entities are neighbors. A common way to define a neighborhood is a common border between two geographical units. If two municipalities i and j share a common border, the single element of the matrix w_ij takes a value of one or zero otherwise. The binary neighborhood relations among all spatial entities form a neighborhood structure which is defined in a contiguous spatial weight matrix W.

With

The contiguous spatial weight matrix W is an NxN matrix where the number of rows and columns corresponds to the total number of municipalities in our sample (N = 1,101). We define the diagonal elements of the matrix as zero. There are many possible definitions of neighborhood; however, we use the common border contiguity definition.

For example, w_ij can take the value of one, if two locations lie within a certain distance to another. Any other boundarymeasure can be taken as well. Another method is the implementation of an inverse distance, so that the neighbourhood is measured in terms of distance rather than connectivity.

The spatial weight matrix is the starting point for most spatial econometric models. An in-depth overview of spatial econometric models is provided by Anselin (1988), Elhorst (2010), and LeSage and Pace (2009). There are three ways to implement a spatial structure (or spatial lag) into an econometric model, namely, the spatial weight matrix in the dependent variable, in the independent variable, and in the error term. Implementation all three types of spatial lag is not feasible and leads to misspecification (Elhorst, 2010). Simply put, a maximum of two types of spatial interactions can be incorporated otherwise the estimation is not identified. LeSage and Pace (2009) argued that implementing spatial structure in the dependent and independent variables is the best way to avoid endogeneity. Ignoring spatial structure in the error term is the lesser evil because it does not affect consistency. However, modeling spatial structure in the error term can improve estimation efficiency. LeSage and Pace (2009) suggested estimating the Spatial Durbin Model (SDM) with spatial structures in the dependent and independent variables to minimize potential endogeneity. For our application, we define the SDM model as follows.

Following the literature, first, we implement a spatial lag in the dependent variable which is expressed by ∑j=1Nwijvotesharej.$\sum\limits_{j=1}^{N}{{{w}_{ij}}voteshar{{e}_{j}}}.$The corresponding spatial autoregressive parameter λ can be interpreted as the average effect of neighboring vote shares on the vote share in municipality i. Note that λ is a scalar. When λ is positive, a high vote share for a party in municipality i reinforces the vote share for the same party in neighboring municipality -i. λ is also positive, when a low vote share further weakens the vote share for the same party in neighboring municipalities. In other words, a positive λ represents the degree of spatial clustering of similar values in the dependent variable at nearby locations.

The case of a negative λ is a rather exotic and less common case.

With

Note that the KPM model does not include the neighborhood effect. r can be interpreted as a spatial heterogeneity parameter which gathers spatial variation in the error term. It improves the efficiency and can be considered a nuisance parameter. ɛ_i is an i.i.d. noise term. If there is no structure in the independent variable, but a spatial structure prevails in the dependent variables as well as the error term, the most efficient and consistent model is the Spatial Durbin Error Model (SDE).

With

For our application, the SDE model proves useful for the CDU party because we do not find significant vote share clustering for the CDU party. For the model selection, we follow the procedure proposed by the literature and estimate three models for each party. First, we estimate a simple OLS model from Eq. (1) without any spatial interactions for both parties. Second, we estimate the SDM model from Eq. (4) for the parties to check the existence of the neighborhood effect. As for the third model, we follow the model selection procedure proposed by LeSage and Pace. We estimate the KPM model from Eq. (5) for the AFD party. Because we do not find significant spatial autocorrelation for CDU vote shares, we estimate Eq. (6) for the CDU party for higher efficiency. For the estimation of the SDM, SDE, and KPM model we use the generalized spatial two-stage least squares estimator (GS2SLS). Ord (1975) and Anselin (1980) provide a detailed description of the GS2SLS estimation procedure. Kelejian and Prucha (1998) provide a detailed example of the GS2SLS estimator for the KPM model. Since we use STATA 14, we follow the estimation procedure proposed by Drukker et al. (2013b).

In this section, we highlight how the German government handles refugee allocation and how we account for non-randomness in the estimation. There is some potential for endogeneity primarily because refugee allocation is not a random process and depends on various factors itself. The asylum seeker management is a top-to-bottom system. This means that the allocation starts at the highest administrative level (federal states) is then followed by the allocation among counties, and ends at the lowest administrative level (municipalities). Note that the allocation of refugees follows different schemes depending on the administrative level. For instance, the allocation regime among states differs from the allocation among municipalities. In practice, every asylum seeker picked up at the German border undergoes a quick check by the border police followed by a registration in a federal asylum seeker database (EASY-System). The EASY-System assigns a number of refugees to each of the 16 states based on a quota, namely, the “Koenigssteiner Schluessel.” The quota allocates refugees based on the state’s population and tax revenues and ensures a rather balanced distribution of the burden.

The distribution rests to one third on the population and two thirds on tax revenues.

The regulation leaves some room for exceptions because the law allows agreements between the municipalities and the counties which can deviate from the quota. According to the Ministry of Social Policy and Integration, two main factors led to deviations from the population quota. First, municipalities that operate IECs are partially exempt from the FlüAG quota.

This regulation ensures that municipalities which operate IECs are not overburdened.

Regarding the overall regression, a low R² = 3.4% confirms that the explanatory power of the model is weak. In other words, the refugee allocation is not completely random but the degree of non-randomness is weak at best. This result is perfectly in line with the official statements on the allocation procedures: municipality population was the main factor for refugee allocation, while some municipalities were slightly affected by the number of IECs and the vacancy rate, due to the exception rules. The GDP and the political affiliation of the mayor are practically irrelevant. Although our estimates suggest that the bias caused by non-randomness is minimal, we control for the bias by adding the predicted values of the estimation from Table 2(y^)$\left( {\hat{y}} \right)$as a regressor to the estimates in the following chapter.

This article aims to study how immigration in voter’s neighborhoods affected the election outcome of the right-wing AFD party and Angela Merkel’s CDU party. We measure the influence of refugee immigration through two effects: the backyard effect, that is, the number of refugee accommodations per 1,000 residents, and the neighborhood effect, that is, the number of refugee accommodations per 1,000 residents in neighboring municipalities. The coefficients a₁ and a₂ represent the magnitude and direction of both effects respectively. The estimates for the AFD party are summarized in Table 3.

Regarding the first coefficient a₁, we find strong evidence for the backyard effect on AFD vote shares. All three models indicate a positive influence of refugee presence on AFD vote shares. The magnitude of the coefficients varies slightly between 0.59 and 0.67. The direction of the effect is clear: A higher refugee presence significantly boosts AFD vote shares. To be precise, one additional refugee accommodation per 1,000 residents increases the AFD vote share by 0.6 percentage points on average.

The true coefficients in the SDM and KPM model will vary slightly since we have to account for a positive autoregressive structure (λ = 0.007 – 0.010) in the dependent variable.

Regarding other policy variables, we find a strong positive correlation between the 2013 election outcome and the 2016 election outcome for both parties. Regions that had a high vote share for either party in the previous election have a similarly high vote share for the same party in the recent election. This indicates some regional and temporal stickiness of party preferences. For both parties, the control variables for the previous election, AFD 13 and CDU 13, exhibit the highest t-values among all variables. The previous election is the driving force for the high coefficient of determination for the CDU party. It explains nearly 80% of the CDU vote share in 2016. Regarding the election turnout, we find that AFD voting is more pronounced in regions with traditionally high voting abstinence. In 2016, the AFD gained more votes in regions that had a low turnout rate in the 2013 election. On the contrary, CDU vote shares are higher in regions with low voting abstinence. This result indicates that right-wing voting is an attractive option in regions with traditionally higher voting abstinence. The AFD is often considered a “catch all” party for protest voters. One alternative for protest voting is voting abstinence. Hence, it is possible that the party mobilized a significant share of nonvoters from 2013, which voted for the AFD party in 2016.

We find some interesting links between the socioeconomic environment and the election outcome. Our results show that right-wing voting is closely related to unemployment. This result is supported by the aforementioned literature. Among all models, the unemployment rate is positively correlated with AFD vote shares at the highest significance level.

Unemployment has also a relatively high coefficient. On average, one additional percentage point in the unemployment rate increases the AFD vote share by 0.87 to 1.08 percentage points. With regard to the municipality output, we find no significant relationship between regional GDP an AFD votes. Although GDP has a positive effect on AFD vote shares in the standard model, the OLS estimator is likely biased (again), since it does not implement spatial structures. The better fitting models, SDM and KPM, show no significant influence of GDP on AFD vote shares. In contrast, the CDU party is unaffected by regional variation in unemployment. However, it exhibits significantly lower vote shares in wealthier regions with high GDP. Regarding other demographic factors, our estimates show that the AFD party obtained more votes in municipalities with a certain demographic milieu. The AFD party was more successful in municipalities with younger age, a higher share of married couples, and a lower level of education. Moreover, the AFD received more votes in regions with higher population density. Surprisingly, the AFD election outcome is unaffected by the share of residents with immigration background. This group consists of people who live in Germany for a longer period of time, and hold German citizenship which makes them entitled to vote. Due to the anti-immigration rhetoric of the AFD party, one could expect that Germans with migration background might averse the party’s position and vote against it, which would result in fewer AFD votes. There is some evidence that refutes this presumption and indicates that some migrants might be strong proponents of the AFD party.

An atricle in Foreign Policy highlights this issue, “Immigrants Are Big Fans of Germany’s Anti-Immigrant Party” from15.01.2020.

With regard to the CDU vote shares, there is no clear-cut demographic environment. The only significant demographic variable is the density of married people which is negatively correlated with CDU vote shares. Variation in age, gender, migration background, and population density does not affect the CDU election outcome significantly. We find some evidence for a positive correlation between a low education level and CDU votes in the SDE model. Overall, the results for the CDU party are a little surprising, since the CDU is Germany’s center party which attracts voters from various social groups. We also do not find a spatial clustering for the CDU vote shares in 2016, which indicates that CDU voters do not cluster in certain regions. Finally, some remarks on the spatial model parameters. The Moran’s I test for spatial autocorrelation finds significant spatial autocorrelation in some of the independent variables as well as the error term. In addition, we find a positive autocorrelation of the dependent variable for the AFD party. Our estimates confirm the existence of spatial autocorrelation. Hence, we conclude that the OLS model is the weakest performing model among the estimates and all estimated spatial models are more consistent and efficient.

This article studies the short-run impact of immigration on right-wing voting behavior. We use the refugee crisis from 2015/2016 as a natural experiment to study the influence of a large and unexpected immigration influx on the voting behavior in German municipalities during the 2016 election. We model the influence of immigration, directly and indirectly, through two channels, that is, the backyard effect and the neighborhood effect. The former effect describes the impact of the number of refugee accommodations per 1,000 residents on the election outcome while the latter effect describes the impact of refugee accommodations per 1,000 residents in neighboring municipalities. Our estimates for the right-wing AFD party are straightforward—the party benefits from the immigration influx since the backyard effect is positive and significant across all models. We find no evidence for the neighborhood effect on AFD vote shares. Our results support the hypothesis that a significant share of the population opposed refugee settlement in their home municipality by voting for the right-wing AFD party. We suppose that voters in municipalities with larger numbers of refugee settlements found the AFD party program more appealing since it opposes refugee settlement in Germany and criticizes Merkel’s refugee policy. With regard to the massive loss of the CDU party in the election, it is surprising that local refugee presence did not directly affect CDU vote shares. Hence, other factors have caused the decline of Angela Merkel’s CDU. According to Haller (2017), the media coverage about the refugee crisis was one major factor that caused CDU vote shares to plummet. He argues that the subjective perception through media consumption was rather important to form an opinion about the matter and stimulated disagreement with the refugee-welcoming policy of the CDU. We do not prove Haller’s argument in our work, however, it is a potential extension to our work in the future. Although we do not find evidence for the backyard effect on CDU votes, we do find some evidence that the neighborhood effect negatively affects CDU vote shares. This is an interesting result because it implies that direct exposure to immigration does not decrease CDU votes while indirect exposure does. The contact hypothesis might partly explain this observation. The hypothesis states that social interaction with immigrants can reduce prejudice and hostility. Having few migrants in the “backyard” while having many immigrants in the “neighborhood” describes a situation where social interaction between natives and immigrants is potentially limited. Note that AFD and CDU voters are two different population segments. Contrary to AFD voters, CDU advocates could be more open to intercultural contact with refugees, which could reflect in an insignificant backyard effect but a significant neighborhood effect. According to the contact hypothesis, intercultural contact can induce empathy and acceptance among different cultural groups. This would explain the insignificant backyard effect and the significant negative neighborhood effect on the CDU. In other words, CDU advocates, who experience refugee presence in their hometown, also have a higher chance to become familiar with refugees and develop empathy and tolerance, while refugee presence in neighboring municipalities physically limits intercultural contact. This would ultimately explain why the CDU received fewer votes in regions which neighbor municipalities with many refugee accommodations. This behavior is also in line with the CDU party programmer of 2016, which has a tolerant and integrative stance toward refugees. Furthermore, we find that CDU obtained more votes in regions that had higher turnout rates in the previous elections. In contrast, the right-wing AFD was more successful in regions, which had low turnout rates in the previous election. This indicates that the AFD party was more attractive in regions that previously had higher voting abstinence. Another interesting result is the spatial clustering of AFD voters. The SDM and KPM models show that regions with high AFD vote-shares at the 2016 election are likely to be neighbor to other regions which prefer the AFD likewise. This is an interesting observation because it raises the question of why right-wing voting is spatially clustered. There seems to be some sort of “contagion effect” of right-wing voting among neighboring municipalities. This result leaves some space for further research. However, one potential explanation could be the public perceptibility of the AFD in regions where the AFD is strong. Regions with higher AFD vote shares have more AFD supporters and therefore, more financial means and manpower. This allows more party advertisements and a professional structure. Since the AFD party is a new party, it is all the more important to make a professional impression on voters, which is easier achievable with solid party funding. We assume that the “professional” AFD presence is perceived in the neighborhood which attracts voters in neighboring municipalities. It would be interesting to study the role of party funding on election results in a spatial context to check this hypothesis.

The analysis of the socioeconomic environment shows that unemployment is a major driving factor for right-wing voting. The AFD benefitted in regions with higher unemployment rates. A large share of the literature on that matter points out that right-wing voting is often reinforced by a high unemployment rate. With regard to the demographic factors, we find strong evidence for AFD support in municipalities with high degrees of married couples, low education, and lower mean age. In light of German history, a recent study shows that political extremism is the second largest concern among Germans in 2015 (RAV, 2016). We presume that the fear of political extremism is pronounced more strongly among older voters. Hence, older generations might be less willing to vote for right-wing parties. Another surprising result is the seemingly indifferent stance toward right-wing voting in regions with many Germans with immigration background. We cannot say whether immigrants in Germany have the same stance toward “new” immigrants as native Germans do or whether native Germans’ voting behavior is not affected by the presence of Germans with immigration background. As a conclusion, the finely scaled spatial methodology proofed to be useful to study the influence of local immigration shocks and local socioeconomic conditions on voting behavior since it outperformed the non-spatial model. It could serve as a useful template for future policy evaluations of various local shocks and spatial clustering of voting behavior.