Do Refugees Impact Voting Behavior in the Host Country? Evidence from Syrian Refugee Inflows in Turkey

Transcription

1 Do Refugees Impact Voting Behavior in the Host Country? Evidence from Syrian Refugee Inflows in Turkey Onur Altındağ Neeraj Kaushal September 21, 2018 Abstract We study how individual political preferences responded to the influx of over 3.5 million Syrian refugees in Turkey during Using a difference-in-differences research design, we compare the political outcomes in geographic areas with high versus low intensity of refugee presence before and after the beginning of Syrian Civil War. To address the endogeneity in refugees location choices, we adopt an instrumental variables approach that relies on the historic dispersion of Arabic speakers across Turkish provinces, taking advantage of the fact that Syrians are more likely to settle in locations with a high percentage of Arabic-speakers in the host population. We find a strong polarization in attitudes towards refugees between the supporters and opponents of the ruling Justice and Development Party (AKP). However, regression analyses of monthly survey data suggest that the massive inflow of refugees induced only a modest net drop in support for the AKP. We show a similarly small, but statistically insignificant impact on election outcomes. JEL Classification: D72, F22, and O15 Keywords: refugees, political preferences, voting. Bentley University, Economics Department and Harvard Center for Population and Development Studies. Corresponding author. oaltindag@bentley.edu Columbia University, School of Social Work and National Bureau of Economic Research. We would like to thank KONDA Research and Consultancy for the data support. We received valuable comments from Mathis Wagner, David Canning, Aytuğ Şaşmaz, Stephen O Connell, Selim Erdem Aytaç, Murat Kırdar, Yulya Truskinovsky, Donald Halstead, and Theodore Joyce. Thanks to seminar and workshop participants at Kings College, Oxford University, Harvard Center for Population and Development Studies, Columbia Population Research Center, Galatasaray University, Koç University, Bentley University, Eastern Economics Association, and Population Association of America for their comments. 1

2 I Introduction Recent years have seen an astounding increase in the refugee population fleeing wars and conflicts (UNHCR, 2017). In 2017, there were 66 million forcibly displaced people worldwide, of whom 23 million were refugees, representing a 48 percent increase since Roughly half the global refugee population currently lives in six countries neighboring the regions of conflict, million of whom reside in Turkey alone. In addition to the burden of resettlement, for which these host countries have limited resources, such large-scale refugee inflows impose economic, social, and political costs that are likely to influence the political will and administrative capability of host societies towards refugee resettlement and integration. In this paper, we study the effect of the Syrian refugee influx on voting behavior in Turkey, which currently hosts the largest refugee population in the world. As of October 2017, there were more than 3.5 million Syrian refugees registered in Turkey, accounting for approximately 4.4 percent of its population. However, little is known about the impact of the refugee influx on political preferences of the Turkish people, an important issue that is likely to influence Turkey s ability and commitment to resettle the vast population of Syrians already in the country and may even have consequences on European politics. 2 Getmansky et al. (2018) is the only study, we know of, that has focused on public attitude towards refugees in a causal framework. They study public perceptions of the refugees and whether 1 According to UNHCR, in 2016, around half of the global population of refugees was hosted in Turkey, Pakistan, Lebanon, Islamic Republic of Iran, Ethiopia, and Jordan (UNHCR, 2016) 2 Turkey has been using Syrian refugees resettlement as a bargaining chip in its negotiations with the E.U. The E.U. in turn has given Euro 6 billion to Turkey to aid the refugee resettlement. 2

3 messages (positive or negative) about the possible effects of hosting refugees affect these perceptions using a survey-experiment they conducted in the summer of 2014 in Turkey. Our study provides novel evidence on the impact of refugee inflows on voting behavior using data from repeated monthly cross-sectional surveys of Turkish citizens political preferences during and the results of three recent national elections in June 2011, June 2015 and November Refugee migration from Syria to Turkey is empirically appealing because (i) the population movement between Turkey and Syria was limited before the Syrian Civil War, rendering both the timing and the scale of migration between two countries plausibly exogenous, and (ii) the regional and within-time variations in refugee resettlement are substantial, which allows an empirical comparison of Turkish provinces with varying levels of refugee presence during the Civil War. Using a difference-in-differences approach, we compare the political outcomes in geographic areas with high and low intensity of refugee presence before and after the beginning of Syrian Civil War. To address the endogeneity in refugees location choices, we adopt an instrumental variables approach which relies on the historic dispersion of Arabic speakers across Turkish provinces based on the 1965 Turkish census, the most recent census that collected information on native language at the province level. Our instrument relies on the fact that Syrians are more likely to settle in locations where the host population is more likely to speak Arabic. We provide direct evidence on the parallel trends assumption that is crucial for causal identification, showing that the trends in voting behavior were similar in regions with varying intensity of Arabic speakers before the civil war intensified in Syria. 3

4 We first document a strong polarization in attitudes towards refugees between the supporters and opponents of the ruling Justice and Development Party (AKP), the architect of the open door policy for Syrian refugees. The empirical analysis of political preferences, however, suggests that the massive inflow of refugees induced a modest drop in support for the ruling AKP. A one percentage-point increase in the share of refugees in the population led to a 0.68 percentage point decrease in support for AKP. Those who left the AKP, moreover, did not swing to the other major political parties but turned indecisive or expressed unwillingness to vote. We additionally show that the massive refugee inflow did not have any impact on Turkish election results between 2011 and These results are robust to different empirical specifications and data restrictions. II Background II.1 Migration and Electoral Behavior Research on the effect of migration flows on voter preferences has primarily focused on western countries and leans towards the general finding that migration flows strengthen anti-immigration attitudes. Barone et al. (2016), for instance, find that migration flows induced voters in Italian municipalities to favor right-wing parties with a more conservative agenda on migration. Otto and Steinhardt (2014) show that in Hamburg, an increase in the number of immigrants and asylum seekers leveraged support for the German far-right whereas in Denmark, Harmon (2017) and Dustmann et al. (2016) show that increased inflows of immigrant workers and 4

5 refugees strengthened the native population s anti-immigration attitudes in political preferences. In a recent study, Vasilakis (2017) show that Syrian refugees triggered a similar sentiment among Greek voters, and increased support for the Golden Dawn, an extreme-right movement with neo-nazi tendencies. Evidence from Austria is mixed: Halla et al. (2017) find that immigration of unskilled workers contributed to the rise of the Austrian extreme right while Steinmayr (2016) argues that the recent Syrian refugee inflows weakened the political support for the same far-right movement. Hangartner et al. (2017), on the other hand, find that the native communities who are exposed to a large number of Syrian refugees in Greek islands have become more hostile towards them as well as Muslim minorities. There are reasons to expect the impact of refugee inflows to be different in host countries in the conflict regions. First, refugees most often arrive in neighboring counties in highly vulnerable conditions, which may make the host community more sympathetic to refugees and therefore minimize the adverse public reaction to their presence than is the case in countries of final resettlement. Second, refugees often share ethnic and religious backgrounds and histories with the host population in neighboring countries. The social interaction between the host and displaced people therefore might increase the trust and sympathy towards the latter, as predicted by the intergroup contact theory (Allport, 1979). Anecdotal evidence suggests that in predominantly Muslim host countries within the MENA region, the ethnic and sectarian composition of the refugee population are central to the acceptance policies of neighboring countries. 3 Refugee migration within the neighboring country 3 See for example Achilli (2015) for Jordan, Arar et al. (2016) for Egypt, and Turner (2015) for Lebanon. 5

6 of resettlement is, on the other hand, usually unregulated and results in geographic clustering of the refugees, which threatens to change the ethnic balance of resettled communities. Clustering combined with arrival of a large number of displaced individuals within very short time intervals might create the perception that refuges are an economic and national security threat. (Stephan and Stephan, 2017; Hainmueller and Hopkins, 2014). Our study differs from the existing literature on several fronts. First, we examine the political impact of refugee inflows between two neighboring countries with largely Sunni Muslim populations and a shared history. Second, we study one of the largest refugee influx since the World War II, a scale that is substantially larger than any of the cases mentioned above. Finally, we analyze the impact on a developing country with limited capacity to smooth the resettlement of the refugee population. Existing literature has solely focused on western countries with relatively well-functioning migration systems and strong state capacities to manage refugee flows than developing countries. Our study is important because it investigates the most prevalent scenario of forced displacement where a large number of refugees, and often the most vulnerable ones, settle is in a developing country, neighboring regions of civil conflicts that create the refugees, with little administrative and economic capacity to absorb them. II.2 Syrian Refugee Inflow Syrian refugees began entering Turkey in April 2011, shortly after the Syrian government violently cracked down on anti-government protests (Erdoğan and Ünver, 6

7 2015). In 2011, a majority of the refugees who left Syria were politically active youth on the government s black list, and many of them returned to Syria as conditions stabilized temporarily (Özden, 2013; İçduygu, 2015). The refugee influx began again in 2012, when ceasefire talks between the Syrian government and the opposition failed and has continued over the past five years (figure 1). In March 2012, the Government of Turkey announced a directive on resettlement of Syrians in Turkey, which was enacted as a Temporary Protection regime in April It assured no forced return and allowed temporary stay for all Syrian refugees. From April 2011 to December 2016, Turkey received over 2.8 million of the 4.8 million refugees who fled Syria; this number is expected to remain high as the Syrian Civil War continues, and with the European Union imposing pressures on Turkey to restrict the flow of refugees into Europe. 4 While the issue of providing work permits to Syrian refugees continues to be debated, in 2014 refugees were granted permission to receive education and health care (Yeginsu, C., 2014). According to the United Nations High Commissioner for Refugees (UNHCR), the average monthly inflow of refugees was roughly 13,000 in 2012, 32,000 in 2013, and 88,000 in Inflows reached a peak in October 2014 after the siege of Kobane by the Islamic State, and declined to 73,000 in 2015, and to 23,000 in 2016 (figure 1). The refugee flows show significant regional patterns, changing the ethnic and sectarian balance of the population in the eastern and southeastern regions of Turkey, where many of the refugees are settling (figure 2). Sunni Arab refugees, for instance, now outnumber Alawites who previously dominated the ethnic Arab 4 In December 2015, the European Union and Turkey reached a deal that involved the E.U. providing $3.2 billion in aid for refugee resettlement and Turkey promising to take action to reduce the flow of migrants from the Middle East to Europe (Economist, 2015). 7

8 population in many of these regions. These demographic changes have created fears that refugee resettlement would sow seeds of ethnic and sectarian strife in eastern and southeastern regions of Turkey (Cagaptay and Menekşe, 2014). A substantial proportion of the refugees initially lived in camps that the Turkish government had built to provide them temporary settlement. However, as the influx increased, it became increasingly difficult to accommodate the refugees in camps, and they began to move out of the southern provinces and into larger cities. As a result, less than ten percent of the refugees live in the camps today. Thus, whereas Turkey initially welcomed the incomers as guests, more recent reports indicate that the Turkish government is engaging in efforts to resettle refugees within Syria (İçduygu, 2015). II.3 Refugee Politics There are four national parties represented in the Turkish parliament: President Recep Tayyip Erdoğan s Justice and Development Party (AKP), which has an authoritarian, pro-islamist, socially conservative, and economically liberal political agenda; the center-left and secular Republican People s Party (CHP); the extreme nationalist Nationalist Movement Party (MHP); and the socialist and pro-kurdish Peoples Democratic Party (HDP). The electorate is consolidated around these four political parties. In the last three general elections, merely 3.5 percent of votes went to other political parties. Voter turnout in Turkey was substantially high during these recent elections with more than 84 percent of the eligible voters casting their votes. AKP has dominated Turkish politics since 2002, securing nearly half 8

9 of the electoral support in the last three general elections: 49.8 percent in June 2011, 40.9 percent in June 2015, and 49.5 percent in November The party has no coalition partner and has been the sole decision-making authority on Turkish foreign policy since Among political parties, AKP has maintained the most explicit policy with respect to refugees. In his speech to refugees in 2014, Erdoğan summarized Turkey s role as being ansar, a historical reference to the people of Medina who voluntarily supported the Prophet Mohammad and his followers after fleeing Mecca. 5 More recently, he promised citizenship to a number of the Syrians who have permanently resettled in Turkey. 6 On the other hand, CHP considers the refugee crisis an outcome of AKP s foreign policy mistakes, including attempts to topple the Syrian regime. In its policy report on Syria, CHP emphasizes the economic cost of an adventurist foreign policy but does not provide a specific framework for refugee resettlement (Cumhuriyet Halk Partisi, 2015). MHP believes that Turkey must host the Syrians refugees who fled the war yet heavily criticizes the government on national security concerns resulting from the government s foreign policy towards Syria. HDP views the Syrian refugee question from a human rights perspective: its policy report on Syrian refugees recommends a change of status for refugees from Temporary Protection to equal permanent residents, abolition of refugee camps as resettlement zones, as well as a number of policies for the social and economic integration of refugees (Halkların Halk Partisi, 2015). Among the four major political parties, CHP stands as the most outspoken critic of Syrians in Turkey. In short,

10 Turkey s major opposition parties have not taken an explicit anti-refugee stance nor do they have a common view of refugees that distinguishes them from the AKP, which is a remarkable difference from the position taken by political parties across Eastern and Western Europe. The civil war in neighboring Syria and the refugee influx have impacted Turkey s geo-political influence and its economy. Both of these factors are strong correlates of voting behavior (Çarkoğlu and Ergen, 2002; Çarkoğlu, 2002; Akarca and Tansel, 2006). Both Ceritoglu et al. (2017) and Del Carpio and Wagner (2015), for example, show that during our study period, the refugee workers lowered the labor cost and replaced Turkish workers in the informal labor market. More significantly, the fear that the refugee influx compromises national security might have caused voters to decrease their support for the AKP, even more so in areas where the refugees migrated. Kibris (2011), for example, provides evidence that during the early 1990s, instances of terrorist attacks in Turkey caused a decline in support for the governing party. Yet another factor that may influence public opinion is the media coverage of refugee issues. Yaylacı and Karakuş (2015) find that the national newspapers that constitute the overwhelming majority of newspaper readership in Turkey, provide frequent yet biased coverage of Syrians depending on their level of support for the government. The major television outlets in Turkey are under the strict control of the government and unconditionally support the government policy on refugees. The impact of national media is likely to mitigate any negative impact of local exposure to refugees. 10

11 III Data The data used in the empirical analysis come from multiple sources that we describe in detail below. III.1 Data on Refugees We obtained data on the number of refugees who fled Syria and those who resettled in Turkey from the UNHCR. The UNHCR began reporting these data in December 2011 and has continued at frequent but irregular intervals. We aggregated the data into a monthly format. The data on refugee population by province for April 2016 and December 2016 come from the Directorate General of Migration Management (GIGM) and Disaster and Emergency Management Authority (AFAD). We also use the geographic distribution of refugees complied by Erdoğan and Ünver (2015) from the same institutions for September As shown in figure 3, the geographic dispersion of refugees is remarkably similar in September 2015, April 2016, and December In each panel of this figure, we plot the dispersion of refugees across provinces in September 2015 on the x-axis. The y-axis in the top panel shows the refugee dispersion in April 2016, and in the bottom panel, it shows the refugee dispersion in December The nationwide refugee population is normalized to 100 in each period. Each circle indicates a province, and is sized proportional to the number of refugees who live in that province. The regression line shows the linear trend weighted by the refugee population. The circles in scatter plots in both panels are clustered remarkably close 11

12 to the 45-degree line as well as the regression line in the top and bottom panels. Put another way, the provinces that host the majority of refugees did not change even though around 725,000 additional refugees arrived from September 2015 to December We estimated the province level monthly refugee inflows by multiplying the nationwide refugee inflow at each period with the share that each province received in September Note that data on refugee population at province level covers only three periods and these numbers merely reflect the cumulative number of Syrians who are registered within the administrative boundaries of a province. Refugee population is highly mobile, and some of the refugees might have left the province or even the country after registration. As a result, data on province-level refugee population could at best serve as a proxy for the intensity of refugee presence. Later, we directly test if the data on geographic dispersion of refugees offers a meaningful variation in native population s exposure to refugees. To calculate the refugee exposure measure, defined as the share of refugees in total population, we used the population of Turkish citizens for , provided by the Turkish Statistical Institute (TurkStat). We interpolated the monthly province populations from the yearly data. We digitized the aggregate data from the 1965 Turkish Census to calculate the percentage of the Arabic native population at the province level. The 1965 Census, conducted in an early stage of the rapid urbanization and mass internal migration to major Turkish cities, provides a reliable proxy for the historical ethnic/linguistic distribution of the population (Gedik, 7 Erdoğan (2014) also provides similar estimations for December As expected, results from the regression analyses are similar irrespective of whether we use the geographic distributions in April or December

13 1997). Later censuses do not include data on mother language or ethnicity. In 1965, Turkey was divided in 67 provinces, the smallest administrative unit in which we can conduct the analysis due to data availability. Over time, additional provinces were carved out raising the total to 81. In our empirical analysis, we assigned the Arabic speaker population density based on the administrative units of To check the sensitivity of the results, we conduct the empirical analysis based on the sample of 67 provinces in 1965 and find that the results are similar, as shown in the empirical section. III.2 Survey on Attitudes towards Refugees Data on local perceptions of refugees come from a field survey of a nationally representative sample of 2,649 respondents aged 18 and above. The survey was conducted in February 2016 by Konda Research and Consultancy, a leading and independent research company in Turkey. It included a range of questions to capture views on the economic and social effects of the refugee presence, government policies towards refugee inclusion in the Turkish society, and overall attitudes towards Syrian refugees. Based on a six-point Likert scale, the respondents ranked the degree to which they agreed or disagreed with a series of statements detailing these attitudes. The survey also asked whether the respondents voted for the ruling party in the most recent general election, held in November For example, Aksaray was a district of Niğde until becoming a province in We assigned the same percentage of Arabic speaking population for both. 13

14 III.3 Survey on the Political Affiliation of Turkish Citizens Konda also conducts a monthly survey to capture the political affiliation of Turkish citizens. These surveys have been quite accurate in predicting the general election results in Turkey since the early 2000s (Akarca et al., 2009; Dağı, 2008). We used 54 monthly surveys 10 conducted from January 2012 to December 2016 that included 149,746 individual observations. The survey instrument included the following question: Which party would you vote for if the elections were held today? The respondents can select a political party, abstain, or express their indecisiveness about the elections. We used the answer to this question to create four binary variables for each of the four major parties that have seats in the current parliament and a fifth variable that captures responses that are in favor of other minor parties, indecisive, or absentee. Due to consolidation of voters around the four main political parties during the study period, the overwhelming majority of the voters in the last category is typically composed of indecisive or absentees. 11 III.4 Data on Election Results We collected the June 2011, June 2015, and November 2015 General Election Results, which are publicly available through TurkStat. These results include provincelevel data on number of voters, votes, and valid votes that each party received. The Turkish electoral system is based on proportional representation that requires 10 Konda did not conduct surveys in some months, which usually corresponds to Ramadan. Thus, we do not have data on 6 months out of 60 months between January 2012 and December Due to the format that Konda provides data, it is impossible to create an exclusive category of indecisive and absentee voters for all the survey months that we have data. 14

15 a ten percent minimum vote share at the national level to secure representation in parliament. The threshold does not apply to independent candidates. Therefore, candidates from political parties who do not expect to receive the ten percent vote share can participate in elections independently. The HDP candidates followed this strategy in the June 2011 General Elections, but not in the subsequent two elections. Due to the data structure, 12 it is not possible to separate the votes for HDP and other independent candidates in the June 2011 election results. The non-hdp independent candidates, however, constituted only 0.57 percent of the vote share in the two following elections. Given the relatively weak support for the non-partisan independent candidates, we combined the HDP and independent votes for all election results. For convenience, we use the term HDP to describe the outcome for this group. IV Identification Strategy Our empirical strategy exploits changes in the Syrian refugee population over time and across provinces to study the effect of refugee influx on voting behavior. The second stage specification is given by: y ijt = α 1 + δ 1j + γ 1t + X iγ + τ 1 S ref jt + ɛ 1ijt (1) where y ijt is a binary indicator of respondent i s (living in province j and interviewed in time t) support for a specific political party (for example AKP). The 12 TurkStat only reports the aggregated number of votes for all independent candidates. 15

16 parameters δ 1j and γ 1t respectively capture the fixed effects for each province and survey month.x i is a vector of individual covariates that include respondent s sex, age, education, ethnicity, whether the respondent is a Sunni Muslim, and whether she considers herself religious, and indicator variables for missing observations for each of these variables. S ref jt is a continuous measure of treatment intensity, defined as the share of refugees in province j s population in period t. The coefficient τ 1 captures the effect of a one-percentage point increase in refugee influx on party affiliation. There is no restriction on the location choice of Syrians who live outside the refugee camps. Because more than 90 percent of refugees live outside camps, the geographic variation in refugee resettlement is potentially endogenous. For example, districts or provinces with booming economies may have attracted refugees as well as influenced political preferences in favor of a certain political party (for instance the ruling party) that voters consider responsible for the robust economic growth. Therefore, regression estimates based on equation 1 could be biased. Moreover, as mentioned before, the estimate of S ref jt involves a potentially large measurement error, as refugees often register at one place and then move to another where they find better opportunities or they may even register in multiple provinces. We use a precisely measured instrument to overcome these threats to the internal validity of the empirical design. Syrian refugees in Turkey are more likely to settle in areas with a historically higher proportion of Arabic speakers. The population share of Arabic natives among Turkish population in 1965, as we document, strongly predicts the geographic distribution of refugees. Further, the total refugee outflow from Syria is plausibly exogenous to Turkish electoral behav- 16

17 ior/preferences given that the intensity of conflict in Syria has remained largely unpredictable throughout the civil war. Thus, we interact the percentage of Arabic native Turkish population in 1965 with the cumulative number of refugees who fled Syria in period t as an instrument to predict S ref jt. The first stage regression is: S ref jt = α 2 + δ 2j + γ 2t + X iλ + π 2 iv jt + ɛ 2ijt (2) where the instrument iv jt is the percentage of Arabic native Turkish population in 1965 weighted by the cumulative number of refugees who fled Syria at period t, and expressed as: iv jt = Arab 1965 j S reftot t (3) The refugee inflow into provinces in Turkey is jointly predicted by: (1) S reftot t, the overall outflow of refugees from Syria in time t, and (2) Arab 1965 j, the share of Arabic-speaking Turkish population in province j in The identifying assumption is that after controlling for both components through province and survey month fixed effects, their interaction is as good as random. We cluster standard errors at the province level to capture the within-province correlations in voter behavior. Our instrument differs from a typical shift-share instrument proposed in Altonji and Card (1991) and Card (2001), who use geographic variation in the past settlement patterns of immigrants to predict future inflows of immigrants. Before the Syrian conflict, migration from Syria to Turkey was negligible. Thus the Arabic 17

18 speaking population in Turkey that generate the geographic variation in our instrument are not Syrian immigrants but rather natives of Turkish Republic which was founded following the collapse of the multi-ethnic Ottoman Empire in Also note that it is not possible to purge the effect of a refugee influx on internal mobility. Previous studies suggest that the refugee influx altered the internal migration patterns of the native population, and therefore, might also have altered the composition of the electorate (Akgündüz et al., 2015). Our analysis captures the impact of the refugee influx on the overall voting pattern, which is a combination of its direct effect on voting behavior as well as its indirect effect via changed composition of native residents due to internal migration. We supplement our empirical analysis with the actual election data for the three most recent elections in June 2011, June 2015, and November We accordingly modify the specification in equation 1 to fit a model for five outcomes: the vote share of each of the four major parties and voter turnout by province and election. The regressions control for province- and election-fixed effects, as well as a vector of time-varying voter characteristics: percentage of voters under 40, percentage with a primary school degree, and share of female voters in each province. These control variables aim to account for, among other things, the increasing share of young voters from 2011 to 2015 who were exposed to the compulsory schooling law and were more likely to graduate from secondary school or above, which might have influenced their voting preferences (Cesur and Mocan, 2018; Gulesci and Meyersson, 2012). As in the previous analysis, the standard errors are clustered at the province level. 18

19 In this research design, the causal interpretation of the results requires, first, the trends in voting behavior across provinces with varying 1965 Arabic speaking population intensity to be similar in the early stages (when the refugee flow is negligible) and diverge with the increasing intensity of flow. We use dynamic differencein-differences regressions to directly observe the validity of this assumption in the empirical section. Second, the differential impact of conflict intensity on provinces with a historically larger ratio of ethnic Arabs should impact the voting preferences exclusively through refugee inflows and not the conflict itself. This is a restrictive assumption given that most of the cross-sectional variation in our sample is driven by provinces near the Syrian border, which might have differentially impacted by the conflict. Under the assumption that the conflict had an adverse effect on the likelihood of supporting the AKP government, the Wald estimator would be upward biased. As a robustness check, we tackle this threat to interval validity by excluding the border provinces from our analysis. As later discussed in detail, the empirical findings remain qualitatively similar with the exclusion. V Results V.1 Descriptive Analysis and Voter Perception of Refugees Table I presents data on the Syrian refugee population as of December 2016, in and outside refugee camps. We report the intensity of refugee presence (share of refugees in total population) in the top 19 provinces with the highest refugee presence and the rest of Turkey. Figure 2 provides a graphical presentation of the same 19

20 data on a province level map. There are a few points to note. First, refugee density is higher in provinces in the southeast of Turkey, specifically those closer to the Syrian border. Second, refugees are more likely to move to industrial areas and coastal provinces with established immigrant smuggling networks such as Mersin, İzmir, İstanbul, and Bursa (Tinti and Reitano, 2016). Overall, refugees are a little over 6 percent of the population in the 19 provinces listed in table 1. In the rest of the country, refugee presence is modest by comparison, representing about 0.5 percent of the population. We first use data from the public opinion survey on Syrian refugees to investigate if there is evidence that citizens opinions of refugees are associated with (i) their electoral preferences, and (ii) the intensity of refugee inflow in their provinces. Note that this is only a one cross-sectional survey and therefore cannot use our main identification strategy to causally study the effect of refugee influx on public attitude towards them. The first two columns in table II show the overall attitudes towards refugees of Turkish citizens who voted for the ruling AKP and those who did not. The third column reports the difference between columns (1) and (2) adjusted for a large set of demographic variables as well as location fixed effects. 13 These statistics show that voters are highly polarized on their perception of refugees. Individuals who mentioned that they voted for the ruling party in November 2015 were less likely to have negative views on the effects of refugees on the 13 The adjusted difference controls for respondents gender, age, education level, ethnicity, whether the respondent considers herself as religious, and whether she is Sunni Muslim, plus indicator variables for missing observations for each of these variables. Location fixed-effects are captured by dummy variables for each district. 20

21 economy and society than those who did not report voting for AKP by large margins. For instance, they were 13 percentage points less likely to agree that job opportunities decreased due to Syrians, 11 percentage points less like to agree that refugees harm the Turkish economy, and 11 percentage points less likely to agree that refugees make cities less safe compared to those who did not vote for the ruling party. The differences in perceptions are similar across all survey questions except on the question that, refugees should integrate to daily life, for which there is no statistically significant difference between the two groups. These differences, moreover, are mostly driven by the extreme points in the Likert scale; that is, non-akp voters are more likely to strongly disagree with a statement sympathetic towards Syrian refugees, and vice versa (figure 4). Importantly, the last row in table II shows that AKP supporters are equally likely to come across refugees in their daily life as compared the rest of the population. This suggests that the stark differences in voter attitudes towards refugees are not driven by differential exposure to refugees. These findings overall suggest that Turkish citizens, similar to other countries, are highly polarized in their perceptions of Syrians, based on their existing party affiliation and independent of where they live. The rows of column 5, table II, contain coefficients from separate regression estimates in which the question listed in the row heading is the dependent variable. The regressions control for a rich set of demographic characteristics and we report the coefficient on the refugee share in province population. The reported coefficient in each cell shows the association between a one percentage-point increase in the refugee share in the population and the likelihood of agreeing with the statements 21

22 on refugees listed in the row heading. Exposure to refugees is associated with negative attitudes, especially on issues related to labor market and national security, which reveal perceptions on the effect of refugee influx on issues directly linked to the lives of the local population. For example, respondents who live in high-intensity refugee areas are substantially more likely to argue that refugees reduced the number of available jobs and express concerns on security. But perceptions on the humanitarian responsibilities towards refugees, such as Turkish aid to Syrians is sufficient, Syrians should be granted a residence permit, Syrian children should receive education in Arabic, and Accepting refugees is a geographic/historic responsibility do not seem to vary with the intensity of refugee presence in the province. Remarkably, a one percentagepoint increase in the share of refugees in total population is associated with a 2.5 percentage-point increase in the likelihood of daily encounter with refugees. This last estimate shows that despite the potential measurement error in the refugee population at province level, our treatment intensity variable clearly captures the differences in exposure to the refugee population. Overall, a substantial proportion of the national population (44 percent) experience the presence of refugees on a daily basis. Further, public opinion towards refugees is strongly divided along political lines and to some extent by geographic proximity to refugees. In the next sections, we investigate whether these observed associations imply a causal effect of refugee presence on political affiliation. 22

23 V.2 Effect of Refugees on Voting Behavior Table III the OLS and 2SLS estimates of the effect of refugee inflows on voter behavior, based on equation 1. OLS estimates from the survey data (panel A) suggest that a one-percentage point increase in refugee presence is associated with a 0.44 percentage point drop in AKP s vote share. There is no detectable impact on the voting share of other political parties. The instrumental variable estimate is slightly larger and shows a 0.68 percentage point decline in the ruling AKP s vote share in response to a one-percentage point increase in refugee presence. The refugee influx has no effect on the electoral fortunes of the other three major opposition parties in the OLS and 2SLS models. The likelihood of not supporting any of the four political parties increases by 0.54 percentage point. In the bottom row of panel A, we report the F-statistic from the first stage results. In line with figure 2, the interaction of the intensity of Arabic speakers with the overall number of refugees strongly predicts the refugee distribution, indicating that the instrumental variable is sufficiently strong (Bound et al., 1995). As mentioned before, the causal interpretation of these results requires the assumption that in the absence of refugee inflows, regions with high- and low-intensity of Arabic speakers would have similar trends in voting behavior. In figure 5, we show the trends in unadjusted rates of political support for AKP and the share of non-supporters of the major four parties, the two outcomes that show significant regression results, from 2012 to 2016 (table III). We compare the 11 provinces that have the highest share of Arabic speakers in 1965 (and drive the variation in our 23

24 instrument) with the rest of the country. 14 Panel (A) shows strikingly overlapping trends until 2014, followed by a clear slight drop in AKP s vote share during the expansion of Islamic State of Iraq and Syria (ISIS) in northern Syria. Panel (B) shows a similar pattern in the reverse direction: the share of respondents unaffiliated with any major political party rises after the acceleration of refugee inflows in 2014 in provinces that are more likely to receive these refugees. In a similar way, we fully interact the Arabic speaker intensity with each survey period and plot the estimated interaction coefficients from a dynamic difference-indifferences model: y ijt = α 3 + δ 3j + γ 3t + X iω + t π t (d t Arab 1965 j ) + ɛ 3ijt (4) where, as before, y ijt indicates the binary outcome for political affiliation, δ 3j and γ 3t capture the province and survey month fixed effects, and the vector X i include characteristics of the survey respondent. d t is a binary indicator for each survey month and is interacted with Arab 1965 j, the population share of Arabic speakers in The estimated coefficients for π t are plotted in figure 6, panels (A) and (B). Despite being less precise, we observe similar trends in the pre-refugee period for both outcomes and the movement of the coefficients afterwards, as expected, follow the same patterns as in figure 5, panels (A) and (B). In other words, the reported 2SLS coefficients in table III appear to be mainly driven by the peak of the refugee movement in 2014 and Hatay, Mardin, Siirt, Şanlıurfa, Osmaniye, Adana, Mersin, Diyarbakır, Van, Kilis, and Gaziantep. 97 percent of the native Arabic speakers in 1965 lived in these 11 provinces. 24

25 In table III, panel B, we provide the 2SLS estimates with the province level data on actual election outcomes. OLS results show that a one-percentage point increase in refugee influx lowered the voting shares of CHP by 0.08 percentage points but increased the voting share of MHP by 0.06 percentage points. The effects are marginally significant and tiny in magnitude, but precisely estimated. The only significant effect for the 2SLS models is the effect on CHP voting share, which is modest and negative. The 2SLS estimates for AKP and HDP lack precision although 95 percent confidence interval of the estimated coefficients from the election data overlap with the survey data results. We also find no discernible effect on voter turnout. Overall, these results suggest that the refugee inflows caused a small decline in support for the government party in shifting the voters into not affiliating themselves with the government party in a survey while it had no impact on election results. Table IV shows the 2SLS results based on survey data from subgroups by ethnicity, gender, age, and education. The estimated coefficients are similar for all demographic groups except when the samples are broken by ethnicity. The negative impact of refugees is larger for ethnic Turks who constitute the majority of the electoral body in Turkey. There are tiny but precisely estimated coefficients for CHP and MHP for some demographic groups, which we interpret cautiously. We conclude that the empirical evidence is too weak to indicate any impact of refugees on the voting preferences for the three opposition parties; our results rather suggest a small and temporary decline in support for the ruling AKP with no cross-party transition of voters. 25

26 V.3 Robustness Checks We conduct multiple robustness checks to corroborate our results. First, we estimate our baseline models with only three covariates: refugee density, province and period fixed effects. The results should be similar to those with the covariates if we are capturing the exogenous variation through location and period fixed effects. Second, we exclude Istanbul to confirm that our results are not driven this province, which is an outlier in most indicators including population, share of economic activity, and the number of Syrian refugees residing. Third, we exclude provinces that were district in 1965 and use the exactly matched sample of administrative units in both 1965 and the analysis period. The results from the three robustness checks for survey data and election results are shown in appendix tables A1, A2, and A3, and are statistically identical with similar coefficients directions and sizes. In appendix table A4, we show the main estimates excluding the border provinces to confirm that, despite being small, the estimated effects are not solely driven by the direct impact of conflict on border provinces. Note that we exclude a substantial variation in our instrument and despite being still correlated with the endogenous refugee resettlement, the instrument is now not sufficiently strong to provide valid statistical inference for the local average treatment effects. 15 We additionally provide a reduced-form estimate by regressing our outcomes on the instrument directly without instrumenting to confirm that the results from the IV estimates are qualitatively similar. The reduced-form estimates from the survey data suggest a decline in support for AKP while no detectable increase in the share 15 See Staiger and Stock (1997) and Stock and Yogo (2002) for a general discussion on weak instruments. 26

27 of supporters for other opposition parties. The other category is composed mostly of indecisive and absentee voters and the coefficient for this outcome positive but similarly imprecisely estimated. Note that this is due to the fact there is little variation left to exploit when we take out the border provinces which constitute both the major destinations and the historically dense settlement areas for Ethnic Arabs in Turkey. The final robustness check involves comparing elections results and the survey data with actual voting behavior. In our main analysis, the respondents in the survey are asked to state their voting preference if the elections were held on the day of the interview, which might be different from who they actually voted for. For most of the monthly surveys, Konda also asked respondents the political party they voted for during the previous general elections. 16 We use this information to create our outcome based on the actual voting behavior to re-estimate the main model using survey data. Appendix table A5, Panel A shows the results from the survey data, which is directly comparable to the actual election estimates that we report again in appendix table A5, Panel B. The results are remarkably similar, indicating that the refugee inflows had a negligible impact on election outcomes. VI Conclusion The Syrian Civil War has caused one of the largest movements of people since World War II and resulted in over 3.5 million refugees entering Turkey. The voters in Turkey, as our study documents, are strongly polarized in their attitudes towards 16 The information is missing 10 consecutive months from June 2014 to March

28 refugees based on their political party affiliation. Using a different survey instrument, Getmansky et al. (2018) document similar alignment of Turkish voters with the party lines in Turkey. Both empirical findings are consistent with a polarized political environment in which party line plays a more important role than the individual exposure or significant information on the subject in forming public opinion (Druckman et al., 2013). Moreover, we find that the local average treatment effects that rely on ethnic/language similarities between the refugees and the local host population show no impact on actual voting behavior of these communities. The immobility of voters across parties might simply reflect the extreme polarization of voters in Turkey. As documented in Hainmueller and Hopkins (2014), if cultural, ethnic and religious differences between the host and refugee populations are drivers of the political repercussions of these inflows rather than their impact on economic wellbeing, the shared demographic background between Syria and Turkey would mitigate the voter response to refugee inflows. Our empirical results align with both arguments. 28

29 References Achilli, L. (2015). Syrian refugees in jordan: a reality check. Migration Policy Centre, Policy Briefs, 2. Akarca, A. T., Başlevent, C., et al. (2009). Inter-party vote movements in turkey: The sources of akp votes in İktisat İşletme ve Finans, 24(285): Akarca, A. T. and Tansel, A. (2006). Economic performance and political outcomes: An analysis of the turkish parliamentary and local election results between 1950 and Public Choice, 129(1-2): Akgündüz, Y., Van den Berg, M., and Hassink, W. (2015). The impact of refugee crises on host labor markets: the case of the syrian refugee crisis in turkey. IZA Discussion Paper No Allport, G. (1979). The nature of prejudice. Addison-Wesley Pub. Co, Reading, Mass. Altonji, J. G. and Card, D. (1991). The effects of immigration on the labor market outcomes of less-skilled natives. In Immigration, trade, and the labor market, pages University of Chicago Press. Arar, R., Hintz, L., and Normal, Kelsel, P. (2016). The real refugee crisis is in the middle east, not europe. goo.gl/aelaus. Barone, G., D Ignazio, A., de Blasio, G., and Naticchioni, P. (2016). Mr. rossi, mr. hu and politics. the role of immigration in shaping natives voting behavior. Journal of Public Economics, 136: