Hate at first sight? Dynamic aspects of the electoral impact of migration: the case of Ukip

In this paper, we test the hypothesis that the causal effect of immigrant presence on anti-immigrant votes is a short-run effect. For this purpose, we consider a distributed lag model and adapt the standard instrumental variable approach proposed by Altonji and Card (1991) to a dynamic framework. The evidence from our case study, votes for the UK Independent Party (Ukip) in recent European elections, supports our hypothesis. Furthermore, we find that this effect is robust to differences across areas in terms of population density and socioeconomic characteristics, and it is only partly explained by integration issues.

Recent studies investigate the causal role of immigration on voting, i.e., how electoral outcomes are influenced by the presence of immigrants in a neighborhood (Otto and Steinhardt 2014;Mendez and Cutillas 2014;Barone et al. 2016;Steinmayr 2016;Halla et al. 2017;Harmon 2018;Brunner and Kuhn 2018;Dustmann et al. 2018). Most of these studies provide evidence that living in an area with a greater number of immigrants increases the probability of voting for parties that promote tighter immigration policies. Nevertheless, several of these studies yield the opposite result (Steinmayr 2016;Colantone and Stanig 2018). Furthermore, the literature focusing directly on natives' attitudes toward immigration, rather than on voting, does not indicate a significant effect caused by the presence of foreign-born individuals in surrounding areas (Scheve and Slaughter 2001;Card et al. 2005). Anecdotal evidence related to the latest US presidential election and the Brexit vote has pointed in the same direction. 1 Basing our work on this puzzling framework, we explore dynamic aspects related to this effect. In particular, we formulate the hypothesis that hostility toward immigration is temporary. Thus, the time dimension might hold the key to understand the way in which immigration affects voting outcomes. Different underlying mechanisms could lie at the root of such negative short-run effects, such as material concerns regarding the adjustment costs of new migrant inflows (the early phases of new immigrant integration) or prejudicial attitudes (initial attitudes not confirmed by experience), both denoting a "hate at first sight" effect.
It is important to note that integration does not occur overnight. Immigration can initially exert increased pressure on the welfare system and labor market, causing natives to react with hostility to new immigrant flows. Over time, as local authorities and the native population learn to cope with the process, the potential initial distressing Fig. 1 Google Trends in the UK by topic. Data originate from https://trends.google.it 1 http://www.motherjones.com/kevin-drum/2016/11/support-trump-strongest-where-illegal-immigrationlowest; http://www.nytimes.com/2016/11/11/opinion/identity-over-ideology.html situation could be mitigated. The extensive literature that analyzes attitudes toward immigration has noted the negative attitudes regarding the effects of immigration on the labor market, crime, and cultural homogeneity (Scheve and Slaughter 2001;Mayda 2006;Facchini and Mayda 2009;Card et al. 2012). These attitudes might not reflect the effects of immigration per se; however, they could be influenced by a failure to anticipate the immigrants' integration.
With respect to prejudicial attitudes, Arrow (1971) suggested that individuals only resort to stereotyping immigrants when they lack information regarding their habits, customs and traditions. Therefore, as soon as direct information regarding new immigrant neighbors is acquired, the negative attitude produced by their presence should cease. A similar argument is that political narratives regarding immigration might be more effective where immigration is more recent, as recent changes in the composition of a neighborhood's population could make them more salient.
After testing for this short-run effect, we will explore several channels related to integration, specifically unemployment, welfare expenditure, and crime. Furthermore, political preferences have been shown to be very sensitive to differences between urban and rural environments (e.g., Cesur and Mocan 2018). In particular, immigration to larger urban centers has generally started before immigration to more rural areas, which can explain why previous studies have found a different effect in these two different contexts (Barone et al. 2016;Dustmann et al. 2018). Considering our dynamic framework and heterogeneity in areas over different dimensions-population density and socioeconomic characteristics-we may derive insights as to why these previous studies found contradictory results. A general problem with these studies is that they search for individual channels through which hostility may arise even though voting data are expressed at a more aggregated level. This shortcoming may result in a misleading interpretation of the empirical results, and we will account for it when commenting on our findings. Indeed, when analyzing electoral outcomes, ecological fallacy issues may be relevant (Rotte and Steininger 2009;King 1997).
To study the dynamic features of the causal effect of immigration, we consider a finite specification of a distributed lag model (referred to as the DL model) and adapt to this model the empirical strategy commonly used in the literature (Halla et al. 2017; Barone et al. 2016). We use a multi-instrument approach that considers the lagged values of the "shift-share" instrument proposed by Altonji and Card (1991). In a paper on the labor market effects of immigration, Jaeger et al. (2018) use a similar approach to disentangle a short-run effect from a long-run effect of immigration. In contrast to their work, we have panel data, and we study and test the specific identification conditions that arise in a panel IV estimation. Finally, instead of imposing a lag structure, we choose the DL specification according to the standard information criteria used in time series analyses. The best lag specification of our model for the UK includes two lags: the present lag and the 2-year lag. As it will be clearer later on, the two coefficients of our DL model can be interpreted as coefficients associated with the overall level of immigrants (independent of their arrival dates) and the recent flows of new migrants, respectively.
We focus on European elections in the UK, a country in which the immigration issue has been central to all of the latest electoral outcomes. We analyze the votes for the UK Independent Party (Ukip), which, over the span of 3 election turnouts from 2004 to 2014, increased the share of its votes from 15.6 to 26.8%. In recent years, this party has proposed a strong anti-EU and anti-immigrant agenda. We will show that attitudes toward migrations in the UK are strongly related to attitudes toward the European integration process, which indicates that votes for Ukip in European elections are particularly appropriate for investigating the causal effect of the presence of immigrants. In Scotland and North Ireland, voting during the European elections has been influenced by nationalistic sentiments more than by migration; thus, we exclude data from these regions from the dataset.
To anticipate our results, the panel data estimation of our finite DL model regarding votes for Ukip shows that the two effects associated with the lag coefficients are both significant and have opposite signs. The presence of new immigrants has a positive effect on votes for Ukip, but there is no positive effect based on the presence of immigrants per se. From our estimates, a one-standard deviation increase in the new flows of immigrants produces an increase in votes that amounts to approximately twothirds of its standard deviation, while an equivalent increase in the share of immigrants entails a decrease in votes for Ukip of more than five-thirds of its standard deviation. This result strongly supports our hypothesis. In contrast to previous studies, in our case, this effect is robust to heterogeneity between areas. At same time, exploring these heterogeneities across different dimensions enables us to go beyond the city-country dualism of the existing results and suggests that other political and cultural factors may be relevant. Furthermore, we find that integration issues fail to fully explain the effect of immigration flows; however, they are indeed relevant in explaining this effect, particularly in combination with changes in welfare expenditure per capita.
In the next section, we provide a review of the literature on the political effects of immigration. We subsequently explain why votes for Ukip at European elections are an appropriate test for the effect of immigration and present the data in Sect. 3. In Sect. 4, we establish the econometric strategy, and in Sect. 5, we present our main results on the hate at first sight effect. In Sect. 6, we run robustness tests, testing for self-selection in political preferences related to internal migration and electoral turnout and estimating our model over other political outcomes (votes for Conservatives and Labour and turnout at elections). In Sect. 7, we consider heterogeneity in areas and explore possible channels of the short-run effect by considering the interactions of migration inflows with economic, social, and public policy variables. The final section presents our conclusions and suggests policy implications.

Related literature
This paper is related to literature examining the effects of immigration on electoral outcomes. Some studies belonging to this branch find a positive effect of the presence of immigrants on votes for parties with clear anti-immigration agendas. More specifically, Otto and Steinhardt (2014) study votes for far-right (anti-immigration) and green (pro-immigration) parties in Hamburg between 1987 and 2009, Barone et al. (2016) investigate electoral outcomes in Italy for the center-right coalition between 2001and 2008, and Halla et al. (2017 analyze votes for FPOE in Austria from 1971 to 2013. Their results suggest that competition in the labor market, welfare issues, and concerns for compositional amenities may drive the effect of the presence of immigrants. It should be noted, however, that these results are not without caveats. In many countries, only specific groups of immigrants cause an increase in votes for anti-immigrant parties. In the aforementioned paper by Halla et al. (2017), only low-skill immigrants are responsible for the effect of immigration, while the opposite effect holds for highly skilled immigrants. In each country, a slightly different category seems to matter. More specifically, effects are found in Spain only for African immigrants (Mendez and Cutillas 2014), in Denmark only for non-Western immigrants (Harmon 2018), and in Switzerland only for immigrants not from historically Protestant and Catholic non-ex-Communist countries (Brunner and Kuhn 2018), which suggest overall that cultural distance and prejudices may be an important explanation. Several results on refugees suggest opposite conclusions altogether. Focusing on the 2015 wave in Austria, Steinmayr (2016) finds that hosting refugees decreases votes for FPOE. In studying the dismantling of the Calais jungle in France, Vertier and Viskanic (2018) find that where new refugee centers are established, Front National received fewer votes in the 2017 Presidential elections. Both these papers interpret their results in terms of the contact theory (Allport 1954), according to which through repeated interactions immigrants and natives learn to cooperate and hostility decreases.
Trying to explore the complex relation between immigration and voting, other studies have searched for nonlinear effects. Mayda et al. (2016) introduce in their main IV regressions on votes for the Republican Party in the USA both the linear and the quadratic term of the share of immigrants. They find that only the quadratic term of the share of immigrants is significant, which indicates that only where the immigration phenomenon has reached large proportions it has an impact on votes for anti-immigrant agendas. The interpretation is that perceived issues related to migration become salient for the natives in a certain area only if the share of immigrants is above a certain threshold. Among nonlinear effects, other researchers have focused on differences between large and small cities. Dustmann et al. (2018) find that the presence of refugees positively affects votes for anti-immigrant parties in Denmark in the years 1986-1998 only if they exclude the bigger municipalities from the sample. This effect is stronger in richer areas with a larger pre-existing presence of immigrants and more crimes. Barone et al. (2016) find the same results for immigration in big municipalities for Italy and suggest an interesting explanation for this pattern. It states that "immigration in big cities may have started sooner than in small municipalities" (pg. 7).
We contribute to the literature on nonlinear effects of immigration on voting by testing the hypothesis that the history of migration matters. Most studies of the effects of immigration on voting behavior use a panel data model in which the main explanatory variable is the share of immigrants. This strategy fails to capture these dynamics. As we subsequently discuss in further detail, an analysis with fixed effects can show the effect of a variation in the number of immigrants on the change in the number of votes, regardless of the size or history of the existing foreign community. However, the process of integrating a foreign population into a receiving society requires time. Social and economic frictions, such as prejudices or integration issues, might emerge at the beginning; however, once the immigrant community has settled, it could be that its presence does not lead to further changes in political outcomes.
We also refer to a small number of papers on the political impact of migrations in the UK. Similar to us, Becker and Fetzer (2016) analyze votes for Ukip in European elections, and they provide evidence that exposure to migration from Eastern Europe after 2004 boosted votes for Ukip. The same evidence has been found regarding Brexit, both in a comprehensive analysis of drivers of votes at a district level (Becker et al. 2017) and specifically with regard to Polish immigration (Viskanic 2017). What distinguishes our paper is the aim to disentangle a short-run effect and causally assess its relevance independent of the nationality of the immigrants.
3 The case of the UK

Political preferences toward Europe and attitudes toward immigration
Ukip was formed in 1993 by members of the Conservative Party who disagreed with the position of their former party on the European integration process. Ukip founding members had the primary goal of having the UK leave the European Union. Nonetheless, over time, they developed clear anti-immigrant positions. In September 2014, Ukip spokeswoman Jane Collins stated while trying to rally support for a law called "British jobs for British workers": "I want to know from the other parties whether they will give employers the right to put British youngsters first in regard to job vacancies." 2 Another example is a debated poster they used for the campaign for Brexit, in which they associated criticism over Europe with the 2015 refugee flows using an image that looked strikingly similar to frames of an old Nazi propaganda video. 3 We argue that there is generally a strict relationship between the two main issues at stake in the European political debate: immigration and the European institutional setting. This relationship justifies our consideration of votes for Ukip in European elections to test the causal effect of immigration. Data from the 2014 European Social Survey provide evidence for the existence of a link between attitudes toward immigration and individuals' attitudes toward the European Union (Table 1). We analyzed UK data from the 2014 survey, excluding observations from Northern Ireland and Scotland. We considered a question that captures attitudes toward the process of European unification. It reads, "Now thinking about the European Union, some say European unification should go further. Others say it has already gone too far. Using this card, what number on the scale best describes your position?". At the lower extreme, i.e., 0, the answer corresponded to "Unification already gone too far"; at the upper extreme, i.e., 10, it corresponded to "Unification should go further." We consider a broad range of variables that might be related to answers to this question, such as the type of job contracts of individuals, attitudes toward the government, the economy, religion and the country, self-reported measures of happiness and trust, and how much subjects watch TV. We also consider a whole set of attitudes toward migration. We do consider not only the standard question "To what extent do you think [country] should allow people from the poorer countries [in Europe or, alternatively, outside Europe] to come and live here?" but also other questions regarding immigrants' effects on the economy, the labor market, and well-being.
We find that self-reported attitudes toward Europe show strong Spearman rank correlations with issues related to immigration. They range from 0.33 to 0.45. We also find moderate but highly significant correlations with answers to other social questions, with none as high as those regarding immigration. It should be noted that this survey was conducted before the 2015 refugee crisis. Therefore, such correlations could not have been driven by the border control emergency and the sense of panic that derived from this situation. Furthermore, there seems to be no substantial difference in the correlations between attitudes toward immigrants from Europe and those from the rest of the world. Answers to these two questions showed a correlation of 0.83. It is therefore not welfare or labor considerations related to European legislation on internal migration or social rights that are responsible for this outcome. Given this evidence, it makes sense to consider votes for Ukip in European parliamentary elections as a dependent variable for testing the effect of the presence of immigrants on votes.

Data and descriptive statistics
Immigration to the UK, defined according to the individual country of birth, has considerably increased, from 8% of the overall population in 2002 to 11% in 2014. Starting from London, it has also expanded to the North-Central regions and some areas in the South-West (refer to Figs. 2 and 3). Greater variation occurred at the end of the period under analysis, namely, from 2012 to 2014 (refer to Fig. 3). At the same time, Ukip has received a large consensus, obtaining, on average, 30% of the votes in the most recent elections (refer to Fig. 4). Particularly interesting is the comparison between the maps showing the variation in the population immigrating to the UK and the electoral results for Ukip. Clearly, the regions where foreigners have arrived in greater numbers are also the areas where individuals have expressed a stronger preference for the anti-immigration party. This correlation does not emerge from the comparison of the maps showing the shares of the immigrants and the shares of votes for Ukip. In contrast, the correlation of the two variables seems to be negative. We How close do you feel to the country 0.10*** 0.23*** − 0.10*** 0.18*** Allow many immigrants from poorer countries in Europe Allow many immigrants from poorer countries outside Europe Immigration good for country's economy Immigrants create new jobs in country Immigrants make country better place to live 0.40*** 0.38*** 0.44*** 0.33*** 0.45*** N 2599 Data originate from 2014 round of the European Social Survey. Values in the table are nonparametric Spearman rank correlations between the corresponding rows and column * p < 0.1; * * p < 0.05; * * * p < 0.01 attempt to better investigate the apparent relationships between immigration and political preferences for Ukip using data from the UK Electoral Commission for the electoral data and data from the Office for National Statistics (ONS) for all of the other variables used in the analysis ( Table 2). The study concerns England and Wales, and it was performed at the Local Administrative Unit level (LAU1-NUTS3) for a total of 345 territorial units. From the total of 348 territorial units, we excluded the LAU of the City of London (which is a small subset of what is commonly considered London's financial district) because, curiously, no corresponding data on immigration are available. We also excluded the Isles of Scilly and the Isle of Wight, for which income estimates are missing. All variables are at the LAU level. We focus on immigrants by country of birth, so individuals who were born in another country and live legally in the UK regardless of their citizenship. This is the most commonly used definition by the OECD and in the literature. Unemployment is the  Authors' elaboration over data from the Electoral Commission and ONS (refer to paper for more details). High education and income are time invariant variables from 2011 UK census; the remaining variables are time variant (2004, 2009, and 2014) estimated rate by the Regional Labour Market Statistics office of the ONS. The other timevariant control variables are the proportion of individuals younger than 35 years old and older than 65 years old, population, and density. The last two variables should capture effects, linear and nonlinear, respectively, related to the urban features. For the pooled OLS estimates, we introduce additional time-invariant controls from the 2011 census data, including household income per year and high education, which is the share of individuals who have at least a bachelor's degree or equivalent and regional fixed effects. Additional variables that we consider in further analyses are the number of crimes per capita and benefit expenditures per capita. Numbers on crimes derive from the ONS and the benefit expenditures from the Department of Work and Pensions of the UK government.

Model specification
To analyze the dynamic aspects of the causal effect of immigration, our econometric strategy extends the methodology used in previous studies, particularly in Barone et al. (2016). This literature has adopted the IV approach used in the analyses of immigration's impact on the labor market. To adapt this approach to a dynamic framework, we use a multi-instrument strategy based on the current and lagged values of immigration. Jaeger et al. (2018) use a similar approach to overcome the limitations of the standard static approach, i.e., that it "conflates these short-and long-run responses (pg.13) to immigration. In contrast to their work, we use panel data. This approach is not costless, as rank and order conditions in the panel case require further specific tests on the instruments. Finally, instead of imposing a lag structure, we choose the better specification according to the information criteria commonly used in time series analysis. We start with the general specification of a distributed lag model: where y i, t is the share of votes for anti-immigration parties in region i at time t, x i, t is the share of immigrants in region i (i.e., the number of immigrants among the overall population) at time t, x i, (t − l) is the share of immigrants at time t − l, w i, t is a vector of control variables, ε i, t is the error term, and L is the total number of lags considered. The coefficients β 0 , …, β L explain the dynamic patterns of the impact of immigrants on voting behavior. In standard panel data model estimation, α and β 0 are estimated. However, if a dynamic process exists, i.e., if not all β 1 , …, β L are null, such a model would not be correctly specified. 4 Therefore, as a first step in our analysis, we verify whether a dynamic process exists. As is standard in time series analysis, to perform the identification analysis of the lag structure, we use the Akaike information criterion (AIC) and the Bayesian information criterion (BIC). We consider the specification that has the lowest AIC and BIC between all of the possible specifications of our model, including lags of up to 4 years. Table 3 shows the results for the possible lag structures with up to 4-year lags. Both tests indicate that the best specification is the one in the 5th row with only the 2-year lag: where l is the 2-year lag. As 2 years is a shorter time span than the electoral cycle of EU elections (the times t at which the variable y is realized are 2004, 2009 and 2014), the lag time dimension does not interact with our panel lag dimension (each realization of the variable x appears only in the equation corresponding to one realization of y).
As usual in finite DL models, prior to using least squares estimation, we first must check that the variable x is not a unit root. We use the Pesaran panel unit root test (Pesaran 2007). Table 4 presents the results of this test and confirms that there is no unit root up to the 2nd lag.
The multi-instrument approach consists of using the same instrument z for both the T realizations of x t and the T realizations of x t-l . We will discuss the validity of the "shift-share" instruments in the next section; however, here we introduce a test to satisfy the order and rank conditions for identification in this multi-instrument approach. As we discuss technically in Appendix 2, identification in this case requires checking for the nonstationarity of the instruments. This is analogous to the OLS case. Table 5 shows the results of the Pesaran unit root test on the instrument z.
Although the specification in Eq. 2 does not significantly change the standard estimation approach, which is common when considering a finite DL structure, the main issue is the interpretation of the results. In this study, we follow a standard approach; we rearrange the model specification to provide an appropriate interpretation of the parameters.
In our one lag case, Eq. 2 can be rewritten as where β s = β 0 + β l and Δx i, t = x i, t1 − x i, (t − l) represent immigration flows in the last l years. In this manner, as a result of the specification in Eq. 3, we can read the coefficient β l as the effect of the most recent migration flows and β s as the effect of the current stock of immigrants. 5 The hate at first sight eventual effect is represented by β l , where the negative sign of the estimation results seems to be anticipated by the name that we give it: recent migration flows increase consensus for anti-immigrants agendas. In contrast, we do not formulate a hypothesis regarding the direction of the overall effect of immigration, which can be negative or positive. 6 If β s is positive, we can conclude that immigration per se increases anti-immigration consensus. If β s is negative, we should conclude that regions with higher levels of immigrants are more prone to pro-immigrant views instead.
Any least squares estimation of the two specifications in Eq. 2 and Eq. 3 will obtain the same results once we consider the linear transformation β̂s ¼ β̂0 þ β̂l, and the estimators would have the same properties. The only difference is that, in the first case, we will obtain the standard errors for β 0 , while in the second case, we obtain the standard errors for β s . We consider the stock and flow specification in Eq. 3 because their coefficients can be directly interpreted; however, we also consider the equivalent for the standard DL specification in Eq. 2 to obtain the standard errors also for β̂0.

Identification strategy
The existing literature raises common concerns with regard to the endogeneity of the migration phenomenon. The primary concern addresses reversal causality. Immigrants could avoid areas with anti-immigrant sentiments and locate where they could more easily integrate. For example, by analyzing the case of Switzerland, Slotwinski and Stutzer (2019) find evidence that foreigners' location choices are significantly affected by natives' attitudes toward immigrants. Thus, in areas with more tolerant attitudes, the presence of immigrants would be larger than what would be predicted purely by push factors. This outcome generates a downward bias because there is an underestimation of the positive effect of immigration on votes for anti-immigrant agendas. The second concerns confounding factors that can drive both sorting decisions and political preferences. For example, Halla et al. (2017) suggested that an increase in votes for right-wing parties could be the result of a shift to greater pro-business attitudes. If these attitudes are determined by better economic 5 It is useful to recall that although panel data estimations, namely, estimations with a within estimator, of the standard specification without lag of Eq. 1 provide estimates of the effects of variations of x on variations of y, the underlying relation that they estimate is the effect of the level of x i, t on the level of y i, t . In contrast, the panel estimations of Eq. 3 would provide estimates of the effects of variations and acceleration of x on the variations in y. 6 From such a perspective, the AIC and BIC test the specifications with only one lag corresponding to the second, fifth and last raw numbers in Table 3, which can be interpreted as a test of the best specification for the length of the "hate at first sight" effect. Cips test assumes cross-section dependence in the form of a single, unobserved common factor performance in one area compared to other areas, it would create a pull factor for sorting decisions by immigrants. More generally, structural changes in the economy create the conditions for both a political shift in preferences and the emergence of a pull factor for immigration. We can assume that this spurious correlation generates an upward bias. The sorting problem can be overcome using an instrumental variable approach. An important body of literature has argued that immigrants choose where to locate according to individual networks. 7 Then, we exploit the tendency of migrants to move to areas where individuals from the same country already live (Bartel 1989), and we predict the actual distribution of the foreign population by observing the past distribution. Therefore, we use the instrument suggested by Altonji and Card (1991), as slightly modified by Cortés and Pan (2014) and Barone et al. (2016): where Imm it is the predicted share of current immigrants in region i. It is equal to the number of immigrants originating from group n at time t at the country level, the net contribution of region i to the total, apportioned by the proportion of individuals originating from the same country and living in region i at time t 0 , all aggregated to the main foreign nationalities present in the UK. Setting in a new country can be easier if individuals from the same foreign ethnicity can share information regarding the labor market and other institutions of the receiving nation. Nonetheless, a vast amount of literature shows that when immigrants become geographically clustered, which gives rise to the so called ethnic enclaves, it can hamper the adjustment process. According to Dustmann and Preston (2007), ethnic concentration can result in a xenophobic attitude, while Bauer et al. (2007) argue that the labor market competition is the highest among immigrants. As a result, the arrival of new immigrants is deterred where the stock is already high, and the dynamics of the migration network is thus better represented by a concave function of past settlement (for an example, see Epstein and Heizler-Cohen 2016). This is what we do in a further specification of our estimating model. 8 Using the past allocation of immigrants, we can argue that reverse causality issues are left out because the current political preferences are typically not anticipated. Furthermore, the time t 0 chosen must be such that the unobserved variables that cause voting today are not correlated with the determinants of immigration in the same year. This condition is reinforced if the voting behavior has changed over the last t − t 0 years. To validate this condition, in line with previous studies, we proceed to the choice of a year 7 Refer to, for example, Beine et al. (2011) and Mayda (2010). 8 A similar identification strategy can also be found in Bianchi et al. (2012). Cips test assumes cross-section dependence in the form of a single, unobserved common factor t 0 preceding the major changes in political attitudes related to the migration issue. This task becomes easy when we examine Ukip. It was founded in 1993 and as a Eurosceptic and anti-immigration party. It has represented a novelty on the political scene in the UK. Therefore, we choose the year 1981, well before the foundation of Ukip. We explicitly test for reverse causality related to this base year by regressing the share of immigrants in 1981 on the outcomes of the referendum on the permanence of the UK in the European Community in 1975. The choice of the 1975 referendum was based on the conditions being very similar to those in 2016. If we do not find a correlation between location choices by immigrants in 1981 and the results of the referendum, then even more so, these location choices should be independent from future political outcomes. The results of regression (1) in Table 6 show that the political preferences expressed at that time did not affect the location choices of past immigrants. In fact, a greater consensus for more open foreign policies does not seem to be correlated with the presence of immigrants in the area. Moreover, public debates and party representatives' positions toward immigration have dramatically changed since 1981. The tones of public discussions of immigration have exacerbated compared to past years. The existence of a pro-immigration political climate in the early 1990s is confirmed by the UK's signature of the Amsterdam Treaty in 1997 and its position in favor of the enlargement of the European Union to Central and Eastern European countries. Instead, an example of the change in the role that immigration plays in political debate is a straight comparison between the slogans used by Prime Ministers of the UK in the mid-1990s and today. Blair stated "a simple way to take the measure of a country is to look at how many want in and how many want out," 9 whereas May currently argues "the aim is to create here in Britain a truly hostile environment for illegal migration". 10 This issue is not the result of a difference between OLS. Errors are clustered at the regional level in parentheses. In regression (1), controls include unemployment and age older than 65 in 1981 at the LAU level. In regression (2), controls include population, population density, unemployment, the percentage older than 65, higher education, and income. The 1975 EU referendum results are at the county council level * p < 0.1; * * p < 0.05; * * * p < 0.01 9 https://www.forbes.com/sites/johnbaldoni/2018/07/04/a-newcomer-to-america-saved-my-fourth-of-july/#3 e1ebac35a6b 10 https://www.telegraph.co.uk/news/uknews/immigration/9291483/Theresa-May-interview-Were-going-togive-illegal-migrants-a-really-hostile-reception.html the Conservative and Labour parties. John Major, the Conservative Prime Minister who preceded Tony Blair (1990Blair ( -1997, also expressed opinions in favor of immigration: "There was a different social value placed on immigration (...) I saw immigration at very close quarters in the 1950s. They shared my house. They were my neighbors. I played with them as boys. I did not see people who had come here just to benefit from our social system. I saw people with guts and the drive to travel halfway across the world in many cases to better themselves and their families. And I think that is a very conservative instinct." 11 Immigration is also currently a hotly debated issue at the European level. Immediately after the win of Leave in the Brexit referendum, Nigel Farage stated at the European Parliament "You're in denial over Mrs. Merkel's call for as many people as possible to cross the Mediterraneanwhich has led to massive divisions between within countries and between countries". 12 We believe that all of these arguments have spread consensus in favor of anti-immigration agendas.
There is also empirical evidence that this shock has occurred, and it is both general and specific with respect to the immigration topic. As seen in regression (2) in Table 6, we show with a pooled OLS estimation over all of our time-variant and time-invariant controls that votes for Ukip do not depend on the results of the 1975 European Referendum. More specifically with respect to migration, in Fig. 5, we show the share of votes for Ukip in 2014 by regions in relation to attitudes toward immigrants in 1983, codified in ranges from 1 (gone too far) to 3 (not gone far enough). This last bit of data originates from the 1983 wave of the British Election Studies, which comprises slightly less than 4000 observations. There is no significant difference in the consensus for Ukip between the areas where individuals were, on average, worried about the number of immigrants living in the UK and the areas where individuals were not worried. Both of these analyses support the idea that there was a political shock between 1981 and the years that we considered for our main analysis.
The figures for immigrants in Local Authority Units in 1981 are obtained from the 1981 census. We eliminated Harrogate from the units observed because it is missing from the census. This census divided immigrants into seven subgroups (Old Commonwealth, New Commonwealth, Pakistani, Irish, European Community, other Europeans and rest of the world), and we exploit this categorization in creating the instruments. Table 7 shows the main results of our analysis. The first column corresponds to OLS estimates of standard Eq. 3 without fixed effects. Here, we exploit both cross section and time variability. As controls, both time-variant variables and time-invariant variables are included. Between these variables are regional and year fixed effects. In contrast to previous studies, we do not find that the share of immigrants has a positive effect on votes for Ukip: the coefficient (− 0.11) is significant at the 10% level and negative.

Results
In column (2), we report the OLS results of the DL model in Eq. 3. The same estimation corresponding to the specifications in Eq. 2 is reported in Appendix 3. Once the appropriate linear transformations are considered, the coefficients are the same. This approach enables us to obtain the standard errors of the original specification. Recent immigration flows have significant and positive effects. A 1 percentage point increase in new flows is associated with an increase of 0.25 p.p. in votes for Ukip. Note that the coefficient associated with the share of immigrants decreases to − 0.25, which indicates that some of the positive effects of the immigrant flows were captured in the previous model by the share of immigrants, which is consistent with the theoretical underpinnings of our econometric analysis.
Column (3) provides the results of the specification with local authority fixed effects. Time-invariant variables are excluded from the analysis. This specification addresses many time invariant endogeneity concerns, such as measurement errors, omitted variables, and reverse causality. The coefficient of recent flows remains significant and greater than the coefficient of the share of immigrants (which remains negative but not significant). Both coefficients are smaller in magnitude when we consider fixed effects, which indicate that omitted variables related to political preferences at the LAU level were likely driving the coefficients up.
The more robust estimations are in columns (4), (5), and (6), in which we complement LAU fixed effects with the instrumental variable approach presented in the previous section, with a 2SLS estimation, a 3SLS estimation or by adding the quadratic term of the instruments. In these estimates, we account for all endogeneity concerns. The Kleibergen-Paap F statistic on multiple instruments when we only consider the linear term of the instruments is 17, which is greater than the level of 10 that typically causes concerns for weak instruments. When we consider the quadratic term of the instruments, the F increases up to 587.45, which indicates the relevance of nonlinear effects in immigrants' exploitation of pre-existing ethnic networks. Both coefficients Errors clustered at regional level in parentheses. Immigrant share is computed over total municipality population. For the FE-IV estimations, we use as an instrument the share of immigrants by LAU in 1981 (more details in the paper). In column (6), we consider as instruments the polynomial of the share of immigrants and immigration flows of degree two. The Hansen test of overidentification of multiple instruments does not reject the null hypothesis of validity of the instruments (chi-sq = 3.62, p = 0.1634) * p < 0.1; * * p < 0.05; * * * p < 0.01 associated with the share and recent flows of immigrants remain significant and with the same sign, although the former is less significant. They both increase in magnitude. Areas where there has been an acceleration of new arrivals by 1 percentage point see an increase in votes for Ukip by 1.1-1.2 p.p. In contrast, an increase of 1 p.p. in the share of immigrants corresponds to 1.7-1.9 p.p. fewer votes for Ukip. To put it differently, a one-standard deviation increase in immigration flows and in the share of immigrants entails an increase in votes for Ukip of 0.68% of its standard deviation and a decrease of 1.71, respectively. This increase in the size of the coefficients is consistent with previous results, specifically with Halla et al. (2017) and Harmon (2018), and its size is similar to Barone et al. (2016) and Dustmann et al. (2018). In particular, it suggests that there is a downward bias for OLS and the FE coefficients of immigration flows and an upward bias for OLS and the FE coefficients of the share of immigrants.

Political preferences, internal migration flows, and turnout
We discuss here two important issues that may affect our results: internal migration flows and the electoral turnout. An important body of the previous literature (Card and Di Nardo 2000;Borjas 2006; Peri and Sparber 2011) has argued that cross-regional analyses are invalidated because natives decide to relocate in response to immigration flows. In the case of political preferences, individuals with more anti-immigrant attitudes could move to other regions because of the arrival of new immigrants. To date, the existing studies have produced scant evidence of the so-called skating rink effect, according to which "each new immigrant knocks a native off the ice" (Card and Di Nardo 2000, pg. 360). We do not make a priori hypotheses; however, we explicitly test for it as a robustness check. In our case, the effect of a negative sentiment related to immigration might have pushed anti-immigrant natives to migrate internally, thus reducing the number of individuals voting for Ukip in areas with high concentrations of immigrants. We add net internal migration flows in our empirical strategy to control for this potential bias. In particular, we distinguish between the internal migrations of natives and foreign-born citizens, both deriving from the Internal Migration Dataset of the ONS, and we regress the net flows of natives on the flows of immigrants. We subsequently use the fitted values, i.e., the relocation decisions of natives due to immigration, as a control.
Similarly, it has been argued that radical right parties' results may be related to electoral turnout. There is mixed evidence regarding this issue; however, some studies have found that radical right parties benefit from low turnout (Immerzeel and Pickup 2015;Stockemer 2017;Abbondanza and Bailo 2018). If individuals who attend the elections have more anti-immigrant preferences due to this effect, taking into consideration the electoral turnout permits to better pin down the coefficient of the level and flows of immigrants. Therefore, we introduce the share of turnout at elections as a control in our main specification. Table 8 shows that internal flows do not have a significant effect on the vote share for the Ukip party, which indicates that internal migrants are not selected with respect to their political preferences. As a result, the addition of the internal flows to the estimating equation does not produce a sizeable change in the other coefficients. The same goes for electoral turnout. While we find a positive relation between turnout and votes for Ukip, it is not significant, and there is no change in the other coefficients.

Immigrants' effects on other political outcomes
As a robustness check, we examine the political outcomes of other parties and turnout in the same European elections. If our results are robust, we should find some consistency between the short-run effects related to immigration flows regarding votes for other parties and their agendas regarding migration.
In Table 9, we consider the fixed effects 2SLS specification of the model (column 4 in Table 7) and change the dependent variable. We test the effects of the share of immigrants and their recent flows on votes for the Conservative Party, votes for the Labour Party and turnout at the elections.
The Labour Party in these years was much in favor of immigration. We previously cited Tony Blair, who viewed immigration as a measure of the success of the UK. There is also speculation that favoring immigration was a deliberative policy of the Labour governments of the 1990s and 2000s. 13 In contrast, Conservatives' view in these years was more nuanced. In a political speech to his party in 2011, David Cameron explicitly stated regarding immigration, "I believe the role of politicians is to cut through the extremes of this debate and approach the subject sensibly and reasonably". 14 He noted Errors clustered at regional level in parentheses. All models are FE-2SLS. Immigrant share is computed over total municipality population. All regressions are estimated with the IV approach explained in Sect. 4.2 (refer to column (4) of Table 7 for the baseline specification) * p < 0.1; * * p < 0.05; * * * p < 0.01 the great contributions of immigrants to the UK's life and at the same time the need to ease the increasing pressure on the UK's borders.
Our results are as consistent with these parties' attitudes toward migration as they were with votes for Ukip. In fact, we do not identify a significant effect of share or flows of immigrants on votes for the Conservative Party. Citizens likely did not see the party's position on immigration as a clear stance on the matter. In contrast, there is a 0.6 p.p. significant and negative effect of recent flows of immigrants on votes for the Labour Party. Given their policies on issues related to immigration, this outcome is not surprising. We do not identify a significant effect on turnout for European elections: immigration flows pushed individuals in local authorities to vote more for Ukip and not abstain from the vote in protest. Note that all of these models have lower goodness of fit than the model on Ukip, ranging from 0.343 to 0.379 compared to 0.685. It truly seems that differences in attitudes regarding migration might explain this difference.

Heterogeneity by areas
The literature to which we contribute has found evidence of relevant heterogeneity in immigrants' effects by socioeconomic features, particularly in the gradient big citiessmall municipalities. The effect of the presence of immigrants is higher in small municipalities than in big cities (Barone et al. 2016;Becker and Fetzer 2016;Dustmann et al. 2018). As noted in the introduction, our distinction between stock and flows may explain this difference because it captures the different immigration history of the areas. If both coefficients remain significant and with the same sign regardless of the size of the municipality, the hate at first sight effect would explain Errors clustered at regional level in parentheses. All models are FE-2SLS. Immigrant share is computed over total municipality population. All regressions are estimated with the IV approach explained in section 4.2 (refer to column (4) of Table 7 for the baseline specification) * p < 0.1; * * p < 0.05; * * * p < 0.01 much of the previous evidence. Therefore, to test this hypothesis, we proceed in three steps. First, as we have electoral results at the LAU level, we use the population density as a proxy of this geographical heterogeneity and compute the same model of our preferred specification adding an interaction term of the stock of immigrants with the density. In this way, we intend to reproduce previous findings. Second, we add an interaction term between population density and the flows of immigrants to test whether the distinction helps explain previous findings. The results from the first two steps are shown in Figs. 6 and 7. The underlying regressions are presented in Appendix 4. In Fig. 6, we represent the effect of the immigrant share by population density. We reproduce previous findings. For lowdensity areas, the effect of the immigrant share is positive; however, it is not significant. In our dataset, 50% of the LAUs have a population density below 0.55; thus, the number of areas where this effect is positive is not negligible. Figure 7 shows that disentangling flows and share of immigrants does not completely explain this result. On one side, the effect of the immigration flows is significant and positive at all levels of population density. Regardless of the population density, the effect has almost a uniform magnitude, albeit slightly decreasing. This result runs counter to previous findings. Disentangling the time components of the effect of immigration isolates a short-run effect, which is robust to heterogeneity in areas. On the other side, the effect of the share of immigrants maintains the same attributes of the previous analysis. It is positive for low-density areas and negative for high-density areas. Therefore, we can conclude that our distinction provides insights on heterogeneity, although it does not fully explain previous findings in the literature.
Third, we further investigate this issue by exploiting a classification of the Local Administrative Level 1 in the 2011 UK Census based on 59 social variables. In particular, we aggregate areas using this classification to form 4 supergroups, as shown  Fig. 8. The first supergroup corresponds to the category of "English and Welsh Countryside". Compared to the UK average, these individuals work more in agriculture, have lower education levels, are more home owners, married, and use more private transports. The second supergroup includes higher density areas with a more tertiary vocation, where the immigration incidence is higher as the weight of single young individuals who use public transports and rent flats. It includes the three categories "London cosmopolitan", "Suburban traits" and "Business and Education centers." The third category is a mixture of the first two supergroups; similar to the first one, it has a lower population density, individuals use more private transport or bicycle and are more home owners, while similar to the second one, it is characterized by a broader service sector and has a higher weight of medium and high education levels. It includes "Prosperous England" and "Coast and Heritage." The last supergroup is the "Mining Heritage and Manufacturing," which comprises the "working class" areas that were the hard core of the Labour Party social base. In this class, immigration is below the average, low education is higher, as well as social renting, and there is a higher presence of divorced and separated.
The results are shown in Table 10, where we use the same specification in column (1) of the main results (Table 7). As in the previous analysis, immigration flows have a positive and significant effect in all cases. Thus, this result is robust not only to heterogeneity in population density but also in social contexts. However, there is a relevant difference in the magnitude of the coefficient by supergroups. This coefficient is strikingly high in the "English and Welsh Countryside." The other supergroup where this effect is higher is the "working class" supergroup, which is consistent with the results on the Labour Party in Table 9. Territorial specific features have an impact on the effect of immigration levels, which is not significant in the two low-density supergroups (columns (2) and (4)). Interestingly, the two supergroups in which the effect is Regressions are FE-2SLS. Refer to Appendix 4. Confidence intervals in the figure are at 95% level significant are very different from each other. "London cosmopolitan, Suburban traits and Business and Education centers" and "Mining Heritage and Manufacturing" differ in education levels and economic activities; the former has higher education levels and more economic activities in services, while the latter has lower education levels and more industry. This evidence suggests that it may be better to go beyond the city-country dualism and explore other possible channels yielding this effect. It suggests that political and cultural factors may be relevant in explaining heterogeneity by subgroups of areas, as differences in both cosmopolitan attitudes and past political preferences may have affected these results.

Integration issues
Failures in integration might be related to competition in the labor market, social benefits, or crime. First, if there is no immediate integration in the local labor market, immigrants may increase unemployment. Second, benefit expenditures per capita may not adjust to the arrival of new immigrants. Third, failures in integration may induce an increase in crimes. Hostility toward immigrants could be caused by any of these issues.
We test for these possibilities by adding an interaction term to the baseline IV model. This term captures the effects of immigration flows in areas where there might have been integration issues. These areas are captured by a dummy that takes the value of 1 if they are characterized by higher than the average growth in the unemployment rate, the expenditures on social benefits per capita, and the crime rate.
The results in Table 11 suggest that although adjustment costs to new migration inflows matter, there is substantially more that is left unexplained. Only interactions with areas that have spent more on benefits are significant. Not surprisingly, it is negative, which indicates that in areas where there is a Errors clustered at regional level in parentheses. All models are FE-2SLS. Immigrant share is computed over total municipality population. All regressions are estimated with the IV approach explained in section 4.2 (refer to column (6) of Table 7 for the baseline specification). Area Classification for Local Authorities: column (1) is over all England and Wales; column (2) is English and Welsh Countryside; column (3) is London Cosmopolitan, Business and Education Centers, and Suburban Traits; column (4) is Prosperous England, Coast, and Heritage; and column (5) is Mining Heritage and Manufacturing * p < 0.1; * * p < 0.05; * * * p < 0.01 reduction in benefit expenditures per capita, immigration flows have a greater positive effect on votes for Ukip. However, the two effects do not compensate for each other. The effect of recent flows remains positive after subtracting the interaction term. The remainder of the interaction terms is not strongly significant, although they go in the right direction and slightly decrease the coefficient associated with immigration flows. Only the coefficient associated with the variation in unemployment is weakly significant, which suggests that previous results in the literature related to the presence of immigrants may also hold for flows. Nevertheless, in contrast to previous literature, we do not draw conclusions regarding individual preferences. Rather, we focus on the identification of local socioeconomic conditions that push the vote for anti-immigration parties. Therefore, what we have stressed is not the existence of competition mechanisms between natives and immigrants but the suitability of a territory to properly integrate the newly arrived. This perspective accounts for the ecological fallacy caveat, according to which one should be very cautious in inferring about individual behaviors from group analyses. Errors clustered at regional level in parentheses. Immigrant share is computed over total municipality population. All regressions are estimated with the IV approach explained in Sect. 4.2 (refer to column (6) of Table 7 for the baseline specification) * p < 0.1; * * p < 0.05; * * * p < 0.01

Discussion and conclusions
This paper analyzes the impact of migrant inflows on votes. It modifies the standard IV strategy used in the literature to test the dynamic aspects of the political effect of immigration, and it considers a distributed lag specification and the correspondent additional tests for the rank and order conditions. We built a panel dataset based on England and Wales at the LAU level (345 territorial units) by matching data from different administrative sources with the electoral data on European elections from 2004 to 2014. Focusing on the UK Independence Party, our results robustly support the hypothesis that the time path of immigration is essential; immigration flows have a positive effect on the political consensus over anti-immigration parties, although this is a short-run effect and it vanishes before the parliamentary term. Accordingly, to understand the effects of immigration on the popular vote by simply analyzing the stock of immigrants per se can be deceptive. In the case of Ukip, the stock of immigrants per se would have exerted a negative effect on votes, although one that was barely significant. However, considering immigrant flows enables us to find evidence that immigration was indeed a key factor in boosting votes for Ukip. As additional results, we find that the short runeffect is independent of the size of cities. At the same time, we find suggestions that considering different subgroups of areas based on social resemblance may provide insights into the cultural and political motives for hostility toward immigrants. Finally, introducing interactions between immigration flows and changes in unemployment, benefit expenditures, and crimes suggests that adjustments in welfare systems might be an issue. We find that areas where there is a quicker increase in welfare after the arrival of new immigrants become less hostile toward new immigrants.
Our results provide insights into policy issues. First, if immigration flows drive hostility toward immigrants more than their share, it undermines some explanations for the emergence of anti-immigrant agendas and their corresponding proposed policies. Racism, changes in the equilibrium in the labor market, and compositional concerns might be less relevant than other mechanisms, although they typically attract significant attention in both economic research and the media. Second, policies should focus more on integration over its cultural, social, and economic sides. Over time, social forces can drive toward integration; however, it would likely be better if policies expedited this process. Third, a positive effect related to immigration flows indicates that there is a need to pay closer attention to how flows are distributed over time and space: it is probably better to allow immigrants to arrive in small waves rather than large ones and distribute recent arrivals in a homogenous manner, rather than having immigrants concentrate in certain areas.
In conclusion, we wish to note that our analysis is necessarily only a starting point. The elements that underlie the short-and long-term effects and the relationship between the two remain to be better investigated. There might be attitudes rooted in culture that could derive from historical events, such as decolonization or racism. However, it could also be that failures in integration policies are responsible for these effects. Moreover, we are not fully aware of the roles played by the media and political narratives. Given the current relevance of "post-factual" phenomena, there is a need to investigate this issue.
In the 1930s, alternative facts and propaganda in Europe were fueling prejudices against ethnic or religious minorities. At that time, similar to now, the same arguments over minorities, tinged with nationalist tones, were propagated about immigrants. There would appear to be nothing new under the sun. At the same time, over the last 20 years, there has been an unprecedented rise in immigration flow across Europe. Both issues determine the need to further investigate the political impact of migration. Our contribution to the subject is the conclusion that the effect of immigration on voting patterns depends on the specific time path of immigration. considered in Sect. 2 could be the result of the different effect of immigration in new, rather than the oldest, destinations of migration.

Appendix 2. Identification of OLS and IV estimations in the case of Eqs. 2 and 3
The finite lag model can be interpreted as a nonlinear case based on the lag operator. Using Wooldridge's (2010, pg. 342) notation, we can write the following: where using the lag operator L (which maps x t onto x t − l ), the function g(x) is equal to Lx for Eq. 2 and to (1 − L)x for Eq. 3. Since the specifications are linear in the parameters (the lag operator is an algebraic function), if we have an instrument z t for x t at any t ∈ (1 − l; 1; …; T − l; T), we can rely on standard 2SLS estimations by adding g(z) as instrument for g(x) (ibid pg. 235). It is analogous to the linear case of two distinct endogenous variables instrumented with two distinct instruments.
Regarding the identification conditions, the order condition is satisfied since we have in each equation the number of excluded endogenous variables, and the number of included exogenous ones corresponds; however, for the rank condition, the dynamic relation between the instruments impose a further condition. Indeed, the covariance matrix of the estimator is given in the two cases of Eqs. 2 and 3, respectively, by the following: x t 0 z t x t 0 z t−l x t−l 0 z t x t−l 0 z t−l ð7Þ If one of the two variables x and z has unit roots, in both cases, the expectation and the plim of this matrix would not have full ranks. In the case in Eq. 7, the rows or the column would correspond; in the second case, one column or one raw number would be null.
Appendix 3. Results using the specification in Eq. 2 We report in Table 12 an alternative version of Table 6 with regressions on the original FDL specification in Eq. 2. By construction, the coefficient of the lagged immigrant share is equal to the opposite of the coefficient of the flows in Table 6, while the coefficient of the first raw number is equal to the sum of the coefficients of the two first raw numbers in Table 6. While standard errors change, the results are all confirmed.

Table 12
Main results using the original FDL specification in Eq. 2 (1) (3) Errors clustered at regional level in parentheses. Immigrant share is computed over total municipality population. For the FE-IV estimations, we use as instrument the share of immigrants by LAU in 1981 (more details in the paper). In column (6), we consider as instrument the polynomial of the share of immigrants and of immigration flows of degree two.
The Hansen test of overidentification of multiple instruments does not reject the null hypothesis of validity of the instruments (chi-sq = 3.62, p = 0.1634) * p < 0.1; * * p < 0.05; * * * p < 0.01 Errors clustered at regional level in parentheses. All models are FE-2SLS. Immigrant share is computed over total municipality population. All regressions are estimated with the IV approach explained in Sect. 4.2 (refer to column (6) of Table 7 for the baseline specification) * p < 0.1; * * p < 0.05, * * * p < 0.01 Stockemer D (2017) The success of radical right-wing parties in Western European regions-new challenging findings. J Contemp Eur Stud 25(1):41-56 Vertier P, Viskanic M (2018) Dismantling the "jungle": migrant relocation and extreme voting in France.
CESifo Working paper 6927 Viskanic M (2017) Fear and loathing on the campaign trail: did immigration cause Brexit?. https://papers.ssrn. com/sol3/Papers.cfm?abstract_id=2941611 Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.