The Effect of Migration on Terror - Made at Home or Imported from Abroad?

We investigate whether the stock of foreigners residing in a country leads to a larger number of terrorist attacks on that country. Our instrument for the stock of foreigners relies on the interaction of two sets of variables. Variation across host-origin-dyads results from structural characteristics between the country of origin and the host, while variation over time makes use of changes in push and pull factors between host and origin countries resulting from natural disasters. Controlling for the levels of these variables themselves and fixed effects for dyads and years, the interaction provides a powerful and excludable instrument. Using data for 20 OECD host countries and 187 countries of origin over the 1980-2010 period we show that the probability of a terrorist attack increases with a larger number of foreigners living in a country. However, this scale effect is not larger than the effect domestic populations have on domestic terror. We find some evidence that terror is systematically imported from countries with large Muslim populations. A larger number of attacks against foreigners in the host country increases the risk of terror from foreigners there. We find that host country policies relating to integration and the rights of foreigners are key to fight terror- stricter policies that exclude foreigners already living in a country increase the risk of terror. High-skilled migrants are associated with a significantly lower risk of terror compared to low-skilled ones, while there is no significant difference between male and female migrants.


Introduction
There is clear and systematic evidence that countries threatened by terrorist attacks respond to this threat to their values by diminishing the very rights they aim to protect in the first place (Dreher et al. 2010). An area particularly prone to human rights restrictions is immigration and asylum policy. Arguably, it is easier to restrict the rights of foreigners in order to increase the (perceived) security of a country's natives than to restrict the rights of these natives (i.e., voters) themselves.
Plenty of evidence suggests that stricter immigration and visa policies are indeed a preferred reaction to terrorist attacks (Fitzpatrick 2002, Martin and Martin 2004, Avdan 2014 After the September 11, 2001 (hereafter 9/11) attacks on the United States, U.S. President George W. Bush issued a Presidential Directive introducing stricter immigration policies to combat terrorism. The new Department of Homeland Security (DHS) was founded in 2003, incorporating the former Immigration and Naturalization Service (INS). The new Department explicitly links immigration policies to anti-terrorism strategies (Kerwin 2005). A number of additional discriminatory measures have since been implemented, among them exceptional powers to the Attorney General to detain foreigners without hearings and proof of guilt if there is "reasonable grounds to believe" these foreigners are involved in terrorist activity, ethnic profiling, and required registration for certain groups of entrants -in particular from Muslim states (Spencer 2007). In 2016, the Republican candidate for the US-Presidency successfully competed on the promise to ban all Muslims from immigration to the United States. Directly after his inauguration he issued a travel ban for seven predominantly Muslim countries.
The United Kingdom equally tightened immigration policies in the wake of 9/11, most notably with the introduction of the Anti-terrorism, Crime and Security Act 2001 (Spencer 2007). Under the Act, the Secretary of State for the Home Department is allowed to order the detention of foreigners based on mere suspicion of terrorist involvement, without trial. 3 As Spencer (2007) summarizes, France, Germany, and Spain, among others, have similarly tightened immigration laws or procedures in response to the terrorist attacks of 9/11. In light of these reactions to terror, evidence that liberal immigration and integration policies or the number of foreigners living in a country foster terrorism is surprisingly scarce.
The only systematic statistical analysis we are aware of reports a negative correlation between migration and terrorist attacks (Bove and Böhmelt 2016). 4 Other previous studies that address the link between terror and migration either examine the effect of terror on migration (e.g., Dreher et al. 2011) or employ data on terrorists with immigration status rather than relying on systematic cross-country time-series data on migration and terror attacks (Kephart 2005, Leiken 2004, Leiken and Brooke 2006). Studies focusing on terrorists with immigrant background find a close link between immigration and terrorism. Given that they do not examine overall flows of immigration but only those cases in which immigrants have been involved in terrorist activity, these studies do obviously not allow drawing conclusions about the benefits and drawbacks of overall migration on terrorist attacks (Spencer 2007). 5 The absence of a causal investigation about whether and to what extent migration induces terror is an important gap in the literature.
We fill this gap and analyze the effect of immigration on terror attacks in an instrumental variable setting. Our instrument for the stock of foreigners relies on the interaction between two sets of variables. Variation across host-origin-dyads results from structural characteristics between the country of origin and the host, while variation over time (and dyads) makes use of changes in push and pull factors between host and origin countries resulting from natural disasters. 6 Controlling for the levels of these variables themselves and fixed effects for dyads and years, the interaction provides a powerful and excludable instrument.
Using data for 20 OECD host countries and 187 countries of origin over the 1980-2010 period we find that terror increases with the number of foreigners living in a host country. This scale effect relating larger numbers of foreigners to more attacks does not imply that foreigners are more likely to become terrorists compared to the domestic population. When we calculate the effect of a larger native population on the frequency of terror attacks by natives, we find 3 The act was deemed unlawful in 2004, which is why the Prevention of Terrorism Act 2005 was passed, allowing the Home Secretary to impose "control orders" on everyone suspected of being involved in terrorism. 4 There is, however, evidence that the number of refugees hosted in a country are correlated with a larger number of terrorist attacks (Milton et al. 2013). 5 The benefits of immigration tend to be ignored when discussed in the context of terrorism (see Fitzpatrick 2002). 6 This follows established practice in the literature on migration (see Alesina et al. 2016, Doqcuier et al. 2016). this effect to be of comparable size. Overall, we thus conclude that migrants are not more likely to become terrorists than the natives of the country they live in.
We refine the basic analysis in several ways. With some exceptions, we do not find migrants coming from Muslim-majority countries and those coming from countries with particularly pronounced terrorist activity to be more likely to be engaged in terrorist activity than other foreigners. We also test whether and to what extent immigration and integration policies change the effect of foreigners on terror. We find that domestic policies relating to the integration and prospects of immigrants as well as immigration policies affect the probability that foreigners turn violent. More specifically, our results show that restrictions on migrants' rights and stricter immigration laws increase terror. Overall, we conclude that host country policies are key to fight terror, but in other ways than commonly perceived. It seems that stricter policies segregating foreigners already living in a country lead to alienation and thus increase the risk of terror. We also find that high-skilled migrants are associated with a significantly lower risk of terror compared to low-skilled ones, while there is no significant difference in terror arising from male compared to female migrants.
The next section discusses the previous evidence linking immigration to terrorism and introduces our hypotheses. We outline our data and methods of estimation in Section 3, and present results in Section 4. Section 5 tests robustness, while the final section concludes and discusses policy implications.

Terror and Migration
While there is no evidence of a systematic effect of immigration on terrorism, plenty of anecdotes and opinion-based writings, in concert with a number of descriptive evaluations of terrorist events exist. 7 Somewhat systematic evidence is offered in the few studies analyzing the vitas of known or suspected terrorists. Among these, Camarota (2002) investigates how 48 foreign-born Islamic terrorists entered and remained in the United States in the 1993-2001 period. Leiken (2004) focuses on 212 suspected and convicted terrorists in North America and Western Europe from 1993-2003. Kephart (2005) covers the immigration histories of 94 terrorists operating in the United States in the 1990-2004 period, while Leiken and Brooke 7 A particularly prominent example of opinion-based "analysis" is Michelle Malkin's (2002) bestseller Invasion, suggesting a range of discriminatory measures against immigrants to prevent the migration of terror. Regarding anecdotes, with the exception of the Oklahoma Bombings in April 1995 all major global terrorist attacks in Western countries in the previous decade have been conducted by immigrants (Angenendt and Cooper 2006). Among the sample of terrorists covered in Leiken (2004), 86 percent were Muslim immigrants, 8 percent were converts, and the remainder mainly African American Muslims.
(2006) coded 373 terrorists belonging to organizations with global reach over the years 1993-

2004.
All these studies find that terrorism is strongly associated with immigration. Camarota (2002: 5) consequently concludes that "there is probably no more important tool for preventing future attacks on U.S. soil than the nation's immigration system." Camarota (2002) summarizes activities of foreigners who immigrated to the United States long before engaging in terrorist activities. According to Camarota, between 700,000 and 900,000 people permanently immigrate to the United States every year. Roughly half of these previously entered the United States on temporary visas. 8 For example prior to 9/11, more than half of the terrorists in Camarota's sample have been long-term residents or naturalized citizens.
Among them are Mohammed Saleh -involved in the planned bombing of the United Nations building in New York, among others, in 1993. Saleh was a legal permanent resident (by marriage). The leader of this planned attack, Siddig Ibrahim Siddig Ali has also been a legal resident by marriage. Iyman Faris, another example, was sentenced to 20 years in prison for his involvement in a planned terrorist attack on the Brooklyn Bridge by al Qaeda in 2003. He entered the United States in 1984 and was naturalized in 1999 (Kephart 2005). Similarly, the metro and rail bombings in Paris during the mid-1990s have been conducted by "legal" French Muslim citizens of Algerian origin (Leiken 2004). The leader of the French cell responsible for the bombings, Khaled Kelkal, e.g., immigrated to France from Algeria as an infant in the 1970s (Leiken 2004). French Algerian Rachid Ramda -alleged mastermind behind the bombingshas been granted asylum in Britain in 1992 (Leiken 2004). In these and all of the other examples provided in Kephart (2005), immigration happened many years before the involvement in any terrorist activity. A more recent example is Najim Laachraoui who is alleged to be involved in the suicide terrorist attack on Brussel's airport in March 2016 (as well as in the Paris attacks of November 2015). He is suspected of building the bombs for both attacks and died in Brussels.
Laachraoui was born in Morocco but migrated to Belgium as a child. 9 Based on terrorists' vitas summarized in the previous literature, in the vast majority of cases, we would expect foreigners engaging in global terrorism to have lived in the country for a longer period of time rather than entering and immediately being engaged in an attack. Rather than entering as a terrorist, it seems likely they immigrate without the intention to be involved 8 Kephart (2005) finds that 16 of the 23 terrorists in her sample who applied for legal permanent residence once being in the United States have been successful in obtaining it, while 20 of 21 were successful in becoming naturalized U.S. citizens. Two thirds of the 94 foreign-born terrorists covered in her sample committed immigration fraud. 9 See http://www.nbcnews.com/storyline/brussels-attacks/najim-laachraoui-what-we-know-about-suspectedbomb-maker-n543996 (accessed November 13, 2016). in terrorism, and only later turn into terrorists. They get into contact with terrorists living in their host country or when returning to their country of origin for holiday or business. 10 In the empirical analysis below we therefore test whether and to what extent the stock of foreigners living in a country is related to the level of terror, rather than focusing on recent entrants.
More specifically, we are interested in whether foreign nationals living in a host country coincide with a larger number of terrorist attacks originating from nationals of the same country on their host country's population. Arguably, the absence of such a pure "scale effect" would be surprising. An increasing number of people living in a country mechanically increases the probability that some of them turn violent (Jetter and Stadelmann 2017). Such correlation is comparable to those between the size of the domestic population living in a country and the number of terrorist attacks pursued by them (Kruger and Maleckova 2003). In light of the scale effect population size has on domestic terror according to the previous literature, the absence of a positive correlation between the number of foreigners and the number of attacks pursued by foreigners would imply that foreigners are less likely to become terrorists than the domestic population. We consider this unlikely. It is therefore important to put the effect of foreigners on the number of foreign attacks in perspective, and compare them to how the number of natives affects terrorism by those natives.
We also allow for the possibility that migrants from different countries engage in terrorist activity to a different extent. The examples in Camarota (2002) and elsewhere suggest that foreigners with Muslim background are particularly likely to engage in terrorist activity.
As Enders and Sandler (2006) point out, the marginal costs of terrorism are particularly low in countries with large Muslim populations, while resources required to conduct terror are plenty.
We therefore test whether the effect of immigrants from Muslim-majority countries differs from those of other countries. We also test whether immigrants from countries where terror prevails are more likely to be involved in terror 11 and to what extent migrants are more prone to engage in terrorism if the host country is engaged in military conflict with the origin country. Conflict 10 As one example, consider the vita of Jose Padilla, who was arrested for suspected terrorist activity in 2002. He was a U.S. citizen and moved from Puerto Rico to Chicago at the age of 5. He converted to Islam in prison (Leiken 2004). Another example is Christian Ganczarski, convicted for his participation in the Djerba, Tunisia, synagogue bombing in April 2002. He immigrated from Poland at the age of 9 and became German national. He later converted to Islam, studied in Saudi Arabia, and undertook six trips to Afghanistan in the 1999-2001 period (Spiegel online, 1/5/2009, http://www.spiegel.de/international/europe/0,1518,599485,00.html, accessed May 5, 2010. Khaled Kelkal, the leader of the French cell responsible for the Paris metro and rail bombings came into contact with Islamism while in French prison (Leiken 2004). The three future 9/11 hijackers from the Hamburg cell came to Germany as legal immigrants and only later came in contact with fundamentalist networks (Leiken 2004). More than 40 percent of the 373 terrorists investigated in Leiken and Brooke (2006) are nationals of Western countries. 11 As Leiken (2004: 87) puts it: "For the production of terrorists what could be more ideal than Algeria -with its modern history of violent political struggle and a vicious fundamentalist resistance movement?". has been shown to either directly increase the risk of a country's citizens being involved in terrorist activity or to make them more violent in general (Montalvo and Reynal-Querol 2005, Esteban et al. 2012, Campos and Gassebner 2013. Regarding terror, Bove and Böhmelt (2016) provide evidence of a spatial spillover among countries. They show that countries closer to countries rich in terror are more likely to experience terror themselves (with "closer" being measured by the number of migrants from a country, among others). Terrorism has also been shown to be an important tool in interstate disputes (Findley et al. 2010). Hence, we expect foreigners born in countries with populations involved in substantial terrorist activity to be particularly violent.
The role of gender and education has also received attention in the previous literature.
While the earlier literature tends to characterize women as victims of terror, more recent discussions acknowledge their role as perpetrators as well (Agara 2015). We therefore examine the role of male and female immigrants separately in addition to investigating their joint effect.
We have however no clear hypothesis regarding the importance of gender for whether or not immigrants turn terrorists. The role of education is equally unclear. While many believe poverty to be among the root causes of terrorism, parts of the previous literature have shown that terrorists are often well educated compared to their peers (Kruger and Maleckova 2003).
We hypothesize that a host country's policies and environment are crucial in the fight against terror. One important dimension concerns the extent to which immigrants are integrated into the culture and society of their host country (Leiken 2004, Rahimi andGraumans 2015).
Well-integrated foreigners are less likely to engage in terror against their host country population. Tensions among the host and foreign populations, to the contrary, will increase the propensity (of foreigners and natives, arguably) to engage in terrorist acts. Most importantly, we expect terrorist groups to have an easier time recruiting foreigners for the fight against the host country's population, if they themselves are the target of political violence from the domestic population. Furthermore, we expect immigrants' prospects to earn their living and obtain positions of respect in their host countries to be crucial. Policies aimed at forced integration -putting pressure on immigrants to assimilate, learn the language of their host country or change the way they dress or exercise their religion -can turn either way. To the extent these policies are successful and result in better integrated immigrants they can be successful in reducing terror in the future. Yet, restrictions and pressure on immigrants on areas of their lives they deem important, can as well raise resistance and alienation and thus achieve the opposite effect.
A second important dimension of host country policies concerns their immigration policies. Policies on immigration are officially at least in part designed to reduce the risk of terror. It is however not clear if stricter immigration policies do in fact reduce terror committed by foreigners, since their effect on foreigners already living in the host country is not well understood. They could in fact be perceived as acts of repression, racism, and humiliation by foreigners already residing in the host country, thereby increasing the tendencies to engage in terrorism. While we cannot test the first mechanism we can test if stricter immigration policies reduce the risk that migrants engage in terror against their host country when immigration restrictions are put in place.

Data and Method
We aim to test whether a larger number of foreigners from a particular country causes more terrorist attacks from people of that nationality in their host country. We define TERRORhot as We take TERROR from the "International Terrorism: Attributes of Terrorist Events" (ITERATE) database (Mickolus et al. 2014). ITERATE provides data on global terrorist acts, including information about the nationality of perpetrators and victims. 13 Our data on migration are taken from the Institut für Arbeitsmarkt-und Berufsforschung's (IAB) brain-drain dataset (Brücker et al. 2013). The IAB defines "immigrants" as the number of foreign-born individuals aged 25 years and older living in a country other than the country they were born. 14 The data are based on harmonized census data 12 Note that we use a binary indicator since 99.5 percent of our dyad-year observations show no transnational terror events, while of the remainder, around 80 percent are one, 15 percent are between 2 and 4, and the remaining 5 percent range between 5 and 17 incidents. 13 Mickolus et al. (2014: 2) define transnational terrorism as "the use, or threat of use, of anxiety-inducing, extranormal violence for political purposes by any individual or group, whether acting for or in opposition to established governmental authority, when such action is intended to influence the attitudes and behavior of a target group wider than the immediate victims and when, through the nationality or foreign ties of its perpetrators, its location, the nature of its institutional or human victims, or the mechanics of its resolution, its ramifications transcend national boundaries." 14 The exception is Germany, for which data on foreign-born population before 2009 are unavailable, so that a citizenship-based definition of foreigners is used (Brücker et al. 2013: 3). Germany differs also as an origin country, since the migrant stocks of East-and West-Germany in other countries have been aggregated prior to unification. The same procedure was implemented for South-and North-Yemen. For a more detailed discussion of the IAB harmonization procedure, see Brücker et al. (2013). of 20 OECD host countries. The dyadic data show the stocks of immigrants from 187 countries of origin in the host countries in 5-year intervals over the 1980-2010 period. We linearly interpolate the years with missing data. Since the stock of foreigners typically evolves slowly over time we expect the noise that we introduce is inconsequential, while allowing us to exploit yearly variation in the terrorist data. 15 We report results without interpolations as a test for robustness.  To put these numbers in perspective, Table 1 reports the total number of terrorist attacks in each OECD country over the 1980-2010 period, along with the percentage of those numbers committed by NATIVES (against NATIVES or FOREIGNERS) compared to FOREIGNERS. As can be seen, the large majority of attacks originates from NATIVES. However, when we divide the number of attacks by the number of NATIVES and FOREIGNERS, respectively, the number of attacks per foreigner dominates by an order of magnitude. Specifically, for every one million people, 4.8 terrorist attacks are conducted by NATIVES, while the corresponding number for 15 Out of the full sample of 170,742 potential dyadic country-years, data for 39 percent are not available. We test robustness in two ways. One, we exclude host-and origin-country observations where inflows or outflows of migrants surge due to the effect of refugee crises (and noise introduced by linear interpolation is consequently most severe). Two, we use the OECD's (2015) International Migration Database instead, where gaps in the data however abound, so that we also rely on linear interpolation there (while the dataset includes up to 34 host and 184 origin countries, the total number of observations is substantially lower compared to the IAB data). This dataset provides two definitions of immigrants, relying on "passport-based" and "foreign-born" definitions. While we would like to include "immigrants" not only in the first, but also in later generations and so be able to test whether children or grandchildren of immigrants are more likely to become terrorists compared to citizens without any recent history of migration, such data are not available. 16 We calculate the number of NATIVES by subtracting the number of FOREIGNERS from the host country's total population, taking data on total population from the World Bank (2016). These data include foreigners, according to the World Bank's definition of the series: "Total population is based on the de facto definition of population, which counts all residents regardless of legal status or citizenship" (World Bank 2016). ITERATE exclusively includes terrorist events in which the location, perpetrator, and victim do not have the same nationality. Terror conducted by NATIVES of country h within h thus exclusively captures events in which NATIVES attack FOREIGNERS. Domestic attacks are those where both the perpetrator and the victim originate from the country the attack takes place (taken from Enders et al. 2011 andGaibulloev et al. 2012 based on data from the Global Terrorism Database, GTD). We test the effect of FOREIGNERS on TERROR estimating the following baseline specification with a linear-probability model (and clustering standard errors at the host-origindyad): where ℎ ′ is a set of time-varying control variables, ℎ are dyadic host-origin fixed effects, are year fixed effects, and ℎ is an error term.
In our main specifications we assume that terrorist attacks react to changes in our explanatory variables in the same year. This is likely to be the case if terrorist attacks are largely based on short-term changes that foreigners expect to affect their situation in the future or if the attacks are direct reactions to recent policy changes. We rerun all specifications including explanatory variables as (lagged) five-year moving averages to allow for longer lags between changes in policies and outcomes and the actions of terrorists.
Following the previous literature on bilateral terror (Blomberg and Rosendorff 2009, Neumayer and Plümper 2009, Plümper and Neumayer 2010 we include the natural logarithm of host and origin GDP as well as their populations as our basic control variables. 18 The resulting dataset covers more than 102,000 dyadic observations from 183 origin countries in 20 OECD countries, over the 1980-2010 period. This basic regression ignores the obvious problem of reversed causality and omitted variables bias. Migrants might choose their host country according to the risk of experiencing terror, but potentially also according to the ease of pursuing attacks there. A large number of omitted variables is arguably related to both terror and migration as well. We still report these basic results for comparison.
We proceed by including a number of interactions that test the more nuanced hypotheses introduced above: where ℎ , −1 represents the variables that we hypothesized to alter the effect of FOREIGNERS on TERROR in Section 2 above. 19 These variables are moving averages over five years, as we expect foreigners to react to a country's general trend in policies rather than year-to-year changes. We lag them by one period, since we assume that the effect of these variables on how migration affects terror is not likely to be immediate.
First, we include a ℎ , −1 indicator that measures the fraction of years a hostorigin pair is in a military conflict over the t-5 -t-1 period, based on data taken from the UCDP/PRIO Armed Conflict Dataset V.4-2015 (Gleditsch et al. 2002, Pettersson andWallensteen 2015).
Our second group of variables exclusively varies at the host-country level. The variable _ ℎ measures the number of terrorist attacks by NATIVES against foreigners in host country h and year t. ℎ is taken from the International Country Risk Guide (PRS Group undated), ranging between 1 and 6, with higher values representing fewer tensions. It measures "the domination of society and/or governance by a single religious group that seeks to replace civil law by religious law and to exclude other religions from the political and/or social process; the desire of a single religious group to dominate governance; the suppression of religious freedom; the desire of a religious group to express its own identity, separate from the country as a whole" (PRS Group undated). We include host country per capita GDP growth ( ℎ ) as proxy for expectations of future well-being. We include indicators of the restrictiveness of immigration, migrant rights, and repression and integration, following the general approach of Mayda (2010) and Ortega and Peri (2013). Like them we measure changes in "restrictiveness" with respect to the first year in our sample, based on data from the dyad-specific DEMIG database of the International Migration Institute (DEMIG 2015, de Haas et al. 2015). 20 In the initial year (usually 1980) restrictiveness is assigned a value of "zero," while in each following year the number of policies that make migration more restrictive is subtracted by the number of policies that make migration less restrictive. Thus our restrictiveness indicators rise in years in which more restrictive policies have been passed than less restrictive ones, and fall in years in which liberalizing reforms dominate. Specifically, we measure policies that either regulate the rights of foreigners living in the respective host country or the degree of surveillance and sanctions employed against them ( ℎ and ℎ ). 21 Higher scores imply that integration policies are more restrictive, fewer rights are granted, and surveillance is more extensive. ℎ covers policy measures that affect government agreements about worker recruitment, programs that resettle refugees, migrants' access to language programs or financial assistance, as well as religious and cultural integration programs, among others. Examples for policies covered by ℎ are controls on the movement and migration status of people (like the construction of fences or introduction of fingerprinting), rules on identification documents, procedures and criteria for the detention of foreigners, and employment permits.
We also use an integration policies index ( ℎ ), built in analogy to the ℎ and ℎ indices and covering restrictions on the naturalization of nonnative speakers, preferential naturalization for natives of particular countries, and regulations of permanent residency or work permits, among others (DEMIG 2015). Higher values on the index imply more restrictive policies.
Finally, we aim to test the effect of the host country's immigration policies. Our indicator is an ordinal measure of the restrictiveness of immigration policies, again based on 20 An obvious alternative to DEMIG is the International Migration Policy and Law Analysis (IMPALA) Database (Beine et al. 2016), which however currently covers only ten years from nine countries. 21 ℎ cover policies that fall into DEMIG's categories recruitment/assisted migration program, resettlement programs, language, housing and cultural integration programs, access to social benefits and socioeconomic rights, access to justice and political rights, access to permanent residency, and access to citizenship (DEMIG 2015 Our instrument relies on the interaction between two sets of variables. Variation across hostorigin-dyads results from structural characteristics between the country of origin and the host, while variation over time (and between countries) makes use of changes in push and pull factors between host and origin countries resulting from natural disasters. Our first-stage regression is as follows:

Important pull and push factors between host and origin countries include a binary indicator
showing whether or not the host and origin country share a common border, a (past or present) colonial relation, a common language (spoken by at least nine percent of the population), the logged great circle distance between them (in kilometers), and the log of the bilateral stock of foreigners in 1960 to capture preexisting networks.
We interact the structural variables with the vector of the total number of natural disasters in host and origin countries in a given year ( ℎ ), assuming that natural disasters in origin countries increase the importance of push factors for migration (Artuc et al. 2015, Docquier et al. 2016, while natural disasters within host countries reduce pull factors.
Note that we control for the number of disasters in the first and second stage regressions, while none of the structural variables forming part of our instrument vary over time, so that they are captured by the host-origin fixed effects in the second-stage regressions.
The regressions with interactions include a second set of instruments. We instrument both ℎ and ℎ * ℎ , −1 with the instruments of equation (3) as well as with these instruments' interaction with ℎ , −1 . The second-stage regression (excluding interactions) then looks as follows: The intuition behind the interacted instruments is based on a difference-in-difference approach: We investigate a differential effect of dyad-specific pull and push factors on the number of terrorist attacks in a year with fewer or more disasters. 23 A natural disaster in a country of origin makes migration to the OECD overall more attractive if this country is closer, has traditional migrant communities, or colonial ties. The dyadic characteristics would then be crucial in determining how many people affected by the disaster decide to migrate in reaction to the shock, and which host country they chose. In analogy, disasters in host countries make them less attractive and welcoming.
Our identifying assumption it that the effect of disasters on terror between a host and origin country does not depend on the dyadic pull and push factors, except for its effect on migration, conditional on the effect of disasters itself, dyad-and year-fixed effects, and other variables in the model. 24 Even if terrorists commit more or fewer attacks conditional on the dyadic characteristics -for example because they target their former colonial powers more often -and would react with more or less terror to natural disasters in either host or origin countries, this would not violate the exclusion restriction. If terror would however systematically vary with the dyad-specific push and pull factors resulting from the number of disasters in host and origin countries our identifying assumption would be violated. Since the actual number of natural disasters in host and origin countries are random to terror incidents, and with respect to dyad-specific structural variables, we consider this unlikely. 23 We follow the previous literature and use the number of natural disasters rather than disaster outcomes such as deaths or destruction (Docquier et al. 2016), since the latter two are more likely to be correlated with terrorist activity in the origin or host country, e.g., blocking relief organizations from distributing emergency relief. 24 Bun and Harrison (2014) and Nizalova and Murtazashvili (2016)

Results
Column 1 of Table 2 shows the results of the baseline regression, estimated with OLS (equation  (Table 2, column 2). We do not include origin-country GDP in this regression, which would be undefined for the domestic terror regression. We also do not include data for the year 1993, for which GTD does not report data on terrorist events. 26 According to the results, there is no significant difference among the two sets of regressions. The average scale effect of the total population on the probability of terror is positive and significant at the one percent level. However, while the point coefficient estimating the difference for terror originating from the native population compared to the total population is large, this difference is not significant at conventional levels.
Column 3 of Table 2 replaces the binary dependent variable with the number of attacks in a country-dyad and year. Again, the difference between average terror and terror by NATIVES is not statistically significant. When we calculate the elasticity at the sample mean of transnational terror incidents (0.028), we find that a one percent increase in the stock of  Columns 2-9 include the interacted variables (and their levels). In addition to migration, we also instrument the interaction of the respective variable with migration (using the variable in levels interacted with our instruments as additional regressors). We have however no suitable instruments for the level of the interacted variable itself. This implies that as long as there is no feedback between the potentially endogenous variables, the coefficients of the interaction terms are estimated consistently, though we cannot interpret the coefficients of the policy variables themselves. Therefore we cannot infer whether or not these policies themselves increase or reduce the risk of terror. 28 Note that feedback between our potentially endogenous interaction variables and terror attacks occurs if attacks would be planned in anticipation of future policy changes -or rises or falls with religious tensions -and would then be executed a year after the change occurs. We consider this unlikely since all our interaction variables are moving averages over the period −5 to −1 .
In columns 2 and 3 we test the importance of conflict between the host and origin countries and of terror against foreigners' in host countries. We introduce ℎ (column 2) and _ ℎ (column 3) and their interaction with the number of foreigners. While we do not find a significant interaction with military conflict, we find that more attacks of natives against foreigners within the host country increase the average effect of migrants on terror, at the one percent level of significance.
Columns 4-9 introduce the variables measuring domestic policies and outcomes and their interactions with the number of foreigners. The results show that fewer religious tensions 28 See again the references we refer to in footnote 24. reduce the probability that foreigners turn violent, at the ten percent level of significance (column 4). We find that laws putting pressure on migrants to integrate increase the probability of terror associated with a rising number of foreigners in a country (column 6). Stricter sanctions on migrants not behaving in line with expectations seem to reduce the threat of terror associated with the number of foreigners (column 8). Stronger restrictions of foreigners' rights however increase the risk of terror coming with a larger migrant population from the same country (column 7). We do not find significant interactions with GDP per capita growth (column 5) and restrictions on immigration (column 9). 29  Figure 4 shows that terror vs. foreigners in the host countries substantially affects the risk of terror arising from any given stock of migrants in a country. At the mean value of terror against foreigners, a one percent increase of the stock of foreigners increases on average the probability of a terrorist attack committed by foreigners by 0.0379 percentage points. The corresponding increase is 0.0445 percentage points at the maximum value of terror against foreigners (17% higher compared to the mean). To the contrary, while the effect of (the absence of) religious tensions is statistically significant (Figure 5), the difference of a one percent increase in the stock of foreigners at mean religious tensions is hardly distinguishable from those at the maximum  significant over the entire range of the distribution, the effect is small in quantitative terms: At the mean value of the index, a one percent increase of the stock of foreigners leads to an increase of 0.0394 in the likelihood of terror compared to an 0.0369 increase at the maximum (corresponding to a 6.4% reduction). While of opposite direction, the difference in percent is thus substantially smaller compared to the integration and rights interactions discussed above. 29 Results are similar when we estimate the regressions with OLS. The exceptions are the interactions with religious tensions (which turns insignificant) and immigration restrictions (which is significant and positive).
Overall, we conclude that migration policies are key in the fight against terror. The optimal mix however is intriguing. Countries that put too much pressure on immigrants to integrate and restrict their rights are likely to achieve the opposite of what they aim for, at least in the short-run. Immigrants already living in the country might turn against their host and get increasingly violent. 30 Tables 4 and 5 investigate whether migrants from some countries are more 'dangerous' than others, focusing on migrants from Muslim and terror-rich countries. In order to allow comparisons between countries rather than within dyad-pairs over time exclusively, we replace the dyad-fixed effects in our regressions with dummies for individual host countries. We add individual country dummies indicating migrants from Muslim and "terror-rich" countries. The resulting coefficients can thus be interpreted as the difference in the average partial effect of migrants from these countries compared to all other countries (i.e., those we either define as non-Muslim or not "terror-rich"). As before, we instrument the stock of foreigners with the interaction of natural disasters and the pull and push factors introduced above. Table 4 investigates whether Muslim foreigners are more or less likely to engage in terrorist activity compared to average non-Muslim foreigners. Our binary indicator is one if Islam is the religion followed by the majority of the country's population according to the CIA World Factbook. 31 We show the marginal effect and its standard error. Given that we are primarily interested in testing whether Muslim foreigners are more likely to engage in terror compared to other foreigners we also show the p-value corresponding to a test of whether the marginal effect of a specific country differs from those referring to the average foreigner from a non-Muslim country. It turns out that foreigners from most Muslim countries do not differ in how they affect terror against their host from the average non-Muslim country ("Reference Group"). 32 The two exceptions are Algeria and Iran. Compared to the (insignificant) average effect of foreigners from non-Muslim countries the marginal effects imply that a one percent increase in the stock of Algerian migrants increases the likelihood of terror by 2.1 percentage points in the average OECD country. The corresponding effect for Iranian migrants is 1.5 percentage 30 As an illustration, consider France. According to the DEMIG (2015) data, France introduced 18 additional restrictions on immigration over the 1991-1994 period. This included prohibiting foreign graduates from gaining employment in France and suppressing work permits for asylum seekers. In 1994, France restricted the access and right of residence for Algerians (DEMIG 2015). France suffered a spell of terrorism in the following year with at least one attack per year committed by an Algerian citizen over the 1995-1999 period. 31 Available at: https://www.cia.gov/library/publications/the-world-factbook/fields/2122.html (last accessed August 11, 2016). 32 Jetter and Stadelmann (2017) show that the probability that Muslims turn terrorists is smaller compared to non-Muslims once population size is taken account of.

points. The former effect is mainly driven by attacks from Algerian fundamentalists and the
Islamic Armed Group Algeria (GIA) which participated in 12 attacks in France in the late 1980s to mid-1990s related to integration issues and fights for an Islamic state. The latter effect is driven by 18 attacks against each France and Germany in the 1980s and early 1990s by Iranian nationals. There is no dominant terror organization behind these attacks in Germany while one third of the French attacks where conducted by Islamic Jihad organizations. 33 Table 5 compares migrants from "terror-rich" countries to the effect of foreigners from the average "non-terror-rich" country. We include the binary variable , indicating that a country is located within the top quintile of the overall terrorist incident distribution of the GTD dataset. 34 Five countries show marginal effects higher than the reference group, at least at the ten percent level of significance. Compared to the average "nonterror-rich" country, migrants from Algeria, Iran, India, Spain and Turkey are all more likely to be involved in a terrorist attack, while migrants from Angola and Cambodia are less likely than the reference group to commit terror. Some background for Algeria and Iran is given above. In Finally, we test whether the composition of migrants matters. Column 1 of Table 6 separately investigates male and female migrants. Column 2 distinguishes migrants with low, medium, and high skills. 35 As additional set of instruments for the stock of male and female migrants we add the interaction of our instruments with the share of male migrants from a country of origin to a specific host country over the entire sample period. For the stock of low skilled, medium skilled and high skilled migrants, we add interactions of our instruments with 33 In our sample, Algerian terrorists conducted 34 terror attacks in total while citizens of Iran conducted a total of 80 attacks. 34 We focus on GTD rather than ITERATE as we are interested in overall terror at the origin-country level rather than exclusively transnational terror exposure or the terror against specific groups. 35 The IAB database defines the skill levels as follows: Low-skilled individuals have received lower secondary, primary or no schooling. Medium-skilled migrants have obtained a high school diploma or equivalent certificate. High-skilled immigrants have tertiary education (Brücker et al. 2013: 4). the shares of low skilled and medium skilled workers among each dyad over the sample period.
As can be seen in Table 6, our instruments are highly relevant.
The results of column 1 show that the risk of terror increases with the number of male immigrants, at the one percent level of significance, but not with the number of female immigrants. The coefficients of the two groups are however not statistically different from each other (p-value: 0.54). Column 2 shows that the risk of terror increases with low-skilled immigrants, but decreases with high-skilled immigrants, the difference between the two being significant at the one percent level. While the previous literature has often argued that terrorists are well educated compared to their peers (Kruger and Maleckova 2003), the same does not seem to hold for immigrants who turn terrorists. This is in line with the game theoretical model of Bandyopadhyay and Sandler (2014), showing that increases in skilled labor quotas generally reduce terrorist attacks in the host country.
In summary, we find a positive scale-effect of larger foreign populations comparable in size to the scale-effect of natives on the likelihood of a terror attack. We find this scale-effect to be more severe when migrants are situated in host countries where terror against foreigners is prevalent and religious tensions abound, when migrant rights are restricted and integration laws are tough. The risk of terror is lower when sanctions against migrants are frequent. We find no significant difference between male and female immigrants on the risk of terror. Highskilled immigrants reduce the risk of terror, while low-skilled immigration increases risk. The next section tests the robustness of these main results.

Tests for robustness
We test the robustness of our main results in a number of important dimensions. First, we include all explanatory variables as (lagged) five-year moving averages to allow for longer lags between changes in policies and outcomes and the actions of terrorists.
Second, we use yearly values for our interaction variables, rather than five-year moving averages.
Third, we test whether and to what extent linear interpolation of the migration data affects our results. Instead of interpolating, we use averages over five years (but no moving average).
Fourth, we test whether and to what extent our results are driven by dyads in which the stock of foreign-born natives changes substantially, for example due to refugee crises and the resulting surge in immigrants. Specifically, we exclude the five percent largest changes in migration in our sample.
Fifth, we employ additional instrumental variables. Following Beine and Parsons (2015) we add deviations in temperature and precipitation 36 as interaction variables to our structural ones. 37 The resulting sample is much smaller, however.
Sixth, we test whether our results are robust to holding factors constant at the periodhost-origin level that have been identified as most robust determinants of terror in the previous literature, such as unemployment, growth, or economic freedom (Gassebner and Luechinger 2011). We therefore include fixed effects for origin-year, host-year and origin-host. We consequently rely on dyadic within-variation exclusively to identify coefficients.
Seventh, we test whether the effect of migrants on terror persists when we investigate terror events committed by both domestic nationals and foreigners, rather than exclusively those conducted by foreigners. The reason is that transnational terror covers attacks by nationals with foreign passports, while our definition of foreigners relates to migrants born abroad. To the extent that attacks originate from migrants already in possession of a domestic passport, the results of our regressions above would underestimate the true effect of migrants. We therefore replicate the regressions focusing on domestic and transnational events.
Finally, we test whether our results are driven by our focus on all terror events rather than focusing on severe events only. Hence we replace all terror events with terror events during which at least one victim got wounded or killed.
As can be seen from Table A-5 in the Appendix most of our results turn out to be robust to these modifications. The effect of a one-percent increase in the stock of migrants on the probability of transnational terrorist attacks ranges between 0.025 percentage points (when we include the additional fixed effects) and 0.044 percentage points (when we use period averages and no interpolation). This is similar to the main estimate of 0.043 percentage points from Table   3 above, that we reproduce in Table A-5 for comparison ("no moving average"). The coefficient obviously increases substantially in magnitude when we estimate the effect of migrants on the risk of transnational and domestic terror ("total terror"). A one percent increase in the stock of migrants increases the risk of terror by 3.24 percentage points. However, this effect includes attacks conducted by non-migrants as well and thus overestimates the risk of terror originating from immigrants. The religious tensions interaction holds only in two of the five robustness tests and changes its direction when we pool domestic and foreign terrorist attacks.
Table A-5 also shows that terror from natives against migrants in the respective host country increases the risk that migrants resort to transnational terror. Surprisingly, the effect on 36 We thank Sven Kunze for sharing his temperature and precipitation data (Kunze 2016). 37 We calculate deviations in temperature and precipitation as the yearly deviations from decade averages.
total terror from migrants is reduced if citizens engage in terror against foreigners. With respect to policies, it turns out that while the harmful effect of strict integration policies and restrictions of migrant rights prevails in any regression, the beneficial effect of strict sanctions turns insignificant in four out of the seven additional regressions.
Finally, the results for severe terror attacks are overall in line with the results from the main regressions (though the interactions with religious tensions and with sanctions are not significant at conventional levels).
Table A-6 turns to the results for gender and skills. Overall, the previous results for transnational terror are highly robust. There is no evidence that male migrants are more likely to lead to terrorist attacks than women, while the risk associated with high-skilled migration is lower compared to those of low-skilled migration in three of the five additional regressions.
The next section concludes and discusses policy implications.

Conclusions
Over the last 15 years a number of countries have substantially tightened immigration laws and introduced policies putting pressure on migrants to integrate into their host countries, including restrictions on migrants' rights as well as surveillance and sanctions. These changes have been caused by expectations that a larger number of foreigners living in a country increases the risk of terrorist attacks in the host country. This paper has put these expectations to the data, for 20 OECD host countries and 187 countries of origin over the 1980-2010 period.
First, we tested the hypothesis that the stock of foreigners residing in a country leads to a larger number of terrorist attacks. Our instrument for the stock of foreigners relies on the interaction between two sets of variables. Variation across host-origin-dyads results from structural characteristics between the country of origin and the host, while variation over time makes use of changes in push and pull factors between host and origin countries resulting from natural disasters. Controlling for the levels of these variables themselves and fixed effects for dyads and years the interaction provides a powerful instrument.
Our results show that the probability of a terrorist attack increases with a larger number of foreigners living in a country. This scale effect relating larger numbers of foreigners to more attacks does however not imply that foreigners are more likely to become terrorists compared to the domestic population. When we calculate the effect of a larger domestic population on the number of times they attack foreigners or natives, we find this effect to be substantially larger.
Overall, we thus conclude that migrants are not more likely to become terrorists compared to the nationals of the country they live in.
Second, we tested whether migrants from countries rich in terror or from Muslimmajority countries affect the risk of terror differently and whether and to what extent host country immigration and integration policies mediate the risk arising from foreigners. We find scarce evidence that terror is systematically imported from countries with large Muslim populations. There is however evidence that immigrants from some countries rich in terror increase the probability of terror in their host country.
Contrary to the expectations of politicians, introducing strict laws that regulate the integration and rights of migrants does not seem to be effective in preventing terror attacks from foreign-born residents. Terrorist attacks have made politicians across the Western world severely diminish the very rights they aim to protect (Dreher et al. 2010), without, it seems, achieving the desired increase in security. To the contrary, repressions on migrants already living in the country alienates substantial shares of the population, which overall increases rather than reduces the risk of terror.
We conclude with two qualifications. First, our results are based on data for the group of migrants from a particular country, but the number of terrorist attacks by all foreigners. While we can thus estimate the risk of terror associated with a larger number of migrants, we cannot test whether migrants from a particular country are themselves engaged in terrorist events. Such analysis would require more detailed (individual-level) data than are currently available for a large sample of countries and years.
Second, an analysis of whether or not migration should be restricted has to involve a broader calculation of its costs and benefits. Driving fast on motorways leads to accidents and fatalities, planes crash and people die, and more people living in cities leads to a larger number of murder cases. Few people favor strict bans on motorways and planes, or cities. In a similar vein, a larger number of people leads to a higher risk that some of them turn terrorists. This holds for native and foreign populations alike, and by itself does hardly qualify as reason to ban migration. Rather, the increased risk of terror has to be confronted by the many other -positive and negative -effects that come with immigration. We leave such analysis for future research.   Notes: Estimates are based on the average number of natives and foreigners within the host countries during the 1980-2010 period (1993 is omitted due to lack of data). The total amount of terror attacks refers to the sum of terror attacks committed within the host country, by nationals against nationals (Enders et al. 2011 andGaibulloev et al. 2012), by nationals against foreigners (ITERATE 2015) and foreigners within the host country regardless of target nationality (ITERATE 2015).  (1) and (2) is binary and indicates that at least one transnational attack occurs in a year. Column (3) uses the number of transnational attacks per year. In column (4) the binary indicator is one if a transnational terror attacks occurs in a given year which results in at least one wounded or killed victim. Column (5) uses the number of those attacks per year. In the case of natives also domestic attacks are included. Robust standard errors clustered on host-origin dyad in parentheses; *** p<0.01, ** p<0.05, * p<0.1. Table 3: Terror andMigration, 2SLS, 1980-2010 (1) (2) (3) (4) (5) (6) (7) (8) (9) Log GDP host -0.0644*** -0.0643*** -0.0534*** -0.0659** -0.0683*** -0.0694*** -0.0717*** -0.0620*** -0.0622*** (0    Notes: The dependent variable is binary and indicates that at least one transnational attack occurs in a year. Robust standard errors clustered on host-origin dyad in parentheses; *** p<0.01, ** p<0.05, * p<0.1.

Variable Source Definition
Transnational terror attacks Mickolus et al. 2014 Sum of yearly incidents of terror attacks from nationals of an origin country within the host country.
Domestic terror attacks Enders et al. 2011, Gaibulloev et al. 2012 Terror from nationals against nationals within the country.
Terror against foreigners (by natives) Mickolus et al. 2014 Terror from nationals against foreigners within the host country.
Transnational terror dummy Mickolus et al. 2014 Dummy that is one if at least one terror attack was committed by a national of an origin country within the host country during a year.
Transnational terror attacks (severe) Mickolus et al. 2014 Sum of yearly incidents of terror attacks from nationals of an origin country within the host country, in which at least one victim was wounded or killed.
Transnational terror dummy (severe) Mickolus et al. 2014 Dummy that is one if at least one severe terror attack was perpetrated by a national of an origin country within the host country during a year.
Log of foreign-born residents IAB Database, Brücker et al. 2013 Log of total bilateral foreign-born residents from an origin country.
Log of foreign-born male residents IAB Database, Brücker et al. 2013 Log of total bilateral foreign-born male residents from an origin country.
Log of foreign-born female residents IAB Database, Brücker et al. 2013 Log of total bilateral foreign-born female residents from an origin country.
Log of foreign-born residents low skilled IAB Database, Brücker et al. 2013 Log of total bilateral foreign-born low skilled residents from an origin country.
Log of foreign-born residents medium skilled IAB Database, Brücker et al. 2013 Log of total bilateral foreign-born medium skilled residents from an origin country.
Log of foreign-born residents high skilled IAB Database, Brücker et al. 2013 Log of total bilateral foreign-born high skilled residents from an origin country.
Log of natives World Bank 2016, IAB Database, Brücker et al. 2013 Log of total population minus the total foreignborn resident stock.  Terror rich country dummy Enders et al. 2011, Gaibulloev et al. 2012 Dummy that is one if a country is in the top quintile of the domestic terror distribution over the whole sample. Notes: Column 1 shows the first stage corresponding to column 1 of Table 3 (including host-origin and year fixed effects). Column 2 includes fixed effects for origin-year, host-year and origin-host (Column 1, row 6, in Table A -5). Robust standard errors in parentheses. Significance levels: *** p<0.01, ** p<0.05, * p<0.1 Notes: None shows the coefficient of log migrants, without interaction. The remaining columns show the coefficient of the interaction. All moving averages includes all explanatory variables as five-year moving averages (lagged by one year). No moving averages uses yearly values for the interaction variables, rather than five-year moving averages. Period averages uses averages over five years (but no moving average). Exclude outliers excludes the five percent largest changes in migration in our sample. Additional instruments adds deviations in temperature and precipitation as interaction variables to our set of instruments. High dimensional FE includes fixed effects for origin-year, host-year and origin-host. Total terror relies on domestic and transnational terror events to code the dependent variable. Severe terror incidents involve at least one victim wounded or killed.  Table A-5. The p-values correspond to t-tests testing the equality in coefficients for male and female migrants and low-and mediumskilled migrants. We do not show results for No moving averages and High dimensional FE. Given that we do not include interactions, there are no moving averages in any of the regressions. When we include the additional fixed effects, the first-stage F-statistic is insufficiently low, so we do not report these (insignificant) results in the table. Severe terror incidents involve at least one victim wounded or killed.