Policy Research Working Paper 8381 Transnational Terrorist Recruitment Evidence from Daesh Personnel Records Anne Brockmeyer Quy-Toan Do Clément Joubert Mohamed Abdel Jelil Kartika Bhatia Development Research Group & Middle East and North Africa Region Office of the Chief Economist March 2018 Policy Research Working Paper 8381 Abstract Global terrorist organizations attract radicalized individuals opportunities and migration costs interact to explain the across borders and constitute a threat for both sending and spatial pattern of foreign participation in the terrorist group. receiving countries. The paper provides plausibly-identified Poor labor market opportunities generally push more indi- evidence on the drivers of transnational terrorist recruitment. viduals to join Daesh, but they hamper recruitment in Using unique personnel records from the Islamic State in Iraq countries far away from the organization’s headquarters, as and the Levant (ISIL, a.k.a. Daesh), it shows how economic migration costs are large and liquidity constraints may bind. This paper is a product of the Development Research Group and the Office of the Chief Economist, Middle East and North Africa Region. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at ichaaldabi@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Transnational Terrorist Recruitment: Evidence from Daesh Personnel Records∗ ement Joubert, Anne Brockmeyer, Quy-Toan Do, Cl´ Mohamed Abdel Jelil, and Kartika Bhatia† JEL classification: F51, E24, E26, Z12 Keywords: transnational terrorism, violent extremism, unemployment, migration costs ∗ We are grateful to Pierre Bachas, Jishnu Das, Shantayanan Devarajan, Rafael Dix-Carneiro, Hideki Mat- sunaga, Daniel Lederman, Steven Pennings, Jacob Shapiro, two anonymous referees and workshop par- ticipants at CSAE, ESOC, LACEA (AL CAPONE), National University of Singapore, the World Bank and the World Congress of the IEA for helpful discussions. We are also grateful to Zaman Al Wasl and Fathi Bayoud for facilitating access to the data on Daesh foreign recruits. Sarur Chaudhary provided excellent research assistance. The findings, interpretations, and conclusions expressed in this work do not necessar- ily reflect the views of the World Bank, its Board of Executive Directors, or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. † Macroeconomics, Trade & Investment Global Practice (World Bank) and Institute for Fiscal Studies; Re- search Department (World Bank); Research Department (World Bank); Human Development Unit (World Bank); ASPIRE India, respectively. 1 Introduction A new wave of terrorism has surged in the past two decades, characterized by transna- tional attacks and global recruitment, and spearheaded by multinational terror groups such as Al-Qaida and the Islamic State in Iraq and the Levant.1 An unprecedented num- ber of foreign fighters - over 25,000 - travelled to Iraq and Syria between the start of the Syrian Civil War in 2011 and September 2016 to fight for Daesh or for the Al-Nusra Front. These foreign fighters also come from a more diverse set of countries than in previous wars. United Nations (2017) reports that, by May 2015, Daesh had recruited fighters from over 100 countries. Some of these fighters have engaged in extreme levels of violence in Iraq and Syria, others have perpetrated terrorist attacks in third countries, and those who ultimately return to their home countries are viewed as threats to domestic security (The Atlantic 2017). Quantitative evidence on the economic drivers of transnational terrorist recruitment is scarce.2 In contrast, domestic terrorism has been the subject of more extensive research, as recently reviewed by Gaibulloev and Sandler (2019). Berman and Laitin (2008) contend that modern religious terrorist groups rely on their ability to limit their recruits’ outside economic opportunities, in contrast to the ideologically-motivated left-wing or nation- alist groups of the past. Empirically, however, evidence on the effect of economic op- portunities on terrorism is mixed.3 Bandyopadhyay and Younas (2011) and Enders and Hoover (2012) further observe that domestic and transnational terrorism may respond differently to local economic conditions.4 For instance, engaging in domestic terrorism can be a part-time occupation and does not require the recruit to travel long distances. By contrast, joining an international terror group involves migration costs in addition to 1 ISIL, a.k.a. ISIS or Daesh, its Arabic acronym. 2 Existing studies have investigated the ideological motivations of foreign recruits (Hegghammer 2010) or analyzed the process of radicalization and recruitment at the individual level (Weggemans, Bakker and Grol 2014, Gates and Podder 2015, Holman 2016). These case studies have gathered invaluable insights into the motivations of foreign fighters through interviews with the fighters and their contacts, yet they do not attempt a quantitative assessment of the drivers of recruitment. 3 See Krueger and Maleˇ ckov´ a (2003), Li and Schaub (2004), Abadie (2006), Krueger (2007), Lai (2007), Krueger and Laitin (2008), Gassebner and Luechinger (2011), Santiford-Jordan and Sandler (2014), and Enders, Hoover and Sandler (2016). 4 These studies present separate cross-country correlations for the two phenomena, but do not delve into the mechanisms that could distinguish them. 1 forgoing earning opportunities at home, a combination of mechanisms that has received little attention in the literature on terrorism. This paper exploits a unique data set of Daesh’s personnel records to study how eco- nomic opportunities and migration costs interact to explain the spatial pattern of foreign participation in transnational terrorist organizations. The data set contains information on 3,965 foreign recruits from 59 countries, including their age and education. Dodwell, Milton and Rassler (2016) estimate that these data account for approximately 30 percent of the total number of foreign recruits who entered Syria between early 2013 and late 2014. Our main explanatory variable is the unemployment rate in the countries of origin of these foreign recruits, a first-order measure of economic opportunity costs. The individual information contained in the Daesh personnel records allows us to move beyond cross-country correlations and control for any observed and unobserved country characteristics that may affect both terrorism participation and labor market op- portunities, such as institutions, government policies, and state capacity (Fearon and Laitin 2003, Sanchez de la Sierra 2019). Specifically, we link the number of Daesh re- cruits from a particular country and education group to the unemployment rate faced by workers in that same country and with the same education level. We run panel regres- sions that include country- and education-level fixed effects so that identification relies on within-country correlations between the schooling gradient of the unemployment rate and the relative number of recruits from each schooling group. Therefore, we contribute plausibly causal estimates of the impact of economic conditions on terrorism participa- tion that are informed by a new data source and a different identification strategy than in the previous literature.5 Theoretically, unemployment has an ambiguous effect on foreign terrorist recruitment. On the one hand, unemployment lowers the economic opportunity cost of participa- tion in terrorist activities and exacerbates grievances against the government (Collier and Hoeffler 2004, Collier and Hoeffler 1998, Blattman and Miguel 2010). On the other hand, 5 Krueger and Maleˇ ckov´ a (2009) propose a related identification strategy to investigate how public opin- ion of residents in one country towards another country predicts the incidence of terror events perpetrated in the latter country by citizens of the former. Their unit of observation is a country dyad, which makes it possible to control for both sending-country and receiving-country fixed effects. 2 unemployed individuals may face liquidity constraints that can hamper their ability to travel to the Mashreq region. This mechanism is more relevant in far-away countries where travel costs are higher. To disentangle the opposing effects of unemployment on terrorist recruitment, we first consider countries in the neighborhood of Iraq and Syria where the role of travel costs should be minimal. For this sample of close countries, we find that higher unemployment rates push more recruits to join Daesh, with a semi- elasticity of 0.16. Given available estimates of the total flow of fighters from that area in the period covered by our data, this estimate implies that 1,200 fewer recruits would have joined Daesh during that time if the unemployment rate had been 1 percentage point lower in all countries in the sample. As more distant countries are added to the analysis, the estimated elasticities drop until they become indistinguishable from zero for coun- tries at a median distance from Iraq and Syria. However, among countries furthest away to Iraq and Syria (located more than 2500 miles away), we find that unemployment rates negatively affect recruitment to Daesh, with a semi-elasticity of -0.15. Therefore, we hy- pothesize that travel costs to Iraq or Syria from such distances are high enough to become a binding constraint for some unemployed individuals wishing to join Daesh. The spatial heterogeneity in the effect of unemployment on recruitment is robust to a large number of alternative specifications, allowing us to discard competing interpre- tations. First, we show that the results hold within sub-samples constituted of Muslim- majority or Muslim-minority countries; when controlling for average wages; with alter- native estimators such as the Poisson Pseudo Maximum Likelihood estimator; and with alternative distance measures. Second, we use data on domestic terrorism across the world to show that the availability of domestic terrorism opportunities is unlikely to ex- plain our results. Third, we show that the heterogeneous effect of unemployment at dif- ferent distance levels is not explained by country-level factors that would be correlated with migration costs. The distance-unemployment interaction in our regression domi- nates competing interactions between unemployment and GDP per capita, the share of the Muslim population, or regional dummies. Therefore, we conclude that the variation in migration costs between countries of origin and the headquarters of the terrorist orga- nization is a credible driver of the spatial heterogeneity of the effect of unemployment on 3 recruitment. Our paper contributes to several strands of literature. First and foremost, our work adds to the emerging scholarship on the economic drivers of transnational terrorism. In addition to Bandyopadhyay and Younas (2011) and Enders and Hoover (2012) mentioned earlier, our paper is closely related to Verwimp (2016) and Benmelech and Klor (2018). Benmelech and Klor (2018) ask a question similar to ours, but use a country-level mea- sure of terrorist recruitment, estimated from a variety of sources such as social media or investigations. Therefore, their results rely on a different source of data and on cross- country, rather than within-country, variation. We nonetheless replicate their results by aggregating our individual records by country as a data check exercise. The study by Ver- wimp (2016) emphasizes the difference in labor market outcomes between EU natives and non-EU immigrants and finds that larger gaps are associated with higher numbers of for- eign fighters. As in Benmelech and Klor (2018), the analysis relies on cross-country vari- ations, which makes it vulnerable to country-level confounders, unlike our fixed-effects estimates. Admittedly, our measure of labor market opportunities is not specific to the Muslim or non-native population as in Verwimp (2016)), but we conduct a large number of robustness checks in section 4.3 to ensure that this is not driving our results. In particu- lar, running our regressions within subsamples of muslim-majority and muslim-minority countries yields similar results. The spatially heterogeneous relationship between local socio-economic conditions and the transnational recruitment of terrorists that we uncover mirrors findings in the interna- tional migration literature that emphasize the non-monotonic relationship between eco- nomic development and migration (Clemens 2014). Our result on geographically close countries — that economic opportunities at home reduce participation in terrorism — is consistent with the literature on micro-economic drivers of violent conflict (Verwimp, ¨ 2018); similar findings emerged in many different local contexts and Justino and Bruck for various forms of violence. For instance, the violence-dampening effect of improved labor market opportunities has been found among youths susceptible to crime in Chicago (Davis and Heller 2019), Liberian ex-combatants (Blattman and Annan 2010), Indian vil- lagers affected by the Maoist rebellion (Fetzer 2019, Dasgupta, Gawande and Kapur 2017), 4 or insurgents in Afghanistan, Iraq, or Pakistan (Guardado and Pennings 2019).6 The rest of the paper is organized as follows. In section 2, we describe the data sources used in the paper and provide evidence that the personnel records on Daesh recruits are consistent with the existing information used in the literature. Section 3 discusses our empirical strategy and section 4 presents the two main results and robustness tests. Section 5 concludes. 2 Data Sources The analysis conducted in this paper combines personnel records on Daesh foreign re- cruits and socio-economic information about the countries of residence of these individ- uals before they joined the terrorist group. 2.1 Daesh personnel records Daesh personnel records were obtained by a number of news organizations including Syria’s Zaman al Wasl (who in turn shared the data with the World Bank), Germany’s ¨ Suddeutsche Zeitung, Westdeutscher Rundfunk, and Norddeutscher Rundfunk, Britain’s Sky News, and NBC News in the U.S.. The latter described a Daesh defector as their source for the documents. Our data are identical to the ones described in Dodwell et al. (2016), who provide a detailed description of their origin and were able to corroborate 98% of the records with data maintained by the U.S. Department of Defense. The data set contains information on 3,965 foreign recruits from 59 countries. The in- formation is on foreign recruits who joined the ranks of the terrorist group in Iraq and Syria rather than on individuals who remained in their home country and pledged al- legiance to the organization. The records include information on a recruit’s country of 6 Berman, Callen, Felter and Shapiro (2011b) on the other hand find a negative relationship between unemployment and localized violence in Afghanistan, Iraq and the Philippines. They suggest that local unemployment can affect conflict by changing civilians’ incentives to side with the government in its fight against insurgencies. In particular, the authors argue that higher unemployment rates could lower violence by lowering the government’s cost of buying information about insurgents from civilians. This mechanism is less relevant in the context of trans-border terrorism, where recruits travel to join the terrorist organization in another country. 5 residence, citizenship, education, age and marital status. Table 1 provides a breakdown of records by country of last residence. Dodwell et al. (2016) estimate that these data ac- count for approximately 30 percent of the total number of foreign recruits who entered Syria between early 2013 and late 2014. All individuals in our sample are male, although the terrorist group is known to have also recruited females (Windsor 2018). Although the nature of the sample selection cannot be precisely established, the distri- bution of countries of origin – our main outcome variable – is highly consistent with the existing publicly available information, which Benmelech and Klor (2018) use.7 Figure 1 shows a high correlation between our personnel records and their estimates, with a slope of 0.78 in the full sample and a slope of 0.99 when we drop one outlier (South Africa). Half of the variation in our data is absorbed by variation in their estimates; most data points are closely aligned with the predicted values from a linear regression. As an additional data check, we reproduce Benmelech and Klor (2018)’s estimations of the country-level determinants of Daesh recruitment in Tables B1 and B2. Table B1 uses a dummy outcome indicating if any recruit is coming from a given country, and Table B2 uses the log of one plus the number of recruits, as in Benmelech and Klor (2018). In both tables, we use our personnel records to construct the outcome variable in columns 1-4 and the expert esti- mates from Benmelech and Klor (2018) in columns 5-8. We find that the predictors for Daesh recruitment are similar in both data sets; these comparisons fail to reveal a bias in our data one way or the other. In contrast to previous studies on terrorism (see e.g. Abadie 2006 and Benmelech and Klor 2018) or on civil conflicts generally speaking (see survey from Blattman and Miguel 2010), we have detailed and plausibly representative individual information on terrorist recruits, which allows us to draw inference from sub-national variation. Specifically, in the Daesh personnel records, individuals report having either no education or primary, high school or university level education. We can thus construct recruitment statistics by country of residence and level of education, distinguishing primary education and below, secondary, and tertiary. After removing observations without either country of residence 7 Their data were published in two reports by the Soufan Group, a strategic security intelligence think tank. They gather official and unofficial counts of the stock of foreign fighters from each country obtained from social media, community sources, or investigations, as of June 2014. 6 or education, we are left with a sample of 2,987 recruits originating from 59 countries.8 Daesh recruits the majority of its fighters from nearby Muslim countries. Table 1 orga- nizes the sample of Daesh recruits by country of last residence, ranking the countries by geographical distance. The first 10 countries in the list account for almost 45 percent of Daesh’s foreign recruits in our data set. Despite a few more distant large providers such as Tunisia, Morocco, or France, recruitment in a country declines with distance, both at the extensive and the intensive margins, after controlling for total and muslim popula- tions (Tables B1 and B2, columns (1) and (2)). This suggests prima facie that migration costs associated with distance may be an obstacle to recruitment by Daesh. Two-third of the recruits are in their twenties (Table 2). In addition, we find that 33.7 percent of the sample is married and 22.1 percent of the recruits have children. Our data also contain characteristics that reflect a recruit’s human capital and indicate that 51.7 percent of the recruits report having a secondary education and 30.6 percent report having a tertiary education. Figure 2 compares the fraction of primary, secondary and tertiary educated recruits in our sample with the proportions observed in the labor force of their country of last residence. In order to obtain stable proportions, we restrict the figure to countries repre- sented by at least ten recruits. A large majority of blue squares and green triangles are above the forty-five degree line, meaning that Daesh recruits are more likely to have a sec- ondary or tertiary education than the average worker in their country of last residence. Conversely, there are fewer recruits that have only a primary education or less, relative to the labor force in their country of last residence. These findings reinforce the conclusions ckov´ of Krueger and Maleˇ a (2003), and later Abadie (2006), Krueger (2007) and Krueger and Laitin (2008) who argued that terrorist recruits are not uneducated, and often come from middle-class backgrounds or have some college education. Another original feature of the data is that they contain information on self-reported knowledge of Sharia, which is available for almost 90 percent of our observations and is recorded as low, intermediate, or high. A large majority of recruits are too ignorant of Islam to be accurately described as religious fundamentalists; only about a third of 8 We do not include the 32 recruits from Iraq and 43 from Syria. 7 recruits report an intermediate or high level of knowledge (Table 2). This observation is consistent with the view held in the literature that religious terrorism is less driven by ideology than it is by kinship and social networks (see discussion in Gaibulloev and Sandler 2019). 2.2 Macroeconomic indicators We combine Daesh personnel information with country-level economic data, also disag- gregated by education levels. We use ILOSTAT data to construct education-level-specific unemployment rates for most countries, yielding 177 country*education-level observa- tions. We use data from 2013 to best match the personnel records on Daesh foreign re- cruits. If data from 2013 are missing, we use the nearest available year.9 To construct wage data, we use the International Income Distribution Data Set (I2D2) to compute median wage by education level for each country. The data set is a global har- monized household survey database compiling data from household surveys and labor force surveys (Montenegro and Hirn 2009). As for the unemployment variable, we take median wage data for the year 2013 and replace the missing values with the closest lead or lag during 2010-2016. Since we will be computing relative wages, we do not attempt to deflate or convert the nominal wage information. When we include the wage, unemploy- ment and education variables together, we are left with only 28 country*education-level observations from 12 countries. For robustness, we also use a second version of the wage variable, specific to the male population between 18 and 36 years.10 Augmenting the data with observations from 109 countries that do not supply Daesh 9 To maximize the number of observations, we use the total unemployment rate in our main results, but obtain qualitatively similar results when using the male unemployment rate or the youth unemployment rate. 10 One limitation is due to recent unemployment and wage rate information not being available for all countries. Table B9 in the Appendix shows the countries for which we have these data, and countries that supply Daesh recruits. Given the lack of sufficient overlap between the unemployment and wage variables, we henceforth proceed in two steps. First, we conduct our analyses using the unemployment variable only, hence omitting the wage variable. If wages and unemployment are uncorrelated, this approach is innocuous. We indeed find that the residuals of unemployment and wages, after partialling out country and education fixed effects, are uncorrelated, as illustrated in Appendix Figure B1. We nonetheless verify in section 4.3 that our results are robust to controlling for wages using the smaller sample of countries where we have both wages and unemployment data by education categories. 8 recruits leads to a final dataset that consists of 168 countries or 504 country*education ob- servations. Table 3 describes the country-level variables we use (total population, Muslim population, per capita GDP, Human Development Index, political freedom measures, cor- ruption index, religion variables and distance to Iraq and Syria) as well as the country-by- education-level variables (unemployment and wage rates). Detailed variable definitions and their sources are provided in Appendix Section A. 3 Empirical strategy Our empirical approach incorporates two main ingredients. First, we leverage our de- tailed individual data on Daesh recruits and propose an identification strategy that we believe is an improvement on the existing cross-country analyses of the economic drivers of terrorism. Second, we exploit variation in the distance travelled by Daesh fighters to join the terror group in Iraq or Syria to provide empirical support for an economic mech- anism specific to transnational terrorist recruitment. To control for unobserved country-level confounders that plagued the earlier litera- ture on the macroeconomic determinants of terrorism, we exploit the unique features of our data – namely the availability of the number of Daesh recruits and the unemployment rate for each country and education category (primary, tertiary and secondary education). This allows us to implement an identification strategy that leverages within-country vari- ation across education groups, hence isolating the causal impact of unemployment on transnational terrorism under weaker conditions than in the previous literature. Specifi- cally, we estimate Nce = α + µc + γe + β · U nempce + ξ · Xce + ce , (1) where the outcome is the number (or log number) of Daesh recruits from country c with education level e, µc and γe represent fixed effects for each country and the three education-level categories; β captures the conditional association of the unemployment 9 rate specific to a country-education cell with the number of Daesh recruits11 ; and ce is an error term. We control for the size of the labor force in the country-education cell, Xce . In additional robustness checks, we will also control for the average wage in each country-education cell. The inclusion of country fixed effects allows us to control for any country-level characteristics affecting individuals’ propensity to join Daesh, such as those related to distance to Iraq and Syria, state capacity, institutions and political representa- tion, as long as the effect of these country-level characteristics on Daesh participation is constant across the three education-level categories. The constant α meanwhile absorbs the mean returns to engaging in violence. We observe that the theoretical prediction about the impact of unemployment on par- ticipation in transnational terrorism is ambiguous. On the one hand, unemployment low- ers the economic opportunity cost of participation in terrorist activities and might also generate or exacerbate grievances against the government. Both predict a positive rela- tionship between unemployment and Daesh enrollment. For simplicity, we refer to this mechanism as the opportunity-cost channel. On the other hand, unemployment can be an obstacle to participation in a transnational terrorist organization, if joining the latter is economically costly and unemployment exacerbates liquidity constraints. The trip to join Daesh indeed constitutes a non-trivial cost (plane ticket, visa, potentially hotel and bus tickets), which most recruits fund out of pocket, with little to no financial support from the organization. The cost of joining the terrorist group is analogous to the cost of migra- tion considered in the labor and migration literature (Ozden, Wagner and Packard 2018), but has not previously been considered in the conflict literature. We henceforth refer to this mechanism through which unemployment may be negatively affect participation in transnational terrorism as the liquidity-constraint channel.12 The liquidity-constraint channel should be stronger for potential recruits from coun- 11 To the extent that psychological and political grievances co-vary with the unemployment rate across education categories, their effect would also be captured by β . 12 Previous studies have highlighted other mechanisms which may offset the positive effect of unemploy- ment on participation in terrorism. Most importantly, Berman, Shapiro and Felter (2011a) find that higher wages are associated with more rather than less violence in Iraq, which is consistent with a community- centric model of participation in violence, whereby higher wages make it harder for the government to financially incentivize communities to participate in counter-insurgency efforts. However, this channel does not apply to our context of transnational recruitment. 10 tries far away from Iraq and Syria, for whom the travel costs are highest. Thus, to dis- tinguish the liquidity-constraints channel from the opportunity-cost channel, we estimate the extended model Nce = α + µc + γe + β · U nempce + δ · U nempce · Distancec + ξ · Xce + ce , (2) where Distancec is the shortest distance in miles from country c to the nearest border point of Iraq or Syria. The liquidity-constraint mechanisms would suggest that the co- efficient δ on the interaction term between distance and unemployment is negative. The relative size of δ compared to β measures the importance of the attenuating effect of liq- uidity constraints to cover travel costs on the role of unemployment as a driver to joining Daesh. The liquidity-constraint channel will be weaker, potentially even absent, in countries at a low geographic distance to Daesh headquarters. Thus, we start our empirical analysis in section 4.1 with a specification of equation 1 restricted to countries that are “close” to Iraq and Syria. This approach minimizes the liquidity constraint channel, allowing us to estimate the effect of higher unemployment on terrorist supply which operates through a lower opportunity cost of joining Daesh and through increased grievances. In section 4.2, we then broaden our analysis to all countries with Daesh recruits, and estimate equation 2 to see how the effect of unemployment changes with distance, providing direct evidence on the liquidity-constraint channel. In section 4.3, we present a battery of robustness tests to show that the distance interaction indeed captures the strength of the liquidity constraints mechanism rather than other country characteristics correlated with distance. 4 Results 4.1 Unemployment and the Opportunity Cost of Joining Daesh To test the theoretical prediction of a positive correlation between unemployment and Daesh recruitment due to the opportunity-cost channel, we first shut down the liquidity- constraint channel by estimating equation 1 in the sample of countries within 500 miles 11 of the nearest border point of Iraq or Syria. This includes immediate neighbors in the Middle East, countries in the Gulf and North Africa, as well as some countries in Central Asia (see Table 1 for the list of countries ranked by distance to Syria). The regression results are displayed in Table 4 and indeed document the positive ef- fect of unemployment on Daesh enrollment in geographically close countries. The un- conditional correlation between unemployment and the (log) number of Daesh recruits is positive, with a point estimate of 0.061.13 In column 2, we add dummies for the three education categories and in column 3 we add country fixed effects, to absorb any country- level factors that do not vary across education groups. The inclusion of these fixed effects doubles the size of the point estimate and strengthens its significance. It suggests the country-level unobservables were biasing estimates downward. In column 4, we addi- tionally control for the size of the labor force so that the main coefficient can be inter- preted as a propensity of joining Daesh. This leads to a slight reduction in the sample size and to a further increase in the point estimate to 0.147. This semi-elasticity of recruitment with respect to the unemployment rate implies that a 1 percentage point reduction in the unemployment rate leads to a 15.8 percent reduction in Daesh enrollment. Dodwell et al. (2016) estimate that the total number of foreign recruits arriving during our sample pe- riod is about 15,000, and our data indicate that around 50 percent of that flow stems from the sample of close countries, as defined here. Thus, our result suggests that around 1200 fewer fighters would have joined Daesh from these countries over the period 2013-2014, if the unemployment rate had been 1 percentage point lower in these countries.14 To anticipate the coming analysis for the full sample, in column 5 of Table 4, we extend our definition of “close” countries by including countries at below median distance from Iraq and Syria. This increases the sample from 12 to 21 countries. The positive associa- tion between unemployment and Daesh recruitment is still present in this sample, but the point estimate is now half the size compared to column 4. This suggests that the effect of 13 Since the left-hand side of the equation is the logarithm of the number of Daesh recruits, it is only defined when such number is strictly positive. Cells that do not have at least one foreign recruit are dropped from the regression. However, in our sample of close countries, almost all of the 36 country-education cells register fighters, leaving us with a sample of 34 observations. We apply Moulton’s parametric correction to re-compute the standard errors in all regressions where cluster size is less than 40 (Moulton 1986). 14 The average unemployment rate in that set of countries is 9.6 percent. 12 unemployment on Daesh recruitment is weaker in more distant countries, a result con- sistent with a liquidity-constraint channel working in the opposite direction. We examine the spatial heterogeneity in the unemployment effect in more detail in the next section. 4.2 Spatial Heterogeneity in the Unemployment-Terrorism Relation- ship For countries close to Iraq and Syria, unemployment is found to increase enrollment in Daesh. For potential recruits from countries that are further away, however, the travel cost to Mashreq countries is higher, meaning that liquidity constraints may become binding for poorer or unemployed candidates. Theoretically therefore, the effect of unemploy- ment on Daesh enrollment should decrease as distance to Iraq and Syria increases; the re- lationship can potentially change sign if the effect of more stringent liquidity constraints dominates the effect of lower opportunity costs of participation. To test this hypothesis, we estimate the extended regression model in equation 2. In this model, the interaction term between unemployment and distance can be a continuous interaction or an interaction with country group dummies based on the distance median, terciles or quartiles across countries. We show results for all specifications, but note that the quartiles-specification is our preferred option, as it is most flexible, allowing the effect of unemployment to be non-linear in distance. Figure 3 graphically illustrates our main result. The different panels plot the residual- ized unemployment rate and log number of Daesh foreign fighters, after partialling out country and education-group fixed effects. Among countries in the first distance quar- tile, which is similar to our initial sample of countries at below 500 miles distance (minus Ukraine), the resulting slope is positive and significant as discussed earlier. In the fourth distance quartile group, the slope is now negative and significant, while it is insignifi- cant in the second and third quartile subsamples. Besides, as Figure 3 makes clear, the slopes we obtain are informed both by cross-country variation within a schooling level and cross-education-group variation within a country. Each one of three education-level- specific clouds of points (triangles, squares and circles) line up individually to create a 13 slope. Similarly, the within-country variation identifies a similar slope, as can be seen by looking at the alignment of the three points for specific countries such as Egypt and Saudi Arabia in Panel A. The regression results are presented in Table 5. In the first column, we use the contin- uous distance interaction, showing that the migration costs indeed attenuate the effect of unemployment on recruitment. In columns 2-4, we repeat this estimation, interacting un- employment with distance median-groups, terciles or quartiles respectively. The results are robust across specifications: the effect of unemployment on recruitment is positive in close countries, then decreases with distance, and becomes negative in distant countries where the liquidity-constraint mechanism dominates. The quartile interactions in column 4 confirm that the positive effect of unemployment is concentrated in the first quartile and the negative effect is concentrated in fourth distance quartile. In the second and third dis- tance quartile, the effect of the opportunity and grievance mechanism is exactly nullified by the liquidity constraints mechanism, so that the association between unemployment and recruitment becomes insignificant.15 Bootstrapped standard errors yield similar re- sults (Table B3). Thus, unemployment is a push factor for Daesh recruitment in countries close to Iraq and Syria, but becomes an impediment to recruitment in distant countries. While we have so far used a log-linear OLS estimation with the log of the number of Daesh recruits (from a given country with a given education level) as the outcome vari- able, Table B4 shows that the results are very similar when estimating a Pseudo Poisson Maximum Likelihood (PPML) model according to Santos Silva and Tenreyro (2006) with the number of Daesh recruits as outcome. This model has the advantage of utilizing all observations from countries with any recruits, whereas the log-linear model uses only country-education cells with any recruits. The PPML thus increases the sample from 105 to 132 observations. Finally, we show that we obtain our main result also within groups of fighters with the same desired occupation within Daesh — fighter, suicide fighter, or administrator. Conceptually, the outside option now includes staying in the home country or joining 15 The results from this regression are visualized in Figure B3, which plots the point estimates β with the 95 percent confidence interval. 14 Daesh in a different role. Columns 5-7 in Table 5 report the results of our main regres- sion specification applied separately to the contingents of fighters, suicide fighters and administrators. The point estimates and the levels of significance differ, but the patterns obtained for the whole sample largely carry through for each separate role. The main effect of unemployment is positive, the interaction with distance is negative, and both co- efficients are of the same order of magnitudes for all three roles and for the whole sample. For fighters, the effect of unemployment is relatively lower than for the other categories, while it is higher for suicide fighters. The point estimates for administrators are not sig- nificant (the number of observations is markedly lower, leading to large standard errors), but very similar to those obtained for the full sample. Our findings highlight the two opposing effects of unemployment on the international recruitment of jihadists. On the one hand, unemployment means lower foregone earnings upon joining Daesh. On the other hand, unemployed candidates in distant countries find it harder to mobilize the financial resources for long-distance travel to reach the terrorist organization. An alternative to international jihad is domestic terrorism, which might provide similar ideological benefits to radicalized individuals without requiring a migra- tion cost (Hegghammer 2013). Indeed, substitution across various types of terrorism is not uncommon, as Enders and Sandler (2004) show in their analysis of substitution be- tween attack types, countries and over time. We thus consider whether the availability of domestic terrorist opportunities could explain part of our results, i.e. explain the negative distance-unemployment interaction. If radicalized individuals in more distant countries substituted joining Daesh with do- mestic terrorism, the occurrence of local terrorist events should have increased more in distant countries relative to less distant countries, in the period in which Daesh was re- cruiting. The substitution effect should be particularly strong in countries with high rates 15 of unemployment. We test this by estimating the following triple-difference model: Ln(Tct ) = α + µc + ρt + β1 · U nempct + β2 · Distantc · P ostt + β3 · Distantc · U nempct + β4 · P ostt · U nempct + β5 · P ostt · U nempct · Distantc + ct , where Tct is the number of terrorist events per country and year from the Global Terror- ism Database, Distantc indicates countries in the fourth distance quartile (the remaining countries are in the second and third distance quartile, as the first quartile is affected by more direct spillovers from Daesh and hence dropped)16 , µc and ρt are country and year fixed effects, P ost indicates the years after Daesh emergence, and the unemployment rate is measured at the country-year level. We control for year and country fixed effects. As the outcome data is at the country level, we cannot run our main specification with education-group disaggregation. Table B5 displays the results. We find that there was indeed an increase in terrorist events in distant countries after Daesh emerged, and the likelihood of a terrorist incident is generally higher in distant countries with high levels of unemployment. However, the coefficient on the triple interaction is always insignificant, suggesting there is no evidence for substitution from Daesh to local terrorism. The results change little when we vary how the Daesh-time indicator P ost is measured as shown in the different columns, or when using a dummy indicating any terrorist event as outcome. In addition, we show in Table B6 that our main results from the model with country and education-group fixed effects are unchanged when controlling for additional interactions between unemployment, dis- tance and domestic terrorism. The coefficients on these additional interactions are not statistically significant. We thus fail to detect any evidence of a substitution between do- mestic and transnational terrorism. 16 These spillovers are also the reason we cannot test for a negative substitution effect on local terrorism in countries close to Iraq and Syria. 16 4.3 Robustness Tests This section presents a number of robustness tests. First, we show that our main results are not driven by one or two influential countries. To do so, we estimate our preferred specification (Table 5, column 4) forty-four (44) times, each time leaving out one country. Figure 4 displays the distribution of point estimates from this exercise. The distribution is clearly concentrated around the main effect we estimate in the full sample, and has short tails. Figure 5 shows results for a similar exercise, in which we drop two countries from our sample in each iteration. We then refute concerns related to the fact that our unemployment variable is not measured among Muslims only. Under the assumptions that Muslims constitute the pool of potential Daesh recruits, and that Muslims face different unemployment rates than non-Muslims, unemployment rates would be mis-measured in countries with large non-Muslim populations. Depending on the correlation between between Mulsim and non-Muslim unemployment rates, and how it varies with distance, the mis-measurement could lead to a falsely significant coefficient or the wrong sign. We provide three pieces of evidence against these concerns. First, figure B2 shows that the Muslim unemployment rate (as measured by Gallup survey data) is strongly correlated with the general unemployment rate.17 Given this positive correlation, the negative effect of unemployment in the fourth distance quartile is prima facie evidence against the measurement error hypothesis, as classical measurement error would bias the coefficient to zero. Third, and crucially, our results are not driven exclusively by Muslim-majority coun- tries, as we demonstrate in Table 6. Columns 4 and 5 in this table split the sample by whether Muslims constitute more or less than 50% of the population. As this leads to a slightly unequal split of the sample, we repeat the exercise in columns 6 and 7 by splitting the sample exactly at the median of the Muslim population share. In all subsamples, the coefficients on unemployment and the unemployment*distance interaction are remark- ably similar, and the standard errors suggest that we cannot reject the null hypothesis 17 Unfortunately, the Gallup measure cannot be used dis-aggregated at the education-category level. 17 that the coefficients in all specifications are identical.18 This robustness check addresses not only the concern about measurement error in the unemployment rate, but also the more general point that the supply function of Daesh recruits could be different between Muslim-majority and minority countries. Conceptually, the labor market opportunity cost of joining Daesh is composed not only of the probability of being unemployed, but also of the wage level available at home to potential recruits. Our main specification does not include wages as a regressor, be- cause schooling-specific wage data are available only for a small subset of the countries producing Daesh fighters. Therefore wages are part of the regression’s error term. If wages are correlated with unemployment (Blanchflower and Oswald 1994), the coeffi- cient on unemployment should be interpreted as the effects of labor market opportunities at home broadly construed, including both unemployment and wages. Note, however, that our specification includes country and education fixed effects. Therefore, the co- efficient on unemployment will be affected by the omission of wages only if these two variables are still correlated after partialling out country and education fixed effects. Fig- ure B1 shows this is not the case for the subset of 28 observations in 12 countries for which schooling-specific wage levels and unemployment rates are available and that register at least one Daesh recruit. Using that subset of observations, we further verify in Table 7 that our results are not driven by wages rather than unemployment. To maximize power in this smaller sample, we focus on the specification that includes a continuous interaction between unemploy- ment and distance. The results for that specification estimated on the full sample are reproduced for comparison purposes in column 1, Table 7. In column 2, we add the loga- rithm of the median wage in each country and education level as an additional regressor. The coefficient on the wage variable itself is not significant, and the impact of unemploy- ment on Daesh enrollment remains qualitatively and quantitatively similar. If the stan- dard errors in column 2 were comparable to those in column 1, the point estimate of the coefficient on unemployment would be statistically significant. This shows that the differ- 18 A similar result holds if we instead restrict to countries such that Muslims account for at least 1 percent of their entire population. There are 41 such countries in our sample. 18 ence in statistical significance between columns 1 and 2 are due to changes in sample size. Indeed, removing the wage regressor but keeping the restricted sample yields estimates comparable to column 2 (see column 3). In column 4, we use an alternative wage variable that takes the median value of wages for males aged 18-36, which is the appropriate com- parison group for Daesh foreign recruits. Here again, the coefficients on unemployment and its interaction with distance remain consistent with our main specification in column 1. Next, we address the concern that our main specification sample is mechanically cen- sored at 0 recruits in a given country-education cell. First, note that a censoring rule based on the total number of fighters from a given country would not be problematic, since the expectation of the error term conditional on that rule would be absorbed in the fixed effects. Using this insight, we find the lowest country-level threshold such that all countries with a number of recruits equal to or above the threshold have recruits in all three education categories. This happens for countries with more than 33 fighters. The result, displayed in column 1 of Table 6, is similar to our main result despite the fact that this restriction lowers the number of countries under consideration to 12 and the total number of observations to 36. Furthermore, columns 2 and 3 show that results are robust to varying either the country- level cutoff or the country-education-level cutoff away from 0. Column 2 uses countries that have at least ten Daesh recruits. This increases the sample to 28 countries. In col- umn 3, we instead consider all countries that have at least one recruit in each of the three education levels being considered, even if they have less than 33 fighters overall. This selection leads to a regression based on 25 countries. Besides these results, the Pois- son regressions in Table B4 are also robust to censoring concerns, as the Poisson uses all country-education cells in countries with at least one fighter. Lastly, we show in Table B7, that our results are highly robust to different distance measures. Indeed, the coefficients on our regressors of interest are very stable, whether we measure distance from a country’s most populous city, or the capital city, or geo- graphic centre, and whether we consider distance to Iraq or to Syria.19 19 We prefer these geographic measures to alternative distance measures such as the cost of a flight ticket, 19 We now turn to concerns that geographical distance might stand in for other country characteristics, which are correlated with distance, and which mediate the effect of un- employment on Daesh recruitment. For example, geographically more distant countries, such as OECD countries, have stronger social welfare systems, so that unemployment does not necessarily generate social and economic exclusion to the point of driving Daesh participation. More distant countries are also less likely to be Muslim-majority countries, and hence less relevant or costlier as a pool for Daesh recruiters. Geographical distance might also capture some more general form of cultural distance, implying non-monetary costs that would not interact with unemployment through credit constraints. Finally, there are very few individuals with only primary education in OECD countries, such that the unemployment rate for this education category is measured more imprecisely and less relevant. Note first that these alternative stories can produce an attenuated or zero effect of unemployment in more distant countries, but not the negative effect that arises in the farthest quartile of countries. We can also specifically rule out those distance confounders that we can measure. In Tables 8 and 9, we conduct a horse race between distance and four alternative variables correlated with distance: GDP per capita, the fraction of Muslims in a country’s pop- ulation, and dummies for the MENA region and the OECD. That is, we interact these alternative variables with the unemployment rate, and test them individually or jointly against the interaction with distance. Only the OECD interaction and Muslim-fraction in- teraction are marginally significant when used individually (colum 3 in both tables), but loose significance once the distance interaction is added (columns 6 and 7). The physical distance interaction thus trumps all other interactions, and is the driving force for our as measuring the latter would require more choices to be made by the researcher, such as the time of the year at which to measure the cost, or how to average across seasonally changing prices. Besides, it is clear that flight costs are strongly correlated with distance. 20 main effect.20 5 Conclusion We used a unique data set on Daesh personnel records to shed light on the determinants of transnational terrorist recruitment. We document the impact of higher unemployment rates on enrollment in the terror group. Exploiting detailed information on foreign re- cruits’ countries of origin and education levels, we are able to establish this finding under weaker identification assumptions than those previously used in the literature. More specific to the question of transnational terrorism, we show that travel costs to Iraq and Syria, which exacerbate liquidity constraints of unemployed candidates, negatively affect enrollment. The tension between opportunity costs and liquidity constraints is novel to the literature on terrorism and applies not only to Daesh but to transnational terrorist recruitment more generally: limited labor market opportunities simultaneously have a substitution effect by lowering the opportunity costs of joining the terror group and an income effect, which exacerbates liquidity constraints for candidates who need to travel long distances to join. This gives rise to spatially heterogeneous effects of economic con- ditions on recruitment. This result is relevant beyond counter-terrorism policy — see e.g. Clemens and Postel (2018) on the relation between foreign aid and migration— and has implications for the design of interventions to limit transnational terrorist recruitment: policies that improve socio-economic outcomes have income and substitution effects that can go in opposite directions. 20 To conduct a more systematic analysis of potential regional differences in our main effect, we show in Table B8 regressions in which we interact unemployment with each region dummy individually, and a fully saturated model with all unemployment*region interactions. There is no region where unemployment has a significant effect, emphasizing again that the relevant driver of the interaction is physical distance rather than institutional characteristics of a country or region. Indeed, each region is spread across various of the distance quartiles. 21 22 6 Tables Table 1: Daesh Recruits by Country of Last Residence Country Region Fighters Fighters per Distance Per- capita Labor Muslim million to Syria GDP Force Proportion (#) Muslims (miles) (USD) (millions) (%) Mean All 58.3 13.1 2,081.4 21,083.9 37.9 51.7 St. Dev. All 128.5 16.4 1,615.5 26,021.3 121.6 43.1 Palestine MENA 21 4.9 174.7 2,992.2 1.0 97.5 Lebanon MENA 14 5.5 190.7 8,389 1.9 59.7 Iraq MENA 32 1 289.8 6,816.6 8.5 98.9 Jordan MENA 56 8.8 332.9 4,656.2 1.9 93.8 Turkey MENA 209 2.8 354.9 10,800.4 27.8 98.6 Georgia Fmr Soviet 3 6.8 573.2 4,274.4 2.0 10.5 Azerbaijan Fmr Soviet 92 10.5 598.1 7,811.6 4.9 98.4 Kuwait MENA 34 12.9 625.4 48,463.2 1.9 86.4 Egypt MENA 203 2.5 735.5 3,264.4 29 94.7 Saudi Arabia MENA 731 28.7 838.9 24,646 11.8 97.1 Iran MENA 13 .2 861.2 6,631.3 26.6 99.7 Bulgaria Europe 1 1.7 910.2 7,656.6 3.3 78 Bahrain MENA 24 27.7 915.5 24,378.9 0.7 70.2 Qatar MENA 9 7.7 977.9 96,077 1.6 77.5 Ukraine Fmr Soviet 3 7.6 1,021.5 3,986.3 23.1 .8 Macedonia Europe 16 32 1,046.6 5,219.5 0.9 33.3 Kosovo Europe 36 22.7 1,112.5 3,890.3 . 95.6 Albania Europe 9 4.8 1,113.7 4,412.3 1.3 58.79 Serbia Europe 1 4.4 1,149.6 6,353.8 3.1 2.8 Turkmenistan Fmr Soviet 5 1 1,170.6 7,480.3 2.3 93.3 Bosnia Europe 4 2.2 1,297.3 4,748 1.5 50.7 Libya MENA 123 19.4 1,418.6 10,454 2.3 96.6 Yemen, Rep. MENA 16 .7 1,456.8 1,408.1 7.3 99 Uzbekistan Fmr Soviet 42 1.6 1,459.1 1,878 13.3 96.5 Austria Europe 1 1.7 1,536.7 50,557.8 4.4 6.8 Poland Europe 1 50 1,538.3 13,776.5 18.3 .1 Sudan SSA 6 .2 1,614.1 1,726.1 12.1 97 Afghanistan Asia 1 0 1,634.8 653.3 8.0 99.8 Tunisia MENA 609 54.4 1,677.6 4,248.9 4.0 99.8 Kazakhstan Fmr Soviet 21 2.4 1,698.6 14,310 9.2 70.2 Note: This table is based on the Daesh personnel records, and lists the number of Daesh recruits by country of last residence, with country characteristics. The data sources are described in Appendix A. 23 Country Region Fighters Fighters per Distance Per- capita Labor Muslim million to Syria GDP Force Proportion (%) Muslims (miles) (USD) (millions) Pakistan Asia 21 .1 1,788 1,275.7 63.6 96.4 Switzerland Europe 2 5 1,796.7 84,669.3 4.7 5 Tajikistan Fmr Soviet 55 7.9 1,799.5 1,048.7 3.6 99 Germany Europe 84 52.5 1,815.5 45,600.8 42.8 2 Sweden Europe 12 26.7 1,975.2 60,283.2 5.1 5 Kyrgyzstan Fmr Soviet 38 7.7 1,984.5 1,282.4 2.7 88.8 Denmark Europe 17 73.9 2,030.8 60,361.7 2.9 4.1 Belgium Europe 26 39.5 2,037.8 46,622.5 5.0 5.9 Netherlands Europe 22 26.7 2,044.4 51,425.1 9.0 5 France Europe 148 29.5 2,066.7 42,571.2 30.1 7.5 Somalia SSA 1 .1 2102 521.2 3 98.9 Norway Europe 4 24.5 2,161.5 10,2910.4 2.7 3.0 Algeria MENA 26 .6 2,239.8 5,491.6 12.1 98.2 Spain Europe 12 6.4 2,350.8 29,370.7 23.4 4.1 Kenya SSA 3 1 2,409.9 1,261.1 17.0 10.0 Britain Europe 63 20.3 2456.3 42,294.9 32.8 4.8 Cameroon SSA 2 .4 2,543.1 1,331.2 8.9 20.9 Ireland Europe 1 14.3 2,612 51,814.9 2.2 1.1 India Asia 6 0 2616.5 1,456.2 487.9 14.2 Morocco MENA 275 8.5 2,649.6 3,153.8 12.3 99.0 Mauritania SSA 1 .2 3,163.5 1,457.8 1.2 100.0 Russia Fmr Soviet 171 18.2 3,374.3 15,543.7 76.9 6.5 China Asia 50 2.3 3,607.7 6,991.9 801.8 1.8 Malaysia Asia 1 .1 4,533.9 10,973.7 13 61.4 South Africa SSA 3 4.6 4,640.6 6,881.8 19.4 1.5 Indonesia Asia 73 .4 5,404.3 3,631.7 122.1 87.2 Canada Americas 18 17.1 5,838.5 52,266.2 19.5 1.9 Trinidad&Tobago Americas 3 38.5 6,373.1 20,217 .7 5.8 United States Americas 11 4.2 6,688.5 52,660.3 159.8 0.8 Australia Asia 13 27.3 7,455.9 67,652.7 12.2 2.2 Note: This table is based on the Daesh personnel records, and lists the number of Daesh recruits by country of last residence, with country characteristics. The data sources are described in Appendix A. 24 Table 2: Summary Statistics of Fighter Characteristics Fighter Characteristics Mean Std. Error N (%) (%) (#) Age < = 20 years 13.8 0.6 3,344 21 -30 years 67.6 0.8 3,344 31+ years 23.8 0.7 3,344 Education Primary 17.7 0.7 2,827 Secondary 51.7 0.9 2,827 Tertiary 30.6 0.9 2,827 Religiosity Level Low 68.7 0.9 2,634 Medium 26.2 0.9 2,634 High 5.1 0.4 2,634 Previous Occupation No Job, Student, Retired or Illegal 27.2 0.8 3,178 Craftsperson, Manual/Ag work, Security 11.9 0.6 3,178 Shop owner, Employee 31.1 0.8 3,178 Manager, Prof. Worker 20.6 0.7 3,178 Jihad Experience 11.0 0.6 3,121 Desired Role Admininstrator 6.8 0.8 1,024 Fighter 54.2 1.6 1,024 Suicide Fighter 39.0 1.5 1,024 Note: This table displays summary statistics on Daesh foreign recruits from the Daesh personnel records used in this paper. 25 Table 3: Descriptive Statistics of Macroeconomic Variables Panel A: Country Level Variable Mean St. Dev Min Max N Distance to Syria 3,254 2,253 174 10,030 168 Per capita GDP (thousand) 14.6 20.8 0.26 113.73 164 Human Development Index 0.68 0.16 0.33 0.94 161 Total Muslim population (millions) 9.67 29.77 0.001 204.85 166 Total population (millions) 42.93 149 0.3 1357 165 Corruption Index 41.79 19.725 8 91 162 Index of political rights 3.543 2.124 1 7 162 Ethnic fractionalization 0.458 0.26 0 0.930 157 Linguistic fractionalization 0.403 0.288 0.002 0.923 154 Religious fractionalization 0.426 0.24 0.002 0.86 158 Average self-reported religiosity 0.743 0.244 0.142 0.998 162 Government Restrictions Index 3.352 2.199 0.2 9.1 164 Social Hostilities Index 2.659 2.494 0 9 164 Panel B: Country-Education Level Variable Mean St. Dev Min Max N Relative wage 0.67 0.31 0.05 1.78 154 Unemployment rate 9.70 7.86 0.10 45.40 359 Note: This table displays summary statistics of country-level and country-education level variables. The data sources are described in Appendix A. The relative wage is normalized to 1 for tertiary education. 26 Table 4: Determinants of Foreign Enrollment in Daesh - Close Countries (1) (2) (3) (4) (5) VARIABLES logNce logNce logNce logNce logNce Unemployment rate 0.061* 0.070*** 0.127*** 0.147*** 0.078* (0.037) (0.026) (0.028) (0.033) (0.040) Total Labor force (log) 0.330 0.041 (0.201) (0.092) Observations 34 34 34 31 51 Mean Nce 36.8 36.8 36.8 40.1 36.6 Number of countries 13 13 13 12 21 Education Dummies N Y Y Y Y Country FE N N Y Y Y Adj. R-squared 1.0e-04 5.4e-02 .86 .86 .87 Note: Linear regression model used. Dependent variable is log of number of foreign recruits to Daesh by country and education category. Columns 1-4 are for countries at less than 500 miles distance from Syria, column 5 is for countries at below median distance from Syria. Standard errors in parentheses, clustered at the country level and corrected for small number of clusters whenever number of clusters ≤ 40 using Moulton correction factor. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. 27 Table 5: Determinants of Foreign Enrollment in Daesh - Distance Interaction and Different Daesh Aspiration (1) (2) (3) (4) (5) (6) (7) logNce logNce logNce logNce logN Fce logN Sce logN Ace VARIABLES Total Total Total Total Unemployment rate 0.668*** 0.069** 0.027 0.084 (0.140) (0.031) (0.058) (0.069) Total Labor force (log) -0.000 0.027 0.030 -0.063 0.515*** 0.479 0.321 (0.082) (0.087) (0.089) (0.075) (0.184) (0.328) (0.686) Interaction between unemployment and Distance to Syria (log) -0.091*** (0.020) Distance to Syria - First Half 0.068* (0.034) Distance to Syria - Second Half -0.050 (0.036) Distance to Syria - First Tercile 0.124*** (0.026) Distance to Syria - Second Tercile -0.014 (0.028) Distance to Syria - Third Tercile -0.082* (0.047) Distance to Syria - First Quartile 0.113*** (0.030) Distance to Syria - Second Quartile 0.009 (0.029) Distance to Syria - Third Quartile -0.008 (0.026) Distance to Syria - Fourth Quartile -0.160*** (0.037) Observations 105 105 105 105 62 45 22 Mean Nce 25.4 25.4 25.4 25.4 x x x Mean N Fce x x x x 7.9 x x Mean N Sce x x x x x 7.5 x Mean N Ace x x x x x x 2.8 Number of countries 44 44 44 44 32 24 13 Country FE Y Y Y Y Y Y Y Education Dummies Y Y Y Y Y Y Y Adj. R-squared .83 .81 .84 .85 .76 .45 .3 Note: Linear regression model used. Dependent variable is log of number of foreign recruits to Daesh by country and education category. Column 5, 6 and 7 include only those that aspire to become fighters, suicide fighters and administrators respectively. Standard errors in parentheses, clustered at the country level and corrected for small number of clusters whenever number of clusters ≤ 40 using Moulton correction factor. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. 28 Table 6: Determinants of Foreign Enrollment in Daesh - Robustness Across Sub-Samples (1) (2) (3) (4) (5) (6) (7) logNce logNce logNce logNce logNce logNce logNce VARIABLES Nc >= 33 Nc >= 10 Nc >= 0 Main effects Unemployment rate 1.012** 0.587** 0.639*** 0.620** 0.668 0.584 0.593** (0.416) (0.221) (0.214) (0.263) (0.432) (0.400) (0.261) Total Labor force (log) 0.071 0.075 0.012 -0.048 -0.082 -0.022 0.058 (0.222) (0.156) (0.108) (0.182) (0.161) (0.155) (0.192) Interaction between unemployment and Distance to Syria (log) -0.141** -0.080** -0.088*** -0.082** -0.087 -0.081 -0.074* (0.057) (0.030) (0.029) (0.038) (0.056) (0.052) (0.038) Observations 36 76 75 55 50 53 52 Mean Nce 65.7 34.4 33.6 39.8 9.6 9.1 42 Number of countries 12 28 25 21 23 24 20 29 Country FE Y Y Y Y Y Y Y Education Dummies Y Y Y Y Y Y Y Adj. R-squared 0.732 0.793 0.838 0.841 0.744 0.746 0.833 Note: Linear regression model used. Dependent variable is log of number of foreign recruits to Daesh by country and education category. The threshold for Nc in column 1 is set such that countries with a number of recruits at or above this thresholds have at least one recruit in all three education categories. In column 2, we include all countries with at least ten recruits. In column 3, we include all countries that have at least one recruit in each education category. Standard errors in parentheses, clustered at the country level and corrected for small number of clusters whenever number of clusters ≤ 40 using Moulton correction factor. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. Table 7: Determinants of Foreign Enrollment in Daesh - Robustness to Wage Controls (1) (2) (3) (4) VARIABLES logNce logNce logNce logNce Unemployment rate 0.668*** 0.443 0.371 0.436 (0.140) (0.415) (0.401) (0.390) Total Labor force (log) -0.000 -0.042 -0.065 -0.051 (0.082) (0.135) (0.131) (0.129) Median wage (log) -0.435 (0.517) Median wage among 18-36 old (log) -0.260 (0.283) Interaction between unemployment and Distance to Syria (log) -0.091*** -0.056 -0.048 -0.055 (0.020) (0.053) (0.051) (0.050) Observations 105 28 28 29 Mean Nce 25.4 6.5 6.5 6.4 Number of countries 44 12 12 12 Country FE Y Y Y Y Education Dummies Y Y Y Y Adj. R-squared .83 .62 .63 .63 Note: Linear regression model used. Dependent variable is log of number of foreign recruits to Daesh by country and education category. Standard errors in parentheses, clustered at the country level and corrected for small number of clusters whenever number of clusters ≤ 40 using Moulton correction factor. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. 30 Table 8: Determinants of Foreign Enrollment in Daesh - Robustness of Distance Interaction (1/2) (1) (2) (3) (4) (5) (6) (7) VARIABLES logNce logNce logNce logNce logNce logNce logNce Unemployment rate 0.668*** 0.324 -0.057 0.745*** 0.558*** 0.002 0.622* (0.140) (0.226) (0.043) (0.193) (0.198) (0.330) (0.312) Total Labor force (log) -0.000 0.069 0.080 0.001 0.009 0.078 0.007 (0.082) (0.108) (0.107) (0.082) (0.083) (0.108) (0.082) Interaction between unemployment and Distance to Syria (log) -0.091*** -0.083*** -0.079*** -0.080*** (0.020) (0.024) (0.024) (0.024) Per capita GDP (log) -0.034 -0.014 -0.006 -0.006 (0.024) (0.025) (0.032) (0.031) Muslim fraction 0.131* 0.053 0.117 0.038 (0.067) (0.074) (0.087) (0.083) Observations 105 105 105 105 105 105 105 Mean Nce 25.5 25.5 25.5 25.5 25.5 25.5 25.5 Number of countries 44 44 44 44 44 44 44 Country FE Y Y Y Y Y Y Y Education Dummies Y Y Y Y Y Y Y 31 Adj. R-squared .83 .81 .81 .83 .83 .81 .83 Note: Linear regression model used. Dependent variable is log of number of foreign recruits to Daesh by country and education category. Standard errors in parentheses, clustered at the country level and corrected for small number of clusters whenever number of clusters ≤ 40 using Moulton correction factor. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. Table 9: Determinants of Foreign Enrollment in Daesh - Robustness of Distance Interaction (2/2) (1) (2) (3) (4) (5) (6) (7) VARIABLES logNce logNce logNce logNce logNce logNce logNce Unemployment rate 0.668*** -0.029 0.045 0.663*** 0.598*** 0.034 0.634*** (0.140) (0.033) (0.030) (0.168) (0.144) (0.045) (0.181) Total Labor force (log) -0.000 0.078 0.083 0.000 0.011 0.082 0.010 (0.082) (0.110) (0.112) (0.084) (0.086) (0.111) (0.089) Interaction between unemployment and Distance to Syria (log) -0.091*** -0.091*** -0.079*** -0.082*** (0.020) (0.022) (0.021) (0.023) MENA region dummy 0.081 0.003 0.022 -0.026 (0.065) (0.069) (0.072) (0.077) OECD region dummy -0.095* -0.047 -0.088 -0.055 (0.052) (0.053) (0.055) (0.055) Observations 105 105 105 105 105 105 105 Mean Nce 25.5 25.5 25.5 25.5 25.5 25.5 25.5 Number of countries 44 44 44 44 44 44 44 Country FE Y Y Y Y Y Y Y Education Dummies Y Y Y Y Y Y Y 32 Adj. R-squared .83 .8 .81 .83 .83 .81 .83 Note: Linear regression model used. Dependent variable is log of number of foreign recruits to Daesh by country and education category. Standard errors in parentheses, clustered at the country level and corrected for small number of clusters whenever number of clusters ≤ 40 using Moulton correction factor. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. 7 Figures Figure 1: Comparison Between Daesh Personnel Records and Expert Estimates A: Full Sample B: Dropping One Outlier SAU 8 TUN Slope: .78 (.12), R2=.54 6 MAR EGY TUR FRARUS SAU TUN Daesh Personnel Records (logs) Daesh Personnel Records (logs) AZE DEU Slope: .99 (.14), R2=.61 6 IDN GBR TJK JOR MAR 4 CHN EGY TUR RUS DZA BEL FRA PAK NLD KAZ CAN DNK MKD AZE DEU AUS LBN IDN USA SWE ESP GBR TJK JOR 4 ALB CHN 2 IND SDN NOR BIH DZA BEL PAK NLD KAZ ZAF CAN DNK MKD CHE AUS LBN ESP USA SWE ALB 2 IRL MYS AUT 0 IND SDN NOR BIH CHE -2 IRL MYS AUT 0 0 2 4 6 8 2 4 6 8 10 Expert Estimates (logs) Expert Estimates (logs) Note: This figure plots the (log) number of Daesh recruits from expert estimates (used in Benmelech and Klor (2018)) against the numbers from our Daesh personnel records. We consider all countries with recruits in panel A and all countries minus South Africa (SAF, an outlier) in panel B. 33 Figure 2: Schooling Attainment Among Daesh Recruits Relative to their Country of Last Residence Note: This figure plots, for each country and education category, the share of individuals that obtained the relevant education level, in the country’s general labor force and among the recruits appearing in our Daesh personnel records. To obtain stable shares, we focus on countries with more than 10 Daesh recruits. 34 Figure 3: Relative Supply of Daesh recruits and Relative Unemployment Rate (a) Countries in Distance Quartile 1 (b) Countries in Distance Quartile 2 1 1 TUR WBG IRN Relative supply of Deash recruits (log) Relative supply of Deash recruits (log) EGY slope=.103 (.032) MKD LBN KAZ .5 .5 BHR SAU SAU slope=-.001 BIH LBN (.029) ALB ALB UKR AZE UKR TUN JOR KWT GEO EGY BIH 0 0 BGR WBG AUT POL SRB JOR AZE JOR GEO TUN IRN TUN KWT AZE TUR KAZ KAZ -.5 -.5 BHR BIH UKR MKD ALB TUR IRN EGY SAU WBG LBN -1 -1 -10 -5 0 5 -10 -5 0 5 Relative unemployment rate Relative unemployment rate Primary Secondary Tertiary Fitted values Primary Secondary Tertiary Fitted values (c) Countries in Distance Quartile 3 (d) Countries in Distance Quartile 4 1 GBR MAR 1 Relative supply of Deash recruits (log) Relative supply of Deash recruits (log) NLD slope=-.125 (.034) IDN slope=0 FRA KGZ USA AUS .5 .5 (.02) SWE DEU BEL DZA CAN GBR DNK CHEFRA RUS IDN DZA PAK 0 IRL IND MYS TTO BEL ESP RUS NOR 0 NOR NLD ESP RUS ESP DEU USA CAN PAK USA -.5 SWE SWE AUS MAR DEU MAR CHE DNK BEL -.5 IDN DZA KGZ -1 NLD FRA -1.5 -1 GBR -10 -5 0 5 -10 -5 0 5 Relative unemployment rate Relative unemployment rate Primary Secondary Tertiary Fitted values Primary Secondary Tertiary Fitted values Note: This figure displays scatterplots of the residuals from a regression of unemployment (log number of Daesh foreign recruits) on country and education-category fixed effects and total labor force. The countries are divided into four quartile samples according to their distance from Syria. Each panel pertains to a different quartile. 35 Figure 4: Distribution of Main Effect Estimates (1/2) (a) Distance Quartile 1 (b) Distance Quartile 3 Value Value 40 40 30 30 Frequency Frequency 20 20 10 10 0 .105 .11 .115 .12 .125 .13 0 -.06 -.04 -.02 0 .02 Unemployment * Distance - First Quartile Unemployment * Distance - Third Quartile (c) Distance Quartile 2 (d) Distance Quartile 4 Value Value 40 40 30 30 Frequency Frequency 20 20 10 10 0 0 -.01 0 .01 .02 .03 -.16 -.14 -.12 -.1 -.08 -.06 Unemployment * Distance - Second Quartile Unemployment * Distance - Fourth Quartile Note: These figures plot the distribution of point estimates βi on the unemployment*distance-quartile interaction term, from the regression ln Nce = α + µc + γe + i βi ln Uce .quartilei + ln LFce + ce , where we re-estimate the model 44 times, leaving one country out at a time. 36 Figure 5: Distribution of Main Effect Estimates (2/2) (a) Distance Quartile 1 (b) Distance Quartile 3 Value Value 600 600 500 500 400 400 Frequency Frequency 300 300 200 200 100 100 0 0 .08 .1 .12 .14 -.08 -.06 -.04 -.02 0 .02 Unemployment * Distance - First Quartile Unemployment * Distance - Third Quartile (c) Distance Quartile 2 (d) Distance Quartile 4 Value Value 600 600 500 500 400 400 Frequency Frequency 300 300 200 200 100 100 0 0 -.05 0 .05 .1 .15 .2 -.2 -.15 -.1 -.05 0 Unemployment * Distance - Second Quartile Unemployment * Distance - Fourth Quartile Note: This figure is identical to Figure 4, except that we leave out two countries in each iteration. 37 References Abadie, Alberto, “Poverty, Political Freedom, and the Roots of Terrorism,” American Eco- nomic Review, May 2006, 96 (2), 50–56. Bandyopadhyay, Subhayu and Javed Younas, “Poverty, political freedom, and the roots of terrorism in developing countries: An empirical assessment,” Economics Letters, 2011, 112 (2), 171–175. Benmelech, Efraim and Esteban F. Klor, “What Explains the Flow of Foreign Fighters to ISIS?,” Terrorism and Political Violence, 2018, 0 (0), 1–24. Berman, Eli and David D. Laitin, “Religion, terrorism and public goods: Testing the club model,” Journal of Public Economics, 2008, 92 (10), 1942 – 1967. , Jacob N. Shapiro, and Joseph H. Felter, “Can Hearts and Minds Be Bought? The Economics of Counterinsurgency in Iraq,” Journal of Political Economy, 2011, 119 (4), 766–819. , Michael Callen, Joseph H Felter, and Jacob N Shapiro, “Do Working Men Rebel? Insurgency and Unemployment in Afghanistan, Iraq, and the Philippines,” Journal of Conflict Resolution, 2011, 55 (4), 496–528. Blanchflower, David G. and Andrew J. Oswald, The Wage Curve, MIT press, 1994. Blattman, Christopher and Edward Miguel, “Civil War,” Journal of Economic Literature, 2010, 48 (1), 3–57. and Jeannie Annan, “The Consequences of Child Soldiering,” The Review of Eco- nomics and Statistics, 2010, 92 (4), 882–898. Clemens, Michael, “Does development reduce migration?,” in “International Handbook on Migration and Economic Development,” Edward Elgar Publishing, 2014, chap- ter 6, pp. 152–185. 38 Clemens, Michael A. and Hannah M. Postel, “Deterring Emigration with Foreign Aid: An Overview of Evidence from Low-Income Countries,” Population and Development Review, 2018, 44 (4), 667–693. Collier, Paul and Anke Hoeffler, “On Economic Causes of Civil War,” Oxford Economic Papers, 1998, 50 (4), 563. and , “Greed and Grievance in Civil War,” Oxford Economic Papers, 2004, 56 (4), 563–595. Dasgupta, Aditya, Kishore Gawande, and Devesh Kapur, “(When) Do Antipoverty Pro- grams Reduce Violence? India’s Rural Employment Guarantee and Maoist Conflict,” International Organization, 2017, 71 (3), 605–632. Davis, Jonathan M.V. and Sara B. Heller, “Rethinking the Benefits of Youth Employ- ment Programs: The Heterogeneous Effects of Summer Jobs,” Review of Economics and Statistics, 2019, Forthcoming. Dodwell, Brian, Daniel Milton, and Don Rassler, “The Caliphates Global Workforce: An Inside Look at the Islamic States Foreign Fighter Paper Trail,” Technical Report, United States Military Academy Combating Terrorism Center West Point United States 2016. Enders, Walter and Gary A Hoover, “The nonlinear relationship between terrorism and poverty,” American Economic Review, 2012, 102 (3), 267–72. and Todd Sandler, “What Do We Know about the Substitution Effect in Transna- tional Terrorism,” in Andrew Silke, ed., Researching Terrorism: Trends, Achievements and Failures, Abingdon: Routledge, 2004. , Gary A Hoover, and Todd Sandler, “The changing nonlinear relationship between income and terrorism,” Journal of Conflict Resolution, 2016, 60 (2), 195–225. Fearon, James D. and David D. Laitin, “Ethnicity, Insurgency, and Civil War,” American Political Science Review, 2003, 97 (01), 75–90. 39 Fetzer, Thiemo, “Can Workfare Program Moderate Violence? Evidence from India,” Jour- nal of the European Economic Association, 2019, Forthcoming. Gaibulloev, Khusrav and Todd Sandler, “What We Have Learned about Terrorism since 9/11,” Journal of Economic Literature, June 2019, 57 (2), 275–328. Gassebner, Martin and Simon Luechinger, “Lock, stock, and barrel: A comprehensive assessment of the determinants of terror,” Public Choice, 2011, 149 (3-4), 235. Gates, Scott and Sukanya Podder, “Social Media, Recruitment, Allegiance and the Is- lamic State,” Perspectives on Terrorism, 2015, 9 (4). Guardado, Jenny and Steven Pennings, “The Seasonality of Conflict,” April 2019. Hegghammer, Thomas, “The Rise of Muslim Foreign Fighters: Islam and the Globaliza- tion of Jihad,” International Security, 2010, 35 (3), 53–94. , “Should I Stay or Should I Go? Explaining Variation in Western Jihadists’ Choice between Domestic and Foreign Fighting,” American Political Science Review, 2013, 107 (1), 1–15. Holman, Timothy, “‘Gonna Get Myself Connected’: The Role of Facilitation in Foreign Fighter Mobilizations,” Perspectives on Terrorism, 2016, 10 (2). Krueger, Alan B., “What Makes a Terrorist: Economics and the Roots of Terrorism (New Edition),” Princeton University Press, 2007. and David D. Laitin, “Kto Kogo?: A Cross-Country Study of the Origins and Targets of Terrorism,” Terrorism, Economic Development, and Political Openness, 2008, pp. 148– 173. and Jitka Maleˇ a, “Education, Poverty and Terrorism: Is There a Causal Con- ckov´ nection?,” Journal of Economic Perspectives, Fall 2003, 17 (4), 119–144. and , “Attitudes and Action: Public Opinion and the Occurrence of Interna- tional Terrorism,” Science, 2009, 325 (5947), 1534–1536. 40 Lai, Brian, ““Draining the Swamp”: An Empirical Examination of the Production of In- ternational Terrorism, 1968—1998,” Conflict Management and Peace Science, 2007, 24 (4), 297–310. Li, Quan and Drew Schaub, “Economic globalization and transnational terrorism: A pooled time-series analysis,” Journal of conflict resolution, 2004, 48 (2), 230–258. Montenegro, Claudio E and Maximilian L Hirn, “A new disaggregated set of labor mar- ket indicators using standardized household surveys from around the world,” 2009. Moulton, Brent R., “Random Group Effects and the Precision of Regression Estimates,” Journal of Econometrics, 1986, 32 (3), 385–397. Ozden, Caglar, Mathis Christoph Wagner, and Michael Minh Tam Packard, “Moving for Prosperity: Global Migration and Labor Markets, Policy Research Report,” Tech- nical Report, World bank 2018. Sanchez de la Sierra, Raul, “On the Origins of the State: Stationary Bandits and Taxation in Eastern Congo,” Journal of Political Economy, 2019, Forthcoming. Santiford-Jordan, Charlinda and Todd Sandler, “An Empirical Study of Suicide Terror- ism: A Global Analysis,” Southern Economic Journal, 2014, 80 (4), 981–1001. Santos Silva, J. M. C. and Silvana Tenreyro, “The Log of Gravity,” The Review of Eco- nomics and Statistics, November 2006, 88 (4), 641–658. The Atlantic, “Where Do ISIS Fighters Go When the Caliphate Falls?,” 2017. United Nations, “Enhancing the Understanding of the Foreign Terrorist Fighters Phe- nomenon in Syria,” 2017. Verwimp, Philip, “Foreign Fighters in Syria and Iraq and the Socio-Economic Environ- ment They Faced at Home: A Comparison of European Countries,” Perspectives on Terrorism, 2016, 10 (6). ¨ , “The microeconomics of violent conflict,” Jour- , Patricia Justino, and Tilman Bruck nal of Development Economics, 2018. 41 Weggemans, Daan, Edwin Bakker, and Peter Grol, “Who Are They and Why Do They Go? The Radicalization and Preparatory Processes of Dutch Jihadist Foreign Fight- ers,” Perspectives on Terrorism, 2014, 8 (4). Windsor, Leah, “The Language of Radicalization: Female Internet Recruitment to Partic- ipation in ISIS Activities,” Terrorism and Political Violence, 2018, 0 (0), 1–33. 42 A Variable definitions Variable name Description Source Country-Education level Variables LogNce Log of number of Daesh recruits from country c by Daesh personnel education categories: No education/Primary, Secondary records and Tertiary level. Authors calculation. LogNFce Log of number of Daesh recruits who aspire to be fighters Daesh personnel from country c by education categories: No records education/Primary, Secondary and Tertiary level. Authors calculation. LogNSce Log of number of Daesh recruits who aspire to be suicide Daesh personnel fighters from country c by education categories: No records education/Primary, Secondary and Tertiary level. Authors calculation. LogNAce Log of number of Daesh recruits who aspire to be Daesh personnel administrators from country c by education categories: No records education/Primary, Secondary and Tertiary level. Authors calculation. Unemployment Number of unemployed persons as a percentage of the ILOSTAT rate total number of persons in the labor force by education categories: No education/Primary, Secondary and Tertiary level. Missing values were replaced from World Bank data. Total Labor force Log of sum of the number of persons employed and the ILOSTAT (log) number of persons unemployed. Median wage Median wage for men of all age groups and men aged 18- International (log) 36 Income Distribution Data Set (I2D2) Country level Variables 1Nc >1 Dummy variable which is one when a country sends at Daesh personnel least one Daesh recruit and zero otherwise. records Distance to Syria Log of air (flying) distance between centroid of a country DistanceCalculator. (log) and centroid of Syria in miles. net Per capita GDP Log of Gross Domestic Product divided by midyear The World Bank (log) population. Data are in current U.S. dollars. Database Muslim Log of Muslim population in a country divided by Pew Research Population (log) (1+1000000). Year: 2010. Center’s The future of global Muslim population, January 2011 Total Population Total population is based on the de facto definition of The World Bank (log) population, which counts all residents regardless of legal Database status or citizenship. The values are midyear estimates and are logged. 43 Human The index is a summary measure of average achievement The World Bank in key dimensions of human development: a long and Database Development healthy life, being knowledgable and have a decent Index standard of living. The HDI is the geometric mean of normalized indices for each of the three dimensions. Index of political Political rights enable people to participate freely in the Freedom House rights political process, including the right to vote freely for distinct alternatives in legitimate elections, compete for public office, join political parties and organizations, and elect representatives who have a decisive impact on public policies and are accountable to the electorate. The specific list of rights considered varies over the years. Countries are graded between 1 (most free) and 7 (least free). Corruption Index The corruption perception index focuses on corruption in Transparency the public sector and defines corruption as the abuse of International public office for private gain. The CPI Score relates to perceptions of the degree of corruption as seen by business people, risk analysts and the general public and ranges between 100 (highly clean) and 0 (highly corrupt). Ethnic Reflects probability that two randomly selected people Alesina et al., 2003 fractionalization from a given country will not belong to the same ethnic group. The higher the number, the more fractionalized society. Linguistic Reflects probability that two randomly selected people Alesina et al., 2003 fractionalization from a given country will not belong to the same linguistic group. The higher the number, the more fractionalized society. Religious Reflects probability that two randomly selected people Alesina et al., 2003 fractionalization from a given country will not belong to the same religious group. The higher the number, the more fractionalized society. Average Proportion of people who agree that religion is an Gallup World Poll religiosity (self- important part of their daily life. reported) Government The Government Restrictions Index (GRI) measures - on Pew Research Restrictions a 10-point scale - government laws, policies and actions Center’s Global Index that restrict religious beliefs or practices. The GRI is Restrictions on comprised of 20 measures of restrictions, including Religion study efforts by governments to ban particular faiths, prohibit conversions, limit preaching or give preferential treatment to one or more religious groups. Social Hostilities The Social Hostilities Index (SHI) measures - on a 10- Pew Research Index point scale - acts of religious hostility by private Center’s Global individuals, organizations and social groups. This Restrictions on includes mob or sectarian violence, harassment over attire Religion study for religious reasons and other religion-related intimidation or abuse. The SHI includes 13 measures of social hostilities. 44 B Supplementary Tables and Figures 45 Table B1: Cross-Country Analysis of Foreign Enrollment in Daesh, Extensive Margin (1) (2) (3) (4) (5) (6) (7) (8) Personnel Records Expert Estimates VARIABLES 1Nc >0 1Nc >0 1Nc >0 1Nc >0 1Nc >0 1Nc >0 1Nc >0 1Nc >0 Total population (log) 0.036 0.022 0.013 0.011 0.082*** 0.056* 0.032 0.029 (0.029) (0.030) (0.031) (0.031) (0.028) (0.030) (0.032) (0.032) Muslim population (log) 0.156*** 0.169*** 0.169*** 0.167*** 0.092** 0.117*** 0.127*** 0.131*** (0.033) (0.040) (0.039) (0.039) (0.037) (0.042) (0.040) (0.041) Unemployment rate 0.013*** 0.011** 0.007 0.008 0.003 0.003 0.002 0.002 (0.005) (0.005) (0.006) (0.006) (0.006) (0.006) (0.006) (0.006) Distance to Syria (log) -0.149*** -0.144*** 0.035 0.039 -0.051 -0.052 0.083 0.080 (0.046) (0.052) (0.074) (0.074) (0.045) (0.054) (0.079) (0.079) Per capita GDP (log) 0.109*** 0.132*** 0.068** 0.127*** 0.108*** 0.013 (0.020) (0.028) (0.031) (0.023) (0.031) (0.040) Human Development Index 0.842** 0.293 46 (0.370) (0.473) Index of political rights 0.026 0.031* 0.033* -0.001 0.015 0.019 (0.017) (0.018) (0.019) (0.016) (0.017) (0.019) Ethnic fractionalization 0.206 0.329* 0.236 -0.350 -0.137 -0.117 (0.163) (0.184) (0.166) (0.235) (0.240) (0.269) Linguistic fractionalization -0.283* -0.283 -0.150 -0.028 -0.136 -0.144 (0.149) (0.191) (0.172) (0.225) (0.262) (0.294) Religious fractionalization 0.193 0.224 0.238 0.243* 0.296** 0.281** (0.141) (0.155) (0.155) (0.143) (0.129) (0.131) Observations 160 148 148 147 160 148 148 147 Adjusted R-squared 0.411 0.412 0.465 0.472 0.301 0.318 0.382 0.381 Mean Outcome .356 .358 .358 .354 .288 .304 .304 .306 Region FE N N Y Y N N Y Y Note: This Table presents linear estimation of Daesh enrollment (dummy) on country-level characteristics. Columns 1-4 and 5-8 respectively replicate columns 1-4 of Table 7 in Benmelech and Klor (2018). In columns 1-4, we use our Daesh personnel records to construct the outcome variable, in columns 5-8 we use the expert estimates from Benmelech and Klor (2018). ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. Table B2: Cross-Country Analysis of Foreign Enrollment in Daesh, Intensive Margin (1) (2) (3) (4) (5) (6) (7) (8) Personnel Records Expert Estimates VARIABLES Log(N+1) Log(N+1) Log(N+1) Log(N+1) Log(N+1) Log(N+1) Log(N+1) Log(N+1) Total population (log) 0.088 0.033 0.060 0.049 0.375*** 0.241* 0.186 0.173 (0.087) (0.084) (0.082) (0.082) (0.132) (0.132) (0.129) (0.129) Muslim population (log) 0.677*** 0.737*** 0.672*** 0.691*** 0.708*** 0.850*** 0.868*** 0.888*** (0.123) (0.141) (0.129) (0.133) (0.188) (0.212) (0.201) (0.207) Unemployment rate 0.029** 0.028* 0.017 0.016 0.033 0.040 0.033 0.032 (0.013) (0.015) (0.014) (0.014) (0.029) (0.031) (0.032) (0.033) Distance to Syria (log) -0.413*** -0.330** 0.371 0.361 -0.370 -0.368 0.237 0.226 (0.126) (0.144) (0.255) (0.255) (0.239) (0.276) (0.457) (0.460) Per capita GDP (log) 0.395*** 0.446*** 0.059 0.736*** 0.623*** 0.087 (0.064) (0.095) (0.097) (0.104) (0.148) (0.175) Human Development Index 1.203 1.695 47 (1.126) (1.993) Index of political rights 0.165*** 0.143*** 0.157*** 0.034 0.106 0.123 (0.063) (0.050) (0.053) (0.092) (0.092) (0.099) Ethnic fractionalization -0.006 -0.065 0.028 -2.280** -1.969* -1.913 (0.566) (0.503) (0.524) (1.081) (1.079) (1.175) Linguistic fractionalization -1.212*** -0.747 -0.797 -0.097 0.005 0.021 (0.425) (0.463) (0.521) (0.944) (1.048) (1.181) Religious fractionalization 0.490 0.702* 0.637 0.971 1.287* 1.220* (0.435) (0.394) (0.400) (0.740) (0.698) (0.732) Observations 160 148 148 147 160 148 148 147 Adjusted R-squared 0.456 0.497 0.593 0.594 0.379 0.414 0.466 0.465 Mean Outcome 1.009 1.033 1.033 1.036 1.436 1.524 1.524 1.534 Region FE N N Y Y N N Y Y Note: This Table presents linear estimation of the number of Daesh recruits (long(N+1)) on country level characteristics. Columns 1-4 and 5-8 respectively replicate columns 1-4 of Table 8 in Benmelech and Klor (2018). In columns 1-4, we use our Daesh personnel records to construct the outcome variable, in columns 5-8 we use the expert estimates from Benmelech and Klor (2018). ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. Table B3: Determinants of Foreign Enrollment in Daesh - Bootstrapped Std. Errors (1) logNce VARIABLES Total Total Labor force (log) -0.063 (0.108) Interaction between unemployment and Distance to Syria -First Quartile 0.113*** (0.035) Distance to Syria - Second Quartile 0.009 (0.082) Distance to Syria - Third Quartile -0.008 (0.033) Distance to Syria - Fourth Quartile -0.160*** (0.051) Observations 105 Number of countries 44 Country FE Y Education Dummies Y Adj. R-squared .85 Note: Linear regression model used. Dependent variable is log of number of foreign recruits to Daesh by country and education category. Standard errors in parentheses, are bootstrapped with 500 replications. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. 48 Table B4: Determinants of Foreign Enrollment in Daesh - Poisson Estimation (1) (2) (3) (4) VARIABLES logNce logNce logNce logNce Unemployment rate 1.105*** (0.361) Total Labor force (log) 0.207 0.140 0.082 0.004 (0.201) (0.143) (0.192) (0.188) Interaction between unemployment and Distance to Syria (log) -0.151*** (0.049) Distance to Syria - First Half 0.072 (0.049) Distance to Syria - Second Half -0.122*** (0.039) Distance to Syria - First Tercile 0.133*** (0.022) Distance to Syria - Second Tercile -0.019 (0.021) Distance to Syria - Third Tercile -0.159*** (0.055) Distance to Syria - First Quartile 0.146*** (0.023) Distance to Syria - Second Quartile -0.006 (0.022) Distance to Syria - Third Quartile -0.050 (0.041) Distance to Syria - Fourth Quartile -0.189*** (0.053) Observations 132 132 132 132 Mean Nce 20.2 20.2 20.2 20.2 Number of countries 44 44 44 44 Country FE Y Y Y Y Education Dummies Y Y Y Y Adj. R-squared .83 .82 .84 .85 Note: Poisson Pseudo Maximum Likelihood Estimator used. Dependent variable is the number of foreign recruits to Daesh by country and education category. Standard errors in parentheses, clustered at the country level and corrected for small number of clusters whenever number of clusters ≤ 40 using Moulton correction factor. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. 49 Table B5: DDD Estimation of Substitution Between Daesh and Domestic Terrorism (1) (2) (3) (4) (5) (6) Outcome: Log(N terrorist events+1) Outcome: 1(Any terrorist event) Post Daesh Definition Post 2011 Post 2012 Post 2013 Post 2011 Post 2012 Post 2013 Unemployment Rate (Fraction) -0.901 -1.109 -1.133 0.025 -0.120 -0.157 (2.092) (2.172) (2.177) (0.668) (0.682) (0.674) Distance * Post Daesh 0.717*** 0.860*** 1.062*** 0.108 0.197 0.278** (0.261) (0.318) (0.359) (0.098) (0.140) (0.125) Distance* Unemployment Rate 13.863*** 15.025*** 16.294*** 2.857* 3.177** 3.319** (4.476) (4.449) (4.405) (1.532) (1.569) (1.577) Unemployment Rate * Post Daesh 0.429 0.972 1.306 -0.214 0.108 0.264 (1.218) (1.294) (1.356) (0.318) (0.357) (0.376) Distance* Unemployment Rate * Post Daesh -2.784 -3.784 -4.874 -0.708 -1.521 -1.918 (2.625) (3.783) (3.707) (1.266) (1.913) (1.644) Observations 1,639 1,639 1,639 1,639 1,639 1,639 Number of countries 149 149 149 149 149 149 Country FE Y Y Y Y Y Y Year FE Y Y Y Y Y Y Note: This table display estimates of equation 4.2. The outcome is the log(N terrorist events +1) in columns 1-3, and a dummy for any terrorist event in columns 4-6, based on the Global Terrorism Database. The Distance dummy indicates countries in the fourth distance quartile. Countries in the first distance quartile are dropped from the analysis, as they may be affected by direct spillovers from Daesh. The P ost dummy indicates years after 2011, 2012 or 2013, as per the column headings. Standard errors, clustered at the country level, are in parentheses. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. 50 Table B6: Controlling for Domestic Terrorism in Main Estimation (1) (2) (3) (4) (5) (6) VARIABLES Log (Nce ) Log (Nce ) Log (Nce ) Log (Nce ) Log (Nce ) Log (Nce ) Unemployment rate 0.668*** 0.678*** 1.328** 0.479** 0.548*** 1.445** (0.140) (0.147) (0.646) (0.181) (0.180) (0.694) Total Labor force (log) -0.000 0.009 0.018 (0.082) (0.086) (0.091) Interaction between unemployment and Distance to Syria (log) -0.091*** -0.090*** -0.175* -0.068*** -0.071*** -0.190* (0.020) (0.021) (0.088) (0.025) (0.026) (0.095) Domestic Terrorism -0.032 -0.759 -0.061 -1.039 (0.052) (0.668) (0.049) (0.712) Domestic Terrorism * Log Distance 0.096 0.130 (0.090) (0.097) Observations 105 105 105 114 114 114 Mean Nce 25.4 25.4 25.4 23.9 23.9 23.9 Number of countries 44 44 44 47 47 47 Country FE Y Y Y Y Y Y Education Dummies Y Y Y Y Y Y Adj. R-squared .83 .83 .83 .81 .82 .82 Note: This table display estimates of our main estimating model, equation 2, with additional interaction terms between unemployment, distance and domestic terrorism. Domestic terrorism is a dummy variable that indicates if any terrorist event took place in the country in 2013. The data is from the Global Terrorism Database. The outcome is the log(N Daesh recruits). Standard errors, clustered at the country level, are in parentheses. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. 51 Table B7: Determinants of Foreign Enrollment in Daesh - Robustness to Different Distance Measures (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) VARIABLES logN logN logN logN logN logN logN logN logN logN logN logN logN Unemployment rate 0.668*** 0.658*** 0.762*** 0.739*** 0.759*** 0.788*** 0.887*** 0.858*** 0.886*** 0.688*** 0.734*** 0.740*** 0.728*** (0.140) (0.219) (0.232) (0.226) (0.234) (0.232) (0.225) (0.229) (0.227) (0.226) (0.224) (0.214) (0.222) Total Labor force (log) -0.000 0.005 0.016 0.043 0.021 -0.019 -0.005 0.026 0.000 -0.003 0.008 0.030 0.013 (0.082) (0.079) (0.085) (0.094) (0.087) (0.076) (0.081) (0.088) (0.083) (0.078) (0.084) (0.090) (0.085) Interaction between unemployment and Distance to Syria (log) -0.091*** -0.086*** -0.099*** -0.095*** -0.098*** -0.102*** -0.114*** -0.109*** -0.114*** -0.089*** -0.094*** -0.094*** -0.093*** (0.020) (0.028) (0.030) (0.029) (0.030) (0.030) (0.029) (0.029) (0.029) (0.029) (0.028) (0.027) (0.028) Observations 105 102 102 102 102 102 102 102 102 102 102 102 102 Mean Nce 25.5 26 26 26 26 26 26 26 26 26 26 26 26 Country FE Y Y Y Y Y Y Y Y Y Y Y Y Y Number of countries 44 43 43 43 43 43 43 43 43 43 43 43 43 Education Dummies Y Y Y Y Y Y Y Y Y Y Y Y Y Adj. R-squared 0.832 0.830 0.831 0.831 0.831 0.849 0.851 0.850 0.850 0.829 0.829 0.830 0.828 Note: Linear regression model used. Dependent variable is log of number of foreign recruits to Daesh by country and education category. The first column replicates our main result from Table 5, column 1. Columns 2-5 measure distance from a country’s most populous city, columns 6-9 measure it from the capital city, columns 10-13 measure it from the country’s geographic center. Columns 2, 6, 10 measure distance to Damascus; columns 3, 7, 11 measure distance to Raqqa; columns 4, 8, 12 measure distance to Mosul; columns 5, 9, 13 measure distance to Tell Abyad (the primary entry point to Daesh territory during the period covered by our data). 52 Table B8: Determinants of Foreign Enrollment in Daesh: Region Interactions (1) (2) (3) (4) (5) (6) logNce logNce logNce logNce logNce logNce VARIABLES Total Total Total Total Total Total Unemployment rate -0.029 0.032 -0.004 0.001 0.003 (0.033) (0.035) (0.025) (0.025) (0.025) Total Labor force (log) 0.111 0.078 0.127 0.128 0.083 0.060 (0.147) (0.110) (0.141) (0.148) (0.123) (0.125) Interaction between unemployment and MENA 0.052 0.081 (0.048) (0.065) Europe -0.032 -0.057 (0.039) (0.055) Former Soviet 0.061 0.094 (0.075) (0.076) Asia -0.018 -0.017 (0.109) (0.104) Americas -0.071 -0.069 (0.045) (0.043) Observations 105 105 105 105 105 105 Mean Nce 25.4 25.4 25.4 25.4 25.4 25.4 Country FE Y Y Y Y Y Y Number of countries 44 44 44 44 44 44 Education Dummies Y Y Y Y Y Y Adj. R-squared .8 .8 .8 .8 .79 .79 Note: Linear regression model used. Dependent variable is log of number of foreign recruits to Daesh by country and education category. Standard errors in parentheses, clustered at the country level and corrected for small number of clusters whenever number of clusters ≤ 40 using Moulton correction factor. ***, **, and * indicate statistical significance at the 1, 5, and 10 percent level, respectively. 53 Table B9: Wages, Unemployment and Daesh Recruits Data Overlap Wages Unemployment Daesh recruits Wages Unemployment Daesh recruits Wages Unemployment Daesh recruits AFG GMB NIC AGO GNB NLD ALB GNQ NOR ARE GRC NPL ARG GTM NZL ARM GUY OMN AUS HKG PAK AUT HND PAN AZE HRV PER BDI HTI PHL BEL HUN POL BEN IDN PRI BFA IND PRK BGD IRL PRT BGR IRN PRY BHR ISL QAT BIH ISR ROM BLR ITA RUS BLZ JAM RWA BOL JOR SAU BRA JPN SDN BTN KAZ SEN BWA KEN SGP CAF KGZ SLE CAN KHM SLV CHE KOR SOM CHL KSV SRB CHN KWT SSD CIV LAO SUR CMR LBN SVK COG LBR SVN COL LBY SWE COM LKA SWZ CRI LSO TCD CUB LTU TGO CYP LUX THA CZE LVA TJK DEU MAR TKM DJI MDA TTO DNK MDG TUN DOM MEX TUR DZA MKD TZA ECU MLI UGA EGY MLT UKR ERI MMR URY ESP MNE USA EST MNG UZB ETH MOZ VEN FIN MRT VNM FRA MUS WBG GAB MWI YEM GBR MYS ZAF GEO NAM ZAR GHA NER ZMB GIN NGA ZWE Note: This table reports for each country whether the wage and unemployment data by education category are available, and whether the country has at least one Daesh recruit (solid markers). 54 Figure B1: Wage and Unemployment Correlation slope=-.711 IDN IDN (.591) IDN 5 ALB ALB ALB KAZKAZ SRB SRB Wages (logs) UKR PAK UKR UKR SRB PAK PAK IND 0 IND IND USA USA USA KGZ KGZ LBN LBN LBN KGZ JOR JOR KSV KSV JOR KSV GEO -5 GEO -2 -1 0 1 2 Unemployment Rate (logs) Primary Secondary Tertiary Fitted values Note: This figures displays the scatter plot of log wages and log unemployment rates, after country and education-level fixed ef- fects are partialled out. The sample includes countries that have at least one Daesh recruit and available wage and unemployment information. 55 Figure B2: General Unemployment versus Muslim Unemployment 100 slope=1.107 Muslim Male Unemployment Rate (%) (.265) 20 40 0 60 80 correlation=.388 0 10 20 30 40 Male Unemployment Rate (%) Note: This figures displays the correlation between Muslim male unemployment and the general unemployment rate, in the Gallup survey data, for countries with a non-zero unemployment rate. 56 Figure B3: Marginal Effect of Unemployment on Daesh Recruitment by Quartiles .2 Marginal effect of Unemployment -.1 -.20 .1 1 2 3 4 Quartile (by Distance) Note: This figures displays the coefficients on the unemployment*distance-quartile interaction, and their 95% confidence intervals, from the estimation in Table 5, column 4. 57