Policy Research Working Paper 9885 Transnational Terrorism and the Internet Quy-Toan Do Nicolas Gomez-Parra Bob Rijkers Development Economics Development Research Group December 2021 Policy Research Working Paper 9885 Abstract Does the internet enable the recruitment of transnational organization’s recruitment of foreign fighters from Tunisia. terrorists? Using geo-referenced population census data The positive association between internet access and Daesh and personnel records from the Islamic State in Iraq and recruitment is robust to controlling for a large set of observ- the Levant—a highly tech-savvy terrorist organization— able and unobservable confounders as well as instrumenting this paper shows that internet access has facilitated the internet access rates with the incidence of lightning strikes. This paper is a product of the Development Research Group, Development Economics. 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://www.worldbank.org/prwp. The authors may be contacted at abonfield@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 Terrorism and the Internet∗ Quy-Toan Do Nicolas Gomez-Parra Bob Rijkers JEL codes : D74, H56, O15 Keywords : Internet, social media, transnational terrorism ∗ We are grateful to Hassen Arouri and Nadia Touihri for their help with identifying and processing relevant data and their constructive feedback. We thank Maleke Fourati, and Andrea Tesei for generously sharing data for this project; Rwan Mahmoud for excellent research assistance; Massimiliano Cali, Jishnu Das, Anne Grant, Marco Manacorda, C ¸ a˘ ¨ glar Ozden, ¨ Berk Ozler, Todd Sandler, Jake Shapiro, Roy van de Weide, and Aaron Zelin for helpful comments and suggestions. This paper has been partly supported by the Umbrella Facility for Trade Trust Fund (financed by the governments of the Netherlands, Norway, Sweden, Switzerland and the United Kingdom). We also acknowledge the financial support from the World Bank research support budget and the Knowledge for Change Program (KCP), a trust-funded partnership in support of research and data collection on poverty reduction and sustainable development housed in the office of the Chief Economist of the World Bank. The findings, interpretations, and conclusions expressed in this work do not necessarily reflect the views of the World Bank or the Inter-American Development Bank, their Boards of Executive Directors, or the governments they represent. The maps displayed in this paper are for reference only. The boundaries, colors, denominations and any other information shown on these maps do not imply, on the part of the World Bank Group and the Inter-American Development Bank, any judgment on the legal status of any territory, or any endorsement or acceptance of such boundaries. Do: World Bank (qdo@worldbank.org); Gomez-Parra: Inter-American Development Bank (ngomezparra@iadb.org); Rijkers: World Bank (brijkers@worldbank.org) 1 Introduction The internet has had a transformative impact on society, promoted trade, and generated growth (Freund and Weinhold, 2004; Litan and Rivlin, 2001). Yet, it has also enabled various forms uller and Schwarz, 2021), reduced confidence in government of crime (Bhuller et al., 2013; M¨ in most countries (Miner, 2015; Guriev et al., 2020), and long raised concerns about its use by violent extremists (Gartenstein-Ross, 2017; Conway, 2017). The Islamic State in Iraq and the Levant (ISIL, also known as Daesh) has been the epitome of tech-savvy terrorism and has attracted an estimated 30,000 foreign recruits from 89 countries to Syria and Iraq (Benmelech and Klor, 2018), about 6,000 of whom came from Tunisia (Berger and Morgan, 2015; Soufan Center, 2015; Benmelech and Klor, 2018), the largest country contingent in per capita terms. While Daesh’s reliance on social media – to spread propaganda, secure financing, plan attacks, proselytize, and recruit– is well-documented (FATF, 2015; Greenberg, 2016; Bloom et al., 2017; Conway et al., 2019; Mitts et al., forthcoming), the internet can also hamper its transnational terrorist recruitment efforts; it can enhance the efficiency of law enforcement or of repression by the state (Rød and Weidmann, 2015; Baum and Potter, 2019; Gohdes, 2020; Mastrobuoni, 2020), positively reshape public views of government (Gunitsky, 2015; King et al., 2017; Qin et al., 2017; Guriev and Treisman, 2019), and increase the opportunity cost of joining the terror group by improving labor market prospects (Autor, 2001; Hjort and Poulsen, 2019). The overall extent to which the internet has enabled the international recruitment of jihadists is, therefore, theoretically ambiguous. This paper addresses this question empirically by using a unique dataset on Tunisians who joined Daesh in Syria and Iraq between early September 2013 and late 2014. The data contain information on recruits’ ages, education levels, and delegations (district) of residence before traveling abroad, allowing us to exploit variation in enrollment both between and within ge- ographical areas, and between age and education groups.1 These granular data are matched with equally granular socio-economic information obtained from the 2014 Tunisia Census of Population and Housing. Thus, for each delegation-education-age cell, we are able to associate 1 Tunisia’s smallest administrative units are the communes, which form the country’s 264 delegations, which in turn aggregate into 24 governorates. 1 the number of Daesh recruits with both the internet access rate as well as other socio-economic variables that may promote radicalization such as unemployment and marital status. Our main finding is that internet access does predict jihadist recruitment. This result is robust to controlling for a large set of potential confounders, including age and education group dummy variables, and delegation fixed effects. While these fixed-effect estimations account for a large set of unobservables that have impeded the causal interpretation of earlier observational studies of ckov´ the drivers of terrorism (see discussions in Krueger and Maleˇ a, 2003; Abadie, 2006; Krueger and Laitin, 2008), we further provide two-stage least squares estimates using the incidence of lightning strikes as instrument for internet access rates (Manacorda and Tesei, 2020; Guriev et al., 2020). The IV estimates are significant and larger than their OLS analogs. We check the robustness of our findings to various alternative specifications, sample restrictions, and standard error corrections that account for spatial correlation. Our most conservative estimate implies that a one percent increase in internet coverage would result in an additional 2 recruits showing up in our sample or roughly 25 individuals overall. These numbers correspond to a .56 elasticity of recruitment with respect to internet access rates. This paper builds on and aims to contribute to two strands of literature that as yet have had little overlap. Our main finding relates to the emerging literature on the political implications and impacts on crime of internet expansion (see Zhuravskaya et al., 2020, for a review). By specifically looking at radicalization into violent extremism, our results also contribute to the literature on the drivers of terrorism in general (Li and Schaub, 2004; Gassebner and Luechinger, ckov´ 2011; Gaibulloev and Sandler, 2019; Krueger and Maleˇ a, 2003), and transnational terrorism in particular (Bandyopadhyay and Younas, 2011; Enders and Hoover, 2012; Benmelech and Klor, 2018; Brockmeyer et al., forthcoming; Hegghammer, 2013). This paper thus adds to the nascent literature on quantitative analyses of the relationship between the internet and violent extremism (Warren, 2015; Bail et al., 2018; Mitts, 2019; Mitts et al., forthcoming). The remainder of the paper is organized as follows. The next section presents our empirical strategy. The data are described in section 3. In section 4, we present and discuss our results; section 5 summarizes our results and draws conclusions. 2 2 Empirical strategy Our main empirical specification identifies the conditional correlation between radicalization into violent extremism and internet access at the delegation-education-age level (see section 3 for a discussion of how the data are constructed). The dependent variable is the inverse hyperbolic sine (ihs) transformation of Rdea , the number of recruits in district d who joined Daesh in Syria and Iraq that have education level e (primary or below, secondary, and tertiary and above) and belong to age group a (where individuals have been grouped in 5-year age categories).2 We estimate asinh(Rdea ) = αInternetdea + βP opulationdea + γXdea + εdea (1) where Internetdea is the share of men in delegation d with education level e in age group a who have access to the internet, P opulationdea is the (log of the) total number of men in a delegation-age-education cell {d, e, a}, and Xdea is a vector of controls. All these independent variables are measured in 2014, the year the data on Daesh recruits are believed to be from (Dodwell et al., 2016a). The coefficient of interest, α, measures the impact of internet access on the propensity to join the terrorist organization. Theoretically, the sign of α is ambiguous. On the one hand, the internet enables Daesh to spread propaganda and thus increase its “labor supply” (Bloom et al., 2017). It also reduces transaction costs associated with recruiting, similar to the effect internet has on licit recruit- ment efforts by regular firms (Autor, 2001). On the other hand, the internet facilitates law enforcement, repression of dissent (Greenberg, 2016; Conway et al., 2019; Gohdes, 2020) and can reduce grievances against the government by positively (re-)shaping citizens’ views (Gunit- sky, 2015; Rød and Weidmann, 2015; King et al., 2017; Qin et al., 2017; Guriev and Treisman, 2019). Which channel dominates is, therefore, an empirical question; the reduced-form impact α reflects the net effect of all these potential mechanisms. 2 The ihs transformation approximates the natural logarithm of the variable for large values, while retaining 0 values (see, e.g., Burbidge et al., 1988; Bellemare and Wichman, 2020), and helps reduce the skewness of the dependent variable. Appendix Table A3 produces regression results with alternative transformations of the Daesh recruit count variable, including the linear Rdea , log(1 + Rdea ), and the dummy variable 1{Rdea > 0}. Given that most cells do not have any recruits, the log transformation log Rdea is not an appropriate alternative. 3 Ordinary Least Squares (OLS) estimates of this parameter, however, will potentially be biased. To alleviate potential endogeneity concerns, we use two different empirical strategies. Fixed effects A first approach consists of controlling for delegation, education, and age group fixed effects. Namely, we write error term εdea as εdea = θd + θe + θa + ζdea . (2) Fixed-effect estimations account for observable and non-observable delegation characteristics; our identification therefore leverages differences between individuals whose age and education differ but who live in the same delegation. This strategy accounts for common unobserved del- egation characteristics such as delegation-level income, local inequality, availability and quality of public services, local government repression, and history. Our fixed-effect estimates might nonetheless be biased if some unobserved cell characteristics were correlated with both internet access and propensity to join Daesh. Similarly, it may fail to address reverse causality: individuals who would want to enlist might actively seek an internet connection to do so. However, given that the total number of Daesh recruits is very small, 0.046 percent of the total population, reverse causality is unlikely to be of first order relevance as it would explain a negligible portion of internet access. Instrumental variables To address the potential biases, we use an instrumental variable strategy, which consists of estimating (1) using two-stage least squares (2SLS), with a first stage of the form Internetdea = γZd + Xdea · η + νdea (3) where Zd is a variable that varies geographically only and affects internet access but is excluded from the second stage. To that end, we follow Manacorda and Tesei (2020), Guriev et al. (2020), and Andersen et al. (2012) and construct instrumental variable Zd , which measures the incidence of lightning strikes on delegation d as proxied by flash density, i.e. the average number of ground flashes per square kilometer per year between 1995 and 2010. The rationale behind the first stage is that lightning 4 strikes increase the user cost of IT related infrastructure because they create voltage surges and dips (Andersen et al., 2011, 2012); accordingly, we expect a negative first-stage coefficient. Our exclusion restriction, therefore, postulates that lightning strikes affect radicalization only through their impact on internet access. However, as discussed by Manacorda and Tesei (2020), lightning strikes might be correlated with other climatic or topographical characteristics like rainfall, which in turn might affect economic development. Since the instrument varies only at the delegation level, controlling for delegation fixed effects is no longer feasible. We nonetheless include in the vector of control variables Xdea a measure of delegation d’s income as proxied by (log) nighttime lights, its (log) area, and the (log) average monthly amount of rain that fell there. Addressing spatial correlation The instrumental variable strategy makes salient that de- ployment of the internet, a network technology, is likely to be spatially correlated. Thus, for both error terms εdea (or ζdea for the fixed-effect estimations) and νdea , we allow for spatial correlation. We first allow εdea and εde a to be correlated, by clustering standard errors at the delegation level, but given the potential spatial dependence in internet access and its instrumen- tal variable, our preferred specifications use the procedure from Conley (1999) to accommodate spatial correlation. 3 Data To estimate equation (1) and its variants, we combine a dataset on Daesh foreign personnel that contains information on Tunisian recruits with socio-economic and other data on the geographic areas those recruits are coming from. Daesh foreign recruits Our main source of data comes from Daesh administrative records on its foreign fighters. The dataset, believed to have been leaked by a defector, covers 3965 male foreign recruits, which constitutes an estimated 30 percent of foreign fighters who joined the organization from early 2013 through late 2014 (Dodwell et al., 2016a).3 3 Dodwell et al. (2016a) checked the information in the database against U.S. Department of Defense intelli- gence records and found high levels of consistency. The data were used in earlier cross-country studies (Dodwell 5 The data have the advantage that recruits report not only their age and education but also where they resided before enrolling; of the 605 Tunisian recruits, we are able to map delegation of residence for 501, of whom 468 also have information on education and age. Our sample for the analysis thus consists of these 468 Daesh recruits.4 The average number of recruits per delegation is 1.9, but the average masks a highly skewed distribution (see Figure 1a): recruits come from 94 delegations only (out of 263), whereas Bizerte, the northernmost city in Africa, had 30, the highest number. The Tunis area was also a recruitment hotspot. Of the 468 Tunisian recruits in our database, 128 are from the governorates of Tunis and Bizerte. In general, recruitment rates are much higher in urban areas. Detailed information on the socio-economic characteristics of recruits (including their age and education), allows us to construct a variable of recruitment that counts the number of Tunisian recruits who resided in delegation d, have education level e, and have reported to be a years old; we therefore created five-year age groups so that each cell (observation) in the dataset is characterized by the triplet {d, e, a}. Table 1 provides both cell-level and delegation-level summary statistics for our variables of interest. The number of recruits in a given cell ranges from 0 to 8. Socio-economic data The 2014 Tunisia Census of Population and Housing is the source of our main independent variables. The census collected information on demographics, educational attainment, marital status, unemployment, television and radio ownership, emigration, and et al., 2016a,b; Jayakumar and Sumpter, 2019; Morris, 2020; Evans et al., 2020; Brockmeyer et al., forthcoming), but to the best of our knowledge they have not been used for country case studies. 4 Using available information on Daesh recruit age, education, occupation before joining, marital status, pre- vious experience with jihad, and self-reported knowledge of Islam, Appendix Table A2 presents tests examining whether there are systematic differences in the characteristics of fighters who were included in the analysis and those who are not. For each variable grouping, the F-test of joint significance of the difference in means between our sample and the data universe fails to reject that these are equal to zero. Our analysis sample thus does not appear to suffer from selection bias. In addition, Brockmeyer et al. (forthcoming) show that the larger database exhibits high correlation with alternative sources of information on Daesh foreign recruits; they look at country aggregates and conclude that the data are representative at the country level. We conduct a similar exercise and compare the reported delegations of last residence for observations in our dataset with the ones obtained by Zelin (2018) on 121 Tunisian recruits who died in Syria and Iraq. The correlation between our data on recruitment per delegation and Zelin’s measure of deceased foreign fighters per delegation is 0.71. Our data thus appear to accurately represent the spatial distribution of Daesh recruitment in Tunisia. 6 Figure 1: Recruits, internet access, and lightning strikes in Tunisia (a) Number of Daesh recruits (b) Internet access (rate) (c) Lightning strikes Stats (N = 263) Stats (N = 263) Stats (N = 263) 0.000 = Min 0.086 = Min 0.022 = Min 30.000 = Max 0.802 = Max 0.306 = Max 0.000 = P50 0.335 = P50 0.182 = P50 1.779 = Avg 0.347 = Avg 0.158 = Avg Legend N in parentheses Lightning strikes Legend (0.27,0.30] (3) N in parentheses (0.24,0.27] (24) Legend Internet access (rate) (0.21,0.24] (16) N in parentheses (0.600,1.000] (17) (0.18,0.21] (90) Number of Daesh recruits (0.500,0.600] (18) (0.15,0.18] (19) (15,35] (5) (0.400,0.500] (49) (0.12,0.15] (29) (10,15] (10) (0.300,0.400] (74) (0.09,0.12] (20) (5,10] (13) (0.200,0.300] (66) (0.06,0.09] (42) [1,5] (66) (0.100,0.200] (38) (0.03,0.06] (15) 0 (169) [0.000,0.100] (1) [0.00,0.03] (4) Notes: Figure 1a maps the total number of Daesh recruits in each delegation. Figure 1b indicates internet access rates (as measured by the share of men that used the internet in the last 3 months) for each delegation. Figure 1c shows the intensity of lightning strikes for each delegation (as measured by the average yearly number of strikes per square kilometers over the period 1990-2010). Appendix Table A1 provides variable definitions. 7 Table 1: Descriptive statistics N Mean Std. Min Max Dev. Panel A: Delegation-age-education {dea} level variables Internet access (rate) 7878 0.437 0.306 0.000 1.000 Married (rate) 7878 0.597 0.377 0.000 1.000 Number of Daesh recruits 7878 0.059 0.366 0.000 8.000 Population – in thousands 7878 0.520 0.579 0.001 5.237 Radio ownership (rate) 7878 0.535 0.165 0.000 1.000 TV ownership (rate) 7878 0.953 0.038 0.000 1.000 Unemployment (rate) 7878 0.088 0.104 0.000 1.000 Panel B: Delegation d level variables Main Area – in thousands 263 0.587 1.855 0.002 26.526 Internet access (rate) 263 0.347 0.141 0.086 0.802 Married (rate) 263 0.566 0.034 0.380 0.639 Mountains (share) 263 0.179 0.288 0.000 1.000 Number of Daesh recruits 263 1.779 4.355 0.000 30.000 Lightning strikes 263 0.158 0.064 0.022 0.306 Radio ownership (rate) 263 0.509 0.144 0.101 0.817 TV ownership (rate) 263 0.949 0.022 0.832 0.983 Unemployment (rate) 263 0.079 0.033 0.027 0.252 1995-2010 nighttime lights – in thousands 263 2.331 1.772 0.122 16.965 1995-2010 rainfall 263 33.657 15.494 5.401 83.211 2014 nighttime lights – in thousands 263 4.407 4.542 0.126 60.279 2014 rainfall 263 31.319 14.817 4.310 84.919 Other outcomes Convicted terrorists 263 0.487 1.223 0.000 11.000 Emigration (rate) 263 0.010 0.010 0.000 0.074 Islamist vote (share) 263 0.279 0.121 0.089 0.678 Jihadist lectures (dummy) 263 0.198 0.399 0.000 1.000 Protests and riots 263 4.719 22.714 0.000 346.0 Notes: Our analysis considers 263 delegations rather than 264 because Douz North and Douz Sud are combined into one reporting area. We only consider male population in our estimations where possible - data on elections and civil unrests are not disaggregated by gender. Variables expressed in thousands are multiplied by 1000 throughout our analyses. Variable definitions are available in Appendix Table A1. internet access.5 The average delegation has a 35 percent (male) internet access rate; the aggregate nation-wide rate is 39 percent because more densely populated urban areas tend to be more connected. Figure 1b shows how geographically heterogeneous internet access is; metropolitan areas closer to the coast have high access rates, while these tend to be low in the interior of Tunisia and in delegations abutting the Algerian border. To match census data with our Daesh recruit data, we aggregate the census data to a level below the delegation, i.e. the cell level. Recall that a cell is characterized by the delegation d, education level e (primary or below, secondary, and tertiary and higher), and a five-year age group a (starting from at 15-19 through 60-64). We construct demographic, unemployment, and internet and media usage variables for each cell by taking the mean among men in that cell. Our data set thus consists of 7890 observations, but the population in 12 of the {d, e, a} cells is equal to zero, so they are excluded them from the estimation sample, which eventually consists of 7878 cell-level observations. Definitions and sources of the variables used in our analyses are provided in Appendix Table A1 and summary statistics in Table 1. Instrumental variables As an instrumental variable for internet access, we use the average lightning strike intensity per square km per year over the period 1995-2010 obtained from Man- acorda and Tesei (2020). They use data collected by the U.S. National Aeronautics and Space Administration (NASA) and processed by the Global Hydrology Resource Center (GHRC). Figure 1c shows the spatial distribution of lightning strikes. The map shows a high degree of spatial correlation, underscoring the need to accommodate this feature in the econometric anal- ysis. Lightning strikes are least prevalent in the south, which is covered by the Sahara desert, and most prevalent in the Chaambi mountains close to the Algerian border and in the Cape Bon peninsula located in the North East. A comparison of the spatial variations in lightning (Figure 1c) and in internet access (Figure 1b) points to a high correlation. Other data We complement census data with delegation-level nighttime light data (Li et al., 2020) for the 1995-2014 period, which we use as a proxy for income. Information on 5 We only consider metrics for the male population where possible, since the sample of recruits comprises men only. In practice, this distinction makes little difference to our results. The correlation between male internet access rate and overall internet access rate is ρ = .94. 9 elevation (mountains) comes from the UN Environment Programme World Conservation Mon- itoring Centre (UNEP-WCMC, 2002). Rainfall data were obtained from the GPCC Global Precipitation Climatology Centre (Schneider et al., 2015). We also collected data on Tunisia’s socio-political landscape, using legislative election data for 2014 from the national election au- thority (ISIE, 2015). Finally, alternative measures of violent extremism comprise information on political protests, riots, and political violence from the Armed Conflict Location & Event Data (ACLED) Project database (Raleigh et al., 2010) and Zelin (2020)’s data on the locations where extremist group Ansal al-Sharia Tunisia (AST) delivered jihadist religious lectures - a vehicle for the group’s proselytization and recruitment activities. We also compiled data from the National List of Persons, Organizations and Entities Associated with Terrorist Crime cu- rated by the anti-terrorism commission, which tracks all individuals indicted for terrorism as listed in the Official Gazette (CNLCT, 2021). It also lists their addresses, which allows us to georeference them.6 4 Results Our main finding that internet access predicts recruitment is shown in Table 2. Columns 1-6 show progressively more stringent OLS estimation results. Column 1 documents a positive and statistically significant correlation between (the inverse hyperbolic sine of- henceforth ihs) the number of Daesh recruits and internet access rates, while normalizing by the population of each cell by controlling for the logarithm of the number of men in a particular delegation-education- age cell as enumerated in the census. Standard errors are clustered at the delegation level. Recognizing that the nature of internet access might lead error terms in (1) to be spatially correlated, the specification in column 1 is replicated in column 2 but following Conley (1999) to account for spatial correlation of the error term with a 50-km buffer area. The results are very similar. Given the scope for spatial correlation in internet access, we henceforth adopt Conley (1999) standard error correction procedures. This association between internet access and jihadist recruitment is robust to including addi- 6 Note that the terrorist acts documented therein were committed between 2004 and 2018; the timespan the data cover is thus different from that covered by our data on Daesh recruitment 10 tional control variables. Column 3 shows that adding delegation-level controls such as income (proxied by nighttime light intensity and rainfall, both averaged over the years 1995-2010), and {dea} cell-level controls including unemployment and marriage rates, education and age- category dummy variables, leads to a reduction in the coefficient on internet access. These control variables not only explain Daesh recruitment but are also correlated with internet ac- cess. Column 4 adds a control for the Islamist vote share in the 2014 legislative elections (ISIE, 2015) as a proxy for political preferences for Islamism; doing so hardly affects the coefficient on internet access. Next, in column 5, we add proxies for extremism notably the occurence Ansar-Al-Sharia-sponsored jihadist lectures (Zelin, 2020), the number of convicted domestic terrorists (CNLCT, 2021), and the incidence of violent protests and riots (Raleigh et al., 2010). These variables could themselves be influenced by internet penetration; the estimated coeffi- cient of interest thus captures the effect of internet access on Daesh recruitment above and beyond the internet’s independent influence on these other forms of radicalization. Delegations where Ansar-Al-Sharia-sponsored jihadist lectures took place are more likely to produce Daesh recruits and the slight reduction in the coefficient for internet access is suggestive that internet might actually have played a role in the decision to locate these lectures.7 Finally, to capture all sources of latent delegation-level heterogeneity, both observable and unobservable, we add delegation fixed effects in column 6, which presents our preferred OLS specification. While less precisely estimated, the coefficient for the internet remains positive and statistically significant at the 5 percent level. In short OLS estimations yield a robust correlation between internet access and the extent of Daesh foreign fighter recruitment. As discussed in Section 2, our fixed-effects estimation in Column 6 accounts for both observable and non-observable delegation-level characteristics. As such, delegation-level variables such as income, civil liberties, availability and quality of public goods, government repression, etc. are accounted for in the delegation fixed-effect variable. However, since we cannot rule out latent heterogeneity, fixed-effect estimations might still be biased. We therefore estimate the relation- ship using 2SLS. We instrument the internet variable using lightning strike counts (Manacorda and Tesei, 2020; Guriev et al., 2020), which measure the average number of lightning strikes 7 In separate analyses not shown here, we however do not find any significant correlation between internet access and the occurrence of jihadist lectures. 11 Table 2: Internet and transnational jihadist recruitment Dependent variable Number of Daesh recruits (inverse hyperbolic sine) OLS 2SLS (1) (2) (3) (4) (5) (6) (7) (8) Panel A: OLS and 2SLS second stage Internet access (rate) 0.105 0.105 0.074 0.075 0.060 0.072 0.318 0.557 (0.017) (0.016) (0.025) (0.025) (0.025) (0.028) (0.159) (0.214) Population (log) 0.026 0.026 0.013 0.013 0.007 -0.002 0.002 -0.014 (0.004) (0.003) (0.004) (0.004) (0.003) (0.004) (0.008) (0.010) Unemployment (percentage) 0.012 0.012 0.002 0.014 0.056 0.096 (0.039) (0.039) (0.038) (0.037) (0.053) (0.060) Married (percentage) -0.033 -0.032 -0.012 0.027 0.001 0.064 (0.040) (0.041) (0.042) (0.042) (0.048) (0.057) Delegation area (log) -0.009 -0.009 -0.003 -0.002 0.009 (0.003) (0.003) (0.002) (0.005) (0.006) 1995-2010 nighttime lights (log) -0.003 -0.003 -0.006 -0.002 -0.001 (0.005) (0.005) (0.005) (0.005) (0.006) 1995-2010 rainfall (log) -0.009 -0.010 -0.009 0.005 0.013 (0.006) (0.007) (0.007) (0.012) (0.014) Islamist vote (share) -0.005 -0.025 -0.081 (0.026) (0.026) (0.036) Jihadist lectures (dummy) 0.067 0.055 (0.010) (0.010) Convicted terrorists 0.010 0.011 (0.004) (0.004) Protests and riots 0.000 0.000 (0.000) (0.000) Panel B: 2SLS first stage Lightning strikes (log) -0.047 -0.043 (0.006) (0.006) Age group (dummies) No No Yes Yes Yes Yes Yes Yes Education group (dummies) No No Yes Yes Yes Yes Yes Yes Delegation FE No No No No No Yes No No Std. err. clustered by delegation Yes No No No No No No No Std. err. correction (Conley, 1999) No Yes Yes Yes Yes Yes Yes Yes Kleibergen-Paap Test 66.183 51.521 N 7878 7878 7878 7878 7878 7878 7878 7878 R2 0.032 0.032 0.074 0.074 0.090 0.228 0.061 0.036 Notes: Standard errors are reported in parentheses, and are clustered by delegation (column 1) or corrected for spatial correlation following Conley (1999) using a 50 km distance cutoff (columns 2-8). OLS estimates are shown in columns 1-6, and 2SLS estimates in columns 7-8. Panel A presents the estimates for the OLS and the second stage of 2SLS. Panel B reports only the coefficient for the excluded variable. Variable definitions are available in Appendix Table A1 and summary statistics in Table 1. 12 per square kilometer and per year over the period 1995-2010.8 This is a strong instrument as is evidenced by the Kleibergen and Paap (2006) test (henceforth KP) statistic of 66 reported in the bottom panel of column 7 of Table 2. As was shown in Figure 1b areas buffeted by more intense lightning strikes have worse internet access. The 2SLS estimate of the impact of the internet on Daesh recruitment is statistically significant at the 10 percent level and is three times as large as its OLS counterpart. Our exclusion restriction hinges on the assumption that lightning strikes affect the propensity to join Daesh only through the availability and quality of internet. As lightning is correlated with rainfall (ρ = 0.63 in the data), which could affect economic development (e.g. through higher agricultural productivity), column 7 controls for averages of nighttime lights and rainfall over the period 1995-2010, which is also the time inter- val the instrumental variable is constructed from. Choosing lagged variables further mitigates the possibility that internet access has had an impact on nighttime lights, which could bias the first stage. In column 8, we revisit these results by controlling for delegation-level measures of radicalization into violent extremism used in column 5; if anything, doing so produces larger estimates for the effect of internet access. Robustness Table 3 shows the results of a number of robustness checks. Where applicable, we present both OLS regression results (with delegation, age, and education fixed effects as in Table 2, column 6) in panel A and 2SLS estimates in panel B, using lightning strikes as instrument for internet access rates (as in Table 2, column 7). We start by estimating three alternative specifications. In Table 2, column 7, we argued that satisfying the exclusion restriction requires controlling for GDP proxies and climatic data. We therefore added rainfall and nighttime light intensity variables as controls, and averaged these variables over the same time span as for the instrument, 1995-2010. We use 2014 averages instead to measure values of GDP and rainfall that are contemporaneous with the dependent variable. The estimated 2SLS coefficient for internet access (panel B, column 1), though slightly lower, is still significant at the 5 percent level. Next, we investigate whether the documented association between internet and jihadist recruitment could be driven by usage of other media. 8 Yearly data on lightning strikes with the same level of geographical detail is not available (Manacorda and Tesei, 2020) and lightning strikes are highly serially correlated (Andersen et al., 2012). 13 Table 3: Robustness tests Dependent variable Number of Daesh recruits (inverse hyperbolic sine) Alternative Sample Spatial specifications restrictions correlation 2014 Other Emi- No No No 20 km 100 km values media gration Tunis mountain outliers radius radius (1) (2) (3) (4) (5) (6) (7) (8) Panel A: OLS with delegation FE Internet access (rate) 0.072 0.074 0.073 0.075 0.075 0.061 0.072 0.072 (0.028) (0.028) (0.028) (0.028) (0.029) (0.025) (0.028) (0.032) Radio ownership (rate) -0.020 (0.022) TV ownership (rate) 0.039 (0.041) Emigration (rate) 0.022 (0.031) N 7878 7878 7878 7248 7489 7788 7878 7878 R2 0.228 0.228 0.228 0.222 0.230 0.208 0.228 0.228 Panel B: 2SLS second stage Internet access (rate) 0.255 0.369 0.312 0.318 0.306 0.359 0.318 0.318 (0.133) (0.177) (0.157) (0.156) (0.166) (0.164) (0.168) (0.165) Radio ownership (rate) -0.058 (0.029) TV ownership (rate) -0.041 (0.061) Emigration (rate) 0.053 (0.044) Kleibergen-Paap Test 85.730 58.512 66.124 70.412 61.072 64.488 138.982 38.306 N 7878 7878 7878 7248 7489 7788 7878 7878 Notes: Standard errors are reported in parentheses and are corrected for spatial correlation following Conley (1999) using 50 km (columns 1-6), 20 km (column 7), or 100 km (column 8) distance cutoffs. All specifications include population (log), unemployment (rate), marital status (marriage rate), area (log), age group (dummies) and education group (dummies) as controls. Column 1 further controls for 2014 nighttime lights (log) and 2014 rainfall (log), while columns 2-8 use 1995-2010 nighttime lights (log) and 1995-2010 rainfall (log) as in Table 2. Panel A reports OLS estimates, which also account for delegation FE. Panel B reports 2SLS estimates that use lightning strikes (log) as an instrument for internet access (rate). Columns 1-3 and 7-8 use the entire sample, while column 4 removes delegations in the Tunis metropolitan area (21 delegations), column 5 removes delegations with the highest coverage of mountains (13 delegations), and column 6 removes delegations with the highest number of Daesh recruits per capita (3 delegations). Variable definitions are available in Appendix Table A1 and summary statistics in Table 1. A “horse race” between internet and TV and radio access helps to discriminate between these competing explanations. Column 2 shows the results when both radio and TV access are included in the regression: in both OLS and 2SLS specifications, neither radio nor TV access have much predictive power. In fact, the coefficient on internet access rises somewhat once these variables are included, which attests to the unique role of the internet as compared to media in general.9 As a third potential mechanism to rule out, column 3 adds a control for emigration, to address the possibility that the association between internet access and recruitment is simply the effect of the internet on emigration. Emigration rates are however not a significant predictor of recruitment and do not affect the internet/Daesh recruitment relationship. This suggests that the mechanism by which internet access enables recruitment of international jihadists operates beyond the lowering of transaction costs. In the next three columns, we remove potential outliers and specific geographic areas. We verify that the relationship we documented between internet access and terrorist recruitment is not driven by Tunis, the capital and most populated and connected city, from which a third of all Daesh recruits come. After delegations in the metropolitan area of Tunis are removed, column 4 indicates that the recruitment-internet correlation is still statistically significant. Column 5 excludes delegations in mountainous areas where jihadists can more easily hide from law enforcement (Zelin, 2019) and internet access is poor. Removing the delegations with mountains covering more than 85 percent of their territory or above the 95th percentile (in doing so, we adopt the Manacorda and Tesei (2020) definition of a mountainous area) does not change the result. Finally, in column 6, we check if outliers in Daesh recruitment – defined as delegations that are in the top 1 percent of the distribution of Daesh recruitment per capita– were driving our results, and they were not. The last set of robustness checks evaluated how standard errors change when different distance cutoffs for the buffer zone are used to compute standard errors following Conley (1999). Recall that we use 50 km in the baseline specifications. Column 7 (resp. 8) shows the estimation result of the same regression with a shorter (resp. longer) radius of 20 km (resp. 100 km). Although standard errors do increase when we widen the radius, the association between internet access 9 The same conclusions can be drawn if radio or TV access are considered separately. 15 and radicalization remains statistically significant at conventional levels. Discussion Our findings resonate with anecdotal evidence on Daesh recruitment practices:10 its strategy entailed monitoring the web to identify prospective recruits, establish first contact, cultivate personal contact and a sense of community, and culminate in a call for action (Berger, 2015). Speckhard and Ellenberg (2020), who interviewed 220 Daesh prisoners, defectors and re- turnees, found that internet played an important role in recruitment, with 8.2 percent of recruits claiming they were recruited solely over the internet. In addition, 11.8 percent mentioned the importance of internet recruiters, 22 percent mentioned being influenced by pro-Daesh content on YouTube, 10.9 percent mentioned Facebook, and 5.9 percent mentioned Twitter. Similarly, when testifying about American jihadists during a U.S. Senate hearing, Mr. Bergen, a respected national security analyst, argued “[t]he only profile that this group really shares is that 53 of the 62 individuals were very active on social media, downloading and sharing jihadist propaganda, and in some cases (...) directly communicating with members of ISIS in Syria” (United States Senate, 2015). What do our results suggest in terms of Daesh’s reliance on the internet for recruitment? To calculate the elasticity of enrollment in the terror group with respect to internet access rates, we use the fixed-effect specification in Table 2, column 4, which yields a more conservative estimate than the larger estimates obtained using 2SLS. Because our left-hand side variable is the inverse hyperbolic sine transformation of the number of Daesh recruits, the regression coefficient does not have a straightforward elasticity interpretation, especially since in our sam- ple the median delegation in Tunisia has zero Daesh recruits. To compute an elasticity, we calculate for each {dea} cell how the number of recruits would increase if internet access in- creased by one percent. The resulting change in (the ihs of) the number of recruits is given by ∆asinh(Rdea ) = αInternetdea × 0.01. The elasticity of Daesh recruitment with respect to 1 sinh[Rdea +αInternetdea ×0.01] internet access rates is thus ε = .01 × Rdea . ˆ = 0.072 With an estimated α 10 Daesh was very active in online forums; it also maintained a well-read online magazine (Dabiq) and various blogs encouraging readers to travel to Syria and Iraq to join the jihad and offering tips on how to travel and avoid getting caught. It also had a strong social media presence (Bloom et al., 2017; Conway et al., 2019). By 2014, at least 46,000 Twitter accounts were overtly supporting Daesh (Berger and Morgan, 2015) and in 2015 Daesh issued over 17 million tweets, about 0.2 percent of all Arabic twitter content. Daesh is also infamous for its graphic footage of beheadings and violent attacks, which were widely viewed, with over 20 percent of the U.S. population having watched at least part of an ISIS video (Redmond et al., 2019). 16 (Table 2 column 5), a one percent increase in the rate of internet access by Tunisian men in- creases the total number of recruits in the sample by an estimated 2.6 individuals, which, with a sample size of 468 recruits, yields a .56 elasticity. Our quantitative results thus complement qualitative accounts on the importance of the internet for Daesh recruitment of foreign fighters. 5 Conclusion By allowing new forms of social interaction and acquisition of information the internet has fundamentally reshaped society. While it is often a force for development, this paper presents evidence that it has also been accompanied by the rise of violent extremist groups such as Daesh, the epitome of tech-savvy terrorism. After combining administrative data from the international jihadist organization with census data from Tunisia, we document quantitatively that the internet has been, on net, a potent facilitator of international terrorist recruitment. Our analysis suggest a 0.56 elasticity of Daesh recruitment with respect to internet access, attesting to the proclivity of transnational terror organizations to (ab)use cutting-edge information and communication technologies. References Abadie, Alberto, “Poverty, Political Freedom, and the Roots of Terrorism,” American Eco- nomic Review, May 2006, 96 (2), 50–56. Andersen, Thomas Barnebeck, Jeanet Bentzen, Carl-Johan Dalgaard, and Pablo Selaya, “Does the Internet Reduce Corruption? Evidence from US States and Across Coun- tries,” The World Bank Economic Review, 2011, 25 (3), 387–417. , , , and , “Lightning, IT Diffusion, and Economic Growth Across US States,” Review of Economics and Statistics, 2012, 94 (4), 903–924. Autor, David, “Wiring the Labor Market,” Journal of Economic Perspectives, March 2001, 15 (1), 25–40. 17 Bail, Christopher A., Friedolin Merhout, and Peng Ding, “Using Internet Search Data to Examine the Relationship between Anti-Muslim and pro-ISIS Sentiment in U.S. Counties,” Science Advances, 2018, 4 (6). 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. Baum, Matthew A and Philip BK Potter, “Media, Public Opinion, and Foreign Policy in the Age of Social Media,” The Journal of Politics, 2019, 81 (2), 747–756. Bellemare, Marc and Casey Wichman, “Elasticities and the Inverse Hyperbolic Sine Trans- formation,” Oxford Bulletin of Economics and Statistics, 2020, 82 (1), 50–61. Benmelech, Efraim and Esteban F. Klor, “What Explains the Flow of Foreign Fighters to ISIS?,” Terrorism and Political Violence, 2018, 0 (0), 1–24. Berger, J.M., “Tailored Online Interventions: The Islamic state’s Recruitment Strategy,” CTC Sentinel, 2015, 8 (10), 19–23. and Jonathon Morgan, “The ISIS Twitter Census: Defining and Describing the Popula- tion of ISIS Supporters on Twitter,” The Brookings Project on US Relations with the Islamic World, 2015, 3 (20), 4–1. Bhuller, Manudeep, Tarjei Havnes, Edwin Leuven, and Magne Mogstad, “Broadband Internet: An Information Superhighway to Sex Crime?,” Review of Economic Studies, 2013, 80 (4), 1237–1266. Bloom, Mia, Hicham Tiflati, and John Horgan, “Navigating ISIS’s Preferred Platform: Telegram,” Terrorism and Political Violence, 2017, 31 (6), 1242–1254. ement Joubert, Mohamed Abdel Jelil, and Brockmeyer, Anne, Quy-Toan Do, Cl´ Kartika Bhatia, “Transnational Terrorist Recruitment: Evidence from Daesh Personnel Records,” The Review of Economics and Statistics, forthcoming. 18 Burbidge, John, Lonnie Magee, and Leslie Robb, “Alternative Transformations to Han- dle Extreme Values of the Dependent Variable,” Journal of the American Statistical Associ- ation, 1988, 83 (401), 123–127. CNLCT, Commission Nationale de Lutte Contre le Terrorisme, “The National List of Persons, Organizations and Entities Associated with Terrorist Crime,” https://bit.ly/ 3olNHPO 2021. Accessed: 2021-04-19. Conley, Timothy, “GMM Estimation with Cross Sectional Dependence,” Journal of Econo- metrics, 1999, 92 (1), 1–45. Conway, Maura, “Determining the Role of the Internet in Violent Extremism and Terrorism: Six Suggestions for Progressing Research,” Studies in Conflict & Terrorism, 2017, 40 (1), 77–98. , Moign Khawaja, Suraj Lakhani, Jeremy Reffin, Andrew Robertson, and David Weir, “Disrupting Daesh: Measuring Takedown of Online Terrorist Material and Its Im- pacts,” Studies in Conflict & Terrorism, 2019, 42 (1-2), 141–160. 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. , , and , “Then and Now: Comparing the Flow of Foreign Fighters to AQI and the Islamic State,” Technical Report, US Military Academy-Combating Terrorism Center West Point United States 2016. Enders, Walter and Gary Hoover, “The Nonlinear Relationship between Terrorism and Poverty,” American Economic Review, 2012, 102 (3), 267–72. Evans, Tyler, Daniel J Milton, and Joseph K Young, “Choosing to Fight, Choosing to Die: Examining How ISIS Foreign Fighters Select Their Operational Roles,” International Studies Review, 2020. 19 FATF, Financial Action Task Force, “Financing of the terrorist organisation Islamic State in Iraq and the Levant (ISIL),” FATF Report 2015. Freund, Caroline and Diana Weinhold, “The Effect of the Internet on International Trade,” Journal of International Economics, 2004, 62 (1), 171–189. Gaibulloev, Khusrav and Todd Sandler, “What We Have Learned about Terrorism since 9/11,” Journal of Economic Literature, June 2019, 57 (2), 275–328. Gartenstein-Ross, Daveed, “Jihad 2.0: Social Media in the Next Evolution of Terrorist Recruitment. Hearing before the Senate Committee on Homeland Security & Governmental Affairs (7 May 2015),” 2017. Gassebner, Martin and Simon Luechinger, “Lock, stock, and barrel: A comprehensive assessment of the determinants of terror,” Public Choice, 2011, 149 (3-4), 235. Gohdes, Anita R, “Repression technology: Internet accessibility and state violence,” Amer- ican Journal of Political Science, 2020, 64 (3), 488–503. Greenberg, Karen J., “Counter-Radicalization via the Internet,” The ANNALS of the Amer- ican Academy of Political and Social Science, 2016, 668 (1), 165–179. Gunitsky, Seva, “Corrupting the Cyber-Commons: Social Media as a Tool of Autocratic Stability,” Perspectives on Politics, 2015, 13 (1), 42–54. Guriev, Sergei and Daniel Treisman, “Informational Autocrats,” Journal of Economic Perspectives, November 2019, 33 (4), 100–127. , Nikita Melnikov, and Ekaterina Zhuravskaya, “3G Internet and Confidence in Gov- ernment,” The Quarterly Journal of Economics, 12 2020. Hegghammer, Thomas, “Should I Stay or Should I Go? Explaining Variation in Western Ji- hadists’ Choice between Domestic and Foreign Fighting,” American Political Science Review, 2013, 107 (1), 1–15. Hjort, Jonas and Jonas Poulsen, “The Arrival of Fast Internet and Employment in Africa,” American Economic Review, March 2019, 109 (3), 1032–79. 20 erieure Ind´ ISIE, Instance Sup´ esultats finaux par cir- ependante pour les Elections, “R´ elections l´ conscription pour les ´ egislatives 2014,” https://bit.ly/30AEnQ5 2015. Accessed: 2021-09-20. Jayakumar, Shashi and Cameron Sumpter, “Southeast Asian Fighters from Islamic State Leaks,” Perspectives on Terrorism, 2019, 13 (6), 58–72. King, Gary, Jennifer Pan, and Margaret E. Roberts, “How the Chinese Government Fabricates Social Media Posts for Strategic Distraction, Not Engaged Argument,” American Political Science Review, 2017, 111 (3), 484–501. Kleibergen, Frank and Richard Paap, “Generalized Reduced Rank Rests Using the Sin- gular Value Decomposition,” Journal of Econometrics, 2006, 133 (1), 97–126. Krueger, Alan and David Laitin, “Kto Kogo?: A Cross-country Study of the Origins and Targets of Terrorism,” in Philip Keefer and Norman Loayza, eds., Terrorism, Economic Development, and Political Openness, 01 2008, pp. 148–173. ckov´ and Jitka Maleˇ a, “Education, Poverty and Terrorism: Is There a Causal Connec- tion?,” Journal of Economic Perspectives, 2003, 17 (4), 119–144. Li, Quan and Drew Schaub, “Economic Globalization and Transnational Terrorism: A pooled Time-Series Analysis,” Journal of Conflict Resolution, 2004, 48 (2), 230–258. Li, Xuecao, Yuyu Zhou, Min Zhao, and Xia Zhao, “A Harmonized Global Nighttime Light Dataset 1992–2018,” Scientific Data, 2020, 7 (168). Litan, Robert and Alice Rivlin, “Projecting the Economic Impact of the Internet,” Amer- ican Economic Review, May 2001, 91 (2), 313–317. Manacorda, Marco and Andrea Tesei, “Liberation Technology: Mobile Phones and Polit- ical Mobilization in Africa,” Econometrica, 2020, 88 (2), 533–567. Mastrobuoni, Giovanni, “Crime is Terribly Revealing: Information Technology and Police Productivity,” The Review of Economic Studies, 2020, 87 (6), 2727–2753. 21 Miner, Luke, “The Unintended Consequences of Internet Diffusion: Evidence from Malaysia,” Journal of Public Economics, 2015, 132, 66–78. Mitts, Tamar, “From Isolation to Radicalization: Anti-Muslim Hostility and Support for ISIS in the West,” American Political Science Review, 2019, 113 (1), 173–194. , Gregoire Phillips, and Barbara Walter, “Studying the Impact of ISIS Propaganda Campaigns,” The Journal of Politics, forthcoming. Morris, Andrea Michelle, “Who Wants to be a Suicide Bomber? Evidence from Islamic State Recruits,” International Studies Quarterly, 2020, 64 (2), 306–315. uller, Karsten and Carlo Schwarz, “Fanning the Flames of Hate: Social Media and Hate M¨ Crime,” Journal of the European Economic Association, 2021, 19 (4), 2131–2167. omberg, and Yanhui Wu, “Why Does China Allow Freer Social Media? Qin, Bei, David Str¨ Protests versus Surveillance and Propaganda,” Journal of Economic Perspectives, February 2017, 31 (1), 117–40. avard Hegre, and Joakim Karlsen, “Introducing Raleigh, Clionadh, Andrew Linke, H˚ ACLED: An Armed Conflict Location and Event Dataset: Special Data Feature,” Journal of Peace Research, September 2010, 47 (5), 651–660. Redmond, Sarah, Nickolas Jones, Alison Holman, and Roxane Cohen Silver, “Who Watches an ISIS Beheading—And Why.,” American Psychologist, 2019, 74 (5), 555. Rød, Espen Geelmuyden and Nils B Weidmann, “Empowering Activists or Autocrats? The Internet in Authoritarian Regimes,” Journal of Peace Research, 2015, 52 (3), 338–351. Schneider, Udo, Andreas Becker, Peter Finger, Anja Meyer-Christoffer, Bruno Rudolf, and Markus Ziese, “GPCC Full Data Reanalysis Version 7.0 at 0.5°: Monthly Land-Surface Precipitation from Rain-Gauges built on GTS-based and Historic Data: Grid- ded Monthly Totals,” https://bit.ly/3BTrP3O 2015. Accessed: 2021-04-28. Soufan Center, “Foreign Fighters: An Updated Assessment of the Flow of Foreign Fighters into Syria and Iraq,” New York, 2015. 22 Speckhard, Anne and Molly Ellenberg, “ISIS in Their Own Words,” Journal of Strategic Security, 2020, 13 (1), 82–127. UNEP-WCMC, UN Environment Programme World Conservation Monitoring Centre, “Mountains and Treecover in Mountain Regions 2002,” https://bit.ly/3FQmIUp 2002. Accessed: 2021-04-28. United States Senate, “Senate Hearing. 114-438 - Jihad 2.0: Social Media in the Next Evolution of Terrorist Recruitment,” 2015. Warren, Camber, “Explosive Connections? Mass Media, Social Media, and the Geography of Collective Violence in African States,” Journal of Peace Research, 2015, 52 (3), 297–311. Zelin, Aaron, “Tunisian Foreign Fighters in Iraq and Syria,” Policy Notes: The Washington Institute for Near East Policy, 2018, (55). , “Not Gonna Be Able To Do It: al-Qaeda in Tunisia’s Inability to Take Advantage of the Islamic State’s Setbacks.,” Perspectives on Terrorism, 2019, 13 (1). , Your Sons are at Your Service: Tunisia’s Missionaries of Jihad, Columbia University Press, 2020. Zhuravskaya, Ekaterina, Maria Petrova, and Ruben Enikolopov, “Political Effects of the Internet and Social Media,” Annual Review of Economics, 2020, 12 (1), 415–438. 23 A Supplemental material [online version only] Table A1: Variable descriptions Variable Description Source Panel A: Delegation-age-education {d, e, a} level variables Age group Dummy variables indicating which age group a individ- The 2014 Tunisia Census of Pop- (dummies) uals in a given {d, e, a} cell belong to. There are 10 dif- ulation and Housing (INS)[a] ferent groups, notably: (i) 15-19 years, (ii) 20-24 years, (iii) 25-29 years, (iv) 30-34 years, (v) 35-39 years, (vi) 40- 44 years, (vii) 45-49 years, (viii) 50-54 years, (ix) 55-59 years, and (x) 60 years or more. Education group Dummy variables indicating which level of education e The 2014 Tunisia Census of Pop- (dummies) individuals in a given {d, e, a} cell have completed. There ulation and Housing (INS)[a] are 3 different groups, notably: (i) primary or below, (ii) secondary, and (iii) tertiary. Internet access Share of men in the {d, e, a} cell that used the internet The 2014 Tunisia Census of Pop- (rate) in the last 3 months ulation and Housing (INS)[a] Married Share of men in the {d, e, a} cell that are married The 2014 Tunisia Census of Pop- (rate) ulation and Housing (INS)[a] Number of Number of individuals in the {d, e, a} cell who joined Daesh personnel records Daesh recruits Daesh in Syria or Iraq between 2013 and 2014. Note that the personnel records only present information on male recruits. Population Number of men in the {d, e, a} cell The 2014 Tunisia Census of Pop- ulation and Housing (INS)[a] Radio ownership Share of men in the {d, e, a} cell that own a radio The 2014 Tunisia Census of Pop- (rate) ulation and Housing (INS)[a] TV ownership Share of men in the {d, e, a} cell that own a TV The 2014 Tunisia Census of Pop- (rate) ulation and Housing (INS)[a] Unemployment Share of men in the {d, e, a} cell that are unemployed The 2014 Tunisia Census of Pop- (rate) ulation and Housing (INS)[a] Panel B: Delegation d level variables Area Geographical area of each delegation d in km2 . Douz Author’s calculations based on North and Douz Sud are combined into one administra- HDX-OCHA[b] tive division. Convicted Number of individuals indicted for terrorist crimes com- National List of Persons, Or- terrorists mitted between 2004 and 2018 from delegation d. The ganizations and Entities As- records include their addresses, which allows us to geolo- sociated with Terrorist Crime cate them. (CNLCT, 2021)[c] Emigration Share of men from delegation d that emigrated abroad. The 2014 Tunisia Census of Pop- (rate) ulation and Housing (INS)[a] 24 Internet access Share of men in the delegation d that used the internet The 2014 Tunisia Census of Pop- (rate) in the last 3 months. ulation and Housing (INS)[a] Islamist vote The share of votes in the 2014 Legislative Elections for Author’s compilation from ISIE (share) Ennahda, the main Islamist party in Tunisia, in delega- (2015)[d] tion d. Jihadist lectures Dummy variable that assumes value 1 if Ansar al-Sharia Author’s compilation from Zelin (dummy) Tunisia (AST) leadership delivered jihadist religious lec- (2020)[e] tures between 2011 and 2016 in the delegation d, and 0 otherwise. Lightning Average annual intensity of lightning strikes per square Manacorda and Tesei (2020) strikes km for the 1995-2010 period in a given delegation d. Married Share of men in delegation d that are married The 2014 Tunisia Census of Pop- (rate) ulation and Housing (INS)[a] Mountains The share of the area in delegation d covered by moun- Author’s calculations based on (share) tains. We calculate the fraction of mountains within a Mountains of the World 2002 delegation as the share of grid cells within the adminis- (UNEP-WCMC, 2002)[f ] trative division that are covered by one of the following three categories of mountains: (1) the elevation is be- tween 1000 and 1500 meters with either a slope of at least 5 degrees or a local (i.e. within a 7 km radius) elevation range exceeding 300 meters, (2) elevation exceeding 300 but lower than 1000 meters and a local (7 km radius) ele- vation range exceeding 300 meters, and (3) Inner isolated areas (≤ 25 sq.km in size) that do not meet criteria (1) or (2) but are surrounded by mountains. Nighttime lights Total sum of the average annual delegation d level night- Author’s calculations based on time light intensities for the 1995-2010 period and for the the Harmonized Global Night- year of 2014. Li et al. (2020)’s dataset contains annual time Light dataset 1992-2018 (Li time series data. Intensities are expressed in digital num- et al., 2020) bers. Number of Number of individuals in delegation d who joined Daesh Daesh personnel records Daesh recruits in Syria or Iraq between 2013 and 2014. Note that the personnel records only present information on male re- cruits. Population Number of men in delegation d The 2014 Tunisia Census of Pop- ulation and Housing (INS)[a] Protests and Cumulative number of political protests and riot events ACLED[g] (Raleigh et al., 2010) riots that took place between 2011 and 2014 in delegation d. Radio ownership Share of men in delegation d that own a radio The 2014 Tunisia Census of Pop- (rate) ulation and Housing (INS)[a] Rainfall Average monthly rainfall in mm for delegation d for the Author’s calculations based on 1995-2010 period and for the year 2014. The GPCC the GPCC dataset[h] (Schneider dataset (2015) contains monthly time series data and is et al., 2015) aggregated according to the period/year being analyzed. Turnout The share of eligible voters in the 2014 Legislative Elec- Author’s compilation from ISIE (share) tions that cast a vote in delegation d. (2015)[d] 25 TV ownership Share of men in delegation d that own a TV The 2014 Tunisia Census of Pop- (rate) ulation and Housing (INS)[a] Unemployment Share of men in delegation d that are unemployed The 2014 Tunisia Census of Pop- (rate) ulation and Housing (INS)[a] Notes: [a] The 2014 Tunisia Census of Population and Housing was conducted by L’Institut National de la Statistique (INS). [b] Tunisia administrative boundaries for delegations are obtained from The Humanitarian Data Exchange (HDX) and produced by The United Nations Office for the Coordination of Humanitarian Affairs (OCHA). Date of access: March 20, 2019 [link]. [c] The National List of Persons, Organizations and Entities Associated with Terrorist Crime is produced by La Commission Nationale de Lutte Contre le Terrorisme (CNLCT, 2021). Date of access: April 19, 2021 [link].[d] Elections in Tunisia are conducted by the Instance Sup´erieure Ind´ependante pour les Elections ISIE (2015). Date of access: September 20, 2021 [link]. [e] The Ansar al-Sharia in Tunisia’s Clerical Establishment dataset is described in Zelin (2020). Date of access: March 25, 2021 [link]. [f] The Mountains of the World 2002 dataset, produced by The UN Environment Programme World Conservation Monitoring Centre (UNEP-WCMC), is described in UNEP-WCMC (2002). Date of access: April 28, 2021 [link]. [g] The Armed Conflict Location and Event Data Project (ACLED) is described in Raleigh et al. (2010) [h] The precipitation data are produced by the Global Precipitation Climatology Centre (GPCC) and described in Schneider et al. (2015). Date of access: April 28, 2021 [link]. 26 Table A2: Tests of sample selection Analysis sample Left-out sample Difference Joint Test Variable N Mean Std. N Mean Std. Diff Std. F-stat. Err. Err. Err. Education group 1.502 Primary or below (dummy) 468 0.241 0.020 114 0.167 0.035 0.075 0.044 Secondary (dummy) 468 0.506 0.023 114 0.544 0.047 -0.037 0.052 Tertiary (dummy) 468 0.252 0.020 114 0.289 0.043 -0.037 0.046 Age group 3.067 15-24 years (dummy) 468 0.038 0.009 118 0.102 0.028 -0.063 0.023 20-24 years (dummy) 468 0.327 0.022 118 0.144 0.325 0.183 0.019 25-29 years (dummy) 468 0.368 0.022 118 0.432 0.046 -0.065 0.050 30-34 years (dummy) 468 0.162 0.017 118 0.169 0.035 -0.007 0.038 35-39 years (dummy) 468 0.058 0.011 118 0.093 0.027 -0.036 0.025 40-44 years (dummy) 468 0.019 0.006 118 0.034 0.017 -0.015 0.015 45-49 years (dummy) 468 0.009 0.004 118 0.008 0.008 0.000 0.009 50-54 years (dummy) 468 0.004 0.003 118 0.000 0.000 0.004 0.006 55-59 years (dummy) 468 0.000 0.000 118 0.008 0.008 -0.008 0.004 60 or more years (dummy) 468 0.015 0.006 118 0.008 0.008 0.006 0.012 Occupation 0.715 Employer/Manager (dummy) 459 0.007 0.004 128 0.008 0.008 -0.001 0.008 Professional (dummy) 459 0.220 0.019 128 0.227 0.037 -0.007 0.042 Manual worker (dummy) 459 0.126 0.016 128 0.180 0.034 -0.053 0.034 Agricultural worker (dummy) 459 0.031 0.008 128 0.016 0.011 0.015 0.016 Military and security (dummy) 459 0.002 0.002 128 0.008 0.008 -0.006 0.006 Owner of a shop (dummy) 459 0.039 0.009 128 0.039 0.017 0.000 0.019 Government employee (dummy) 459 0.007 0.004 128 0.000 0.000 0.007 0.007 Private sector employee (dummy) 459 0.255 0.020 128 0.250 0.038 0.005 0.044 Craftsperson (dummy) 459 0.102 0.014 128 0.055 0.020 0.048 0.029 Illegal (dummy) 459 0.002 0.002 128 0.000 0.000 0.002 0.004 Student (dummy) 459 0.129 0.016 128 0.156 0.032 -0.028 0.034 Retired (dummy) 459 0.002 0.002 128 0.000 0.000 0.002 0.004 No work (dummy) 459 0.078 0.013 128 0.063 0.021 0.016 0.026 Civil Status 1.538 Divorced (dummy) 391 0.010 0.005 123 0.000 0.000 0.010 0.009 Married (dummy) 391 0.179 0.019 123 0.236 0.038 -0.057 0.041 Single (dummy) 391 0.811 0.020 123 0.764 0.038 0.047 0.041 Religious knowledge 1.230 Basic (dummy) 407 0.705 0.023 113 0.779 0.039 -0.074 0.048 Intermediary (dummy) 407 0.265 0.022 113 0.204 0.038 0.062 0.046 Advanced (dummy) 407 0.029 0.008 113 0.018 0.012 0.012 0.017 Experience with Jihad 0.161 No (dummy) 432 0.951 0.010 125 0.960 0.018 -0.009 0.021 Yes (dummy) 432 0.049 0.010 125 0.040 0.018 0.009 0.021 Notes: Variables are constructed from self-reported information provided by recruits that is available in the Daesh personnel records database. The “analysis” sample consists of observations included in the analysis; the “left-out” sample consists of observations dropped because of missing information on either delegation of residence, education, or age. In the last column, for each group of variables, the F-statistic reports the result of a test of joint significance of all variables in that group. 27 Table A3: Alternative transformations of the dependent variable Dependent variable: Number of Daesh recruits ihs linear log (1+...) dummy (1) (2) (3) (4) Panel A: OLS with delegation FE Internet access (rate) 0.072 0.121 0.056 0.048 (0.028) (0.047) (0.022) (0.021) N 7878 7878 7878 7878 R2 0.228 0.207 0.228 0.216 Panel B: 2SLS second stage Internet access (rate) 0.318 0.555 0.247 0.217 (0.159) (0.264) (0.123) (0.123) N 7878 7878 7878 7878 Kleibergen-Paap test 66.183 66.183 66.183 66.183 Notes: Standard errors are reported in parentheses and are corrected for spatial correlation following Conley (1999) using a 50 km distance cutoff. All specifications control for population (log), unemployment (rate), married (rate), area (log), 1995-2010 nighttime lights (log), and 1995-2010 rainfall (log). Panel A presents OLS estimates, which also account for delegation FE. Panel B presents 2SLS estimates that use lightning strikes (log) as an instrument for internet access (rate). Column 1 replicates the baseline result in Table 2 with the inverse hyperbolic sine transformation of the number of Daesh recruits as dependent variable. Column 2 simply uses the number of Daesh recruits as dependent variable. Column 3 takes the logarithm of 1 + the number of Daesh recruits as dependent variable. Column 4 uses as dependent variable a dummy variable that assumes value 1 if the number of Daesh recruits is strictly positive, and 0 otherwise. Appendix Table A1 provides complete definitions of variables and Table 1 displays summary statistics. 28