• Non ci sono risultati.

Hidden attraction to radical Islam and terrorist attacks

N/A
N/A
Protected

Academic year: 2022

Condividi "Hidden attraction to radical Islam and terrorist attacks"

Copied!
41
0
0

Testo completo

(1)

Hidden attraction to radical Islam and terrorist attacks

Dean Fantazzini

, Marcella Nicolini

, Fabio Sabatini

Ÿ

May 30, 2020

Abstract

We nd a positive relationship between terrorist attacks and the public's interest in radical Islam and violent extremism. Our worldwide dynamic panel analysis over the period 2004 - 2015 shows that attacks perpetrated in the name of Islam correlate with online searches for a set of keywords that lead to radicalizing propaganda on the Internet.

Attacks are more strongly associated with searches for terms that evoke violent actions, such as beheadings, or explicitly related to the jihad, likely revealing attraction to the terrorists' message.

Keywords: Islamist extremism; terrorism; Google searches.

JEL Classication: F50; H56.

We thank Elena Esposito, Domenico Giannone, Robert Hefner, Daniele Massacci, Jeremy Menchik, Paolo Pinotti, Andrea Pozzi and seminar participants at the META 2018 workshop in Nizhny Novgorod for useful comments and suggestions. Luca Biglieri, Emma Cambieri, Antonella Cozzolino, Simona Fabris, Emanuela Padellaro, Vittoria Senzalari, and Vittorio Spinelli provided excellent research assistance. Marcella Nicolini thanks the Institute for the Study of Muslim Societies & Civilizations at Boston University for hospitality.

All errors and omissions are our own.

Moscow State University, Moscow School of Economics. Email: fantazzini@mse-msu.ru.

Corresponding author. University of Pavia, Department of Economics and Management. Email: mar- cella.nicolini@unipv.it.

ŸSapienza University of Rome, Department of Economics and Law, and IZA. Email:

fabio.sabatini@uniroma1.it

(2)

1 Introduction

Terrorists seek the public's attention to gain new followers and proselytize, eventually improving their harming capacity. By keeping people focused on terrorist claims, the media coverage of the attacks unintendedly encourages the planning of new missions (Jetter, 2017;

2019). Detecting the public's interest in violent stances and understanding to what extent it relates to attacks can help to monitor radicalization and prevent its escalation into violence.

However, measuring the celebrity of terrorists and the appeal of their message is challenging for several reasons. The supply of terrorism-related news does neither capture the active in- terest of the public nor its views, which widely dier with the bias of news sources (Gentzkow and Shapiro, 2004). Curiosity and support for extremist ideas tend to be taboo and under- stated, thereby frustrating any attempt of detection through survey data and experimental methods (Stephens-Davidowitz, 2014).

To uncover the attraction to violent extremism and study its relationship with terrorist attacks, we exploit online searches for a set of sensitive keywords as a measure of the interest in radical Islam and violent extremism. We pick 47 sensitive keywords and collect the monthly volume of their Google searches in all the countries that faced a terrorist attack from 2004 to 2015, according to the Global Terrorism Database (GTD). The set includes terms denoting curiosity for radical Islam and keywords implying a sympathetic attitude towards terrorists' claims, or explicitly evoking violent actions typical of the jihad such as the killing of indels.

As the GTD does not reveal the background of the attacks, we make an additional eort to document the origins and motivation of the terrorist groups involved in each incident. The

nal database contains detailed information about 21,710 attacks perpetrated in the name of Islam by self-proclaiming Islamist groups worldwide.

By exploiting the panel dimension of the data, we show that online searches for sensi- tive keywords are positively correlated with Islamist attacks. The association is robust to controlling for cross-country correlations and survives several robustness checks, including placebo tests and instrumental variables estimates. When we distinguish between keywords,

(3)

we nd that attacks are more signicantly and sizably associated with keywords recalling violence, such as beheadings, or explicitly related to the jihad, such as Abdullah Azzam, a Sunni scholar and founding member of Al Qaeda, also known as the Father of Global Jihad

(Edwards, 2017).

The limited variability of the instruments and possible exceptions to the assumption of their exogeneity suggests caution in interpreting our results as causal. Despite this limitation, our analysis improves the understanding of the relationship between terrorism and the active interest in violent extremism. Previous studies suggest that the media coverage of the attacks feeds violent extremism as the attention of the public encourages terrorists to schedule new actions. Our approach oers a way to uncover the public's attention to extremist ideas and documents its relationship with terrorists' missions. The link between the occurrence of the attacks and searches likely capturing support for the jihad and attraction to violence is striking. This nding points to the possibility that the attention of the public to radical Islam and violent extremism reveals a substratum of public opinion potentially fertile for radicalization, which is impossible to detect with survey data and controlled experiments and has implications for the prevention of terrorism.

This paper bridges two strands of literature. The rst strand deals with the relationship between media and terrorism (Gentzkow and Shapiro, 2004; Rohner and Frey, 2007; Jetter, 2017; 2019; El Ouadghiri and Peillex, 2018; Jetter and Walker, 2018). We contribute to this

eld in two substantive ways. We use Google searches to overcome the diculty of detecting the celebrity and popularity of terrorists, and we show that the public's active attention to terrorists' claims robustly correlates with the attacks. Our eort is more in general related to the abundant literature studying the sources and the outcomes of terrorism, including poverty (Abadie, 2006), trade restrictions (Amodio et al., 2020), , inequality (Benmelech et al., 2012), deterrence (Jaeger et al., 2012) and retaliation between opposing sides (Jaeger and Paserman, 2008), well-being (Metcalfe et al.; Clark et al., 2020), and the assimilation of Muslim immigrants (Gould and Klor, 2016).

(4)

The second strand encompasses studies using online searches to track economic behaviors (Baker and Fradkin, 2017; Berger et al., 2018), the spreading of viral diseases (Pichler and Ziebarth, 2017, Lampos et al., 2020), and hidden social phenomena such as racism (Doleac and Stein, 2013; Stephens-Davidowitz, 2014; Giulietti et al., 2019), contraceptive use and abortion (Kearney and Levine, 2015), religiosity (Bentzen, 2019), interest in Brexit leave (Born et al., 2019), and prosocial attitudes (Guriev and Melnikov, 2016). We connect the two strands by introducing the use of online searches in the study of radicalization and terrorism.

The rest of the paper proceeds as follows: in the next section, we describe our data on online searches of sensitive keywords and terrorist episodes. We then explain our econometric approach in Section 3. Section 4 presents the analysis of the relationship between the online interest in Islamist extremism and terrorist attacks. Section 5 provides some robustness checks. In Section 6, we briey summarize and discuss our main ndings. Section 7 concludes.

2 Data

2.1 Measuring interest in radical Islam and violent extremism

As of January 2019, almost 4.4 billion people were active Internet users, roughly correspond- ing to 59% of the global population.1 With around 90% market share worldwide between 2010 and 2015, Google is by far the world's leading search engine.2 Starting from January 2004, the company has made available information on the queries that people search on the engine with a tool called Google Trends (hereafter GT).3 For any specic keyword, GT provide an index of the volume of queries entered into Google in a given geographic area. The query index is the ratio between the total query volume for the keyword in question over a certain period and the total number of queries in the same period. At the country level, Google pro- vides search trends with a monthly frequency. In any selected period, the maximum query

1See www.statista.com.

2Data sourced from www.statista.com/statistics/216573/worldwide-market-share-of-search-engines/

3GT can be inspected at https://trends.google.com/trends/.

(5)

share is normalized to be 100, and the query share at the initial date is normalized to be zero (Choi and Varian, 2012). By aggregating millions of searches with a monthly frequency over a long period across a large sample of countries, GT allow constructing reliable indicators of a broad array of behavioral variables.

Empirical evidence shows that GT consistently and strongly detect many phenomena for which conventional data are publicly available with a delay, or at a lower frequency.

For example, queries consistently detect economic behaviors such as job searches (Baker and Fradkin, 2017), the receipt of food stamps (Fantazzini, 2014), the issuance of bad checks (Eksi et al., 2017), deposit reallocation (Fecht et al., 2019), car sales for specic brands (Fantazzini and Toktamysova, 2015), and the use of Uber services (Berger et al., 2018), to name just a few. With some caution, online searches have also proven to be useful in tracking the spreading of viral diseases like H1N1 (Aguero and Beleche, 2017), Covid-19 (Lampos et al., 2020), and other u-like illnesses (Pichler and Ziebarth, 2017).

However, it is in the detection of usually undercover interests and behaviors that GT play a crucial role. Online searches are generally conducted in private, elude social control, and allow people to express socially taboo, anti-social, or even criminal interests and attitudes that they would hardly reveal in public. For example, MacInnis and Hodson (2015) use GT to study how interest in pornography relates to religiosity and conservatism across American states.

Google-based measures of racism strongly predict black mortality (Chae et al., 2015), black homeownership (Harris and Yelowitz, 2018), access to public services (Giulietti et al., 2019), and the performance of Barak Obama in 2008 and 2012 U.S. presidential elections (Stephens- Davidowitz, 2014). Google data have also proved useful to detect or predict contraceptive use and abortion (Kearney and Levine, 2015), suicides (Ma-Kellams et al., 2015), drug abuse (Perdue et al., 2018), and the propagation of sexually-transmitted diseases such as syphilis (Young et al., 2018) and HIV (Young and Zhang, 2018).

The spreading of curiosity and sympathy for the claims of terrorists is a typically hidden social pattern that Google searches can help uncover. Research into the relationship between

(6)

the Internet and terrorism has so far focused on the role of social networking sites (Weimann, 2015; Aly et al., 2016). Tracking jihadi propaganda on social media is useful for detecting ex- tremists' communication and enacting counter-terrorism measures (Zeitzo, 2017). However, radicalized individuals' activities on platforms like Facebook and Twitter do not necessarily capture the attention of the general public in radical Islam and violent extremism. People likely radicalize as a result of the spreading of feelings sympathetic to extremists across soci- ety (Koomen and Van Der Pligt, 2016), a phenomenon that eludes detection through social media (Ahmed and Lloyd George, 2016). Moreover, jihadi propaganda and recruitment have progressively moved to the dark web (Weimann, 2016), which cannot be easily accessed by ordinary Internet users. Search engines, by contrast, can help to detect the active attention of the public to extremist ideas (Jetter, 2019).

In our empirical analysis, we use a list of keywords that, according to the Centre on Religion & Geopolitics (CRG) at the Tony Blair Foundation, are likely to lead to extremist content if entered into a search engine. The CRG rst developed a set of words and phrases that might be used in research on radical Islam or in deliberate attempts to access violent extremist content. The list was then employed as a basis to ascertain further potentially associated keywords through specic tools such as Google Keyword Planner, Google Suggest phrases, and Google-related searches. The nal list comprised 143 English keywords and the corresponding Arabic translations.

Based on the search frequency data and a qualitative assessment, the CRG then selected a set of 47 keywords. The set is the sum of two subgroups. The rst includes popular keywords that were searched on average at a minimum frequency of 500 times per month. The second group includes keywords explicitly related to violent extremism. Ahmed and Lloyd George (2016) conducted a Search Engine Result Page (SERP) analysis on the keywords to ascertain the extent to which they return extreme content. Though being non-violent and mostly related to political Islam, keywords belonging to the rst group return extremist content in a remarkable share of cases, anticipating the risk of unintentional exposure to jihadi propa-

(7)

ganda. The second set comprises keywords that, though being less popular, carry a higher potential for risk. Overall, 44% of the material was explicitly violent. In contrast, counter- narrative content outperformed extremist material in only 11% of the results generated.4

Several sites aiming to present legitimate Islamic culture were also found to host extremist material. For example, Kalamullah.com and WorldofIslam.info - which can be retrieved by entering our selection of keywords into a search engine - contain resources for ordinary Muslims and those interested in Islam, such as online versions of the Quran, collections of Hadith literature, and a variety of Islamic books.5 However, these sites also provide jihadi manuals and audio lectures by known jihadis. Overall, the Google searches for the chosen set of keywords returned a vast array of fundamentalist material spanning from non-violent extremism to explicitly violent jihadism.

We use the GT for the 47 English keywords to measure the active interest in Islamist fundamentalism in 149 countries over the period 2004-2015. Figure 1 shows on a map the volume of searches for the English keywords worldwide.

4Ahmed and Lloyd George (2016) looked at the rst two pages of results, as 92% of Google trac is on page one, with page two only receiving 5% of the trac and page three 1% (Chitika Insights, 2013). The authors split the extreme content into three categories: violent, non-violent, and political Islamist. Websites were categorized as violent if they contained either images of graphic violence or exhortations to violence.

The political Islamist category included content expressing a specic anity to a particular Islamist group.

The non-extreme content regarded counter-narrative or neutral websites, such as news websites or portals targeting researchers and students. Websites deemed to be extreme but non-violent were those expressing anti-Semitic, homophobic, racist, or sectarian views without a depiction of, or incitement to, violence.

5From March 2020, these websites addressed the Islamic public by also providing a collection of prayers invoking divine protection from the Covid-19 pandemic.

(8)

Figure 1: Google Trends monthly averages for the mean of the 47 search terms

In line with Ahmed and Lloyd George (2016), we also collect the GT of the Arabic translation of the 47 keywords. This approach does not want to confuse the Arab world with Islamist fundamentalism, nor it claims that terrorism is an Arab, or Islamic, phenomenon.

The reason for also exploring Google searches in Arabic lies in the fact that Islamist terrorists' codewords are often in Arabic, given the inherent relationship of this language with Islam and the Quran. Figure 2 reports data on the volume of searches for the Arabic keywords for 149 countries globally. As evident, searches for sensitive keywords do not correspond exclusively to Arab countries. Instead, the global keyword frequencies in Arabic include searches by Arabic-speaking Internet users worldwide.

A variety of people may be searching for keywords that return extremist content, from academics to journalists. We do not claim to be measuring the online search activity of perspective terrorists. Instead, we assume that Google searches measure the interest of the public in radical Islam and violent extremism, which may reveal the development of a sub- stratum of public opinion potentially fertile for radicalization. Previous studies employing Google searches to assess the hypothetical outcomes of social phenomena all share the as- sumption that trends for specic keywords can proxy the incidence of hidden social patterns in the population (e.g., Stephens-Davidowitz, 2014; Kearney and Levine, 2015; Guriev and Melnikov, 2016).

(9)

Figure 2: Google Trends monthly averages for the mean of the 47 search terms in Arabic

Table 1 reports the list of keywords. We provide a glossary with a brief description of their meaning in Appendix A.1.

2.2 Terrorist attacks

Data on terrorist attacks are drawn from the Global Terrorism Database (GTD) maintained by the National Consortium for the Study of Terrorism and Responses to Terrorism (START) at the University of Maryland. The database collects information on the attacks perpetrated around the world from 1970 through 2018. Including more than 190,000 cases over the whole period, and 80,694 over our period of analysis (2004-2015), the GTD is the most com- prehensive unclassied database on terrorist attacks.6 Dierently from other databases on terrorism, the GTD includes systematic data on domestic as well as transnational and inter- national terrorist incidents. For each attack, the database provides information on several variables, among which the date and location, the weapons used, the nature of the target, the number of casualties (distinguishing victims into deaths and wounded), and, when available, the group or individual responsible for the attack.

We make an additional eort to collect further information on the 999 terrorist groups

6See https://www.start.umd.edu/gtd/about/ and LaFree and Dugan (2007) for more information on the database.

(10)

Table 1: List of the 47 keywords

Most popular Most extreme

Apostasy Abdullah Azzam

Apostate Amaq Agency

Apostates Apostate Islam

Ayman al-Zawahiri Apostates in Islam

Caliphate Beheadings

Crusader Crusader Army

Crusaders Crusaders against Islam

Dabiq Dabiq Pdf

Dabiq Magazine How to do Jihad

Ibn Taymiyyah Ibn Taymiyyah Jihad

Islamic State Inspire Magazine

Jihad Jewish Coalition

Jihad meaning Jihad for Ummah

Kar Jihad in the Quran

Khalifah Meaning Khalifah

Khilafah Khilafah Syria

Kuar Killing Apostates

Martyr Killing Indels

Martyrdom in Islam Killing Kuar

Martyrs Mujahid

Mujahideen Preparing for Jihad

Shahada Radah

Suicide vest Soldiers of the Caliphate

Taghut

(11)

identied in the GTD that have been active in the 2004-2015 period. We classify groups according to their ideology or intent. Inspecting each terrorist event allows us to enrich the GTD database by further categorizing the attacks according to the ideology, religious background, and motivation of the perpetrating group. Out of 80,694 events, it is possible to identify the group's background in 34,907 cases (43.3% of cases). Religious fundamentalist organizations have caused two-thirds of these events, and 97.8% of the attacks with a religious identity have been perpetrated by self-proclaiming Islamist groups, corresponding to 21,711 registered attacks. We document the distribution of terrorist attacks with identied authors across religions in Table A.2 in Appendix A.3.

Figure 3 reports the incidence of Islamist attacks. They are widespread globally: more than 60 countries in the world have been hit at least once in the 2004-2015 period. However, if we consider countries with a high incidence of this phenomenon, we see that only a few countries are severely aected.

Figure 3: Average monthly islamist attacks

3 Econometric approach

We collect data on all countries that faced at least one terrorist attack of any kind over the period 2004-2015. The sample includes 149 countries, with data collected at monthly

(12)

frequency. However, given that there were no Islamist attacks in many countries, as evident from Figure 3, we present results both in the full sample and in the subset of 63 countries in which at least one Islamist attack took place. Our panel is characterized by a wide cross- sectional dimension, but also a long time span, thanks to the availability of monthly data for 12 years (t = 144). To investigate the relationship between online searches for sensitive keywords and the outcomes of Islamist terrorism, we regress the attacks on the GT of those keywords.

While the maps discussed in Section 2 present a static picture of the spatial distribution of the Islamist attacks and the GT, we want to investigate the presence of cross-sectional dependence over time.7 Such correlation may arise from a globally common shock that has a heterogeneous impact across countries, such as the 9/11 attack to the Twin Towers. Or it can be the result of local spillover eects between countries. Searches for sensitive keywords might increase not only in the country hit by a terrorist attack but also in the neighboring countries, where the media are likely to give wide coverage to terrorists, thereby involving the attention of the public.8 The cross-section dependence (CD) test statistic (Pesaran, 2004) shows a positive cross-country correlation of 0.190 for Islamist attacks and 0.228 for GT: the econometric specication will take into account these patterns. Inspecting the stationarity properties of our variables of interest suggests that they are not trending over time, thus ruling out a spurious correlation driven by trends in our data.9 Finally, the Wooldridge (2002) test for serial correlation in panel data reveals the presence of rst order autocorrelation in the residuals.

These features of our data lead us to choose the dynamic common correlated eects model (DCCE) by Chudik and Pesaran (2015), which is a dynamic model that also takes

7Cross-sectional dependence in macro panel data has received growing attention in the panel time series literature. See Baltagi (2005) for an introduction.

8See for example: https://www.bbc.com/news/world-europe-38415287

9We implement two panel unit root tests on our data: a Fisher-type unit root test, which combines the p-values from independent Augmented Dickey and Fuller (1979) (ADF) tests on the single series and the Im-Pesaran-Shin (IPS) test (Im et al., 2003). Results are conrmed with the Pesaran (2007) panel unit root test in the presence of cross-section dependence.

(13)

into account the presence of cross-sectional dependence. We thus estimate the following model:

IAct = αc+ λcIAct−1+ βc0GTct+X

l

δ0clIAt−1+X

l

θ0clGTt+ εct (1) εct = γc0ft+ uct

where IAct is the number of Islamist attacks perpetrated in country c at time t; αc are the country xed eects; IAct−1 is the lagged value of IA; GTct is an aggregate measure of Google searches, obtained as the mean of the 47 English keywords search volumes in country cat time t; IAt−1 and GTt and their lags l are cross-section averages included to account for cross-sectional dependence. The error term εct is composed of an unobserved common factor ft, with heterogeneous factor loading γc, and an independent and identically distributed (IID) component uct. The heterogeneous coecients βc are randomly distributed around a common mean. The DCCE runs separate regressions for each country. The coecients of interest are obtained as the unweighted means of the individual coecients: stacking λcand βc in πc = (λc, βc), we can write that the DCCE coecients are πDCCE = N1 PN

c=1πc, where N is the number of countries (either 149 or 63, in our setting). Note that the estimated coecients for the cross-section averages δcl and θcl are not interpretable in a meaningful way: they merely blend out the biasing impact of the unobservable common factor and are therefore not reported in the subsequent analysis.

We delve deep into our results by considering alternative measures of GT s. First, we use the rst factor of a principal component analysis (PCA) of the 47 keywords.10 Then, we check whether our results dier considering the Google searches for the keywords in Arabic.

We then consider dierent sub aggregations of keywords. A rst natural classication is the one into popular and extreme keywords, as proposed by Ahmed and Lloyd George (2016). We add to this distinction several alternative aggregations to capture dierent aspects

10PCA is used to reduce the number of variables in an analysis: it provides a set of uncorrelated linear combinations of the variables that contain most of the variance. We consider here only the rst component.

(14)

of the online searches for Islamist and extremist content. A further avenue to deepen our analysis focuses on the outcomes of the attacks. We present evidence on the relationship between Google searches for sensitive keywords and some measures of the damage caused by terrorists' missions.

In the robustness section, we provide several placebo tests to control further for the possibility that our results capture a spurious correlation, either working on the dependent or the explanatory variables. First, we consider the overall number of attacks, irrespective of the religious or ideological motivation behind them. We expect the latter to be poorly or not at all correlated to the GT for our keywords. Then, we follow D'Amuri and Marcucci (2017) and use a neutral word, such as dos, and check its relationship with Islamist attacks.

Lastly, we explore which keywords are highly correlated with our target variable, as provided by Google Correlate. Such placebo keywords do not exhibit any meaningful relationship with Islamist attacks.

As an additional robustness check, we discuss the possibility of instrumenting Google searches. Following Jetter (2019), we assume that natural disasters can monopolize the attention of the media and the interest of the public, thereby lowering the relative query volume for our sensitive keywords. This strategy relies on the assumption that disasters aect terrorist attacks exclusively through the reaction they prompt in the general public.

The analysis provides encouraging results but has some limitations that we discuss in detail in Section 5.2.

4 Results

We present our main results using alternative measures of the volume of Google searches in Table 2.11 Column 1 in Panel A reports our preferred specication, with the mean value of

11Diagnostic tests support the choice of the DCCE estimator. The combined evidence from the Fisher- type and the IPS panel unit root tests conrms that the series are stationary. The Pesaran (2004) test for cross-sectional dependence suggests signicant cross-sectional correlations for all our variables. The Pesaran (2007) panel unit root test in the presence of cross-sectional dependence also rejects the null hypothesis of

(15)

the 47 keywords in English. In column 2, we present the results with the rst factor of a principal component analysis (PCA) of the 47 dierent keywords. We observe a positive and statistically signicant coecient for GT . The lagged value of the dependent variable, IA, also displays a positive and signicant coecient, as expected.

To delve more in depth into the relationship between Google searches and the attacks, we distinguish between keywords. First, we adopt the classication between popular and

extreme keywords proposed by Ahmed and Lloyd George (2016), which is reported in Table 1. We compute the average volume of Google searches for the two groups. Both popular and

extreme keywords signicantly and positively relate to the attacks, with extreme keywords displaying a larger coecient. We then check whether our results dier considering the online searches for the 47 keywords in Arabic. Results reported in columns 5-8 of Table 2 show that searches in Arabic are always not statistically signicant.12

homogenous non-stationarity. For a discussion on alternative econometric choices and the robustness of our

ndings across dierent estimators, see the Econometric Appendix A.2 and the results thereby reported.

12For the subsequent analysis, we will not report estimates considering GT in Arabic as results are sys- tematically not signicant. Results are available upon request.

(16)

Table2:AlternativedenitionsofGTs PanelA(1)(2)(3)(4)(5)(6)(7)(8) CountrieshitbyEnglishArabic anyterroristattackMeanFactorPopularExtremeMeanFactorPopularExtreme GTt0.048**0.083***0.020**0.031**-0.004-0.0030.010-0.029 (0.019)(0.031)(0.009)(0.012)(0.022)(0.021)(0.012)(0.023) IAt−10.062***0.062***0.062***0.061***0.052***0.052***0.054***0.057*** (0.012)(0.012)(0.012)(0.013)(0.012)(0.012)(0.012)(0.012) R20.8260.8180.8280.8030.8220.8200.8230.802 R2(MG)0.8250.8230.8250.8180.8250.8250.8260.821 #Observations21,15821,15821,15821,15821,15821,15821,15821,158 #Countries149149149149149149149149 PanelB(1)(2)(3)(4)(5)(6)(7)(8) CountrieshitbyEnglishArabic IslamistattacksonlyMeanFactorPopularExtremeMeanFactorPopularExtreme GTt0.108***0.187***0.042**0.080***-0.013-0.0130.022-0.075 (0.041)(0.069)(0.019)(0.030)(0.054)(0.050)(0.028)(0.059) IAt−10.145***0.147***0.145***0.144***0.125***0.126***0.129***0.134*** (0.025)(0.025)(0.025)(0.026)(0.025)(0.025)(0.025)(0.026) R20.8260.8140.8290.8010.8190.8180.8200.802 R2(MG)0.8250.8230.8250.8180.8250.8250.8250.821 #Observations8,9468,9468,9468,9468,9468,9468,9468,946 #Countries6363636363636363 Notes:DCCEestimates.Thedependentvariableisthenumberofislamistattacks.Meanistheunweightedmean oftheGTsforthe47keywords,FactoristherstfactorfromaPCAonthe47keywords.ForthePopularand Extremegroupings,seeTable1.Standarderrorsinparentheses,***,**,and*denote,respectively,10%,5%,and 1%signicancelevels.

(17)

Panel B of Table 2 shows the results on the subsample of 63 countries that experienced at least one Islamist attack in the period under consideration. Across dierent specica- tions, results reported in Panel B conrm those on the full sample: the sign and statistical signicance of coecients do not change, but we notice a slight increase in their size. As countries that did not face attacks perpetrated in the name of Islam are now excluded from the sample, the specication in Panel B likely captures the relationship between the interest of the public to Islamist extremism and Islamist attacks more neatly. Not surprisingly, the estimated coecients are slightly larger in this subsample.

To better understand the positive correlation between Google searches and terrorist at- tacks perpetrated by self-proclaiming Islamist organizations, we distinguish the keywords into six groups according to their meaning. The rst two categories include keywords referring to the ISIS and Al Qaeda, respectively. In the third group, we focus on apostasy: in radical Islam, apostasy is not necessarily understood as the personal choice of renouncing a religious belief. Instead, it is often considered as a public act of political secession from the Muslim community. This view makes apostates one of the favorite targets of jihadi propaganda (Tor- res Soriano, 2010). The fourth group comprises keywords evoking violent actions such as

beheadings. The fth and sixth categories contain terms related to religious doctrines and the history of political Islam, respectively.13 For each group, we compute the average volume of Google searches and assess its relationship with terrorist attacks perpetrated in the name of Islam.

Panel A of Table 3 reports results in all countries that faced an attack. The sets of keywords related to the two most prominent Islamist terrorist groups, Al Qaeda and ISIS, are not signicant.14 Queries referring to the concept of apostasy in Islam, such as kar and

kuar, signicantly and positively correlate with the attacks. The index of violent keywords, including queries like beheadings and killing indels, signicantly and positively relates to the attacks. The aggregations of keywords comprising the doctrinal and historical terms are

13Table A.3 in the Appendix denes the six categories in detail.

14Results are unaected if we exclude the term Caliphate from the ISIS index.

(18)

instead not statistically signicant.15 Panel B of Table 3 reports the results on the subsample of countries aected by Islamist attacks in the period under consideration. Again we observe an increase in the size of the estimated coecients.

Table 3: Alternative aggregations of GTs

Panel A (1) (2) (3) (4) (5) (6)

Countries hit by

any terrorist attack ISIS Al Qaeda Indels Violent Religion History

GTt -0.002 0.002 0.005* 0.035*** 0.006 0.002

(0.006) (0.005) (0.003) (0.014) (0.006) (0.002) IAt−1 0.060*** 0.065*** 0.063*** 0.062*** 0.064*** 0.061***

(0.012) (0.013) (0.013) (0.012) (0.013) (0.012)

R2 0.815 0.793 0.794 0.814 0.790 0.825

R2 (MG) 0.823 0.815 0.814 0.820 0.817 0.824

# Observations 21,158 21,158 21,158 21,158 21,158 21,158

# Countries 149 149 149 149 149 149

Panel A (1) (2) (3) (4) (5) (6)

Countries hit by

Islamist attacks only ISIS Al Qaeda Indels Violent Religion History

GT -0.009 0.004 0.011* 0.088** 0.012 0.002

(0.013) (0.012) (0.006) (0.035) (0.012) (0.005) IAt−1 0.140*** 0.154*** 0.149*** 0.149*** 0.150*** 0.142***

(0.025) (0.027) (0.027) (0.026) (0.027) (0.026)

R2 0.820 0.794 0.798 0.809 0.796 0.826

R2 (MG) 0.825 0.815 0.816 0.819 0.817 0.822

# Observations 8,946 8,946 8,946 8,946 8,946 8,946

# Countries 63 63 63 63 63 63

Notes: DCCE estimates. The dependent variable is the number of islamist attacks. For the denition of the dierent groupings, see Table A.3. Standard errors in parentheses,

***, **, and * denote, respectively, 10%, 5%, and 1% signicance levels.

Finally, we explore how Google searches for our sensitive keywords relate to indicators of

15The doctrinal terms are not specic to Islam and could generically refer also to other Abrahamic religions, such as Christianity. Results do not change if we extend the Religion index by including Shahada, a doctrinal term specic to Islam.

(19)

the outcomes of terrorist missions. To this end, we repeat the analysis using the number of killed, wounded, and overall aected people as dependent variables.16 We report the results in Table 4: GT for the English keywords are positive and signicant on the full set of countries (Panel A) and in the subset of countries aected by an Islamist attack (Panel B).

Table 4: Robustness with dierent measures of violence

Panel A (1) (2) (3)

Countries hit by

any terrorist attack y=killed y=wounded y=victims

GTt 0.014 0.014* 0.019*

(0.009) (0.008) (0.011)

yt−1 0.053*** 0.047*** 0.058***

(0.012) (0.011) (0.012)

R2 0.881 0.907 0.880

R2 (MG) 0.543 0.418 0.488

# Observations 21,158 21,158 21,158

# Countries 149 149 149

Panel B (1) (2) (3)

Countries hit by

Islamist attacks only y=killed y=wounded y=victims

GT 0.034 0.034* 0.045*

(0.021) (0.018) (0.026)

yt−1 0.124*** 0.112*** 0.136***

(0.025) (0.023) (0.026)

R2 0.880 0.907 0.879

R2 (MG) 0.542 0.418 0.488

# Observations 8,946 8,946 8,946

# Countries 63 63 63

Notes: DCCE estimates. The dependent variable is respec- tively, the log of killed persons, the log of wounded per- sons, the log of total victims (equal to the sum of killed and wounded). Standard errors in parentheses, ***, **, and * denote, respectively, 10%, 5%, and 1% signicance levels.

16We compute the overall number of aected individuals as the sum of the killed and injured people.

(20)

5 Robustness

5.1 Placebo tests

To further check the robustness of our results, we discuss two placebo tests for the main

ndings reported in column 1 of Table 2. First, we test whether Google searches are signif- icantly and positively correlated with non-Islamist attacks. In Columns 1 and 2 of Table 5, we report results using all kinds of attack as dependent variable, including those perpetrated by non-Islamist terrorists. In columns 3 and 4, we remove Islamist attacks from the pool.

We see that the attacks are never aected by searches for our sensitive keywords, neither in English nor in Arabic.

As a second exercise, we consider a term that is completely unrelated to the Islamist attacks. We follow D'Amuri and Marcucci (2017) and choose the term dos, which is an acronym for Disk Operating System.17 We expect Google searches for the term dos to be unrelated to Islamist attacks. As reported in column 5 of Table 5, the coecient is not statistically signicant. We also repeat the exercise with the GT for google nding again that it is not correlated with the occurrence of the attacks. Panel B shows that the results of the placebo tests do not change when only considering the subsample of countries aected by Islamist attacks.

17It also stands for Department of State in the U.S., Demokratska opozicija Srbije, a political party in Serbia, and the IATA code for the Dios airport in Papua New Guinea.

(21)

Table 5: Placebo tests

Panel A (1) (2) (3) (4) (5) (6)

Countries hit by

any terrorist attack y=all attacks y=non-Islamist attacks y=Islamist attacks

GTEnglisht 0.007 -0.013

(0.063) (0.054)

GTArabict 0.095 0.077

(0.065) (0.061)

dost 0.001

(0.001)

googlet -0.047

(0.106)

yt−1 0.138*** 0.132*** 0.138*** 0.132*** 0.059*** 0.059***

(0.019) (0.019) (0.020) (0.020) (0.013) (0.012)

R2 0.663 0.675 0.671 0.674 0.815 0.794

R2 (MG) 0.822 0.824 0.777 0.780 0.818 0.825

# Observations 21,158 21,158 21,158 21,158 21,158 21,158

# Countries 149 149 149 149 149 149

Panel B (1) (2) (3) (4) (5) (6)

Countries hit by

Islamist attacks only y=all attacks y=non-Islamist attacks y=Islamist attacks

GTEnglisht 0.066 0.015

(0.141) (0.118)

GTArabict 0.228 0.196

(0.147) (0.137)

dost -0.001

(0.002)

googlet -0.135

(0.264)

yt−1 0.254*** 0.244*** 0.254*** 0.243*** 0.140*** 0.138***

(0.031) (0.031) (0.033) (0.032) (0.026) (0.025)

R2 0.663 0.671 0.671 0.669 0.820 0.795

R2 (MG) 0.825 0.827 0.783 0.785 0.818 0.825

# Observations 8,946 8,946 8,946 8,946 8,946 8,946

# Countries 63 63 63 63 63 63

Notes: DCCE estimates. Columns 1-4 present placebo tests adopting alternative dependent variables. Columns 5 and 6 present placebo tests adopting dierent explanatory variables. The dependent variable is the total number of terrorist attacks (columns 1-2), the number of non- Islamist attacks (columns 3-4), while in columns 5 and 6 is the standard measure adopted throught

(22)

Finally, we identify the Google search keywords for each country that are most correlated with Islamist attacks over 2004-2015 but are not necessarily related to the Google searches for sensitive keywords. To this aim, we took advantage of an application developed by Google, named Correlate.18 Correlate is an online automated method for query selection that determines which queries best mimic a given trend of data, in our case the occurrence of Islamist attacks, for a given temporal pattern. Before being discontinued in 2019, the tool was available for 50 countries, some of which did not experience any Islamist attack during the observation period. In some cases, this tool could not provide a correlated query, possibly because of the low variability of our dependent variable measuring the attacks.19 Table 6 reports the most highly correlated search keyword found for each country, its English transliteration, if necessary, and the correlation with the series of Islamist terrorist attacks.

The exercise yields quite diversied results.

We see that the highest correlations in each country range from a maximum of 0.9996 in Germany to 0.6418 in Indonesia. More interestingly, while for three countries the search terms found by Google Correlate do meaningfully relate to terrorist attacks (bombs in london at 0.9881 in the UK, terroriste paris at 0.9849 in France, and attentat maroc at a lower 0.8106 in Maroc), in the remaining cases we observe highly correlated terms that are completely disconnected with terrorist attacks, such as binnen i in Germany, which is a word-internal capital I, i.e., a non-standard, mixed case typographic convention used to indicate gender inclusivity for nouns having to do with persons, by using a capital letter I inside the word.20

18Google correlate was available at www.google.com/trends/correlate/. See Mohebbi et al. (2011) for further details on this application. Google shut down Correlate on December 15, 2019, allegedly due to low usage.

19See Vanderkam et al. (2013) for further details on the algorithms which power Google correlate.

20As an example, LehrerInnen means teachers of both sex, while Lehrerinnen is the standard term for female teachers.

(23)

Table 6: Google Correlate

Country Correlation Keyword and English translitteration

Bulgaria 0.9608 Boyan Ivanov

China 0.6770 betfair

Egypt 0.9173 How to work email crosse re (T)

France 0.9849 terroriste paris

Germany 0.9996 binnen i (word-internal capital I) Indonesia 0.6418 jadwal motogp di trans7

Israel 0.6922 Orna Banai

Malaysia 0.7417 lk988 citibet net

Morocco 0.8106 attentat maroc

Saudi Arabia 0.9594 And get rid of the abandonment (T)

Thailand 0.6680 blue waes

Turkey 0.8474 sultangazi kiralik daire (sultangazi for rent apartment) United Kingdom 0.9881 bombs in london

5.2 Instrumental variables approach

We have so far documented the existence of a positive correlation between Islamist attacks and GT . To establish a causal link, we resort to an instrumental variables approach. To this purpose, we must build on an exogenous source of variation of the volume of online queries for our sensitive keywords. The scope for our search is limited, as measures of the instruments must be available with monthly frequency across the world. To address this issue, we focus on the impact of natural and technological disasters. There is evidence that disasters exogenously prompt a strong behavioral response in the public, for example by aecting religious beliefs (Belloc et al., 2016), time preferences (Callen, 2015), risk attitudes (Hanaoka et al., 2018), and support for redistribution (Gualtieri et al., 2019).

Following Jetter (2017; 2019) and Jetter and Walker (2018), we assume that catastrophic events monopolize the attention of the media and the interest of the public, thereby lowering the relative query volume for our sensitive keywords. This assumption is consistent with

(24)

evidence that natural disasters compete with other newsworthy stories in getting media coverage and catching the curiosity of people (Eisensee and Strömberg, 2007). This implies claiming that there is no way through which disasters aect terrorist attacks other than the reaction they prompt in the attention of the public. It is reasonable to assume that natural disasters randomly harm the military capacity of terrorist groups and counter-terrorist forces.

Moreover, as disasters are exogenous, terrorist organizations cannot accordingly schedule their missions in advance to avoid overlapping and competition with other stories in the media coverage of the attacks (Jetter, 2019).

The Centre for Research on the Epidemiology of Disasters (CRED) at the Université Catholique de Louvain (UCL), collects data on catastrophic events worldwide with a daily frequency into the Emergency Events Database (EM-DAT).21 The random occurrence of disasters provides a monthly, exogenous, source of variation of our explanatory variable. Our instruments are two dummies equal to one if a natural or a technological disaster occurred in country c at time t. The results reported in Table 7 suggest that the instrumental variable approach supports our main ndings. Google searches for sensitive keywords signicantly and positively correlate with the number of Islamist attacks.

However, there are reasons to handle this nding with caution. The estimates only regard those countries that faced at least one disaster where at least one Islamist attack occurred over the 2004-2015 period. The intersection between the two sets reduces our sample to 58 countries. The rare occurrence of natural and technological disasters in many of the countries hit by Islamist attacks yields a limited variability of the instruments. Also, there are circumstances under which our assumption on the exogeneity could be challenged.

For example, disasters could harm the State's capacity to react to attacks by requiring the deployment of military forces for civilian purposes. This circumstance may perhaps encourage terrorists to expedite scheduled missions. Even if we nd no proof of such a mechanism in our dataset, we need further information to validate the orthogonality assumption denitively.

21See: www.emdat.be

(25)

Table 7: Instrumental variable approach

DCCE-IV

GTt 1.708*

(0.929)

IAt−1 0.157***

(0.030) Sargan statistic Chi2(53)=24.332

(0.999)

# Observations 8,236

# Countries 58

Notes: The dependent variable is the number of islamist attacks. Standard errors in parentheses, ***, **, and * denote, respectively, 10%, 5%, and 1% signicance levels.

6 Discussion

Our panel analysis reveals a systematic relationship between Islamist attacks and online searches for a set of sensitive keywords over the period 2004-2015 worldwide: online searches signicantly and positively relate to attacks. Dierent kinds of user may search for some of the keywords we employ in the empirical analysis, from students who want to deepen their knowledge of political Islam to journalists and scholars interested in researching violent extremism. Online searches may also detect the celebrity of terrorist organizations. However, we also use keywords explicitly denoting an attraction for fanaticism and violence that are likely searched by sympathizers of the jihad.

Distinguishing between keywords allows us to better understand the positive relation- ship between Google searches and the attacks. We divide our set of 47 keywords into two groups. The rst group contains the keywords with an average monthly frequency higher than 500 searches in the UK. Such searches are not explicitly violent and mostly regard po-

(26)

litical Islam. The SERP analysis conducted by Ahmed and Lloyd George (2016), however, shows that these seemingly non extremist searches often return extreme content aimed at promoting radicalization among Internet users explicitly. The second group contains more extreme queries, which are mostly related to violent extremism and systematically lead to fundamentalist, militant and explicitly violent content (Ahmed and Lloyd George, 2016).

We nd that the online searches for both groups of queries are signicantly associated with the attacks. The coecient for the popular keywords signals that the celebrity of terrorists among the public is signicantly and positively correlated with the attacks. This result is consistent with previous evidence suggesting that media coverage and celebrity encourage terrorists to plan new missions (Jetter, 2017; 2019; Frey and Osterloh, 2018; Jetter and Walker, 2018). The signicant and positive coecient of the queries for the extreme keywords, on the other hand, supports the interpretation that online searches may detect not only the celebrity, but also the popularity of the extremist message. The spreading of curiosity and sympathy for radical Islam and violent extremism may signal the formation of a fertile ground for radicalization. The fact that, through search engines, potential sympathizers easily come into contact with the jihadi propaganda, may further nurture extremism triggering a cycle of radicalization. Rearranging the keywords into dierent groups based on their meaning allows us to understand the relationship between the interest in Islamist extremism and the attacks more in depth. This part of the analysis reveals that attacks are more signicantly and sizably associated to keywords evoking violent actions, such as beheadings, or explicitly related to the Jihad, such as Abdullah Azzam. Beheading was a method of execution in pre-modern Islamic law. Jihadist organizations use beheading as a method of killing captives. Since 2002, groups such as Al Qaeda and ISIS have been mass broadcasting beheading videos as a form of propaganda (Campbell, 2007). Abdullah Azzam was a Sunni scholar, also known as the

Father of Global Jihad (Edwards, 2017). Despite having laid the foundations of Al Qaeda and being considered as the spiritual mentor of Osama Bin Laden, Azzam is not as known to the general public as his disciple. He died assassinated in 1989, well before the rst spectacular

(27)

actions of Al Qaeda, and his name mostly circulates among researchers and sympathizers of the global jihad doctrine (Maliach, 2010). The nding that the online searches for such specic keywords signicantly correlate with the occurrence of Islamist attacks is striking and calls for a more profound investigation into the transmission mechanisms that could allow the popularity of extremist groups to turn into support and feed radicalization.

The limited variability of the instruments and possible exceptions to the assumption of their exogeneity suggest caution in interpreting our results as causal. Despite this limita- tion, our work improves the understanding of the relationship between terrorism and the active interest in radical content. Previous studies suggest that the media coverage of violent actions feeds further attacks, as having the public's attention encourages terrorists to plan new missions. Our approach oers a way to uncover the attention of the public to violent extremism, and documents that people's interest in the terrorists' message is indeed signi- cantly associated to the attacks. The association is even stronger for searches likely denoting a propensity for violence and support for the jihad. Measuring such a fatal attraction could help detecting, and perhaps preventing, the spreading of sympathy for terrorist ideas.

7 Conclusions

In this paper, we propose a new approach to investigate the relationship between the pub- lic's attention to radical Islam and terrorist attacks. The empirical analysis documents a link between the online searches for a group of sensitive keywords and the occurrence of Islamist attacks. Through dierent model specications we show that Google searches for the selected keywords positively correlate with attacks in a panel of countries over the period 2004-2015. Overall, our evidence suggests the possibility that growing interest in violent ex- tremist content may reveal a substratum of public opinion potentially favorable to individual radicalization.

Our ndings extend a small but growing literature studying the relationship between the

(28)

media and terrorism. Previous studies mostly focus on the role of social media and the press coverage of terrorist missions. We add to this eld by highlighting the usefulness of also focusing on the general online landscape that users can access through search engines and employing a new method for detecting the public's interest in radical Islam and violent extremism. The literature has shown that attacks more likely occur when terrorists expect the media to give enough coverage to their actions as the attention of the public brings new opportunities for gaining followers and nurturing radicalization among potential sympathizers (Rohner and Frey, 2007; Jetter, 2017; 2019). We uncover the attention of the public to radical Islam and violent extremism and show that it is indeed systematically and robustly correlated with attacks perpetrated in the name of Islam by self-proclaiming Islamist groups.

Previous literature on radicalization has often focused on the detection of the jihadi propaganda on social networking sites. Our research urges the need to also focus on the decentralized provision of information on the world wide web. Monitoring the broader online landscape helps in detecting the active attention of the public to extremist ideas. Knowing how interest in violent extremism is spread among Internet users is essential not only because it provides a measure of the exposure of the public to violent and illegal content, but also as it may signal the sedimentation of a substratum of sympathy for terrorist claims susceptible of turning into a ground for radicalization. Given previous evidence on the impact of media coverage on the popularity of terrorists, our results call for further investigation on the possibility that the attention raised by the attacks in the media and their audience could fuel a cycle of violence. By raising media coverage and catching the attention of the public, terrorist attacks could awake dormant sympathizers whose radicalization may, in turn, feed new attacks.

References

Abadie, A. (2006). Poverty, political freedom, and the roots of terrorism. American Economic Review, 96(2):5056.

(29)

Aguero, J. M. and Beleche, T. (2017). Health shocks and their long-lasting impact on health behaviors: Evidence from the 2009 H1N1 pandemic in Mexico. Journal of Health Economics, 54:4055.

Ahmed, M. and Lloyd George, F. (2016). A war of keywords. how extremists are exploiting the internet and what to do about it. Technical report, Centre for Religion & Geopolitics and Digitalis, London, UK.

Aly, A., Macdonald, S., Jarvis, L., and Chen, T. (2016). Violent extremism online: New perspectives on terrorism and the Internet. Routledge.

Amodio, F., Baccini, L., and Di Maio, M. (2020). Security, trade, and political violence.

Journal of the European Economic Association, DOI: 10.1093/jeea/jvz060.

Baker, S. R. and Fradkin, A. (2017). The impact of unemployment insurance on job search:

Evidence from Google search data. Review of Economics and Statistics, 99(5):756768.

Baltagi, B. (2005). Econometric analysis of panel data. John Wiley & Sons.

Belloc, M., Drago, F., and Galbiati, R. (2016). Earthquakes, religion, and transition to self-government in Italian cities. Quarterly Journal of Economics, 131(4):18751926.

Benmelech, E., Berrebi, C., and Klor, E. F. (2012). Economic conditions and the quality of suicide terrorism. The Journal of Politics, 74(1):113128.

Bentzen, J. S. (2019). Acts of God? Religiosity and natural disasters across subnational world districts. The Economic Journal, 129(622):22952321.

Berger, T., Chen, C., and Frey, C. B. (2018). Drivers of disruption? Estimating the Uber eect. European Economic Review, 110:197210.

Born, B., Müller, G. J., Schularick, M., and Sedlá£ek, P. (2019). The costs of economic nationalism: Evidence from the Brexit experiment. The Economic Journal, 129(623):2722

2744.

Callen, M. (2015). Catastrophes and time preference: Evidence from the Indian ocean earth- quake. Journal of Economic Behavior and Organization, 118:199214.

Campbell, L. J. (2007). The use of beheadings by fundamentalist Islam. Global Crime, 7(3-4):586614.

Chae, D. H., Clouston, S., Hatzenbuehler, M. L., Kramer, M. R., Cooper, H. L. F., Wilson, S. M., Stephens-Davidowitz, S. I., Gold, R. S., and Link, B. G. (2015). Association between an Internet-based measure of area racism and black mortality. PLoS ONE, 10(4):e0122963.

Chitika Insights (2013). The value of Google result positioning. Technical report.

Choi, H. and Varian, H. (2012). Predicting the present with Google trends. Economic Record, 88(S1):29.

(30)

Chudik, A. and Pesaran, M. H. (2015). Common correlated eects estimation of heteroge- neous dynamic panel data models with weakly exogenous regressors. Journal of Econo- metrics, 188(2):393420.

Clark, A. E., Doyle, O., and Stancanelli, E. (2020). The impact of terrorism on individ- ual well-being: Evidence from the Boston marathon bombing. The Economic Journal, (https://doi.org/10.1093/ej/ueaa053).

D'Amuri, F. and Marcucci, J. (2017). The predictive power of Google searches in forecasting US unemployment. International Journal of Forecasting, 33(4):801816.

Dickey, D. A. and Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series with a unit root. Journal of the American Statistical Association, 74:427431.

Doleac, J. L. and Stein, L. C. D. (2013). The visible hand: Race and online market outcomes.

The Economic Journal, 123(572):F469F492.

Edwards, D. B. (2017). Caravan of martyrs. Sacrice and suicide bombing in Afghanistan.

University of California Press, Oakland, California.

Eisensee, T. and Strömberg, D. (2007). News droughts, news oods, and U.S. disaster relief.

Quarterly Journal of Economics, 122(2):693728.

Eksi, O., Gurdal, M. Y., and Orman, C. (2017). Fines versus prison for the issuance of bad checks: Evidence from a policy shift in Turkey. Journal of Economic Behavior and Organization, 143:927.

El Ouadghiri, I. and Peillex, J. (2018). Public attention to "Islamic terrorism" and stock market returns. Journal of Comparative Economics, 46(4):936946.

Fantazzini, D. (2014). Nowcasting and forecasting the monthly food stamps data in the US using online search data. PLoS ONE, 9(11):e111894.

Fantazzini, D. and Toktamysova, Z. (2015). Forecasting German car sales using Google data and multivariate models. International Journal of Production Economics, 170(A):97135.

Fecht, F., Thum, S., and Weber, P. (2019). Fear, deposit insurance schemes, and deposit reallocation in the German banking system. Journal of Banking and Finance, 105:151165.

Frey, B. S. and Osterloh, M. (2018). Strategies to deal with terrorism. CESifo Economic Studies, pages 698711.

Gentzkow, M. A. and Shapiro, J. M. (2004). Media, education and anti-americanism in the Muslim world. Journal of Economic Perspectives, 16(3):117133.

Giulietti, C., Tonin, M., and Vlassopoulos, M. (2019). Racial Discrimination in Local Public Services: A Field Experiment in the United States. Journal of the European Economic Association, 17(1):165204.

(31)

Gould, E. D. and Klor, E. F. (2016). The long run eect of 9/11: Terrorism, backlash, and the assimilation of Muslim Immigrants in the West. The Economic Journal, 126(597):2064

2114.

Gualtieri, G., Nicolini, M., and Sabatini, F. (2019). Repeated shocks and preferences for redistribution. Journal of Economic Behavior and Organization, 167:5371.

Guriev, S. and Melnikov, N. (2016). War, ination, and social capital. American Economic Review, 106(5):230235.

Hanaoka, C., Shigeoka, H., and Watanabe, Y. (2018). Do risk preferences change? Evidence from the great East Japan earthquake. American Economic Journal: Applied, 10(2):298

330.

Harris, T. F. and Yelowitz, A. (2018). Racial climate and homeownership. Journal of Housing Economics, 40:4172.

Im, K. S., Pesaran, M. H., and Shin, Y. (2003). Testing for unit roots in heterogeneous panels. Journal of econometrics, 115(1):5374.

Jaeger, D. A., Klor, E. F., Miaari, S. H., and Paserman, D. M. (2012). The struggle for Palestinian hearts and minds: Violence and public opinion in the second intifada. Journal of Public Economics, 96(3-4):354368.

Jaeger, D. A. and Paserman, D. M. (2008). The cycle of violence? An empirical analysis of fatalities in the Palestinian-Israeli conict. American Economic Review, 98(4):15911604.

Jetter, M. (2017). The eect of media attention on terrorism. Journal of Public Economics, 153:3248.

Jetter, M. (2019). The inadvertent consequences of al-Qaeda news coverage. European Economic Review, 119:391410.

Jetter, M. and Walker, J. K. (2018). The eect of media coverage on mass shootings. IZA Discussion Paper No. 11900.

Kearney, M. S. and Levine, P. B. (2015). Media inuences on social outcomes: The impact of MTV's 16 and pregnant on teen childbearing. American Economic Review, 105(12):3597

3632.

Koomen, W. and Van Der Pligt, J. (2016). The Psychology of Radicalization and Terrorism.

Routledge, New York, NY.

LaFree, G. and Dugan, L. (2007). Introducing the global terrorism database. Political Violence and Terrorism, 19:181204.

Lampos, V., Moura, S., Yom-Tov, E., Edelstein, M., Majumder, M., Hamada, Y., Rangaka, M. X., McKendry, R. A., and Cox, I. J. (2020). Tracking COVID-19 using online search.

arXiv:2003.08086.

(32)

Ma-Kellams, C., Or, F., Baek, J. H., and Kawachi, I. (2015). Rethinking suicide surveillance:

Google search data and self-reported suicidality dierentially estimate completed suicide risk. Clinical Psychological Science, 4(3):15.

MacInnis, C. C. and Hodson, G. (2015). Do American states with more religious or conserva- tive populations search more for sexual content on Google? Archives of Sexual Behavior, 44(1):137147.

Maliach, A. (2010). Abdullah Azzam, al-Qaeda, and Hamas: Concepts of Jihad and Istishhad.

Military and Strategic Aairs, 2(2):7995.

Metcalfe, R., Powdthavee, N., and Dolan, P. (2011). Destruction and distress: Using a quasi- experiment to show the eects of the September 11 attacks on mental well-being in the United Kingdom. The Economic Journal, 121(550):F81F103.

Mohebbi, M., Vanderkam, D., Kodysh, J., Schonberger, R., Choi, H., and Kumar, S. (2011).

Google correlate whitepaper.

Perdue, R. T., Hawdon, J., and Thames, K. M. (2018). Can big data predict the rise of novel drug abuse? Journal of Drug Issues, 48(4):508518.

Pesaran, M. H. (2004). General diagnostic tests for cross section dependence in panelss. IZA Discussion Paper No. 1240.

Pesaran, M. H. (2006). Estimation and inference in large heterogeneous panels with a mul- tifactor error structure. Econometrica, 74(4):9671012.

Pesaran, M. H. (2007). A simple panel unit root test in the presence of cross-section depen- dence. Journal of applied econometrics, 22(2):265312.

Pesaran, M. H. and Smith, R. (1995). Estimating long-run relationships from dynamic heterogeneous panels. Journal of Econometrics, 68(1):79113.

Pichler, S. and Ziebarth, N. R. (2017). The pros and cons of sick pay schemes: Testing for contagious presenteeism and noncontagious absenteeism behavior. Journal of Public Economics, 156:1433.

Rohner, D. and Frey, B. S. (2007). Blood and ink! The common-interest-game between terrorists and the media. Public Choice, 133(1-2):129145.

Stephens-Davidowitz, S. (2014). The cost of racial animus on a black candidate: Evidence using Google search data. Journal of Public Economics, 118:2640.

Torres Soriano, M. R. (2010). The road to media jihad: The propaganda actions of Al Qaeda in the Islamic Maghreb. Terrorism and Political Violence, 23(1):7288.

Vanderkam, D., Schonberger, R., Rowley, H., and Kumar, S. (2013). Nearest neighbor search in Google Correlate.

(33)

Weimann, G. (2015). Terrorism in Cyberspace: The Next Generation. Woodrow Wilson Center Press with Columbia University Press, New York, NY.

Weimann, G. (2016). Going dark: Terrorism on the dark web. Studies in Conict and Terrorism, 39(3):195206.

Wooldridge, J. M. (2002). Econometric Analysis of Cross Section and panel Data. MIT Press, Cambridge, Massachussets, London, UK.

Young, S. D., Torrone, E. A., Urata, J., and Aral, S. O. (2018). Using search engine data as a tool to predict syphilis. Epidemiology, 29(4):574578.

Young, S. D. and Zhang, Q. P. (2018). Using search engine big data for predicting new HIV diagnoses. PLoS ONE, 13(7):Article e0199527.

Zeitzo, T. (2017). How social media is changing conict. Journal of Conict Resolution, 61(9):19701991.

(34)

Appendix

A.1. Glossary

The denitions provided hereafter are sourced from the ARDA (Association of Religion Data Archives)22 and Wikipedia, and are marked by an (A) and (W), respectively.

Abdullah Azzam: (1941 - 1989) also known as Father of Global Jihad, was a Palestinian Sunni Islamic scholar, theologian, and founding member of Al Qaeda. (W)

Amaq Agency: is a news outlet linked to the Islamic State of Iraq and the Levant (ISIL), it is often the rst point of publication for claims of responsibility by the group. (W)

Apostasy: Departing or falling away from a religious faith. (A)

Apostates in Islam: Apostasy in Islam is commonly dened as the conscious abandonment of Islam by a Muslim in word or through deed. It includes the act of converting to another religion or non-acceptance of faith to be irreligious, by a person who was born in a Muslim family or who had previously accepted Islam. The denition of apostasy from Islam, and whether and how it should be punished are matters of controversy among Islamic scholars.

(W)

Ayman al-Zawahiri: (1951 - ) is the current leader of Al Qaeda (assumed oce in 2011).

(W)

Caliphate: an Islamic state under the leadership of an Islamic steward with the title of caliph (Arabic: khal fah), a person considered a religious successor to the Islamic prophet Muhammad and a leader of the entire ummah (community). (W)

Crusades: Medieval military campaigns of the eleventh through fteenth centuries waged by Christians to recapture Jerusalem from Muslims. (A)

Dabiq Magazine: an online magazine used by the Islamic State of Iraq and the Levant (ISIL) for Islamic radicalization and recruitment. It was rst published in July 2014 (until July 2016) in a number of dierent languages including English. (W)

Ibn Taymiyyah: (1263 - 1328) a medieval Sunni Muslim theologian. His iconoclastic views

22http://www.thearda.com

(35)

on widely accepted Sunni doctrines such as the veneration of saints and the visitation to their tomb-shrines made him unpopular with the majority of the orthodox religious scholars of the time, under whose orders he was imprisoned several times. He has become one of the most inuential medieval writers in contemporary Islam, where his rejection of some aspects of classical Islamic tradition are believed to have had considerable inuence on contemporary Wahhabism, Salasm, and Jihadism. His controversial fatwa allowing jihad against other Muslims is referenced by Al Qaeda and other jihadi groups. (W).

Inspire Magazine: an English language online magazine reported to be published by the organization Al Qaeda in the Arabian Peninsula (AQAP) since 2010. (W)

Islamic State: a type of government primarily based on the application of shari'a (Islamic law), dispensation of justice, maintenance of law and order. From the early years of Islam, numerous governments have been founded as Islamic. However, the term Islamic state

has taken on a more specic connotation since the 20th century. The concept of the modern Islamic state has been articulated and promoted by ideologues such as Ayatollah Ruhollah Khomeini, Israr Ahmed or Sayyid Qutb. Like the earlier notion of the caliphate, the modern Islamic state is rooted in Islamic law. It is modeled after the rule of Muhammad. However, unlike caliph-led governments which were imperial despotisms or monarchies (Arabic: ma- lik), a modern Islamic state can incorporate modern political institutions such as elections, parliamentary rule, judicial review, and popular sovereignty. (W)

Jihad: A term derived from Arabic that means to struggle. For Muslims, there are two types of Jihads: the greater struggle is the internal spiritual battle between the believer and his/her nature, and the lesser struggle is the physical battle against the enemies of Islam.

Muslim extremists and critics of Islam emphasize jihad as a holy war, while most Muslims do not. (A)

Jewish Coalition: The Republican Jewish Coalition (RJC), formerly the National Jewish Coalition, founded in 1985, is a 501(c)(4) political lobbying group in the United States that promotes Jewish Republicans. The RJC is one of the most important voices on conservative

Riferimenti

Documenti correlati

PCNSL: Primary CNS lymphoma; PET: Positron emission tomography; PiB: Pittsburgh compound B; PK11195: N-butan-2- yl-1-(2-chlorophenyl)-N-methylisoquinoline-3-carboxamide;

This kind of survey is used to test how the perception of a good rec- ommendation relates with the number of item shown to the user in order to understand how strong is an algorithm

Infine si possono registrare carenze derivanti dal fenomeno di liquefazione del terreno (§2.1.1 del presente lavoro di tesi). In tal caso gli interventi devono riguardare un’area più

circulation: a State which is defined as a complex institutional assemblage and a site of political practices "which seek to deploy its various institutions and

stati, e per la sua capacità di renderne miglior conto, è dunque da riconsiderare con assoluto favore l’ipotesi di un frammento di commedia ajrcai'a. 182 sulla base di

Si le temps de l’apogée médiéval n’est sans doute pas le temps de la primauté vénitienne, il n’empêche que les premières décennies du XIV e siècle

Although preoperative chemotherapy in colorectal liver metastasis patients may affect the rate of positive resection margin, its impact on survival seems to be limited.. In the