From Butter to Guns, from Guns to Butter? Causality Between Terrorism and Economic Development

From Butter to Guns, from Guns to Butter? Causality Between Terrorism and Economic Development Thomas Gries Daniel Meierrieks y Abstract This contribution investigates the causal relationship between terrorism and economic development, taking a comprehensive global perspective. We use two di erent methods to test for homogenous and heterogenous Granger causality between terrorist activity and economic development for a panel of 83 countries between 1972 and 2007. We nd no evidence of a simple causal link from economic conditions to terrorist activity. We also detect no convincing evidence that terrorism causes signi cant economic damages. These results are supported by variance decompositions and robust to empirical modi cations. Our ndings thus corroborates more skeptical views on the terrorism-economy nexus. Keywords: terrorism, economic development, panel causality analysis JEL Classi cation: N74, O10 Corresponding Author. University of Paderborn, Department of Economics, Warburger Straße 100, 33098 Paderborn. Ph.: +49-(0)5251-60-2113, fax: +49-(0)5251-60-3540, e-mail: thomas.gries@notes.uni-paderborn.de. y University of Paderborn, Paderborn, Department of Economics, Warburger Straße 100, 33098 Paderborn, Germany. Ph.: +49-(0)5251-60-2115, fax: +49-(0)5251-60-3540, e-mail: daniel.meierrieks@notes.uni-paderborn.de. 1

1 Introduction In popular discourse on the roots of terrorism a lack of butter has often been linked to more guns, i.e., economic underdevelopment has been labeled as one of the main causes of terrorism. At the same time, more guns have also been argued to lead to less butter, meaning that terrorism has been repeatedly named as an obstacle to economic activity. While these lines of argumentation are not without controversy (with skeptical voices arguing that other factors matter far more strongly to the genesis of terrorism and the patterns of economic development), they essentially say that the causal dynamics between terrorism and the economy may be more complex. That is, causation may run from terrorism to the economy or vice versa (where it is still unclear whether such causal links actually exist and matter). However, most empirical studies investigating the terrorism-economy nexus assume a xed causal connection between the two, either only estimating the e ects of terrorism on the economy or vice versa. Few studies recognize the potential existence and associated problems of reverse causality. As one example, Abadie (2006) uses an instrumental variable approach when estimating the e ect of per capita income on terrorism to avoid reverse causation, consequently nding that economic variables do not matter strongly to terrorism. Gries et al. (2010) use a time-series approach to assess the causal links between terrorism and economic growth in Western Europe. Their ndings suggest that while in some countries economic success has contributed to a reduction in terrorism, the e ects of terrorism on the economy have not been signi cant. While the latter study directly considers both directions of causation simultaneously, its analytical focus is limited to a speci c selection of countries, making it di cult to arrive at (globally valid) policy conclusions. This contribution seeks to further investigate the question of causation between terrorism and economic development, adopting a global perspective. Correctly taking into account panel integration and cointegration, we employ two di erent methods to test for homogenous and heterogenous Granger causality between various indices of terrorist activity and economic development for a panel of 83 countries between 1972 and 2007. Previewing our results, we nd no evidence that more butter means fewer guns. That is, our results do not show that there is simple causal link runs from (poor) economic conditions to terrorism. We also detect no convincing evidence that terrorism actually causes signi cant economic damages, so we do not nd that more guns necessarily lead to less butter. These results are supported by variance decompositions and robust to further empirical modi cations. In general, our ndings are thus strongly in line with more skeptical views on the terrorism-economy nexus that consider both directions of causation. The remainder of this contribution is organized as follows. In Section 2 we discuss the academic literature on the links between terrorism and economic development. We highlight that the causal relationship between these variables is a priori unclear and also present more skeptical views on the terrorism-economy nexus. In Section 3 we describe the data used to investigate causation in the 2

terrorism-economy nexus. In Section 4 we present our methodology and empirical ndings. We conclude in Section 5. 2 Causality Between Terrorism and Economic Development 2.1 E ects of Terrorism on Economic Development Terrorism is commonly de ned as the premeditated use of violence (or the threat thereof) of non-state actors (i.e., terrorist groups) to obtain political goals through the intimidation of a large audience or government (e.g., Sandler and Enders, 2008). Evidently, terrorists have political or social goals that are not enforceable in the ordinary political process. Through violence terrorists aim at media attention and political and economic destabilization to achieve these ultimate goals (e.g., Melnick and Eldor, 2010; Schelling, 1991). Economic destabilization is a central intermediate goal of terrorists as it weakens their enemy (i.e., the government). The higher the present and anticipated future costs of terrorism in terms of economic damages, the likelier it is that the opposing government concedes to (some of) the terrorists demands. The government weighs the costs of terrorism against the costs of accommodating those demands, with the likelihood of making concessions rising with the probability of more violence (Sandler and Enders, 2008). This mechanism opens up a causal link from terrorism to reduced economic activity. Terrorism may impair economic activity through direct damages (e.g., loss of lives, property damages), reducing an economy s stock of human and physical capital through destruction. The indirect costs of terrorism resulting from the reaction of economic agents to terrorism are also relevant to understanding of the consequences of terrorism. For instance, terrorism is anticipated to hurt certain sectors of an economy (e.g., tourism) that are particularly exposed to terrorism that increases costs and changes consumer demand (e.g., Bird et al., 2008). On an aggregate level, terrorism may distort national levels of consumption, investment, government spending and savings. For instance, terrorism creates uncertainty, leading to the potential postponement of (costly) long-term investments and thus reduced growth (e.g., Bird et al., 2008). As another example, a government may increase spending on security at the expense of (more productive) spending on education or infrastructure, where such a diversion of public spending may impede future growth, potentially resulting in development di erentials. On an international level, the threat of terrorism may lead to the diversion of international capital and trade ows. For instance, Abadie and Gardeazabal (2008) note that terrorism negatively a ects an economy s foreign direct investment position by posing a (costly) risk and reducing the returns to investment. The resulting drawback of international capital may hurt economic development, in particular when foreign nance is an important engine 3

of economic growth (e.g., in developing economies). 1 A number of empirical studies indeed nd that terrorism has, inter alia, a negative e ect on tourism (e.g., Drakos and Kutan, 2003; Enders et al., 1992), on macroeconomic variables such as consumption, investment and public spending (e.g., Blomberg et al., 2004a; Crain and Crain, 2006; Eckstein and Tsiddon, 2004; Gaibulloev and Sandler, 2008) and on the international ow of goods and capital (e.g., Abadie and Gardeazabal, 2008; Enders and Sandler, 1996; Nitsch and Schumacher, 2004). Such e ects help to explain why several studies nd that terrorism is universally detrimental to economic development on regional, national and global levels (e.g., Abadie and Gardeazabal, 2003; Blomberg et al., 2004a; Crain and Crain, 2006; Eckstein and Tsiddon, 2004; Gaibulloev and Sandler, 2008, 2009). 2 While intuitive, the notion that terrorism homogeneously produces economic damages is, however, not without controversy. In fact, a number of empirical studies nd that the e ects of terrorism on foreign direct investment (Enders et al., 2006) and economic development and performance in general (e.g., Blomberg et al., 2004a; Gaibulloev and Sandler, 2009; Gries et al., 2010; Tavares, 2004) are rather modest and short-lived. There are several ways of explaining why terrorism may fail to cause reduced economic activity, which are linked to varying degrees of economic vulnerability and size of attacked economies and di erences in the intrinsic motivation for attacking. These factors may explain why causation from terrorism to the economy does not follow a (negative) homogenous but rather heterogenous pattern, where the causal link may only matter to some less-advanced economies, if there is any signi cant causal link at all. As Sandler and Enders (2008) point out, the e ects of terrorism are anticipated to have only localized e ects. For instance, terrorist attacks may only hurt the tourism industry in a certain part of a country. When countries face the threat of terrorism, substitution of economic activity away from vulnerable to less vulnerable sectors of the economy is likely to take place (accompanied by quick adjustments in price and capital allocation), in particular when the economies in question are well-diversi ed. Also, markets (notably, nancial markets) may be able to anticipate the risk of terrorism, so they are able to work rather e ciently even when faced with attacks (e.g., Chen and Siems, 2004). Such e cient market behavior should mitigate any negative e ect of localized terrorism on national economic development. Similarly, businesses can adapt to terrorism, anticipating its negative consequences, e.g., by means of business decentralization (Frey, 2009). Such forward-looking behavior of economic agents should further reduce the impact of terrorism on development. In 1 Our argumentation of links from terrorist violence to poor economic development is similar to that proposed by Collier (1999) for the negative e ects of civil war on economic conditions. Bird et al. (2008) and Sandler and Enders (2008) discuss the potential macroeconomic consequences of terrorism in more detail. 2 These empirical results also resemble the ndings on the negative e ects of political instability (e.g., Jong-A-Pin, 2009) and civil war (e.g., Murdoch and Sandler, 2002) on economic conditions. 4

general, terrorism may fail to cause economic damages because its immediate e ects are almost always localized and attacked economies may be su ciently resilient (with economic agents reacting to terrorism with, say, substitution). This implies that when attacked economies are su ciently large and diversi ed, they should be able to absorb the risk of terrorism, so that the (most probable) negative e ects of terrorism on certain sectors of an economy will not translate into reduced economic development on the national level (measured, e.g., in real GDP per capita). In fact, even the high costs of the 9/11 attacks (in absolute terms) can be considered small when seen in relation to the economic size of the US (cf., e.g., Sandler and Enders, 2008). As Sandler and Enders (2008) note, small and less diversi ed countries facing signi cant terrorist activity ought to experience the most severe macroeconomic consequences. However, even when countries are economically vulnerable to terrorism, development may still not be interrupted when economic destabilization is not the main goal of an attacking terrorist group. For instance, terrorist activity by domestic terrorists against foreigners may have political rather than economic goals. As a prominent example, domestic terrorist groups may target Americans living/working in their countries (e.g., military personnel) to gain political in uence at home, e.g., since attacks against American interests are likely to erode American support for the government the domestic terrorist group opposes (Neumayer and Plümper, 2010). Furthermore, as Sanchez-Cuenca and De la Calle (2009) show, many terrorist organizations systematically attack combatants (i.e., the military and police). Again, such attacks obviously have a political/military motive (i.e., the political/military weakening of the government) and may not have the same adverse e ects on the economy as actions against economic targets (e.g., attacks on hotels etc.). That is, the reason why terrorist groups attack (which may di er across countries) may be rather political and/or military. Consequently, when terrorism has predominantly non-economic goals, its economic e ects should be rather marginal. In general, some theoretical and empirical contributions argue that terrorism exerts a negative and homogenous causal e ect on economic development, e.g., by damaging an economy s global investment position or distorting national levels of consumption and investment. This view is, however, challenged by other studies which nd that the causal e ect from terrorism to the economy is heterogenous and may very well be unsubstantial. This skeptical view argues that the e ects of terrorism on the economy are small, short-lived and strongly dependent upon the socio-economic size and vulnerability of an attacked economy and the motives of attacking terrorist groups. 2.2 Impact of Economic Development on Terrorism Poor economic development is often argued to be a root cause of terrorism, in particular in popular discourse (cf. Krieger and Meierrieks, 2010). The intuition is that poor economic development creates grievances among the economically disenfranchised, in turn leading to violence (terrorism) that serves to change 5

the status quo. In an economic sense, terrorists are rational actors who consider the costs, bene ts and opportunity costs associated with violence to choose the optimal level of terrorism (e.g., Sandler and Enders, 2004). Taking such a perspective, it is possible to theoretically understand how low levels of economic development may translate into terrorist violence. Blomberg et al. (2004b) o er an economic model of terrorism along such lines. They argue that the state of the economy matters to the extensity of internal con ict. In poor economic times (e.g., when growth is slow) economically disenfranchised groups are more likely to resort to terrorism in order to make their voice in the economy heard. When economic underdevelopment abounds, terrorists may nd it more promising to engage in violence in order to participate economically (so that the bene ts of terrorism increase), in particular as they are less likely to nd non-violent (economic) alternatives (so that the opportunity costs of violence are comparatively low). Also, low levels of economic development may mean that the state has only limited means to counter terrorism because government capacity is anticipated to be a function of an economy s capacity (so that the costs of terrorism decrease). The considerations outlined above provide a theoretical causal pathway (via a terrorist s calculus) through which poor economic conditions may translate into terrorism, resembling the popular notion that terrorism is rooted in poverty. Conversely, economic success is expected to reduce terrorist activity. 3 Indeed, several studies nd that terrorism is less likely to originate in countries that exhibit comparatively high rates of economic growth and levels of per capita income (Blomberg and Hess, 2008; Blomberg et al., 2004c; Lai, 2007; Muller and Weede, 1990). Furthermore, Burgoon (2006) argues that terrorism and the economy are also indirectly linked through social welfare systems. He nds that terrorism is less likely to originate in countries with developed social welfare systems because these very systems are able to ameliorate the poor socio-economic conditions which may otherwise lead to violence. That is, the studies discussed above o er support for the popular notion that there is a causal link from development to terrorism, arguing that terrorism is homogeneously linked to poor national economic conditions. However, the hypothesis that terrorism is predominantly swayed by economic conditions is challenged by a number of empirical studies, as summed up by Krieger and Meierrieks (2010). While these studies control for the e ect of economic factors on the genesis of terrorism, they fail to provide evidence of a connection between them. Instead, they nd that terrorism is more strongly linked to non-economic factors. For instance, Abadie (2006), Krueger and Laitin (2008), Krueger and Maleckova (2003) and Kurrild-Klitgaard et al. (2006) do not nd that the production of terrorism is signi cantly linked to poverty, slow growth and income inequality. Instead, they nd that terrorism is more strongly 3 This reasoning is consistent with studies on the causes of civil war which have also found that this form of violence becomes less likely in phases of strong economic activity (e.g., Collier and Hoe er, 2004). 6

connected to a country s political and institutional environment (e.g., the degree of political repression). Further studies reviewed by Krieger and Meierrieks (2010) have also assessed the links between terrorism and, e.g., political instability, ethnic con ict and international integration (e.g., Piazza, 2008; Sanchez- Cuenca, 2009). Again, most of these studies fail to nd a consistent connection between terrorism and poor economic conditions; rather, they nd that terrorism is linked to other variables. Evidently, when non-economic (e.g., political or ethno-religious) factors are important to the genesis of terrorism, the e ect of economic development on terrorism should rather be marginal. In general, a popular notion argues in favor of a homogenous causal link from economic underdevelopment to terrorist activity. While some empirical studies have produced evidence in support of this, there is also a more skeptical view. While this latter view does not necessarily rule out the possibility of heterogenous causation from economic development to terrorism (so that this link matters to some, but not all countries), it strongly rejects the possibility of homogenous causality from economic conditions to terrorist activity. Instead, this skeptical view argues that non-economic factors (e.g., political or ethno-religious repression) sway a terrorist s calculus much more strongly than economic factors. 3 Data In order to empirically assess the causal links between terrorism and economic development and to arrive at globally valid conclusions, we compile data on terrorist activity and economic development for 83 countries for the period 1972 to 2007. Note that we predominantly focus on those countries that have actually been hit by major terrorist activity in the past. We expect terrorists to be more successful in destabilizing the economy when they carry out a prolonged and violent campaign rather than isolated attacks. For the period of observation we take three-year averages for all variables indicating economic development and terrorism to take into account cycle e ects in terrorist activity and economic development. Also, taking averages reduces some of the methodological problems associated with the detection of causality, as we shall discuss below. A list of countries used in this contribution is given in the appendix. Indicating Terrorist Activity Raw data for terrorist activity comes from the Global Terrorism Database (GTD) introduced by LaFree and Dugan (2007). The main advantage of this database is that it contains information on transnational and domestic terrorist events, unlike other datasets (notably ITERATE) that focus on transnational terrorism only. 4 Even though some scholars have raised doubts about the quality of the 4 The academic literature commonly di erentiates between domestic and transnational terrorism. The former involves only one country (i.e., the terrorists homeland), whereas the latter involves at least two countries (e.g., because terrorist groups strike outside their home- 7

GTD data (e.g., Gaibulloev and Sandler, 2009: 363), we believe that its use is advantageous in particular because we are not limited to an investigation of the causal nexus between economic conditions and transnational terrorism. For this analysis we do not follow the classic distinction between domestic and transnational terrorism. As argued by Sanchez-Cuenca and De la Calle (2009), the distinction between domestic and transnational terrorism may not be suitable because it truncates the data in unnecessary (and potentially misleading) ways. For instance, transnational terrorism may be directed against foreign policy, foreign dominance or foreign a liation with and support for domestic governments (e.g., Neumayer and Plümper, 2010; Savun and Phillips, 2009). Focussing on international terrorism may thus lead to an underestimation of the role of economic factors in explaining terrorist activity in favor of political ones. For this analysis, we follow Sanchez-Cuenca and De la Calle (2009) and de ne terrorism (homeland terrorism) as all terrorist activity carried out by nonstate actors (i.e., terrorists) operating within their homeland, regardless of the nationality of the target. 5 We believe that our broader de nition of homeland terrorist activity allows us to more accurately investigate the terrorism-economy nexus, in particular as we are able to avoid drawing general conclusions while only focusing on truncated data (i.e., on transnational terrorism). As an example, we record all terrorist activity by ETA (Euskadi Ta Askatasuna) attacks in Spain (ETA s homeland) and the victims/targets of these attacks (regardless if they are Spanish, American etc.) in a given year. 6 Given that a number of potential indicators of terrorist activity have been suggested in the empirical literature, we rely on three distinct proxies. When constructing these variables, we explicitly take into account the mediating role of country size (population size) in the terrorism-economy nexus by adjusting the terrorism indicators for country size. On the one hand, the e ect of terrorism is anticipated to matter more strongly to a small than to a large country (e.g., Gaibulloev and Sandler, 2009; Sandler and Enders, 2008) due to the generally higher socio-economic resiliency of larger (in population terms) economies. On the other hand, it is also reasonable to adjust for country size when assessing the (economic) causes of terrorism, given that population size is consistently identi ed as a strong predictor of terrorist activity (cf. Krieger and Meierrieks, land or because foreign interests are targeted inside the terrorists homeland). Commonly, the relation between domestic and transnational terrorist attacks is assumed to be ca. 8:1 to 9:1, so domestic terrorism substantially outnumbers transnational terrorism (cf. Sanchez-Cuenca and De La Calle, 2009). 5 Note that in some cases of our sample (e.g., Peru, Sri Lanka) terrorism has been used as a strategy of insurgency, so that such terrorist activity can also be considered as part of a civil war (cf. Merari, 1993). 6 This de nition of terrorism does not consider unclaimed terrorist attacks (where the perpetrators are unknown). It also does not include terrorism that is imported from a third country. For instance, we do not take into account PKK terrorism (i.e., Turkish terrorism) in Western European countries in the 1990s. From our point of view this kind of imported terrorism should not be linked to poor economic conditions in the country where the attack occurs. Rather, such attacks may be motivated, e.g., by foreign policy (cf. Savun and Phillips, 2009). 8

2010). First, we use the number of homeland terrorist attacks per 100 000 inhabitants (ATT ), similar to, e.g., Crain and Crain (2006) and Tavares (2004). This proxy indicates the frequency of terrorism, adjusted for country size. Second, we employ the number of victims of homeland terrorist attacks per 100 000 inhabitants (VICT ), again as in, e.g., Crain and Crain (2006) and Tavares (2004). Victims of homeland terrorism are individuals reported as being wounded or killed in a homeland terrorist attack. This indicator measures the intensity of terrorism. Third, we create a homeland terrorism index (TERR) similar to that proposed by Eckstein and Tsiddon (2004). This index is de ned as the natural logarithm of the index that is equal to the mathematical constant e plus the equally weighted and population-adjusted number of terrorist attacks by groups operating within their homeland and the sum of the individuals wounded or killed in these very attacks (regardless of their nationality). 7 This index takes into account the frequency and ferocity of while correcting for country size. Indicating Economic Development In order to indicate the level of economic development we use the straightforward measure of the (logged) real gross domestic income per capita (GDP). This variable is commonly used in empirical studies to assess the in uence of terrorism on economic development (e.g., Crain and Crain, 2006) and the economy s impact on terrorism (e.g., Lai, 2007; Tavares, 2004), even though it may be argued that GDP does not necessarily re ect all negative e ects of terrorism on economic variables (e.g., unemployment, absolute poverty) nor all of terrorism s potential economic roots (e.g., poverty, income inequality). The data for this variable comes from the most recent version of the PENN World Table provided by Heston et al. (2009). 4 Empirical Methodology and Results In this section we describe our empirical methodology and results. We rst investigate the stationarity properties of the data by means of panel unit root tests and also discuss the issue of panel cointegration. Treating the data for economic development and terrorism correctly (as it follows from the panel unit root tests), we then test for panel causality. Using panel data, we are able to incorporate more observations (even when using averaged data) than in pure time-series analyses, thus obtaining more precise estimates. As we are able to use both cross-sectional and time-series data, we also, inter alia, arrive at more e cient results regarding causation, at the same time (potentially) controlling 7 Formally, our terrorism index T ERR for country i in year t is de ned as: T ERR i;t = ln(e + attacks i;t + victims i;t ). population i;t population i;t 9

for country-speci c e ects and reducing identi cation problems (cf. Hurlin and Venet, 2001). For an empirical analysis of the causal link between terrorism and economic development we generally follow Granger s (1969) de nition of causality which implies causation in a statistical but not necessarily philosophical sense. Testing for Granger causality in bivariate panel systems, our ndings indicate whether past (i.e., lagged) values of terrorism (economic development) help to signi cantly explain present values of economic development (terrorism), also taking into account past values of the dependent variable. As discussed in the literature review, causality may run (i) unidirectionally from terrorism to economic development, (ii) unidirectionally from development to terrorism or (iii) in both directions at the same time (bidirectional causality). Finding no causal links between the two variables implies that (iv) there is no signi cant link between them, so that terrorism and economic development are driven by other factors instead. This nding would reinforce more skeptical views regarding the terrorism-economy nexus. We use two distinct methods to test for causality in a panel setting. First, we test for homogenous panel causality using the method proposed by Holtz-Eakin et al. (1988). Second, we test for heterogenous panel causality following Hurlin and Venet (2001). We believe that it is appropriate to present the ndings from both causality analyses, given that there is still no consensus on the superiority of any of these empirical techniques. We complete our main analysis by presenting variance decompositions obtained from a set of panel vector autoregressions. Finally, we discuss some extensions to our main empirical work and the robustness of our ndings. 4.1 Panel Unit Root and Cointegration Both procedures to test for Granger causality in a panel framework consider covariance-stationary systems in order to make correct causality inferences. That is, we need to investigate the stationarity properties of the series for economic development and terrorism. Finding that the series are integrated of the same order may also raise the question of panel cointegration. Given that traditional (time-series) unit root tests are known to have little power, we resort to two recently developed methods to detect stationarity in a panel framework, namely the LLC test (Levin et al., 2002) and the IPS test (Im et al., 2003). While the LLC test assumes a common unit root process within the panel, the IPS test assumes individual unit root processes. The test results are reported in Table 1. Table 1 here The test results indicate that the GDP series is I(1) and only becomes stationary after taking rst di erences. By contrast, the tests show that all terrorism series are I(0). 8 In order to make correct causality inferences it is 8 Other panel unit root tests (e.g., Breitung, 2000) also strongly suggest that the GDP series is I(1), whereas the terrorism series are I(0). These results are not reported to save space. 10

thus necessary to adjust the GDP variable by applying the di erence lter, so to remove any non-stationarity. We therefore use the rst di erence of the (logged) GDP (GDP) series in the subsequent analysis. Note that because the GDP and terrorism series are not found to be integrated of the same order, the issue of panel cointegration does not arise. 4.2 Testing for Homogenous Panel Causality Methodology We rst test for homogenous panel causality, following Holtz-Eakin et al. (1988). Here, we consider a time-stationary bivariate system of the following form: X ijt = 0 + Y it = 0 + px 1 X ijt l + l=1 px 1 X ijt l + l=1 px 2 Y it l + i + u it (1) l=1 px 2 Y it l + i + v it (2) X is the measure of terrorism for country i over t periods in the j-th form (representing the three proxies for terrorist activity). Y is our proxy for economic development (log-di erence of the real GDP per capita). and are the intercepts or regression coe cients, and are the country-speci c xed e ects, and u and v are the error terms. The lag order of the explanatory variables may run from l = 1 to p. The lag order p should meet (in line with the usual rule of thumb) the requirement that T i > 5 + 2p (with T i indicating the number of time spans available for country i). Given that we take three-year averages for X and Y, T = 12. After taking the rst di erences for logged GDP per capita, we arrive at T = 11. Using a di erence GMM estimator (as described below), T = 10. Thus, we work only with a lag order of p = 1. Considering that the e ect of terrorism on the economy is anticipated to be rather short-lived, a cautious lag length is called for. We estimate the above system in rst di erences to eliminate any individual e ects (and intercepts) from the model, avoiding a common problem in dynamic panel analyses stemming from the correlation between the error terms, the xed e ects and the lagged dependent variable. We use the di erence GMM estimator (e.g., Arellano and Bond, 1991) to obtain consistent estimations through the use of instrumental variables. For this approach, taking averages is particularly suited so to avoid using too many instruments when estimating the corresponding GMM models (cf. Roodman, 2009). The validity of the instruments is checked relying on the Hansen test for overidentifying restrictions and the Arellano-Bond test for rst and second order serial correlation (M1/M2 test). Following the empirical application of the afore described methodology by Holtz-Eakin et al. (1988), our approach to testing for Granger causality in a l=1 11

panel framework involves the following steps. For any X j (terrorism) to grangercause any Y (economic development), in Equation (2) there has to be a signi - cant e ect of the lagged values of X j;t l on Y t (i.e., 1 6= 0), conditional upon the lagged values of Y t l. We examine whether there is a signi cant e ect through a Wald test for noncausality which checks whether the lagged values of X j;t l are jointly signi cant with respect to Y t. The sign of 1 hints at the nature of the causality e ect. For instance, 1 < 0 means that higher past values of X j (terrorism) cause lower present values of Y (economic development). In order to examine whether past values of Y cause present value of X j, we have to check if 2 6= 0 in Estimation (1), again using a Wald noncausality test. We can also assess the nature of the causality e ect by investigating the signs of 2. Note that through this procedure we test for homogenous causation between two variables. Either there is a causal relationship between the two variables that is uniform across the whole sample (e.g., 1 6= 0) or there is not (so that, e.g., 1 = 0). Results The results of the panel Granger causality test following Holtz-Eakin et al. (1988) are reported in Table 2. They indicate that terrorist activity exerts a negative, uniform and signi cant e ect on economic development, regardless of which indicator of terrorism is used. By contrast, we never nd that economic development has a substantial causal impact on terrorist activity. Note that the diagnostic tests (M1/M2 and Hansen test) show that all models are well speci ed and that the vector of instruments is appropriate. Table 2 here Finding a signi cantly negative impact of homeland terrorist activity on economic development (measured as GDP ) ts in previous studies arguing that terrorists are successful in achieving economic destabilization through a number of potential channels (e.g., Abadie and Gardeazabal, 2003; Crain and Crain, 2006; Eckstein and Tsiddon, 2004; Gaibulloev and Sandler 2008). While we cannot assess the exact economic channels from terror to reduced economic activity, we can speculate that terrorism may distort, e.g., public spending and private saving and investment behavior, and international capital and trade ows, thus impairing economic development. The absence of a signi cant causal link from economic development to homeland terrorist activity is in line with other studies that do not suggest that terrorism is rooted in poor economic conditions (e.g., Abadie, 2006; Krueger and Laitin, 2008; Krueger and Maleckova, 2003). Instead, they emphasize that terrorism is determined by, e.g., political, institutional and demographic factors (cf. Krieger and Meierrieks, 2010). That is, our ndings do not lend support to studies that emphasize the fundamental role of economic conditions in determining terrorist activity. 12

4.3 Testing for Heterogenous Panel Causality Methodology Testing for homogenous causation with the methodology presented above may entail certain problems which we can circumvent by testing for heterogenous panel causality, following the methodology introduced by Hurlin and Venet (2001) and Hurlin (2005). 9 First, through rst-di erencing the above methodology eliminates any country-speci c xed e ects from the estimation (and corresponding inferences about causation). However, it is very likely that such e ects are highly signi cant and matter to the terrorism-economy nexus. For instance, a country s level of institutional quality and economic vulnerability may determine how prone its economy is to terrorist activity, where such differences may be captured in the xed e ects. We therefore explicitly consider such e ects in the empirical methodology described below. Second, testing for causation in the sense of Holtz-Eakin et al. (1988) implies that causation is truly homogenous. Granger (2003: 70) argues that this is a strong assumption. In particular, the dynamics of the terrorism-economy nexus are likely to di er across countries, as already discussed above. For instance, terrorism may damage less-diversi ed economies far stronger than those that are more diversi ed, meaning that causation from terrorism to economic development matters only to the former countries (so that causation is indeed heterogenous). That is, assuming homogenous causation may generally lead to false conclusions about causality. With the methodology described below, we are able to identify heterogenous causal relationships between terrorism and the economy, potentially allowing for the identi cation of those subsets of countries for which related causal dynamics are truly important. To test for heterogenous panel Granger causality, we consider a time-stationary bivariate system of the following form: X ijt = Y it = px 1 X ijt l + l=1 px l=1 (p) 1 X ijt l + px l=1 (p) 2 Y it l + i + u it (3) px 2 Y it l + i + v it (4) The notation is the same as above, as are the variables used. Following the constraints made by Hurlin and Venet (2001), the autoregressive slope coe cients (i.e., 1 and 2 ) are identical for all cross-sections. The regression coe cients (i.e., 2 and 1 ) are allowed to vary across individual cross-sections, thus allowing for causal variation across countries. 10 Note again that the xed 9 This methodology has been employed and discussed in Hood et al. (2008) and Lin and Ali (2009). 10 Given that we only work with a lag length of p = 1, we do not have to consider any constraints regarding the di erent coe cients from time period to time period and across di erent lag lengths. However, related constraints are discussed in, e.g., Hurlin (2005), Hurlin and Venet (2001) and Hood et al. (2008). l=1 13

e ects are not dropped when proceeding with the causal analysis. Testing for heterogenous panel causality between terrorism and economic development involves the following steps. First, we estimate a restricted model which only includes the lagged values of the independent variable (i.e., P P 1 X ijt 1 and 2 Y it 1 ) and the country-speci c xed e ects ( i and i ) to explain present values of the dependent variable. We store the residual sum of squares obtained from this estimation (RSS 2 ). Then, we estimate the unrestricted model (i.e., Equations (3) and (4)) which also includes information on lagged values of the independent variable. Again, we save the residual sum of squares from this model (RSS 1 ). Finally, we calculate an F -statistic to make causality inferences from the following equation: F = (RSS 2 RSS 1 )=(N p) RSS 1 =[N T N(1 + p) p] ; (5) where N is the number of cross-sections (countries), p is the number of lags (=1) and T is the number of time periods (three-year averages). We evaluate the signi cance of the F -statistic using an F -distribution with Np; NT N(1+p) p df (cf. Hurlin and Venet, 2001). A signi cant F -statistic implies that the independent variable Granger causes the dependent variable in at least one crosssection. This hypothesis evidently implies that causation does not necessarily follow a homogenous pattern. An insigni cant F -statistic indicates that there is no causal relationship for any cross-section. Results The results of the panel Granger causality test following Hurlin and Venet (2001) are reported in Table 3. As in our test for homogenous panel causality, we do not nd that economic development has a substantial causal impact on terrorist activity, regardless of which proxy for homeland terrorist activity is employed. In sharp contrast to the ndings from the homogenous panel causality analysis, our ndings now also show no signs of a causal link from homeland terrorist activity to economic development. Table 3 here We thus nd robust evidence that homeland terrorist activity (with respect to its frequency and ferocity) is not rooted in poor economic conditions. Note that even though we allow for causal variations across countries, we do not nd that economic development (measured as GDP ) exerts a causal e ect on terrorism (measured as AT T, V ICT and T ERR) in any cross-section. The ndings from the homogenous panel causality are thus con rmed even when allowing for heterogenous causation. This nding lends some support to those studies attributing only a negligible role to economic conditions in determining terrorist activity. Evidently, non-economic factors matter more strongly to the genesis of terrorism. 14

Finding that terrorism does not have a causal e ect on economic development using the heterogenous panel causality technique is at odds with previously reported results. We believe that this discrepancy is a consequence of the inclusion of country-speci c e ects. The country-speci c factors may capture a variety of e ects in uencing the terrorism-economy nexus and shaping the patterns of economic development (e.g., the country-speci c degrees of trade openness and government spending). Once these factors are incorporated (by allowing for distinctive intercepts in the model), the causal e ect of terrorism on economic development vanishes. That is, testing for heterogenous panel causality we nd little support for the notion that economic development is negatively impacted by terrorism, contrary to some previous studies (e.g., Crain and Crain, 2006). Rather, our ndings seem to be consistent with more skeptical studies on the e ect of terrorist activity on economic conditions (at least, when it is measured as GDP ). These contributions (e.g., Sandler and Enders, 2008; Tavares, 2004) argue that related e ects are rather short-lived and/or modest and that other factors matter more strongly to economic performance and development. For instance, Tavares (2004) nds that once additional controls (e.g., for trade openness) are factored in, the in uence of terrorism on economic growth becomes marginal, while the negative e ects of natural disasters and currency crises on growth survive such modi cations. Evidently, even though terrorists may aim at economic destabilization, according to this empirical method there is little evidence that they actually succeed. 4.4 Variance Decompositions In order to further investigate the terrorism-economy nexus, we conclude our main analysis by carrying out a set of variance decompositions. Here, we set up a panel vector autoregression (VAR) model using the data described above. 11 As in Love and Zicchino (2006), we consider the following panel VAR: Z it = A 0 + A 1 Z i;t l + i + e t ; (6) where Z it is a two variable vector (e.g., fgdp; T ERRg or fgdp; AT T g) indicating any possible combination of economic development and homeland terrorist activity. A 0 and A 1 are a vector of intercepts and regression coe cients. As in the panel analyses described above, all variables are treated as endogenous. The panel VAR approach also allows for the incorporation of unobserved individual heterogeneity through i (country-speci c e ects). e t is the error term. Due to their econometric setup, variance decompositions complement our panel causality ndings. Here, the main objective of the panel VAR approach is to perform variance decompositions to investigate how the variation in one variable is explained by the shock to another variable over time. For instance, variance decompositions help us to understand how much of the forecast error 11 Following Love and Zicchino (2006), we time-demean the data and then create forward mean-di erenced variables via a Helmert transformation. 15

variance of economic development (GDP ) can be explained by exogenous shocks to homeland terrorism (e.g., AT T ). The results of the variance decompositions derived from the panel VAR are reported in Table 4. These ndings mostly mirror the earlier presented panel causality ndings. On the one hand, we nd that homeland terrorist activity (measured by three distinct proxies) is virtually unresponsive to innovations in economic development over a 10- and 20-year time horizon. This is consistent with our earlier ndings that there is no substantial causal e ect running from economic conditions (GDP ) to homeland terrorism. On the other hand, we also nd little support that economic development reacts strongly to shocks in terrorist activity. At best, we nd a moderate e ect from T ERR to GDP, but not for the other two terrorism proxies. This result matches our earlier causality ndings that do not attribute a particularly strong role to terrorism in shaping economic development, especially after country-speci c e ects are factored in. Considering the individual response functions (not reported) we nd that the e ect of terrorism is rather short-lived, contributing to the view that terrorist activity on its own is not an important determinant of economic development on a long-run, global scale. 4.5 Robustness Table 4 here We assess the robustness of our empirical ndings by employing di erent proxies of terrorist activity. As de ned above, AT T, V ICT and T ERR indicate homeland terrorist activity, i.e., these variables contain information on domestic and transnational terrorist activity at the same time, in contrast to the classic distinction between these two forms of terrorism that is common in the academic literature. As already discussed before, it is possible that the objectives of transnational terrorism are di erent from the those of domestic terrorism. For instance, attacks by homeland terrorist groups against foreign targets may not be motivated by a desire to destabilize the economy. Rather, they may be regarded as a protest against foreign policy or foreign support for domestic governments (e.g., Neumayer and Plümper, 2010). By contrast, purely domestic terrorism may be more strongly rooted in economic conditions and aim at economic damage, so that a mixture of domestic and transnational terrorism may actually lead to misleading causality inferences. We thus create proxies for terrorist activity similar to AT T, V ICT and T ERR, either incorporating only information on domestic or transnational terrorism (i.e., terrorist attacks on either domestic or international targets). Running a series of panel causality tests using the methodology described above with these additional variables gives a picture that is very similar to the one presented before (results not reported). That is, we do not nd evidence of a statistically relevant causal e ect from economic development to domestic or transnational terrorist activity. Also, a negative causal link from domestic 16

or transnational terrorism to economic development is only present when employing the panel causality test of Holtz-Eakin et al. (1988) but not when using the approach by Hurlin and Venet (2001) which also factors in country-speci c e ects that may belittle the e ect of terrorism on economic conditions. When using proxies that only indicate domestic or transnational terrorist activity, we neither nd that one form of terrorism is more successful in achieving economic destabilization (or is more focussed on this goal) nor that one form of terrorism is more strongly rooted in economic circumstances than the other. Creating a more complete variable indicating homeland terrorist activity has not led to incorrect causality inferences. As a further robustness check, we also analyzed whether the terrorismeconomy nexus is contingent upon national levels of economic development by creating two sub-samples indicating the developed and the less-developed world. 12 Again running a series of panel causality tests (results not reported), we do not nd that the terrorism-economy nexus di ers between the two subsamples. While we nd no evidence that terrorism is rooted in economic development for the developed and less-developed world, we nd some (but not convincing) evidence that terrorism may cause economic damages. Generally, these results do not indicate that the terrorism-economy nexus is dependent upon economic development. These ndings provide additional support for the main panel causality ndings presented before. 5 Conclusion In this contribution we investigated the causal links between terrorism and economic development. We speci cally considered that more guns may mean less butter and that less butter may mean more guns at the same time. Taking into account the issues of panel integration and cointegration, we used two di erent methods to test for homogenous and heterogenous Granger causality between various indices of terrorist activity and economic development in a panel framework for 83 countries during 1972-2007. We found no evidence that terrorism is rooted in poor economic conditions. 13 We also found no convincing evidence that terrorism substantially impairs economic development. 14 These results were supported by variance decompositions and robust to some empirical modi cations. In general, our ndings were thus strongly in line with more 12 To create these two sub-samples, we followed the recent classi cation of the World Bank. However, given that some countries (e.g., Algeria, Peru) have strengthened their economies during our period of observation (1972-2007), making it di cult to classify them unambiguously, we included them in both sub-samples. 13 Due to our de nition of terrorism (which does not include imported transnational terrorism) our ndings do not necessarily contradict previous ndings (e.g., Krueger and Laitin, 2008) which suggest that economic success may attract (the import of) transnational terrorist attacks (i.e., attacks from abroad). 14 Note that there may be a stronger negative e ect of terrorism on economic development once we also factor in the impact of cases of imported transnational terrorism (e.g., the attacks on 9/11) and unclaimed terrorism. 17