How Much Terror? Dissidents, Governments, Institutions and the Cross-National Study of Terror Attacks

Transcription

1 How Much Terror? Dissidents, Governments, Institutions and the Cross-National Study of Terror Attacks Will H. Moore, Ryan Bakker, and Daniel W. Hill, Jr. May 15, 2013 Previous versions of this research have been presented to the Centre for the Study of Civil War Working Group at Peace Research Institute, Oslo; the SPAM seminar at Notre Dame, the START Center at the University of Maryland, the Departments of Political Science at Essex Univeristy, Florida State University, and University of North Texas; and the annual meetings of the International Studies Association (2010, New Orleans, LA) and Peace Science Society (2008, Claremont, CA). We appreciate comments and feedback on this project from John Ahlquist, Dave Armstrong, Jason Barabas, Bethany Barrett, Bill Berry, Halvard Buhaug, David Cunningham, Kathleen Cunningham, Scott Edwards, Mike Findley, James Forest, Scott Gates, Jeff Gill, Hȧvard Hegre, Brian Lai, Jim Piazza, Chris Reenock, Dave Siegel, Hȧvard Strand, and Joe Young. Professor, Department of Political Science, Florida State University. Assistant Professor, Department of Political Science, University of Georgia. Assistant Professor, Department of International Affairs, University of Georgia.

2 Abstract Our knowledge of the set of concepts that influence the number of terror attacks experienced by different countries is rudimentary. Existing work on the cross-national incidence of terror focuses upon the structural characteristics of polities, economies, and societies, and fails to place competition between dissidents and states center stage. This study addresses that shortcoming by proposing a theoretical framework that places competition between dissident groups and government at the center of the analysis, then scours the literature for existing arguments and develops hypotheses by evaluating the implications of those arguments within a dissident/state-centered framework. We evaluate the resulting hypotheses using Bayesian statistical techniques and a broad scope of relevant data. The results demonstrate that government and dissident behavior have larger effects on the cross-national incidence of terror attacks than the variables that populate existing research.

3 Introduction This study develops and tests, in a cross national time series (CNTS) data structure, hypotheses about the number of terror events that occur in a given country in a given year. Cross-national statistical research on the use of terror focuses almost exclusively upon structural characteristics of economies, polities, and societies (e.g., Lai, 2003; Abadie, 2004; Li, 2005; Piazza, 2006; Lai, 2007; Piazza, 2008b; Chenoweth, 2010; Caruso and Schneider, 2011; Krieger and Meierrieks, 2011; Gassebner and Luechinger, 2011). 1 More than a decade ago the first author advanced the same point with respect to both non-violent and violent dissent: I contend that the parties to these conflicts... pay attention to the behavior of the other parties when forming their own strategic plans. Hence my critique of [earlier] studies is that they are national attribute rather than strategic behavior models (Moore, 1995, p. 131). In the intervening years strategic, rational choice models have become quite common in the study of violent dissent, and a large and vibrant literature on civil wars has blossomed. Yet cross-national analyses of civil wars continue to focus upon the structural characteristics of economies, polities, and societies not the tactical choices of states and dissidents (e.g., Hegre et al., 2001; Sambanis, 2001; Reynal-Querol, 2002; Fearon and Laitin, 2003; Collier and Hoeffler, 2004). This study is motivated by the cross-national literature s lack of attention to dissident groups adoption of terror tactics as a function of the coercive behavior of the state and the tactical choices made by other dissident groups. Krieger and Meierrieks (2011) and 1 The work of Walsh and Piazza (2010) is a notable exception. They focus on the relationship between government coercion and terror attacks and find that the former stimulates the latter. Enders and Sandler s work is another exception to this tendency, though their cross-national work on the number of terror events has focused on trends in the global system (e.g., Enders and Sandler, 1993, 1999; Sandler and Enders, 2004; Enders and Sandler, 2006). 1

4 Gassebner and Luechinger (2011) provide comprehensive reviews of this literature, the first conducting a thorough review of published findings and the second performing an Extreme Bounds Analysis (Leamer, 1985, 2008) of essentially the same variables described in the former. To move beyond this understanding we contend that researchers will do well to engage the widely recognized fact that terror is but one tactic that dissident groups use to press their claims against states. Existing large-n analyses fail to leverage this point to construct, and test, hypotheses about the behavioral covariaties of terror attacks. This study does so. Why is ignoring the behavior of governments and other dissident groups a problem for studies of the cross-national correlates of terror? First, terror attacks are a tactical choice: dissidents might instead use non-violent protest, non-terror violence (e.g., guerrilla attacks), or some combination of these. Studying the incidence of terror attacks absent information about groups use of other tactics may bias findings. Second, it is difficult to imagine that researchers believe state coercion has no impact upon terror attacks, yet nearly all studies fail to include government repression in their models. Lastly, the policy implications of existing studies are muted: structural characteristics are, by definition, difficult to change. Behavior, however, can be revised by choice. Each of these reasons suggests that the large-n, crossnational study of terror attacks will be enriched by the development of hypotheses about the impact of government and dissident behavior. A considerable body of theoretical and empirical research has established that states and dissidents respond to one another s conflictual behavior (e.g., Lichbach, 1987; Davenport, 1995; Tilly, 1995; Francisco, 1996; Moore, 2000; Thyne, 2006; Davenport, 2007; Shellman, 2009; Pierskalla, 2010; Ritter, 2013). 2 This should not be surprising if one assumes that actors 2 See, also, Gupta and Mundra (2005); Brym and Araj (2006) and Araj (2008) for accounts of the impact of Israeli government repression upon Fatah s, Hamas s and Islamic Jihad s suicide bombing campaigns, and Dugan and Chenoweth (2012) for all types of terror attacks in Israel. 2

5 behave strategically, that is they try to anticipate the behavior of other relevant actors. This means that scholars analyzing terror tactics should examine factors that influence dissidents expectations about the government s response to these tactics. For any dissident group these expectations will be influenced by 1) the behavior of the government towards all dissident groups, itself included, and 2) the behavior of other dissident groups towards the government. Unlike existing research on the annual incidence of terror attacks that occur in different countries we focus attention upon the resource allocation decision of dissident groups across tactics available to them, one of which is terror. A micro-foundation focus is unusual for a study of the cross-national incidence of terror, and we demonstrate that it produces an empirical payoff. As we explain in the following section, we begin by considering how a representative dissident group will shift resources from a different tactic toward terror as a function of information that is available to all political actors in the country. This theoretical orientation directs attention to (1) the behavior of the state and other dissident groups, and (2) the political institutions that would constrain the behavior of both the state and other dissident groups. Doing so permits us to develop several new hypotheses as well as integrate a number of existing hypotheses about the impact of political institutions upon the aggregate number of terror attacks countries will experience. At first blush it may appear that such an approach would suffer from the ecological fallacy problem, but this is not so. The reason is that rather than expect hypotheses made at one level to exhibit the same pattern at a more aggregated level, we explicitly describe what the microfoundational argument leads us to expect to observe at the country year level: our hypotheses describe the expected impact on the tactical decisions of a randomly selected group dissident, aggregated across all dissident groups. Naturally, the behavior of the state and other dissidents groups within political institutions are not the only variables that might influence the number of terror attacks countries will experience. We explored existing cross-national studies to identify the structural char- 3

6 acteristics of economies and societies that we would need to control for. That literature primarily uses transnational terror events as the dependent variable (e.g., Li, 2005; Piazza, 2006, 2008a; Gassebner and Luechinger, 2011; Krieger and Meierrieks, 2011). In their review of the large-n, cross-national literature Krieger and Meierrieks (2011, p. 19) could find only two studies that examine only domestic terror attacks, and three studies that pool both domestic and transnational attacks. The existence of the Global Terrorism Database (GTD; LaFree and Dugan, 2007a) makes an analysis of the covariates of all terror attacks (i.e., both domestic and transnational) possible, and though transnational terror attacks comprise only an estimated 15% of all attacks (LaFree and Dugan, 2007b), like others (e.g., Piazza, 2011) we rely upon those studies to identify plausible candidates for control variables. 1 Country Level Terror and Dissident Tactics We are interested in the total number of acts of terror committed by all of the dissident groups that exist in a given country, and thus use the term dissident to describe that set of all groups. Dissident groups are any collection of people who make public demands upon the government using means that are not sanctioned by the state as legitimate or routine. 3 We assume that dissident groups allocate their resources among three tactics: terror, violent non-terror, and non-violence. We rely upon Schelling s classic definition of terror as: violence intended to coerce the enemy rather than to weaken him militarily (Schelling, 1960, p. 17). To conceptually distinguish terror tactics from other forms of violence we focus upon whether the tactic would degrade the state s coercive capacity (non-terror violence), or whether it would not (terror). Terror tactics are acts of violence that target a victim for the purpose of imposing 3 Lobbying, voting, or writing to state representatives are legitimate and/or routine, and often are not public. Protest rallies, sit-ins, boycotts, and other forms of symbolic protest are non-violent dissident activities. 4

7 costs on a broader audience. That is, the purpose of the attack is less the direct damage it will do to the target s coercive capacity than the costs it imposes upon an audience that witnesses the attack. Non-violent dissident activity seeks to pressure the state to change policies via economic, political or social coercion, but does not use violence to produce that pressure (Stephan and Chenoweth, 2008; Chenoweth and Stephan, 2011). The total number of domestic terror attacks in a country year is, by definition, the sum of the terror attacks each group in that country commits in a given year. The argument above discusses the tactical choices of a single group. So to develop hypotheses about what we will observe at the country year level we need to move from the decisions of dissident groups to a national level event count. We build this argument by assuming that a common information set is available to governments, dissidents, and the people in society. Doing so allows us to sketch a story about how governments and dissidents revise their beliefs based on their interactions with each other as well as the interactions they observe between relevant actors. We contend that the proportion of resources devoted by a group to a particular tactic is influenced by the following elements of the common information set: (1) the state s use of violent repression and its non-violent coercive activity, (2) the amount of non-violent activity by other dissident groups, (3) the amount of violent (non-terror) activity by other dissident groups, and (4) the amount of terror activity by other dissident groups. Below we make arguments about how individual groups will respond to their beliefs, given that those beliefs are formed by consulting the common set of information available to all actors. Our lament concerning the literature on the use of terror is that it focuses on macro-level structure rather than the behavior of governments and dissidents. One consequence of this is that institutions are often treated as factors that simply produce demand for violent dissent rather than rules that influence choices by constraining, and creating expectations about, the behavior of actors in a strategic situation. The point is that it is not institutions per se that are generating terror; it is the expectations they create about the government s response 5

8 to, and the effectiveness of, various alternative tactics. With that in mind, we also advance arguments about how the institutional context in which governments and dissidents interact affects the choice of dissidents to use terror tactics. We focus on four democratic institutions (participation, contestation, veto points, and liberal institutions that protect freedoms of association, press, religion and speech), and also turn to a relatively neglected feature of domestic politics that we believe impacts the tactical choices of dissident groups: the size of the coercive bureaucracy. To move from arguments about a randomly selected group s resource allocation across tactics to expectations about the number of terror attacks we would observe at a country year level one must imagine calculating an aggregate measure of the tactical distribution of resources across all dissident groups in that country year. Either the total or the average would be serviceable. We expect that, ceteris paribus, as the total (or average) proportion of resources that dissidents in a given country commit to terror attacks in a given year, the larger will be the number of terror attacks that occur in that country year. The arguments we develop below, then, focus on the expected impact that changes in the information available to dissidents will have upon the proportion of group resources they would invest in terror tactics. 1.1 Government Coercion It is widely agreed that states coercive behavior influences the behavior of dissidents (e.g., Lichbach, 1987; Enders and Sandler, 1993). Many scholars contend that violent repression raises the costs of dissent (e.g., Gurr, 1970; Tilly, 1978), and some researchers have argued that when the cost of one tactic rises relative to others, dissidents will switch to the less costly tactic (e.g., Lichbach, 1987; Moore, 1998; Enders and Sandler, 2006). We noted above that we assume that several dissident groups exist and that to develop their beliefs about the relative costs of the tactics they will use public knowledge of government coercion toward 6

9 every group, not just coercion against their group. We thus hypothesize that state coercion will reduce the level of terror we observe. 4 As often happens in conflict literature, counter-arguments exist to support the conjecture that coercion stimulates, rather than reduces, dissent. These are grievance based arguments, and the authors who advance them do not dispute that coercion is costly to dissidents. Instead, they argue that coercion alienates people who would have been indifferent or supported the government. (e.g., Gurr, 1970; Rasler, 1996; Walsh and Piazza, 2010). Several scholars maintain that arbitrary targeting exacerbates this tendency (e.g., Mason and Krane, 1989; Mason, 2004; Kalyvas, 2006; Findley and Young, 2007), and Wilkinson (2001) in particular advocates that counter insurgency is a political, rather than military, struggle and that coercion is less effective than investigative police work and prosecution. Each of these arguments suggests that state coercion will increase the aggregate level of terror we observe. 1.2 Dissident Competition It seems plausible that the tactical choices of other dissident groups influence the choices made by a given dissident group. We have assumed that there is a common information set available to states, dissidents, and the people. We thus need to construct an argument about how the tactics other dissident groups use influence the tactical choices of a given group, and how that would affect the aggregate number of terror attacks we would observe at the country year level. The key is to search for or develop arguments that discuss the impact that the level of non-violent and violent dissent have upon a group s proclivity to turn to 4 Lichbach (1987) and Enders and Sandler (2006) contend that coercion targeting a specific tactic will lead to a reduction in that tactic, but a rise in alternatives. Because we cannot determine in our data what tactic a given act of coercion targeted, we cannot evaluate this more nuanced hypothesis. However, we contend that of the three tactics, terror attacks are most likely to draw a coercive response. That contention supports the hypothesis we advance above. 7

10 terror. By focusing exclusively on the impact of structural characteristics, existing large-n studies have not invested much attention in this issue. To turn the development of theory about terror attacks to the tactical choices of dissident groups, it is imperative to address this issue. A prominent argument in the literature on the use of terror by dissidents suggests that groups that compete for the allegiance of the same supporters can get into an outbidding process where each pursues more violent tactics in an effort to show that they are the strongest group (Tillion, 1960, pp. 50-2, Hutchinson, 1972, p. 391, Rabushka and Shepsle, 1972, Della Porta, 1995, pp , Brubaker and Laitin, 1998, p. 434, Bloom, 2004; Kydd and Walter, 2006). The implication is that terror attacks stimulate more terror attacks. Because we are aggregating the data at a coarse temporal level the hypothesis is not well captured by inclusion of a lagged term (i.e., T error t 1 ) in a regression. Instead, the implication is that there will be over-dispersion in the event count. Another implication of the outbidding argument is that violent (non-terror) dissent will stimulate terror attacks. 5 What impact will non-violent protest activity by other groups have upon a given group s use of terror? Non-violent protest relies on mass mobilization: many people need to participate (e.g., DeNardo, 1985; Chong, 1991; Schock, 2005; Chenoweth and Stephan, 2011). It is conceivable that the more non-violent protest occurs in society, the more successful that tactic will appear to be, thus making the relative cost of terror tactics rise. Thus we conjecture that non-violent protest may reduce the number of terror attacks during a country year. Once again, counter-arguments are available. Mullins and Young (2010) argue that the more common is anti-state protest and violence, the more legitimated all forms of anti- 5 As above, we are oversimplifying existing theories. Enders and Sandler (2006) have theorized about acts of terror, and Lichbach (1987) has theorized generally, that when government repression reduces the effectiveness of a tactic dissidents change their tactics. A superior version of this argument, then, claims that outbidding is more likely when terror is successful. Unfortunately, the data required to evaluate that hypothesis in a CNTS setting have not been collected. 8

11 state protest and violence become. They show that crime and even capital punishment are positively associated with terror events. In addition, expected utility arguments suggest that dissidents beliefs about the probability of success will influence the likelihood that they challenge the state. The more anti-state activity there is, the weaker the state appears to be, and that will increase groups beliefs about the likelihood of successfully challenging the state. If observing greater levels of non-violent dissent leads groups that have been using terror tactics to conclude that the state is weak, then that group might infer that more terror attacks will produce concessions. Thus, as with government coercion, we cannot rule out the possibility that non-violent protest may increase the aggregate number of terror attacks. 1.3 Institutions as Context As noted above, political scientists have shown great interest in the impact of political institutions upon the incidence of terror (e.g., Eubank and Weinberg, 1994; Eyerman, 1998; Pape, 2003; Li, 2005; Savun and Phillips, 2009; Chenoweth, 2010; Krieger and Meierrieks, 2011; Young and Dugan, 2011). Li (2005, pp ) provides a useful overview of arguments about the impact of press freedom, observance of human rights, executive constraints, and voting rules and party systems ability to manage competition and conflict. Here we add to our argument by considering how various democratic institutions, and the size of the coercive bureaucracy, influence the common information set available to dissident groups and increase/reduce the amount of resources these groups devote to terror tactics and, by way of our argument, increase/reduce the aggregate level of terror we observe Contestation and Participation Contestation and participation are the two dimensions of Dahl s (1971) conceptualization of polyarchy. Contestation refers to the legality of expressing preferences over policy and leadership that conflict with those of the current government, while participation is defined 9

12 as the proportion of the public that is allowed to take part in political processes. The more welcome competing points of view are within institutionalized political forums, the higher contestation is. The greater the proportion of the population allowed to engage in institutionalized political activity, the greater the level of participation. We add to Dahl s definitions the costs citizens incur for engaging in political activity. With respect to participation this means the average cost of participating in leadership selection. The lower this cost is, on average, the higher participation is. Contestation then relates to the cost of becoming a candidate for public office. The lower this cost is, on average, the higher contestation is. Allowance of widespread participation and easy access to public office are likely to reduce violent dissent, including acts of terror. First, high levels of each create a means by which groups can pursue political demands and have grievances assuaged (Gurr, 1970, pp ). Second, the existence of legal pathways for pursuit of policy and leadership change delegitimizes the use of violence as a tactic to pursue political ends (Gurr, 1970, pp ). Perhaps more importantly, dissident groups in societies where non-violent dissent is banned are likely to infer that non-violent dissent is unlikely to be effective. This is because the legality of an action correlates, by definition, with a government s willingness to use force to punish that action. Where peaceful means of expressing policy preferences have been outlawed they are likely to be met with repression and prove ineffective. If we believe that dissidents substitute violence for non-violence when the latter is met with repression (Lichbach, 1987; Moore, 1998), it follows that where non-violent dissent is illegal violent dissident tactics, terror included, will likely become attractive. We therefore expect that the aggregate number of terror attacks will be decreasing in contestation and participation. Again, counterarguments exist. In particular, Chenoweth (2010, p. 19) contends that political competition increases terrorist activity. The idea motivating this hypothesis is that mobilization is high in the most competitive regimes, thereby encouraging conventional and unconventional forms of political activity (p. 19). Chenoweth tests her hypothesis with 10

13 the ITERATE transnational terror attacks data using variables that measure both political competition and participation. Her results support the hypothesis: both competition and participation have a positive impact upon transnational attacks. Li (2005) also points out that increasing the number of people who can influence policy gives dissidents greater incentives to employ terror because it creates the expectation that the government will concede when faced with terror and reduces the costs of such tactics. It has the former effect because executives are more concerned with the welfare of the public when the public has some control over their tenure in office. It has the latter effect because both the number of potential targets and the difficulty of protecting them increase with the size of the government s winning coalition. 6 Thus, we have competing expectations regarding the effect of contestation and participation on terror attacks: higher levels of each will either increase, or decrease, the aggregate number of terror attacks Veto Players Veto players refers to the number of actors in the political system whose consent is needed to change policy (Tsebelis, 2002). As noted above, Tsebelis shows that the status quo will prevail longer the greater the number of veto players. The idea is straight forward: the more actors there are who can block change, the less likely change in policy becomes. There are two ways in which the number of veto points might influence terror attacks. First, it may be the case that a large number of veto points leads to deadlock (i.e., zero policy change), which would greatly diminish the ability of the government to actively address grievances in the population and thus may increase the violent dissent, including terror attacks (Young and Dugan, 2011). Second, it has been argued that executives facing many 6 Though Li s argument and empirical test focus on government constraints (i.e., veto players), the institutional feature driving this argument is the size of a government s winning coalition, which relates directly to participation. 11

14 veto players will be constrained in their ability to combat dissidents through curtailment of civil liberties and other repressive means, thus the state s response is expected to be relatively mild and, consequently, the use of terror becomes more likely (Crenshaw, 1981; Wilkinson, 2001; Li, 2005). Both of theses argument imply that terror attacks will increase with the number of veto players Association, Press, Religion, and Speech Liberal democratic ideals include freedom to associate, to publish ideas, to worship deities of choice, and speak publicly about politics. Arguments can be constructed for either a positive or a negative impact of these institutions on terror. On the negative side, scholars have argued that civil liberties produce responsive political systems that forestall the need to turn to terror tactics, and this hypothesis has found support in large-n studies of transnational terror (e.g., Krueger and Maleckova, 2003; Kurrild-Klitgaard, Justesen and Klemmensen, 2006; Krueger and Laitin, 2008). 7 Alternatively, others have argued, and found empirical support, for the argument that liberal institutions facilitate all types of dissent, including terror attacks (Eubank and Weinberg, 1994; Lai, 2007; Chenoweth, 2010). Given these arguments, we expect that the extent to which government institutions protect freedoms of association, press, religion, and speech will either increase, or decrease, dissidents use of terror tactics Coercive Capacity of the State One institution that has received little attention is the coercive bureaucracy. A monopoly on the legitimate use of force is generally thought to be one of the defining characteristics 7 Existing empirical work largely finds that liberal democratic institutions are associated with lower levels of transnational attacks (Krieger and Meierrieks, 2011, p. 11), but as noted above, these are a small portion of total attacks. 12

15 of a modern nation-state (Weber, 1964). Yet scholars studying dissident and government behavior have largely neglected this aspect of the state and focused instead on the autocratic/ democratic distinction which, while important, is often orthogonal to the coercive capacity of the state. That is, the strength of coercive institutions varies independently of political democracy and is something that deserves explicit attention. In line with our injunction to focus on factors that influence expectations about the behavior of others, we argue that the number of terror attacks a country experiences will be a function of the size of the state s coercive bureaucracy. One notable exception to the tendency to ignore coercive bureaucracies is Gurr (1988). He proposes a model in which the coercive institutions of the state may become deeply embedded over time. The argument is that states which have used coercion successively will be more likely to employ it in the future. Also, states that successfully use coercion will continue to invest in the bureaucratic structures they created to prosecute coercion. We argue that where coercive bureaucracies are large dissidents will be hesitant to resort to violent tactics, including terror tactics. Gurr s argument implies that a large coercive bureaucracy indicates successful use of repression in the past and a high propensity to use repressive means in the present. 8 If we believe that dissidents consider the likely response of the government when choosing their tactics this implies that the size of the coercive bureaucracy should be negatively correlated with dissident violence, including terror tactics. Additionally, some have argued that a lack of coercive capacity is a precondition for, or at least contributes to, the outbreak of large-scale, violent rebellion (e.g., Skocpol, 1979; Fearon and Laitin, 2003). If terror tactics are viewed as merely one possible manifestation of violent dissent then coercive capacity should have some impact on their use. These arguments lead to the expectation that as the capacity of the coercive bureaucracy increases, the aggregate number of terror attacks decreases. 8 See Blomberg, Hess and Weerapana (2004b) for a similar claim. 13

16 1.4 The Socio-Economic Context, and Other Important Factors To this point we have restricted our attention to the dissident sector s tactical choices as a response to the state s behavior within the context of political institutions. We complete our specification by accounting for the broader economic and societal factors shown to be important by previous work. Macroeconomic performance is associated with grievances: the lower the level of macroeconomic performance, the greater the average level of grievances among people in a country. While grievances are not sufficient to produce dissident activity, the intensity and scope of grievances should covary positively with the amount of dissident activity, including acts of terror (e.g., Gurr, 1970; Crenshaw, 1981; Blomberg, Hess and Weerapana, 2004b,a). 9 An opportunity cost argument is also available to motivate this claim: the weaker is macroeconomic performance, the lower is the average opportunity cost of devoting labor to challenging the state (e.g., Collier and Hoeffler, 2004; Caruso and Schneider, 2011). That said, Krieger and Meierrieks (2011, p. 10) report that studies have found little empirical support in favor of a relationship between macroeconomic performance and transnational terror attacks. Given that these results might be a function of the focus on transnational terror attacks, however, we include macroeconomic performance as a control variable. Turning to the structural characteristics of the society, a vibrant civil society tends to be associated with a well-functioning state and so should be associated with less violent dissent. That is, the more vibrant the level of civic association is within a state, the better the state will function as a whole (Putnam, Leonardi and Nanetti, 1993; Putnam, 2000). Putnam argues that civil association is positively correlated with tolerance and equality, thus we would expect higher levels of civil society to lead to lower levels of violence. There is, however, a dark side to civil society, (Levi, 1998; Putnam, 2000; Uslaner, 2002). It is true that more associational ties within a group strengthen intra-group trust, 9 Though see Eyerman (1998) and Abadie (2004) for contrary evidence. 14

17 but this can also lead to a weakening of the ties between groups. That is, high levels of intra-group association can lead to more inter-group violence. Thus, the strength of civil society might increase, or decrease, the number of terror attacks that occur in a country. Ethno-linguistic composition refers to how heterogeneous the population of a state is. There is a large literature on the effect of ethnic heterogeneity on large-scale, violent civil conflict (Horowitz, 1985; Posen, 1993; Ellingsen, 2000; Sambanis, 2001; Reynal-Querol, 2002; Fearon and Laitin, 2003; Habyarimana et al., 2007). While the theoretical links between ethnic heterogeneity and conflict remain somewhat obscure (Saideman, 2010) and the empirical findings concerning this relationship are mixed, we conjecture that if ethnic diversity is thought to be a cause of violence between governments and dissidents then we should not rule out by fiat the possibility that it will exhibit some relationship with terror attacks (Krieger and Meierrieks, 2011, p. 12). 10 Physical quality of life is expected to exhibit a negative relationships with terror attacks. This variable has been shown to be positively correlated with state failure (Goldstone et al., 2010), and Piazza (2008b) finds state failure has a positive impact on terror attacks. As such, the higher the level of quality of life, the lower we might expect terror incidence to be. Also, the size of the population has been shown to have a positive impact upon terror attacks (Krieger and Meierrieks, 2011; Gassebner and Luechinger, 2011). Finally, a state s involvement in external conflict may affect the occurrence of terror attacks (Drakos and Gofas, 2006). Involvement in war tends to produce competition and conflict at home (Stohl, 1976, 1980; Rasler, 1986; Olzak, 1992). 10 Piazza (2011) reports the interesting finding that countries in which ethnic minority groups are economically marginalized are considerably more likely to experience domestic terror attacks than countries in which any ethnic minority groups do not experience such discrimination. 15

18 2 Statistical Analyses Testing the conjectures above requires measures of government coercion, violent and nonviolent dissent, contestation, participation, veto points, freedoms of association press religion speech, the size of the coercive bureaucracy, and the five socio-economic structures discussed above. A number of challenges confront cross national empirical work of this type: noisy and missing data; complex concepts that are not easily measured with one indicator and/or an absence of direct measures of concepts, and cross-sectional v. temporal co-variation (Zorn, 2001; Green, Kim and Yoon, 2003). To minimize the threat to our inferences that these issues raise we adopt a multi-pronged approach which employs techniques that are not individually novel, but have not, to the best of our knowledge, been used in combination before. In order to make the analysis accessible we limit discussion of technical details in the body of the study. The Supporting Information to the study contains those details. 2.1 Noisy and Missing Data As is common in CNTS datasets, ours is full of noisy indicators, many of which have missing values. For these reasons we employ Bayesian factor analysis to estimate several of our concepts. We describe our operational indicators in the Supporting Information. Here we describe the approach we took to construct the latent measures. Bayesian measurement models are not new to comparative/international politics research (e.g., Treier and Jackman, 2008). These models have two distinct advantages over classic factor analytic techniques they directly estimate the latent concept and provide measures of uncertainty for these estimates. When using these latent variables in subsequent predictive models, this uncertainty propagates directly into estimation. Classic approaches to measurement do not directly estimate values of the latent concept. Rather, they recover values of the latent variable indirectly from correlations among the observed indicators. Since these 16

19 values are estimates, treating them as observed overstates one s confidence in their values when including them as independent variables in subsequent predictive models. 11 Yet another advantage of the Bayesian factor model is the ability to avoid listwise deletion. Bayesian factor models permit one to use differing numbers of observed indicators for each unit (i.e., countries). Whereas listwise deletion produces biased sets of data, we are able to retain every observation for which we have values for most, but not all, of our variables. This is particularly helpful when one s data includes country-years from the developing world and/or new countries that do not cover the entire time-span of the analysis (e.g., new states born upon the collapse of the Soviet Union). For example, if we have four indicators of Civil Society of country A and six indicators for country B, the Bayesian model will yield estimates of the latent trait for both countries. The estimates will almost certainly have larger standard errors for the country with fewer indicators, which is a desirable result since we should be more uncertain when we have less information. The classic approach would, in most instances, delete all cases with missing values on any observable indicators, which would potentially create sample selection bias, among other problems. In fact, for our sample, a classic measurement approach would discard nearly all of the observations, as we illustrate below. For a list and description of our operational indicators please see the Supporting Information. One last comment about the latent measurement models warrants brief mention. It has become standard in cross-national research to provide a series of robustness checks. These checks largely evaluate the robustness of inferences to changes in operational indicators. Note that using latent measurement models is a viable alternative to robustness checks, especially if one has a number of possible indicators for several different concepts, as we do here. Gassebner and Luechinger (2011) conduct an analysis using every possible combination of indicators that have been utilized in cross-national studies of terror attacks. Few scholars 11 For a detailed discussion of Bayesian measurement models see Jackman (2009). 17

20 are willing to take on such a large effort, and latent measurement models offer an intriguing alternative to doing so. 2.2 Cross-sectional Covariation Overdispersion is a common feature of count data in the social sciences. Specifically, we often encounter count data in which the variance of the count increases with the level of the count. This makes intuitive sense in that cases with, on average, small counts have smaller variances than cases with larger counts. Often many cases in one s data will experience zero or very few events while some experience a large number of events. As noted above, our argument leads us to expect overdispersion: at the annual level of temporal aggregation outbidding among groups that use terror tactics will produce overdispersion in event counts of terror attacks. Visual inspection of Figures 3 and 4 in the Supporting Information indicates that our data are characterized by overdispersion, but there are more precise methods than visual inspection to evaluate the hypothesis. In order to model overdispersed count data, researchers commonly employ the negative binomial model. This model allows the variance of the dependent variable to be bigger than the mean by including an additional parameter relative to the Poisson model. When the data are conditional exchangeable 12 and overdispersed, this model is usually superior to the Poisson in terms of model fit and predictive power. In our data, however, we have the additional complication of panel-specific overdispersion. That is, not only are the data overdispersed, but the degree of overdispersion varies by country. To model this countryspecific overdispersion we estimated a Poisson-lognormal mixture model with random effects for overdispersion. The Poisson-Lognormal model allows us to estimate the expected number of counts per country-year as a multiplicative function of the covariates described above and 12 Conditional exchangeability refers to the assumption that, conditional on the covariates in the model, the errors are independent of one another (Lindley and Novick, 1981). 18

21 a country-specific, normally distributed random effect. This model has the added virtues of being relatively simple to estimate in the Bayesian context and is generally preferable to the negative binomial (Cameron and Trivedi, 1998, pp Winkelmann, 2008, pp. 134). We show below that it yields a considerable improvement in fit over the Poisson or a negative binomial with a single overdispersion parameter. 13 Finally, the presence of covariates that do not vary across time, or vary very little, within a country poses one additional complication. Multilevel modeling is an increasingly popular regression technique in political science (Steenbergen and Jones, 2002), particularly within the context of the CNTS data, and the Bayesian approach is highly amenable to such models (Gelman and Hill, 2006). Our model, then, includes several covariates that vary across both time and space, and several that are time-invariant. The predictive model, then, is: y it P oisson(λ it ) log(λ it ) = α i + X it β + Z it γ + ɛ it α i Normal(µ i, σ 2 ) µ i = α 0 + W i δ + υ i Where the α i s are the country-specific overdisperson parameters, X is a matrix of latent variables, Z is a matrix of observed variables and ɛ is a vector of errors. In the second-level equation, the α i s are modeled as a function of time-invariant country-level characteristics, W i and a country-specific error υ i. During estimation values of X, the set of latent variables, are drawn from their respective posterior distributions from the measurement models. This allows us to include in our predictive estimation the uncertainty in the latent variable esti- 13 Given that the random effect is an additive combination of information not directly included in the model, the assumption of normality is theoretically more defensible. In the Supporting Information we briefly discuss why the Poisson lognormal is preferred to the negative binomial, and report results obtained from estimating a negative binomial regression. 19

22 mates. Further technical details of the model and estimation are provided in the Supporting Information. 3 Results from the Statistical Models We present our results in two subsections, the first of which presents the coefficient estimates. We then discuss the substantive effects. The predictive model is estimated via Monte Carlo Markov Chain (MCMC) in WinBUGS (Lunn et al., 2000). Once the model converged, we then took the means and standard deviations of the posterior distributions of the regression coefficients and random effects. 14 These quantities are summarized in Table 4. The model ran for 50, 000 iterations and the first 10, 000 were discarded as burn-in. The model showed signs of convergence according to a variety of diagnostics available in WinBUGS and the Convergence Diagnosis and Output Analysis (CODA; Plummer et al., 2006) software. 3.1 Coefficient Estimates Figure 1 presents estimates (as dots), and 95% credible intervals (as lines), for the coefficients from our model. Given that the model is non-linear, the effects of the covariates on the expected counts of terror attacks depend on the values of the other covariates. That said, these coefficients do indicate the mean effect of a one-unit change in X on the log of the dependent variable, holding all other variables constant. All four of the Behavior variables, two of the Institutions variables, and two of the five socioeconomic structure variables have posterior densities for which at least 95% lies above or below zero. 15 Six variables appear not to influence the incidence of terror attack: 14 These are the Bayesian analogues to the frequentist coefficient estimates and standard errors. 15 These coefficient estimates are of limited interest. Because we adopt a Bayesian, rather than a frequentist, approach to inference in this study, we do not perform null hypothesis tests (see Gill, 1999), though readers who wish to do so can use the credible intervals (CIs) in Table 4 in the same manner one uses confidence 20

23 Figure 1: Results, Poisson-Lognormal Model Behavior: Govt Coercion Non violent Dissent Violent Dissent Int'l Violence Institutions: Contestation Participation Veto Assoc/Pres/Rel/Speech Coercive Bureaucracy Structure: Civil Society Econ Performance Ethno Ling Frac Population Phys Qual of LIfe Coefficient Estimates 21

24 Participation, Veto Players, the basket of civil liberties, Macroeconomic Performance, Ethnolinguistic Fractionalization, and Physical Quality of Lifes. 16 All but two of the variables that do have an impact increase the annual incidence of terror: both International Behavior and the Size of the Coercive Bureaucracy are associated with lower numbers of terror attacks. Government Coercion, Violent Dissent, Non-violent Dissent, Contestation, Strength of Civil Society, and the Size of the Population are all positively associated with the number of terror attacks that occur in a country. 3.2 Substantive Effects We first evaluate both the blocks of variables (i.e., Behavior, Institutions, and Socioeconomic Structures), thus examining the relative impact that each block of variables has upon model fit, and then the impact of individual variables. We present Deviance Information Criterion (DIC) for each of four model specifications, the full model (all blocks), the behavior only, institutions only and structures only models. DIC is an appropriate measure of model fit for comparing nested or non-nested models, particularly when estimating multi-level models via MCMC. Similar to the AIC, models with smaller DICs are prefered in terms of there ability to accurately reproduce the observed data (Spiegelhalter et al., 2002). The results in Table 1 show that the full model offers a large improvement in fit over the null model. More importantly, the results demonstrate that, in comparison to the null model, the block of dissident and state behavior variables make the largest contribution to explaining variance in the number of terror attacks each country experiences, that the Political Institutions block has the second largest influence, and that the block of Socioeconomic Structures has the smallest. Notice that the model using the structural block only intervals in a frequentist null hypothesis test. 16 It should be noted that for some of these variables, for example Veto and civil liberties, 90% of the posterior distribution lies above zero. 22

25 Table 1: DIC for Variable Blocks Model DIC Random Intercepts 80,412 Structural Block 80,527 Institutions Block 78,701 Behavior Block 73,653 Full Model 71,966 actually fits the data slightly worse than a model with only random intercepts. This is powerful evidence in favor of the focus upon dissident and state behavior that we advocate in this study. To evaluate the impact of individual variables we set all variables to their observed value for that case and then increased a given X by one standard deviation, 17 and examined the resulting change in the predicted count. It is common, when calculating marginal effects in non-linear models, to hold the values of all other independent variables at a fixed value. Rather than hold the other independent variables constant at a fixed value, however, we use the observed values for each of the other variables. Doing so allows us to conduct the type of counter-factual analyses one often thinks of: how many fewer/greater terror attacks would we expect if we increased the value of X in country A? 18 Conventional statistical models estimate only the mean effect of each X upon the dependent variable, and a confidence interval around that mean, holding the others Xs constant. In Bayesian statistics one estimates the full distribution of the impact of each X upon Y, holding the other X s constant. We depict in Figure 2 the means of those distributions, with 95% credible intervals, across all country years This is true for all variables except our measure of International Violence, which is a binary variable. We changed it from 0 to We describe in the Supporting Information the simulation approach we used to calculate these effects. 19 Readerswho prefer a tabular summary will find one in the Supporting Information. 23

26 Figure 2: Substantive Effects, Poisson-Lognormal Model Behavior: Govt Coercion Non violent Dissent Violent Dissent Int'l Violence Institutions: Contestation Participation Veto Assoc/Pres/Rel/Speech Coercive Bureaucracy Structure: Civil Society Econ Performance Changes in Expected Counts 24

27 3.2.1 Behavior Variables Figure 2 indicates that a one standard deviation increase in government coercion produces an average (across all country years) of seven additional terror attacks. Keep in mind that our measure reflects the overall level of government coercion toward society. This finding is consistent with that of Walsh and Piazza (2010), with the arguments made by researchers such as Rasler (1986) and Mason (2004) that the extent to which the government represses a population has more of a stimulative, than deterrent, impact upon violent dissent generally, and Wilkinson s (2001) argument about terror specifically as well as the case study work of Gupta and Mundra (2005); Brym and Araj (2006); Araj (2008). Non-Violent dissent also has a significant, positive impact on the use of terror tactics. Across the full sample an increase of one standard deviation produced an average increase of roughly one terror attack (see Figure 2). This suggests that non-violent dissent is a complement to, not a substitute for, terror attacks, and supports the argument that groups that adopt terror tactics will perceive states as vulnerable as the number of non-violent protests rise, and commit more attacks. Violent dissident activity also has a positive impact upon terror attacks, while international conflict is associated with fewer attacks Political Institutions We now turn to an examination of the effects of political institutions. Three of the five variables exhibit a substantial impact on the use of terror, though only contestation has greater than a one event impact. Figure 2 indicates that contestation is associated with an expected increase in the number of terror attacks, offering further support for Chenoweth s (2010) argument that political competition stimulates all types of political activity, including the use of terror tactics. Increasing the value of Contestation by one standard deviation, while holding all other variables at their observed values, produces an average increase in the number of terror 25

28 attacks of almost three. This result cuts against the more commonly advanced hypothesis that institutions which lower the general public s costs of entry into the political arena reduce grievances and discourage the use of violent tactics to promote policy change. Participation, on the other hand, does not have a consistent impact. Turning to Veto, Figure 2 indicates that increasing the value of the veto points variable by a standard deviation increases the expected number of terror attacks by about 1, an effect consistent with Li (2005) and Young and Dugan (2011). Like participation our measure of freedoms of association, press, religion, and speech does not have a consistent impact on the number of terror attacks a country experiences. Our final institutional variable, capacity of the coercive bureaucracy, exhibits a negative, deterrent relationship with the decision of dissidents to employ terror tactics: the mean across all country years is an expected reduction of one terror attack. This provides support for our conjecture, based on Gurr s (1988) argument concerning the rise of garrison states, that a large coercive bureaucracy indicates that past uses of state repression were successful in quelling dissent and that the state is likely to respond to dissent with repression in the present, and should thus discourage the use of terror by dissidents Structural Characteristics of Economy and Society As noted above, two of the five socio-economic structure variables were found to exhibit no systematic relationship with the number of terror attacks occurring within a given country year: Ethno-linguistic Fractionalization and physical quality of life. Neither of those variables vary substantially over time, and we thus used them as level two variables to model our random, country-specific intercepts. One of the socio-economic structural variables that did change over time exhibit a positive association with the number of expected terror attacks. Figure?? shows that a one standard deviation increase in the Strength of Civil Society produces roughly two additional terror attacks, on average, which supports the dark side 26

29 of civil society view advanced by Levi (1998); Putnam (2000) and Uslaner (2002). A one standard deviation increase in Macroeconomic Performance, however, produces neither a consistently positive nor negative impact. 3.3 Summary Our global analysis of the co-variates of terror attacks demonstrates the value of theorizing about the choice to adopt terror tactics as a process influenced by the states coercive behavior and the other violent and non-violent dissident behavior within that country. We have done so taking into account the context of both political institutions and socio-economic structures. An evaluation of the impact of each set of variables demonstrates that behavior explains more of the cross-national variation in terror attacks than do either institutions or structural characteristics. Turning to specifics, we find that government coercion tends to stimulate terror attacks as do other acts of dissent (both violent and non-violent), and that there is overdispersion consistent with the outbidding dynamic. Turning to political institutions, which have been of considerable interest to political scientists, our findings indicate that electoral contestation is positively associated with terror attacks, and veto points exhibits a smaller positive associations with terror attacks. Coercive bureaucracy, on the other hand, has a negative, deterrent, impact. Lastly, with respect to the most studied covariates structural characteristics of the economy and society we found that both the strength of civil society and population are positively associated with the incidence of terror. Having reviewed our findings we hasten to add that this study represents a baseline or beginning. Krieger and Meierrieks (2011) and Gassebner and Luechinger (2011) provide very valuable analyses of the covariates of (trans national) terror, but those studies focus upon strucutral characteristics of the economy and society and, to a lesser extent, political institutions, ignoring the coercive behavior of states, the behavior of other dissident groups, and the capacity of the coercive bureaucracy. As both of those studies summarize existing 27

30 cross-national statistical work they merely reflect the state of the field. We justify our claim that this study sets a baseline, then, on the grounds that we have not only argued that the study of cross-national terror attacks can be enriched by theoretically placing the decision to carry out such attacks within the context of the behavior of states and dissidents, but marshalling statistical evidence that strongly supports our theoretical claims. That said, we trust that readers will readily identify opportunities to both challenge and extend our arguments and findings, and we welcome such critique and development within this literature. 4 Implications & Conclusion In concluding we reflect upon this study within the context of the broader disaggregation movement in studies of intranational violent conflict. During the past five plus years there has been a considerable increase in studies that either spatially or temporally disaggregate the country year unit of observation and conduct statistical hypothesis tests in their study of violent conflict (e.g., see Holmes, Piñeres and Curtin, 2006; Cederman and Gleditsch, 2009; Raleigh and Hegre, 2009; Sánchez-Cuenca, 2009; Balcells, 2011). This is important work: while there is value in cross-national studies such as we have reported here, it is equally important that we study the processes that produce terror attacks (and other forms of political violence) as they unfold over time and across space. The year is not a special unit of time, nor is the country a special unit of space. Data are frequently collected in such a way that countries and years are convenient as units of observation, but more and more scholars are becoming familiar with the collection, and analysis, of events data that can be assembled using a variety of spatial and temporal units of observation (e.g., GDELT, see Leetaru and Schrodt, 2013). Further, prominent arguments in the theoretical literature imply that analyzing information about individual dissident groups, rather than information aggregated across all groups in the country, is crucial for understanding and explaining 28

31 terror attacks. We believe that the findings from studies such as this are most valuable as a baseline against which findings from spatially and/or temporally disaggregated studies can be compared. This is especially important if we are to develop a meaningful capacity to inform policy debates. We believe that this study can provide a valuable baseline for such research, and we hope others will find our case compelling. 29

32 References Abadie, Alberto Poverty, Political Freedom, and the Roots of Terrorism. In NBER Working Papers National Bureau of Economic Research. Abouharb, M. Rodwan and Annessa L. Kimball A New Dataset on Infant Mortality Rates, Journal of Peace Research 44(6): Aitchison, John and CH Ho The Multivariate Poisson-log Normal Distribution. Biometrika 76(4): Araj, Bader Harsh state repression as a cause of suicide bombing: the case of the Palestinian Israeli conflict. Studies in Conflict & Terrorism 31(4): Balcells, Laia Continuation of Politics by Two Means: Direct and Indirect Violence in Civil War. Journal of Conflict Resolution 55(3): Bank, World World Bank Development Indicators.. Banks, Arthur S Cross-National Time-Series Data Archive.. Binghamton: Center for Social Analysis. URL: Bernhard, Michael, Timothy Nordstrom and Christopher Reenock Economic Performance, Institutional Intermediation and Democratic Breakdown. Journal of Politics 63(3): Blomberg, S. Brock, Gregory D. Hess and Aakila Weerapana. 2004a. Economic conditions and terrorism. European Journal of Political Economy 20(2): Blomberg, S. Brock, Gregory D. Hess and Aakila Weerapana. 2004b. An economic model of terrorism. Conflict Management and Peace Science 21(1):

33 Bloom, Mia M Palestinian Suicide Bombing: Public Support, Market Share, and Outbidding. Political Science Quarterly 119(1): Brubaker, Rogers and David D. Laitin Ethnic and Nationalist Violence. Annual Review of Sociology 24(1): Brym, Robert J. and Bader Araj Suicide bombing as strategy and interaction: The case of the second intifada. Social Forces 84(4): Cameron, A.C. and PK Trivedi Regression analysis of count data. Cambridge Univ Press. Caruso, Raul and Friedrich Schneider The Socio-Economic Determinants of Terrorism and Political Violence in Western Europe ( ). European Journal of Political Economy 27:S37 S49. Cederman, Lars Erik and Kristian S. Gleditsch Introduction to Special Issue on Disaggregating Civil War. Journal of Conflict Resolution 53(4): Chenoweth, Erica Democratic Competition and Terrorist Activity. Journal of Politics 72(1): Chenoweth, Erica and Maria J. Stephan Why civil resistance works: The strategic logic of nonviolent conflict. Ithaca: Cornell University Press. Chib, Siddhartha and Rainer Winkelmann Markov Chain Monte Carlo Analysis of Correlated Count Data. Journal of Business & Economic Statistics 19(4): Chong, Dennis Collective Action and the Civil Rights Movement. Chicago: University of Chicago Press. 31

34 Cingranelli, David L. and David L. Richards The Cingranelli-Richards (CIRI) Human Rights Dataset.. URL: Clague, Christopher, Philip Keefer, Stephen Knack and Mancur Olson Contract- Intensive Money: Contract Enforcement, Property Rights, and Economic Performance. Journal of Economic Growth 4(2): Collier, Paul and Anke Hoeffler Greed and grievance in civil war. Oxford Economic Papers 56(4): Crenshaw, Martha The Causes of Terrorism. Comparative Politics 13(4): Dahl, Robert Polyarchy: Participation and Opposition. New Haven: Yale University Press. Davenport, Christian Multi-Dimensional Threat Perception and State Repression: An Inquiry Into Why States Apply Negative Sanctions. American Journal of Political Science 39(3): Davenport, Christian State Repression and Political Order. Annual Review of Political Science 10:1 27. Della Porta, Donatella Social Movements, Political Violence, and the State: A Comparative Analysis of Italy and Germany. New York: Cambridge University Press. DeNardo, James D Power in Numbers: The Political Strategy of Protest and Rebellion. Princeton: Princeton University Press. Drakos, Konstantino and Andreas Gofas In Search of the Average Transnational Terrorist Attack Venue. Defence and Peace Economics 17(2):

35 Dugan, Laura and Erica Chenoweth Moving Beyond Deterrence The Effectiveness of Raising the Expected Utility of Abstaining from Terrorism in Israel. American Sociological Review 77(4): Ellingsen, Tanja Colorful Community or Ethnic Witches Brew? Multiethnicity and Domestic Conflict During and After The Cold War. Journal of Conflict Resolution 44(2): Enders, Walter and Todd Sandler The Effectiveness of Antiterrorism Policies: A Vector-Autoregression-Intervention Analysis. American Political Science Review 87(4): Enders, Walter and Todd Sandler Transnational Terrorism in the Post-Cold War Era. International Studies Quarterly 43(1): Enders, Walter and Todd Sandler The Political Economy of Terrorism. Cambridge University Press. Eubank, W.L. and L. Weinberg Does Democracy Encourage Terrorism? Terrorism and Political Violence 6(4): Eyerman, Joseph Terrorism and Democratic States: Soft Targets or Accessible Systems? International Interactions 24(2): Fearon, James and David Laitin Ethnicity, Insurgency and Civil War. American Political Science Review 97(1): Findley, Michael G. and Joseph K. Young Fighting Fire with Fire? How (Not) to Neutralize an Insurgency. Civil Wars 9(4): Francisco, Ron Coercion and Protest: An Empirical Test in Two Democratic States. American Journal of Political Science 40(4):

36 Freedom House Freedom in the World. New York: Freedom House. Gassebner, Martin and Simon Luechinger Lock, stock, and barrel: a comprehensive assessment of the determinants of terror. Public Choice 149(3): Gelman, Andrew and Jennifer Hill Data Analysis Using Regression and Multilevel/Hierarchical Models. New York: Cambridge University Press. Ghosn, Faten, Glenn Palmer and Stuart A. Bremer The MID3 Data Set, : Procedures, Coding Rules, and Description. Conflict Management and Peace Science 21(2): Gibney, Mark, Linda Cornett and Reed M. Wood Political Terror Scale. Available online at: Acessed August Gill, Jeffrey The insignificance of null hypothesis significance testing. Political Research Quarterly 52(3): Goldstone, Jack A., Robert H. Bates, David L. Epstein, Ted Robert Gurr, Michael Lustik, Monty G. Marshall, Jay Ulfelder and Mark Woodward A Global Model for Forecasting Political Instability. American Journal of Political Science 54(1): Green, Donald P., Soo Yeon H. Kim and David Yoon Dirty Pool. International Organization 55(2): Gupta, Dipak K and Kusum Mundra Suicide Bombing as a Strategic Weapon: An Empirical Investigation of Hamas and Islamic Jihad. Terrorism and Political Violence 17(4): Gurr, Ted Robert Why Men Rebel. Princeton: Princeton University Press. 34

37 Gurr, Ted Robert War, Revolution, and the Growth of the Coercive State. Comparative Political Studies 21(1): Gurr, Ted Robert Minorities at Risk: A Global View of Ethnopolitical Conflict. Washington, D.C.: United States Institute of Peace. Habyarimana, James, Macartan Humphreys, Daniel N. Posner and Jeremy M. Weinstein Why Does Ethnic Diversity Undermine Public Goods Provision? American Political Science Review 101(4): Hegre, Håvard, Tanja Ellingsen, Scott Gates and Nils Peter Gleditsch Toward a Democratic Civil Peace? Democracy, Political Change and Civil War, American Political Science Review 95: Henisz, Witold J The Institutional Environment for Economic Growth. Economics and Politics 12(1):1 31. Heston, Alan, Robert Summers and Bettina Aten Penn World Table Version 6.2. Center for International Comparisons of Production, Income and Prices at the University of Pennsylvania. URL: Holmes, Jennifer S., Sheila Amin Gutiérrez Piñeres and Kevin M. Curtin Drugs, Violence, and Development in Colombia: A Department-Level Analysis. Latin American Politics and Society 48(3): Horowitz, Donald L Ethnic Groups in Conflict. Berkeley: University of California Press. Hutchinson, Martha Crenshaw The Concept of Revolutionary Terrorism. Journal of Conflict Resolution 16(3):

38 Jackman, Simon Bayesian Analysis for the Social Sciences. Vol. 846 Wiley. Jenkins, J. Craign and Charles L. Taylor The World Hanbook of Political and Social Indicators IV. Available at: Kalyvas, Stathis N The Logic of Violence in Civil War. Cambridge University Press. Keefer, Philip DPI 2004, Database of Political Institutions: Changes and Variable Definitions. Washington, DC: Development Research Group, The World Bank. variabledefinitions.pdf. Krieger, Tim and Daniel Meierrieks What Causes Terrorism? Public Choice. Krueger, Alan B. and David D. Laitin Kto Kogo?: A cross-country study of the origins and targets of terrorism. Terrorism, economic development, and political openness pp Krueger, Alan B. and Jitka Maleckova Education, Poverty, and Terrorism: Is There a Causal Connection? Journal of Economic Perspectives. Kurrild-Klitgaard, Peter, Morgens K. Justesen and Robert Klemmensen The political economy of freedom, democracy and transnational terrorism. Public Choice 128(1): Kydd, Andrew H. and Barbara F. Walter The Strategies of Terrorism. International Security 31(1): LaFree, Gary and Laura Dugan. 2007a. Global Terrorism Database, : Codebook and Data Documentation. Ann Arbor: Interuniversity Consortium for Political and Social Research. ICPSR Study #

39 LaFree, Gary and Laura Dugan. 2007b. Introducing the global terrorism database. Terrorism and Political Violence 19(2): Lai, Brian Examining the number of incidents of terrorism within states, paper presented at the American Political Science Association Meetings, Philadelphia, PA. Lai, Brian Draining the Swamp : An Empirical Examination of the Production of International Terrorism, Conflict Management and Peace Science 24(4): Leamer, Edward E Sensitivity analyses would help. American Economic Review 75(3): Leamer, Edward E Extreme Bounds Analysis. In The New Palgrave Dictionary of Economics, ed. Steven N. Durlauf and Lawrence E. Blume. London: Palgrave McMillan. Leetaru, Kalev and Philip A. Schrodt GDELT: Global Data on Events, Location and Tone, State College, PA: Penn State Events Data Project. URL: Levi, Margaret A State of Trust. Trust and governance 1: Li, Quan Does Democracy Promote or Reduce Transnational Terrorist Incidents? Journal of Conflict Resolution 49(2): Lichbach, Mark Deterrence or Escalation? The Puzzle of Aggregate Studies of Repression and Dissent. Journal of Conflict Resolution 31: Lindley, D.V. and M.R. Novick The role of exchangeability in inference. The Annals of Statistics 9(1):

40 Lunn, David J., Andrew Thomas, Nicky Best and David Spiegelhalter WinBUGSa Bayesian modelling framework: concepts, structure, and extensibility. Statistics and Computing 10(4): Marshall, Monty and Keith Jaggers Polity IV Project: Political Regime Characteristics and Transitions, Data Users Manual. Center for Systemic Peace. Available online at: pdf. Mason, T. David Caught in the Crossfire: Revolutions, Repression, and the Rational Peasant. New York: Rowman & Littlefield Publishers. Mason, T. David and Dale A. Krane The Political Economy of Death Squads: Toward a Theory of the Impact of State-Sanctioned Terror. International Studies Quarterly 33(2): Moore, Will H Action-Reaction or Rational Expectations? Reciprocity and the Domestic-International Conflict Nexus during the Rhodesia Problem. Journal of Conflict Resolution 39(1): Moore, Will H Repression and Dissent: Substitution, Context and Timing. American Journal of Political Science 45(3): Moore, Will H The Repression of Dissent: A Substitution Model of Government Coercion. Journal of Conflict Resolution 44(1): Mullins, Christopher W. and Joseph K. Young Cultures of Violence and Acts of Terror: Applying a Legitimation-Habituation Model to Terrorism. Crime & Delinquency forthcoming(0):

41 Olzak, Susan The Dynamics of Ethnic Competition and Conflict. Stanford: Stanford University Press. Pape, Robert A The Strategic Logic of Suicide Terrorism. American Political Science Review 97(03): Piazza, James A Rooted in Poverty?: Terrorism, Poor Economic Development, and Social Cleavages. Terrorism and Political Violence 18(1): Piazza, James A. 2008a. Do Democracy and Free Markets Protect Us From Terrorism? International Politics 45(1): Piazza, James A. 2008b. Incubators of Terror: Do Failed and Failing States Promote Transnational Terrorism? International Studies Quarterly 52(3): Piazza, James A Poverty, minority economic discrimination, and domestic terrorism. Journal of Peace Research 48(3): Pierskalla, Jan Henryk Protest, Deterrence, and Escalation: The Strategic Calculus of Government Repression. Journal of Conflict Resolution 54(1): Plummer, Martyn, Nicky Best, Kate Cowles and Karen Vines CODA: Convergence Diagnosis and Output Analysis for MCMC. R News 6(1):7 11. URL: Political Risk Group. nd. The International Country Risk Guide, Part II. New York:. busguide04.pdf. Posen, Barry The Security Dilemma and Ethnic Conflict. In Ethnic Conflict and International Security, ed. Michael Brown. Princeton University Press pp

42 Putnam, R.D., R. Leonardi and R.Y. Nanetti Making democracy work: Civic traditions in modern Italy. Princeton UniversityPress Princeton, NJ. Putnam, Robert Bowling alone. The collapse and revival of American community. New York: Simon and Shuster. Rabushka, Alvin and Kenneth A. Shepsle Politics in plural societies: a theory of democratic instability. Columbus, Ohio: Merrill Publishing Company. Raleigh, Clionadh and Hȧvard Hegre Population size, concentration, and civil war. A geographically disaggregated analysis. Political Geography 28(4): Rasler, Karen A War, Accommodation, and Violence in the United States, American Political Science Review 80(3): Rasler, Karen A Concessions, Repression, and Political Protest. American Sociological Review 61(1): Reynal-Querol, Marta Ethnicity, Political Systems, and Civil Wars. Journal of Conflict Resolution 46(1): Ritter, Emily Hencken Policy Disputes, Political Survival, and the Onset and Severity of State Repression. Jounral of Conflict Resolution forthcoming. Saideman, Stephen Personal Communication. Status Update post on Will Moore s FaceBook Wall, Monday 1 February, 7 am. Sambanis, Nicholas Do Ethnic and Nonethnic Civil Wars Have the Same Causes? A Theoretical and Empirical Inquiry (Part 1). Journal of Conflict Resolution 45(3):

43 Sánchez-Cuenca, Ignacio Explaining temporal variation in the lethality of ETA. Revista Internacional de Sociología 67(3): Sandler, Todd and Walter Enders An economic perspective on transnational terrorism. European Journal of Political Economy 20(2): Savun, B. and Brian J. Phillips Democracy, Foreign Policy, and Terrorism. Journal of Conflict Resolution 53(6): Schelling, Thomas C Arms and Influence. New Haven: Yale University Press. Schock, Kurt Unarmed insurrections: People power movements in nondemocracies. Minneapolis: University of Minnesota Press. Shellman, Stephen M Taking Turns: A Theory and Model of Government-Dissident Interactions. Saarbücken: VDM Verlag. Singer, J. David Reconstructing the correlates of war dataset on material capabilities of states, International Interactions 14(2): Singer, J. David, Stuart Bremer and John Stuckey Capability Distribution, Uncertainty, and Major Power War, In Peace, War, and Numbers, ed. B.M. Russett. Beverly Hills: Sage pp Skocpol, Theda States and Social Revolutions: A Comparative Analysis of France, Russia, and China. New York: Cambridge University Press. Spiegelhalter, D.J., N.G. Best, B.P. Carlin and A. van der Linde Bayesian measures of model complexity and fit. Journal of the Royal Statistical Society. Series B (Statistical Methodology) 64(4):

44 Steenbergen, Marco and Bradford S. Jones Modeling Multilevel Data Structures. American Journal of Political Science 46: Stephan, Maria J. and E. Chenoweth Why civil resistance works: The strategic logic of nonviolent conflict. International Security 33(1):7 44. Stohl, Michael War and Domestic Political Violence: The American Capacity for Repression and Reaction. Beverly Hills: Sage Publications. Stohl, Michael The Nexus of Civil and International Conflict. In Handbook of Political Conflict: Theory and Research, ed. T.R. Gurr. New York: The Free Press pp Taylor, Charles and David Jodice World Handbook of Political and Social Indicators III. New Haven: Yale University Press. Thyne, Clayton L Cheap Signals with Costly Consequences: The Effect of Interstate Relations on Civil War. Journal of Conflict Resolution 50(6): Tillion, Germaine Les Ennemis-Complémentaires. Paris: Éditions de minuit. Tilly, Charles From Mobilization to Revolution. Massachusetts: Addison-Wesley. Tilly, Charles State-Incited Violence, Political Power and Social Theory 9: Treier, Shawn and Simon Jackman Democracy as a Latent Variable. American Journal of Political Science 52(1): Tsebelis, George Veto Players: How Political Institutions Work. Princeton: Princeton University Press. Uslaner, E.M The moral foundations of trust. Cambridge University Press. 42

45 Van Belle, Douglas A Press Freedom and Global Politics. Westport: Praeger. Vanhanen, Tatu The Polyarchy Dataset: Vanhanen s Index of Democracy.. URL: detail//9649/42472.html Walsh, James Igoe and James. Piazza Why respecting physical integrity rights reduces terrorism. Comparative Political Studies 43(5): Weber, Max The Theory of Social and Economoic Organization. New York, NY: Free Press. Wiik, Robert Democratisation in the Name of Civil Society. A Quantitative Analysis of the Impact of Non-governmental Organisations on Democratisation PhD thesis. Wilkinson, Paul Terrorism Versus Democracy: The Liberal State Response. Second ed. London: Routledge. Winkelmann, R Econometric analysis of count data. Springer Verlag. Young, Jospeh.K. and Laura Dugan Veto Players and Terror. Journal of Peace Research 48(1): Zorn, Christopher Estimating Between and Within Cluster Covariate Effects, with an Application to Models of Inernational Disputes. International Interactions pp

46 Supporting Information for Moore, Bakker & Hill s How Much Terror? Dissidents, Governments, Institutions and the Cross-National Study of Terror Attacks May 15, 2013 This Supporting Information provides a considerable amount of detail and supporting information for our study. We first discuss the operational indicators and Bayesian measurement models we estimated. We then discuss the econometrics and estimation of our predictive model, and offer a considerably more detailed discussion of the substantive effects reported in the study. With respect to those substantive effects, in the study we focus on the sample mean of the substantive effects. Yet, our decision to model the overdispersion as a random effect permits us to produce country specific substantive effects, and below we explore those as well. Single Variable Indicators To measure five of our concepts we use a single indicator that was measured directly: the number of terror attacks; non-violent dissident activity; violent dissident activity; interstate violence; and ethno-linguistic fractionalization. Before estimating fully Bayesian measurement models to construct indicators for our concepts we first conducted confirmatory factor analysis, using 3 or more single indicators for each concept, to determine whether the relationships between the single indicators were strong enough to justify estimating a measurement model. Our rule for determining whether to proceed with a Bayesian model was as follows: if the Eigenvalue for the first common factor estimated was > 1 we estimated a Bayesian mea- 1

47 surement model. If this condition was not met we used a single indicator for that concept. 20 If a Bayesian model was estimated, we then conducted face validity checks by checking the simple correlation between the latent variable scores generated from the factor analysis and those generated from the Bayesian model. 21 Terror Events To measure the number of terror events observed in a given country in a given year we use the Global Terrorism Database (GTD 1.1) available through the Interuniversity Consortium for Political and Social Research (ICPSR). 22 Specifically, we used the GTD 1.1 events data to produce an event count of all types of terror events in each country-year. 23 Most studies of terror events have used the ITERATE data. We use GTD 1.1 because the ITERATE data code only transnational events, which they define as events where the perpetrators and victims are from different countries. Due to this emphasis, the ITERATE data exclude the entire category of domestic terror such as the Oklahoma City bombing, the 1986 Olympics bombing in Atlanta, and so on. The GTD 1.1 codes all terror events regardless of the nationality of the perpetrators and targets and is thus allows one to create a more valid variable for studying the production of terror events. Figures 3 and 4 use a dot to depict the average number of terror events that occurred in 20 Below we do not discuss measures that were considered for use but omitted based on initial factor analysis. 21 For factor scores from the Bayesian model we used the last draw from the posterior distributions of the latent variables. 22 The GTD 1.1 data are available online at: See the website at for a description of the larger project from which these data are developed. 23 The GTD allows one to break down terror events by a number of types. This analysis groups all types of events. 2

48 each country. The gray line depicts the range of the 5th to 95th percentile of the distribution of the number of events in that country. Because there are so many countries we divide the plots in two, the first one of which displays countries with greater than the median average number of terror events, and the second of which records the countries with less than the median average number of terror attacks per year. Figure 3: Mean # of Terror Events per Country, Upper Half (Non) Violent Dissident Activity To measure non-violent and violent dissident activity we initially sought to create latent measures based upon multiple indicators from the Cross National Time Series Archive (CNTSA, 3

49 Figure 4: Mean # of Terror Events per Country, Lower Half 4

50 Banks, 2001), the Minorities at Risk project (Gurr, 1993) and the World Handbook of Political Indicators III (Taylor and Jodice, 1983). Unfortunately, initial confirmatory factor analysis indicated that these indicators do not share much covariance, and so we created an index by summing the events that we categorized in each category. We use the domestic conflict event data which the CNTSA codes from the daily files of the New York Times. Two variables measure non-violent tactics: anti-government demonstrations 24 and general strikes. 25 We sum these two measures to create an index of non-violent dissent. We employ two measures of violent tactics from the CNTSA data : riots 26 and acts of guerrilla warfare. 27 We sum these two measures into a single index. 28 Interstate Violence To measure the government s violent conflict behavior vis-à-vis other states we use a dichotomous indicator of MID involvement drawn from the COW data (Ghosn, Palmer and Bremer, 2004). 24 CNTSA defines these as Any peaceful public gathering of at least 100 people for the primary purpose of displaying or voicing their opposition to government policies or authority, excluding demonstrations of a distinctly anti-foreign nature. The source for this, and the following, variable definitions is the Cross-National Time Series Data Archive Variable Location and Variable Descriptions electronic file. 25 General strikes are Any strike of 1,000 or more industrial or service workers that involves more than one employer and that is aimed at national government policies or authority. 26 Riots are Any violent demonstration or clash of more than 100 citizens involving the use of physical force. 27 Acts of guerrilla warfare are Any armed activity, sabotage, or bombings carried on by independent bands of citizens or irregular forces and aimed at the overthrow of the present regime. 28 The CNTSA data also contain a measure of assassinations which we omit from the summary index due to a relative lack of variation. 5

51 Ethno-linguistic Composition of Society Finally, to measure of ethno-linguistic composition of society we use the Ethno-Linguistic Fractionalization index as updated by Fearon and Laitin (2003). This variable ranges from 0 to 1 and measures the probability that two randomly drawn members of the population will be from the same ethno-linguistic group. Thus 0 represents perfect heterogeneity and 1 perfect homogeneity. Latent Measures Missingness For each country-year in our data, we collected a variety of observable indicators across a wide range of latent concepts. Figure 5 shows, for each country, the proportion of years that would be missing if using a classic measurement model (i.e., normal factor analysis). The map makes it clear that executing the analysis below would not be possible outside of a Bayesian framework: the classic model would retain, at best, roughly 3.5% of the observations, and that for only a handful of countries. Bayesian Estimation of Latent Variables The model used to estimate the latent measures is a Bayesian latent variable model in which the observed values y of the latent concept X are modeled with a multivariate normal distribution. All observed indicators were standardized and the prior distributions of the latent variables were assumed standard normal in order to identify the models. Standardizing the observed indicators eliminates the need for a constant, so a single coefficient β is estimated for each y in every model. In some cases restrictions were placed on the β parameters (i.e., restricting a single β to be positive or negative) in order to orient the latent trait in an intuitive fashion (i.e., where higher values of the latent variable indicate more of the trait 6

52 Figure 5: Proportion of Years With 1 Indicator Missing (Classic Measurement Models) 7