Does Money Matter? Terrorism and Income Distribution

Transcription

1 Does Money Matter? Terrorism and Income Distribution A Thesis submitted to the Graduate School of Arts & Sciences at Georgetown University in partial fulfillment of the requirements for the degree of Master of Public Policy in the Georgetown Public Policy Institute By Carroll Joseph Ganier, B.S.E.E. Washington, DC April 16, 2009

2 I would like to thank my thesis advisor, Professor Nada O. Eissa, for her guidance. ii

3 Does Money Matter? Terrorism and Income Distribution Carroll Joseph Ganier, B.S.E.E. Thesis Advisor: Nada O. Eissa, Ph.D. Abstract Do income contractions among higher income earners of a country explain increases in terrorist violence by nationals from that country? At the individual and national levels, poverty has been rejected as a determinant of terrorism, but questions about how wealth influences terrorism remain. I test whether terrorism by a country s nationals is associated with quintile income per capita in that country via regression on these variables and controls for political rights, geography, unemployment, linguistic fractionalization, and population. While I do find an association between first and second quintile income levels and terrorism, these associations are not controlled by inequality or poverty. iii

4 Contents 1 Introduction 1 2 Literature Review Rational Choice Model Conceptual Framework The Government Terrorist Organizations Potential Terrorists Additional Considerations Empirical Model Choice of Data and Variables Description of Datasets ITERATE World Bank University of Gothenburg: Quality of Government Center for International Development General Measures of Geography UNU-WIDER Descriptive Statistics Descriptive Statistics on Terrorism Descriptive Statistics on Independent Variables Regression Implementation Hypothesized Effects iv

5 5 Results Autocorrelation Diagnosis Probit Regressions Explanatory Interactions Conclusion 41 v

6 List of Tables 1 Casualties and fatalities from international terrorism by incident, Casualties and fatalities from international terrorism by country-year, Descriptive statistics of independent variables over 212 countries, Unknown attacker nationality by incident, Number of known attacker nationalities by incident, OLS regression of the sum of casualties and fatalities from terrorism Regression of residuals from equation?? on lagged residuals Probit of country s nationals conducting attack causing large property damage Probit of country s nationals conducting attack causing casualties or fatalities Prais-Winsten regression of the sum of casualties and fatalties from terrorism Continuation of Table??, Prais-Winsten regression of the sum of casualties and fatalties from terrorism Prais-Winsten regression of the sum of casualties and fatalties from terrorism Continuation of Table??, Prais-Winsten regression of the sum of casualties and fatalties from terrorism vi

7 1 Introduction Understanding the determinants of terrorism matters because the appropriate anti-terrorism policy prescriptions differ depending on the causes. Explanations of terrorism based on irrationality and psychosis imply terrorists who cannot be deterred. To defeat this kind of terrorism, governments might emphasize interdiction over economic development. Alternately, explanations of terrorists as opportunistic rent-seekers might make the case for aid to terrorism-fighting governments. Developing a theory of terrorism requires interrogating the mechanisms that facilitate terrorist organizations and the mechanisms of terrorist organizations themselves. Finding a strong association between a country s income (or distribution of income) and terrorism would indicate that economic development may be a useful tool to reduce terrorism or that education without development might be a contributing factor to the problem. Terrorist groups do not operate in isolation from the economy around them. They draw scarce resources from the environment and cannot be immune to the constraints that extend from that scarcity. Analogous income tradeoffs apply to potential terrorist recruits. In this thesis, I investigate whether income contractions in particular income quintiles explain terrorist violence by nationals from a country. The question is do income contractions among workers whose income is not at the bottom of the distribution, but rather above the median of income earners in a country, increase the likelihood of residents participation in terrorist violence? An important complication in terrorism research is defining terrorism. There is no agreed upon standard definition within terrorism research, and this lack of consensus is an issue for discussion itself within the field. There were over one hundred definitions of terrorism in 1988, and the subsequent proliferation of terrorism research has seen the creation of more [Gordon2004]. This paper does not put forth an alternative definition of interna- 1

8 tional terrorism, instead using the same definition as the terrorism dataset, ITERATE. I only consider international terrorism, solely because of data limitations. In addition, this thesis only addresses non-state sponsored terrorism because the suspected economic connection between middle income earners and terrorism would not be present for individual attackers who work on behalf of a state. In Section 2, I review the literature on the causes of terrorism and explain the warrant for using a rational choice model of terrorists. Section 3 describes the theoretical model of a terrorism labor market that forms the basis for my specification. I describe the datasets, descriptive statistics, and regression specification in Section 4. Section 5 presents the results of these regressions. The last section gives my conclusion. I find an association between income levels in the bottom two quintiles and terrorism, but I find no evidence that inequality or poverty have any effect on this association. I do not find evidence of an association between higher income in middle or upper-middle quintiles and less terrorism. 2 Literature Review Terrorists have diverse motivations and causes, and there are commensurately many explanations of the roots of terrorism. These explanations have developed over time away from a belief in the fundamental abnormality or irrationality of terrorists toward the belief in a more ordinary psychological, political, or economic subject. In spite of the violence of terrorism, attempts to develop a psychological type for terrorists have not been successful at finding common a profile with any serious pathology. Martha Crenshaw finds in her review of the research into terrorist psychology that most analysts of terrorism do not think that personality factors account for terrorists behavior [Crenshaw2000]. John Horgan reviews the psychological research into terrorism describing how the search for the terrorist pathological profile proceeded in the late seventies and 2

9 eighties but failed to identify a pathological type [Horgan2005]. Instead, Horgan cites the fact that terrorist organizations can be complex and require discipline and loyalty as evidence that the individual terrorists must be psychologically capable of performing in those organizations. While terrorists do not share a common psychological profile, terrorist organizations do show particular group traits. Fathali Moghaddam identifies isolation of the group from larger society as the one of the most important factors explaining a groups turn to violence. This factor is salient because it facilitates an extreme view of justice and the moral permissibility of violence. [Moghaddam2004]. A second path of explanation for terrorist motivation focused on recruits being motivated by their relative or absolute poverty. Researchers examined whether the lack of opportunity or a sense of outrage at deprivation might motivate terrorism. Krueger and Maleckova find no connection between low GDP per capita and terrorism nor any special effect of lack of political freedoms on poverty s link to terrorism [Krueger and Malečkova2003]. Blomberg, Hess, and Weerapana link macroeconomic downturns to terrorist attacks in democratic, high-income countries, but they do not find a link between poverty and terrorism [S. Brock Blomberg2004]. Fair and Shepherd examined the Pew Research Center s survey of Muslims in fourteen countries to search describe the support for terrorism found there [Fair and Shepherd2006]. They found that their proxies for poverty were associated with being less likely to support terrorism. At the level of the individual terrorist, terrorist organizations use psychological vulnerability and personal crisis as catalysts for recruitment rather than rather than economic distress [Pedahzur and Perliger2006]. While there are associations between the economic factors and terrorism, the actual terrorists are not the poorest and are not necessarily from the poorest countries. 3

10 2.1 Rational Choice Model Current microeconomic models of the terrorist and the terrorist organization approach them as rational actors. These models then try to layout the mechanisms influencing rational members of those organizations. Intuitively, the successful terrorist groups will be ones whose methods, members, and immediate goals depend on available resources. The rational actor model does not require any special proclivity for violence. Unfortunately, in almost all cases, the individual level data about the terrorist organizations is lacking, so these models go untested. Rational sympathizers might participate in a terrorist organization in exchange for the benefits of club goods. Iannaccone describes a model in which strict religious sects might attract members in the face of competition from more permissive religions [Iannaccone1992]. To deter shirkers, the club enforces sacrifices and religious proscriptions on its members. Eli Berman extends this idea specifically to terrorist groups where the shirker-deterring club selection process can serve the dual purpose of drawing lower wage, higher loyalty members into its ranks [Berman2003]. Violent collective action requires this kind of loyalty to succeed. Berman and Laitin use the same club model of terrorist groups to explain why Palestinian terrorists tend to use suicide attacks instead of conventional violence against risky targets [Berman and Laitin2005]. Joseph Felter notes that the desperate conditions and low social service provision in Palestinian territory makes Hamas membership an attractive proposition because of their provision of social services to members [Felter2006]. Social service provision by terrorist organizations can be seen not only as a tool to disguise or front for terrorist activities, but also as a tool to win the hearts and minds of the supporting population. Magourik describes one model of how social service provision signals to a target population that the terrorist group is not simply seeking economic rents from its violent struggle [Magouirk2008]. By devoting resources to community service, the 4

11 group demonstrates that its utility function includes the welfare of the community. By comparison, a group that does not devote resources to public service might be suspected of being as exploitive of the target population as the current regime. Others have looked at how terrorists might interact with the labor market both in recruiting in allocating labor within a terrorist organization. Efraim Benmelech and Claude Berrebi found that terrorist organizations show an awareness of human capital in their assignment of older and better educated terrorists to more difficult, high yield targets [Berrebi2007]. Berrebi uses individual level data of Palestinian terrorists to identify a positive association between education and terrorism, while poverty is negatively associated with terrorism [Berrebi2007]. Ethan Bueno de Mesquita posits a rational choice model with interactions between government, terrorists, and a recruitment population in which the terrorist organizations recruit the best members available from the population, while recruits balance traditional economic opportunities against antipathy for government oppression [Bueno de Mesquita2005]. The scarcity of information about how terrorist groups operate necessitates assumptions about what factors drive terrorist group behavior. Although many studies take terrorist financing as an exogenous influence on operations, terrorist resource gathering interests both government counter-terrorism agents as well as modelers of terrorist organizations. Hutchinson and O Malley considered the development of criminal fundraising capabilities by terrorist organizations as state-sponsorship of terrorism has declined [Hutchinson and O Malley2008]. Despite topological similarities between organized crime and terrorism, they believe that the difference in priorities between criminal organizations and terrorist organizations will limit the potential for cooperation. Enders and Sandler model terrorist group decision making using a household production function to estimate the substitution effect in terrorist decision making [Sandler2004]. They find evidence of 5

12 inter-temporal substitution by looking at the bursty pattern of terrorism between high and low periods surrounding the 1986 US bombings of Libya. The patterns of terrorism have shown evidence of a sensitivity to costs via substitution among modes of attack, but there remains considerable research to be done to understand how strong income and substitution effects would be for terrorist organizations. 3 Conceptual Framework To estimate the effect of the distribution of income on terrorism, I use a conceptual framework that links economic conditions and terrorism. To make this consistent with empirical evidence showing terrorists tend to be better educated than the recruitment pool they come from, I assume that terrorist organization recruit the best-quality person their resources allow [Bueno de Mesquita2005]. The focus therefore is on the recruitment process of terrorist organizations. The model has three decision making members: the government, the terrorist organization, and the potential terrorist recruits. 3.1 The Government The government selects a level of counter-terrorism a to maximize its utility. The government sees disutility from both terrorist violence v() and the cost of counterterrorism k(a). While counterterrorism directly increases the cost of producing violence, it also hurts the economy by a proportion τ(a). In this way, government crackdowns on terrorism have an ambiguous effect on the level of terrorism. The benefit of increased security can be outweighed by increased participation in terrorism. U G = v(r, θ, a) k(a) 6

13 3.2 Terrorist Organizations The terrorist group select the level of violence v within this model. The terrorist organization chooses the amount of violence given the constraints of (the endogenous) average available recruit quality θ, the resource cost r of the violence, and the level of counterterrorism chosen by the government a. The terrorist organization accepts the most highly qualified recruits to volunteer. As the quality of available recruits increases, the terrorist group will invest more resources in terrorism because the marginal yield on those resources will be greater. The total resources available to the terrorist organization are limited to the opportunity cost c(r) of those resources rather than any specific endogenous boundary. More clearly, resources available and the related opportunity cost function are exogenous to the model. In practice, terrorist funding is unlikely to be truly exogenous. Particularly in environments where terrorists raise both funding and recruits in the same place, factors affecting the wages of recruits are likely to endogenously affect resource levels too. Alternately, the cost of human capital might be disconnected from resources because of fundraising among an expatriate community or the transnational appeal of some causes. Funding raised in one country might be used to hire recruits in a second country for an attack in a third. U T = v(r, θ, a) c(r) 3.3 Potential Terrorists The potential recruits are modeled as choosing between participation in terrorism and participation in the traditional economy. Recruits possess a certain level of ability θ i that positively affects both their effectiveness as terrorists and their income from economic activity. As ability increases, economic benefits increase faster than terrorism benefits, so 7

14 there is a level of ability beyond which a person will not participate in terrorism. In actuality, terrorist recruitment might not strictly by economic opportunity, however. For example, different individuals might respond to crackdowns differently (different u(a, θ i )) despite equivalent human capital. Recruits gain ideological benefits e(a) from participation in terrorism that increase as crackdowns by the government become more intense. They or their families also receive financial compensation from the terrorist organization as compensation. This compensation u(a, θ i ) is increasing with the intensity of the crackdown and the individual s ability. The benefits from economic participation are a function of ability and exogenous economic conditions γ. Additionally, the economic harms of counterterrorism τ(a) reduce the benefits of regular economic opportunities. Individuals seeing greater utility from participation in terrorism than non-participation will choose to participate, and these will tend to be individuals of lower ability (low θ i ) and greater motivation for terrorism (high e(a)). 3.4 Additional Considerations The differences in the terrorist organizations operating among different countries complicates application of this model. For example, if a country is particularly successful at counter-terrorism it might tend to have less-experienced terrorist organizations than another country. Over time, these differences might snowball as the experience of the terrorist organization increases the efficiency of its production of violence. This is similar to the assumption that the opportunity cost curves for resources are the same across terrorist organizations and time. In both cases, the most important component of the assumption is that these differences do not correlate with variables of interest. I assume that quintile income contractions are uncorrelated with differences in the efficiency of violence or the opportunity cost of resources. 8

15 U v i = e(a) + u(a, θ i ) U e i = (1 τ(a))f(γ, θ i ) This conceptual framework interprets macroeconomic conditions both exogenously and endogenously, and these economic conditions influence the choice of the recruitment population to participate in terrorism. The government endogenously selects crackdowns as responses to terrorism. The model assumes that the direct damage from terrorism is unimportant to the economy at large, that the violence v does not directly have an influence on the economy. Several additional limitations on the model will be imposed. Apart from the violence they cause, the terrorists are unobservable; to simplify, the opportunity cost curves of resources are unrealistically assumed to be constant over time and organization. Furthermore, the individual human capital of terrorists is unmeasured; so this study will instead use income quintile as the index. Income quintile becomes a proxy for the ability of the individuals within that quintile. This study assumes, therefore, that the shifts that happen when real individuals move from one income quintile to another correspond to changes in ability. Quintile income contractions become the term for overall utility of economic participation Ui e because the quintile income contractions include the influence of crackdowns, exogenous economic shocks, and ability for those individuals. 4 Empirical Model I hypothesize that while a country s income-per-capita for a country will not have an association with terrorism, the third or fourth quintiles will show a negative association 9

16 between income and the terrorist violence caused by nationals from that country. Income in other quintiles would be expected to have no associated effect on violence. Income quintile acts as a proxy for a potential recruit s human capital level θ with income in that quintile serving as a market opportunities f(γ, θ i ). I lack a strong proxy for the critical term for government crackdowns a, so endogeneity is a serious concern. My regression is specified as follows: V iolence cy = β 0 + β 1 ln(quintileincome q cy) + β 2 maleunemploy cy + β 3 ln(population cy ) +β 4 highp olrights + β 5 lowp olrights + β 6 areakmsq c + β 7 tropicar c +β 8 elevation c + β 9 langf ract c + β c langf ractsq + ɛ (1) The core specification was also tested for misspecification by RAMSEY reset test which rejects the hypothesis that there are no nonlinear terms missing from the specification. However, trial and error introduction of nonlinear terms of the control variables led only to the inclusion of the square of linguistic fractionalization. I regress violence by nationals of a country on each quintile s income and a set of control variables, using OLS with heteroskedasticity-robust standard errors and Prais- Winsten correction for autocorrelation. I run probit regressions on the occurrence of a terrorist attack causing a casualty and on the occurrence of a terrorist attack causing a large amount of monetary damage. 4.1 Choice of Data and Variables I use several datasets for this analysis, including economic, human rights, government, and geographic data. Finally, I rely most heavily on the ITERATE terrorism dataset. The ter- 10

17 rorism dataset s unit of analysis is an individual terrorist incident, while the other datasets have a unit of analysis of a country, typically over a year. In almost all of the datasets, there are large sections of missing data, making missing data selection an additional concern. While the government and economic datasets are organized per country per year, the terrorist dataset is organized by attack. To perform the described analysis, I merge the datasets for economic, government, and terrorism variables into a single dataset with the unit of analysis being a time series, cross sectional dataset of all of the countries of the world. The political upheaval at the fall of the Soviet Union would have complicated the definition of countries. Because of this, I chose to use 1990 as the beginning of the time series. For countries that unified, the unified country was created in the data as if it had existed since Data for the separate constituent parts were set to missing when additive constructions were not possible. I chose the logarithm of income by quintile as the variables of interest because they captured both the size and wealth of the economy overall and the distribution by quintile. Income quintile acts as a proxy for broad tiers of human capital. This assumes that economies that have similar levels of income in their quintiles have similar levels of human capital for the individuals in their quintile. This is a weak assumption. To the extent that countries in the data have different levels of human capital with the same level of income for a given quintile, this will confound results. Other proxies for human capital were sought such educational attainment, but sufficient panel data could not be located. Illiteracy had previously been used by other authors and discarded as not correlating with terrorism. I construct these quintile incomes by taking the natural logarithm of the product of gross domestic product data and a quintile s income share. Gross domestic product data comes the World Bank while the income distribution data is drawn from the World Bank, Luxembourg Income Study, and the UNU WIDER. This breadth of sources in turn draws 11

18 from a variety of income distribution surveys techniques. Unit of analysis, definitions of income, and demographic populations surveyed vary among studies. This patchwork of survey methods introduces measurement inconsistency among methods [Hoover1989]. This may introduce bias to the regression if the factors associated with one survey method correlate with contributors to terrorism. I also chose to include unemployment both because previous studies have identified unemployment as a determinant of terrorism [Hon2004]. In addition, the Bueno de Mesquita model directly implies that unemployment should be important as a contributing factor in terrorism. Higher unemployment provides a wider pool of individuals whose talents are available to the terrorist labor market. Unfortunately, the unemployment rate is not broken out by income quintile. Male unemployment in particular was chosen because most terrorists are male. Labor force participation was not included on the assumption that individuals who have elected not to participate in the labor market are either unlikely to be recruited into a terrorist organization on the margins for the same reasons they are not in the labor force. Countries with high levels of civil liberties or political rights may be easier for terrorist groups to operate, organize, and recruit within. Alternately, these freedoms might tend to provide an outlet for political expression that forestalls violent external terrorism. Because of these, I use indicators for high and low levels of political rights as measured by a score of one for a free society and six or seven for an unfree one. I include Political rights from Freedom House based on their inclusion by previous authors [Abadie2004, Krueger and Malečkova2003]. In addition, the inclusion of a variable for very restrictive countries captures the possibility that these countries prevent the formation of active terrorist groups or possibly engender additional terrorist activity through repression. In determining the choice of a proxy for crackdowns a, I also tried the use of the 12

19 physical integrity and freedom of movement variables in regressions not included in this paper. Political freedom control variables serve as proxies for government crackdowns, but the lack of political freedoms may or may not represent focused counter-terrorism efforts rather than general political restrictions against a non-terrorist population. I include the logarithm of country population as a control variable because many of the control variables as well as the variables of interest are not scaled to the size of a country. I expect that a wealthy, small country should be the source for less terrorism than a wealthy, large country. Previous authors have also found population to be a statistically significant factor in terrorism [Krueger and Malečkova2003] Geographic data for elevation, surface area, and percentage of area in the tropics were included based on previous literature [Berman and Laitin2007, Abadie2004]. Mountainous terrain and tropical climate encourage civil wars or insurrections because of better terrain for those actions. Flat and open terrain may encourage terrorism because of the difficulty of hiding an insurgency in that kind of terrain. However, for international terrorism, the importance of geography in hiding a terrorist group may be reduced because the country might be less likely to oppose the groups operations. Moreover, nationals of a country may not even base their terrorist organizations from within their home country, making national geography less important than in domestic terrorism. I also include regional dummies to capture regional fixed effects. I use three different dependent variables to separately analyze whether the variables of interest affect the occurrence of terrorist violence or the amount of violence. First, I create a positive, continuous dependent variable of the sum of all injuries and fatalities from international terrorism caused by the nationals of a country in a year. When an attack was perpetrated by attackers known to be from multiple countries, the violence was assigned evenly among all known countries involved. Assumptions about the proportion 13

20 of responsibility are necessary because exact proportions of attacker nationalities are not included in the dataset. In a later Section 4.3, I discuss why this decision should not affect my results. I drop all attacks unattributable to a country of national origin. I assume that the selection function for whether perpetrators are attributed is uncorrelated with the error term. This paper also assumes that the attribution by national origin is correct in the dataset. I consider attacks as taking place in the year the attack begins. For the probit regressions, I use one binary indicator for whether a country s nationals were involved in an international terrorist attack causing casualties in a year and another for whether a country s nationals were involved in an international terrorist attack causing property damage of at least ten thousand dollars. The ITERATE dataset limits the analysis via three important selection effects. First, the data include only international terrorist incidents, although the premises of this paper do not expect that international data behave differently. This omission, combined with variation among datasets in the breadth of their definition of terrorism does affect the number of incidents and amount of violence for analysis [Gary LaFree2006]. Second, IT- ERATE draws information only from open-source information on terrorist attacks. Media sources are open to distortion because of inaccuracy, bias, and audience context, though these concerns also pose problems for any of the indirect sources of data [Silke2004]. The selection effect based on the newsworthiness of a country s terrorism. Terrorist violence not covered by media sources will not be included in the ITERATE dataset, and the standards of newsworthiness may have changed over time as media consumers become jaded. Silke specifically notes the selection bias would have worsened since mid-1996 when some FBI data sources were no longer made available to ITERATE. [Sandler and Enders2004] I expect that my OLS results should be more robust against this bias because highly violent terrorist attacks are presumably more likely to be included in the dataset whereas 14

21 very small or non-violent terrorist attacks are less likely. Finally, not all terrorist incidents have the same level of information available. In particular, the nationality of attackers or the organization responsible is not always known. This factor also hints at the further selection problem that incidents whose perpetrator is unknown might not be recognized as terrorism and included [Gary LaFree2006] La Free et. al. also raise the issue that conflicting media reports may not be reconcilable by the event compiler. 4.2 Description of Datasets I merged several datasets to produce conflict, economic, political, and geographic data at the country-year level. Although, most of the underlying datasets are at the country level of observations, a country is not necessarily a politically independent entity. I use the country-year as the unit of analysis to avoid several problems with data quality and endogeneity. Using the country-year as the unit of analysis avoids problems of recruiting substitution by a terrorist group among countries, temporarily inactive groups, or groups that are disrupted before perpetrating an attack ITERATE This paper will draw data on terrorist events from the numeric version of the International Terrorism: Attributes of Terrorist Events (ITERATE) dataset, compiled by Edward F. Mickolus, Todd Sandler, Jean M. Murdock, and Peter A. Flemming. The dataset lists individual international terrorist incidents beginning in 1968 through the present day, but this study will utilize only data from 1990 through Geographic coding of terrorist incidents spreads across hundreds of geographic and political entities. ITERATE considers many islands as separate from their nationality. For the purposes of this paper, the key information about each terrorist attack is the country of origin of the attackers, the number 15

22 of fatalities, the number of casualties, the amount of property damage, and the year of the attack. Additionally, incidents described as hoaxes in the ITERATE dataset are not included in the descriptive statistics and do not have injuries or fatalities involved. ITERATE has its own definition of terrorism [Mickolus et al.2004]: The use, or threat of use, of anxiety-inducing, extra-normal violence for political purposes by any individual or group, whether acting for or in opposition to established governmental authority, when such action is intended to influence the attitudes and behavior of a target group wider than the immediate victims and when, through the nationality or foreign ties of its perpetrators, its location, the nature of its institutional or human victims, or the mechanics of its resolution, its ramifications transcend national boundaries. International terrorism is such action when carried out by individuals or groups controlled by a sovereign state, whereas transnational terrorism is carried out by basically autonomous non-state actors, whether or not they enjoy some degree of support from sympathetic states. Victims are those individuals directly harmed by the terrorist incident. While a given terrorist action may somehow harm world stability, citizens of nations must feel a more direct loss than the weakening of such a collective good World Bank The World Development Indicators dataset from the World Bank provides the bulk of the economic statistics used for this paper. All data is taken from 1990 through 2007, where available. The World Bank does not collect these statistics according to a uniform schedule or distribution across time and countries. While some of the fields are panel data taken across multiple countries at regular intervals, there are certain countries for 16

23 whom the panel data is absent despite the country s importance to the study of terrorism. In other cases, the variable of interest is surveyed less than annually. The World Bank provides data for 209 economies where an economy is a country or a separate component of a nation s economy for which that nation reports data separately (e.g. Hong Kong) [Wor2008] For this paper the information used from the World Development Indicators were: GDP, population, income share by quintile, male unemployment, poverty rate University of Gothenburg: Quality of Government As a source for a variety of control variables on the incidence of government crackdown on terrorism or general suppression, this paper draws from the University of Gothenburg: Quality of Government collection of datasets [Teorell et al.2008]. All data were at the country level unit of analysis. Alesina, Devleeschauwer, Easterly, Kurlat, and Wacziarg have data on linguistic fractionalization within a country from 189 countries sampled in 2001 [Alesina et al.2003]. These values are used for the entire time series as a partial substitute for country level fixed effects. As an omitted variable, this assumes that changes in linguistic fractionalization over time do not correlate with economic effects. The square of linguistic fractionalization was also found to be statistically significant after misspecification testing and was therefore also included. Variables for ethnic and religious fractionalization were tested but excluded from the model. Freedom House s Freedom in the World indicators are used to provide data on civil liberties and political rights. Freedom House has rated 193 countries and 15 territories since 1973 [dat2007]. I use the Freedom House data on political rights rather than some others because the Freedom House data has fewer countries or years missing than the other datasets. 17

24 4.2.4 Center for International Development General Measures of Geography The Center for International Development at Harvard University offers a General Measures of Geography dataset with mean elevation, tropical land area percentage, and area data for most of the countries in the data [Gallup et al.2001]. These values are not adjusted over time, but I assume that changes in these values over time are unrelated to economic indicators. This assumption is weak, however because the primary reason for changes in geography are political and military events that might be quite correlated with economic factors or terrorism. However, where geographic alterations occur during the time period, the changes to data have been backdated as if the changes happened before Furthermore, in all cases, terrorism is not associated with those countries during the period when the changes occurred. In some cases, like Russia and Kazakhstan there are periods of significant terrorism later in the time series UNU-WIDER To address concerns about insufficient income distribution data, I supplemented the World Bank data with additional data from the World Institute for Development Economics Research of the United Nations University (UNU-WIDER) [dat2008]. This dataset, the World Income Inequality Database version 2.0, compiles income distribution and inequality data from a wide variety of sources. This dataset included Gini indices, decile income shares, and quintile income shares that were used. All data were at the country level unit of analysis. 18

25 4.3 Descriptive Statistics Descriptive Statistics on Terrorism The dependent variable, terrorist violence, is the sum of casualties and fatalities from international terrorism between the years The pattern of terrorism is highly variable with most incidents having no casualties or fatalities associated with them even when hoaxes are excluded. The descriptive statistics of terrorism by incident are shown in Table 1. Of the 3,844 incidents with fatality data, no one was killed in 2,895 of them. With injuries, the distribution is similar with 2,973 of the 3,776 incidents with casualty data having no injuries. At the same time, there is a wide variation in the outcome of terrorist incidents. Because they are such extreme outliers, the 9/11 attacks are excluded from the regression analysis. Table 1: Casualties and fatalities from international terrorism by incident, Incidents Mean Std. Dev. Min Max Missing Fatalities 3, , Injuries 3, , The distribution of terrorism across country-years is similar to that of the incident level analysis. Most country-years have no terrorist violence, with 3,483 country-years having no fatalities out of 3,813 observations. As can be seen in Table 2, the variation in the level of violence by country-year is also wide. Even without 9/11, there are observations with hundreds of fatalities. In the case of injuries, 3,521 observations have no injuries from international terrorism. 19

26 Table 2: Casualties and fatalities from international terrorism by country-year, Mean Std. Dev. Min Max Fatalities Injuries Total Descriptive Statistics on Independent Variables All of the independent variables are collected at the country level in Table 3. There are 212 countries in the dataset, although some are not truly independent countries. In most cases, these extra countries are preserved in the data because both the World Bank and ITERATE tracked data for them separately. Table 3: Descriptive statistics of independent variables over 212 countries, Observations Mean Std. Dev. Min Max ln(qy1) ln(qy2) ln(qy3) ln(qy4) ln(qy5) munemp 1, ln(pop) 3, highpr 3, lowpr 3, areakmsq10k 2, ,657 tropicar 2, elev100m 2, langfrac 181* >$1 a day poverty rate * The 181 observations for language fractionalization are 181 countries from a single cross section in These values are repeated across the entire time range. 20

27 4.4 Regression Implementation To estimate the association between the amount of violence and quintile income, I use OLS to regress violence by nationals of a country on the variables of interest and controls. After some testing, I use a main specification (1) as the basis against which I test interactions, check for problems, and conduct the probit regressions. I arrived at the core model (1) through aggregation of results from the literature which included geographic data, regional dummies, linguistic fractionalization, population, unemployment, and political rights [Abadie2004, Krueger and Malečkova2003, Hon2004]. I experimented with these control variables to check for robustness across different configurations. After some manipulation of the variables of interest, I chose to use log of income by quintile. Alternative arrangements of quintile income, such as separating GDP per capita from the individual income share percentages, were highly collinear. Other formulations were difficult to interpret, particularly with interactions. I construct the present specification (1) of the log of income per capita in a quintile by taking the logarithm of the product of the income share with GDP per capita. I chose the control variables for the main specification (1) through trial-and-error addition and removal and addition based on previous literature. I checked for heteroskedasticity by examining scatter plots of dependent violence variables versus each independent variable. By visual inspection, only one independent variable, population, appeared to exhibit heteroskedasticity. There was a rapid decline in variation of terrorism for nations with population below one million people. This may have been a product of selection bias in the ITERATE dataset or unobservable qualities of smaller national communities being less prone to external terrorism violence. To correct for this problem, robust OLS regressions were used to generate all results. I also examined the data for outliers and influence. I did not identify any country, 21

28 country-year, or region whose influence alters the sign or statistical significance of the effect of income in the first or second quintiles. A plot of residuals against leverage reveals Sierra Leone as the country with the most leverage. Omitting Sierra Leone does not affect the general tenor of the results, however. Omitting of Sierra Leone from the core model (1) increases the effect size of income in the first quintile from 28.9 to 34.3 casualties or fatalities per one percent increase and decreases the effect size of income in the second quintile from to casualties or fatalities for a one percent increase. Of more concern as outliers are the 9/11 attacks. As an individual event, 9/11 outstrips other attacks enormously, and while it does not affect the variables of central interest because there is no data on income distribution for those country-years, it does influence the control variables considerably. With 9/11 included, there is no statistically significant autocorrelation. While other terrorist groups might have previously had aspirations to commit such mega-terrorism, none had previously succeeded. Consequently, it is difficult to estimate the determinants of smaller variations in terrorism with 9/11 included. I have kept 9/11 for the purposes of the probit regressions but removed it for the OLS regressions. A number of countries with high residuals lack leverage because of missing data in the variables of interest. Large amounts of terrorist violence came from the Russia, Jordan, and Indonesia in years when no data on income distribution were available. The World Bank dataset has no income distribution data for the West Bank and Gaza, Saudi Arabia, and Iraq, but these countries have years with large amounts of violence. Without data on income distribution In these cases, the absence of relevant data leads to their lack of influence on the effect size of quintile income. Missing observations are endemic throughout the independent variables. I employed missing data indicators to keep the number of observations constant across all specifications and regressions. Missing observations are not randomly distributed; the missing observa- 22

29 tions are concentrated in developing countries and those with fewer people. Data missing because a country is small should not bias results because these entities also tend to be responsible for only a small part of terrorism and the variation in terrorist violence. On the other hand, observations missing because of the level of development cause concern because there may be omitted variables affecting both terrorism and missingness of the income distribution values. Such a correlation causes bias and upsets standard errors. In about a third of incidents from 1990 to 2007, the nationalities of perpetrators are either completely unknown or only some of their nationalities are known. I show this breakdown in Table 4. If none of the nationalities of an attack are known, I do not assign the outcome of the attack to any country, so the incident is not included in the regression analysis. Because the ITERATE dataset draws only from publicly available sources, the dataset may not include the final outcome of the investigation into the terrorists identities. My assignment decision affects a significant minority, sixteen percent, of the overall violence. If it were the case that the terrorists who remain unidentified were more capable than those who were identified, this assignment choice would bias results to reduce the apparent effect of independent variables associated with terrorist effectiveness. These incidents without any known nationalities are included in the descriptive statistics of terrorism. If only some of the nationalities involved in a terrorist attack are known, the unknown nationalities are ignored as if only terrorists from the identified countries were involved. This choice was made on the assumption that unidentified terrorists are of the same nationality as the identified terrorists. This has the effect of over-representing nationalities of known terrorists compared to countries whose terrorists are unknown. This includes only four percent of the overall terrorist violence. ITERATE also lists only three nationalities involved in any particular incident 1. As 1 Between 1990 and 2007, only nine incidents had four or more nationalities involved; only a single casualty and a single fatality come from incidents known to have had more than three nationalities involved. 23

30 Table 4: Unknown attacker nationality by incident, Nationalities Some Unknown No Known Total All Known Nationalities Nationalities No. Incidents 3,925 2, ,157 % Incidents No. casualties or fatalties 21,526 17, ,498 % casualties or fatalties shown in Table 5, attacks with only one or no known nationalities are the majority of attacks and violence. Table 5: Number of known attacker nationalities by incident, Three Known Two Known One Known No Known Total Nationalities Nationalities Nationality Nationalities No. Incidents 3, , % Incidents No. casualties or fatalties 21,526 3,790 1,349 12,889 3,498 % casualties or fatalties ITERATE does not provide how many terrorists of each nationality were involved in an attack, but it does list the nationalities in order of the number of terrorists in the attacking party of that nationality. To divide the casualties and fatalities from an incident among the nationalities involved, an assumption about the proportion of the nationalities was necessary. For the results presented in this paper, responsibility for the attacks was divided equally among all known nationalities. As a sensitivity test, this paper repeats the main specification (1) using different allocations of violence to participating nationalities. The primary allocation method assigns 24

31 violence evenly to all participating countries. The second method assigns the violence unevenly. For example, a terrorist attack with 10 fatalities and three countries contributing terrorists to the attack in the order A, B, C would have the violence assigned with 1 fatality to country A, 2 fatalities to country B, and 7 fatalities to country C. The results of the main specification do not vary significantly with the application of these changes to the dependent variable. Because I use time series data, I test for autocorrelation by regressing residuals from the model (1) on lagged residuals. Correlation among consecutive error terms could be caused by a terrorist campaign in one year leading to more intensive government interdiction in the next. Contrariwise, lulls in activity might spark additional - or more violent - actions by a group in order to stay relevant or retain attention. A Prais-Winsten OLS regression will be used to correct for autocorrelation. This regression method also allows for fixed effects and heteroskedasticity robust standard errors Hypothesized Effects The literature on terrorism and human capital suggests a negative association between income beyond the bottom two quintiles and terrorism. This negative association with income would be spread out across quintiles because different countries have different equilibrium points for the labor markets for terrorists. The more developed economies would tend to have terrorists who were closer to the median levels of human capital because of higher returns to human capital in those economies making terrorism comparatively unattractive. To the extent that this association was shown, I expected that separation of countries by development indicators would show that countries with greater income inequality or more extreme poverty would tend to have negative associations with terrorism in higher quintiles. More developed countries are predicted to have negative associations 25