ASSESSING THE INFLUENCE OF POLITICAL INSTABILITY ON COUNTRIES LIKELIHOOD OF SUFFERING TERRORIST ATTACKS

Transcription

1 ASSESSING THE INFLUENCE OF POLITICAL INSTABILITY ON COUNTRIES LIKELIHOOD OF SUFFERING TERRORIST ATTACKS A Thesis submitted to the Faculty of the graduate School of Arts and Sciences of Georgetown University in partial fulfillment of the requirements for the degree of Master of Public Policy in Homeland Security By Alex Michael Winograd, B.A. Washington, DC April 2007

2 The research and writing of this thesis is dedicated to my wife and family who have provided unlimited amounts of love and support throughout this process. Many thanks, Alex ii

3 Copyright 2007 by Alex Michael Winograd All Rights Reserved iii

4 ASSESSING THE INFLUENCE OF POLITICAL INSTABILITY ON COUNTRIES LIKELIHOOD OF SUFFERING TERRORIST ATTACKS Alex Michael Winograd, B.A. Thesis Advisor: Richard Hayes, Ph.D. ABSTRACT The 9/11 attacks forced U.S. policy makers to recognize the growing threat posed by terrorism and redefine their policy priorities. Since 1998, the global number of reported terrorist incidents increased from a low of 1,153 in to a high of 5,008 in The number of identified active terrorist groups also climbed from a low of 82 in to a high of 152 in Even more alarming, the ratio of attacks to active groups increased from a low of attacks per group in 1999, to a high of attacks per group in Given the increasing rates of terrorist group development and attacks, policy makers must implement a strategy to address short-term security needs and longer-term threats. President Bush introduced the National Strategy for Combating Terrorism in (updated in 2006) detailing two strategic objectives: 1. Eliminating active terrorists, denying them resources including WMDs, safehavens, funding, and access to targets, and creating an inhospitable global environment for violent extremism (commonly referred to as the War of Ideas ). 2. Eliminating the conditions causing individuals to radicalize. Eliminating the conditions causing radicalization is difficult because there is little consensus on what causes radicalization. Previous research examines factors such as poverty, levels of education, unemployment rates, religion, and forms of government as drivers of radicalization. 2 This thesis assesses the effects of political instability on a country s probability of being attacked and the aggregate number of terrorist attacks countries experience. Using a combination of twostage least-squares regression, Linear Probability Modeling, and Tobit analysis, I control for factors including poverty, education, geography, form of government, religion, and unemployment rates. I find political instability is insignificant in explaining terrorist attack data. iv

5 TABLE OF CONTENTS Section Page Number 1. Introduction... 1 Defining Terrorism... 1 Combating Terrorism Drivers of Population Radicalization... 3 Examining Structural Factors Contributing to Terrorism... 4 Assessing Political Instability as a Terrorism Risk Factor Methodology Strengths and Weaknesses of the Aggregate Dataset Description of the Two-Stage OLS Model Methodological Challenges Defining the Linear Probability and Tobit Model Specifications Assessing the LPM and Tobit Models Country Specific Factors Form of government Specific Factors Population Specific Factors Conclusion Considerations for Future Research Endnotes Appendix 1: Instrumental Variables for Two-Stage Least Squares Regression Appendix 2: Base Model Variables Appendix 3: Description of Source Materials Appendix 4: Descriptive Statistics for Non-Binary Variables Appendix 5: Bivariate Correlations Matrix Appendix 6: Stepwise Analysis Appendix 7: Observations Missing Form of Government and Unemployment Rate Data Appendix 8: Observations with Inputted Regional Values for Missing Data Bibliography v

6 TABLES, FIGURES AND CHARTS Title Page Number Table 1: Definitions of Terrorism 1 Table 2: Variables Incorporated into the Predicted Instability Model 14 Table 3: Controlling Factors within the Models 15 Table 4: Skewness Index Values for Non-Normally Distributed 18 Variables Table 5: Sample Observations by Variable 19 Table 6: Sample Observations by Variable 19 Table 7: Summary of High Independent Variable Correlations 20 Table 8: Variables Excluded from the Tobit and LPM Analyses 21 Table 9: Coefficient Estimates for the LPM with Robust Standard Errors 22 Table 10: Coefficient Estimates for the Tobit Model with robust Standard Errors 22 Table 11: Sub-Categories of Government Disempowerment 27 Figure 1: The World Bank Governance Dataset 10 Figure 2: Assessing the Impact of Democracy on the Attacked Variable 26 vi

7 INTRODUCTION The 9/11 attacks were the first attacks on the U.S. mainland since World War II. The perceived suddenness of the attacks and the magnitude of the destruction, forced the American public and policy makers to recognize the nation s vulnerability to attack. Following 9/11, policy makers passed legislation to improve domestic security and the President redefined the U.S. s foreign policy, issuing several strategies to eliminate terrorism. 3 The policies enacted today will define the U.S. s role in the world. Decision makers need to understand the threat facing the U.S. to make effective policy and secure the country s interests. Defining Terrorism One of the challenges policy makers face is the lack of consensus regarding what constitutes a terrorist act. Most definitions of terrorism define terrorist acts by examining the characteristics of attacks. While analysts generally agree terrorist acts are committed by sub-state groups (if a country were to conduct a terrorist attack it would be an act of war), there is a lack of agreement regarding the nature of terrorist attacks, the goals of terrorist attacks, and the targets of terrorist attacks, making a uniform definition impossible. Table 1 provides a summary of common definitions for terrorism: Table 1: Definitions of Terrorism Source Nature of the Attack Goal of the Attack Target U.S. State Department 4 Violence Political Noncombatants Robert Pape 5 Violence Intimidate or frighten Louise Richardson 6 Rand Corporation 7 Violence, or the threat of violence Violence, or the threat of violence Politically Inspired. Sends a message through a symbolic act to target audience Politically motivated. Creates an atmosphere of fear, are designed to coerce others, and achieve maximum publicity 1 Population of interest Civilians Generally directed against civilian targets

8 Definitions can be narrow or broad depending on how the analyst defines the nature, goal, and target of the attack. For example, the U.S. State Department restricts terrorism to violent acts, with political goals, targeting non-combatants. 8 Robert Pape uses a broader definition expanding the target group, and expanding attacks to include those with non-political objectives. 9 Louise Richardson and the Rand Corporation use the broadest definitions including the threat of violence, and broaden objectives to include generating fear. Richardson restricts the target population to civilians but states military personnel can qualify as civilians if they are not actively engaged in a conflict. The Rand Corporation allows for both civilian and non-civilian targets. 10 This thesis seeks to analyze terrorism in its broadest sense and therefore I use the Rand Corporation s definition of terrorism, as well as its dataset. Combating Terrorism The burgeoning literature on terrorism provides a plethora of explanations for the causes of terrorism. Initial theories postulated poverty is a driver of terrorism. In 2002, President George W. Bush asserted, we fight against poverty because hope is an answer to terror. 11 This supposition was incorporated into the National Strategy to Combat Terrorism, where the President indicated the U.S. would work to, diminish the underlying conditions that terrorists seek to exploit. 12 Intuitively the assumption that poverty contributes to higher rates of terrorism makes sense. It is easy to imagine poverty leading to desperation and a willingness to undertake extreme actions. Yet, in the 2006 National Strategy to Combat Terrorism the President removed poverty as a causal factor of terrorism. 13 This reversal in policy likely resulted from multiple studies demonstrating poverty, at both the individual and the country level, does not correlate with terrorism. 14 2

9 In the 2006 National Strategy to Combat Terrorism the President identifies the absence of democracy as a primary driver behind terrorism stating, Effective democracy provides a counter to each [condition], diminishing the underlying conditions terrorists seek to exploit. 15 Preliminary research indicates the President has reason to believe a country s form of government influences terrorist activities. Alan Krueger s analysis shows that democracies are more likely to be targeted for attack but are less likely to be attacked directly (i.e., embassies in other nations will be attacked rather than the country itself). 16 In contrast, Alberto Abadie s results indicate the relationship between political freedom/democracy and the risk of terrorism is non-linear decreasing at both ends of the political spectrum. 17 This non-linearity indicates that there may be something inherent to the transition from democracy to autocracy, and vice versus, that increases the number of attacks that a nation is likely to experience. Transitions from one type of government to another are often characterized by high levels of political instability. This may indicate that measures of political instability are better predictors of terrorist attacks than the form of a nation s government. The conclusion that the form of a nation s government drives both the likelihood of attack and the number of attacks on a given country is a critical component underlying the President s policy of democratization. If democratization can prevent/eliminate terrorism then the U.S. should pursue this policy. If political instability is a better determinant of the likelihood of terrorist attack, then policy makers will need to weigh whether other policies will be more effective in combating terrorism. DRIVERS OF POPULATION RADICALIZATION While there is a growing body of literature surrounding the drivers behind terrorist attacks, and the factors leading to terrorist recruitment, little of this data is definitive. This ambiguity makes it difficult for policy makers to make informed decisions regarding how 3

10 governments should respond to the escalation in the number of terrorist attacks and the number of terrorist groups worldwide. A variety of factors seem to correlate with the number of terrorist attacks a nation experiences and its risk of terrorism. Factors fall in one of three categories psychological, based in individual rational choice, and structural: 1. Psychological Factors Specify and explain why individuals join terrorist organizations, terrorist group dynamics, and how participants affect the commission of terrorist acts Individual Rational Choice Use the theory of individual choice based in economics to understand why individuals become terrorists and complete terrorist attacks Structural Factors Posits the causes of terrorism are found in the environmental, political, cultural, social, and economic structure of societies. 20 While all three categories may influence some aspect of terrorism, only structural factors assess country specific factors that influence terrorist activities. Psychological factors and individual rational choice focus on individual terrorists decisions and do not provide insight at a national level. Therefore, this thesis focuses on structural factors that make nations more likely to be victims of terrorist attacks. Examining Structural Factors Contributing to Terrorism Some terrorist groups form to compel an occupying force to withdraw forces from territories the terrorists consider to be their homeland, while others form to address local or regional concerns. 21 Therefore, many terrorist organizations are sub-national, and recruit and attack within the same country (although the target of the attack may be another nation). 22,23 Structural variables that provide insight into the likelihood a nation s population is susceptible to terrorist recruitment and radicalization may provide insight into whether a country is more likely to be attacked by terrorists. Structural variables fall into two general categories. The first is population specific factors such as educational levels, religious beliefs, age distribution, absolute and relative poverty levels, and unemployment rates. The second is country specific factors such 4

11 as a nation s geography, population distribution, government type/levels of political freedom, and measures of political instability. Population Specific As mentioned previously, many terrorist organizations recruit and operate within the same country. Measurement of population specific factors that predispose individuals to radicalization may enable better predictions regarding a country s vulnerability to terrorist attack. One theory regarding why populations radicalize is extreme poverty, either in absolute terms (i.e., very low income) or in relative terms (i.e., compared to others), predisposes individuals toward terrorist philosophies. Yet, most studies have not proved this relationship exists or that it influences the number of terrorist attacks on a given country. For example, Abadie finds relative levels of wealth (both at the country level and at the individual level) do not correlate with a country s risk of terrorism (using a terrorism risk index as the dependent variable). 24 His findings echo those of other experts including Krueger and Maleckova, and Berrebi. 25 Additionally, if poverty drives terrorism, then terrorism should be much more prevalent in African countries, which are the poorest in the world. 26 Nonetheless, poverty data continues to improve and I control for measures of absolute poverty and income dispersion in this analysis. Another hypothesis is educational levels are predictors of populations susceptibility to radicalization. In his analysis of Palestinian terrorists, Berrebi determines educational levels and participation in terrorism positively correlate and are statistically significant at p <.05 (p values are the mathematical probability that the data are the result of random chance. In this case there is less than a five percent chance that this correlation is due to chance). 27 Qualitative evidence supports his findings. For example, between 1996 and 1999, Nasra Hassan, a relief worker for the United Nations, interviewed 250 militants and associates of militants involved in the Palestinian cause. 28 These interviews showed terrorist groups had an overabundance of recruits, 5

12 requiring them to screen candidates for desirable characteristics; including educational levels. Finally, hate groups, which are often compared with terrorist organizations, are more likely to operate in areas where adult populations have higher levels of education. A similar relationship may exist for terrorist groups. 29 There are many explanations for the high levels of education exhibited by some terrorists. One is high levels of educational attainment may signal one s commitment to the cause and one s ability to prepare for and complete assignments. A second is those with higher levels of education are more aware of injustice in the world, making it more likely that they will be passionate about terrorist causes. This, in conjunction with the higher wages this population earns making them less preoccupied with ensuring they have the bare essentials needed for survival may make it more likely they will join or form terrorist organizations. Given evidence that terrorists are often well educated this analysis includes a variable that controls for the percent of population with post secondary education. Another factor that may increase the likelihood populations will be radicalized is whether citizens can engage in other productive activities, such as work. 30 Terrorist groups sometimes seek to engage underutilized populations by providing social services. For example, Hamas and Hezbollah garner public support by meeting the social needs of potential recruits, establishing hospitals, schools, orphanages, and providing work. They provide social services in areas with high unemployment rates such as Egypt, Lebanon, and the West Bank enabling them to enhance their ability to recruit underutilized young men. 31 Given these behaviors, it is possible unemployment rates may provide a means of gauging a society s vulnerability to radicalization and by extension to terrorist attacks and as such I control for unemployment rates within this thesis. Over the past 30 years the number of religiously based terrorist groups increased dramatically. 32 Evidence suggests religion can be a powerful driver behind terrorist recruitment, 6

13 recruit radicalization, and terrorist attacks. Religion enhances the stature of terrorist groups, legitimizing their tactics and objectives through the use of religious texts. 33 This better positions them to market their causes to recruits by offering intangible spiritual benefits in exchange for recruits willingness to fight and die. Krueger and Maleckova s study supports these assertions finding each of the four major religions (Islam, Christianity, Buddhism, and Hinduism) positively correlate with the incidence of terrorist attacks. 34 Religion also serves as a basis for determining where terrorist groups attack. For example, members of one religious group may target members of other religious faiths (e.g., the IRA targeting Protestants). Since religion seems to correlate with terrorism, this analysis controls for the effect that majority religions (religions that comprise more than 50 percent of a nation s population) have on a country s vulnerability to terrorist attack. Country Specific In her analysis of terrorism, Martha Crenshaw asserts geographic factors influence terrorist attack patterns. She claims cities are more likely than rural environments to facilitate terrorism. Urban environments allow terrorists several advantages over countrysides including logistic superiority, support, and access to new recruits. 35 Abadie, in his 2004 paper, agrees geographic factors influence terrorism asserting, It is well known that certain geographic characteristics may favor terrorist activities. Areas of difficult access offer safe haven to terrorist groups, facilitate training, and provide funding through other illegal activities. 36 In his study, Abadie controls for geographic characteristics total country area, average elevation, and proportion of the country area in tropical weather and finds all three are statistically significant in explaining changes in global terrorism risk ratings. 37 Given the conflicting conclusions drawn by Crenshaw and Abadie (urbanization versus remoteness) I include control variables for both of these factors in my analysis. 7

14 Another theory is that the distribution of a nation s population across different age ranges may predispose them to violent crime and terrorism. Data tracked by the Bureau of Justice Statistics shows that 62 percent of violent crime in the U.S. is committed by individuals younger than the age of Columbia found that 59.1 percent of violent perpetrators are younger than 40 years old. 39 Additionally, in the National Security Studies Center analyzed terrorist incidents between 1993 and 2002 finding that suicide and non-suicide terrorists are underrepresented in the youngest and oldest age groups but dominate the middle one (ages 22-27). 40 As such, it is possible countries with large proportions of their populations between the ages of 18 and 34 may be more likely to be victims of terrorism and as a result I control for age dispersion within my analysis. Several attempts have been made to correlate form of government with a nation s vulnerability to terrorism. Experts disagree as to whether free societies (democracies) or repressed societies (authoritarian states) are more vulnerable to terrorist attacks. Kreuger and Maleckova, using the State Department s Annual List of International Terrorist Events and the ITERATE data set (two of three primary datasets tracking terrorist attack data), determine countries with greater civil liberties are less likely to generate international terrorists and by extension be the victims of terrorist attacks. 41 Crenshaw seemingly agrees, stating terrorism stems from individuals having a lack of opportunity for political participation and from government use of unexpected and unusual force in response to protest or reform. 42 Pape disagrees with Crenshaw and Krueger and Maleckova pointing out seven democratic countries were the target of most suicide attacks over the past 20 years. 43 A third view is the risk of terrorism, as it correlates with measures of political freedom, is non-linear, with terrorism falling at both extremes of the freedom spectrum. 44 Given the mixed results surrounding the relationship 8

15 between the number of terrorist attacks and form of government further exploration of this issue is required. Political Instability as a Predictor of Terrorism Risk One possible explanation for the mixed results surrounding the relationship between form of government and terrorist attacks is previous studies misspecified their analytical models. Specifically, the measures of form of government may be biased by the omission of accurate measures of political instability. Attempts to control for political instability show mixed results. Krueger finds people from stable countries are more likely to be represented in suicide attacks, either as target or perpetrator, than the average person. 45 Crenshaw states the most salient political factor in the category of permissive causes of terrorism is a government s inability or unwillingness to prevent terrorism inefficiency or leniency can be found in a broad range of all but the most brutally efficient dictatorships, including incompetent authoritarian states such as tsarist Russia. 46 If she is correct, form of government is not a good indicator of whether a country is likely to be attacked by terrorists and other factors should be evaluated. Assessing Political Instability as a Terrorism Risk Factor Literature does not provide definitive guidance on how political instability should be defined or measured. Some studies differentiate the variable based upon factors exogenous to the system of governance (e.g., unplanned political instability) and factors endogenous to the system of governance (e.g., planned political instability such as regular elections). 47 Others seek to divide political instability into categories assessing the magnitude of the imperative facing the government (i.e., crises versus regime unpopularity). 48 Another approach is to analyze political instability by creating more extreme and less extreme indexes (e.g., violent versus non-violent). 49 Finally, some analysts have conducted research into political instability by examining specific 9

16 variables that impact overall levels of political instability and incorporating them directly into their regressions. 50 Each of the approaches listed above are inadequate for the examination of the relationship between terrorist attacks and political instability. Most of the indexes developed in previous studies incorporated terrorism as a measure of political instability creating endogeneity. Since the dependent variable is the number of terrorist attacks on a given country, incorporating this as a measure of political instability will artificially inflate the correlation between instability and terrorism biasing the results of the analysis. Additionally, many of these approaches artificially limit the dimensions of political instability to only its most extreme manifestations (e.g., riots, coups, etc.) and ignore its more subtle manifestations such as planned changes of power and peaceful demonstrations. Finally, while attempts to incorporate multiple dimensions of political instability in models provide greater specificity regarding the individual measures driving the results, this approach provides an imperfect assessment of the correlation of political instability in aggregate. This thesis relies on the aggregate index developed by the World Bank in its Governance dataset (see Figure 1). The World Bank defines political instability as perceptions of the likelihood a government will be destabilized or overthrown by unconstitutional or violent means. 51 This index measures instability within the political regime, public support of the political regime, and politically motivated violence against the regime. By using this Figure 1: The World Bank Governance Dataset The Worldwide Governance Indicators are based on a longstanding research program of the World Bank Institute and the Research Department of the World Bank, initiated in the late 1990s by Daniel Kaufmann and Aart Kraay. Indicators, which capture six key dimensions of governance, have been compiled since 1996 and measure the quality of governance in well over 200 countries, based on 31 data sources produced by 25 different organizations worldwide. indicator, this thesis analyzes how the aggregate effects of political instability influence terrorist 10

17 attack patterns. That being said, use of this indicator poses specific challenges. Specifically, it incorporates terrorism as a measure of political instability. A discussion of this variable and the challenges it poses is provided in the methodology section. METHODOLOGY My analysis quantifies terrorism based upon the number of terrorist attacks within a country in a given calendar year. Specifically, my dependent variables of interest are the number of terrorist attacks within a country annually (Attacks), and a binary dummy variable equaling one (1) if a country experiences a terrorist attack (Attacked). In order to analyze these variables I developed both a linear probability (LPM) and a Tobit model. The LPM allows analysis of the Attacked variable because it generates more accurate predicated values than a Logit model (Logit enables estimation of a binary dependent variable without the creation of predicted values less than zero or greater than one); as measured through scatter plot analysis. Tobit models are econometric models that describe the relationship between a dependent variable that cannot take on values smaller than zero and designated independent variables. Most countries are not victims of terrorist attacks which means country terrorism datasets have large numbers of observations with a value of zero. Since attack data is highly skewed estimates are downward biased when generated using ordinary least squares analysis (OLS uses the same technique as Linear Probability Modeling but is used to analyze continuous dependent variables while Linear Probability Modeling only applies to analysis of binary dependent variables. It finds a line that closely approximates the data to enable the user to predict likely outcomes given the independent variables included in the model). The Tobit model corrects for this effect. Many terrorist have long-range planning cycles, often taking years to plan and execute attacks. Therefore, data over short periods of time (0-5 years) should provide consistent results. An examination of the data obtained for and reveals that correlations between the 11

18 dependent and independent variables are consistent across these two time periods, indicating that conclusions drawn from this dataset should be generally applicable. Strengths and Weaknesses of the Aggregate Dataset The dataset compiled for this analysis draws from eight separate databases (for background information on these databases refer to Appendix 3). 1. Cross National Time Series (Land Area, Predicted Political Instability, Percent of Population in Urban Areas) 2. Polity IV (Form of Government) 3. World Bank Governance Dataset (Political Instability Index) 4. United States Census Bureau (Percent of Population Ages 15-34) 5. World Bank Development Indicators (National Unemployment Rate, Poverty Gap US$1 1, Gini Index 2 ) 6. U.S. State Department International Religious Freedom Report (Islam, Christian, Otherrel 3 ) 7. UNESCO Institute for Statistics (Percent of Population with Post-Secondary School Education) 8. Rand Corporation Dataset (Number of terrorist attacks) This data set has several strengths. First, it incorporates data on a large number of countries resulting in a sample of more than 377 observations for use in the analysis. Second, the data sources have all been extensively used in critically reviewed research and are recognized as high-quality sources of information. 52 That being said, there has been vigorous debate regarding the quality of the variable of interest; the Political Instability Index. Critics of the Political Instability Index assert it is biased towards the views of business elites. 53 Since those elites have interests, indicators will be biased to favor countries that are most supportive of those interests. While this is an understandable criticism, the structure of the 1 The poverty gap is the mean shortfall from a given standard, expressed as a percentage of the stated standard. This measure reflects the depth of poverty as well as its incidence. 2 The Gini index is used to measure income inequality. Zero corresponds to perfect income equality (i.e. everyone has the same income) and one corresponds to perfect income inequality (i.e. one person has all the income). 3 Religion variables are binary variables equal to one when the identified religious group comprises 50 percent or more of the country s total population. 12

19 dataset makes it unlikely indicators are biased in this manner. Indicators are based on three crosscountry surveys, six sets of ratings produced by government and multilateral organizations, and eleven data sources produced by non-governmental organizations. The creators of the indicators assert there is a strong correlation between all forms of data obtained. 54 Critics also claim the indicators contain measurement error. Statistical models show measurement error would have to be implausibly strong to account for the observed correlation between sources. 55 Description of the Two-Stage OLS Model The political instability index draws upon terrorism as a defining element, which poses a methodological challenge for this study. This relationship causes endogeneity, (an endogenous change is one that comes from inside the model and is explained by the model itself) which biases the results. To eliminate this bias I used a two-stage linear regression to estimate predicted values for the Political Instability variable. These predicated values exclude measures of instability caused by terrorism enabling unbiased estimation of the model. To correct for endogeneity within the models I identified instrumental variables that enable me to estimate levels of political instability in the absence of terrorism. I used two instrumental variables to generate predicted values. The first is an interval variable capturing three dimensions of political instability aggregated from the Cross National Time Series (CNTS) dataset: 1. Civil protest 2. Politically motivated aggression 3. Instability of the regime. The second is a dummy variable that is equal to one (1) if the CNTS dataset was missing data for an observation. I incorporated nine dimensions of political instability captured by the CNTS dataset into the interval variable. They were selected based on their ability to meet one or more of the following four criteria: 1. Variables were not be defined as a type of attack within the Rand Corporation dataset. 13

20 2. Variables were not associated with violence, but were associated with political instability. 3. Variables were not associated with terrorists, but were associated with political elites. 4. Violent acts were not correlated with terrorism but were correlated with political instability. Table 2 provides a list of the variables used to generate predicted values of Political Instability (for definitions reference Appendix 1): Table 2: Variables Incorporated into the Predicted Instability Model Civil Protest Riots, Strikes, Anti-government demonstration Politically Motivated Aggression Coups, Purges Instability of the Regime Executive changes, Government crises, Major constitutional changes, Cabinet changes Since the variables I examined measure frequency of events rather than magnitude, I summed them creating an aggregated interval variable called Predinst. Predinst is not normally distributed and was transformed into the log of Predinst (Lpredinst) in order to approximate a normal distribution. Lpredinst correlates with the Political Instability Index (Chow test 4 generated: F = 11.9, d.f. 2,363 ) and is uncorrelated with the error term generated in the base model (P <.203) making it an acceptable instrumental variable (good instruments are correlated with the variable of interest and uncorrelated with residuals generated by the base model). Using Lpredinst I generated predicted values for the political instability index (Instabboth) with the following model: Instabboth=β o +β 1 controlling variables in the base model+β 2 Lpredinst+β 3 Lpredinstmiss+u Controlling Variables in the Base Model When conducting the LPM and Tobit analysis I controlled for population specific and country specific variables believed to correlate with terrorist attacks (for definitions of population 4 The Chow test is an econometric test of whether the coefficients in two linear regressions on different data are equal. In this instance I demonstrate that a model incorporating the Predinst variable as an independent variable captures a statistically significant amount of variation beyond the variation captured by a model not incorporating the Predinst variable. 14

21 specific and country specific variables reference Appendix 2). Table 3 provides a list of population and country specific variables controlled for within the model. Table 3: Controlling Factors within the Models Population Specific Poverty Gap Ratios at US$1 (Povgap) Log of Poverty Gap Ratios at US$1 (Lpovgap) Gini Index (Gini) Post-Secondary School Enrollment Rates (Postsec) National Unemployment Rates (Unemp) National Unemployment Rates Dummy Variable; Equals one if data is missing (Unempmiss) Religion (Christian, Islam, and Otherrel) Proportion of the Population Age (Ageprop) Country Specific Country s Land Area (Area) Log of Country s Land Area (Larea) Country s Population Density (Popdens) Log of Country s Population Density (Lpopdens) Form of Government (Government) Form of Government Squared (Govsqr) Form of Government Dummy Variable; equals one if data is missing (Govmiss) Proportion of Population in Urban Areas (Urban) Predicted Instability Index (Instabboth) The Povgap variable captures the absolute number of people living on less than US$1 per day, while Gini captures how wealth is distributed within a nation, enabling me to capture both absolute and relative poverty levels. Additionally, evidence suggests terrorist organizations recruit in economically depressed areas indicating countries with high unemployment rates may have populations prone to radicalization. Recent trends indicate religiously based terrorist groups are increasing in number. 56 The dominant religion (religion that comprises more than 50 percent of the population) of a nation may provide an indication as to the likelihood its populations will radicalize. Finally, as mentioned earlier data tracked by the Bureau of Justice Statistics, the Columbian government and other think tanks indicates that younger populations may be predisposed to violence and would therefore make countries with large numbers of citizens between the ages of 15 and 34 more vulnerable to terrorism. 57 Regarding incorporation of country specific variables, countries with large geographic areas needing to be policed may have insufficient resources to do so effectively. Additionally, 15

22 densely populated countries with large proportions of their citizens living in cities may be more prone to terrorist attacks. Form of government variables may be significant because of the documented relationship between political violence and the political environment. Previous research shows that political violence falls in highly autocratic and highly democratic society. LPM and Tobit analyses assume that the marginal effect of the independent variable on the dependent variable is constant (i.e., the change caused in the dependent variable does not vary across all values of the independent variable). This assumption is false for nonlinear functions where a change in the dependent variable for a given change in the independent variable depends on the starting value of the independent variable (for example a change from zero to one in the Government variable may have a positive correlation with the dependent variable while a change from eight to nine may have a negative correlation with the dependent variable). 58 Previous research indicates that the relationship between political violence and form of government is parabolic and consequently I incorporated the form of government squared (Govsqr) variable in order to control for bias generated by this relationship. 59 Finally, I incorporate several dummy variables in order to control for variance in the dependent variable explained by observations missing data for one or more of the independent variables. While these variables lack substantive meaning in that they only indicate whether there is an underlying pattern to the missing data. If observations missing data are random they should not be statistically related to the dependent variable (missing observations would be randomly distributed and in effect cancel each other out). If this is not the case, it is possible that observations data are capturing variance in the dependent variable caused by some other factor. Incorporating dummy variables to control for this possibility enabled me to identify the significance of the Govmiss variable and conduct a more in depth study of observations missing 16

23 form of government data as they might relate with terrorist attack data (Reference Appendix 7 for the observations missing data controlled for by the Unempmiss and Govmiss variables). Methodological Challenges When conducting this analysis I confronted three methodological challenges. First, some of the interval variables in this model are not distributed normally (a key assumption of regression analysis is that data is normally distributed. If that is not the case then results will be inaccurate). Second, several of the independent variables are correlated, which may create multicollinearity within the model (if independent variables are highly correlated with each other it will distort the model and generate biased results). Third, my analysis addresses issues at a national level and not all variables have data for all countries. An additional consideration was the possibility that use of panel data for this analysis would create unobserved country specific factors that influenced results. In order to address this possibility I incorporated fixed effects transformations into the model (the fixed effects model is a statistical model that quantifies variation in the dependent variable that is attributed to characteristics of the units/observations under analysis). Results generated using this technique revealed that the fixed effects estimator is statistically significant but distorts both the LPM and Tobit models, thereby reducing their abilities to generate accurate predictions. Consequently, my final analysis does not incorporate a fixed effects term. Non-Normal Distribution of Data A methodological challenge confronting this analysis is correcting for non-normal distribution of the data. Analysis reveals that several variables are skewed, indicating that they are not normally distributed. Table 4 identifies those variables that are non-normally distributed and their associated skewness index values (the skewness index is a standardized measure of the balance of the distribution of data about the mean. Normally distributed data is located equally 17

24 both above and below the mean. The sign of the skewness value indicates the direction of the bias.) Typically, distributions are approximately normal if the skewness index has an absolute value of less than three (for complete analyses of descriptive statistics for all non-normally distributed variables refer to Appendix 4). Table 4: Skewness Index Values for Non-Normally Distributed Variables Variable Attacks Area Popdens Predinst The high skewness index values associated with each of the variables above indicates that they may bias the results of the analysis. I used the following approaches to adjust for this effect: 1. Attacks The non-normal distribution of the attack variable was controlled for through use of the Tobit model (The Tobit model corrects for bias generated by the non-normal distribution of data by multiplying estimators by a weighted value equal to the proportion of the data not equal to zero). 2. Area, Popdens, Predinst These variables are skewed to the right of the normal distribution and were transformed by calculating their log values resulting in the creation of the Larea, Lpopdens, and Lpredinst variables. I used two approaches to control for skewness because of the large number of observations with zero terrorist attacks. It is impossible to take the log of zero and a consequence of transforming the dependent variables using the technique would be a dramatic reduction in the number of observations incorporated into the analysis. Further, the Tobit model only adjusts for skewed dependent variables and cannot control for skewness in the independent variables. 18

25 Missing Data Table 5 and Table 6 summarize the number of countries reporting data for each of the independent variables incorporated into the model and how inclusion of additional variables influences the cumulative sample (Definitions of the variables listed below are included in Appendix 2): Table 5: Sample Observations, by Variable Variable Total N Cumulative Sample Ageprop Attacks Urban Instability Area Popdens Predinst Christian, Islam, Otherrel Government Unemp Postsec Gini Povgap 28 6 Table 6: Sample Observations, by Variable Variable Total N Cumulative Sample Ageprop Instability Urban Area Attacked Popdens Predinst Christian, Islam, Otherrel Government Unemp Postsec Gini Povgap I used three techniques to address missing data within my dataset. First, where regional data was available for variables (Urban, Unemp) I inputted those values for observations with missing data (Appendix 8 contains a list of observations with inputted values). 5 Second, several of the variables with large numbers of missing observations were eliminated during the stepwise analysis process (this process is discussed on page 21). Third, I assigned missing observations values of -11 (Government), -1 (Unemp), or 0 (Lpredinst) for variables not eliminated by the stepwise analysis and without regional data. I then created dummy variables equal to one (1) if 5 Two variables reported regional data (Urban and Unemp). I inputted regional unemployment data for 17 observations and regional urbanization data for 32 observations. 19

26 the original variables assumed the inputted values to control for bias created by inputting values for missing observations. Multicollinearity Use of the Pearson s r statistic (The Pearson s r statistic reflects the degree of linear relationship between two variables. It ranges from +1 to -1 and a correlation of +1 means that there is a perfect positive linear relationship between the two variables) shows that several pairs of variables are highly correlated, having an r statistic of greater than 0.5 (for a complete listing of bivariate correlations refer to Appendix 5). These correlations involve five variables Political Instability, Government, Urban, Post Secondary Enrollment, and the Gini index. Table 7 provides a summary of problematic variable relationships: Table 7: Summary of High Independent Variable Correlations Variable 1 Variable 2 Correlation Government >.7 Postsec >.5 Christian >.5 Ageprop >.3 Islam >.4 Predicted Political Instability Government Postsec Christian Islam >.5 >.5 >.4 Urban Povgap >.5 Postsec >.5 Postsec Povgap >.5 Ageprop >.5 Urban Ageprop >.3 Gini Ageprop >.4 While these relationships create the potential for multicollinearity within the model these interactions were resolved through the bivariate and stepwise analysis discussed later in this section. 20

27 Defining the Linear Probability and Tobit Model Specifications I used a combination of two analytical techniques to identify variables likely to have statistically significant interactions with the dependent variables. Bivariate analysis enabled me to eliminate variables that were uncorrelated with the dependent variable from the model while stepwise analysis enabled me to identify those variables that provided the most explanatory power to the LPM and Tobit models dependent variables. Bivariate and Stepwise Analysis First, I conducted bivariate analysis using the dependent variables and each independent variable individually. This process eliminated the Popdens, Povgap, Lpovgap, Postsec, Otherrel, and Ageprop variables from the LPM and Tobit models. Having eliminated the variables that failed to provide significant explanatory power, I used stepwise regression to identify the variables that provided the greatest explanatory power to the model. Based on this process, the variables in Table 8 were eliminated from the Tobit and LPM models (reference Appendix 6 for tables detailing the results of the stepwise process): Table 8: Variables Excluded from the Tobit and LMP Analyses LPM Tobit Christian Christian Gini Gini Islam Government Poptotal Instabboth Unemp Islam Unempmiss Poptotal Thus, I estimated the following unrestricted LPM and Tobit models using robust standard errors to control for heteroskedasticity within the model (heteroskedasticity dependent and/or varying distribution of residuals causes standard errors to be biased because it violates the assumption that residuals are independent and identically distributed. Robust standard errors relax that assumption making them more accurate when heteroskedasticity is present): 21

28 LPM: Tobit: Attacked=β o +β 1 Larea+β 2 Lpopdens+β 3 Urban+β 4 Government+β 5 Govsqr+β 6 Govmiss+ β 7 Y+β 8 Islam+β 9 Instabboth, robust Attacks=β o +β 1 Poptotal+β 2 Larea+β 3 Lpopdens+β 4 Urban+β 5 Unemp+β 6 Unempmiss+ β 7 Government+β 8 Govsqr+β 9 Govmiss+β 10 Y+β 11 Instabboth, robust ASSESSING THE LPM AND TOBIT MODELS In addition to the unrestricted models detailed above, I estimated three restricted models for both the LPM and Tobit analyses. I used these restricted models to ascertain the consistency of the results generated by the unrestricted model and to control for multicollinearity between the Government and Instabboth variables. Tables 9 and 10 summarize the Beta coefficients and associated P values for the LPM and Tobit models. Table 9: Coefficient Estimates for the Linear Probability Model with Robust Standard Errors Variable Restricted* Instabboth** Government*** Unrestricted**** Coeff P Coeff P Coeff P Coeff P Intercept Larea Lpopdens Urban Government Govsqr Govmiss Y Islam Instabboth * R 2 =.2734; ** R 2 =.2503; *** R 2 =.2656; **** R 2 =.2812 Table 10: Coefficient Estimates for the Tobit with Robust Standard Errors Variable Restricted* Instabboth** Government*** Unrestricted**** Coeff P Coeff P Coeff P Coeff P Intercept Poptotal Larea Lpopdens Urban Unemp Unempmiss Government Govsqr Govmiss Y Instabboth * Pseudo-R 2 =.0878; ** Pseudo-R 2 =.0883; *** Pseudo-R 2 =.0904; **** Pseudo- R 2 =

29 Given the high correlation between the Instabboth and Government variables ( r =.77) incorporating the two variables into a single model creates multicollinearity and results in biased standard errors. Therefore, I discarded both the unrestricted and restricted models for the LPM analysis, and the unrestricted model for the Tobit analysis. Having thus corrected for multicollinearity I selected the government model for both LPM and Tobit analyses because it maximizes the R 2 values, avoids multicollinearity, and has a concurrence rate (concurrence measures a model s ability to generate predictions that accurately forecast reality; in this case whether a country was or was not attacked) for the LPM framework. The data shown in Tables 9 and 10 demonstrate that the two models are consistent with one another. Many of the same variables provide significant explanatory in both the LPM and Tobit models. Nonetheless, these models do differ indicating that some factors may influence the aggregate number of attacks on a given country but not the overall probability that they will be attacked and vice versus. The following section assesses the influence statistically significant variables (variables with p values of <.05) have on the measures of terrorist attacks and explores the ramifications these results may have for future policy. As expected, there is substantial overlap between the significant variables in the LPM and Tobit models. The LPM framework identifies six variables associated with a significant (p <.05) portion of the variation in the Attacked dependent variable and the Tobit model identifies five variables that are associated with a significant portion of the variation in the Attacks dependent variable. Variables that are significant in both models include Larea, Urban, Lpopdens, and Govmiss. Variables fall into three general categories: country specific, form of government specific, and population specific factors. 23