Repression, Dissent, and the Onset of Civil War: States, Dissidents, And

Florida State University Libraries

Electronic Theses, Treatises and Dissertations The Graduate School

2008 Repression, Dissent, and the Onset of Civil War: States, Dissidents and the Production of Violent Conflict Joseph Young

Follow this and additional works at the FSU Digital Library. For more information, please contact [email protected] FLORIDA STATE UNIVERSITY

COLLEGE OF SOCIAL SCIENCES

REPRESSION, DISSENT, AND THE ONSET OF CIVIL WAR: STATES, DISSIDENTS AND THE PRODUCTION OF VIOLENT CONFLICT

JOSEPH K. YOUNG

A Dissertation submitted to the Department of Political Science in partial fulﬁllment of the requirements for the degree of Doctor of Philosophy

Degree Awarded: Summer Semester, 2008 The members of the Committee approve the Dissertation of Joseph K. Young defended on June 23rd, 2008.

Will H. Moore Professor Directing Dissertation

Richard Feiock Outside Committee Member

William Berry Committee Member

Dale L. Smith Committee Member

Jason Barabas Committee Member

The Oﬃce of Graduate Studies has veriﬁed and approved the above named committee members.

ii This dissertation is dedicated to my parents, Mark and Jora Young, who were my ﬁrst teachers, to my wife, Melissa Scholes Young, for her support, dedication and sense of humor throughout this process, and to my children, Isabelle and Piper, who help me remain focused on what is most important.

iii ACKNOWLEDGEMENTS

I am grateful to my advisor, Dr. Will H. Moore. The process of writing this dissertation, navigating the academic world, and securing a tenure-track position was challenging; he guided me through and remained committed even when I struggled. I also would like to express my gratitude to Dr. Dale L. Smith who counseled me, taught me about the discipline, and treated me like family. The dissertation is vastly improved because of the comments made by Dr. William Berry and Dr. Jason Barabas. They both provided an important perspective; both are outstanding scholars, methodologists, and people. I would also like to thank Dr. Paul Hensel for his kindness and encouragement throughout this long process. Finally, I want to acknowledge the Political Science Department as a whole as I believe that it is something larger than the sum of its parts.

iv TABLE OF CONTENTS

ListofTables...... vii

ListofFigures ...... ix

Abstract ...... xii

1. INTRODUCTION ...... 1

2. CONCEPTUALIZING STATE CAPACITY ...... 9 2.1 Introduction...... 9 2.2 Deﬁning State Capacity ...... 11 2.3 The Dimensions of State Capacity ...... 13 2.4 Conclusion ...... 21

3. A PROCESS THEORY OF CIVIL WAR ONSET ...... 24 3.1 Introduction...... 24 3.2 WhyCivilWar? ...... 26 3.3 Theory...... 27 3.4 PathtoCivilWarModel...... 37 3.5 Additional Implications ...... 38

4. TWO-STAGEEMPIRICALTESTS ...... 43 4.1 Introduction...... 43 4.2 ResearchDesign ...... 44 4.3 Estimation ...... 53 4.4 Conclusions ...... 62 4.5 Appendix ...... 64

5. STRUCTURAL EQUATION MODELING ...... 67 5.1 Introduction...... 67 5.2 Estimation ...... 69 5.3 Results ...... 72 5.4 Discussion...... 78 5.5 Conclusion ...... 79

v 6. CIVIL WAR, CAUSAL INFERENCES, AND PROPENSITY SCORE MATCH- ING...... 81 6.1 Introduction...... 81 6.2 Preprocessing Civil War Data ...... 87 6.3 Matching Using Time-Series Cross-Sectional Data ...... 91 6.4 Conclusion ...... 95

7. INSURGENCY AND COUNTERINSURGENCY: UNDERSTANDING THE MICRODYNAMICS OF STATE-DISSIDENT INTERACTION ...... 97 7.1 Introduction...... 97 7.2 Explaining Modern Insurgency ...... 98 7.3 Support and Insurgency ...... 99 7.4 Iraq and Counterinsurgency ...... 105 7.5 Counterinsurgency Operations in Iraq ...... 109 7.6 ResearchDesign ...... 109 7.7 Estimation ...... 115 7.8 Results and Discussion ...... 117 7.9 Conclusions ...... 120 7.10 Appendix ...... 121

8. INTERNATIONAL RELATIONS AND CIVIL WAR ...... 128 8.1 Introduction...... 128 8.2 State Capacity and International Relations ...... 130 8.3 External Conﬂict and Civil War Onset ...... 134 8.4 The International System and Civil War Onset ...... 135 8.5 ResearchDesign ...... 136 8.6 Results ...... 140 8.7 Conclusions ...... 145 8.8 Appendix ...... 145

9. CONCLUSION...... 149

BIOGRAPHICALSKETCH ...... 168

vi LIST OF TABLES

2.1 Attributes of State Capacity ...... 23

4.1 Summary Statistics ...... 52

4.2 Summary Statistics–Imputed Data ...... 53

4.3 Impact of Resources, Job Insecurity, Previous Repression and Previous Dissi- dent Activity on Repression ...... 56

4.4 Impact of Support, Military Capacity, Previous Dissident Activity, and Previous Repression on Dissident Activity ...... 57

4.5 Impact of Repression, Dissident Activity, and Residuals from Previous Equa- tions on Civil War Onset (Unrestricted Model) ...... 58

4.6 Impact of Repression and Dissident Activity on Civil War Onset (Restricted Model) ...... 59

4.7 Impact of Fearon and Laitin Variables on Civil War Onset Using Imputed DataFrom1976–1999 ...... 64 4.8 Impact of Fearon and Laitin Variables on Civil War Onset using Non-Imputed Datafrom1976-1999 ...... 65

4.9 Civil War Onset–PRIO/Uppsala Coding of Civil War ...... 65

5.1 R2 forEachDependentVariable...... 72

5.2 Expectations for Variables From Theory ...... 73

5.3 Expectations and Outcomes for Variables From Theory ...... 77 6.1 Comparing Means in Unmatched Sample ...... 88

6.2 Comparing Matched and Unmatched Samples ...... 89

6.3 Civil War–Matched and Unmatched Samples ...... 90

6.4 Comparing Year Matched and Unmatched Samples ...... 92

vii 6.5 Civil War–Matched and Year Matched Samples ...... 94

7.1 Count of Counterinsurgent Operation Days in the Regions of Iraq ...... 109 7.2 Summary Statistics for Central Region ...... 115

7.3 Summary Statistics for South Region ...... 115

7.4 Summary Statistics for North Region ...... 116

7.5 Poisson Auto Regressive Model–Central Region ...... 118 7.6 Negative Binomial Models ...... 119

7.7 ARMAModels ...... 126 7.8 Negative Binomial–All Regions ...... 127 8.1 Civil War Onset and International Relations ...... 139

8.2 Probability of Civil War ...... 144 8.3 Samplefrom1945to1999 ...... 147 8.4 International Relations Variables and Spatial Lags ...... 148

viii LIST OF FIGURES

2.1 Percentage of Observations Experiencing Onset of Civil War Given Varying LevelsofGDPPerCapita ...... 10

2.2 Spectrum of Societal Support and Policy Implementation Costs ...... 14

2.3 Societal Support at the Individual Level: Distance Between Individual Citi- zen’s and the State’s Preferred Policy ...... 15

2.4 The Dimensions of State Capacity ...... 18

2.5 Three-Level Concept of State Capacity ...... 20

3.1 Hypothesized Relationship between Job Insecurity and Repression and Re- pression and Societal Support ...... 29 3.2 The Distribution of Societal Support ...... 33

3.3 Spectrum of Societal Support and Policy Implementation Costs ...... 34

3.4 Hypothesized Relationship among Dissidents, Dissident Acts, and Societal Support ...... 35

3.5 Hypothesized Relationship between Resources and Expected Success of Dis- sident Activities and the Expected Success of Dissident Activities and the Number of Dissident Acts ...... 36

3.6 RepressionBlock ...... 39

3.7 Dissident Activity Block ...... 39

3.8 Societal Support Block ...... 40

3.9 Job Insecurity Block ...... 40

3.10CausalModelofCivilWar...... 41

4.1 Predicted Probability of Civil War–Repression at Mean ...... 60

4.2 Predicted Probability of Civil War–Repression at Maximum ...... 61

ix 4.3 Comparing Regression Coeﬃcients ...... 66

5.1 PathModelofCivilWarOnset ...... 68 5.2 Direct and Indirect Eﬀects ...... 69 5.3 Coeﬃcient Estimates for Civil War Onset Predictors ...... 74

5.4 Coeﬃcient Estimates for Repression Predictors ...... 75

5.5 Coeﬃcient Estimates for Dissident Activity Predictors ...... 76 5.6 Coeﬃcient Estimates for Societal Support Predictors ...... 76

5.7 FinalPathModel...... 78 6.1 Matching on the Propensity to be a State with High Repression ...... 85 6.2 Comparison of Reduction in Bias Between Matching Techniques ...... 93

6.3 Comparison of Coeﬃcients Between Matched Imputed Data and Matched Non-Imputed Data ...... 96

7.1 High Support for Counterinsurgency Operations ...... 100

7.2 Medium Support for Counterinsurgency Operations ...... 101

7.3 Low Support for Counterinsurgency Operations ...... 101

7.4 A Continuum of Force used in Counterinsurgency ...... 102 7.5 The Eﬀect that Counterinsurgency has on Insurgent Attacks Given High Support for the Counterinsurgents ...... 104

7.6 The Eﬀect that Counterinsurgency has on Insurgent Attacks Given Low Support for the Counterinsurgents ...... 105 7.7 Stylized Support for Counterinsurgency among Contending Groups in Iraq . 107

7.8 Daily Count of Coalition Deaths in the Central Region of Iraq ...... 111 7.9 Daily Count of Coalition Deaths in the Northern Region of Iraq ...... 112

7.10 Daily Count of Coalition Deaths in the Southern Region of Iraq ...... 112

7.11 Frequency Distribution of Coalition Deaths ...... 123

8.1 Three Level Concept of State Capacity ...... 131 8.2 The Eﬀect of a Defensive Alliance on the Probability of Civil War Onset . . 141

8.3 The Likelihood of Civil War Onset when Foreign Policy Similarity is Zero . . 142

x 8.4 The Likelihood of Civil War Onset when Foreign Policy Similarity is One . . 142

8.5 The Eﬀect of MIDs on the Probability of Civil War Onset ...... 143 8.6 The Likelihood of Civil War Onset when System Power Concentration is Stable144

8.7 The Likelihood of Civil War Onset when System Power Concentration Shifts (Maximum) ...... 144

xi ABSTRACT

The prevailing wisdom among scholars of civil war is that weak states, or resource-poor states, are the most prone to this form of political violence. Yet, a large portion of resource- poor states never experience civil war. What can account for why resource-poor states, like El Salvador, are prone to civil war while resource-poor states, such as Bhutan, are not? I oﬀer a theory of civil war onset that explains how dissidents and states interact to produce civil war. This theory moves beyond structural explanations and explains how the choices of states and dissidents jointly produce violence. From the theory, I derive the expectation that states that repress their citizens are the most likely to kill citizens and to generate dissident violence. In short, the resolution to the puzzle is: State leaders from resource-poor states, who choose to repress, are the most likely to generate violence that exceeds the civil war threshold. This insight not only resolves an academic puzzle but when tested provides a model with better in-sample prediction of civil war than previous models.

After explicating the theory and discussing concepts, I empirically evaluate the hypotheses implied by these arguments using a large cross-national dataset including a global sample from 1975 to 1999. I utilize structural equation modeling as well as two-stage procedures to estimate the direct and indirect effects of variables outlined in the theory. Using a novel approach to reducing bias in my data, time-dependent propensity score matching, I isolate the causal effects of repression on a state’s likelihood of experiencing civil war. I then extend the insights of the model to other forms of political violence including interstate conflict and insurgency and offer hypotheses relating to current debates over counterinsurgency policy and the relationship between state making and interstate war.

xii CHAPTER 1

INTRODUCTION

Since the Peloponnesian War, observers of conflict have attempted to understand its causes. The Great Wars or World Wars at the beginning of the 20th century renewed academic interest in exploring the causes of war in hopes of preventing future world-wide conflicts. The Correlates of War project, begun in the 1960s, attempted to explain international conflict based on the assumptions that “there are indeed regularities in the origins of different types of wars, and that they are discoverable...those regularities are relevant to the future as well as the past...[and] knowledge generated by this...research will also help in showing us how to apply and utilize whatever discoveries are made” (Singer, 1972, 244). In addition to international war, understanding terrorism, civil war, and other violent forms of political contention is a recently salient academic concern. Civil war is a particularly deadly form of conflict that has claimed the lives over 16 million people since 1945 (Fearon and Laitin, 2003). Civil war has lasting impacts well after the conflict concludes. Because of the destruction of infrastructure and public health systems, Ghobrah, Huth and Russett (2003, 200) find that civil wars “raise the the subsequent risk of death and disability from many infectious diseases,” and possibly “increase...homicide, transportation accidents, other injuries, and cervical cancer.” Like most academic questions, there are a variety of contending explanations for why civil war begins. Resolving these disputes is critical as the causes are linked to the solutions for mitigating the likelihood that civil war occurs. After the fall of the Soviet Union and the subsequent interest in civil war, especially wars with an ethnic component, conventional wisdom developed concerning the onset of civil war. The standard story claimed that Soviet- style states kept the lid on deep seated ethnic differences; and when the lid was removed, the pot boiled over (Kaplan, 2001). Huntington (1993, 22) made a related claim, arguing that

1 “the principal conflicts of global politics will occur between nations and groups of different civilizations.” Ethnic arguments struggle to explain why in some circumstances ethnic differences are associated with conflict whereas in other places they are not. More importantly, as Fearon and Laitin (1996, 715) argue, “a good theory of ethnic conflict should be able to explain why...peaceful and cooperative relations are by far the more typical outcome than is large- scale violence.” Posen (1993) offers one potential resolution to this puzzle: when a state can no longer provide internal security, a state of anarchy, not unlike the one characterizing the structure of the international system, creates incentives for ethnic groups to provide their own security. The arming of any particular ethnic groups for defense leads to insecurity and the arming of other ethnic groups. This process explains how ethnicity can lead to war. It is not clear, however, why ethnicity is the relevant dimension over other potential ways individuals may mobilize. Elbadawi and Sambanis (2002) find empirical support for the claim that ethnicity helps explain the prevalence of civil war. Sambanis (2001), however, disaggregates civil violence and makes the claim that ethnic/religious, or identity, conflicts have different causes than non-identity conflicts. While popular accounts of ethnicity’s effect on civil war have been influential, empirical evidence is mixed and recent work casts doubt on this relationship (Fearon and Laitin, 2003). Other academic work, supported by the World Bank, claims that identity, or a larger category dubbed grievance can not explain civil war. Instead, Collier and Hoeffler (1998, 563) claim that, civil “war occurs if the incentive for rebellion is sufficiently large relative to the costs.” According to Collier and Hoeffler (2001, 564), all states have groups who feel aggrieved, but “the incidence of rebellion is not explained by motive, but by the atypical circumstances that generate profitable opportunities.” Opportunities for predation occur, according to Collier and his colleagues, when rebels can easily procure diamonds, oil, or some other valuable natural resource. These opportunities are maximized and “rebellions may occur when foregone income is unusually low” (Collier and Hoeffler, 2001, 569). In sum, the causal mechanism for the relationship between so-called ‘lootable’ resources and civil conflict relates to making the value of controlling the state high relative to the costs for attempting to take it. These mechanisms, however, offer weak microfoundations; it is unclear how aggregated individual choice under constraint leads to the specified outcome. Macro-structures, like the lack of state revenues and the availability of lootable resources,

2 might provide context for where rebellion is likely, but these structures do not explain why some group choses to rebel while another does not and can not explain when it will happen. More importantly, many states have these structural conditions and do not produce civil war. Fearon and Laitin (2003, 75) build on the opportunity logic and claim that “the main factors determining both the secular trend and the cross-sectional variation in civil violence...are not...broadly held grievances but, rather, conditions that favor insurgency.” In brief, structural characteristics provide opportunities for potentially successful violent mobilization against the state. Similar to the so-called greed explanations, Fearon and Laitin assert that state weakness is associated with states that experience civil war, but their approach also lacks the ability to predict when. In all of the opportunity approaches to understanding civil war, the state is a passive or nonexistent actor without any clearly defined preferences or incentives.1 It is unclear whether works in this tradition treat the state in the Marxist fashion, and therefore conceive of the state as merely a tool of the societal elite, or whether the state is relegated to a background condition that marginally affects rebel incentives. Either way, the state is a passive actor that does not make its own choices. On the whole, both opportunity and ethnicity are influential approaches to understanding civil war. Each can explain some variation in why some state can experience this type of violence, but both lead to unanswered questions. Opportunity approaches can sort countries at risk from those that are relatively safe from civil war, but they can not tell us when civil war may occur. Ethnicity and religious differences have the same problem. In the empirical work that attempts to look at the relationship between ethnicity and civil war, most of the measures that scholars use have the same value over time. In other words, the fact that Belgium has two major ethnic groups has not changed much over time. For this fact to predict when civil war should happen, something else must change. Neither approach explicitly treats the state as an actor with interests of its own. While most civil war scholars and scholars of revolution (Skocpol, 1979; Goodwin, 2001) begin their analysis by looking at

1I prefer to use the term opportunity over greed. First, this is consistent with some of the work that claims that all political behavior needs both willingness and opportunity of the agent (Cioﬃ-Revilla and Starr, 1995). In other words, an individual must have the motivation and means to act. States usually provide the political opportunity for civilians to act, but other motivations are necessary for these citizens to want to act.

3 the state, neither opportunity nor ethnicity approaches tell us much about the preferences of this crucial actor. In this dissertation, I develop a model of the state and its interactions with citizens and dissidents to make predictions about where and when civil war may occur. Opportunity approaches, as I note above, explain which states are most susceptible to civil war. I incorporate some of the insights from this approach when attempting to understand how the resources available to the state affect opportunities for organized violence. I improve on these arguments by offering an explanation for the individual motivations of the leaders of the state. I argue that certain leaders, who’s job tenure is threatened, are incentivized to use violence in an attempt to maintain power. State repression, spurred by a concern for job security, starts a process of violence between state and dissidents that culminates in civil war.2 My theory incorporates both the reasons for the opportunities for rebellion as well as the willingness of dissidents to use violence against the state. The theory that I offer expands the understanding of civil war in three ways. First, I offer a formal definition of civil war that identifies the need to think about this phenomena as a process of violence. While the formal definition I offer is consistent with the way that most scholars view the concept, my approach directs attention towards investigating how states and dissidents interact to produce civil war. Second, I argue for a multidimensional conceptualization of state capacity. Previous opportunity arguments equate state weakness with the lack of resources. I argue that state weakness, or state capacity more generally, is a function of resources and the support that the state receives from society. This conceptualization can help explain why some resource-poor states produce civil war while others do not. Third, I provide a statistical model that improves our ability to predict civil war. Since, by definition, civil war occurs when the deaths from political violence between states and dissidents exceeds a given threshold, I argue for a different approach to modeling the process. Using two-stage and structural equation models that treat dissident violence and state repression as endogenous, I show that the effect that structural conditions have on the likelihood of civil war is mediated by states and dissidents acting. Because I use multiple approaches to estimation, I am more confident that the results are less dependent

2Goodwin (2001) similarly argues that both state weakness and repression are associated with the onset of insurgency or what he calls revolution. However, as I explain later, his account is distinct from mine in a number of ways. Most importantly, his state-centered approach fails to explain the microfoundations that my theory oﬀers.

4 upon assumptions from these statistical models. Beyond these three contributions, I also investigate the process of violence at the microlevel using original data from the current conflict in Iraq (Chapter 7). In Chapter 6, I suggest a new technique for drawing causal inferences when using time-series cross- sectional data. Finally, I connect the study of civil war and the process of violence leading to this outcome with international relations and show how the international system, external conflict, and alliances affect the likelihood of civil war for a state (Chapter 8). I begin the dissertation by discussing state capacity and its relationship to civil war. In Chapter 2, I define the state and state capacity and then compare my approach with other scholars who are interested in the relationship between state strength and political violence. Next, I discuss how state capacity relates to civil war and why the current conceptualization is inadequate. Third, I offer a conceptualization of state capacity that includes both an economic and political basis for state power based on the dimensions–resources and societal support. Building on Goertz (2006), I construct what is termed a three-level concept of state capacity that includes the dimensions of capacity and potential indicators that could proxy for these dimensions. The goal of this conceptualization is to generate new hypotheses that can help explain the variation among weaker states that do or do not experience civil war. In Chapter 3, I develop a process theory of civil war. The prevailing wisdom among scholars of civil war is that weak states, or resource-poor states, are the most prone to this form of political violence. Yet, a large portion of resource-poor states never experience civil war. What can account for why resource-poor states, like El Salvador, are prone to civil war while resource-poor states, such as Bhutan, are not? I offer a theory of civil war onset that explains how dissidents and states interact to produce civil war. This theory moves beyond structural explanations and explains how the choices of states and dissidents jointly produce violence. From this theory, I derive the expectation that states who repress their citizens are the most likely to kill citizens and to generate dissident violence. Consequently, the resolution to the puzzle is: State leaders from resource-poor states, who choose to repress, are the most likely to generate violence that exceeds the civil war threshold. In Chapters 4 through 6, I empirically test hypotheses derived from the theory developed in Chapter 3. Since the theory that I offer suggests that repression and dissent mediate the relationship between structural covariates and the onset of civil war, standard single-equation approaches to estimation will not produce unbiased estimates. Hibbs (1973, 5), writing well-

5 before the current literature on civil war, claims that “the causal processes which are at the root of cross-national differences in levels of violence are much too complex to be adequately captured by a single-equation formulation.” Following Hibbs, I estimate the relationships outlined in the theory using a variety of means. First, I estimate two-stage models that take into account the relationship between variables like democracy and mountainous terrain on repression before estimating repression’s effect on civil war onset. This approach allows me to use information from the repression and dissent equations when estimating the likelihood of civil war. Conventional two-stage models are most often used with continuous endogenous variables. In this case, civil war onset is a limited dependent variable and requires logit or probit estimation. Because of this, I estimate a two-stage conditional maximum likelihood model that takes the binary dependent variable into account. Next, in Chapter 5, I estimate a structural equation model that can also deal with a binary endogenous variable. The advantage of this form of estimation is twofold: I can see if the relationships are consistent across different modeling choices with different assumptions, and I can estimate the indirect effect of variables on the onset of civil war. If results hold across the different modeling strategies, then I can be more confident that the results are not due to assumptions embedded in the models. The indirect effect that variables have on the likelihood of civil war are those effects that a particular variable has through another variable. If ethnicity increases dissident activity, which in turn increases the likelihood of civil war, then by using this approach, I can estimate this effect. In Chapter 6, I attempt to better isolate the causal effect that repression has on the onset of civil war by creating a more experimental-like condition where I can compare a group of observations that are treated with high levels of repression to a control group. I use a statistical technique, propensity score matching, to create groups that more closely resemble randomly selected treatment and control groups. Although most matching occurs in cross-sectional data, I offer an potential innovation that extends propensity score matching to studies that employ time-series data. While the empirical tests are all conducted on large cross-national data, I also consider conflict within a single state to extend some of the propositions from the theory and to apply the theory to other forms of political violence. Since Biblical times, strong states or empires have battled insurgencies who use unconventional warfare to combat the power of their opponent. Why have some of these attempts at counterinsurgency been successful

6 while others have failed? This question is particularly salient in present day Iraq. One of the primary tactics Americans are using to diffuse the Iraqi insurgency is direct military strikes to reduce the capabilities of insurgents. Using large scale detentions and military operations, the US has attempted to reduce the amount of violence by reducing the supply of insurgents. I argue in this chapter that this strategy has not achieved the desired results in Iraq because the US and its partners have not taken into account that different operations affect populations differently. Building on the process theory of violence from Chapter 3, I show how different levels of support for counterinsurgency help explain why some forms of counterinsurgency actually stoke–rather than mitigate–violence. After describing support’s effect on the efficacy of counterinsurgency, I evaluate the arguments regarding counterinsurgency using an original data set of US counterinsurgent operations in Iraq. In the conclusion, I offer policy recommendations for the Iraqi case. In developing a theory of civil war onset that is built on the relationship between states and dissidents, I make some necessary simplifying assumptions. For example, I assumed that external actors had no influence on the process of violence that leads to civil war. In Chapter 8, I relax this assumption, and investigate how alliances, external conflict, and other factors influence the violent relations between states and dissidents. While many studies look at how intervention by third parties affects the duration of civil war, few look at how external actors affect the onset of civil war. Adding different actors to the model raises issues concerning how international relations relate to civil war onset. Recent research explores the transnational factors affecting civil war focusing on the role of bad neighbors, refugees, and other region specific processes. While most agree that conflict processes do not respect physical borders, few studies explain how international relations affects civil war. Relatively little is known about how the character of the international system affects the likelihood of domestic conflict. Using insights from international relations and the process theory of civil war outlined in Chapter 3, I identify several hypotheses related to the international system and civil war. I evaluate these hypotheses using a times-series cross-sectional research design, confirming that a reduction in the concentration of power in the international system increases the likelihood of civil war, that increased external conflict in the form of militarized interstate disputes increases the likelihood of civil war, and that having foreign policy similarities with great powers reduces the likelihood of civil war.

7 In Chapter 9, the conclusion, I discuss some of the policy implications. As I mentioned above, the study of civil war has direct implications for how leaders might reduce the risk of civil war for their own state as well as for other states in the international system. The implications of the theory that I oﬀer as well as the supporting empirical evidence suggest that repression serves as an early-warning sign for civil war. While resource-poor states are at risk, leaders who face job insecurity within these states are the most prone to use repression leading to civil war. Finally, I discuss some of the limitations of the project and potential extensions of the project to other topics such as terrorism.

8 CHAPTER 2

CONCEPTUALIZING STATE CAPACITY

2.1 Introduction

Why do civil wars begin? Ethnic diversity, inequality, and opportunity are all potential answers to this question. The most prominent explanation in the current research on civil war relates to state power.1 Current scholarship on the relationship between the power of the state, or state capacity, and the onset of civil war emphasizes resources as the primary component of state strength.2 According to Fearon and Laitin (2003, 80) higher state resources per capita “should be associated with a lower risk of civil war onset because...it is a proxy for a state’s overall ﬁnancial, administrative, police, and military capabilities.” States with high resources are also able to ‘see’ their populations because they have bureaucratic penetration of society. Resources allow states to build roads, to establish police stations, to take a census, and to engage in other tasks that increase the state’s vision. Subsequent studies conﬁrm this basic relationship and focus on trying to uncover other relationships (e.g. Hegre and Sambanis, 2006).3 Figure 2.1 displays a graph of the percentage of the total observations for civil war onset given varying levels of Gross Domestic Product (GDP) per capita. Additionally, it shows the percentage of country-year observations where no onset occurs per category of GDP per capita.4 As the chart shows, while many of the non- onset years consist of countries with a very low GDP per capita (0-$1000), a disproportionate number of civil war onsets come from this group.5 At the upper end of the income categories,

1In the context of interstate war, Kugler and Domke (1986, 39) assert that “[p]ower plays a role in a wide range of relationships but is perhaps most critical for the study of conﬂict.” 2I use the terms capacity, power, strength, and capabilities interchangeably. 3One such study attempt to unpack the complex relationship between ethnicity and civil war onset (Blimes, 2006). 4In other words, the chart is a frequency distribution broken down across onset and non-onset years. 5This amounts to nearly 50% of the total number of onsets or 51 out of 111 in the Fearon and Laitin

9 0.6

0.5

0.4

0.3

0.2 Percentage of Observations

0.1

0 0-1000 1001-2000 2001-3000 3001-4000 4001-5000 5001-6000 >6000 Gross Domestic Product Per Capita

% Total Onsets % Total No Onsets

Figure 2.1: Percentage of Observations Experiencing Onset of Civil War Given Varying Levels of GDP Per Capita

where GDP per capita is greater than $6,000 per year, civil war onset is an extremely rare occurrence. Since there are many countries in this category and thus many observations, is is surprising how few onsets occur.6 The negative relationship between state resources, or low GDP per capita, and the onset of civil war is robust across diﬀerent lists of civil wars as well as diﬀerent sets of covariates (Fearon and Laitin, 2003; Sambanis, 2004).7 Although this relationship is well established,

(2003) data. 6Only six states with a GDP per capita over $6,000 experienced civil war onset since 1945. 7This is setting aside the notion of the ‘resource curse’ for now. Resources in the previous studies mean national income or the income available for taxation by a state. The eﬀect that resources such as oil or

10 civil war scholars can not explain why some states with resources experience civil war while many others do not. In the figure above, over 25% of the cases that did not experience civil war onset had a GDP per capita of less than $1,000. Even more troubling for a solely resource-based explanation is that GDP is what can be termed a ‘near time-invariant variable’.8 In other words, GDP is slow moving and present observations are highly correlated with the recent past. Using GDP to predict civil war may be helpful in sorting among states. What it can not predict is when civil war will occur. To understand civil war onset, we need a theory that explains who (which states) and when (which year). In this chapter, I define the state, state capacity, and compare my approach with other scholars who are interested in the relationship between state strength and political violence. Next, I discuss how state capacity relates to civil war and why the current conceptualization is inadequate. Third, I offer a conceptualization of state capacity that includes both an economic and political basis for state power based on the dimensions–resources and societal support. Building on Goertz (2006), I construct what is termed a three-level concept of state capacity that includes the dimensions of capacity and potential indicators that could be used to measure these dimensions. The goal of this conceptualization is to generate new hypotheses that can help explain the variation among weaker states that do or not experience civil war. 2.2 Defining State Capacity

Max Weber’s definition of the state as the organization that monopolizes the legitimate use of force within a territory is the most influential definition of the state. As Skocpol (1985, 7) notes, states are “compulsory associations” made up of “[a]dministrative, legal, extractive, and coercive” organizations. Citing Stepan (1979), Skocpol (1985) claims that states are more than governments and not simply the tool of the ruling elites in society. This statist approach contrasts with Marxists who assume that the state does not have its own interests, nor autonomy, and is merely an arm of the dominant class. According to Skocpol (1985, 8), the benefit of the statist approach is that it allows the analyst to compare and contrast the state with other organizations and agents in society and “requires us to see [the state] as diamonds has on the likelihood of civil war is explored in Ross (2004); Snyder and Bhavnani (2005); Ross (2006); Snyder (2006). 8See Plumper and Troeger (2007) for a discussion of the problems of estimation when using time invariant and rarely changing variables.

11 much more than a mere arena in which social groups make demands in political struggles or compromises.”9 The modern state has a multitude of tasks and operates in many arenas. Among other things, states tax, determine rules, and enforce laws. Whether states and the leaders who govern these political units are able to change or to enforce rules in any given area is based on their capabilities.10 I define state capacity as the ability of the individuals comprising the state to enact policy outputs consistent with their preferences. This definition is similar to Geddes (1995, 14), who defines state capacity as the state’s ability to “translate preferences into action.” Tilly (2003, 41), among others, defines the state’s capacity as the “extent to which governmental agents control resources, activities, and populations within the [state’s] territory.”11 This definition of state capacity potentially leads to tautology when attempting to use the concept as an independent variable to explain phenomena, such as civil war. Inherent in concepts of political violence is the notion that a particular segment of society is not able to be controlled. Otherwise, the state could make the costs of violence prohibitive. This same logic holds for resources. Without some material ability to challenge the state, violence can not occur. Both the definition that I offer and Geddes’ are analytically distinct from political violence. As I will show in Chapter 3, whether a state is able to implement its preferred policies is related to civil war, but it is not related by fiat or definition.

2.2.1 State Capacity and Civil War

Definitions of state capacity also assume that capacity is a continuum ranging from weak to strong.12 States that are more able to implement their preferred policies are stronger than states that struggle to make public policy.13 Collier and Hoeffler (2001, 569) cite military capabilities as an important dimension of state capacity related to civil war onset. They claim that “atypical weak government military capability” provides the ripe opportunity 9This movement to bring the state back in conceives of the state as an independent variable that influences a phenomena of interest rather than just as a dependent variable that is explained by prior causes. See the essays in Skocpol (1985) for a discussion of this approach. 10The capacity of states to extract, use the military, or perform some other function are often discussed as if these are different kinds of capacity worthy of different causal explanations. I assume that the power of the state is essentially fungible or can be transferred to any policy area that the leaders of the state wish to change or enforce. Therefore, I only discuss capacity and not a host of potential varieties of capacity. 11Tilly uses the term government rather than state. 12Nettl (1968) was one of the first to make this distinction between strong and weak states. 13At times, I use the terms weak and strong to denote a relation not an absolute state of being. In other words, strong states can implement preferred policies more easily than weak states.

12 for rebels to pursue civil war. By this logic, weak states, or states that have fewer military capabilities, are more likely to provide opportunities for rebellion. As outlined above, Fearon and Laitin (2003, 75-76) also suggest that weak states “render insurgency more feasible and attractive due to weak local policing or inept and corrupt counterinsurgency.” They claim, moreover, that “brutal and indiscriminate retaliation...helps drive noncombatant locals into rebel forces” (76). For Fearon and Laitin, this brutal and indiscriminate response is caused by structural conditions that make it difficult for counterinsurgents to be effective. In sum, although Fearon and Laitin and Collier and Hoeffler offer slightly different causal mechanisms,14 both authors hypothesize that the weaker the capacity of the state, the more susceptible it is to civil war. In both cases as well, a single dimension–resources–determines whether a state is more or less capable.15

2.3 The Dimensions of State Capacity

As discussed above, using only the resource dimension of state capacity yields a concept that tends to covary with civil war onset. As Figure 2.1 demonstrates empirically, most civil wars occur in states that are resource poor, and states with the most resources have few civil wars. Current theories help sort among states and ﬁnd that resource poor states are most likely to produce civil war. To link state capacity to civil war, however, we must also explain when civil war will occur. In addition, we need to sort the resource-poor states who are civil-war prone and those that are not. Providing a more complete conceptualization of state capacity is one way to attempt to improve the explanatory power of a theory that relates state capacity to civil war. The dimensions of concepts direct the analyst towards hypotheses relating that concept to other concepts. Choosing dimensions that generate empirically veriﬁable hypotheses is thus one way to distinguish whether a particular conceptualization is more useful than another (Goertz, 2006). I argue that the concept state capacity has two dimensions: resources and societal support.16 In other words, a state may have abundant resources but this alone does

14I follow Little (1991) and define causal mechanisms as “a series of events governed by lawlike regularities that lead from the explanans to the explanandum.” In other words, a causal mechanism is a chain of events linking cause and effect. 15As previously mentioned, resources are linked to the onset of civil war through different processes. 16Many other scholars offer different components of state capacity. The dimensions I choose are relevant then for hypotheses, explanations and causal mechanisms. The development of the concept in this manner is itself an exercise in theory about which attributes constitute the fundamental basis for this concept

13 Figure 2.2: Spectrum of Societal Support and Policy Implementation Costs

not determine its ability to implement policy. Resources are the economic basis of state capacity17 while societal support is the political component of state capacity.18 Other works on conceptualizing state capacity identify both the economic and political sources of state power (Kugler and Domke, 1986; Jaggers, 1992). Kugler and Domke (1986, 42) claim that strong states or states that tend to win interstate war “are those who have the resources and the political capacity to mobilize and maintain a war effort.” Jaggers (1992, 29) argues that “state building [or the state’s ability to accumulate power] requires not only a high level of national capabilities but also a high level of political capacity.” Both authors are arguing for a conception of state capacity that includes an economic and political basis for state power to better explain the relationship between the concept and conflict. I define societal support as the costs of implementing policy imposed on the state by

(Goertz, 2006). In the appendix to the chapter I provide a chart of the dimensions involved in a sample of conceptualizations of state capacity. 17Measuring economic strength is easier, of course, than measuring political power. See (Lukes, 1974) for a discussion related to the difficulty of measuring political power. 18Jackman (1993) calls societal support legitimacy. He claims that “a regime is legitimate to the extent that it can induce a measure of compliance from most people without the resort to the use of physical force” (98). In contrast, Eckstein (1971) defines legitimacy as “the extent to which a polity is regarded by its members as worthy of support” (50). To the extent that legitimacy matches Jackman’s definition, I conceive of support as analogous to legitimacy. Eckstein’s concern with a state being worthy of support is further from my conception of support. Based on my conception, support is generated through habit as opposed to some ideological or affective manner.

14 Figure 2.3: Societal Support at the Individual Level: Distance Between Individual Citizen’s and the State’s Preferred Policy

citizens in a polity.19 Societal support is a continuous concept where high levels make government policies less costly to enact than when support is low.20 Figure 2.2 displays a spectrum of societal support that ranges from negative to positive support. This spectrum also inversely relates to the costs incurred by a state in an attempt to implement its preferred policies. I assume that the greater the distance between an individual’s preferred policy and a state’s preferred policy, the less support the state receives from the individual.21 In Figure 2.3, leaders of a state have a preferred policy outcome and two citizens, A and B, have diﬀerent preferences. Whether the citizen is to the left (L in Figure 2.3) or right (R in Figure 2.3) is unimportant. What matters is that the further the citizen is from the state, the less support it has for the policy. In this case, citizen A’s preferred policy is closer to the state’s policy and, thus, they support the state more. Individual support aggregated at the societal level generates a distribution of support.22

19This definition is similar to Nordlinger (1987, 368) or “societal support...[is] the noncoerced behavior of societal actors that is on balance politically supportive of the state’s preferences and the ‘corresponding’ public officials.” Nordlinger’s definition, however, does not identify a negative pole for the concept constraining support to only the positive end. In my conceptualization, I offer a negative pole that includes citizens who might need to be coerced into accepting the state’s policy outputs. Technically then, support is the inverse of these costs the state must incur. 20I describe this process in more detail in Chapter 3. 21The particular dimension over which the participants disagree is sufficiently general. It may be a left- right spectrum, liberal-conservative, hawk-dove, or any other dimension where citizens may have differing preferences. 22As Little (1991, 39) notes, this explanation is called aggregative as it “attempts to account for social patterns as the aggregate result of the rational actions performed by large numbers of participants.”

15 The mean of this distribution provides an estimate of how much support for the state there is in any given society.23 Support is an important dimension of state capacity as the more support a state receives, the more of its preferred policies it can implement. As Mason and Krane (1989) demonstrate, state reliance on force in generating citizen compliance with government policy is more costly then using political means. Some scholars, such as Jackman (1993), distinguish power from force. For Jackman (1993, 30) the exercise of force is...an admission that compliance cannot be induced by other noncoercive means.”24 I do not necessarily divide them into separate concepts. Instead, force is a signal of low political power.25 Jackman also maintains that support is independent of regime type. The support dimension of state capacity then refers to the degree of capacity rather than the regime type. A classic quote from Huntington (1968, 1) is consistent with this point: “The most important political distinction among countries concerns not their form of government but their degree of government.” Kugler and Domke (1986, 44) similarly argue that, “politically capable nations need not be free, democratic, participatory, or endowed with any of the characteristics attributed to modern governments in the normative sense.” Support can be generated in a number of non-forceful ways that are not inherently conditional on regime type.26 Levi (1988) ﬁnds that societal support, or what she terms quasi-voluntary compliance (QVC), is necessary as it reduces the transaction costs of generating revenues, thus increasing state capacity. It is termed voluntary as tax-payers choose to pay their taxes, but it is actually quasi-voluntary as there are potential punishments for violators (Levi, 1988). This type of compliance is predicated on the individual citizen’s belief that the leaders of the state can credibly commit to deliver promised goods in exchange for the tax revenues. In addition, these citizens must feel their neighbors are also contributing

23In Chapters 3 and 7, I also look at how changing this distribution aﬀects support and dissident responses to state repression. 24Arendt (1970, 56) takes this a step further by asserting that “it is insuﬃcient to say that power and violence are not the same. Power and violence are opposites; where the one rules absolutely, the other is absent.” 25Unlike domination or totalitarian control, support the way Jackman and I conceive of it, requires some persuasion by the state to generate compliance. As Jackman notes, this is similar to the Gramscian notion of hegemony: the state is able to constrain the set of feasible policy alternatives. 26The threat of force underlies many of the goals of any state. Tax-evaders, speeders, and law-breakers are always threatened with force. The more the state must use force, however, the less it is able to generate compliance through political means. In other words, the more the state must place citizens in jail, punish them, and use repression, the more costs the state is incurring to implement its policies.

16 to the provision of public goods and not shirking.27 Citizens prefer not to attain the ‘sucker’s payoff’ which results when they contribute to tax revenues while their neighbors do not. QVC in this conception is a form of support or a way to induce compliance through political means rather than through force. Given these two conditions, states must provide citizens with some return for their revenue. The simple contract between citizens and states provides states with revenue in return for personal security. At the most basic level citizen security relates to personal integrity rights and property rights. How can the state credibly constrain itself from violating these rights in pursuit of revenue? North and Weingast (1989) claim that the solution is to create institutions that can constrain the state and provide remedies for the aggrieved members of the population.28 The critical mechanism for linking citizens and support for institutions is their belief that state institutions are reliably constrained and can solve the commitment problem (Denzau and North, 1994). Jackman equates the development of legitimate institutions with the development of state capacity. Figure 2.4 offers a graphical representation of the two attributes of state capacity identified above. Figure 2.4 depicts the continuous nature of each concept and shows how their intersection creates stronger or weaker states. The space created by the intersection of the two dimensions is state capacity. The further from the origin, or where resources and societal support are equal to zero, the more strength the state possesses. In Figure 2.4, the weakest state is in the bottom left quadrant and lacks both resources and societal support. This state lacks resources and uses repression in an attempt to induce compliance. As we move from left to right, resources available to the state increase, but societal support remains constant. A state that has important natural resources or high amounts of foreign assistance may be in this area. For the states in this region, implementing policy is difficult as using repression leads to greater costs. In the upper left quadrant are states that engender societal support but lack resources. These states are resource poor, but do not face the high costs generated by passive or negative support. States that maximize both support and resources are found in the upper right corner. These strong states have

27In the language of game theory, this type of social situation is often called an assurance game where collective action is dependent upon other individuals reliably participating in the act. See (Chong, 1991) for a detailed discussion of collective action and assurance games. 28Following North (1990), I conceive of institutions as the formal and informal rules that govern human interaction.

Figure 2.4: The Dimensions of State Capacity

the resources to implement policies and face fewer costs in implementing these policies. Since the dimensions of state capacity are both continuous, the area that corresponds to the diﬀerent state labels is not clearly demarcated. The corners of the state capacity space are more clear, but the space in between is what Goertz (2006) calls the gray zone. In fuzzy logic,29 this gray zone represents the area where membership is classiﬁed as more or less rather than as complete. States that are in the gray zone have some score or likelihood of being in the category of interest, in this case, being strong or weak. As Goertz (2006, 27) notes, “the core attributes of a concept constitute a theory of the ontology of the phenomenon under consideration.” The core attributes of a concept identify both what constitutes the concept, or its ontology, and identify causal mechanisms that link the concept to other concepts. In this case, my contention is that resources and societal support constitute state capacity. Resources and support are then the mechanisms that explain the relationship between state capacity and other concepts like civil war.30

29See Ragin (2000) for a detailed discussion of fuzzy logic. 30Deriving the core attributes of a concept is not simply an arbitrary choice. As Goertz (2006) identiﬁes, there are many ways to conceive of any empirical referent, but some attributes are useful for generating testable hypotheses while others are not. We could categorize animals in the world according to size, color, or some other attribute. Using the schema we have, helps explain animal behavior or why humans and other mammals can survive in colder climates while reptiles must seek warm conditions. Using size or color as our core attributes of what makes an animal can not generate testable hypotheses related to warm and cold weather behavior. From a Lakatosian (1977) standpoint, the conceptual framework that can generate

18 Adding another dimension or attribute to the concept of state capacity addresses the problem of what Munck and Verkuilen (2002) term a minimalist definition. They address this problem in relation to minimal definitions of democracy. These concepts are minimal as they include few attributes. For democracy, this single attribute is usually competitive elections, and other attributes, like citizen participation or rule of law, are often overlooked. Ultimately, the goal is to avoid excluding attributes that are theoretically relevant.31 Most scholars in the study of civil war apply a minimalist definition of state capacity that I argue excludes important information.32 Three-Level Concepts and State Capacity

Figure 2.5 offers a three-level explanation for the concept of state capacity consisting of the basic level, the secondary level, and the indicator level. According to Goertz (2006, 53), the reason for having three levels in explicating concepts is that “usually the basic level is too abstract and complex to be directly converted into the indicator/data level.” The purpose of the secondary level, or identifying the dimensions of the concept, is to provide “theoretical linkages between the abstract basic level and the concrete indicator/data level” (53). The arrows represent that each indicator is a potential substitute for the secondary level dimension.33 At what Goertz (2006) calls the basic level is the concept.34 Below the basic level, is the secondary level which includes the dimensions of the concept. In the Figure 2.5, resources and societal support are at the secondary level and constitute the concept of state capacity. The AND operation between the two dimensions of state capacity signify that they are both necessary for constituting a capable state. In addition, the two are jointly sufficient.35 empirically verified hypotheses that subsumes other schemas and provides at least one more is preferable to other approaches. 31On the other hand, adding too many attributes leads to a maximalist definition and risks including attributes that are not theoretically relevant. 32This is not the case for the theorists of the state such as (Kugler and Domke, 1986; Nordlinger, 1987; Levi, 1988; Migdal, 1988; Jaggers, 1992; Jackman, 1993; Goodwin, 2001). 33The arrows do not represent causality. In principal component analysis, arrows linking indicators to concepts indicate that they cause or reveal the concept. In factor analysis or latent variable models, the arrows run the other way from unmeasured variable (the dimension of the concept) to the indicators. This three level framework is more consistent with latent variable models but does not use the same conventions. 34Munck and Verkuilen (2002) just call this level the concept. They also have a three level framework that includes concepts, attributes of concepts, and components of attributes. 35This construction also assumes that resources and societal support are of equal weight. To the extent

Figure 2.5: Three-Level Concept of State Capacity

The third level is the indicator level. As Goertz (2006, 62) notes “the indicator level links the more theoretical analysis in the basic and secondary levels to the more practical requirements of converting these ideas into empirical practice.” The indicators are proxies for each dimension of the concept. At the indicator level, the analyst has to be concerned with using valid measures.36 Additionally, at the indicator level, several measures of each dimension of state capacity are potentially substitutable for each other. The OR operator the one dimension is more important than the other, some weighting scheme can be employed to increase the importance of the particular dimension. This also assumes that each dimension of capacity is not a perfect substitute for the other. Substitutability can also be thought of as a continuum where some mix of the two dimensions is necessary and suﬃcient to constitute the concept. At the negative end of this continuum, no amount of the resource dimension can compensate for lack of support and vice versa. At the positive end, these dimensions are perfectly substitutable. While I would not argue substitutability for resources and support at either extreme, their substitutability shades toward the negative pole. For example, a state can not achieve the highest levels of strength by only maximizing a single dimension. 36Carmines and Zeller (1979, 12) deﬁne validity as whether “an indicator of some abstract concept...measures what it purports to measure.”

20 signiﬁes that a state with resources can acquire this dimension through tax revenue, foreign assistance, or through primary resource extraction. For example, a state with low national income can increase this dimension by seeking foreign aid.37 A state with adequate resources can have one or more of these indicator level characteristics.38 Societal support can be measured in a variety of ways as well potentially including: the degree of protection of private property, the ability of the state to extract revenue from the population, the amount of state violence against the citizens, or amount of citizen violence against the state. These indicators are again substitutes for the societal support dimension.39

2.4 Conclusion

Using a three level approach to concepts helps construct an explicit schema for identifying relevant dimensions of a concept and and the indicators used to measure these dimensions. As data that purports to measure concepts like democracy abounds, the need to be explicit about the components of the concept increases (Munck and Verkuilen, 2002; Goertz, 2006).40 Sartori (1970, 1033) claimed in 1970 that political science was growing more technically sophisticated but that a “large majority of political scientists qualify as pure and simple unconscious thinkers.” Unconscious thinkers, for Sartori (1970, 1033), had little concern “with the logical structure and procedure of scientific enquiry.” Modern quantitative political science is assuredly more concerned with logic and science, but remains somewhat vague when dealing with concepts. At the core of modern comparative politics and international relations are concepts like legitimacy, state capacity, democracy, and conflict that are contested and often vaguely used in verbal and empirical work. While it may be difficult to ever get the majority of political scientists to agree on what concepts like democracy are, having scholars be clear about concepts and how they relate to other concepts improves both quantitative and qualitative work alike. In the next chapter, I develop a theory of state-dissident interaction that explains the variance in resource-poor states experiencing civil war. Both dimensions of capacity,

37In Chapter 8, I investigate how states increase their capacity by appealing to external actors and the impact that this has on the likelihood of civil war. 38Goertz (2006) argues that substitutability should be most common at the indicator level and that is a common method to involve unique characteristics, such as culture, into a causal explanation. 39I discuss the choice of indicators in greater detail in Chapter 4. 40Munck and Verkuilen (2002) also advocate thinking about aggregation, or how to combine the indicators that represent these dimensions.

21 resources and societal support are integral components of the theory, and help to explain the process of violence that leads to civil war.

Appendix

In Table 2.1, I list several prominent scholarly works that address the concept of state capacity. Next to the scholar’s work, I provide a checkmark if their conceptualization of state capacity includes this particular dimension. The dimensions include: state resources, legitimacy, institutional capability, job insecurity, and autonomy. This sample is not meant to be exhaustive of the ways that state capacity can be conceived. Instead, the table is illustrative of some of the commonalities as well as divergence in diﬀerent conceptualizations of state capacity.

22 Table 2.1: Attributes of State Capacity

Author Resources Legitimacy Institutional Cap. Job Insecurity Autonomy Geddes (1995) X X X 23 Goodwin (2001) X X X X Gurr (1986) X X X X Jackman (1993) X X X Levi (1988) X X X X Migdal (1988) X X X X Nordlinger (1987) X X CHAPTER 3

A PROCESS THEORY OF CIVIL WAR ONSET

3.1 Introduction

Recent scholarship has established that the condition of being a weak state raises the probability of civil war (Fearon and Laitin, 2003).1 Weakness is generally measured as lack of resources or low GDP. UN statistics show that of the 50 countries in the world with the lowest GDP, nearly 60% experienced civil strife of varying intensity and duration in the 1990’s. While this is a high percentage, the experience of almost 40% of these resource-poor states remains unexplained.2 These statistics, coupled with the academic ﬁndings, are the foundation for an interesting puzzle: Why do some weak states experience civil wars whereas many others do not? Consider the experiences of El Salvador and Bhutan. Each state is commonly considered weak, each has diﬃcult terrain, neither was a democracy until recently, yet one has seen a bloody, protracted civil war while the other is relatively tranquil.3 El Salvador is a weak state with a per capita national income that has hovered around $2,000 since the 1970s. In the mid 1970s, the military leaders of El Salvador responded to increased political mobilization by dissidents with harsh repression. In 1975, for example, 15 protesters were killed while demonstrating against the Miss Universe Pageant. The interaction between the state and dissidents began in the early 1970s as dissident groups openly criticized and organized against the repressive military rule (Mason and Krane, 1989, 187-188), and by 1979 a state of open

1The deﬁnition of a weak state is one that lacks the capability to implement its preferred policies. This deﬁnition is discussed and expanded in later sections. 2The temporal variance is even larger as the probability of onset for a particular state is almost always closer to zero than to one. 3Fearon and Laitin (2003) use the variables mountainous terrain, GDP, and a large change in regime score, among other variables, to proxy the concept of weak state. These variables all make an insurgency against a state more likely/viable.

24 civil war existed. The civil war lasted over a decade claiming more than 50,000 lives (Lacina and Gleditsch, 2005). Like El Salvador, Bhutan is considered a weak state. It is a landlocked country with a per capita GDP that fluctuated between $400 and $1,000 US dollars between 1988 and 2005.4 Since its inception in 1971, the King and other leaders of the state have exercised uncontested control throughout the territory. Even though it has an even lower per capita GDP and also has rough mountainous terrain, Bhutan has avoided the civil war that El Salvador has experienced.5 These are not isolated cases. For example, civil war has occurred in Rwanda, Cambodia, Angola, Afghanistan, Sierra Leone, and many other countries. What do these countries all have in common? All of the above states are commonly termed weak states. Yet, other equally weak states such as Bhutan, Cameroon, Ecuador and Burkina Faso have avoided civil wars. Why? Although El Salvador and Bhutan are matched on some important civil war-enhancing characteristics, like population and difficult terrain, an important difference between the two is the Salvadoran military’s choice to repress the population. I argue that states that lack resources and societal support are at the most risk for onset of civil war. State leaders make choices, such as repression, which reduce support from society and increase the likelihood of active dissent. Similar to previous work, I argue that states with low capacity are at the most risk of civil war. Contrary to this work, however, I conceptualize state capacity as involving more than just resources; it also involves societal support. Extant work has focused primarily on state weakness and failed to incorporate strategic interaction between the state and dissidents. In contrast, I offer a model that explicitly identifies the process of violence between states and dissidents that links weak states to civil war. This model specifies the microfoundational motives of the central actors, the state and the dissidents. State repression both leads to rising violence levels and pushes civilians into active dissent. Once dissidents are mobilized and challenge the state, civil war becomes more likely. The answer to the puzzle then is: High levels of repression occur in some resource-poor states, and it is these states that are most likely to experience civil war. In sum, I argue that the likelihood of a state experiencing civil war is a function of state–

4GDP estimates come from the Penn World Tables. 5Both state also lack lootable resources such as diamonds or oil.

25 dissident interaction; this is an important claim that, though non-controversial on its face, has been ignored in the literature.

3.2 Why Civil War?

The prevailing wisdom suggests that civil war is most likely when the conditions that favor insurgency are present in a particular state (Fearon and Laitin, 2003). These conditions affect the relative strength of the insurgency and include such characteristics as a financially weak state, rough terrain, large populations, and the newness of a state.6 In short, Fearon and Laitin (2003) conclude that structural conditions like mountains and large populations increase the strength of an insurgency relative to the state, which increases the probability of civil war. Financial weakness, as proxied by GDP is critical; it affects the state’s ability to perform counterinsurgency, use discriminate violence, and provide local policing. For Fearon and Laitin, financially weak states, or resource-poor states, are the most prone to civil war because they are unable to perform the tasks that reduce the effectiveness of potential insurgent adversaries. Fearon and Laitin conceive of state weakness as unidimensional: the resources available to the state determine its capacity. In sum, Fearon and Laitin argue that conditions that favor insurgency increase the relative strength of an insurgency, which raises the probability of civil war. To test the hypotheses implied by the theory, they focus on the conditions favoring insurgency and treat the relative strength of the insurgency as a latent unmeasured variable. While they find a strong statistical relationship between low GDP and the onset of civil war, Fearon and Laitin’s theory struggles to elucidate how low GDP, a fairly time-invariant measure, can explain civil war onset occurring in country j in year i and not some other year i +1 or i − 1. Their theory also fails to explain why civil war occurs in some weak states and not others. While some financially weak states descend into civil war, many others avoid it. Resources alone do not determine a state’s ability to implement its preferred policies. Fearon and Laitin provide a structural model of factors that correlate with the onset of civil war. I improve on their work by advancing a theory that explains how some states and dissidents interact to produce violence beyond a given threshold. Although they define civil

6Fearon and Laitin include other items that likely aﬀect the balance between state and insurgents, such as foreign support, contraband, and foreign sanctuary, but the above conditions are most central to the their theoretical story.

26 war in a similar way, they do not present a model which offers an explanation for why states and dissidents kill people beyond a given threshold.7 In the next section, I offer a theory which provides a link between the choices of individuals composing the state and dissident groups and the onset of civil war. I also identify another dimension to state capacity–societal support. In the theory offered below, I show how this concept affects the choices made by dissidents and their propensity to produce violence.

3.3 Theory

Most scholars conceive of civil war as violent interaction between states and dissidents that produces deaths exceeding some threshold (Small and Singer, 1982; Singer and Small, 1994; Collier and Hoeffler, 2001).8 Below I offer a verbal and formal definition of civil war that is consistent with previous conceptions but highlights the need for investigating state-dissident interactions (Moore, 1995; Tilly, 1995; Shellman, 2006).

Let: ‘CW’ represent civil war and ‘Deaths from PV’ or ‘DPV’ represent deaths from political violence

DPVs Deﬁne: CW ≡ DeathsfromPV > τ| DPVs+DPVd ≥ .10, where τ is a cumulative death threshold, usually 1000 battle deaths, DPVs represents the number of state agents among 9 the dead, and DPVd represents the number of dissidents among the dead. Often civilians are the largest portion of DPVd when state repression misses the intended target. State 10 repression thus increases DPVd while dissident violence increases DPVs.

7According to Fearon and Laitin (2003, 76), civil war involves the following criteria: “(1) fighting between agents...of a state and organized, nonstate groups who sought either to take control of a government, to take power in a region, or to use violence to change government policies. (2) The conflict killed at least 1000 over its course with a yearly average of at least 100. (3) At least 100 were killed on both sides (including civilians attacked by rebels).” 8If the violence were solely by states or dissidents, terms such as mass killing, politicide, or terrorism might be used. Collier and Hoeffler (2001) require that at least 5% of the killings have to be attributable to each side. Sambanis (2004) offers some other criteria for creating an operational definition such as: the dissidents must be based within the same territory as the state, the state must be internationally recognized, and the state must have a minimal population. 9Including the conditional ratio of state to dissident dead ensures that genocides, politicides or other cases of mass killing of unarmed civilians are not counted as civil wars. In this project the conditional ratio is implied throughout but left out for notational convenience. 10 Dissident violence can also increase DP Vs when dissidents target civilians as was the case in the Algerian civil war as well as in Peru during the Shining Path insurgency. Civilians are often the largest portion of the dead as control over civilian populations is central to the struggle between states and dissidents (Kalyvas,

27 Deﬁne: DeathsfromPV = f(DissActivity, Repression).11

Expressed verbally, the probability of the onset of civil war is a function of state–dissident interaction. This seems intuitive, but it has been overlooked in this literature. Much of the civil war literature attempts to ﬁnd correlates that increase the likelihood of passing a given threshold while ignoring that civil war is a process generated by state and dissident violence. Developing a theory relating state capacity, or the ability of the state to implement its policies, to civil war must explain how dissident, civilian and state agent deaths are generated and pushed beyond this threshold. In the next section, I develop some assumptions about state actors and dissidents and in the process identify the micromotivations that push these actors in certain ways that are more likely to generate large numbers of deaths and, thus, result in civil war.

3.3.1 Modeling the State

In developing a process model of civil war onset, I identify three central actors: the state, citizens and dissidents. In this section, I build on the work of Levi (1988) and others to outline how states pursue their goals. In the next section, I outline the preferences of citizens and dissidents and discuss how interaction with the state affects their choices. The first task is to understand the motivations of individuals composing the state. Levi (1988)’s model of how state actors maximize revenue given constraints helps explain variance in revenue production, an important component of state capacity. Building on Levi (1988), I make two assumptions about state actor preferences to derive a model of state actor behavior. I modify her second assumption and explain how this affects the hypotheses derived from the assumptions. First, I assume that the state and the polity are collections of rational individuals and that the preferences of the state are not necessarily the same as the preferences of individuals in the polity. Second, I assume that state actors prefer to maximize revenues given constraints and that the most salient constraint is survival.12 These assumptions are consistent with what others term the predatory state model (Geddes, 1995;

2006). Sympathetic populations to either side are often counted among the dead for that particular side. 11As noted above, the conditional ratio of deaths is suppressed. The full expression is: DPVs (DeathsfromP V | DPVs+DPVd ≥ .10) = f(DissActivity, Repression). 12This diverges from Levi as she does not make a claim about leader survival being more important than the other constraints on revenue maximization. Bueno de Mesquita et al. (1999) make a similar assumption about the leaders of a state. They claim that a leader’s main goal is to survive or to stay in oﬃce.

28 Levi, 1988; Mann, 1993; Migdal, 1988).13 One important implication of these assumptions is that the leader’s job insecurity, or the leader’s expectations about maintaining office in the future, affects the state’s decisions. Because state leaders are rational, they prefer to use less costly means to stay in office. If survival is at stake, however, leaders will discount the value of retaining power in the future and pursue strategies that maintain the leader’s present position but are potentially more costly for the leader in the future.14 One such strategy is to repress citizens. States use repression both to respond to behavioral challenges from dissident groups and as a tool to generate compliance with its policies (Davenport, 2007).

Figure 3.1: Hypothesized Relationship between Job Insecurity and Repression and Repres- sion and Societal Support

Repression, however, is costly to the state as it aﬀects support for the state and its policies. I conceive of support for the state as a continuous dimension with the poles ranging from low to high. Each citizen has some value for supporting the state and repression

13Marxists and pluralists have a different conception of the state and of the preferences of the actors that compose the state. For Marxists, the state is merely a tool of the capitalist class with no independent preferences. Pluralists see the state as an arena for contention among societal groups. Pluralists differ from the Marxists in that they do not expect one class or group to dominate government. They are similar in that they expect the ruling class preferences to be synonymous with state preferences. 14Both Levi (1988) and Cheibub (1998) use this concept in their theories of state revenue accumulation. Cheibub (1998) operationalizes the concept by predicting hazard rates for losing office dependent on time, past cumulative leader changes, and economic growth. Both refer to this concept as the discount rate. To avoid confusion with how game theorists use the term, I instead refer to job insecurity.

29 lowers that value by imposing costs on citizens in an attempt to generate compliance. Each additional unit of repression moves more members of society away from the state’s preferences. Repression thus decreases societal support and in turn increases the costs of implementing a state’s preferred policy (see Figure 3.1).15 The above assumptions and discussion imply the following hypotheses:

Hypothesis 1 The greater a state leader’s job insecurity, the more repression used by the state.

Hypothesis 2 Increasing repression reduces societal support.

Repression is used as a response to the actions of the dissidents. Davenport (2007) claims that the state always responds to behavioral threats (dissident activity) with coercion; he refers to this as the Law of Coercive Responsiveness. According to Davenport (2007, 7), “The consistency of this ﬁnding is quite astonishing in a discipline where very few relationships withstand close scrutiny.” Consistent with Davenport’s claim, this model expects that behavioral challenges to the state always elicit a response because survival for state leaders is imperative.

Hypothesis 3 States respond to increasing dissident activity with increasing repression.

Davenport’s claim about the violent response of states to dissident actions receives a good deal of empirical support. Another area of the repression literature that has a solid empirical foundation relates to the relationship between development (or state resources) and repression. Similar to previous research on the effect of state resources on repression (e. g. (Mitchell and McCormick, 1988)), Poe and Tate (1994) find that states with high resources tend to repress less. They hypothesize that states who have high resources create a political climate where repression is unnecessary. Poe, Tate and Keith (1999) expand the temporal domain of their previous study; and while some other relationships change, they find continued support for the resource-repression relationship. Consistent with these findings, I expect that states with high resources are less likely to repress as they face fewer potential challenges to their policy goals.

15Repression is also costly for states as they have to maintain an apparatus capable of repressing. This state agency requires some portion of the budget that could be spent in other areas.

30 Hypothesis 4 The more resources a state possesses, the less repression the state uses against its citizens.

This discussion and hypotheses outline the motivations of the state and explain why some choose to repress. Insecure leaders use this tool to maintain power even though it makes governing in the future more costly. Additionally, state leaders respond to threats to their political survival with some form of violence. States with high resources tend to face fewer of these violent challenges. The next task is to model dissidents and how they respond to state actors.

3.3.2 Modeling the Dissidents

Because civil war is an outcome of violent interaction between states and dissidents, the other important actor to model is the group of dissidents that oppose the state. Since dissidents come from the pool of civilians, I identify three actors in a given polity: the state, the dissident, and the civilian. Civilians are the population within a given territory controlled by the state. When they engage in dissident activity, civilians are dissidents.16 Dissidents produce dissident activity such as rioting, protests, terrorism, guerrilla warfare, and any other action that directly opposes the state and its policies. The state refers to the leaders who direct and decide policy for the polity. As stated earlier, I assume members of the polity are rational. In addition, I assume that a distribution of support for the state exists such that civilians can have varying levels of support for their leaders. The shape of this distribution can be normal. In other words, a few people have preferences highly consistent with the state, a few are radically opposed, and the majority have preferences that are in between these poles. Changing the shape of the distribution aﬀects the aggregate support for state leaders. When most people support the state, the distribution is skewed left. When few people support their leaders, it is skewed right. Some important factors that aﬀect this distribution include the institutions within the state and ethnic fractionalization. Democratic institutions may increase overall support for the state as state leaders attempt to make policies consistent with the median voter’s preferences. While even in a democracy

16I use the terms civilian, individual, person, and member of the population interchangeably. I reserve the term dissident for when an individual acts against the state.

31 some people’s preferences are extremely far from the median voter and thus have low levels of support for these policies, the frequency of these disaffected people should be lower in democracies than in authoritarian regimes. As Bueno de Mesquita et al. (1999, 2003) argue, states with democratic institutions should produce policies that are beneficial to a larger number of citizens. These institutions produce better policies, from the perspective of increasing general welfare, as leaders’ political survival depends on maintaining a large winning coalition, defined as the minimum number of people sufficient to maintain power.17 Since more people are receiving benefits from the state, support for democratic leaders is, on average, likely higher than support for authoritarian leaders.18 The ethnic composition of society may also affect the support for a given leader. Since ethnic groups may have a different preference for policies than the state (especially if it is a rival ethnic group), the more groups that exist, the less likely it is that citizens will have preferences closely matched to the state. As stated previously, the further citizen’s preferences, influenced by ethnicity, are from the state, the more skewed the distribution will be towards low support.19 Most importantly, changes in the mean of this distribution affect the number of civilians willing to become dissidents and engage in anti-state behavior (see Figure 3.2). As each individual’s support decreases, the likelihood that more of these individuals become dissidents increases. The mean value of individual support for the state leaders and their policies within a polity is then referred to as societal support. Societal support is conceived as the costs of policy implementation imposed on the state

17This subset, the winning coalition, comes from what Bueno de Mesquita et al. (1999) call the selectorate, or the total number of people in society who have some say in selecting leaders. 18In Bueno de Mesquita et al. (1999) terms, institutions that have a large W/S ratio, should produce more public goods and thus benefit a larger portion of society. 19Recent opinion polls from Iraq provide some corroboration for this approach. Support for the nascent state is influenced by which ethnic group the respondent belongs. According to a Program on International Policy Attitudes poll taken January 21, 2006, in general the Shia and Kurds do not support attacks against the state (about 3% of Shia and 1% of Kurds). The Sunni, however, have a lower average level of support for the state with over 24% supportive of attacks against the central government. Support for attacks against the US–led coalition was much higher across the ethnic groups and followed similar differences in levels of support with 88% of Sunnis supportive, 41% of Shia, and only 16% of Kurds. More troubling is that when asked if the newly-elected parliament will be a legitimate representative of the Iraqi people, only 6% of Sunni replied ‘yes’ compared to 81% of Kurds and 90% of Shia.

32 Figure 3.2: The Distribution of Societal Support

by society.20 Support as discussed above can be arrayed along a spectrum (see Figure 3.3).21 The poles of this spectrum are positive and negative support. Societal actors can positively support a state policy through what Levi calls quasi-voluntary compliance (QVC) or what I term positive support. Individuals can also provide passive support or support that requires actual or threatened coercion or sanctions. Finally, individuals in the polity can provide negative support or mobilize against state policies.22 The distance between the preferences of a citizen and the state determines whether they positively, passively, or negatively support the regime. As Figure 3.3 demonstrates, this distinction among types of support is important in explaining the costs to state leaders of implementing policies. In short, positive support costs less than passive support induced by actual or threatened force. Support for the state is lower after the state represses the population, and the distribution of support shifts (from the grey distribution to the black in Figure 3.2). Additionally, more people move further from the state’s preferences thus increasing their likelihood of producing dissident acts. Therefore, as societal support decreases, the number of dissidents increase (see Figure 3.4).23

20This deﬁnition is similar to Nordlinger (1987). 21This Figure was ﬁrst introduced in Chapter 2. 22While I conceive of individual actors positively, passively, or negatively supporting a policy, I am more interested in the distribution of these types in a polity. A polity with a distribution with a higher mean value for support is conceived of as providing high levels of societal support and conversely a distribution with passive or negative support has a lower mean value of societal support. 23Figure 3.4 as well as the other Figures that display relationships between concepts are meant to convey

33 Figure 3.3: Spectrum of Societal Support and Policy Implementation Costs

In short, as the number of people mobilizing against the state and its policies increases, the costs of implementing policies increase. In addition, as support shifts away from the state, more civilians are likely to become dissidents. This leads to the following hypothesis:

Hypothesis 5 A decline in societal support leads to more dissident activity.

Citizens can be unsupportive of the state without taking action. The farther, however, a citizen’s preferences are from state policies, the more likely they are to engage in dissident acts. This reinforces the above hypothesis as increasing the number of civilians with preferences that diverge greatly from the state increases the likelihood of producing dissidents. Previous work on violence by civilians with extreme preferences generally terms civilians as moderates and dissidents as extremists (Kydd and Walter, 2002; Lake, 2002). The goal of the extremists is to invoke harsh responses from the state to push the preferences of the moderates closer to those who prefer violence. While the terms are diﬀerent from this model, the expectations are the same: members of society with extreme preferences use violence to pursue their goals. When states respond with violence, it pushes civilians (or moderates) towards the dissidents (or extremists).24 the direction of the relationship rather than an explicit statement of functional form. 24DeNardo (1985) makes a similar claim in regards to why dissidents use terror. According to DeNardo (1985, 235) “terrorism can sometimes stimulate mobilization by provoking untoward repression against innocent people.”

Figure 3.4: Hypothesized Relationship among Dissidents, Dissident Acts, and Societal Support

Another important factor which influences a civilian’s decision to become a dissident is the probability of success. Considering that most citizens prefer to be on the winning side of a violent conflict, the expected benefits of joining a dissident group are greater when success is likely while the expected costs are greater when success is unlikely. Figure 3.5 offers a graphical representation of this claim. As Figure 3.5 shows, when expected success rises, so does the number of dissident acts against the state. A dissident can approximate the probability of success by evaluating the resources available to the state for the repression of dissent. A state that has a strong coercive apparatus can repress dissent.25 A state that expends resources on its military is expected to reduce the likelihood of generating dissidents and dissident activity. From this discussion, I offer the following hypothesis:

Hypothesis 6 The greater the likelihood of dissident success, the more likely it is that a civilian produces dissident activity.

The above deﬁnitions, assumptions and hypotheses oﬀer a model for the interaction between states and dissidents that can explain why some states engage in the processes that lead

25Gurr (1988) argues that states that have demonstrated strong internal repressive capacities dissuade dissidents from mobilization.

35 Figure 3.5: Hypothesized Relationship between Resources and Expected Success of Dissident Activities and the Expected Success of Dissident Activities and the Number of Dissident Acts

to civil war while others do not. In sum, the more support a state receives, the fewer the number of dissidents. Therefore, the more support a state receives, the fewer civilians will be willing to commit violent acts. In addition, the greater the military capacity of the state, the less likely dissidents expect their activity to be successful. The fewer challenges there are to the state, the less likely the state is to repress. With less violence by the state and fewer dissident acts by the opposition, it becomes less likely that combined deaths will be pushed beyond the threshold indicating civil war. Although many scholars conceive of civil war as violent conﬂict between the state and at least one opposition group within the same territory that surpasses a minimal death threshold,26 this theory is novel as it oﬀers an explicit process whereby state-dissident interactions produce this outcome. Additionally, this proposed explanation for civil war accounts for why state weakness leads to civil war based on a multidimensional conception of state capacity. As the above hypotheses and discussion suggest, societal support is an important dimension to consider when attempting to explain the relationship between state capacity and civil war. Building a theory that models this dimension helps explain why some weak states are more susceptible to incubating civil wars than others and provides microfoundations for this phenomenon.

26See Sambanis (2004) for a discussion of deﬁnitions of civil war.

36 3.4 Path to Civil War Model

Figure 3.10 is a causal model of the relationship among major variables involved in the theory. The arrows offer direction of impact of one variable on another.27 The variables also have subscripts denoting the sequence of events ranging from t − 1 to t. Because the variables operate in sequence and do not just affect the dependent variable of interest, they can not be modeled using a single equation and require the use of multiequations (Berry, 1984). Using a single equation overlooks the mediating effects of dissident activity and repression. In other words, dissident activity and repression intervene between concepts such as job insecurity and societal support and the onset of civil war. In addition, some of the effects of structural conditions that Fearon and Laitin identify are also likely mediated by repression and dissident activity. As they argue, “The numerical weakness of the insurgents implies that, to survive, the rebels must be able to hide from government forces” (Fearon and Laitin, 2003, 80).28 Large populations and mountains allow dissidents/insurgents to hide and to survive. Unless dissidents/insurgents can survive, they are unable to produce violent activity. In terms of my argument, these factors increase their ability to produce dissent and thus have an effect on dissident activity. Dissident activity, in turn, increases the likelihood of producing violent interaction between states and dissidents that then exceeds the civil war threshold. On the state side, mountainous terrain and a large population decrease the state’s ability to use targeted effective counterinsurgency as it is more difficult for the state to find dissidents in these conditions. Again, repression is

27Several relationships that exist in the figure that are not developed in the theory include the effect of previous values of variables on their present values. Bureaucratic inertia and production of resources within a state make it highly likely that last year’s resources affect the level of this year’s resources. As Gurr (1986); Poe and Tate (1994) claim, a similar logic explains the effect of previous repression on present repression. Put otherwise, as states use repression, they develop institutions and budgets that use this tool and become more willing to use repression in future interactions with dissidents. Dissident Activity is also positively related to previous dissident activity as well as previous support helps explain present support for state leaders. Last year’s job insecurity affects this year’s insecurity and thus must be incorporated in the model. Finally, I use Cheibub (1998)’s framework for modeling a leader’s job insecurity or that this insecurity is a function of time, the previous number of leader changes, and economic growth. 28While Fearon and Laitin do not define what an insurgent is, they define insurgency as “...a technology of military conflict characterized by small, lightly armed bands practicing guerrilla warfare from rural base areas.” We can assume that an insurgent is one who practices this form of violence. Dissident is a more general term than the way Fearon and Laitin use insurgent. While they refer to mostly rural activists, many of the conflicts they include in their data involve a sizable urban component. Insurgents thus in their data are members of organized groups violently challenging the state. This conception is nearly identical to my use of dissident and in this project the two terms are synonymous.

37 mediating the relationship between the structural conditions and the onset of civil war. To express the argument explicitly, consider the following system of equations. While the equations for the probability of civil war and resources are fairly straightforward, the relationships among variables in the societal support, repression, and dissident activity blocks are more complicated. Figures 3.6–3.9 oﬀer graphical representations of the blocks and Figure 3.10 oﬀers a depiction of the entire system.

P r(DPV >τ)t = f([+]Represst, [+]DissActt) (3.1)

Represst = f([+]DissActt−1, [−]Resourcest, [+]JobInsecurityt, [+]Represst−1) (3.2)

DissActivt = f([−]Represst−1, [−]SocSupt, [−]MilitaryCapt, [+]DissActt−1) (3.3)

SocSupt = f([+]Institutionst, [−]Ethnicfract, [−]Represst−1, [+]SocSupt−1) (3.4)

JobInsecurityt = f([+]Leader∆t, T imet, [−]EconGrowt, [+]JobInsecurityt−1) (3.5)

Resourcest = f([+]Resourcest−1) (3.6)

While I have described the relationship among many of the concepts in the theory and oﬀered explicit expectations about the nature of this association, in the next chapter I focus on describing and estimating a subset of these equations. In addition, I discuss how to test the hypotheses derived from the process model of civil war. 3.5 Additional Implications

The theory that I oﬀer for the onset of civil war provides an explicit causal path for why some weak states develop civil war while others do not. Previous state capacity arguments tend

38 Figure 3.6: Repression Block

Figure 3.7: Dissident Activity Block

to think of this concept in simple terms and leave out important variations among states. States can be stronger or weaker than other states based upon both the resources available to the state and the amount of support the state receives from society. Some weak states can avoid civil war, not by simply strengthening their repressive apparatus but through engendering support from society. Increasing support should decrease dissident activity and thus the need for future repression. Similar to the arguments put forth here, Goodwin (2001) claims that states with weak

Figure 3.8: Societal Support Block

Figure 3.9: Job Insecurity Block

infrastructural power who repress their citizenry are at most risk for revolution.29 For Goodwin (2001, 30), “...violent...repression [by the state] of certain groups, tends to “push” these oppressed groups into revolutionary movements.” He claims that understanding the state and the political context created by the institutional conﬁguration of the state explains revolution. For example, Goodwin (2001, 143) asserts that “the institutional conﬁgurations and practices of Central American states best explain...the relative success

29Goodwin (2001, 9) deﬁnes revolution as “any and all instances in which a state or political regime is overthrown and thereby transformed by a popular movement in an irregular, extraconstitutional, and/or violent fashion.” He also maintains that regimes which are patrimonial are at most risk to lose in these revolutionary situations. His deﬁnition of revolution is somewhat analogous to a successful outcome for a rebel group engaged in a civil war.

40 Figure 3.10: Causal Model of Civil War

or failure of Central America’s revolutionary movements.” In contrast, my approach looks at the incentives of the leaders of the state and how leaders interacting with dissidents generate political violence. My project unpacks the microfoundations of the process while his is a static structural view of states with little interest in their strategic interaction with dissidents. While this chapter’s focus is civil war, the theory applies to other types of violence that occur between states as well as to violence between states and non-state actors. State

41 strength as determined by support and resources has an important eﬀect on the likelihood of a state repressing human rights, engendering terrorism, or making war with another state. Therefore, this theory is applicable and could be extended to provide insights to these literatures. Where states enjoy societal support, they are stronger than a resource-alone approach may predict.

42 CHAPTER 4

TWO-STAGE EMPIRICAL TESTS

4.1 Introduction

In the previous chapter, I outlined some specific causal mechanisms relating certain types of states and state leaders and the choices that they make to the onset of civil war and thereby identified a number of testable hypotheses drawn from a process model of civil war. Whereas previous explanations for civil war onset often make claims about the association between a variable of interest and civil war, I offer a specific process where leaders in weak states who feel their job is insecure use violence to maintain their office and spark conflict that spirals to civil war. In this chapter, I discuss how to test the hypotheses. In the research design section, I outline the temporal and spatial domain of the study and discuss potential threats to valid inference. After discussing operationalization, I explain the problem of missing data and describe a strategy to address it. Next, I compare estimation strategies and argue for a two-stage approach. Then, I present results and provide simulated quantities of interest to show the substantive effects of key variables. Finally, I discuss comparing these results with other models and how these results potentially influence the study of civil war. In the Appendix, I display other models and results including a comparison of the regression coefficients from the main model presented in this chapter to Fearon and Laitin (2003)’s results.

43 4.2 Research Design

4.2.1 Temporal and Spatial Domain

The temporal domain for the study is 1976–1999.1 Since my theory is not limited to this time period, I do not expect that extending the data to 1945 would drastically change my inferences.2 Civil war has been more common since the end of World War II (Fearon and Laitin, 2003), and changes in norms of sovereignty or international structure may explain this increase from the previous time period (Jackson, 1990; Ayoob, 1995). In addition, concerns about the impact of the end of Cold War are mitigated by the sample including time before and after this historic change. The spatial domain includes a sample of 162 states in the international system.3 This sample includes most countries in the world but excludes many of the micro-states and small island states. Excluding these states likely has little impact on the inferences that I am able to draw on key variables. They are likely, however, to influence some of the controls as these states are unlikely to pass the violence threshold necessary to produce civil war. This means I am excluding observations for variables that correlate with the dependent variable and have low values on population and tend to be new states. The coefficient for population may be underestimated as inclusion of these microstates would increase the slope of the relationship between population and the likelihood of civil war.4 Finally, the coefficient for a new state’s effect on the likelihood of civil war may be overestimated as these micro-states are generally new to the system but lack civil war. Adding more observations to the data that have no civil war but are new states should reduce the slope of this relationship.

1Measures of state repression, a key variable in the analysis, only begin in 1976. Fearon and Laitin’s data, which include a large portion of the control variables that I use, end in 1999. 2I am less certain that civil wars occurring before World War II are applicable to this theory. While previous civil wars involved states and dissidents interacting, norms of sovereignty, the prevalence of political democracy, and the inability of states to use scorched-earth policies in pursuing counterinsurgency are all significant changes in this modern era. 3I have 162 countries in the data versus 161 for Fearon and Laitin. This reflects a change in the Correlates of War coding of Germany as three distinct countries: East and West before 1990 and a unified Germany from 1990 to the present. 4King, Keohane and Verba (1994, 131) demonstrate this potential relationship by showing how excluding low values for an independent variable that also has a low value on the dependent variable can under-predict the slope of the relationship between the variables.

44 4.2.2 Operationalizing Concepts Civil War

As discussed in Chapter 3, I conceive of civil war as the violent interaction between a state and dissident group(s) that exceeds a given threshold. An operational definition also requires some other stipulations to sort civil war from other forms of political violence. The standard operational definitions require that both the state or a group representing the state fight with a group or groups from society within the defined territory of the state. This struggle must exceed a certain death threshold, usually 1,000, in a defined time period, and applies to states beyond some minimal size, usually 500,000 people. To make sure that the inferences are not dependent upon the coding decisions of a particular measure of civil war onset, Sambanis (2002) advises using measures from different sources. I use several different indicators of onset to make sure that relationships are not sensitive to coding issues. Fearon and Laitin (2003) have one of the more widely used codings of onset. Gleditsch et al. (2002) have alternative measures that code civil war onset, intermediate armed conflicts and minor armed conflicts. Using their data on civil war as well as Sambanis (2004)’s data should make the results less sensitive to coding criteria. In the estimation tables, I display results for Fearon and Laitin’s onset variable (Onset).5 I present models with Sambanis (2004) and Gleditsch et al. (2002)’s data in the appendix.

Repression

Repression occurs when states use violence against citizens to induce compliance with policies. Some have called this state terrorism (Gurr, 1986; Poe and Tate, 1994), but in this study repression refers to these acts that violate the personal integrity of citizens within the polity. Torture, murder, disappearance, and political imprisonment are all examples.6 Two separate indicators are used to measure this concept. Gibney and Dalton (1996) oﬀer a Political Terror Scale (PTS), ranging from 1 to 5, which measures the level of repression in a given society. Low levels correspond to states where people are not imprisoned for their political views, torture is infrequent or nonexistent, and

5Results from models using the other two measures were very similar. 6Broader deﬁnitions of human rights incorporate facets beyond simple personal integrity rights such as economic or social rights. The theory as derived does not necessarily deal with violations of these rights. Death, disappearance, and imprisonment clearly take a member of the dissident group and remove them from the set thus reducing their ability to create acts against the state.

45 state murder of civilians is rare. High levels correspond to periods where states use these techniques frequently against a large portion of society. This scale is coded separately using two sources. One is coded from State department country reports (PTS S) while the other is coded from Amnesty International reports (PTS A). I estimate models using both measures independently but only display results for the PTS S (Repress). The measures correlate positively, highly, and provide extremely similar results. Cingranelli and Richards (1999) provide a diﬀerent measure of repression of human rights that is comparable to the PTS. It is highly correlated with PTS (.78) and measures whether states violate personal integrity rights as described above. Their Physical Integrity Rights index (PHYSINT) has four components: torture, imprisonment, disappearance, and extrajudicial killing. The Cingranelli and Richards (1999) index also taps into the ordering and frequency of repression; they demonstrate that governments are more likely to torture and to imprison than to murder and to disappear. Since both PTS and PHYSINT use state department country reports and Amnesty International Reports to code their data, it is not surprising that they correlate highly. I use the PTS S in the main models, but the results for PHYSINT are similar.

Dissident Activity

Dissident activity refers to acts that challenge the state outside of formal institutions. Examples include protests, riots, terrorism, and guerrilla tactics. These data are diﬃcult to attain cross-nationally and over time. Fortunately, Banks (1996) codes data on dissident activity that span from the early 1900’s to the present. I created a composite index of dissident activity (DissAct) using incidents of government crisis, or threats to a government’s survival, assasinations, or politically motivated murder of a state actor, and guerilla warfare, or armed attacks against the state. I then logged this score assuming that diﬀerences in dissident activity had a larger impact on the likelihood of civil war at low levels than at high levels.7 I purposefully chose only the violent acts by dissidents as these lead to state deaths while nonviolent protest and large strikes can not.8

7I also used the raw dissident activity measure and had very similar results. 8An alternative specification uses these measures in the index as well without significant differences.

46 Societal Support

The presidential (or other leader’s) approval rating within a country is a common measure of support for a state leader. While this information is available for the US and more developed democracies, comparable cross-national measures are not available, especially over time. To measure a state’s level of societal support, I employ two strategies. The first approach I use to create a proxy for societal support (Support) is to use Kugler and Domke (1986)’s indicator referred to as relative political capacity (RPC). RPC is constructed using three steps. First, one regresses the tax ratio (i.e. government tax revenues 9 Mining Agriculture Exports divided by GDP), on time, GDP , − GDP , and GDP . Second, one calculates the predicted values for the tax revenue, which are assumed to reflect a government’s potential Actual Government Revenue revenue. Third, one takes the ratio Potential Government Revenue to calculate RPC. When this measure is greater than 1, a state collects more tax revenues than expected, whereas when this measure is less than 1, the state is inefficient in extracting tax revenues. RPC has two attributes: the ability of the state to collect resources and the degree to which the state leaders can control society. RPC is a better proxy for support than for resources as difficulties in collecting resources are not necessarily related to a state possessing them. I argue that where states lack societal support they struggle to generate compliance with tax policy. Lemke (2007), among others, has claimed that this measure is a valid indicator of developing country capacity, but can not adequately proxy for support or capacity in developed countries as policy preferences determine the variance in revenue accumulation among developed states.10 Since most civil war onsets occur in these developing countries, RPC may be better at linking lack of support to onset of civil war than support to civil peace. Second, I use a commitment to property rights as a proxy for societal support. Where people feel that their money is secure in financial institutions regulated by the state, I assume support is high. Where people are unsure about the state’s commitment to protecting their assets, support is low. I use the indicator developed by Clague et al. (1999) called

9In the literature on American state politics and policy, this concept is referred to as “tax eﬀort” or the amount of taxes a state collects relative to the tax base available to the state. 10This claim was articulated during personal communication with the author. Sweden, for example, has a larger RPC than the US based on diﬀerent general preferences for state revenue accumulation. See Jackman (1993) for a similar critique.

47 contract-intensive money (CIM). This indicator is the ratio of currency held in banks to the total money supply. Advantages of this indicator over other approaches to property right’s protections and societal support is its temporal and cross-sectional coverage as well as its objectivity. Other similar indicators rely on subjective scores by country-level experts and lack broad cross-sectional as well as temporal coverage.11 When contracts are enforced by a central authority and property rights are respected by the government, citizens are more likely to keep their funds in banks. Where CIM is high, citizens believe that the state is committed to property rights protection whereas in states with low CIM the opposite is the case. In the estimation section, I display results for RPC.12

Institutions

To measure the concept of democratic institutions, I use the Polity (Democracy) data (Mar- shall and Jaggers, 2001). The concept of democratic institutions ranges from institutions that allow individuals to participate and also allow meaningful elite competition to institutions that exclude large segments of the population and do not allow contestation for higher office.13 While there are many available measures of democracy (see Munck and Verkuilen, 2002), Polity offers an index that uses clear and detailed coding rules and is comprehensive across time and space. I subtract the autocracy score from the democracy score to yield a measure that varies from -10 to +10 with high values corresponding to democratic institutions and low values corresponding to autocratic institutions. Bueno de Mesquita et al. (2003, 135) argue that the loyalty norm or W/S is the important institutional configuration that might explain taxation, leisure, economic growth and general support for the leader. I use this variable, W/S, as an alternative institutional measure. In the estimations, I display only Democracy, but similar effects were found for both measures.

11See for example the International Country Risk Guide data available online at http://www.icrgonline.com. 12CIM is highly correlated with GDP and inclusion of this variable inﬂates standard errors for both variables. Results for this measure are consistent with RPC but miss conventional levels of signiﬁcance (t=1.214). 13See Dahl (1971) for a thorough description of these attributes of democratic political systems or what he refers to as polyarchy.

48 Ethnic Fractionalization

The most common measure of ethnic fractionalization is the ethnolinguistic fractionalization (Ethnic F rac) index which provides the probability that two individuals drawn from a population are not from the same ethnic group. Fractionalization is thought to increase grievances for groups excluded from the state and thus reduce support for the state. This variable is taken from Fearon and Laitin (2003)’s data set.

Job Insecurity

Job Insecurity, as previously outlined, is the leader’s belief concerning his/her ability to retain office. Operationalizing this concept is somewhat difficult, but Cheibub (1998) offers a useful start. Cheibub (1998, 359-360) defines Job Insecurity (Job Insecurity) as the risk of losing office “given the length of tenure in office, the rate of economic growth, and the past rate of executive turnover in each country.” Leaders have high job insecurity when they have greater risks of losing office. In the language of survival analysis, Job Insecurity is the hazard of losing office. Using Cheibub’s original data, I replicated this measure, and then I extended it to other states and time periods.14 Using a survival model, I estimate the time to losing office for each leader in all of the countries in the sample.15 I then predict the hazard rate of losing office, given the change in GDP and cumulative changes in the chief executive, for each leader-year.16 This measure ranges from 0 to 1. Low values correspond to a low probability that the leader will lose office while high values suggest that the leader’s tenure is extremely insecure.

Resources

Resources and their link to civil war have been the topic of a robust literature (see Ross, 2006). The focus of this literature has been on lootable and nonlootable resources and their varying eﬀects on the likelihood of civil war. When I use the term resources, I am primarily

14In the years where our data overlap, our two measures correlate at 0.998. We both assume a Weibull distribution for the hazard. In other words, we assume that the hazard function is monotonically increasing or decreasing. 15I follow Cheibub in using this parametric form. Other parametric models yield similar hazard rates. 16Leader-year and country years are the same in the sample. Operationally, the ﬁrst year for a new leader is the year after the other leader loses oﬃce to avoid multiple country year records.

49 concerned with the nonlootable resources available to the state.17 To measure state resources, I use gross domestic product per capita (GDP ).

Military Capacity

To measure the concept of military capacity, I use the indicator military expenditures (MilExpend) from the Correlates of War Project.18 In short, the more a country spends on its military, the less likely civilians will violently challenge the state.

Controls

To ensure that results are comparable to other estimations, I use a basket of control variables from Fearon and Laitin’s (2003) dataset. These variables include the presence of a prior civil war (W art−1), logged population (P opulation), logged mountainous terrain (Mountains), having noncontiguous territory (NonContig), oil exporter (Oil), the ﬁrst two years of existence for a new state (New State), a major change in the regime score for a state (Instability), and religious fractionalization (Relig F rac).19 The variables that are most likely to covary with both my independent variables and dependent variable thus needing to be controlled include: population, new state, democracy, prior civil war, and instability. Finally, I correct for temporal dependence in my data by using a peace years variable that counts the time between onsets and cubic splines that smooth this function (Beck, Katz and Tucker, 1998).20

4.2.3 Missing Data

One diﬃculty in studying conﬂict in general and civil war in particular is that often data necessary to estimate models is missing for key indicators. The analyst has a choice to exclude cases where any of the variables used in the regression are missing (list-wise deletion), or to

17States, however, can generate resources in a variety of ways that are potentially substitutable including but not limited to foreign aid, tax revenues, colonial expansion, and natural resources. 18An alternative is to use military personnel, but having people does not necessarily mean that the military is effective. 19See Fearon and Laitin (2003) for a detailed discussion of each of these variables. 20Carter and Signorino (2006) offer an alternative to Beck, Katz and Tucker (1998)’s approach that uses the peace years variable (a time counter), this variable squared, and then this variable cubed to model time-dependency. The advantage of this approach is that it does not assume any parametric form for the time-dependency, and the coefficients are more easily interpretable than using cubic splines. I implemented both approaches and the estimates were very similar. Neither approach affects the main coefficients of interest and none of the coefficients of the time variables are statistically significant.

50 attempt to impute the missing values to avoid selecting out cases that may correlate with values of the dependent variable. In general, to impute data we must assume that the data are missing at random (MAR).21 That is, we assume missingness is due to some observable quantities.22 If data are missing at random, drawing inferences from data culled from listwise deletion can be biased and inefficient. Using imputation where a relatively large set of potential predictor variables are used to fill in missing observations makes the MAR assumption more plausible (Horton and Kleinman, 2007). Using Fearon and Laitin’s data as well as data from other sources creates a database with many covariates that can be used in predicting missing values. I use Amelia II (Honaker and King, 2006), a program that imputes time-series cross- sectional data, to create five separate datasets that are then combined and used in the estimations.23 Summary statistics for the data with missing values and for the imputed data are reported in Table 4.1 and Table 4.2. The standard approach to testing hypotheses in the quantitative study of civil war involves using a logit or probit model with a long list of covariates that explains the likelihood of a state producing civil war. Sometimes the effect of time is modeled (or controlled) (Beck, Katz and Tucker, 1998) and occasionally the relationship between the likelihood of civil war and a particular covariate, such as democracy, is re-specified, but rarely are conditional, mediating, or intervening processes explicitly modeled. At times this is due to lack of explicit theory that requires it, and other times these relationships are simply ignored. Both Achen (2002) and Ray (2005) argue for models that directly tie theory and estimation while attempting to remain parsimonious. Achen

21King et al. (2001) describe the difference between data that are MAR, missing completely at random (MCAR), and nonignorable (NI). Data that are MCAR are missing due to randomness and imputation can not help predict values for these observations. For example, according to King et al. (2001, 50-51) “an MCAR process is one in which respondents decide whether to answer survey questions on the basis of coin flips.” NI data are missing because of some unobserved factor. If data are NI, then either multiple imputation or listwise deletion yield biased estimates. Citing Little and Rubin (1989); Little and Schenker (1995), King et al. (2001, 51) note that ”[a]nalyses using listwise deletion are relatively inefficient, no matter which assumption characterizes the missingness, and they are also biased unless MCAR holds. Inferences based on multiple imputation are more efficient than listwise deletion (since no observed data are discarded), and they are not biased under MCAR or MAR.” 22This type of missingness is sometimes labeled covariate-dependent missingness. 23Other approaches exist to impute these data. See Horton and Kleinman (2007) for a review of these approaches and programs available to compute the imputations. I also imputed data using Royston (2005)’s ICE routine and the estimates were nearly identical.

51 Table 4.1: Summary Statistics

Variable Mean Std. Dev. N Onset 0.018 0.131 3588 Repress 2.636 1.150 3171 DissAct 0.370 0.583 3517 Support (RPC) 0.998 0.546 2915 Job Insecurity 0.191 0.166 3222 MilExpend 5.314 2.590 3442 W art−1 0.181 0.385 3588 Resources (GDP) 4.359 4.642 3440 P opulation 9.128 1.471 3432 Mountains 2.091 1.431 3588 Noncontig 0.159 0.365 3588 Oil 0.160 0.366 3588 Democracy 0.010 7.567 3559 New State 0.015 0.123 3588 Instability 0.152 0.360 3588 Ethnic F rac 0.408 0.284 3588 ReligF rac 0.381 0.219 3588 Possible N=3588 argues for simply using three regressors while Ray (2003, 4) claims that only variables that are confounding, or “antecedent third factor[s] that [bring] about a statistical association or correlation between two other variables,” and make an association spurious should be ‘controlled’ for. Similar advice is proffered by King, Keohane and Verba (1994). The standard approach to estimating models attempts to deal with omitted variable bias by including a list of ‘usual suspects’.24 While more variables are included in the specification to mitigate this potential pitfall, little attention is paid to the effect of including these variables. As Clarke (2005) demonstrates, including or excluding variables in a multivariate equation can either increase or decrease bias, and the researcher can not be sure the exact effect in any given case. Clarke advocates using the logic of research design rather than the logic of control variables to estimate models, meaning that theory should guide inclusion of relevant variables rather than ad hoc statistical speculations. Controlling for intervening variables, or variables that intervene in the causal process of

24Omitted variable bias occurs when some omitted factor covaries with both the critical independent variable and the dependent variable making the correlation between the two spurious.

52 Table 4.2: Summary Statistics–Imputed Data

Variable Mean Std. Err. of Mean N Onset 0.018 0.002 3588 Repress 2.592 0.021 3588 DissAct 0.369 0.010 3588 Support (RPC) 1.001 0.010 3588 Job Insecurity 0.188 0.003 3588 MilExpend 5.138 0.427 3588 W art−1 0.181 0.006 3588 Resources (GDP) 4.300 0.079 3588 P opulation 9.134 0.025 3588 Mountains 2.091 0.024 3588 Noncontig 0.159 0.006 3588 Oil 0.160 0.006 3588 Democracy 0.008 0.123 3588 New State 0.015 0.002 3588 Instability 0.152 0.006 3588 Ethnic F rac 0.408 0.005 3588 Relig F rac 0.381 0.004 3588 N=3588 x 5 Imputed Data Sets variables A and B, can lead to the elimination of the correlation between them, and this can lead to a false impression of no relationship between these variables. Since my theoretical model identifies the intervening factors between some classic predictors of civil war, such as resources, controlling for other variables in the theory would be a specification error. To be consistent with my theoretical model, dissident activity and state repression need to be treated as mediating variables or variables that intervene between conditions that affect these variables and the likelihood of civil war onset. It is in this spirit that I proceed and discuss the estimation techniques for drawing inferences. 4.3 Estimation

Since repression and dissident activity are directly aﬀected by other variables in the model, utilizing a single equation to estimate the eﬀects of these regressors on the likelihood of civil war would give biased estimates (Kennedy, 2003).25 Instead, I estimate equations predicting

25Since repression and dissent likely correlate with the error term, ﬁnding a suitable instrumental variable is a way to solve this potential threat to valid inference. A good instrumental variable is correlated with the

53 dissident activity and repression, and then use this information when estimating an equation predicting the onset of civil war. While two-stage techniques are appropriate for estimating these three equations, most two-stage techniques assume that the dependent variable is continuous. As Achen (1986, 49) has shown, if the second stage is a continuous dependent variable, the standard errors can be adjusted based on a weighting factor. If, however, the second stage is a binary variable, adjusting the standard errors is exceptionally complicated. This assumption is justified for both dissident activity and repression, but the onset of civil war is a binary indicator.26 As previously mentioned, estimating two-stage models with binary endogenous variables is more complicated than two-stage estimation with continuous variables. Two techniques provide consistent coefficient estimates when faced with this situation (Alvarez and Glasgow, 1999). First, a two-stage probit least squares (2SPLS) can be estimated that is similar to ordinary two-stage estimation. The first step is to estimate reduced-form equations by regressing each continuous endogenous variables on all exogenous variables. Next, predicted values for these reduced form equations are used as instruments for the endogenous variables. For the probit or logit equation, a similar technique is used except rather than predict the endogenous variable using OLS, probit is used. The predicted values for all three are then substitutes for the endogenous variables in the final equations. Rivers and Vuong (1988) offer an alternative estimator, two-stage conditional maximum likelihood (2SCML), that provides consistent coefficient estimates as well as consistent standard errors and an explicit test for endogeneity.27 In estimating the 2SCML, one first estimates the equations for the continuous endogenous variables. Next, residuals are saved for these equations. Third, these residuals are used in the logit/probit equation. Finally, another logit/probit model without the residuals is estimated. A likelihood ratio test (LR test) of the unrestricted model (with residuals) versus the restricted model (without residuals) is independent variable but uncorrelated with the error term (Wooldridge, 2000). In two-stage approaches, all of the exogenous variables in the system make a suitable instrument for the endogenous variable of interest. In practice, these exogenous variables are used to create an estimated value for the endogenous variable and this prediction serves as the instrument (Kennedy, 2003). 26Dissident activity ranges from 0 to 3.5 but can take on all values in between. Repression ranges from 1 to 5. While the values are ordinal, having at least five categories makes treating the variable as continuous plausible. Previous research using this measure does treat it as continuous and uses OLS regression to estimate models of repression (Poe and Tate, 1994; Poe, Tate and Keith, 1999). 27Alvarez and Glasgow (1999) were among the first to implement this approach in the political science literature.

54 implemented. The null hypothesis is that the variables are exogenous. Thus, rejecting the null allows for conﬁrmation that the variables are endogenous. I am primarily interested in estimating how certain factors aﬀect both state repression and dissident actions and how these actions by the central actors relate to the onset of civil war. To do so, I employ the 2SCML model. In the theory, I have expectations for other relationships as well.28 Below, I identify three equations to be estimated. Most of the hypotheses from the theory section relate to the equations dealing with dissident activity, repression, and civil war. In the second portion of the empirical tests, I estimate each of these equations from the theory section rather than this limited system and use techniques to estimate the entire series using limited information maximum likelihood.

Represst = β1JobInsecurityt+β2Resourcest+β3DissActt−1+β4Represst−1+controls+ǫ

DissActt = β5Supportt + β6DissActt−1 + β7Represst−1 + β8MilExpendt + controls + ǫ

P r(CW )t =β9DissActt + β10Represst + β11Repressresiduals + β12DissActresiduals + controls + ǫ

As I mentioned above, I calculate residuals from the equations estimating repression and dissident activity and then add these residuals to the final logit equation. In the next section, I also discuss the outcome of the likelihood ratio test.29 After estimating econometric models using 2SCML, I also perform a host of robustness checks to make sure that both adjusting the specification has little effect on inferences and that the models are robust to different codings of independent and dependent variables.

4.3.1 Results

Tables 4.3-4.6 display the results of the 2SCML estimations. I present the results from the model explaining the factors that aﬀect repression (Table 4.3), the results from the

28In the theory section of the dissertation, I outline several more hypotheses that relate to other equations in the system. I also do not estimate the effect that repression has on support (Hypothesis 2) in this estimation. An OLS regression supports this claim as well as limited-information structural equation models. 29In the second set of empirical tests (Chapter 5), I use structural equation modeling techniques to derive coefficient estimates for all of the variables in the system of equations from the theory section. Structural equation modeling (SEM) is useful for “...estimat[ing], and test[ing] complex multivariate models, as well as to study direct and indirect effects of variables in a given model” (Raykov and Marcoulides, 2006, 7).

55 Table 4.3: Impact of Resources, Job Insecurity, Previous Repression and Previous Dissident Activity on Repression

Variable Coefficient (PCSE) Job Insecurity 0.270∗∗ (0.079) ∗∗ Represst−1 0.506 (0.046) ∗∗ DissActt−1 0.141 (0.028) Resources -0.037∗∗ (0.005) Support -0.143∗∗ (0.030) MilExpend -0.000 (0.000) P opulation 0.079 (0.012) Mountains 0.025∗ (0.010) Noncontig -0.145∗∗ (0.033) Oil 0.169∗∗ (0.034) Democracy -0.021∗∗ (0.003) New State -0.122 (0.219) Instability 0.089∗ (0.038) Ethnic F rac -0.146∗∗ (0.056) Relig F rac -0.033 (0.045) Constant 0.695∗∗ (0.105) Significance levels : ∗ : p< 0.05 ∗∗ : p< 0.01 N=3588 estimation including factors that influence dissident activity (Table 4.4), the results from estimating the reduced form equation for civil war onset which includes the residuals from the previous estimations (Table 4.5), and the same model without the residuals (Table 4.6). First, I present the results from the repression and dissident activity models, then I build to the civil war onset model. As I argued in Chapter 3 most scholar conceive of civil war as the production of violent deaths between states and dissident groups that exceeds a given threshold. Few, however, attempt to model this process. The results that I present below support this argument. Violence by states and dissidents begets more state and dissident violence. The leader’s job insecurity is important for incentivizing leaders to repress, and a reduction in support for the state spurs dissident activity. Repression has a large impact on why some weak states develop civil wars while others do not. Table 4.3 displays the results for the repression equation. To reiterate, the factors identified in Chapter 3 that influence repression include job insecurity, resources, previous

56 Table 4.4: Impact of Support, Military Capacity, Previous Dissident Activity, and Previous Repression on Dissident Activity

Variable Coeﬃcient (PCSE) Support -0.033∗ (0.014) ∗∗ Represst−1 0.064 (0.011) ∗∗ DissActt−1 0.432 (0.044) MilExpend -0.001∗ (0.000) Job Insecurity 0.010 (0.029) Resources 0.000 (0.002) P opulation 0.006 (0.006) Mountains 0.016∗∗ (0.005) Noncontig 0.068∗∗ (0.023) Oil -0.021 (0.023) Democracy 0.056∗∗ (0.001) New State 0.115 (0.106) Instability 0.088∗∗ (0.025) Ethnic F rac. 0.029 (0.027) Relig F rac -0.070† (0.038) Constant -6.535∗∗ (0.909) Signiﬁcance levels : † : p< 0.10 ∗ : p< 0.05 ∗∗ : p< 0.01 N=3588 repression, and previous dissident activity. All other exogenous variables are included in the reduced form equation. Hypothesis 1–that Job Insecurity is positively related to Repress–is supported. Moving Job Insecurity from its minimum to its maximum leads to a 0.27 unit increase in Repress when holding other variables at their mean. In other words, increasing Job Insecurity from its minimum to its maximum results in an 11% increase in the expected value for Repress holding other variables constant. Hypothesis 4 also receives support as

Repress increases when DissActt−1 increases. When holding other variables at their mean, a unit increase in DissActt−1 leads to 0.141 increase in the expected value for Repress. As expected, states respond to last period’s dissident acts with repression. Similar to Poe and Tate (1994); Poe, Tate and Keith (1999), I ﬁnd that the more Resources a state possesses, the less likely it is to use repression. Results for the dissident activity equation are shown in Table 4.4. The top four variables are key factors hypothesized to inﬂuence dissident activity. All other exogenous variables are included in the reduced form equation. An important claim from the theory presented in

57 Table 4.5: Impact of Repression, Dissident Activity, and Residuals from Previous Equations on Civil War Onset (Unrestricted Model)

Variable Coeﬃcient (Std. Err.) Repress 1.543∗∗ (0.377) DissAct 0.066 (0.615) DissActresiduals 0.824 (0.715) Repressresiduals -0.505 (0.537) ∗∗ W art−1 -3.035 (0.571) Resources -0.264∗∗ (0.075) P opulation -0.001 (0.141) Mountains 0.058 (0.121) Noncontig 1.560∗∗ (0.476) Oil -0.063 (0.346) Democracy 0.043 (0.029) New State 3.195∗∗ (0.664) Instability 0.220 (0.357) Ethnic F rac 1.225∗ (0.587) Relig F rac -0.675 (0.838) Constant -8.610∗∗ (1.151) Signiﬁcance levels : ∗ : p< 0.05 ∗∗ : p< 0.01 N=3588

Chapter 3 is that a reduction in societal support leads to increased dissident activity. Based on the results from Table 4.4, Support is negatively related to DissAct (Hypothesis 5). The marginal effect is not large, but the relationship is in the expected direction. Societal support appears to marginally decrease dissident activity. Last period’s repression, Represst−1, is positively related to DissAct, supporting the idea that dissidents and states respond to each other’s previous actions. This finding lends support to the notion that this period’s repression may kill dissidents, but it also helps create the next period’s dissidents. MilExpend is negatively related to DissAct, lending support to Hypothesis 6: Potential failure discourages dissident activity. The military capacity of the state seems to be a deterrent to dissidents challenging their leaders. The more states spend on their military, the less likely dissidents will mount violent challenges to their authority. Finally, the results for the civil war onset equations are presented in Table 4.5 and Table 4.6. Table 4.5 displays the results for the unrestricted model, or the model that includes the residuals from the dissident activity equation and the repression equation. Table 4.6 offers

58 Table 4.6: Impact of Repression and Dissident Activity on Civil War Onset (Restricted Model)

Variable Coefficient (Std. Err.) Repress 1.100∗∗ (0.189) DissAct 0.828∗∗ (0.261) ∗∗ W art−1 -3.040 (0.550) Resources -0.309∗∗ (0.072) P opulation 0.052 (0.125) Mountains 0.068 (0.117) Noncontig 1.321∗∗ (0.474) Oil 0.098 (0.352) Democracy 0.026 (0.024) New State 2.642∗∗ (0.486) Instability 0.143 (0.358) Ethnic F rac 1.053∗ (0.536) Relig F rac -0.470 (0.791) Constant -8.045∗∗ (1.117) Significance levels : ∗ : p< 0.05 ∗∗ : p< 0.01 N=3588 the results for the restricted probit model, or the model without the residuals. A likelihood ratio test of the two models rejects the null of exogeneity (χ2 = 13.01, p< 0.01) providing support for the presence of endogeneity in the restricted probit model.30 DissAct and Repress are both positively associated with the likelihood of civil war onset, but only Repress is significant. In addition, Repress has a large substantive effect. The residuals are not significant, but the LR test provides support for their joint inclusion in the model. While the coefficient for Repress is clearly larger than many other variables in the model, substantive interpretation of this coefficient is more difficult in a logit model. To display the effects of Repress on the likelihood of civil war, I simulate predicted probabilities using Clarify (King, Tomz and Wittenberg, 2001) for the onset of civil war over the range of possible values of GDP while setting repression at its mean and the maximum.31 Figures 4.1 and 4.2 display the results of the simulations. On the Y axis is the predicted probability of civil war and on the X axis is GDP per capita in US dollars. The possibility

30The critical value is 5.99 for a model with 2 degrees of freedom or two residual terms. 31I hold all the controls at their mean. The simulations are performed 1000 times and draw values from the 5 imputed data sets.

59 Figure 4.1: Predicted Probability of Civil War–Repression at Mean

of civil war when holding repression at its mean, even among the most resource-poor states, is highly unlikely. When repression is held at its mean the expected probability of civil war remains less than 2% throughout the range of values for GDP . Even when GDP is close to zero, the predicted probability of civil war is a little more than 1%, plus or minus about 0.5%. When Repress is at its mean, the probability that a state produces civil war approaches zero as soon as GDP exceeds about $5,000. Civil war is a rare event for all states. When increasing repression to its maximum, the results change markably. In Figure 4.2, Repress is held at its maximum or at 5. As the figure shows, the effects of increasing Repress are quite strong. The probability that a state with GDP close to $1,000 produced civil war is nearly 30%. Uncertainty around this estimate is quite large as the 95% confidence interval ranges from about 10% to over 50%. While our confidence in a precise point estimate is less than what it might be, this simulation shows that the probability of civil war drastically increases for so-called weak states or resource-poor states when repression is increased to its maximum. This effect is large and the effect of maximum repression does not approach zero until GDP reaches about $7,000. The graphs are truncated as GDP ranges from near zero to above $30,000. As both of these graphs show, at about $5,000 or $7,000, the likelihood of civil war approaches zero and subsequent increases in GDP have nearly no effect on the

60 Figure 4.2: Predicted Probability of Civil War–Repression at Maximum

likelihood of civil war. Only Iran during this time period32 had a GDP greater than $6,000 and experienced a civil war.33 There are a few interesting results to note from the group of control variable from Fearon and Laitin. First, GDP still has a negative significant effect, but this effect is reduced by 23% (from -0.344 to -0.264) when the residuals from the previous equations are included along with Repress and DissAct.34 Second, many of the structural variables become insignificant suggesting that factors, such as mountains and populations, have a marginal impact on states’ and dissidents’ abilities to repress or dissent but have no direct impact on the likelihood of civil war. Opportunity approaches, like Fearon and Laitin and Collier and Hoeffler, claim that ethnicity does not play an important role in the onset of civil war. In these results, the

32The UK also fits this description, but their onset occurred in 1969 which is outside the temporal domain of this study. 33Interestingly, as Przeworski notes, no democracy with a per capita GDP above $6,055 has ever transitioned to an authoritarian regime. Argentina is the sole democracy above $6,000 that made such a transition. 34I am comparing the coefficients from Fearon and Laitin’s full model to my unrestricted model with a limited time period. If I compare Fearon and Laitin’s results while restricting the sample to the same period ( 1976-1999) to the unrestricted model, the results are even more substantive; there is a 38% reduction (from -0.426 to -0.264) in the size of the coefficient. These results are available in the Appendix.

61 impact of Ethnic F rac is positive and significant. Similar to Blimes (2006), I find an effect for ethnic fractionalization when choosing a different modeling strategy than Fearon and Laitin. A final interesting note is that states that export oil tend to repress more but are not more likely to generate civil war. Because these states have a resource advantage over dissidents, they may be able to stave off civil war even though few citizens support the state.35 Although many scholars think of civil war as the production of violent deaths between states and dissident groups, few modeling strategies reflect this. The results here support this approach. Previous repression and dissident activity affect current repression and dissident activity. Societal support and job insecurity both affect the choices made by dissidents and states respectively. Finally, repression has a large effect on the likelihood of civil war onset.

4.4 Conclusions

What the theory and statistical model provide is an answer to the puzzle of why civil war only occurs in some weak states: We are most likely to witness civil war in weak (low resource) repressive states. This group of weak states are susceptible to dissident challenges and they increase the likelihood of civil war via repression that reduces popular support and thus stokes the process of violence. Similar weak states like Zambia or Malawi have extremely low GDP, but their governments have average repression levels in this sample that are much lower than states like Iran, El Salvador, or Somalia. Returning to Figure 4.1, Bhutan had an average GDP from 1976 to 1989 of $571 and an average repression score of 1.6. As the ﬁgure shows, Bhutan’s predicted probability for civil war is close to 1% plus or minus about 0.5% holding other variables constant.36 El Salvador, on the other hand, had a repression score (4.3) that approaches the maximum in the few years prior to civil war in 1979. In addition, their GDP averaged $2,099 over this time period. Based on these values, their expected probability for civil war from 1976 to 1977 was 13% while holding other variables at their mean.37 In 1978, El Salvador’s expected

35Whether economic crisis, or when the state’s resources are strained, leads to civil war in oil states is an interesting extension of this idea not yet explored in the literature. 36This is an overestimate as Bhutan’s average repression score is nearly a point below the mean. 37The 95% conﬁdence interval ranges from 3.4% to 34%. While this is a wide conﬁdence interval, the percentage is always much higher than the expected values when repression is held at its mean. In addition, the upper bound is further from the mean than the lower bound.

62 probability of civil war was over 30%.38 While this percentage is not approaching certitude, the difference in likelihood between Bhutan and El Salvador experiencing civil war during this time is quite large. Beyond looking at the substantive impact of repression, it is important to look at the predictive ability of a model of civil war (Ward and Bakke, 2005). Since scholars and policy makers are concerned with where civil war may happen next, building accurate predictive model is a chief concern. Previous studies of civil war have focused on uncovering significant findings rather than the ability to accurately predict cases of war or not war. The use of the Receiver Operating Characteristic (ROC) curve is one way to compare the in-sample predictive ability of a model.39 The ROC curve provides a graph that shows how the model may make Type I or Type II errors given different cutpoints. In the case of civil war, a Type I error is predicting civil war when one did not occur. In contrast, a Type II error is predicting no civil war, when one actually happened. Other methods, such as the percent correctly predicted (PCP), are sensitive to the threshold that one uses when establishing a prediction of 1 versus 0. The area under the ROC curve gives the percentage of cases correctly predicted and provides an estimate of model fit (Beck et al., 2001). The area under the curve for Fearon and Laitin’s full model using data from 1945-1999 is 0.760. In the sample from 1976-1999, the prediction slightly improves to 0.796. For the model that includes the measures of repression, dissident activity, and residuals from the prediction equations for the two variables, this number increases to 0.886. Including splines and years at peace only marginally increases the predictive ability to 0.890. The results from the ROC curve show that a process theory of civil war also provides better in-sample predictions of the likelihood

38The 95% confidence interval ranged from 6% to 73%. 39Several methods can be used in the binary dependent variable context to predict the potential outcome of interest. These methods are often used to assess model fit as opposed to using a R2 or pseudo R2. One way to do this is to use the percent correctly predicted or PCP. Each observation generates a predicted probability, pî, for the binary dependent variable. Whenp î is ≥ 0.5, the model predicts a 1 for the outcome of interest; and whenp î is < 0.5, the model predicts a 0 for the outcome of interest. Different cutpoints besides 0.5 can be used to either assess 1’s and 0’s or to test the sensitivity of the predictions to the cutpoint, but 0.5 is the standard value used. After generating these predictions, dividing the number correctly predicted by the sample and multiplying by 100 gives the PCP. PCP has been criticized as an observation has the value 0.501 is treated the same as one that has the value 0.999. The percentage reduction in error (PRE) is another alternative but is similar to the PCP. The ePCP and Receiver Operating Characteristic (ROC) curves are two alternatives that provide similar information as the PCP and PRE but overcome some if their faults. The PRE and PCP assumes that both Type I and Type II errors are equally bad. Increasing or decreasing the cutpoint privileges either the Type I or Type II error. A ROC curve gives this information for all possible cutpoints.

63 of a country experiencing civil war onset. Building a model of civil war should take into account that it is a process between the state and dissidents that does not simply happen but builds through the interaction of the actors doing the killing. Because of this, previous studies have not provided adequate microfoundations for the onset of civil war. In the next chapter, I use structural equation modeling to establish whether the results are dependent upon model assumptions or not. In addition, structural equation modeling allows for an investigation of indirect eﬀects of variables on civil war and estimation of all the relationships expressed in the theory from Chapter 3.

4.5 Appendix

Table 4.7: Impact of Fearon and Laitin Variables on Civil War Onset Using Imputed Data From 1976–1999

Variable Coeﬃcient (Std. Err.) ∗∗ W art−1 -1.251 (0.428) Resources -0.324∗∗ (0.074) P opulation 0.265∗∗ (0.092) Mountains 0.194† (0.104) Noncontig 0.673† (0.392) Oil 0.444 (0.310) Democracy 0.006 (0.021) New State 2.466∗∗ (0.462) Instability 0.633† (0.358) Ethnic F rac 0.758 (0.512) Relig F rac -0.417 (0.729) Constant -6.535∗∗ 0.909) Signiﬁcance levels : † : p< 0.10 ∗ : p< 0.05 ∗∗ : p< 0.01 N=3588

Results for the Sambanis coding for civil war are similar to the results for PRIO/Uppsala and Fearon and Laitin. Two important diﬀerences matter. First, the Sambanis data show strong temporal dependence. I included splines and years at peace measures and time, time squared, and time cubed to deal with this issue. Second, the war years lag perfectly predicts the dependent variable and is dropped from the analyses.

64 Table 4.8: Impact of Fearon and Laitin Variables on Civil War Onset using Non-Imputed Data from 1976-1999

Variable Coeﬃcient (Std. Err.) ∗ W art−1 -0.965 (0.401) Resources -0.426∗∗ (0.113) P opulation 0.283∗∗ (0.096) Mountains 0.187† (0.106) Noncontig 0.372 (0.454) Oil 0.419 (0.383) Democracy 0.015 (0.023) New State 2.572∗∗ (0.518) Instability 0.692∗ (0.317) Ethnic F rac 0.554 (0.539) Relig F rac -0.517 (0.703) Constant -6.320∗∗ (0.988) Signiﬁcance levels : † : p< 0.10 ∗ : p< 0.05 ∗∗ : p< 0.01 N=3289

Table 4.9: Civil War Onset–PRIO/Uppsala Coding of Civil War

Variable Coeﬃcient (Std. Err.) Repress 0.941∗ (0.424) DissAct -2.084∗∗ (0.731) DissActresiduals 3.344 (0.764) Repressresiduals -0.311 (0.451) ∗∗ W art−1 2.070 (0.468) Resources -0.124 (0.086) P opulation 0.117 (0.146) Mountains 0.178 (0.153) Noncontig 0.712 (0.519) Oil 0.351 (0.467) Democracy 0.017 (0.033) New State 2.827∗∗ (0.883) Instability 0.523 (0.397) Ethnic F rac 0.198 (0.671) Relig F rac 0.155 (0.877) Constant -8.813∗∗ (1.478) Signiﬁcance levels : † : p< 0.10 ∗ : p< 0.05 ∗∗ : p< 0.01 N=3527

65 Repress ● DissAct ●

Young Repress_res ● Dissact_res ● ● War_t−1 ● ● Resources ● ● Population ● ● Mountains ● ● Noncontig ● ● Oil ● ● Democracy ● ●

Fearon and Laitin New State ● ● Instability ● ● Ethnic Frac ● ● Religious Frac ● −4 −1 1 3

Figure 4.3: Comparing Regression Coeﬃcients

66 CHAPTER 5

STRUCTURAL EQUATION MODELING

5.1 Introduction

The previous chapter used two-stage estimation techniques to investigate the effects of repression and dissent on the onset of civil war. Because these variables are mediating the impact that other key structural variables like mountainous terrain and ethnicity have on civil war onset, controlling for them or simply estimating a single equation with repression, dissident activity, mountainous terrain, and ethnicity could lead to biased inferences (Ray, 2003).1 Besides using a two-stage procedure, another approach is to use structural equation modeling (SEM).2 The basic idea behind structural equation modeling is to use multiple equations to estimate the relationships among variables. Using SEM, I can estimate the direct and indirect effects of key variables on the onset of civil war as well as other variables of interest such as repression and dissident activity. Figure 5.1, first introduced in Chapter 3, shows the path diagram of the theory proposed. In the two-stage estimation, I had to focus on estimating the effects of variables on repression and dissent, then use this information in estimating the onset of civil war. SEMs allow me to estimate the relationships among all variables and to probe the robustness of the results given a different modeling strategy. Whereas the two-stage models assume that the independent variables in the model are measured without error, SEMs “take into account potential errors of measurement in all observed variables” (Raykov and Marcoulides, 2006). SEMs include

1As Gujarati (2003, 718) notes that “in simultaneous-equation models one may not estimate the parameters of a single equation without taking into account information provided by other equations in the system.” If this violated, the resulting estimates are “not only biased but also inconsistent.” 2Path analysis is one type of SEM where the analyst only regresses observed variables on potential explanatory variables. When these variables are latent, terms such as conﬁrmatory factor analysis, structural regression models, or latent change models is used. For a discussion of the diﬀerences among these approaches to SEM see Raykov and Marcoulides (2006).

67 error terms for the variables that are potentially measured with error and estimate the variances of these terms. Two-stage approaches and SEMs share many assumptions in common such as linear associations among variables.3 This is a particular problem for estimating civil war as a variable since it is a binary endogenous variable. I discuss how to deal with this problem below.

Figure 5.1: Path Model of Civil War Onset

I first show the different effects of variables within each ‘block.’4 In other words, I show the results of regressing each endogenous variable on its predictors, then build up to the large model. Next, I discuss the direct and indirect effects of resources on the onset of civil war. In the conclusion, I compare these results with the two-stage results and make suggestions for further modeling choices.

3At times, theoretical implications can point the analyst towards models that match theory with assumptions. In this case, it is unclear whether the assumptions of two-stage models or SEMs are more realistic. If the results are robust, I can be more conﬁdent that the estimates are not model dependent. 4Sometimes this term is used in SEM to denote a group of variables that have a recursive relationship in a larger nonrecursive model. I use the term to denote the single equation for an endogenous variable with its direct predictors.

68 5.2 Estimation

As identified above, I use structural equation modeling techniques to derive coefficient estimates for all of the variables in the system of equations from the theory section. Before discussing the results, I outline a few key terms: direct effects, indirect effects, total effects, limited information estimation, polyserial correlations, and latent variable modeling. After introducing these I terms, I discuss how they relate to choosing an estimation procedure for evaluating the model outlined in Chapter 3. First, a direct effect is the uninterrupted impact that one variable has on another. In other words, it is the effect that some X has on Y. An indirect effect is the effect that one variable has on another that is mediated by one or more other variables. For example, if X affects Z which in turn affects Y (which is depicted using the arrow diagram in Figure 5.2), X’s effect on Y can be described as indirect through Z. Moreover, we can speak of X’s total effect on Y, which is defined as the sum of X’s direct effect on Y plus X’s indirect effect on Y through Z. In the most simple version of path analysis, Ordinary Least Squares (OLS) regression can be used to estimate these direct, indirect and total effects (Asher, 1976). The mediating variables, such as Z, are regressed on X. Y is then regressed on X and Z, and the analyst can compute the direct effects of X and Z as well as the indirect effect of X on Y though Z.

X Y

Figure 5.2: Direct and Indirect Eﬀects

A standardizing or weighting procedure may then be used so that these quantities can be multiplied to find the indirect effects. In the OLS context, to find the indirect effects one multiplies the standardized or unstandardized path coefficients. For example, if the

69 standardized path coefficient for X’s effect on Z is 0.25 and the standardized path coefficient for Z’s effect on Y is 0.3, then we can multiply these coefficients to find the indirect effect of X on Y through Z. If the standardized path coefficient for X’s effect on Y is 0.5, then the total effect for X on Y is 0.5+(0.25 ∗ 0.3) = 0.575. In substantive terms, taking into account both the direct and indirect effects, this means that, on average, a one standard deviation increase in X leads to a 0.575 standard deviation increase in Z.5 As mentioned in the previous chapter, this computation becomes much more complicated when either the mediating or dependent variable is not continuous. Fortunately, Muthén (1984) offers a solution by using what are termed polyserial- correlation coefficients, or the degree of association between ordinal and continuous variables. Using this measure of association and insights from latent variable modeling framework (Muthén and Muthén, 2004), it is possible to estimate direct and indirect effects for SEM with ordinal and dichotomous mediating and dependent variables.6 Latent variable modeling assumes that some latent, or unmeasured variable, underlies an ordinal, categorical, or dichotomous manifest, or measured, variable. In the case of civil war onset, the measured variable is dichotomous: coded one when civil war onset occurs and zero otherwise. A latent variable model of civil war onset assumes that there is an underlying propensity of a state ranging from zero to one that generates the observed coded values of the manifest variable. Although this underlying propensity can not be observed, at some point, we can observe the state change (a state experiences civil war onset). Usually, one needs to specify a threshold or cutpoint, for this change from one to zero. The important point to consider is that some underlying continuous dimension explains whether a state is coded as one or zero.7 Estimating this dimension then finding the correlation between the continuous measure and this underlying propensity is another way to respond to the problem of having a correlation between a continuous and categorical/dichotomous variable. To estimate the series of equations, two broad approaches to the structural equations can be utilized: full-information and limited information. Limited information methods estimate the structural parameters of each equation separately taking into account only the restrictions placed on the variables in that equation (Gujarati, 2003, 762). In contrast, full

5Standardizing variables is necessary to assess indirect eﬀects within a model. 6The Mplus software package implements a limited information weighted least squares routine that allows for these calculations. 7This way of thinking also ﬁts precisely with the conception of civil war onset outlined in Chapter 2.

70 information methods estimate all of the equations in a model simultaneously taking into account all restrictions placed on each equation (Gujarati, 2003, 763). Full information techniques potentially offer greater efficiency, but they also can increase bias if any of the equations are mis-specified. Some evidence suggests that the potential gains in efficiency are modest as compared to the likelihood of increasing bias in the coefficient estimates (West, 1986). Because of the sensitivity of full-information methods to specification error, I utilize a limited information technique when estimating the structural equations. Limited information techniques include least squares, maximum likelihood, and asymptotically distribution free (or weighted least squares.) The first two approaches, assume some form of normal distribution for the variables involved whereas the weighted least squares approach does not. The weighted least squares approach, however, requires a large sample to produce consistent estimates. As discussed above Muthén (1984) describes an estimator that utilizes a limited information weighted least squares approach to estimating structural equation models with categorical dichotomous variables. I estimate this series of equations using Muthén (1984)’s estimator:8

P r(CW )t = β1Represst + β2DissActt + ǫ1 (5.1)

Represst = β3DissActt−1 + β4Resourcest + β5JobInsecurityt + β6Represst−1 + ǫ2 (5.2)

DissActt = β7Represst−1 + β8SocSupt + β9Rt + β10DissActt−1 + ǫ3 (5.3)

SocSupt = β11Democracyt + β12Ethnicfract + β13Represst−1 + β14SocSupt−1 + ǫ4 (5.4)

JobInsecurityt = β15Leader∆t+β16T imet+β17EconGrowt+β18JobInsecurityt−1+ǫ5 (5.5)

8Equation 5.5 is estimated separately using survival models to predict the job insecurity for a leader in a given country-year building on Cheibub (1998).

71 Resourcest = β19Resourcest−1 + ǫ6 (5.6)

Some other estimation details are important. To compute standard errors, I utilize bootstrapping techniques. In sum, these standard errors are computed by taking a resample from the original sample, calculating the beta, then repeating (Efron and Tibshirani, 1986).9 The standard error of this sampling distribution is the bootstrapped standard error. For this model, I used 1000 simulations to generate the standard errors. In the next section, I provide the results for the structural equation models.

5.3 Results

Table 5.1 provides the R2 for each endogenous, or dependent, variable. As is well known, this statistic provides one measure of how much variation in the endogenous variable is explained by its regressors. It is not meant, however, to provide the only evidence of model ﬁt. The limitations of this statistic are well-documented, and including lagged regressors helps inﬂate this value (King, 1986). One interesting thing to note is that the variation in Resources is almost completely explained by last year’s value. Recall that this measure is argued to be the best predictor of civil war onset by Fearon and Laitin (2003). Since this variable is nearly time-invariant, it lends credibility to my earlier critique of Fearon and Laitin’s model as only able to explain cross-sectional rather than temporal variation in civil war onset.

Table 5.1: R2 for Each Dependent Variable

Observed Variable R-Square Onset 0.197 Repress 0.753 DissAct 0.437 Support 0.832 Resources 0.997 RMSEA = 0.000

Another goodness-of-ﬁt measure for SEMs is the Root Mean Square Error of Approxima-

9This is often done with replacement as the resample should be the same approximate size as the original sample.

72 tion (RMSEA). This index evaluates the extent to which the model fits the data.10 RMSEA is often preferred to other tests as it is one of the least affected by sample size. When the results from this test are less than 0.05, the model “reasonably” fits the data (Browne and Cudeck, 1993). The results from the RMSEA for this model suggests that this is the case.

Table 5.2: Expectations for Variables From Theory

Variable Dependent Variable Expectation Repress Onset + DissAct Onset + Resources Repress - Job Insecurity Repress + Represst−1 Repress + DissActt−1 Repress + Resources DissAct - Support DissAct - Represst−1 DissAct + DissActt−1 DissAct + Represst−1 Support - Democracy Support + Ethnic F rac Support + Supportt−1 Support +

The first set of results of the model relate to the definition of civil war as an outcome of state-dissident interaction.11 Both coefficients for DissAct andRepress are positive,12 but only repression is significant.13 I also estimated a model with Resources directly

10In any given sample, the RMSEA is calculated in the following way: π = p(T − d)/(dn) where n is the sample size minus 1, T is the test statistic, and d is the degrees of freedom for the model (Raykov and Marcoulides, 2006). 11It is conventional in path models to use directional arrows to signify one variable affecting another. If the arrow is two sided, this implies reciprocal causation and is termed a nonrecursive path model. Since this model is hierarchical with variables operating in sequence, it is recursive. To be recursive, we must also assume that the error terms are uncorrelated with exogenous variables and with each other (Berry, 1984). This assumption is difficult to uphold in this context as the error terms could be autocorrelated. The violation of this assumption potentially leads to biased coefficient estimates and deflated variances (Gujarati, 2003). This could lead to a false rejection of a null hypothesis. Another convention is to use circles to denote latent variables and squares for manifest variables. I use squares as all variables are manifest. I also omit the error terms from the diagrams. Recursive models are also identified, meaning that it is possible to solve the system that includes a group of unknown parameters. 12In Figure 5.2, I provide a table of each variable’s expected direction of effect on the dependent or endogenous variable of interest. 13In all of the path diagrams, significant (p<.05, two-tail test) coefficients are bolded. The standard

73 influencing Onset. The coefficients are nearly identical for Repress (0.373) and DissAct (0.197). Repress, again, is the only coefficient that achieves conventional levels of statistical significance (The Resources coefficient is -0.239 and standard error is 0.150). Since these coefficients are standardized using the variances of the continuous latent variable, they can be interpreted like standardized coefficients from a path model. In other words, a one standard deviation increase in the repression indicator is expected to increase the likelihood of civil war by 0.381 standard deviations.14 Repress thus has a strong, statistically significant effect on Onset. This result is consistent with the two-stage outcomes.

0.200 Dissident (0.198) Activity 0.381 Onset of (0.103) Civil War Repression

Figure 5.3: Coeﬃcient Estimates for Civil War Onset Predictors

Figure 5.4 displays the results of the regressors on Repress. All of the coefficients are in the expected direction, but the Job Insecurity measure fails to achieve statistical significance (t-score of 1.013).15 Resources held by the state reduce expected repression: a one standard deviation increase in Resources is expected to decrease Repress by 0.082 standard deviations. Previous repression and previous dissident activity like in the dissident activity equation above are strongly positively related to repression in the present period. A one standard deviation increase in Represst−1 increases the expected value for Repress by 0.755 standard deviations. Early quantitative work on this subject, such as Hibbs (1973), found a positive relationship between state repressive acts and violent dissident responses. According to Hibbs (1973, 91) the “consequence of repression is often a response of yet greater violence by errors are also listed in parentheses directly below the coefficients. 14For the substantive interpretations of the standardized effects, all other variables are held at their mean. 15This is the same variable used in the previous estimations. When I use a more fully specified measure that includes more covariates to predict the likelihood of losing office, the results are stronger and significant (0.121, t = 2.287). This alternative specification includes the covariates: the number of irregular leadership changes since 1945, the size of the winning coalition, and the age of the leader.

74 -0.082 Resources (0.019)

0.010 Job (0.009) Insecurity 0.755 Repression (0.017) Repression t-1 0.183 (0.029)

Dissident Activityt-1

Figure 5.4: Coeﬃcient Estimates for Repression Predictors

its recipients.” Past repression is positively associated with repression and the reason as mentioned by (Gurr, 1988, 49) is that “elites who have secured state power and maintained their positions by violent means are disposed to respond violently to future challenges.” Moreover, this is rooted in rational calculations or as Gurr continues “[s]uccesful use of coercion enhances leaders’ assessment of its future utility.” For the dissident activity equation, all the coefficients are in the expected direction and most achieve statistical significance. Only the effect of Resources can not be distinguished from zero. A one standard deviation increase in Support, on average, reduces DissAct by 0.117 standard deviations. Previous repression, or Represst−1, and previous dissident activity, or DissActt−1, are both positive and significant lending support to the claim that a process of violent tit-for-tat spawns dissident activity. When DissActt−1 increases by one standard deviation, DissAct is expected to increase by 0.473 standard deviations. As Figure 5.6 demonstrates, the predictors for Support are all in the expected direction, and only Represst−1 is not statistically significant (t = −0.975). Democracy increases support for the state and ethnic fractionalization reduces it. Both effects are relatively small compared to the large coefficient for the lag. Models estimated without the lag have larger coefficients for all the predictors and have significant estimates for Represst−1.

Since Resources is solely a function of Resourcest−1 (Equation 5.6), I do not provide a path diagram. Job Insecurity is also estimated using a survival model. The coeﬃcients for

75 -0.005 Resources (0.014)

-0.117 Support (0.054) 0.134 Dissident (0.015) Activity Repression t-1 0.473 (0.029)

Dissident Activityt-1

Figure 5.5: Coeﬃcient Estimates for Dissident Activity Predictors

-0.006 Repressiont-1 (0.005)

0.002 Democracy (0.001) -0.026 Support (0.008) Ethnic Frac 0.878 (0.033)

Supportt-1

Figure 5.6: Coeﬃcient Estimates for Societal Support Predictors

the predictors are significant and the results are robust to different parametric approaches to estimating these models. I use the Weibull model to remain consistent with Cheibub (1998). Figure 5.7 offers one final path diagram for the coefficients from the structural equation model that aggregate results from Figures 5.5 through 5.6.16 Table 5.3 also provides a list of whether each hypothesized relationship found statistical support. Repression, as discussed

16I include the relationships that are both statistically significant as well as those that are not. The results that are significant remain bold, and the arrows are black. In contrast, the insignificant results are normal font, and the arrows are grey.

76 Table 5.3: Expectations and Outcomes for Variables From Theory

Variable Dependent Variable Expectation Supported? Repress Onset + Yes DissAct Onset + No Resources Repress - Yes JobInsecurity Repress + No Represst−1 Repress + Yes DissActt−1 Repress + Yes Resources DissAct - No Support DissAct - Yes Represst−1 DissAct + Yes DissActt−1 DissAct + Yes Represst−1 Support - No Democracy Support + Yes Ethnic F rac Support + Yes Supportt−1 Support + Yes above, is a major determinant of whether a state produces civil war. Previous repression and previous dissident activity increase repression. Resources available to the state dampen repression. Support reduces dissident activity, and support is increased through democratic institutions. Having an ethnically heterogeneous society reduces support. The final group of effects to discuss relate to indirect effects of key variables. While dissident activity’s direct effect on civil war was not statistically distinguishable from zero, this is not the case for its indirect effect. Dissident Activityt−1, through Repress, is positively associated with Onset (0.069, t =3.214). In other words, a one standard deviation increase in Dissident Activityt−1 through Repress, on average, leads to a 0.069 standard deviation increase in the likelihood of civil war. Previous dissident activity’s effect on civil war onset is thus mediated by repression. Since dissidents using violence increases repression, this indirectly increases the probability of civil war onset. Resources also have an indirect effect on the likelihood of civil war that has been previously overlooked in models of civil war onset. Resources reduce the likelihood of civil war indirectly (-0.032, t = −2.513). These effects sum the indirect effects of resources mediated by both repression and dissent. Most of the effect is captured by the indirect effect through repression (-0.031).17 Resources both

17Other indirect eﬀects such as Ethnic Fractionalization’s eﬀect on Dissident Activity can also be

77 Democracy Ethnic Frac

0.002 -0.026

Supportt-1 0.878 Support

-0.117

Dissident 0.473 Dissident Activity Activityt-1 0.200

0.183 -0.005 Resources Onset of Civil War

-0.006 0.134 -0.082 0.381

Repressiont-1 0.755 Repression

0.010 Job Insecurity

Figure 5.7: Final Path Model

reduce the likelihood that states use violence and that dissidents challenge states and thus indirectly dampen the path to civil war onset.

5.4 Discussion

While the results from the SEM model generally corroborate the findings from the two-stage models, the results from a few variables are not consistent. Most importantly, the results from dissident activity and job insecurity have the expected sign but do not achieve statistical significance. Alternative specifications of job insecurity are significant and in the expected direction suggesting that more work should be done on creating a measure of leader insecurity that aims to closely match the concept with the indicator. In addition, one alternative is to find different measures of dissident activity and either substitute them or use factor analysis calculated. In most cases, these other effects are fairly small. For example, a one standard deviation increase in Ethnic Fractionalization decreases Societal Support by 0.001 standard deviations, on average.

78 to extract the common “dissident activity” factor from a set of indicators related to this factor. SEMs can also be estimated simultaneously with this factor analysis.

5.5 Conclusion

As I mention in Chapter 4 most civil war scholars use single-equation methods. For some theories of civil war onset, single-equations may be suitable as long as mediating or intervening factors are not included in the specification. Most statistical models of civil war assume that all independent variables are exogenous. Usually independent variables are pitted against each other in competition to explain the variance in the dependent variable. But as Blalock (1969, 36) argues “...as soon as one allows for the possibility that some of the “independent variables” cause the others, this reasoning breaks down and must be replaced by simultaneous-equation procedures.” As Ray (2003, 7) claims, this is a problem because when intervening variables are controlled for, this “will eliminate the statistical correlation between th[e] original independent and dependent variables.” I argue in Chapter 4 and elsewhere that by definition civil war models need to account for the mediating effects of repression and dissent. While other scholars have similarly complex theories of onset, they often do not concern themselves with estimating series of equations. Why not? First, the convention of using a logit/probit model is fairly easy to do. Fearon and Laitin (2003) and Collier and Hoeffler (2001) estimate their models of civil war using this approach, and these papers are by far the most influential on the course of this research program.18 Second, by making their extensive data publicly available, Fearon and Laitin (2003) have created a ‘path dependent’ scenario where subsequent analyses are repeatedly compared to their initial estimates. Since these estimates were found using single-equation logit models, this creates incentives for subsequent scholars to do the same. If all of the theories of civil war onset did not include intervening factors, then this approach may be valid for these theories. If, however, more complex theories are identified, the statistical tests should attempt to match the theory. As more scholars move towards specifying microfoundations for political behavior, the need to match the specifics of the theory with the statistical estimator increases (Achen, 2002). While using SEMs and two-stage models may be more difficult,

18According to a research Google Scholar search, Fearon and Laitin (2003) has been cited 746 times and Collier and Hoeﬄer (2001) has been cited over a 1000 times.

79 they are necessary if the theories that one espouses fit these techniques. While Chapters 4 and 5 have used alternative approaches to estimation to assess the process model of civil war, the next chapter takes a step back and investigates pre-processing the data. Instead of attempting new estimation techniques, I instead look at the causal effect of repression as it is the most robust finding from these models. Sorting out the effect that repression has on the onset of civil war can be difficult for many reasons. Classic regression attempts to find this impact by estimating correlations and “controlling” for confounding factors. Regression sorts the data by group and looks at the impact that a group with maximum repression has on civil war onset, a group with minimum repression, and all values in between. The problem with this is that the group with maximum repression might have values for other key variables that are systematically higher or lower than the values for these variables in the minimum repression group (or any of the groups in between). If this is the case, it is hard to sort out the impact that repression has when some other key variable is systematically larger or smaller. A way to mitigate this problem is to try and isolate the treatment effect of repression on civil war onset by creating groups of data that have similar values or distribution of values for other key variables. Propensity score matching techniques allow analysts to isolate the treatment effects of variables while balancing other covariate distributions across the comparison groups. In the next chapter, I use propensity score matching to find what the treatment effect of repression is on the onset of civil war. In addition, I propose an approach to propensity score matching that takes into account time-series data.

80 CHAPTER 6

CIVIL WAR, CAUSAL INFERENCES, AND PROPENSITY SCORE MATCHING

6.1 Introduction

One potential benefit of an improved conceptualization of state capacity and of a theory that provides microfoundations for civil war is to produce a model that better predicts civil war. Current estimation strategies focus on finding variables that covary with the likelihood of civil war onset with little regard for whether the model reasonably predicts civil war onset (Ward and Bakke, 2005). As Ward and Bakke (2005) demonstrate, the canonical model (Fearon and Laitin, 2003) predicts exactly zero onsets of civil war.1 A naive model that predicts zero for every country for every year would be correct 98.3% of the time. Fearon and Laitin’s model provides no improvement over the naive model. With this in mind, aside from producing improved specifications of current models, I show how conceiving of repression as a treatment improves our predictions of civil war. In this chapter, I apply a treatment to a model of civil war to see how isolating this causal effect increases or decreases the likelihood of the onset of civil war. The treatment is having high repression. I expect that a state that is treated with high repression should be more likely to experience civil war than one that has not. By establishing balance among the covariates, I can isolate the impact that having a high level of repression has and reduce the dependence on modeling assumptions and specification in producing these effects (Ho et al., 2007). The two-stage and structural equation modeling approaches also dealt with the problem of endogeneity or that the values of repression and dissident activity are dependent upon other predictors in the model or correlated with the error term. Matching

1This assumes that an observation has to generate a predicted value above 0.5 to count as predicting an onset.

81 also helps to mitigate this problem (Trujillo, Portillo and Vernon, 2005; Mocan and Tekin, 2006).2 Matching attempts to mirror a randomized experiment where two groups are almost identical except that one group receives a certain treatment (Rosenbaum and Rubin, 1983). This treatment effect can then be disentangled from the effects of other variables. Using repression as a treatment, I match observations using propensity scores and compare those that received the treatment with those that have similar propensity scores but did not receive the treatment. Ward and Bakke (2005) matched on prior civil war and instability and were able to increase their ability to predict civil war. While their matching procedure improved the predictive ability of the model, Ward and Bakke (2005) did not offer any particular theory for why those treatments were important in explaining the onset of civil war. In previous chapters I have shown that states that repress are more likely to generate civil war through the process of state-dissident interaction. While the last two chapters estimated the causal effect of repression as it relates to the onset of civil war, among other relationships, the results may be dependent on modeling choices and assumptions embedded within the statistical models. Each model, two-stage and SEMs, require some restrictive assumptions. Pre-processing creates a treatment variable, in this case–repression, that is “closer to being independent of the background covariates, which renders any subsequent parametric adjustment either irrelevant or less important” (Ho et al., 2007, 2). In addition to reducing model dependence, I also address the problem of matching using time-series cross-sectional (TSCS) data. Previous studies that use matching and TSCS data have not dealt with potential problems caused by having units that are not independent over time. Since matching assumes independence of units, I propose a solution that matches by year then aggregates each yearly match into one final sample. After outlining this proposed solution to matching using TSCS data, I also match using imputed data and compare these estimates to non-imputed unmatched and matched data. In the conclusion, I discuss the implications of this approach and future directions for research.

2Instrumental variables are often used to deal with the problem of endogeneity. Finding good instruments, however, is exceedingly diﬃcult in the social sciences. Trujillo, Portillo and Vernon (2005) use both propensity score matching and instrumental variables to deal with endogeneity and ﬁnd very similar results.

82 6.1.1 The Goal of Preprocessing

One way to improve predictions is to exclude irrelevant information. Mahoney and Goertz (2004) provide a framework for establishing which negative (no war, no civil violence) cases to include in qualitative studies that often focus on the positive (war, civil violence) outcomes of interest. Their advice is to use negative cases where the outcome of interest is possible. In quantitative political science, to avoid selection bias, most scholars include all potential cases regardless of whether the outcome of interest is possible. Matching is one potential solution in quantitative political science that increases the balance and overlap of the empirical distributions of treatment and control groups (Gelman and Hill, 2007).3 The best way to observe causal effects in any study is to have both a control and a treatment group that are randomly assigned and include randomly selected subjects that differ only due to the application of the treatment. Such an experimental technique ensures no systematic difference between the units in the control versus treatment group; therefore, the change in the dependent variable can be attributed to application of the treatment (D’Agostino, 1998). In sum, to make valid inferences from comparisons between treated and control groups, it is important that the groups have similar empirical distributions or areas of common support. Most of our studies in political science are what Ho et al. (2007) term observational as they lack at least one of the three key aspects of classic experimental research: random selection of subjects from the population, random assignment to treatment, and a large n or sample size. Regression models allow for an approximation of the experiment but often lack data on confounding factors, the treatment effect might be measured with error, and the direction of causation may be unclear. Matching can be accomplished through the use of a propensity score. Propensity scores assign a conditional probability to an observation of assignment to the treatment group given a set of covariates (Rosenbaum and Rubin, 1983; Dehejia and Wahba, 1999). A logit or probit model is estimated to predict this probability, and the propensity score equals this probability. In political science, Ho et al. (2007); Ward and Bakke (2005) have used this method to preprocess their data in an attempt to have less biased estimation of causal

3Other ways of adjusting the sample include using matched samples, stratiﬁcation, or covariance adjustment (D’Agostino, 1998). Propensity score matching is an improvement on these techniques because it includes more information when matching observations.

83 eﬀects. Barabas (2004), another political scientist, uses matching to compare sample means between treatment and control groups on survey responses.

6.1.2 Matching and Reducing Model Dependency

As Ho et al. (2007) note, most quantitative research proceeds by collecting data, specifying a particular model, estimating, re-specifying the model, estimating again, recoding data, offering a different functional form for a variable of interest, then estimating yet again. In the final paper, the author offers a subset of the different estimations that are clearly dependent upon the choices made and the assumptions embedded in the final models displayed. This creates estimates that are highly model dependent (Ho et al., 2007). One solution is to estimate a large number of regressions and to report some range for the parameter estimates (Leamer, 1983; Sala-I-Martin, 1997). This method still requires adherence to a similar modeling choice (such as logit, probit, OLS, etc) while varying the inclusion of some subset of relevant covariates. Since preprocessing of data reduces the link between the treatment effect and the observed covariates, the parametric analyses following this procedure become “irrelevant or less important” (Ho et al., 2007, 2). Achieving balance between the treated and untreated observations, or making sure that they have similar distributions, can be done by ensuring the regions of the distributions that do not overlap are pruned from the data set. This is particularly important when attempting to make inferences about counterfactuals. When these counterfactuals are far from actual values of the data, the conclusions from the model are based on speculation (King and Zeng, 2007). Matching is most helpful when we are able to achieve balance in the propensity score distributions between treated and control groups, thus limiting the effects of other variables in attempting to estimate our privileged causal effect. Achieving balance should reduce the data’s dependence on any particular model or assumptions while selecting cases based on values for the independent variable rather than values of the dependent variable.

6.1.3 Matching and Onset of Civil War

The treatment for this study, graphically displayed in Figures 6.1, is having high repression

(Tr). I expect that a state that is treated with high repression should be more likely to experience civil war than one that is not.

84 Figure 6.1: Matching on the Propensity to be a State with High Repression

Let: Y1=outcome of unit i exposed to treatment and

Y0=outcome of unit i not exposed to treatment

Let Ti be the treatment variable for unit i, where i =1, 2, 3...n. Tiǫ{0, 1} X = the set of pre-treatment covariates

Then, Yi = TiYi ∗ 1+(1 − Ti)Yi ∗ 0.

Verbally, this means that causal effect for the unit exposed to treatment is equal to the difference between the effect of the treated outcome versus the untreated outcome. In this case, again, the treatment relates to high levels of repression. This notation refers to the underlying logic of counterfactuals or that an outcome may be different if one single treatment were different while all other covariates are essentially the same (Morgan and Winship, 2007). Since it is impossible to re-run history or to have perfect control when we use observational data, propensity scores are the way that the analyst can isolate this causal effect. This assumes that all relevant differences between the treatment and control groups is captured by the observables. This is expressed as: (Y0, Y1) ⊥ T |X (Rosenbaum and Rubin, 1983). If this assumption holds, treatment assignment is then independent of outcomes. While this is difficult to fully implement in observational studies, Rosenbaum and Rubin (1983) claim that we can assume that treatment assignment is strongly ignorable.

85 6.1.4 Generating Propensity Scores

Propensity scores provide a single scalar value that provides information about the likelihood that an observation has received treatment. Generating propensity scores is done in many ways. Regardless of how a propensity score is generated, the goal is always the same: achieve balance between the treatment and the control groups. The most easily understandable form of matching is exact matching. In exact matching, treated observations are matched with untreated observations when they have the exact same values on the observed covariates. This can be done with or without replacement. Replacement means after a treatment observation is matched with an untreated observation, the untreated observation is placed back in the pool of potential matches. A single untreated observation can then be matched to several treated observations to ensure that those treated observations are not discarded. Replacement is especially useful when a dataset has far fewer treated than untreated observations. In practice, exact matching can be extremely difficult with a dataset that includes many coavariates. Instead, nearest neighbor matching is commonly instituted to approximate this approach. After randomly ordering treated and control units, treated subjects are matched to untreated subjects with the closest propensity score. Both are then removed (no replacement) and the process continues (D’Agostino, 1998).4 A third technique that builds on nearest neighbor matching uses lowest distance between treated and control units to match. The difference between this so-called caliper matching and nearest neighbor is that the treated and control units must be a specified minimum distance from each other in order to be matched. Otherwise, the units are left unmatched. Other forms of matching, like subclassification and genetic, are also available, and within in each type discussed a myriad of options are available to individualize the procedure. The goal, as always, is to reduce the bias between treatment and control groups. Evaluating which technique is superior must keep this goal in mind.

4This is also referred to as 1-1 nearest neighbor matching. Some other rules such as matching three controls to one treatment (3-1 matching), can also be used.

86 6.1.5 Evaluating Matching Techniques

One standard way to evaluate the outcome of a matched sample is to compare the mean values for the covariates among the treated versus the control groups.5 Comparing these values between the matched and unmatched sample provides information about how successful the procedure was at reducing bias between the groups. Some programs provide a percent bias reduction in each covariate (Ho et al., 2004) or a total percent bias reduction or both (Leuven and Sianesi, 2003). While one convention is to use a t-test to assess balance, this approach is fraught with problems. As Ho et al. (2007) note, a hypothesis test assumes that some theoretic population exists. Since matching is attempting to achieve balance within this sample, t-tests are not very informative. Instead, an outcome that decreases bias between the means of the treatment and control groups is preferable to one that doesn’t but has a t statistic above 2. After matching I use the percent reduction in bias to assess how well the matching improves balance between the treated and control groups. While any reduction in bias improves our causal estimates and reduces model dependency, the goal again is to reduce bias between the groups as much as possible.

6.2 Preprocessing Civil War Data

As discussed above, small differences in variables that have a large causal effect can confound the results for the treatment variable when the subgroup containing the treatment has different average values for this effect. The necessary first step is to evaluate whether large differences in means between the treatment and control groups exist. Recall that the treatment for this study is a state that employs high repression against its population. Operationally, the treatment is equal to one when the Political Terror Scale (PTS) is four or five. Values from one to three are coded zero.6 Table 6.1 displays the difference between the means for the covariates in the sample between the treated and untreated groups. The mean for DissAct and W art−1 are much larger in the treated group than control. GDP is also much smaller. Another striking

5Reducing the variance is also important. 6The mean of the Political Terror Scale for the sample is 2.5. Based on the coding of the PTS, one and two are clearly low repression and four and ﬁve are high repression. I chose to code three as low repression as it is close to the mean in the sample.

87 Table 6.1: Comparing Means in Unmatched Sample

Variable Mean Treated Mean Control |Mean Diff| DissAct 0.850 0.247 0.603 W art−1 0.564 0.083 0.481 GDP 2.288 4.835 2.547 P opulation 9.936 9.001 0.935 Mountains 2.716 1.901 0.814 Noncontig 0.149 0.174 0.025 Oil 0.200 0.156 0.044 New State 0.006 0.008 0.002 Instability 0.254 0.123 0.131 Ethnic F rac 0.408 0.396 0.087 Relig F rac 0.358 0.382 0.023 Democracy -1.781 0.420 2.201 difference is in the level of Democracy. States that use repression thus are more likely to experience dissident activity, have had war last year, be less democratic, and have low GDP. Disentangling the effect then of repression from these other covariates requires increasing balance between the treatment and control groups so that I can validly attribute the causal effect of repression on the likelihood of civil war. Table 6.2 offers the different means for the groups in the matched sample and compares them to the means from the unmatched sample from Table 6.1. In addition to showing the means and the difference in means between the groups, the table displays the percent reduction in bias from the unmatched to the matched sample. The matched sample was processed using a nearest neighbor matching with the psmatch2 program for Stata (Leuven and Sianesi, 2003).7 Observations outside common support for both the treatment and control variables were discarded.8 Of the 3,025 observations, 150 were discarded including 118 from the control group and 32 from the treatment group leaving a sample size of 2,875. As Table 6.2 shows, the reduction in bias between the matched sample and unmatched

7Using a caliper, or predetermined distance, observations are selected if they fall within this distance from the treated observation. The control unit that has the lowest Malahanobis distance is then selected for matching. Units are treated one-by-one and the sample was randomly ordered and shuffled several times. I tried several different caliper values and found 0.06 to lead to the greatest reduction in bias across the covariates in the matched sample. 8As Ho et al. (2007, 17) note, discarding these data can actually improve the efficiency of the estimates. Unless the observer knows the true model, discarding heterogeneous observations can improve estimation.

88 Table 6.2: Comparing Matched and Unmatched Samples

Variable Sample Mean Treated Mean Control Mean Diﬀ % Bias Reduction DissAct Unmatched 0.841 0.245 0.596 Matched 0.778 0.711 0.067 89% W art−1 Unmatched 0.571 0.082 0.488 Matched 0.549 0.529 0.020 96% GDP Unmatched 2.290 4.856 2.566 Matched 2.304 2.265 0.029 99% P opulation Unmatched 9.951 9.001 0.950 Matched 9.940 9.912 0.028 97% Mountains Unmatched 2.712 1.905 0.814 Matched 2.665 2.431 0.234 71% Noncontig Unmatched 0.150 0.174 0.024 Matched 0.156 0.149 0.007 69% Oil Unmatched 0.202 0.153 0.049 Matched 0.198 0.191 0.007 86% New State Unmatched 0.006 0.008 0.002 Matched 0.006 0.007 0.001 83% Instability Unmatched 0.255 0.123 0.132 Matched 0.239 0.244 0.005 97% Ethnic F rac Unmatched 0.486 0.397 0.089 Matched 0.487 0.459 0.028 69% Relig F rac Unmatched 0.361 0.381 0.020 Matched 0.364 0.363 0.001 98% Democracy Unmatched -1.703 0.553 2.256 Matched -1.582 -1.858 0.276 88% sample is quite substantial. All of the covariates achieve better balance in the matched sample and some, such as P opulation, W art−1, and Instability, are nearly perfectly balanced after preprocessing. I tried several other techniques, including exact matching, other types of nearest neighbor, and optimal matching, and none was as successful at reducing bias in the covariates. I also obtained matching estimates using the MatchIt program (Ho et al., 2004). Again, all of the various matching techniques reduce bias and are preferable to the unmatched sample, but the caliper matching was the most eﬀective at reducing bias. Average bias across the covariates for the unmatched sample is over 45%. After matching, the average bias is reduced to 4.6%. After achieving better balance, I then estimate two logit models: one using the matched sample and a second using the unmatched sample. Table 6.3 displays

89 Table 6.3: Civil War–Matched and Unmatched Samples

Variable Coefficient (Std. Err.) Matched Unmatched Matched Unmatched Repress 2.415∗∗ 2.326∗∗ (0.489) (0.373) DissAct 1.053∗∗ 1.026∗∗ (0.198) (0.246) ∗∗ ∗∗ W art−1 -4.041 -2.767 (0.623) (0.491) GDP -0.530∗∗ -0.525∗∗ (0.082) (0.135) P opulation -0.288∗∗ 0.085 (0.087) (0.115) Mountains -0.146† 0.009 (0.080) (0.124) NonContig 1.688∗∗ 0.894† (0.491) (0.547) Oil 1.244∗∗ 0.166 (0.296) (0.430) Democracy 0.112∗∗ 0.029 (0.018) (0.027) New State 4.722∗∗ 3.127∗∗ (0.694) (0.722) Instability 0.152∗∗ 0.113 (0.270) (0.342) Ethnic F rac 0.139 0.347 (0.449) (0.584) Relig F rac 2.649∗ -0.549 (0.600) (0.826) Significance levels : † : p<.10 ∗ : p< 0.05 ∗∗ : p< 0.01 N=2875, 3025 the results of the logit estimations. These results are largely consistent with the unmatched sample, but there are some notable exceptions. P opulation, Relig F rac, and Mountains switch signs and become significant. Four variables increase their coefficients and decrease their standard errors leading to statistically significant findings: NonContig, Oil, Democracy, and Instability. These results do show that the unmatched sample underestimates the causal effects of many of the variables and as stated previously leads to changes in inference. The positive change for repression is a modest 4%, but the coefficient for noncontiguous territory increases by 89%. While the Ethnic F rac coefficient also changes, I do not discuss this change as the confidence interval includes the potential of no effect. After achieving better balance, the inferences for repression should be less affected by other potentially confounding factors. With this matched sample, it is important to consider the improvement in predictive ability of the model. Compared to the Fearon and Laitin (2003) model, the logit model with the matched sample increases the area under the ROC curve from 0.80 to 0.87.9 This is one piece of evidence that suggests the estimation using

9The area under the ROC curve for the 2SCML likelihood model was 0.89.

90 the matched sample and repression and dissent indicators is better at predicting ones and zeroes than the Fearon and Laitin sample.10 One benefit of using matching is that it is now possible to calculate the average treatment effect on the dependent variable. In this case, I can estimate the average effect of repression as a treatment on the probability of civil war onset. The value for the average treatment effect on the treated (ATT) for the sample is 0.062 with a standard error of 0.010. In substantive terms, the effect of being treated with repression increases the likelihood of civil war to about 6.2% when holding all the other variables at their means. Since the balance between the control and treated groups has been improved, we can be confident that confounding factors play less of a role in explaining this treatment than in the unmatched sample. Less than 2% of the observations are coded 1 for a civil war onset with more than 98% coded as 0. A simple estimate of whether a state should experience a civil war in this year would be about 1.8%. Applying the treatment of repression increases that estimate more than threefold. While it is difficult to ever predict war (Gartzke, 1999), knowing whether a state represses or not decreases our uncertainty about the process. While the results from matching are encouraging, one assumption of matching has been violated using TSCS data: independence of units. Within this data, each panel consists of a state from 1976-1999.11 Matching could potentially occur between observations from the same panel that are not independent. In the next section, I propose a solution to this problem. In short, I match year by year to eliminate temporal correlation, then combine the years to produce a final year matched sample.

6.3 Matching Using Time-Series Cross-Sectional Data

A central assumption to most matching models is that the units are independent. In a time-series context, this assumption is clearly violated. For example, an observation for Japan in 1982 is clearly not independent of Japan’s observation for 1981. Most previous attempts at matching are cross-sectional in nature or ignore the problem (Ward and Bakke, 2005).12 In the matching above, I attempted to minimize the problem by sorting the order

10The two-stage approach had a slightly better ROC–0.89. 11Because of states entering or leaving the system and missing data, not all panels have this temporal domain. 12Li, Propert and Rosenbaum (2001) oﬀer a solution to the problem called balanced risk set matching. A unit in the sample who receives the treatment at time t is matched with another unit in the sample at the same time t who has not been treated. Their optimal matching technique to balance the risk sets is computationally intensive, and Lu (2005) oﬀers a similar risk set matching technique that is based on using

91 Table 6.4: Comparing Year Matched and Unmatched Samples

Variable Sample Mean Treated Mean Control Mean Diff % Bias Reduction DissAct Unmatched 0.841 0.245 0.596 Matched 0.711 0.643 0.068 89% W art−1 Unmatched 0.571 0.082 0.488 Matched 0.465 0.461 0.004 99% GDP Unmatched 2.290 4.856 2.566 Matched 2.449 2.375 0.074 97% P opulation Unmatched 9.951 9.001 0.950 Matched 9.863 9.781 0.082 91% Mountains Unmatched 2.712 1.905 0.814 Matched 2.597 2.303 0.293 64% Noncontig Unmatched 0.150 0.174 0.024 Matched 0.141 0.110 0.031 -38% Oil Unmatched 0.202 0.153 0.049 Matched 0.215 0.187 0.028 43% New State Unmatched 0.006 0.008 0.002 Matched 0.004 0.003 0.001 52% Instability Unmatched 0.255 0.123 0.132 Matched 0.227 0.191 0.036 72% Ethnic F rac Unmatched 0.486 0.397 0.089 Matched 0.469 0.453 0.016 82% Relig F rac Unmatched 0.361 0.381 0.020 Matched 0.357 0.368 0.011 43% Democracy Unmatched -1.703 0.553 2.256 Matched -1.835 -2.078 0.243 89% of observations randomly by year so that matching would be most likely to occur within the same year and thus between independent units. One way to avoid violating this assumption is to match by year then to combine each yearly matched data into one final data set. This procedure eliminates the temporal link between any observations.13 Using this process, a new year matched sample is created. a Cox proportional hazards model to predict the likelihood of unit i receiving treatment at time t. The hazard rate of a treated observation is used as the propensity score to match that observation with a similar hazard rate of a unit that has not been treated. Lu (2008) suggested, in personal correspondence with the author, that this approach would not be useful using TSCS data as the treatment could occur multiple times in each panel, and it is not clear what the treatment effect would be after a long period of time. 13I wrote a Stata do-file that matches within each year of the data, dropping other years from the analysis. It then combines each year’s information into a larger data set for subsequent analyses.

92 This year matched sample also greatly reduces overall bias from over 45% to 7%. The balance between treated and control groups is improved by at least 40% from the unmatched sample for all variables except NonContig. Bias actually increases 38% (from -6.2% to 8.5%). Figure 6.2 compares the bias reduction between the diﬀerent matching techniques. For six variables the reduction is quite similar between matching and year matching (DissAct,

W art−1, GDP , P opulation, Mountains, and Relig F rac). In ﬁve cases, the matching procedure is better at reducing bias (Non Contig, Oil, Democracy, New State, and Ethnic F rac). Finally, in one case (Instability), year matching better reduces bias.

100

89 96 99 97 71

80 89 89

60 96 96 % Reduction 99 99 in Bias 40 97 97 20 71 71 0

-20

-40

Oil GDP DissAct War (t-1) Relig Frac Population Mountains NonContig Democracy New State Instability Ethnic Frac Variable

Matched Sample Year Matched Sample

Figure 6.2: Comparison of Reduction in Bias Between Matching Techniques

93 Table 6.5: Civil War–Matched and Year Matched Samples

Variable Coeﬃcient (Std. Err.) Matched Year Matched Matched Year Matched Repress 2.415∗∗ 2.797∗∗ (0.489) (0.723) DissAct 1.053∗∗ 1.333∗∗ (0.198) (0.236) ∗∗ ∗∗ W art−1 -4.041 -4.517 (0.623) (0.852) GDP -0.530∗∗ -0.386∗∗ (0.082) (0.111) P opulation -0.288∗∗ -0.505∗∗ (0.087) (0.131) Mountains -0.146† -0.212∗ (0.080) (0.105) NonContig 1.688∗∗ 3.335∗∗ (0.491) (0.696) Oil 1.244∗∗ -0.034 (0.296) (0.505) Democracy 0.112∗∗ 0.112∗∗ (0.018) (0.025) New State 4.722∗∗ 5.895∗∗ (0.694) (1.137) Instability 0.152∗∗ -1.279∗∗ (0.270) (0.400) Ethnic F rac 0.139 3.916∗∗ (0.449) (0.834) Relig F rac 2.649∗ 0.277 (0.600) (0.764) Signiﬁcance levels : † : p<.10 ∗ : p< 0.05 ∗∗ : p< 0.01 N=2875, 2748

The results from statistical estimation (Table 6.5) using this sample are very similar to Table 6.3 with some notable exceptions. Repression still has a large positive and significant effect on civil war onset. In the year matched sample, the coefficient increases to 2.797 or by 16% as compared to the matched sample and 20% as compared to the unmatched sample. The effect that having noncontiguous territory has on the probability of civil war increases in expected magnitude by almost 98% from 1.688 to 3.335. The effect of being a new state increases after year matching by 25% or from 4.722 to 5.895 as compared to standard matching. When comparing GDP ’s coefficient in the matched and unmatched sample to the year matched sample, it decreases by about 25%. After year matching, the effect of GDP on onset of civil war is reduced further. The negative effects for both P opulation and Mountains are both supported and enhanced in the year matching. Confidence in both estimates increases as does the size of both coefficients. Ethnic F rac and Relig F rac change in magnitude and significance across the different matching approaches. It seems that they are competing to explain the same variation. In the unmatched sample, they correlate at 0.36. The area under the ROC curve is even larger for the year matched sample (0.93) as compared to the matched sample (0.87), the two stage estimation (0.89), and the Fearon

94 and Laitin model (0.80). While all of these matching and non-matching models used non-imputed data, recent evidence suggests combining multiple imputation and matching can further reduce bias in estimation (Hill, 2004; Mattei, 2008). If the assignment of treatment is related to the reasons for covariate missingness, the subsequent analyses may be biased. To estimate models with multiple imputation and matching, I first impute the data and generate five data sets.14 After generating the data, I perform matching on each of the data sets, then estimate models for each matched sample. I then combine the five separate estimates.15 Figure 6.3 offers a comparison of coefficients for estimations for the imputed matched sample and the non-imputed matched sample.16 In most cases, the estimates are similar. The key variable, Repress, is very similar across the two samples. Oil is positive but not significant in the estimations using imputed data. In contrast, Instability is negative and significant using the imputed sample. Ethnic F rac and Relig F rac both swap strength and significance from the imputed to non-imputed samples. These effects are the least consistent across all the different models. 6.4 Conclusion

Using propensity score matching is increasingly popular in political science and other disciplines as a way to isolate the causal impact of particular variables when other factors potentially confound these eﬀects. In International Relations, most large-N studies use time-series data to test hypotheses. As matching increases in use, dealing with the potential problem of matching using data that are temporally correlated is imperative.17 I proposed in this chapter a simple algorithm to deal with the problem of temporally dependent observations. To reiterate, the goal of matching is to sever the link between treatment status and the observed variables (i.e. achieve better balance). The year matching procedure helped solve the problem of non-independent units, but did not achieve better

14Hill (2004) describes two ways that matching with multiple imputation might be done. My approach is consistent with her second approach. Using monte carlo simulations, Hill demonstrates both reduce bias over non-imputation. 15See Honaker and King (2006) for a discussion of this computation. Thanks to Jason Barabas for sharing a spreadsheet that makes these calculations. 16This ﬁgure is generated based on code provided by Kastellec and Leoni (2007). 17Spatial correlation is another potential challenge to valid inference. To the extent that units are not independent across space, assumptions of matching could also be violated. This is another area of potential contribution.

95 ● Repress ●

● DissAct ●

● War_t−1 ●

● Resources ●

● Population ●

● Mountains ●

● Noncontig ●

● ● Oil Non−Imputed Data ●

● ● Democracy Imputed Data ●

● New State ●

● Instability ●

● Ethnic Frac ●

● Religious Frac ●

−5 −4 −3 −2 −1 0 1 2 3 4 5 6

Figure 6.3: Comparison of Coeﬃcients Between Matched Imputed Data and Matched Non- Imputed Data

balance than the simple matching procedure. This suggests a potential tradeoﬀ between achieving better balance and adhering closely to this assumption.18 Since observations are potentially matched to other observations in the same panel using simple matching, I suggest using the year-matching approach even given the potential reduction in balance.

18Since it is unclear which approach outperforms the other in small samples, a logical next step is to set-up a Monte Carlo study to compare the diﬀerent matching techniques using simulated data.

96 CHAPTER 7

INSURGENCY AND COUNTERINSURGENCY: UNDERSTANDING THE MICRODYNAMICS OF STATE-DISSIDENT INTERACTION

7.1 Introduction

Following the conventional combat phase of the war in Iraq in May of 2003, total victory for the United States and its coalition seemed imminent. The Iraqi regime crumbled with little resistance and the prospects for stabilizing the country seemed hopeful. Within a few months, however, it became obvious to the United States and its coalition partners that a viable insurgency had developed in Iraq in opposition to the goals of the Coalition. The US and its partners found themselves locked in a familiar modern asymmetric war scenario. Insurgency and guerilla warfare, contrary to conventional wisdom, are not new types of conflict. Insurgent revolts against the Syrians by Judas Maccabeus around 160 B.C. are chronicled in the Bible’s Book of Daniel (Beckett, 1999). Hannibal, over two thousand years ago, attacked the Roman Empire using tactics of asymmetrical war to bankrupt the much larger and more powerful foe. In recent history, insurgencies have spread throughout Latin America, Asia, and Africa and have been incredibly costly for governments and colonial powers. Although there is renewed interest in insurgency, the classic question about how best to defeat one remains unanswered. Conventional modern militaries, especially the American military, are exceptionally well-prepared for fighting conventional wars. They are less prepared, however, to engage in asymmetric battles of indeterminate length. Why do some counterinsurgency campaigns pacify while others stoke insurgency? While previous studies of counterinsurgent efficacy have focused on particular cases or large comparisons, I investigate the microdynamics of insurgency and look at the interaction

97 of insurgents and counterinsurgents over time in a single state. The question identified above is salient in present day Iraq. One of the primary tactics Americans are using to diffuse the Iraqi insurgency is direct military strikes to reduce the capabilities of insurgents. Using large scale detentions and military operations, the US has attempted to reduce the amount of violence by reducing the supply of insurgents. This assumes that the strength of the opponent is determined by its material capabilities. Using insights from the theory developed in Chapter 3, I argue that we need to consider the interaction between states and dissidents to understand why some counterinsurgency campaigns fail. I argue in this chapter that this strategy has not achieved the desired results in Iraq because the US and its partners have not taken into account that different operations affect populations differently. Building on process models of violence, I show how different levels of support for counterinsurgency help explain why some forms of counterinsurgency actually stoke–rather than mitigate–violence. Where states respond to dissidents with violent repression, they reduce support for the state and encourage violent dissident responses. After outlining how taking into account support affects the efficacy of counterinsurgency, I evaluate the arguments about counterinsurgency using an original data set of US counterinsurgent operations in Iraq. I also offer a brief narrative about the Iraqi case and the ongoing operations. In the conclusion, I offer policy recommendations for the Iraqi case.

7.2 Explaining Modern Insurgency

Counterinsurgency is often pursued as either a battle for hearts and minds or as war of attrition. Hearts and minds approaches to counterinsurgency assume that the causes of insurgency relate to “modernization,” “rising expectations,” or “foreign meddling” (Shafer, 1988). The solution then is to “transform the passive parochial into a least a subject...of a modern state...[by] improv[ing] the state’s capacity to affect the hearts and minds of its citizens” (Shafer, 1988, 62). According to adherents of this approach, governments can attract the minds and hearts of people by providing security and economic development. In contrast, Leites and Wolf Jr (1970) offer a model of counterinsurgency that focuses on cost/benefit analyses as a way to neutralize an insurgency. Using a systems approach, Leites and Wolf argue that the key to ending terror/insurgent outputs is to raise the costs of inputs such as food, recruits and arms through the use of coercion. According to Leites and Wolf Jr

98 (1970, 42) costs and benefits determine citizen action or “limiting damage or enhancing gain may be a sufficient explanation for the behavior of the population, without recourse to more elusive explanations concerning putative preferences or sympathies.” This approach suggests raising the costs of citizen compliance with insurgents and their goals through coercion. At worst, this approach leads to a draining of the insurgent swamp through attrition and harsh coercion. Galula (1964), in contrast to Leites and Wolf, highlights the need for attracting the support of the population. Galula (1964, 7-8) claims that “if the insurgent manages to disassociate the population from the counterinsurgent, to control it physically, to get its active support, he will win the war because, in the final analysis, the exercise of political power depends on the tacit or explicit agreement of the population or, at worst, on its submissiveness.” Building on works such as Galula, I identify societal support as a key component that explains the rise and fall of insurgencies. Leites and Wolf Jr (1970), however, also identify a key point that insurgents must have resources to make violent outputs. Without targeting inputs, creating security is difficult. Finally, hearts and minds approaches that focus on support are consistent with my perspective as well as Galula’s argument. While improving the economy may be helpful in reducing the incentives to rebel, it still can not explain why in some poor states insurgencies flourish while in others they languish. In the next section, I develop a model of support that explains why counterinsurgent approaches sometimes stoke insurgencies. 7.3 Support and Insurgency

I argue that the key to understanding the rise of modern insurgency lies with the support of citizens within a polity. As discussed in Chapter 2, societal support is conceived of as a continuous dimension that represents the distance between the state and individual’s preferences for policies. States that have a high degree of support have citizen and state preferences that are closely aligned. Where citizen preferences, on average, greatly diverge from the state, there is low support for the state’s policies. Many possible levels of support exist between these poles. Thinking about the diﬀerent levels of support in society can help explain why some counterinsurgent campaigns are successful while others fail. I assume that the larger the distance between the state and citizen’s preferences, the more likely bargaining will fail and

99 violence will be used to pursue opposition group’s preferred policies. It follows then that when an individual’s support for the state is low, the likelihood of that citizen using violence to pursue their policy objective increases. Where the average level of support for the state and its policies are low, more people are likely to take up arms against the state to pursue their policy goals. Just as considering how support for the state and its policies can help uncover why counterinsurgency can succeed or fail, thinking about a population’s support for an insurgency can give us insight into the strength of an insurgency. Whereas in earlier sections of this project I assume a general policy space for representing societal support, in this portion, this space is more concise. The state has a policy to counter the insurgents. Citizens have some degree of support for this policy. The closer their preferences are to the state, the more people support the policies implemented by the state. Figure 7.1 represents a proposed situation where most people support the state and its policies. The average level of support is close to the state’s preferred policy and few people have widely divergent preferences.

Number of Citizens

Low High

Support

Figure 7.1: High Support for Counterinsurgency Operations

Figure 7.2 oﬀers a scenario where people have less support for the state’s policies than in the previous example, but few people are at the far end of the policy spectrum. Figure 7.3 provides an example of a situation where few people support the state and its policies. The average level of support is far from the state’s preferred policy. This makes it more likely that these individuals will use violence to pursue their goals.

100 Number of Citizens

Low High

Support

Figure 7.2: Medium Support for Counterinsurgency Operations

Number of Citizens

Low High

Support

Figure 7.3: Low Support for Counterinsurgency Operations

As outlined in Chapter 3, repression pushes individuals further from the state’s preferences. People who are negatively inﬂuenced by house-to-house searches and detentions are pushed further from the state. As the group of people who have extreme preferences increases, the state becomes more likely to experience violent dissent.1 In previous chapters, I assume that the state’s repressive response is uniform. In other words, a unit of repression has essentially the same impact on dissidents regardless of the

1Recall that this assumes that each individual who has extreme preferences is capable of producing a ﬁxed amount of dissident acts. The implication is that the more the dissidents population grows, the more dissident acts the state should expect to experience.

101 form. Part of the reason for this assumption is related to making a more parsimonious theory. Relaxing this assumption, however, allows me to investigate if different responses by the state lead to varying actions by dissidents. In the context of counterinsurgency, states can choose to use a variety of techniques that involve different degrees of militarization. Counterinsurgency operations can be arranged along a continuum based upon the degree of force used by the counterinsurgents (See Figure 7.4). At the low end of the operations spectrum are policing actions, such as patrols, weapons confiscating, and targeted detentions.

Cordon Weapons And Direct Clearing Checkpoints Search Military

Low High Degree of Force

Figure 7.4: A Continuum of Force used in Counterinsurgency

Establishing checkpoints where civilians are questioned, potentially searched, and possibly detained is a degree more forceful than simple patrols or confiscating weapons. Cordon and Search or Knock and Search operations are even more forceful than checkpoints.2 Cordon and Search operations involve sectioning off a particular area of a city or town and not allowing entry or exit.3 Following this sectioning off of a town or area, troops search house to house for insurgents and weapons in an attempt to denigrate the capacity of insurgents to make violence. Finally, the most forceful type of counterinsurgency is a direct military response. Raids, air strikes, and assaults on homes or neighborhoods are the most

2Cordon and Search operations and other more forceful operations are often coupled with checkpoints, patrols, and other less invasive operations. In this continuum, I assume that the higher level operations likely include the lower levels as well. In other words, checkpoints are often established before and during and a Cordon and Search Operation. Thanks to retired Lt. Colonel Glynn Ellis for sharing this information with me. 3For a detailed description of Cordon and Search operations see the Army Field Manual FMI 307.22. Found online: http://www.globalsecurity.org/military/library/policy/army/fm/3-07-22/index.html

102 invasive type of counterinsurgent techniques. These direct assaults have the potential to affect insurgents and civilians alike. Since deadly force is used, the consequences of mistakes can be drastic. With this continuum in mind, I discuss how each approach predicts the degree to which counterinsurgent operations affect insurgent activity. As discussed above, counterinsurgency operations can be arranged along a degree of force continuum. Operations range from Weapons Clearing to Checkpoints to Cordon and Search to Direct Military. This continuum is helpful in understanding the paradoxical nature of counterinsurgency–sometimes pursuing insurgents vigorously creates more insurgents. Counterinsurgency accomplishes two things: it removes dissidents from the set of people willing to use violence against the state, and it imposes costs on people who are not dissidents yet inadvertently receive punishment from the state. The tradeoff between these two results helps determine whether counterinsurgency incites or destroys an insurgency. If an operation removes more dissidents from the set of people willing to use violence than it creates people willing to use violence, then dissident acts should decrease. As acts of the counterinsurgents move up the scale of force, they are more likely to push receivers of this force further from the preferences of the state. The more forceful the action, the greater the impact that this will have on the aggrieved individual. In Chapter 3, I discuss how increases in repression lead to decreases in support for the state. This decline in support pushes more civilians towards dissident activity. Similarly, counterinsurgent operations that use state violence push members of society away from the counterinsurgent’s preferred policies. When these actions are less violent, then dissident activity may be stunted without generating new recruits. At low ends of the continuum, actions by counterinsurgents reduce the ability of dissidents to produce violent acts. Weapons clearing removes weapons from would-be insurgents. Checkpoints also allow for weapons removal while also allowing counterinsurgents the ability to detain potential insurgents. While establishing checkpoints may have negative consequences, like detention of innocent people, delays in traffic, disruption of normal activities, these costs are significantly less than cordon and search and direct military approaches. The level of support in society also influences whether more forceful attacks by counterinsurgents pacify or inflame insurgencies. When support for counterinsurgents is high, imposing costs on the population may have a negligible effect on the number of insurgents.

103

Support after counterinsurgency operation

Support before counterinsurgency operation

Number of Citizens

Low High

Support

Figure 7.5: The Eﬀect that Counterinsurgency has on Insurgent Attacks Given High Support for the Counterinsurgents

Figure 7.5 shows a depiction of such an event.4 If counterinsurgency leads to a reduction of support but creates relatively few new insurgents, then the operation leads to a reduction in the strength of the insurgency.

Hypothesis 7 The greater support within a polity for the state, the less likely counterinsurgent operations will increase insurgent attacks.

The relationship between operations and future insurgent attacks is more complicated when support in society is low. Since more citizen’s preferences are divergent from the state, increasing repression pushes more people even farther away. Repression removes members of the insurgents from the pool of people willing to use violence, but it also creates more people likely to use violence. The draining of the swamp is overcome by the ﬁlling of the pool. Again, this eﬀect is more dramatic as more force is used in counterinsurgency. At the lowest levels, enough inputs may be removed from the insurgents to reduce future attacks while

4This mechanism is similar to the one proposed in Chapter 3. The main diﬀerence is that I allow for diﬀerent distributional forms of support.

104

Support after counterinsurgency operation

Support before counterinsurgency operation

Number of Citizens

Low High

Support

Figure 7.6: The Eﬀect that Counterinsurgency has on Insurgent Attacks Given Low Support for the Counterinsurgents

also avoiding the costly violence that increases the likelihood of more insurgent violence. At the highest degrees of force, when support is low, counterinsurgency creates more insurgents. Based on this discussion, I oﬀer the following hypothesis:

Hypothesis 8 When support within a polity for the state is low, more forceful counterinsurgent operations will increase insurgent attacks while less forceful operations decrease insurgent attacks.

In the next section, I apply this argument about insurgency and counterinsurgency to the current campaign in Iraq. I show how support by different groups leads to different expectations about the potential efficacy of counterinsurgency.

7.4 Iraq and Counterinsurgency

Figure 7.7 represents a stylized set of distributions of the support by diﬀerent groups for the US-led counterinsurgency policies. Since an accurate census has not been done

105 recently, estimates of the population are uncertain. Over 20 million people live in Iraq, the majority are Shia with estimates ranging from 60-70% of the population. Another 15-25% of the population are Kurds and about 10-15% are Sunni Arabs. The stylized depiction of the groups roughly mirrors these demographics. In addition, a recent public opinion poll conducted by the World Public Opinion consortium found that 88% of Sunnis approved of attacks against the US led force, whereas 41% of Shia approve, and only 16% of Kurds.5 This provides some corroboration for the different average levels of support for US led counterinsurgency as depicted in Figure 7.7.6 Another poll conducted by ABC, the BBC, and other media organizations has found consistent support for attacks on the US-led Coalition that hovers around 50%. This poll does not, however, provide a breakdown of responses based on ethnic groups. Regardless, this result is consistent with a majority Shia population who are split on attacks, strong support among one minority group (Kurds), and weak support among another (Sunnis).7 In addition, disaggregating the state’s response across time and using data from a global sample of countries requires a great deal of information that is not available. The interaction between states and dissidents over time within a single country provides an opportunity to investigate the impact that a variety of state responses have on dissident activity. Iraq provides a fertile ground for addressing these issues. Since the fall of the regime in 2003, the US-led Coalition has battled an insurgency that is opposed to the counterinsurgent’s goals and policies. In many contexts, insurgent- counterinsurgent dynamics primarily play-out in the domestic context and do not involve external actors. Iraq includes an externally-supported counterinsurgent campaign that complicates the basic model of attempting to understand these dynamics. One important implication is that support for external actor policies (e. g. US-led Counterinsurgency) should be generally less than for domestic state policies. Nationalism is a strong force for internal cohesion, thus external agents should inherently expect more resistance than domestic forces. With that said, I can still investigate the effects of different

5Details about the poll can be found at: http://www.worldpublicopinion.org. These numbers are akin to negative support or are the inverse of support for the policies of the US-led Coalition. 677% of Sunni approve strongly of attacks on the US and its partners while only 9% of Shia and 8% of Kurds approve strongly. 7When the poll was ﬁrst conducted in 2004, only 17% supported attacks. All subsequent results ﬂuctuate around 50%. These poll results can be found online at: http://www.globalpolicy.org/security/issues/iraq/poll/2008/0308opinion.pdf.

106

Shia

Number of Kurds Citizens Sunni

Low High

Support

Figure 7.7: Stylized Support for Counterinsurgency among Contending Groups in Iraq

state responses on stoking dissident activity. Even though civilians may be more likely to become dissidents against external forces; if my theory is accurate, there still should be variation in dissident responses to state repression across the regions of Iraq. Iraq also provides a way to expand the number of observations by observing counterinsurgency across three distinct regions–the Kurdish North, the Sunni-dominated Center, and the Shia-led South. These regions all experience counterinsurgent operations to varying degrees. They also have varying levels of support for US-led counterinsurgency. While selecting Iraq as the location to test hypotheses derived from a process theory of state-dissident interaction is not a large-cross national study, it still allows for results that are potentially generalizable. Because of the variation over time and the three regions involved, this study is not simply a case study or a study where n=1.8

8As King, Keohane and Verba (1994) claim “[a] unit may be the single nation called ‘India[,]’ [b]ut ‘India’ as a case can provide numerous observations of the relationship between [the independent and dependent variables] if we consider the diﬀerent parts of India.” Considering observations over time also increases the

107 7.4.1 Hypotheses

While I developed some general hypotheses in the argument section, here I outline several speciﬁc hypotheses related to counterinsurgency in Iraq. Since the Kurdish northern areas of Iraq have high degrees of support for both the US-led invasion and the subsequent eﬀorts at countering the insurgency, I expect that this region is one of high support. Recall above, that counterinsurgency operations in polities with high support do not create more insurgents than they remove. Therefore I expect that:

Hypothesis 9 In the Kurdish region of Iraq, counterinsurgency operations reduce insurgent attacks regardless of the degree of militarization

In the southern portion of Iraq which is Shia dominated, I expect that support is lower than in the Kurdish North and greater than in the Sunni Center. As outlined above, Shia support for the US-led invasion and counterinsurgency is moderate and is likely more normally distributed. Given this distribution, expectations for the Southern region of Iraq are similar to the north. The exception is that more forceful forms of counterinsurgent operations may stimulate more violence than in the North.

Hypothesis 10 In the Shia region of Iraq, counterinsurgency operations reduce insurgent attacks except at the higher levels of force

Finally, the central region of Iraq is home to the Sunni populations. Baghdad remains somewhat mixed but the areas surrounding the capital remain Sunni-dominated. These areas lost the most from the US-led invasion and US-led counterinsurgency, and support for US- led counterinsurgency operations are low. Since this is the case, forceful operations should create more insurgents than they remove. Some evidence in support of this claim exists. For example, a CIA assessment of Iraq in November of 2003 suggested that more aggressive counterinsurgency tactics, such as the photos released of bound women and children during a raid in an Iraqi town in the central region, helped swell the ranks of the insurgency.9 At the lower end of the degree of force continuum, however, counterinsurgent operations decrease insurgent attacks. sample size. 9Regan, Tom. “A Daily Weblog of the Post-9/11 World.” Christian Science Monitor. Found online: http://www.csmonitor.com/2003/1112/dailyUpdate.html

108 Hypothesis 11 In the central region of Iraq, counterinsurgency operations using a high degree of force increase insurgent attacks while counterinsurgency operations using a low degree of force decrease insurgent attacks

In the next section, I describe the scope of counterinsurgent operations in Iraq. After providing a description of the these operations, I test these hypotheses using data from Coalition counterinsurgent operations in Iraq.

7.5 Counterinsurgency Operations in Iraq

Since the end of major combat operations in Iraq until the end of September of 2005, there were 145 declared counterinsurgency operations (COIN ops). Over 80% of the operations or 119 took place in the center region of the country which is predominantly Sunni (See Table 7.1 for the number of days that each region experienced an operation). The Southern region only saw 13 (9%) counterinsurgent operations including the assault of Najaf. In the North, the US and it allies launched 16 (11%) COIN ops to deter the insurgents.

Table 7.1: Count of Counterinsurgent Operation Days in the Regions of Iraq

Operations Center South North Total None 531 789 808 2128 Present 349 84 63 496 Total 880 873 871 2624

COIN ops vary in their composition, scope, and duration. They all have similar goals though. The intent is to dismantle the insurgent’s capacity to carry out attacks. In Iraq the eﬃcacy of these COIN ops is likely due in part to the context in which the operations are undertaken. In the next section, I discuss how to design research to test these claims in Iraq.

7.6 Research Design

The US and its coalition partners have faced an insurgency in Iraq since May of 2003. Counterinsurgent operations have been used in an attempt to pacify the insurgents. Since the inception of the insurgency, data on counterinsurgent operations and Coalition deaths have been released by the US Department of Defense. Iraq then provides a unique laboratory

109 in which to test these various hypotheses. The spatial domain of this study is the state of Iraq which is composed of 18 muhafazat or governorates. These political units can be lumped into three regions that roughly reflect the ethnic composition of the area: Central– Sunni, Southern–Shia, and North–Kurdish. Iraq has three main ethnic groups which are geographically concentrated in different areas. The Kurds reside in the north of the country, the Sunnis in the West and Central region of the country and the Shia in the Southeast. These regions are not completely homogeneous. Saddam’s regime, for example, attempted to cleanse the Kurdish population from Kirkuk and moved as many as 200,000 Kurdish people from the city and replaced them with Sunnis from the center of the country. Balad is a Shia city in the heart of mainly Sunni dominated country. Baghdad is also a fairly heterogeneous large urban area with the largest group being Shia. In general, though, dividing the state into these regions is practical for the following reason. Since insurgents attempt to attack and then hide among the population, I expect that most insurgent activity occurs within areas where insurgents are not ethnically different than the population. In Ramadi, I expect the Sunni insurgents are attacking the coalition and in Najaf, the Shia are behind the insurgency. This, of course, is not always the case,10 but in most examples this assumption fits with the distribution of ethnic groups. In the models, I use dummy variables for each region or estimate the model within regions.11 The temporal domain of the study is the time period following the end of major combat operations, or May 1, 2003 to September 30, 2005. The purpose of the model is to explain how counterinsurgent operations affect insurgent activities. Below I discuss the concepts under investigation as well as the measures used to operationalize the concepts.

7.6.1 Dependent Variable

The dependent variable for this study is insurgent success or how successful the insurgency is at creating Coalition casualties. This concept is operationalized as the number of Coalition deaths in a particular region on a given day.12 I chose to use deadly attacks rather than

10Notable exceptions are in the north where Sunnis and Kurds live in close proximity as well as around Baghdad where Shia and Sunnis are also closely clustered. 11Each approach yields similar results. Since the center region has many more observations than the other two regions, this region provides the most information for the study. 12Figures 7.8 through 7.10 demonstrate the daily counts of Coalition deaths in each region. The Y axes for the graphs have diﬀerent ranges. Daily deaths in the Center region occur at higher rates than in the North or South. In the North, one incident was particularly deadly. Otherwise, the South and North experience

110 number of attacks or numbers injured as the insurgent’s goal is to use deadly force.13 Coalition deaths is then a proxy for insurgent success. I assume that the insurgents want to kill coalition soldiers to weaken the resolve of the counterinsurgents as well as increase the audience costs for the stronger party. If the insurgents are unsuccessful in sustaining a ﬂow of casualties for the stronger party, then they can not maintain the conﬂict. Figures 7.8, 7.9, and 7.10 show the number of attacks over the temporal domain of the study for each region. The central region has experienced the most casualties while the North and South have been more sporadic. 20 15 10 Coalition Deaths 5 0

01apr2003 01oct2003 01apr2004 01oct2004 01apr2005 01oct2005 Date

Figure 7.8: Daily Count of Coalition Deaths in the Central Region of Iraq

7.6.2 Independent Variables

To explain the cause of Coalition deaths I include the following covariates: the type of counterinsurgent operation, the eﬀect of the type of operation over time, and previous counts of Coalition deaths. similar daily rates. 13One other consideration is the reliability of the data. Number of attacks or numbers wounded are much less reliable than numbers killed. The number killed are reported by the Department of Defense through press releases and are compiled by http://icasualties.org/oif/

111 iue71:DiyCuto olto etsi h otenRgo fIraq of Region Southern the in Deaths Coalition of Count Daily 7.10: Figure iue79 al on fCaiinDah nteNrhr eino Iraq of Region Northern the in Deaths Coalition of Count Daily 7.9: Figure

Coalition Deaths Coalition Deaths

0 1 2 3 4 5 01apr2003 0 5 10 01 Apr03 01oct2003 01 Oct03 01apr2004 01 Apr04 112 Date Date 01oct2004 01 Oct04 01apr2005 01 Apr05 01oct2005 01 Oct05 Type of COIN Op

As discussed earlier, I conceive of COIN ops along a continuum from low to high coercion. In practice, I code four type of operations that range from low to high degree of force. The four types of operations in Iraq are “Weapons Clearing,” “Checkpoints,” “Cordon and Search,” and “Direct Military.” First, Weapons Clearing operations, or when counterinsurgents clear roads of IEDs, dig up weapons caches or dismantle bombs in public spaces, are used to dismantle insurgent capacity. These operations affect few to no members of the population and are the least militarized or coercive. Second, Checkpoints are used to block roads, limit access in or out of cities, and to search vehicles for potential arms and explosives. Checkpoints involve a greater degree of force than Weapons Clearing but still affect the population less than other forms of COIN ops. Third, Cordon and Search operations consist of blocking off sections of a town or city and conducting systematic searches of civilian homes. These operations require a greater military presence and are more intrusive than Weapons Clearing or Checkpoints. Finally, Direct Military operations consist of raids, destroying safe houses, engaging insurgents, and direct assaults on cities and towns. Direct Military operations require the highest level of force and intrusion on the population. I include a series of variables that are coded 1 when these type of operations occur and 0 when they are absent.14 This information is coded from Department of Defense press releases and summaries from globalsecurity.org.15

Eﬀects of COIN Ops

I expect that where counterinsurgent operations occur the likelihood of casualties is higher. This should happen for three reasons. First, having troops in the field exposes them to possible insurgent attacks. Convoys, supply routes, bases, and many other areas are also targets, but it is highly plausible that participating in a counterinsurgent operation increases the likelihood of a casualty. Second, a selection effect occurs where counterinsurgent 14In practice this variable ranges from 0 to 3 as on some days, there are multiple direct military operations. Recoding the variable so that it is 0/1 leads to similar estimates. I leave the variable as 0 to 3. 15A list of so-called Iraqi Pacification Operations can be found at http://www.globalsecurity.org/military/ops/iraq ongoing mil ops.htm. I code the type of operations from descriptions from this list and supplemented this information with DoD press releases. I also had two undergraduate students code the types of operations. The intercoder agreement exceeds 86%, Pearson chi square and Fisher exact tests confirm that this agreement is not by chance across the four categories and within each category. When the coders disagreed, after consulting the original text describing the operation, I made a final coding decision.

113 operations happen where there is a high degree of insurgent activity. These operations, in other words, take place in the more dangerous parts of the country. Third, these operations likely anger the population and push some members of the population to violence. After these operations take place, their effects can also be estimated. As discussed above, these missions are regularly disclosed by the US Department of Defense and collected online by globalsecurity.org. I use the verbal descriptions of the operations to discern beginning and ending dates for the operations.16 I began a counter variable that begins the day after an operation ends and counts down from 30 to 1.17 Each decay function is assigned to a specific type of operation. Direct Military operations have decay functions that are grouped together in a single variable. Some days have multiple operations of the same type occurring at the same time. In this case, I sum all the decay functions for these operations. Again, I assume that the effects of these operations on decreasing insurgent activity is additive. In other words, an additional operation affects insurgent activity at the same rate as the previous operation. In all there are four counter variables for the four type of operations that count down the days following an operation. If the counter variables are negatively associated with deaths, then the operation reduces insurgent activities.18

Controls

On average, in the central region, 1.3 Coalition soldiers are killed per day. This amounts to roughly four soldiers every 3 days. That rate is much lower in the other two regions (See Tables 7.2-7.4 for summary statistics). In the central region, 92% of all days had between 0 and 3 deaths. In less than 1% of the days in the sample are deaths greater than 10.19 I also include a time counter that begins at the end of major combat operations, May 1st, 2003, and ends on the last day of the sample, September 30th, 2005.

16Several operations lacking end dates are not included in the analysis, this lead to potential problems if something nonrandom is causing the lack of the DOD specifying an end date for the operation. 17This assumes that the eﬀect of the operation decays in an additive way. I attempted to log the decay and found similar results providing initial support for this additive decay. 18For example, as the counter decreases from 30 to 1 or as the eﬀect of the operation diminishes, deaths increasing from 1 to 7 leads to a negative association 19To control for these outliers, I include dummy variables for the day that these deaths occur. The days in the central region include a Chinook helicopter downed in Fallujah and other days during the Fallujah 2004 operation. A downing of a helicopter in Mosul on November 15th, 2004 was also an extremely deadly incident claiming the lives of 15 Coalition soldiers.

114 Table 7.2: Summary Statistics for Central Region

Variable Mean Std. Dev. Min. Max. Deaths 1.311 2.136 0 19 Direct Military 0.384 0.660 0 3 Cordon and Search 0.065 0.266 0 2 Checkpoints 0.044 0.206 0 1 Weapons Clearing 0.083 0.276 0 1 Direct Military Effect 29.614 28.522 0 134 Cordon and Search Effect 13.396 17.189 0 94 Checkpoint Effect 4.351 10.298 0 63 Weapons Clearing Effect 6.526 12.037 0 61 N=855

Table 7.3: Summary Statistics for South Region

Variable Mean Std. Dev. Min. Max. Deaths 0.143 0.560 0 5 Direct Military 0.098 0.331 0 2 Cordon and Search 0.011 0.103 0 1 Checkpoints 0.001 0.034 0 1 Weapons Clearing – – – – Direct Military Effect 5.483 11.066 0 55 Cordon and Search Effect 2.193 6.312 0 30 Checkpoint Effect 0.548 3.296 0 30 Weapons Clearing Effect – – – – N=848

7.7 Estimation

Since the data are time-series, I am faced with the decision of whether to treat the error structure as a nuisance and correct these non-normal errors or to explicitly model the dynamics. A third option is to assume a white noise process for the residuals (Mitchell and Moore, 2002). I explore three options: an autoregressive moving average (ARMA) approach, an event-count approach, and an autoregressive event-count approach. Since there are serious threats to inference using both the ARMA and event count approaches Brandt, Williams and Fordham (2000), I discuss these estimations in the appendix and concentrate

115 Table 7.4: Summary Statistics for North Region

Variable Mean Std. Dev. Min. Max. Deaths 0.245 0.722 0 11 Direct Military 0.069 0.253 0 1 Cordon and Search 0.007 0.084 0 1 Checkpoints – – – – Weapons Clearing – – – – Direct Military Effect 6.058 11.924 0 66 Cordon and Search Effect 2.748 7.033 0 30 Checkpoints Effect – – – – Weapons Clearing Effect – – – – N=846 the discussion here on the autoregressive event-count approach The dependent variable is a count of Coalition deaths in a given region-day. The standard approach consists of estimating one of the familiar count models–a poisson or negative binomial regression model (NBRM). The problem for this study, however, is that the data are time-series. One solution is to use lagged counts of the dependent variable to correct for possible autocorrelation. Brandt, Williams and Fordham (2000) show that using a NBRM with lagged dependent variables suffers from two problems. First, these models imply that the growth rate of the process “is the exponentiated coefficient on the lagged dependent variable.” So using these lags may model a time series process, “but not data that are dynamic” (825). This means that this type of lagged NBRM is only appropriate for time- series with exponential growth and no dynamics. I use the autocorrelation function (ACF) plot as a diagnostic to aid in model specification. I find that the regions display different dynamics. Since this data, in the central region, is dynamic but lacks exponential growth, a Poisson Exponentially Weighted Moving Average Model (PEWMA) or a Poisson Autoregressive Model (PAR) may be more appropriate.20 Since this data is mean reverting the PAR is a better candidate.21 The distribution of the dependent variable is also different depending upon which region is estimated. In the North and South, I do not estimate PAR models and instead use NBRM models without the lags

20See Brandt and Williams (2001) and Brandt, Williams and Fordham (2000) for discussions of these models. 21The ACF plot shows a weak but signiﬁcant short-memoried process.

116 as these regions show little dynamics or growth.

7.8 Results and Discussion

PAR model estimates for the Central region are reported in Table 7.5. These results are fairly consistent across different modeling specifications (see Appendix). Since the PAR models an explicit time-series process, interpretation of the coefficients is different than in the Poisson or NBR models. Brandt and Williams (2001) explain that the correct interpretation of the PAR requires evaluation of both the parameter estimate as well as the estimated values of the ρ parameter. The implication of this is that NBRM and poisson coefficients will be biased if the data generating process is a dynamic, temporal count (Mitchell and Moore, 2002). After evaluating different specfications, I chose the PAR (3), or the Par with three autroregressive terms, because this specification limited the Akaike Information Criteria (AIC)22 relative to the PAR (1) and (2).23 The type of operation has both a significant and large substantive effect on the count of Coalition deaths. Three of the four operation variables are statistically significant and all have the predicted sign. The more invasive operations lead to more Coalition deaths while the less coercive operations reduce the expected Coalition deaths in a given region-day. The presence of a Direct Military operation increases the expected Coalition death count by 43% while the presence of a Checkpoint or Weapons Clearing operation decrease the expected count by 49% and 26% respectively. Cordon and Search operations also increase the expected count but by less than direct military, and this result is not statistically significant. The effect of the operations over time is a bit less clear. The effect of Direct Military operations is close to statistical significance (t-score of -1.57), suggesting that the effect over time of the operations is to dampen Coalition deaths. An interesting result occurs when looking at the effect over time of the Direct Military operation as well as the effect during the operation. Having a Direct Military operation increases expected Coalition death counts 49%. Each day that an operation occurs, the expected death counts increase. The day after the operation, the expected counts of Coalition deaths decreases but by much less than the

22See the Appendix for a detailed discussion of the AIC and BIC. 23I chose the PAR (3) without the Time variable over the PAR (2) with Time and the PAR (3) with Time as this specification minimized the AIC. The coefficients were all similar as well as their level of statistical significance. Due to multicollinearity, I drop the Spike variables as the model would not converge when these three were included.

117 Table 7.5: Poisson Auto Regressive Model–Central Region

Variable Coefficient (Std. Err.) % ∆ in Counts Direct Military 0.404∗∗ (0.063) 43% Cordon and Search 0.025 (0.244) Checkpoints -1.057∗∗ (0.398) -49% Weapons Clearing -0.436† (0.257) -26% Direct Military Effect -0.0036 (0.002) Cordon and Search Effect 0.0037 (0.292) Checkpoints Effect 0.0003 (0.007) Weapons Clearing Effect -0.0126∗ (0.006) -27% Rho (1) 0.086† (0.048) Rho (2) 0.128∗∗ (0.042) Rho (3) 0.073† (0.043) Intercept 0.447∗∗ (0.108) N = 855 AIC = 2647 Significance levels : † : 10% ∗ : 5% ∗∗ : 1% increase caused by the operation. In other words, participating in the operation may negate any benefits of the operation over time. The effect of Weapons Clearing operations clearly dampens expected counts of Coalition deaths. During the Weapons Clearing operations, expected counts decrease by 26%. The day after the operation concludes, expected counts decrease by 27% and each day there is a changing level of decrease in expected counts. Based on these results, the least militarized operation has the largest effect on dampening Coalition deaths after conclusion of the operation. The effects of Cordon and Search operations seem to increase the expected count of Coalition deaths, but this result misses standard levels of statistical significance. The effect of Checkpoints is very close to zero and is far from statistical significance. These results provide some support for Hypothesis 8: More forceful operations, Direct Military, increase the number of Coalition deaths, and less forceful operations, Weapons Clearing, decrease the number of Coalition deaths. While the results are not conclusive for Cordon and Search operations, they do suggest that more force in a region of low support generates further Coalition casualties. The results for the other two regions (see Table 7.6) are inconclusive. Most of the covariates are insignificant. Time is significant and positive

118 Table 7.6: Negative Binomial Models

Coalition Deaths Center Region South Region North Region IRR Std. Er. IRR Std. Er. IRR Std. Er Direct Military 1.315∗∗ 0.967 0.880 0.424 1.039 0.382 Cordon and Search 0.930 0.164 1.000 0.000 0.634 0.753 Checkpoints 0.494∗∗ 0.134 1.000 0.010 – – Weapons Clearing 0.886 0.168 – – – – Direct Military Effect 1.000 0.00177 1.004 0.012 1.017∗ 0.008 Cordon and Search Effect 1.003 0.00281 0.986 0.0233 0.997 0.132 119 Checkpoints Effect 0.998 0.00534 1.041 0.0466 – – Weapons Clearing Effect 0.993† 0.004 – – – – Spike1 17.648∗∗ 17.514 – – – – Spike2 21.046∗∗ 21.001 – – – – Spike3 24.879∗∗ 24.646 – – – – Spike4 – – – – 33.18∗ 49.66 Lag Deaths 1.063∗∗ 0.0278 – – – – Lag 2 Deaths 1.060∗∗ 0.0220 – – – – Time 1.001∗∗ 0.0002 1.003∗∗ .0006 1.001∗∗ 0.0004 alpha 0.910 0.099 7.668 1.745 2.12 0.484 LRTEST X2=285 LRTEST X2=120 LRTEST X2=57 Obs=853 PsR2=0.05 Obs=848 PsR2=0.04 Obs=846 PsR2=0.03 Significance levels : † : 10% ∗ : 5% ∗∗ : 1% suggesting that casualties may be trending upward, but the other relationships are generally not distinguishable from zero. One finding that is suggestive is that the effect of Direct Military operations in the North decreases Coalition casualties. This is consistent with Hypothesis 9 that operations, regardless of the degree of force, will decrease subsequent Coalition deaths. The lack of findings for the southern region could be due to lack of data as deaths in this region are more sporadic.

7.9 Conclusions

If support is an important component of assessing the ability of an insurgency to survive and grow, then we can expect that different strategies for dealing with insurgencies will depend on the support that populations have for government policies. In areas in Iraq such as Al An-Bar province where support is low, militarized operations and detentions have no effect and may fuel the flames of the insurgency. In the Southern region of Iraq where support is moderate, COIN ops may deter would-be insurgents as well as impose costs on present insurgents. Finding the optimal balance is likely important to both eliminating threats as well as avoiding the creation of new insurgents. In areas such as Ramadi or Fallujah, the US military pursued policies consistent with a highly coercive approach without deterring insurgents or squelching the supply of new recruits. Having a constant military presence without the more invasive military raids in the region is a recent policy that seems to be more effective than attempts to punish insurgents and collaborators.24 Checkpoints and Weapons Clearing may both reduce the effect of the insurgency as well as keep Coalition troops from deadly situations. In this chapter, I apply the argument developed from Chapter 3 to Iraq. The inferences that I draw, however, are likely generalizable to other cases where an insurgent group faces a more powerful authority. If these results and model are accurate, whenever an authority attempts to pacify an insurgency, they need to take into account support in determining which strategies to employ. These results are, of course, tentative and are one sample of a much larger phenomenon. Investigating the microdynamics of insurgencies should be a goal of conflict scholars as we build towards a better understanding of how these processes develop and evolve. More quantitative studies of the effects of different types of COIN ops in different

24Kirk Semple, ”U.S. Forces Try New Approach: Raid and Dig In.” New York Times, 5 December 2005.

120 contexts, such as Afghanistan, Colombia, or in other locations where counterinsurgency campaigns occur, would provide additional evidence to conﬁrm or inﬁrm the results from this study. 7.10 Appendix

Some alternative modeling approaches that I attempt include estimating autoregressive moving average models (ARMA) and count models. I first estimate the parameters of an ARMA model. This model takes into account lagged values of the dependent variable as well as current and lagged values of the stochastic error component. These models tend to be good representations of stationary variables using relatively few parameters. The results of an Augmented Dickey Fuller (ADF) and Phillips Perron Test reject the null hypothesis that a unit root is present. Therefore, using an ARIMA approach is not necessary. An ARMA (p, q) model includes p lagged dependent variables and q lagged error terms. A time-series model was specified for each region of Iraq–North, Central, and South. The results of the ARMA models for the three regions are displayed in Table 7.7. Because of the different groups opposing the Coalition forces in each region, each model has a different process generating the dependent variable. To identify the process generating the death counts in each series, I use a combination of the plots of the autocorrelation function (ACF) and the partial autocorrelation function (PACF). Visual inspection of these plots suggests that in the central region, there is a need for modeling at least the first few lags of the dependent variable as well as several lags of the error term. To identify the exact number, I used both the Akaike Information Criterion (AIC) test and the Schwartz Bayesian Information Criteria (BIC or SBC). These tests are goodness-of-fit measures that measure the function of the variance of the model residuals. The general formulas for each measure are: AIC = T ln (sum of squared residuals) + 2n BIC = T ln (sum of squared residuals) + n ln(T) where n = number of parameters (p + q) t = number of usable observations (Enders, 2004, 69) As the model fit improves, the AIC and BIC should be smaller. These measures also account for the number of parameters involved in a model as adding parameters can be one way to improve the overall fit of a model in standard OLS. I add different numbers of AR

121 and MA lags until I was able to minimize both the BIC and AIC. In addition, when models had similar AIC or BIC scores, I chose the specification that had significant values for the AR and MA coefficients. For the central region, the best fitting model includes three lags of the dependent variable (Lag Deaths, Lag 2 Deaths, and Lag 3 Deaths). In the Northern Region, the BIC and AIC identified a AR1 model without any MA terms as the most parsimonious and best fitting. Finally, the Southern region is also best modeled with one lag of the dependent variable. Including or excluding other AR or MA lags to the specification for each times series either added insignificant parameters to the model or increased both the BIC and AIC. After estimating the best fitting models, I evaluate the residuals. Since the data is event count with a large number of 0’s and 1’s, the residuals are non-normal. The consequence of having non-normal errors could be bias, inefficiency in the estimates or both. Since the error component is supposed to be the random or nonsystematic component of a regression model, violations of the assumption can lead to incorrect inferences. As Brandt, Williams and Fordham (2000) argue using an ARMA model with event count for yt can be flawed for several reasons. First, unless the event counts are very large, the distribution is unlikely to be normal. Second, using differencing to induce normality relies on the assumption that these events are independent of each other. In this data as in most event counts, this is an untenable assumption. Visual inspection of the residuals as well as tests of skewness and kurtosis confirm that the errors are non-normally distributed. King (1988) observed this bias in OLS estimates for event count data and argued for the use of Poisson models for event count data in political science. Since the underlying distribution of the data is a count, non-negative and non-integer values are impossible. Using the Poisson model may provide more unbiased and consistent estimates even in large samples (King, 1988). The sole parameter identifying the Poisson Regression Model (PRM) is the mean (µ) count of the distribution. Larger values for µ, shifts the mean of the distribution to the right and makes the distribution more normal (Cameron and Trivedi, 1986; Long and Freese, 2006). PRM assume independence of events; that is, each event at time t is not dependent on that same event at t + 1. In this data, the mean of the coalition deaths across the three regions is less than the variance, a condition referred to as overdispersion. Since the data are overdispersed, a negative binomial regression model (NBRM) is an alternative. The NBRM does not assume independence or equidispersion

122 2006 2000 1500 1000 Frequency

412 500

167 100 28 12 17 12 6 10 3 1 3 4 3 1 1 1 0 0 5 10 15 20 Coalition Deaths

Figure 7.11: Frequency Distribution of Coalition Deaths

(µ=σ2). The Poission and NBRM have similar coefficients, but the Poisson tends to bias standard errors downward (Long and Freese, 2006). The α parameter from the NBRM can be used to test which model is preferable. Since the dispersion parameter is positive and significant, there is evidence of overdispersion. A likelihood ratio test can then be computed which confirms that there is overdispersion in all of the regions. In this case, the NBRM is the appropriate model to use. Another consequence of overdispersion may be excess zeroes. Based on a histogram of values for the dependent variable (See Figure 7.11), it is evident that the data is not normally distributed and that there are many zeroes. The concern, though, is whether these zeroes are affecting the estimation of the NBRM. If so, then using a zero-inflated negative binomial regression model (ZINB) is more appropriate. A Vuong test which provides a test of which model is a better fit for the data–the NBRM or the ZINB–favors the use of the NBRM. Substance and theory should also guide this choice as there is a possibility that a Coalition soldier in Iraq can be attacked on any given day.25 Frequent patrols, traveling on supply lines, and fixed known locations of bases make it unlikely that any soldier’s probability of

25A zero-inﬂated model would be more appropriate here if I assume that there were two types of zeroes on a given day. One zero is no death that occurred with probablility one, and the other is a no death that occurs with some probability less than one.

123 being attacked is always 0. Participating in operations likely increases a soldier’s probability, but so does driving kitchen equipment through Baghdad or making a trip to the airport for a soldier leaving the theater of conflict. Since the NBRM is the appropriate model to use for count data, I interpret the results from Table 7.6. I report the results from Table 7.7 also to show that across these different specifications, the results for the parameter estimates are consistent.26 The first set of key independent variables, relate to the presence of different types of operations each region- day. The base category is ‘no operation’ so all of the coefficients for the variables can be interpreted relative to a day without a COIN OP. The presence of a Direct Military operation in the central region increases the expected count of coalition deaths by over 31%. In short, the presence of the most forceful form of COIN Ops increases the expected daily count of coalition deaths. I can not draw an inference form the coefficient for the Cordon and Search operations. The two less intrusive forms of COIN Ops have a dampening effect on Coalition deaths relative to days without operations. The Checkpoints coefficient is highly significant while the Weapons Clearing coefficient is not. The presence of a Checkpoint operation reduces the expected count of coalition deaths by over 50% in a central region day. In the other two regions, none of the coefficients are close to statistical significance. Several factors may be at work. First, these regions face far fewer number of operation days, leading to fewer number of data points to analyze (See Table 7.1). Second, the insurgency in the center is different than the ones in the south and north. Given different actors, interests, and support, these actors may respond differently to the same operations. More data might help sort out which factor explains the lack of significant findings. For three of the types of operations, the effect that that the operations have over time or the coefficients for the time after the operation occurs suggest that they dampen coalition deaths and that effect decreases over time. Only the coefficient for the Weapons Clearing operations is significant. The coefficient, or the incident rate ratio, is essentially one for the other three types of operations meaning that the variable has no effect on the dependent variable. In addition, this estimate is not statistically significant. A one unit increase in the Weapons Clearing variable leads to an expected decrease in the expected deaths of about

26Changing the modeling strategy adjusts the standard errors. In a few cases, however, the parameter estimates are also different among the different approaches to modeling this data. These differences are discussed in the next section.

124 0.1%. In practical terms, this means that the day after a Weapons Clearing Operation ends, the expected count of coalition deaths should decrease by 3%. After then 10 days, this effect should be 2%. These effects hold only for the Center region as both the North and South have insignificant coefficients.27 Besides the Time and Spike variables, the only variable in the model for the North region that achieves statistical significance is Direct Military Effects. After a direct military operation, expected Coalition counts actually increase by 60%. After ten days, this effect dampens so that deaths increase by 40%. The lag counts for the Center Region offer some insight into the process of violence. A coalition death at t − 1 increases the expected count for t by a little over 6%. In addition, a coalition death at t − 2 increases the expected counts for t by 6%. Lagged deaths had an indeterminate effect in the other two regions. Since lagged deaths have little effect and are not significant in the South or North, they were dropped from the analysis. Time also has a significant coefficient suggesting that counts are increasing, albeit very slowly, over the length of the sample period. Dummy variables for the spikes in Coalition Deaths also improves the fit of the model and leads to large significant coefficients.28

27When I lump all three regions in a single model, the results are consistent with the central region. See Table 7.8. 28The three Spike variables decrease the AIC.

125 Table 7.7: ARMA Models

Coalition Deaths Center Region South Region North Region Coef. Std. Er. Coef. Std. Er. Coef. Std. Er Direct Military 0.557** 0.124 -0.031 0.106 0.007 0.010 Cordon and Search 0.025 0.339 -0.241 5.414 -0.050 0.500 Checkpoints -0.849 0.768 -0.054 67.16 – – Weapons Clearing -0.299 0.379 – – – – Direct Military Effect 0.001 0.004 0.000 0.002 0.004† 0.002 Cordon and Search Effect 0.004 0.006 -0.004 0.005 -0.001 0.004 Checkpoint Effect -0.007 0.013 0.003 0.007 – – 126 Weapons Clearing Effect -0.012 0.009 – – – – Spike1 15.109 14.594 – – – – Spike2 16.202 15.843 – – – – Spike3 18.524 12.552 – – – – Spike4 – – – – 10.653 10.096 Time 0.0009† 0.0006 0.0004** 0.0001 0.0003* 0.0001 Intercept 0.745* 0.371 -0.369* 0.160 -0.390 0.287 Lag Deaths 0.162** 0.0278 0.050† 0.0297 0.0824** 0.0215 Lag 2 Deaths 0.152** 0.0256 – – – – Lag 3 Deaths 0.063** 0.0253 – – – – /sigma 1.77** 0.0295 0.550** 0.009 0.613** 0.010 Obs=855 X2=247 Obs=848 X2=20.87 Obs=846 X2=25

Signiﬁcance levels : † : 10% ∗ : 5% ∗∗ : 1% Table 7.8: Negative Binomial–All Regions

Variable Coefficient (Std. Err.) Direct Military 1.110 (0.120) Cordon and Search 0.986 (0.207) Checkpoints 0.606 (0.306) Weapons Clearing 0.928 (0.258) Direct Military Effect 0.999 (0.002) Cordon and Search Effect 1.008∗ (0.003) Weapons Clearing Effect 1.007 (0.004) Checkpoints Effect 1.011† (0.006) Other Operations 0.884 (0.205) Lag Deaths 1.100∗∗ (0.028) Lag 2 Deaths 1.041 (0.025) Lag 3 Deaths 1.072∗∗ (0.023) North 1.682∗∗ (0.217) Center 5.762∗∗ (0.798) lnalpha 0.427∗∗ (0.087)

N 2615 Log-likelihood -2201.387 2 χ(14) 565.584 Signiﬁcance levels : † : 10% ∗ : 5% ∗∗ : 1%

127 CHAPTER 8

INTERNATIONAL RELATIONS AND CIVIL WAR

8.1 Introduction

Recent arguments related to civil war onset pay particular attention to transnational factors. Refugees, regional factors, and bad neighbors have all been cited as reasons domestic models of civil war onset are insufficient in explaining the origin of this form of political violence (Murdoch and Sandler, 2002; Salehyan and Gleditsch, 2006; Gleditsch, 2007). As most scholars of conflict processes attest, political violence does not respect borders. While the turn to transnational factors is a positive step in understanding civil war onset, the impact that international relations has on civil war has been understudied. Does the nature of the system, alliances, militarized interstate disputes, and other interstate behavior affect the likelihood of civil war for a particular state? In this chapter, I offer an answer to this question by applying some insights from both international relations theory and a process theory of civil war. In developing a theory of civil war onset, I assume the interaction of states and dissidents that ultimately produces civil war is not influenced by external actors. In the sections that follow, I relax this assumption and investigate how external actors and systemic factors influence the process of violence within states. Different theories of international relations predict that conflict behavior between states is more or less likely depending on the configuration of power within the international system (Waltz, 1979; Organski and Kugler, 1980). Less is known, however, about how the international system affects domestic conflict.1 This chapter identifies several system-wide and interstate factors that affect the likelihood of generating civil war. I offer four ways that

1A notable exception is Thyne (2006) who argues that signals from third parties aﬀect the bargaining between states and dissidents and thus the potential for civil war onset.

128 international relations affects the likelihood of civil war. First, changes in system-wide power concentration lead to increases in uncertainty over the distribution of power in the system. This uncertainty increases the likelihood that dissidents challenge the state and spark the process that leads to civil war. Second, states that attempt to build capacity also attempt to eliminate internal and external challenges to their authority. Therefore, I expect where states are involved in international conflicts or militarized interstate disputes they are likely to also be involved in civil wars. Third, states who are more closely aligned with the great powers are able to supplement their domestic capacity and thus stave off domestic contention that leads to civil war. Fourth, states that build alliances–specifically defensive alliances–are less likely to generate civil war. As states build capacity through defensive agreements with other states they are better able to avoid the onset of civil war. After hypothesizing how international relations might encourage civil war, I empirically evaluate these claims using a large cross-national database. The many variants of realism generally presuppose that the state is the unit of analysis when attempting to understand national level conflict behavior. Recent scholarship, however, has emphasized the utility of employing the leader(s) of the state as the unit of analysis (Bueno de Mesquita et al., 1999; Goemans, 2000; Bueno de Mesquita et al., 2003; Chiozza and Goemans, 2003). To understand the interplay between international relations and civil war, I utilize insights from the process theory of civil war and the relationship between international factors and civil war while also placing emphasis on the impact of state leaders as key actors. In the next section, I identify the key dimensions of state capacity and show how states can maximize each dimension in multiple ways. Most importantly, support from resource- rich states can increase the state capacity of otherwise weak states. After identifying the dimensions of state capacity, I demonstrate that system characteristics and alliances can supplement an otherwise weak state’s capacity. Next, I show how interstate conflict behavior that attempts to build the state or what Tilly (1985) calls state-making contributes to domestic conflict. I then test these arguments using a time-series cross-sectional design that also controls for domestic explanations for civil war. The results for the usual suspects of control variables for civil war are consistent with previous models and the hypotheses for international relations’ impact on civil war onset are supported. In the conclusion of the chapter, I suggest some ways to enrich the study of civil war utilizing insights from

129 international relations.

8.1.1 The Two Dimensions of State Capacity

As outlined in Chapter 2, state capacity can be conceptualized as a two-dimensional phenomenon.2 A state’s capacity is determined by both the resources that it possesses and by the support it receives from society. Each of these dimensions contributes to capacity and many diﬀerent indicators can measure these dimensions. Figure 8.1 oﬀers a three level explanation for the concept of state capacity. Resources and societal support constitute the concept of state capacity. For a state to have the highest levels of strength, it must possess both support and resources. Neither dimension can fully compensate for the other. At the indicator level, however, several measures of each dimension are potentially substitutable for each other. Measuring resources and support can be done in multiple ways. The OR operator between the indicator level items suggests that a state can acquire resources through tax revenue, foreign assistance, or through primary resource extraction.

A state with adequate resources can have one or more of these indicator level characteristics.3 Domestic resources are most often considered when attempting to measure a state’s resources. International factors, however, can also substitute at the indicator level. For example, states that lack domestic resources can join an alliance or seek aid from a third party. Either approach can supplement an otherwise resource-poor state. Societal support can be measured as the degree of protection of private property or by the amount of state violence against the citizens. These two indicators are substitutes for the societal support dimension.4 In the next section, I discuss how these dimensions of state capacity coupled with interstate relations aﬀects the likelihood of civil war. 8.2 State Capacity and International Relations

As I outlined in Chapter 3, civil war is a process of violence between states and dissidents that is aﬀected by the strength of the state, the mobilization of dissent, and the choice of 2I deﬁne state capacity as the ability of the individuals comprising the state to enact policy outputs consistent with their preferences. 3Goertz (2006) argues that substitutability should be most common at the indicator level and that is a common method to involve unique characteristics, such as culture, into a causal explanation. 4I discuss the choice of indicators in greater detail in Chapter 4.

130

Figure 8.1: Three Level Concept of State Capacity

the state to use violence to counter dissident mobilization. Additionally, job insecurity for leaders and societal support indirectly affect the onset of civil war by increasing repression and dissident mobilization. This process is solely discussed endogenous to the actors and state without regard for external factors. This chapter brings in the external actors and systemic factors that influence state-dissident interaction and the onset of civil war. Like Putnam (1988) and Starr (1994) I assume that state leaders are concerned with “domestic audiences and consequences as well as external audiences and consequences” Starr (1994, 485). One area where bringing in international actors matters relates to the resources available to state leaders. Put otherwise, the resource dimension of state strength is affected by more than just Gross Domestic Product. Several indicators are potentially substitutable for this dimension (Most and Starr, 1984). When relaxing the assumption that state capacity is

131 confined to tax revenues or natural resources, new insights emerge. More specifically, states that are weak due to lack of national income or natural resources can satisfy the resource dimension of state capacity by seeking foreign assistance. Egypt, for example, has received well over $38 billion dollars of military aid from the United States since making peace with Israel.5 This aid increases the Egyptian state’s capacity to implement its policies better than a simple measure, like GDP, might predict. Another externally-focused way that a state might increase its capacity is to form an alliance.6 The reasons that states form alliances are to increase security or to increase policy goals, and they may face tradeoffs between these goals depending on the power differential of alliance members (Morrow, 1991). Fearon (1997) claims that formal military alliances provide credible signals to other actors that intervention by the contracted parties is likely. Leeds (2003) finds evidence that the type of alliance commitment matters, and more importantly for this discussion, mutual defensive pacts reduce the likelihood of militarized disputes initiated against the countries with this type of alliance. In the context of civil war, defense pacts matter for two reasons. First, states that receive alliance assurances from outside powers can focus resources on domestic sources of conflict rather than on external enemies. A substitution effect occurs where resources devoted to military endeavors abroad can instead be turned inward. Second, because states with defensive alliances have a credible commitment from another state to intervene in defense against external enemies, this assurance could translate to defense against internal aggressors as well. While there are no large cross-national studies of third-party intervention into domestic unrest or domestic collective violence short of civil war, anecdotal evidence from recent history suggests that states with defensive alliances are able to increase the state’s ability to contain insurgencies and thus reduce the likelihood of building to civil war.7 For example, in the Philippines, the insurgent group Abu Sayyaf has attempted to spark a larger conflict, but US aid and training has allowed the Philippines to be more effective in countering this threat. Both mechanisms identified above suggest that having a

5This figure is from the Federation of American Scientists website and estimates the amount received from the US from 1978 to 2000. Available online at: http://www.fas.org/asmp/profiles/egypt.htm 6I follow Leeds et al. (2002, 238) and define alliances as “written agreements, signed by official representatives of at least two independent states, that include promises to a aid a partner in the event of military conflict, to remain neutral in the vent of conflict, to refrain from military conflict with another, or to consult/cooperate in the event of international crises that create a potential for military conflict.” 7For a study of state intervention into civil wars see Regan (2000).

132 defensive alliance should reduce the likelihood that internal challengers to the state will be able to initiate civil war. Alliances aggregate capabilities and depending on their content can have offensive or defensive implications. The state offering the alliance and/or assistance receives some benefit in return related to their own foreign policy goals. During the Cold War, the United States provided financial assistance to many countries throughout the world to counter communist insurgencies which in turn strengthened the state. In Chile and Guatemala and a host of other countries throughout Latin America and the world, US aid helped keep the state leaders in power. In a cross-national study of foreign aid allocation, Alesina and Dollar (2000) find that political alliances are a major determinant of which countries receive foreign aid. While Morrow (1991) and others are generally concerned with alliances that counter external threats, the weaker states in an asymmetric alliance are likely just as concerned with internal threats. Extending the reasons for weak states to enter into these asymmetric alliances based on fears about internal dissent is a logical extension of Morrow’s work. If we assume that leaders value survival above all else, then a loss in what Morrow terms autonomy is worthwhile in relation to the gains in security for the leaders of the state. Traditionally weak states tend to gain the most from from aligning themselves with more powerful states. These asymmetric defensive alliance agreements should increase the ability of the state to militarily defeat its internal enemies and stave off would-be challengers from rebelling against state authority. This, in turn, keeps the current leaders in power. The more resources that the stronger state can provide the weaker the state, the more effective the alliance will be at strengthening the weaker state. This discussion leads to the following two hypotheses:

Hypothesis 12 States with defensive alliances are less likely to experience civil war onset

Hypothesis 13 States that are more closely aligned with the Great Powers are less likely to experience civil war onset

In the next section, I discuss how external conflict behavior or interstate conflict is related to the likelihood of internal conflict. More specifically, I argue that states involved in war- making are also more likely to be in the process of state-making.

133 8.3 External Conﬂict and Civil War Onset

As Tilly (1985) and Cohen, Brown and Organski (1989) claim, the state attempts to consolidate its power by ridding itself of internal and external rivals. States do this to increase their ability to extract revenue from the population and monopolize the market for this revenue (Tilly, 1985). According to Tilly (1985, 181), “a state that successfully eradicates its internal rivals strengthens its ability to extract resources, to wage war, and to protect its chief supporters.” States that face internal challenges to authority are still in the process of what Tilly (1985) calls state-making or the process of eliminating domestic rivals to the monopolization of violence. In the European context of state-building, state making was often coupled with war-making, or the attempt to neutralize external challenges to state authority. Tilly’s work in this area primarily relates to the European experience. More recent work in other areas, such as Sub-Saharan Africa or Central America, focus on how interstate rivalries affect the state’s ability to extract (Thies, 2006, 2007) and largely support Tilly’s claims derived from Europe. Jackson (1990) claims that one important difference between the European and current developing world experience in state-building is that European states could and often did fail. In the current era, Jackson argues norms of sovereignty allow frail, illegitimate states to subsist and refers to these entities as quasi-states. Ayoob (1995) points the blame towards the shorter time-horizons on modern developing states. Where European states went through the state-building process over centuries, many developing countries are expected to build a capable state in decades or less. The pressure to eliminate both internal and external rivals should be even greater in today’s international system even though the means to do so are potentially more limited. In sum, the relationship between internal and external conflict in the European experience was positive. State leaders perform both of these tasks for the intermediate goal of increasing state revenue with the final goal of ensuring political survival. Developing states face similar pressures that are potentially even greater. Based on this discussion, we should expect the following hypothesis:

Hypothesis 14 States currently experiencing external conflict (militarized interstate disputes) are more likely to experience internal conflict (civil war) than states that do not have external conflict.

134 Beyond alliances and external conflict, the nature of the system might also influence domestic conflict. In the next section, I discuss how changes in the concentration of power affect the likelihood of internal conflict or more precisely–civil war.

8.4 The International System and Civil War Onset

During the Cold War, states turned to the great powers, the US and Soviet Union, to help with internal challenges. From Latin America to Asia, both the US and Soviet Union gave money to governments to increase these weaker state’s power vis-a-vis their internal rivals. At times, each great power gave to different sides of a conflict, like in Angola, while at other times, one Great Power overwhelming supported a partisan without much intervention from the other great power, like in Chile. Previous work demonstrates that intervention on each side of an existing civil war tends to increase the duration of the conflict but that partisan intervention tends to reduce the length of the conflict (Balch-Lindsay and Enterline, 2000). The bipolar international system remained relatively stable for about 40 years and thus system power concentration remained stable as well. This international system created a great deal of certainty among actors about who held power in the system, their levels of capabilities, and thus the likelihood of challenging them militarily. With the fall of the Soviet Empire, the concentration of power in the system dramatically shifted. Germany reunified, China’s power increased, and Russia’s power was greatly diminished.8 First, system power concentration shifted from over 25% in 1988 before the fall of the Berlin Wall to about 22% in 1992 after the dissolution of the Soviet Union. The lack of a strong eastern bloc reduced the concentration of the power and led to large changes in system-wide power in 1992 through 1994. With the lack of super power contention in the developing world came a reduction in external military support to each side in civil wars. Instead, intervention became less likely as the rationale behind intervention (fear of communism) no longer existed. Because of this, states in the post-Cold War era expect less assistance from great powers when there is not great power competition. More importantly, changes in system power concentration lead to uncertainty among the actors in the system. Power transition theorists such as Organski and Kugler (1980) claim that high levels of system concentration should decrease the likelihood of war. As the

8I assume here that power is synonymous with the share of system capabilities measured by the Correlates of War index of national material capabilities.

135 disparity between the great power or hegemon in the system and its challengers increase, the likelihood of war decreases. Parity or near-parity creates uncertainty over which side would prevail in a war. In contrast, defensive realists like Waltz (1979) suggest that high levels of system concentration increase the likelihood of war. Having states or blocs of states at parity creates a status quo situation which should lead to fewer wars in the system. Regardless of the absolute level of power capabilities, both power transition theorists and realists believe that changes in the status quo can lead to violent conflict. Morton and Starr (2001), more in line with power transition theories, find empirically that increases in uncertainty caused by changes in system power concentration increase the likelihood of interstate war. Uncertainty over who would prevail in an interstate conflict may also translate to the domestic sphere. Changes in system power concentration may lead to greater uncertainty over how state’s respond to domestic conflict. As noted above, uncertainty over aid or involvement from major powers may increase the likelihood that dissidents challenge states that lack material capabilities. When states can appeal with greater certainty to system leaders for help, dissidents should be less likely to challenge the state. Based on this logic, increases in uncertainty or changes in system power concentration are expected to increase the likelihood of civil war.

Hypothesis 15 Changes in the concentration of power in the international system increase the likelihood of domestic conﬂict

The insights derived from international relations theory and the process theory of civil war provide testable hypotheses amenable to empirical investigation. In the next section, I discuss how to design tests for the above hypotheses. 8.5 Research Design

The temporal domain of the study is 1975 to 1999. This temporal period is post-World War II, and much of the analysis is deliberately conﬁned to the post-war period. While the years from 1945-1975 are not included because of missing data issues, I do not expect that adding these data would change the inferences drawn from the study.9 One advantage of this

9One robustness check I could use is to impute the repression data for the 1945-1975 period, then estimate models including the entire post-war period. Although imputing data has become common among statisticians and increasingly political scientists (King et al., 2001; Honaker and King, 2006), imputing an entire cross-section of data is not common.

136 temporal domain is that it includes both the time before and after the fall of the Soviet Union and its satellites. The period from 1989-1991 can be thought of in time series language as an intervention or major change in system structure and power. Having data that includes this period as well as the time pre and post intervention allows me to treat it as a natural experiment. In other words, some exogenous shock occurs that is similar to the application of a treatment in a controlled experiment. For example, the institution of metal detectors in airports in 1973 served as a treatment when trying to understand skyjackings (Enders and Sandler, 1993). When metal detectors were instituted skyjackings dropped dramatically across the globe. The only difference then before this exogenous treatment was the lack of metal detectors. The time period that followed only differed due to the application of this treatment. Similarly, the fall of the Soviet Union is an exogenous shock that allows for a great deal of control in comparing the periods before and after the fall.10 As discussed in Chapter 4, the spatial domain includes 162 countries representing nearly all countries in the world. Some of the smaller countries are excluded. This may affect our inferences about the effects of key independent variables as we may expect that the effects that that international system and changes in the system effect smaller states even greater than the larger ones that can insulate themselves more from international processes. If this is true, then the effect of excluding these cases may have the effect of dampening or causal effects or may be smaller than if we included small states that are influenced by changes in the system. Four concepts are discussed in the above theory that need operationalization before testing: the concentration of power in the international system, foreign policy similarity for a state with stronger states, defensive alliances and external conflicts. The concentration of power in the system was first operationalized by Singer, Bremer and Stuckey (1972, 26-27). They measure the concentration of power in the international system by focusing on the top five to eight major powers, then they “compute the standard deviation of the observed percentage shares. [Next, they] divide that figure by the maximum standard deviation of the percentage shares; that maximum would occur if one nation held 100 percent of the

10Similar to a lab experiment, there are still problems related to drawing causal inferences from natural experiments. We can not be completely sure that this intervention is what is causing the diﬀerence in outcomes in the diﬀerent periods under study. Since we can not re-run history, we still face the fundamental problem of causal inference (Morgan and Winship, 2007) or that we can not observe the counterfactual world where the Soviet Union does not fall or metal detectors are not instituted.

137 shares, and the others had none at all.” The final number is between 0 and 1 and at the poles represents a perfect equitable distribution of power among the great powers (0) to all of the power being concentrated in one of the great powers (1). Changes in this measure can be measured by taking the first difference or simply by subtracting concentration at t from concentration at t − 1.11 Measuring foreign policy similarity between a state and the great powers can be done in several ways. Most measures assume that foreign policy similarity can be operationalized by observing alliance commitments (Altfield and de Mesquita, 1979; Signorino and Ritter, 1999). These alliance commitments are the set of commitments that a state has to the great powers in a given year. Commitments can vary by degree ranging from no alliance to a mutual defense pact.12 Bueno de Mesquita (1975) argues that states with similar alliance portfolios share similar security interests and advocates the use of a measure of association called τB. Signorino and Ritter (1999) advocates the use of a spatial measure, or S, of alliance similarity rather than an associational measure. I use both indicators to proxy this concept. Results are similar for both measures. In the estimation results, I report estimations that include S as the measure of a state’s foreign policy similarity to the system leader. To measure defensive alliances, I use a measure from the Alliance Treaty Obligations and Provisions (ATOP) data set (Leeds et al., 1999). This indicator is coded as one if country A has an alliance commitment from another country B in event of an attack. While the ATOP data include many forms of alliances, I am only concerned with defensive alliances as these are the form that could increase domestic capacity for the state. The same basket of control variables from the Fearon and Laitin (2003) data are included: W art−1, GDP , P opulation, Mountains, NonContig, Oil, Democracy, New State, Instability,Ethnic F rac, Relig F rac.13 In addition, I include the repression and dissent variables from the process model of civil war (See Chapter 4).

11In the original formulation of this measure t was subtracted from t − 5 to reﬂect the slower movement in systemic power shifts. Looking for this movement in a yearly time frame is a relatively conservative approach and should bias against ﬁnding a relationship between civil war onset and system capabilities shifts. 12In between these poles, a state can have an entente or a neutrality/nonaggression pact with a great power. 13For a full description of these variables see Fearon and Laitin (2003).

138 Table 8.1: Civil War Onset and International Relations

Variable Coeﬃcient (Std. Err.) Defense -0.870† (0.467) F oreign P olicy -2.176† (1.299) MIDs 0.377∗∗ (0.144) System Movement 21.019∗∗ (6.104) ∗∗ W art−1 -3.669 (0.829) GDP -0.433∗∗ (0.152) P opulation -0.248 (0.168) Mountains 0.189 (0.174) NonContig 1.609∗ (0.719) Oil -0.412 (0.584) Democracy 0.061 (0.041) New State 2.460 (1.143) Instability 0.432 (0.423) Ethnic F rac 0.580 (0.849) Relig F rac -1.709† (1.027) DissAct 0.485 (0.838) Repress 1.776∗∗ (0.520) Dissresiduals 0.380 (0.894) † Repressresiduals -1.083 (0.566) Intercept -5.531∗∗ (1.824)

N 2478 Log-likelihood -133.844 χ2 116.870 Signiﬁcance levels : † : 10% ∗ : 5% ∗∗ : 1%

8.5.1 Estimation

Since the dependent variable is dichotomous (civil war onset, no civil war onset), I use a logit model to estimate the eﬀects of the independent variables on the likelihood of civil war. In addition, the previous residuals from the repression and dissent equations are used as regressors to deal with endogeneity concerns (Alvarez and Glasgow, 1999). I also estimated separate models using splines and years at peace variables (Beck, Katz and Tucker, 1998) or t, t2, and t3 (Carter and Signorino, 2006). The data did not exhibit time dependence and thus these approaches are subsequently dropped from the analyses.

139 8.6 Results

Table 8.1 displays the results of the logit model using controls from Fearon and Laitin, repression and dissent, residuals from the repression and dissent equations, and the key international relations independent variables. The independent variables of interest Defense, F oreign P olicy, MIDs, and System Movement are all at the top of the table. First, Hypothesis 12 or states that have defensive alliances will be less likely to experience civil war receives some support. The sign for the measure of defensive alliance is negative, suggesting support for the claim. Since the result just misses the conventional 5% level (p< 0.062), the result is suggestive but not quite definitive. Similarly Hypothesis 13 or that states aligned with great powers will be less likely to experience civil war onset receives marginal support. The coefficient for F oreign P olicy is both negative and close to significant (p < 0.094) suggesting that the more similar a state’s alliance portfolio is to the great powers in the system, the less likely they are to experience civil war. Hypothesis 14 or the assertion that states who have external conflict are more likely to have internal conflict is also supported. Since the coefficient is both positive and significant, I can infer that as the number of militarized interstate disputes that a state initiates increases, the more likely the state is to experience internal conflict in the form of civil war onset. Predatory theories of the state make similar claims about the relationship between state making and state building, and these findings support this general line of research. Finally, Hypothesis 15 claims that movement in the concentration of power in the international system should increase the the probability of civil war onset. Based on the statistical evidence, this claim is supported.14 While the above hypotheses are generally supported after the empirical tests, it is difficult to draw substantive conclusions about the variables from a logit model. Since interpreting the coefficients from the models can be complicated, I discuss the effect that the variables have on the probability of civil war using simulations.15 The first substantive effects relate to the relationship between defensive alliances and the

14I also estimated models using the Upsalla/PRIO codings for civil war. Using these data, I find all of the above relationships between international relations to be consistent or even stronger (based on the size of the coefficient as well as significance) except for the relationship between defensive alliances and the Upsalla/PRIO coding of civil war onset. This relationship is indeterminant from zero. 15To simulate these quantities of interest, I use a simulation procedure similar to Clarify (King, Tomz and Wittenberg, 2000).

140 0.04

0.035

0.03

0.025

0.02 Pr(Civil War) 0.015

0.01

0.005

0 Alliance Absent Alliance Present

Figure 8.2: The Eﬀect of a Defensive Alliance on the Probability of Civil War Onset

onset of civil war (See Figure 8.2). Defensive alliances reduce the likelihood of civil war. A defensive alliance can reduce the likelihood of civil war from 1.5% to just above 0.5% (See Figure 8.2).16 States with the lowest levels of GDP can reduce their likelihood of civil war by entering into defensive alliance pacts.17 At levels of GDP above roughly $6,000, having a defensive alliance has no impact on the expected probability of civil war onset.18 Figures 8.3, and 8.4 also provide some substantive results for the independent variables of interest. In Figures 8.3 and 8.4, I show the effects of different levels of foreign policy similarity with the system leader and how this affects the probability of civil war onset. Similar to being involved in a defensive alliance, having high foreign policy similarities with system leaders reduces the likelihood of civil war onset. Also similar, as GDP rises to above

16These quantities are found using King, Tomz and Wittenberg (2000), holding all variables at their means. I also held GDP at a lower level (25th percentile). 17The expected probability of civil war is less for states with defensive alliances as compared to those without these pacts. The error bars in Figure 8.2, however, overlap suggesting uncertainty over this result. 18GDP’s effect on the likelihood of civil war creates what Berry, Esarey and DeMeritt (2008) call a compression effect. That is, since GDP has a large impact on civil war especially as high values push the probability close to zero, variables that also influence this relationship have little impact at high levels of GDP.

141 ne hnFrinPlc iiaiyis Similarity War Policy Zero Civil Foreign of when Likelihood The Onset 8.3: Figure h ubro Iscagsfo eot ih,tepoaiiyo ii a incr war civil of probability the Whe eight, war. points. to civil percentage zero of 4 probability from in nearly changes change MIDs the of on impact number largest the the have intersta state chang militarized a of a number by given ho the war initiated in I civil Changes of simulations, interest. probability Using of in variable change independent onset. the the calculate war and civil mean of their at probability variables the affect interest whe of situation variables a creates stability System GDP. of unlikely. levels very expect low is the at war increase only shifts but power war sample. system civil the the in of graphs, amount above maximum the the to by fluctuating again is Similar it data. when the versus in stable cases is concentration the of 1.5% summa about the only (see event in rare Uncertaint occurs very it a mean. 4); is their onset Chapter at war in held a Civil are by high. variables initiated fairly other MID’s is all estimates of and these number 8 the to civ when 0 of 3% from probability over changes year to expected 1% the than shows, less graph from the increases As disputes. interstate militarized away. fade effects these $6,000

19 Pr(Civil War) al . ffr oesbtnieeet httk noacuthwcagsin changes how account into take that effects substantive some offers 8.2 Table power system when onset war civil of likelihood the show I 8.7, and 8.6 Figures In o numbers increasing over war civil for probabilities predicted some offers 8.5 Figure h 5 ofiec nevli arywd,bti ee nldsze includes never it but wide, fairly is interval confidence 95% The 0 .05 .1 .15 0 2 Gross DomesticProduct 4 6 19 (1000sofDollars) ic ohGPadRpeso aefil ag ffcson effects large fairly have Repression and GDP both Since 8 10 142 ne hnFrinPlc iiaiyis Similarity War Policy One Civil Foreign of when Likelihood The Onset 8.4: Figure

Pr(Civil War) 0 .05 .1 .15 0 o hsi refralo h results the of all for true is This ro. 2 Gross DomesticProduct 4 6 (1000sofDollars) dlikelihood ed ystatistics ry edisputes te tt na in state around y 8 ae by eases ecivil re lwar, il dall ld in e the 10 n f te ffcsaefil agnladices sIuealwrlvlo D ( GDP of level lower a use I as per increase of 15 and effect nearly marginal by the war fairly percentile, civil are of 25th effects probability its Other the at increases eight placed to is zero GDP from MIDs If are variables marginal. these when fairly variables are independent mean in changes war, civil of probability the l ftepoaiiisaeaddo utatdfo hsbsln.The t baseline. at this are from points. variables subtracted percentage all or the when added in are model changes probabilities the the in of war all civil of probability baseline for ones the except i cu.Teae ne h O ie h ecnaeo ae orcl pr correctly cases of percentage the gives o ROC an where the war under civil no area diffe predicting w The given is civil error errors predicting occur. II is II Type did error, or Type negative I or false Type A I or positive occur. Type not false make a may case, this model In the cutpoints. how shows that graphs are r estv oweeteaaytcosst sinthe assign g to they as chooses ROC analyst than the useful where less to are sensitive correc approaches are percent These on measure. focus similar model other some of or zeroes. accuracy and predictive ones in-sample predicts the model evaluate a well to how of assessment an provide curves ROC 20 n a ocmaemdl st s eevrOeaigCaatrsi RC C (ROC) Characteristic Operating Receiver use to is models compare to way One o oecmlt ecito fRCcre e hpe 4. Chapter see curves ROC of description complete more a For iue85 h ffc fMD ntePoaiiyo ii a Onset War Civil of Probability the on MIDs of Effect The 8.5: Figure rinPolicy P oreign F Pr(Civil War) 0 .05 .1 .15 0 h pe on nldszr.Teefiue r hne rmthe from changes are figures These zero. includes bound upper The . 2 # ofMilitarizedInterstateDisputes 4 143 6 success eaentpretcags hyare they changes, percent not are se erma–.% hsmasthat means This mean–0.2%. heir 8 upit ncnrs,ROCs contrast, In cutpoint. 10 rweeoedid one where ar etg points. centage e al 8.2). Table See l predicted tly stactually nset dce and edicted hnein change a te ways Other re . urves enerally tthe at rent 20 sStable War Concentrationis Civil Power of System Likelihood when The Onset 8.6: Figure rdcieacrc r vrhdwdb ohrpeso n D hc te which onset. improveme GDP war marginal and civil The repression of predictors both time. best by the overshadowed of are 91% accuracy about wo predictive other zeroes In and 0.91. ones is predicts area disse this correctly repression, relations, include including international which Finally, for 3 variables 0.89. Chapter four is in the curve Fearo estimated and the models standard under the area the In the For dissent, 0.80. and is 2001). repression ROC al., the et under area (Beck the fit model, model of estimate an provides l aibe tma,PoaiiispeitduigCaiy(2000) Clarify using predicted Probabilities mean, at variables All MIDs rinPolicy P oreign F ytmMovement System Defense Pr(Civil War) 0 .05 .1 .15 0 aibeChange Variable 2 Gross DomesticProduct 4 0 0 0 0 6 → → → → al .:Poaiiyo ii War Civil of Probability 8.2: Table .%14.9% 3.9% 8 06 -2.2% -0.6% 1 a .%3.2% 0.9% Max 04 -1.1% -0.4% 1 8 ∆ rC) D mean GDP Pr(CW), 10 144 hfs(Maximum) War ConcentrationShifts Civil Power of System Likelihood when The Onset 8.7: Figure

Pr(Civil War) 0 .05 .1 .15 0 2 ∆ RC) D 25% GDP PR(CW), Gross DomesticProduct 4 6 d,ti model this rds, dt ethe be to nd n Laitin and n 8 t in nts 10 nt 8.7 Conclusions

If Glenn Palmer is correct that civil war and the study of why states bring about civil war onset is the young person’s international relations,21 then using international relations to explain civil war onset is a natural progression. While I have highlighted the effects of alliances, foreign policy similarity, militarized interstate disputes, and system wide power concentration, there are still many questions left unanswered. First, does the changing nature of the international system affect domestic conflict processes? To explore this question, I estimate models utilizing data from 1945-1999 (see Table 8.3).22 The only one of the international relations variables that drastically changes is F oreign P olicy suggesting that the results for this variable are time-sensitive and that changes in the system may indeed matter for how foreign policy similarity affects the likelihood of civil war onset. Second, with both MIDs and alliances, I focus on their effects on state capacity and thus restrict my emphasis to types of MIDs or alliances that would impact a state’s capacity. Of course, there are different types of MIDs and different types of alliances that might influence the likelihood of domestic conflict. Theories that directly relate to these relationships could further unpack the relationship between civil war and international relations.

8.8 Appendix

Scholars such as Gleditsch (2007) argue that transnational factors, such as contagion, explain why states in some regions are more prone to civil war than states in other regions. If contagion is both explaining weak states and onset of civil war, then the inferences we draw from this study could be spurious. I estimated a series of models using diﬀerent coding procedures for contagion of civil war. The measure that I use to proxy this contagion is a spatial lag of civil wars in neighboring countries.23 The following procedure is used to construct the spatial civil war lag: A NT xNT block- diagonal weight matrix records the distance between all states within 1000 km.24 All bilateral

21Palmer made these remarks during his presidential address at the 2007 meeting of the Peace Science Society in Columbia, South Carolina. 22I could not use the repression/dissent data as they do not go back to this period. Therefore, these models should be treated as exploratory rather than as conﬁrming. 23Thanks to Mike Findley for sharing these data. 24The weight matrices are from Gleditsch and Ward (2000).

145 distances greater than 1000 km are recorded as zero. The effect of civil war in the states in the surrounding region are aggregated by multiplying the weight matrices and a temporally lagged, NT x1 measure of civil war in these respective states. The result is a spatial lag of civil war at t − 1 in surrounding states that captures the potential diffusion of civil wars from nearby states. Several different specifications can be used to test the robustness of the variable. Specifications that are based on direct contiguity or some predetermined distance, such as 200 km or 600 km, can also be used. Results from models using different spatial lags are extremely similar to the original models. The civil war diffusion variables are all insignificant except in Model 3 and have little effect on the key variables of interest. In addition, other specifications including changing the distances had little effect. Most surprising, however, is the negative sign in front of the spatial lags. This suggests that civil wars in neighboring countries last year reduce the likelihood of civil war in this country today. Part of the reason for this negative sign is likely due to collinearity. The spatial lag, not surprisingly, is correlated with GDP, repression, dissident activity, population and other determinants of civil war. While a simple bivariate positive correlation exists between the spatial lag and onset of civil war including other factors causes this relationship to disappear.

146 Table 8.3: Sample from 1945 to 1999

Variable Coeﬃcient (Std. Err.) Defense -0.627∗ (0.282) F oreign P olicy 0.776 (0.552) MIDs 0.395∗∗ (0.089) System Movement 16.475∗∗ (4.182) ∗ W art−1 -0.806 (0.363) GDP -0.293∗∗ (0.076) P opulation 0.082 (0.093) Mountains 0.238∗ (0.107) NonContig 0.339 (0.325) Oil 0.640∗ (0.319) New State 0.659 (0.649) Instability 0.682∗ (0.269) Democracy 0.001 (0.020) Ethnic F rac 0.538 (0.480) Relig F rac -0.016 (0.613) Intercept -5.724∗∗ (0.957)

N 5097 Log-likelihood -350.592 χ2 97.310

147 Table 8.4: International Relations Variables and Spatial Lags

Model 1 Model 2 Model 3 Variable Coef. (Std. Err.) Coef. (Std. Err.) Coef. (Std. Err.)

148 Defense -1.163 (0.506) -1.237 (0.517) -1.299 (0.516) F oreign P olicy -2.031 (1.355) -2.139 (1.358) -2.152 (1.341) MIDs 0.428 (0.163) 0.441 (0.165) 0.455 (0.163) System Movement 22.874 (6.369) 23.637 (6.425) 23.946 (6.415) SpatialLag (200 km) -0.261 (0.191) – – – – SpatialLag (600 km) – – -0.257 (0.164) – – SpatialLag Contiguity – – – – -0.475 (0.216) CHAPTER 9

CONCLUSION

My purpose in this dissertation is to provide a theory for the process of violence that leads to civil war. Development of this theory helps to resolve some academic puzzles relating to why civil war begins, how external actors affect the process, and how microlevel dynamics can stimulate or mitigate the process. Whereas the conventional wisdom argues that states are susceptible to civil war based on their resources, I offer an explanation for why resource-poor states can avoid civil war and why some states with adequate resources develop civil war. Furthermore, the dominant approach in the study of civil war is to specify a single equation and estimate direct relationships between the covariates and onset of civil war. I argue and then empirically demonstrate that a model which incorporates state and dissident actions as mediating variables between structural conditions and civil war onset is better at in-sample predictions of civil war. Because I use a variety of methodological techniques, the inferences that I draw are less dependent upon the assumptions of any particular approach. Whereas many of the predictors from recent civil war studies are structural variables that are either time-invariant or nearly time-invariant (e.g. mountainous terrain and GDP per capita), repression varies greatly between countries as well as over time within the same country. Because of this variability across time and space, levels of repression help predict in which country-year civil war is most likely to occur. This finding is encouraging from a policy standpoint because repression is a tool that states can control whereas mountains and national income or more difficult to manipulate. Knowing which states are repressing also helps analysts predict when civil war may occur. States that are resource-poor are all at risk, but the ones who in any given year are also repressing their populations are most at risk. My work suggests further exploration into the causes of state repression in the tradition of Davenport (1995); Davenport and Armstrong

149 (2004); Gurr (1986, 1988); Moore (1998, 2000). If repression helps predict civil war, then understanding why states repress leads to policy suggestions for reducing state repression and, in turn, the probability of civil war. Even after creating treatment and control groups of data and ensuring that the other variables were balanced between these groups, the effect of repression on the likelihood of civil war is strong and positive. This leads to the inference that states who were treated with repression were more than three times as likely to experience civil war than similar states who were untreated. I dedicate a large portion of the dissertation to explaining cross-national phenomena, but in Chapter 7, I investigate repression and dissent within the context of a single state over time. While I explore micro-level violence in Chapter 7, there is still a good deal more work that looks at the particular dynamics between states and dissidents needed. One prominent example is Kalyvas (2006) who examines the violent interaction between states and dissidents in the Greek Civil War. Since how a scholar aggregates time may affect the inferences that they draw, it is useful to look at multiple levels of aggregation (Shellman, 2004). While I explore country-year, and region-day data, there are many other temporal aggregation possibilities that need to be considered. Our theories of political violence are not yet sophisticated enough to inform us what level of aggregation is most appropriate; we need to proceed in a manner where we use different levels to probe the consistency of relationships. Additionally, I offer some clarification of concepts, such as support and state capacity, that are often associated with civil war specifically and political violence more generally. I applied this theory to interstate war, insurgency and counterinsurgency and suggested applications to terrorism. Terrorism is a particularly puzzling form of political violence as it seems to be directed at states that possess large amounts of resources and allow political participation, like the United States, the United Kingdom, Spain, Israel, and Germany. In the future, I plan to extend some of the theoretical insights from this project to explain why groups use terrorism as opposed to other forms of violent dissent. In brief, I believe that systems where small groups have extreme preferences are more likely to encourage terror as opposed to mass political violence or institutional participation. Where many people’s preferences are divergent from the state, civil war should be more likely than terrorism. Where a few disaffected people have extreme preferences, terrorism may be used to induce more repression from the state in hopes that this response will create more dissidents (Lake,

150 2002). Scholars of political violence, political regimes, and international relations may find interest in some of the work in this dissertation, but I also offer implications for statecraft, foreign policy and policy-making. In a recent article, Walt (2005, 23) claims that a literature which hopes to connect international relations theory with policy makers largely sees “a wide gap between academic theories of international relations and the practical conduct of of foreign policy.” For Walt (2005), this gulf between theory and practice exists because of ‘incentives structures’ for academics and the ‘complexity of making policy’ for foreign policy professionals. Newsom (1996, 65) claims that “because of both the emphasis on theories and models and the everlasting search for original dissertation topics, much of the research for the Ph.D. in international relations seems removed from contemporary issues” and thus relevance to policy-makers. At times, the academic community also strays from the relevant form of violence facing the international system. Mack (2002, 515) blames the “academic security studies community [for] continu[ing] to focus on interstate wars, while tending to ignore the 90% plus of armed conflicts that take place within, not between, states.” Fearon and Laitin (2003) note that between the end of World War II and 1999, “a conservative estimate of the total dead as a direct result of [civil wars] is 16.2 million, five times the interstate toll.” Harbom and Wallensteen (2007), using data from the International Peace Research Institute in Oslo (PRIO), reinforce Fearon and Laitin (2003)’s assertions and show that no violent conflict resulting in interstate war has occurred since 2003. Furthermore, the trend in these data since 1946 shows an increasingly large proportion of ongoing conflicts remain intrastate.1 And even though civil wars last longer than interstate wars on average, civil war is a particularly destructive form of violence as it affects the growth and health of a state long after the fighting stops (Ghobrah, Huth and Russett, 2003). While some academic studies may not be intimately connected to policy-making, this is not true for most studies related to civil war. As Mack (2002, 515) notes, “Understanding violent conflict is a necessary condition for preventing it.” Whereas some scholars, or so- called bridge-builders (Eriksson and Sundelius, 2005), attempt to make their work policy

1A small portion of these intrastate conﬂicts become internationalized or the condition where another state provides military personnel to one of the warring factions within a state (Harbom and Wallensteen, 2007).

151 relevant, nearly all who work on civil war strive to make claims about how to avoid this outcome for a state.2 Since in this larger project, I offer an explanation for how states and dissidents violently interact to produce civil war, I implicitly make policy suggestions about ending this process. As I argue in Chapter 3, leaders who have incentives to repress often use this tool to remain in office. Promoting economic growth and avoiding irregular leadership changes are ways to reduce the leader’s job insecurity. Whereas some of the influential literature on state repression found that developed democracies tend to repress less (Davenport, 1995; Davenport and Armstrong, 2004; Poe and Tate, 1994; Poe, Tate and Keith, 1999), short of regime change and a massive increase in national income, it is unclear how states could use policy to change the conditions that make repression more likely.3 Another potential policy implication is that resources available to the state can reduce the likelihood of violent challenges to its authority. While national income per capita tends not to vary much over time, other sources of resources, such as foreign aid or alliances, are potential tools that can increase this dimension of state capacity. In contrast to the resource dimension of state capacity, societal support should be more fluid and amenable to change by policymakers and external actors. One way to increase support is to change how regimes channel political participation. Based on the logic developed in Chapters 2 and 7, how support is distributed in society can influence how many participants are willing and likely to use violence. Where political institutions encourage ethnic politics, class-based politics, or other rigid or immutable coalitions, civil war and other forms of violence should be more likely. When large groups of people are incentivized to have preferences that are widely divergent, according to the model I develop, civil war is more likely. This is consistent with some of the prescriptions offered by Horowitz (1985) for dealing with ethnic conflict. Horowitz (1985) claims that federalism, autonomy, and other structural mechanisms for reducing conflict over the central government work, conditional on some important factors related to ethnicity. According to Horowitz (1985, 603) “federalism can either exacerbate or mitigate ethnic conflict...depend[ing] on the number of component states in a federation,

2The same claim could be made, of course, for studies that attempt to understand the causes of interstate war. 3In fact, the process of democratization could increase repression in the short term (Davenport, 2004).

152 their boundaries, and their ethnic composition.” One of the critical insights that Horowitz identifies is that the creation of cross-cutting cleavages helps states transcend ethnicity. In Nigeria, the creation of federalism “created incentives for political actors to see at least a few all-Nigeria issues in terms of competition among states, rather than among ethnic groups” (Horowitz, 1985, 612). In my framework, disputes that center around ethnicity create clear poles whereas disputes where individuals are cross-pressured should lead to less space between positions among contending groups. Posner (2004) demonstrates in his analysis of Chewa and Tumbuka people that sometimes ethnicity matters for political conflict while other times it does not. According to Posner (2004, 529), “if the cultural cleavage defines groups that are large enough to constitute viable coalitions in the competition for political power, then politicians will mobilize these groups and the cleavage that divides them will become politically salient.” Again, in my framework, where distinct ethnic groups create a bimodal distribution of support for the policies of the state, mass violent conflict in the form of civil war is more likely. Much of the recent literature on state failure or political instability also suggests that way a regime is configured matters for avoiding state collapse. The Political Instability Taskforce (formerly the State Failure Taskforce), using data spanning from 1955-2003, find that states that have a high degree of factionalism, or political participation that favors ethnic or parochial agendas, tend to have a higher likelihood of political instability. State failure is often defined as the complete collapse of government authority (King and Zeng, 2001). Milliken and Krause (2002) among others, however, claim that the concept of state failure is too closely associated with some of the outcomes that it attempts to explain such as civil war, genocide and regime change. I agree with this critique and conceive of state failure as a point in the state capacity space that is close to or exactly at the weakest level (See Figure 2.4 in Chapter 2).4 Where factionalism is rampant, state leaders receive less support, and thus, have less capacity to implement their preferred policies. The policy implication that I offer is that to build capacity, states need to build support. In some contexts, this requires creating political environments that do not allow policies to shift dramatically. Moreover, political systems should not allow policy to move a great distance from a status quo close to most groups in society’s preferences, thus reducing the possibility of violent conflict. Where

4State failure can be more precisely conceptualized as a space near the origin or where a state has both a minimum level of support and resources.

153 political systems already face wide divisions of groups of preferences, institutional means for reducing the spoils of winning elections or of leadership of the state could move the status quo away from some group’s ideal point but closer to a point where more groups are more supportive of the state. Newsom (1996, 62) claims that “the greatest barriers to any broad influence of academic analyses on international affairs today are the language and patterns in which results are presented.” Academics often create an internal language and as Mack (2002, 517) notes “much of the difficulty arises because very few researchers in this growing field [of civil war research]...bother to ‘translate’ their work in such a way as to make it accessible to policy- makers.” A notable exception is the work on the democratic peace which has influenced all recent U.S. presidents. As former US President Bill Clinton said in his 1994 State of the Union address, “Ultimately, the best strategy to ensure our security and to build a durable peace is to support the advance of democracy elsewhere. Democracies don’t attack each other.”5 While most academic work falls short of the impact of the democratic peace, I hope that this project not only solves some puzzling theoretical questions but also informs policy-makers regarding which states are most at risk for civil war and how to reduce the likelihood that people in these states experience the misery associated with this destructive human interaction.

5http://www.washingtonpost.com/wp-srv/politics/special/states/docs/sou94.htm

154 REFERENCES

Achen, Christopher. 1986. The Statistical Analysis of Quasi-Experiments. Berkeley: University of California Press. 4.3

Achen, Christopher. 2002. “Toward a New Political Methodology: Microfoundations and ART.” Annual Review of Political Science 5:423–450. 4.2.3, 5.5

Alesina, Alberto and David Dollar. 2000. “Who Gives Foreign Aid to Whom and Why?” Journal of Economic Growth 5(1):33–63. 8.2

Altﬁeld, Michael and Bruce Bueno de Mesquita. 1979. “Choosing Sides in War.” Interna- tional Studies Quarterly 23:87–112. 8.5

Alvarez, Michael and Garrett Glasgow. 1999. “Two-Stage Estimation of Non-Recursive Choice Models.” Political Analysis 8:147–165. 4.3, 27, 8.5.1

Arendt, Hannah. 1970. On Violence. New York: Harvest Books. 24

Asher, Herbert B. 1976. Causal Modeling. Beverly Hills: Sage Publications. 5.2

Ayoob, Mohammed. 1995. The Third World Security Predicament. Boulder, CO: Lynne Rienner. 4.2.1, 8.3

Balch-Lindsay, Dylan and Andrew J. Enterline. 2000. “Killing Time: The World Politics of Civil War Duration.” International Studies Quarterly 44:615–642. 8.4

Banks, Arthur. 1996. Cross-National Time-Series Data Archive. Binghampton, Electronic File: Center for Social Analysis, State University of New York, Binghampton. 4.2.2

Barabas, Jason. 2004. “How Deliberation Aﬀects Policy Opinions.” American Political Science Review 98:687–701. 6.1.1

Beck, Nathaniel, David Epstein, Simon Jackman and Sharyn O’Halloran. 2001. “Alternative Models of Dynamics in Binary Time-Series-Cross-Section Models: The Example of State Failure.” Paper presented at the 2001 Annual Meeting of the Society for Political Methodology, Emory University, Atlanta, GA. 4.4, 8.6

Beck, Nathaniel, Jonathan N. Katz and Richard Tucker. 1998. “Taking Time Seriously: Time-Series–Cross–Section Analysis with a Binary Dependent Variable.” American Jour- nal of Political Science 42:1260–1288. 4.2.2, 20, 4.2.3, 8.5.1

155 Beckett, Ian F. 1999. Encyclopedia of Guerrilla Warfare. Oxford: ABC-Clio Inc. 7.1

Berry, William. 1984. Nonrecursive Causal Models. Beverly Hills, C.A.: Sage. 3.4, 11

Berry, William, Justin Esarey and Jacqueline DeMeritt. 2008. “Testing for Interaction in Binary Logit and Probit Models: Is a Product Term Essential?” Presented at the Annual Meeting of the Political Methodology Association, Penn State University. 18

Blalock, Herbert. 1969. Theory Construction. Englewood Cliﬀs, NJ: Prentice-Hall. 5.5

Blimes, Randall. 2006. “The Indirect Eﬀect of Ethnic Heterogeneity on the Likelihood of Civil War Onset.” Journal of Conﬂict Resolution 50:536–547. 3, 4.3.1

Brandt, Patrick and John T. Williams. 2001. “A Linear Poisson Autoregressive Model: The Poisson AR (p) Model.” Political Analysis 9(2):164. 20, 7.8

Brandt, Patrick, John T. Williams and Benjamin Fordham. 2000. “Fordham, and Brian Pollins. 2000.Dynamic Modeling for Persistent Event-Count Time Series..” American Journal of Political Science 44(4):823–43. 7.7, 20, 7.10

Browne, M. W. and R. Cudeck. 1993. “Assessing Ways of Assessing Model Fit”. In Testing Structural Equation Models, ed. Kenneth Bollen and J. Scott Long. Newbury, CA: Sage pp. 136–162. 5.3

Bueno de Mesquita, Bruce. 1975. “Measuring Systemic Polarity.” Journal of Conﬂict Resolution 19:187–216. 8.5

Bueno de Mesquita, Bruce, Alastair Smith, Randolph Siverson and James Morrow. 2003. The Logic of Political Survival. Cambridge, MA: MIT Press. 3.3.2, 4.2.2, 8.1

Bueno de Mesquita, Bruce, James Morrow, Randolph Siverson and Alastair Smith. 1999. “An Institutional Explanation for the Democratic Peace.” American Political Science Review 93:791–807. 12, 3.3.2, 17, 18, 8.1

Cameron, A.C. and P.K. Trivedi. 1986. “Econometric Models Based on Count Data: Comparisons and Applications of Some Estimators and Tests.” Journal of Applied Econo- metrics 1(1):29–53. 7.10

Carmines, Edward G. and Richard A. Zeller. 1979. Reliability and Validity Assessment. Sage Publications. 36

Carter, David B. and Curtis S. Signorino. 2006. “Back to the Future: Modeling Time Dependence in Binary Data.” Unpublished Manuscript. 20, 8.5.1

Cheibub, Jos´e. 1998. “Political Regimes and the Extractive Capacity of Governments: Taxation in Democracies and Dictatorships.” World Politics 50:349–376. 14, 27, 4.2.2, 8, 5.3

156 Chiozza, Giacomo and H. E. Goemans. 2003. “Peace Through Insecurity: Tenure and International Conﬂict.” Journal of Conﬂict Resolution 47:443–467. 8.1

Chong, Dennis. 1991. Collective Action and the Civil Rights Movement. Chicago: University of Chicago Press. 27

Cingranelli, David L. and David L. Richards. 1999. “Measuring the Pattern, Level, and Sequence of Government Respect for Human Rights.” International Studies Quarterly 43:407–417. 4.2.2

Cioﬃ-Revilla, Claudio and Harvey Starr. 1995. “Opportunity, Willingness and Political Uncertainty: Theoretical Foundations of Politics.” Journal of Theoretical Politics 7(4):447. 1

Clague, Christopher, Philip Keefer, Stephen Knack and Mancur Olson. 1999. “Contract- Intensive Money: Contract Enforcement, Property Rights, and Economic Performance.” Journal of Economic Growth 4:185–211. 4.2.2

Clarke, Kevin. 2005. “The Phantom Menace: Omitted Variable Bias in Econometric Research.” Conﬂict Management and Peace Science 22:341–352. 4.2.3

Cohen, Youssef, Brian Brown and A. F. K. Organski. 1989. “The Paradoxical Nature of State Making: The Violent Creation of Order.” American Political Science Review 75:901–910. 8.3

Collier, Paul and Anke Hoeﬄer. 1998. “On Economic Causes of Civil War.” Oxford Economic Papers 50(4):563–573. 1

Collier, Paul and Anke Hoeﬄer. 2001. “Greed and Grievance in Civil War.” World Bank Working Paper. 1, 2.2.1, 3.3, 8, 5.5, 18

D’Agostino, Ralph B. 1998. “Propensity Score Methods for Bias Reduction in the Com- parison of a Treatment to a Non-Randomized Control Group.” Statistics in Medicine 17:2265–2281. 6.1.1, 3, 6.1.4

Dahl, Robert A. 1971. Polyarchy. New Haven: Yale University Press. 13

Davenport, Christian. 1995. “Multi-Dimensional Threat Perception and State Repression: An Inquiry into Why States Apply Negative Sanctions.” American Journal of Political Science 39(3):683–713. 9

Davenport, Christian. 2004. “The Promise of Democratic Paciﬁcation: An Empirical Assessment.” International Studies Quarterly 48(3):539–560. 3

Davenport, Christian. 2007. “State Repression and Political Order.” Annual Review of Political Science 10:1–23. 3.3.1, 3.3.1

157 Davenport, Christian and David Armstrong. 2004. “Democracy and the Violation of Human Rights: A Statistical Analysis from 1976 to 1996.” American Journal of Political Science 48(3):538–554. 9

Dehejia, Rajeev H. and Sadek Wahba. 1999. “Causal Eﬀects in Nonexperimental Studies: Reevaluating the Evaluation of Training Programs.” Journal of the American Statistical Association 94:1053–1062. 6.1.1

DeNardo, James. 1985. Power in Numbers: the Political Strategy of Protest and Rebellion. Princeton, NJ: Princeton University Press. 24

Denzau, Arthur and Douglass North. 1994. “Shared Mental Models: Ideologies and Institutions.” Kyklos 47:3–31. 2.3

Eckstein, Harry. 1971. The Evaluation of Political Performance. Beverly Hills, CA: Sage Publications. 18

Efron, Brad and Robert Tibshirani. 1986. “Bootstrap Methods for Standard Errors, Conﬁdence Intervals, and Other Measures of Statistical Accuracy.” Statistical Science 1:54–75. 5.2

Elbadawi, Ibrahim and Nicholas Sambanis. 2002. “How Much War Will We See?: Explaining the Prevalence of Civil War.” Journal of Conﬂict Resolution 46(3):307. 1

Enders, Walter. 2004. Applied Econometric Time Series. New York: John Wiley & Sons. 7.10

Enders, Walter and Todd Sandler. 1993. “The Eﬀectiveness of Antiterrorism Policies: A Vector-Autoregression-Intervention Analysis.” The American Political Science Review 87(4):829–844. 8.5

Eriksson, Johan and Bengt Sundelius. 2005. “Molding Minds That Form Policy: How to Make Research Useful.” International Studies Perspectives 6(1):51–71. 9

Fearon, James D. 1997. “Signaling Foreign Policy Interests: Tying Hands Versus Sinking Costs.” Journal of Conﬂict Resolution 41(1):68. 8.2

Fearon, James and David Laitin. 1996. “Explaining Interethnic Cooperation.” American Political Science Review 90(4):715–35. 1

Fearon, James and David Laitin. 2003. “Ethnicity, Insurgency, and Civil War.” American Political Science Review 97:75–90. 1, 2.1, 5, 2.1, 2.2.1, 3.1, 3, 3.2, 7, 3.4, 4.1, 4.2.1, 4.2.2, 4.2.2, 19, 5.3, 5.5, 18, 6.1, 6.2, 8.5, 13, 9

Galula, David. 1964. Counterinsurgency Warfare: Theory and Practice. New York: Praeger Publishers. 7.2

158 Gartzke, Erik. 1999. “War is in the Error Term.” International Organization 53(3):567–587. 6.2

Geddes, Barbara. 1995. The Politician’s Dilemma: Building State Capacity in Latin America. Berkeley: University of California Press. 2.2, 3.3.1

Gelman, Andrew and Jennifer Hill. 2007. Data Analysis Using Regression and Multi- level/Hierarchical Models. New York: Cambridge University Press. 6.1.1

Ghobrah, Hazem Adam, Paul Huth and Bruce Russett. 2003. “Civil Wars Kill and Maim People—Long After the Shooting Stops.” American Political Science Review 97(02):189– 202. 1, 9

Gibney, Mark and Matthew Dalton. 1996. “The Political Terror Scale.” Policy Studies and Developing Nations 4:73–84. 4.2.2

Gleditsch, Kristian Skrede. 2007. “Transnational Dimensions of Civil War.” Journal of Peace Research 44:293–309. 8.1, 8.8

Gleditsch, Kristian Skrede and Michael D. Ward. 2000. “War and Peace in Space and Time: The Role of Democratization.” International Studies Quarterly 44:1–29. 24

Gleditsch, Nils Petter, Peter Wallensteen, Mikael Eriksson, Margareta Sollenberg and Havard Strand. 2002. “Armed Conﬂict 1946-2001: A New Dataset.” Journal of Peace Research 39:615–637. 4.2.2

Goemans, Hein. 2000. War and Punishment. Princeton, NJ: Princeton University Press. 8.1

Goertz, Gary. 2006. Social Science Concepts: A User’s Guide. Princeton, NJ: Princeton University Press. 1, 2.1, 2.3, 16, 2.3, 30, 2.3, 2.4, 38, 3

Goodwin, Jeﬀ. 2001. No Other Way Out: States and Revolutionary Movements, 1945-1991. Cambridge: Cambridge University Press. 1, 2, 32, 3.5, 29

Gujarati, Damodar. 2003. Basic Econometrics. New York: McGraw Hill, Inc. 1, 5.2, 11

Gurr, Ted Robert. 1986. “The Political Origins of State Violence and Terror: A Theoretical Analysis”. In Governmental Violence and Repression: An Agenda for Research, ed. Michael Stohl and George Lopez. New York: Greenwood Press. 27, 4.2.2, 9

Gurr, Ted Robert. 1988. “War, Revolution, and the Growth of the Coercive State.” Comparative Political Studies 21:45–65. 25, 5.3, 9

Harbom, Lotta and Peter Wallensteen. 2007. “Armed Conﬂict, 1989 2006.” Journal of Peace Research 44(5):623. 9, 1

Hegre, Havard and Nicholas Sambanis. 2006. “Sensitivity Analysis of Empirical Results on Civil War Onset.” Journal of Conﬂict Resolution 50(4):508. 2.1

159 Hibbs, Douglas A. 1973. Mass Political Violence: A Cross-National Causal Analysis. New York: John Wiley & Sons. 1, 5.3

Hill, Jennifer. 2004. “Reducing Bias in Treatment Effect Estimation in Observational Studies Suffering from Missing Data”. Technical report Working paper series, School of International and Public Affairs, Columbia University, New York. 6.3, 14

Ho, Daniel, Kosuke Imai, Gary King and Elizabeth Stuart. 2004. “MatchIt: Match- ing as Nonparametric Preprocessing for Improving Parametric Causal Inference.” http://gking.harvard.edu/matchit. 6.1.5, 6.2

Ho, Daniel, Kosuke Imai, Gary King and Elizabeth Stuart. 2007. “Matching as Nonparamet- ric Preprocessing for Improving Parametric Causal Inference.” Political Analysis 15:1–38. 6.1, 6.1.1, 6.1.2, 6.1.5, 8

Honaker, James and Gary King. 2006. “What to do About Missing Values in Time Series Cross-Section Data.” Unpublished Manuscript. 4.2.3, 15, 9

Horowitz, Donald. 1985. Ethnic Groups in Conﬂict. Berkeley: University of California Press. 9

Horton, Nicholas J. and Ken P. Kleinman. 2007. “Much Ado About Nothing: A Comparison of Missing Data Methods and Software to Fit Incomplete Data Regression Models.” American Statistician 61:79–90. 4.2.3, 23

Huntington, Samuel. 1993. “The Clash of Civilizations?” Foreign Aﬀairs 72:22–50. 1

Huntington, Samuel P. 1968. Political Order in Changing Societies. New Haven: Yale University Press. 2.3

Jackman, Robert. 1993. Power Without Force: The Political Capacity of Nation States. Ann Arbor: University of Michigan Press. 18, 2.3, 32, 10

Jackson, Robert H. 1990. Quasi-states: Sovereignty, International Relations, and the Third World. Cambridge: Cambridge University Press. 4.2.1, 8.3

Jaggers, Keith. 1992. “War and the Three Faces of Power: War Making and State Making in Europe and the Americas.” Comparative Political Studies 25(1):26. 2.3, 32

Kalyvas, Stathis. 2006. The Logic of Violence in Civil War. New York: Cambridge University Press. 10, 9

Kaplan, Robert. 2001. The Coming Anarchy: Shattering the Dreams of the Post Col War. New York: Vintage. 1

Kastellec, Jonathan and Eduardo Leoni. 2007. “Using Graphs Instead of Tables in Political Science.” Perspectives on Politics 5(04):755–771. 16

160 Kennedy, Paul. 2003. A Guide to Econometrics. Malden, MA: Blackwell Publishers. 4.3, 25

King, Gary. 1986. “How Not to Lie With Statistics: Avoiding Common Mistakes in Quantitative Political Science.” American Journal of Political Science 30:666–687. 5.3

King, Gary. 1988. “Statistical Models for Political Science Event Counts: Bias in Conven- tional Procedures and Evidence for the Exponential Poisson Regression Model.” American Journal of Political Science 32(3):838–863. 7.10

King, Gary, James Honaker, Anne Joseph and Kenneth Scheve. 2001. “Analyzing Incomplete Political Science Data: An Alternative Algorithm for Multiple Imputation.” American Political Science Review 95:49–69. 21, 9

King, Gary and Langche Zeng. 2001. “Improving Forecasts of State Failure.” World Politics 53(4):623–58. 9

King, Gary and Langche Zeng. 2007. “When Can History be Our Guide? The Pitfalls of Counterfactual Inference.” International Studies Quarterly 51:183–210. 6.1.2

King, Gary, Michael Tomz and Jason Wittenberg. 2000. “Making the Most of Statistical Analyses: Improving Interpretation and Presentation.” American Journal of Political Science 44:347–361. 15, 16

King, Gary, Michael Tomz and Jason Wittenberg. 2001. “Making the Most of Statistical Analyses: Improving Interpretation and Presentation.” American Journal of Political Science 44:347–361. 4.3.1

King, Gary, Robert Keohane and Sidney Verba. 1994. Designing Social Inquiry: Scientiﬁc Inference in Qualitative Research. Princeton: Princeton University Press. 4, 4.2.3, 8

Kugler, Jacek and William Domke. 1986. “Comparing the Strength of Nations.” Comparative Political Studies 19:39–69. 1, 2.3, 2.3, 32, 4.2.2

Kydd, Andrew and Barbara Walter. 2002. “Sabotaging the Peace: The Politics of Extremist Violence.” International Organization 56:263–296. 3.3.2

Lacina, Bethany Ann and Nils Petter Gleditsch. 2005. “Monitoring Trends in Global Combat: A New Dataset of Battle Deaths.” European Journal of Population 21:145–165. 3.1

Lakatos, Imre. 1977. The Methodology of Scientiﬁc Research Programmes: Philosophical Papers Volume 1. Cambridge: Cambridge University Press. 30

Lake, David. 2002. “Rational Extremism: Understanding Terrorism in the Twenty-ﬁrst Century.” Dialogue IO 1:15–29. 3.3.2, 9

Leamer, Edward E. 1983. “Let’s Take the Con Out of Econometrics.” The American Economic Review 73(1):31–43. 6.1.2

161 Leeds, B.A., J.M. Ritter, S.M.L. Mitchell and G. Andrew. 2002. “Alliance Treaty Obligations and Provisions, 1815-1944.” International Interactions 28(3):237–260. 6

Leeds, Brett Ashley. 2003. “Do Alliances Deter Aggression? The Inﬂuence of Military Alliances on the Initiation of Militarized Interstate Disputes.” American Journal of Political Science 47(3):427. 8.2

Leeds, Brett Ashley, Jeﬀrey M. Ritter, Sara McLaughlin Mitchell and Andrew G. Long. 1999. “Alliance Treaty Obligations and Provisions, 1815-1944.” International Interactions 28(28):237–260. 8.5

Leites, Nathan and Charles Wolf Jr. 1970. Rebellion and Authority: An Analytic Essay on Insurgent Conﬂicts. Chicago: Markham Publishing Company. 7.2

Lemke, Doug. 2007. “Personal Correspondence.”. 4.2.2

Leuven, Edwin and Barbara Sianesi. 2003. “PSMATCH2: Stata Module to Perform Full Mahalanobis and Propensity Score Matching, Common Support Graphing, and Covariate Imbalance Testing.”. http://ideas.repec.org/c/boc/bocode/s432001.html, Version 1.2.3. 6.1.5, 6.2

Levi, Margaret. 1988. Of Rule and Revenue. Berkeley: University of California Press. 2.3, 32, 3.3.1, 14

Li, Yunfei Paul, Kathleen J. Propert and Paul R. Rosenbaum. 2001. “Balanced Risk Set Matching.” Journal of the American Statistical Association 96:870–882. 12

Little, Daniel. 1991. Varieties of Social Explanation: An Introduction to the Philosophy of Social Science. Boulder, CO: Westview Press. 14, 22

Little, J. Rodrick and Donald Rubin. 1989. “The Analysis of Social Science Data with Missing Values.” Sociological Methods & Research 18(2-3):292. 21

Little, J. Rodrick and Nathaniel Schenker. 1995. “Missing Data.” Handbook of Statistical Modeling for the Social and Behavioral Sciences pp. 39–75. 21

Long, J. Scott and Jeremy Freese. 2006. Regression Models for Categorical Dependent Variables Using Stata. College Station: Stata Press. 7.10

Lu, Bo. 2005. “Propensity Score Matching with Time-Dependent Covariates.” Biometrics 61:721–728. 12

Lu, Bo. 2008. “Personal Correspondence.”. 12

Lukes, Stephen. 1974. Power: A Radical View. London: Macmillan. 17

Mack, Andrew. 2002. “Civil War: Academic Research and the Policy Community.” Journal of Peace Research 39(5):515. 9

162 Mahoney, James and Gary Goertz. 2004. “The Possibility Principle: Choosing Negative Cases in Comparative Research.” American Political Science Review 98:653–669. 6.1.1

Mann, Michael. 1993. The Sources of Social Power: The Rise of Classes and Nation-States, 1760-1914. New York: Cambridge University Press. 3.3.1

Marshall, Monty G. and Keith Jaggers. 2001. “Polity IV project: Political Regime Char- acteristics and Transitions, 1800-2004.” Retrieved from http://www.bsos.umd.edu/polity. 4.2.2

Mason, David and Dale Krane. 1989. “The Political Economy of Death Squads: Toward a Theory of the Impact of State-Sanctioned Terror.” International Studies Quarterly 33:175– 198. 2.3, 3.1

Mattei, Alessandra. 2008. “Estimating and Using Propensity Score in Presence of Missing Background Data: An Application to Assess the Impact of Childbearing on Wellbeing.” Statistical Methods and Applications Forthcoming:1–17. 6.3

Migdal, Joel. 1988. Strong Societies and Weak States. Princeton, NJ: Princeton University Press. 32, 3.3.1

Milliken, Jennifer and Keith Krause. 2002. “State Failure, State Collapse, and State Reconstruction: Concepts, Lessons and Strategies.” Development and Change 33(5):753– 774. 9

Mitchell, Neil and James McCormick. 1988. “Economic and Political Explanations of Human Rights Violations.” World Politics 40(4):476–498. 3.3.1

Mitchell, Sara McLaughlin and William H. Moore. 2002. “Presidential Uses of Force During the Cold War: Aggregation, Truncation, and Temporal Dynamics.” American Journal of Political Science 46(2):438–52. 7.7, 7.8

Mocan, H. Naci and Erdal Tekin. 2006. “Catholic Schools and Bad Behavior: A Propensity Score Matching Analysis.” The BE Journal of Economic Analysis & Policy 5(1):13. 6.1

Moore, Will. 1995. “Action-Reaction or Rational Expectations? Reciprocity and the Domestic-International Conﬂict Nexus During the ’Rhodesia Problem’.” Journal of Con- ﬂict Resolution 39:129–167. 3.3

Moore, Will. 1998. “Repression and Dissent: Substitution, Context, and Timing.” American Journal of Political Science 42:851–873. 9

Moore, Will. 2000. “Repression and Dissent: A Substitution Model of Government Coercion.” Journal of Conﬂict Resolution 44:107–127. 9

Morgan, Stephen L. and Christopher Winship. 2007. Counterfactuals and Causal Inference. New York: Cambridge University Press. 6.1.3, 10

163 Morrow, James D. 1991. “Alliances and Asymmetry: An Alternative to the Capability Aggregation Model of Alliances.” American Journal of Political Science 35:904–933. 8.2

Morton, Jeﬀrey S. and Harvey Starr. 2001. “Uncertainty, Change, and War: Power Fluctuations and War in the Modern Elite Power System.” Journal of Peace Research 38(1):49. 8.4

Most, Benjamin and Harvey Starr. 1984. “International Relations Theory, Foreign Policy Substitutability.” World Politics 36:383–406. 8.2

Munck, Gerardo and Jay Verkuilen. 2002. “Conceptualizing and Measuring Democracy: Evaluating Alternative Indices.” Comparative Political Studies 35:5–34. 2.3, 34, 2.4, 40, 4.2.2

Murdoch, James C. and Todd Sandler. 2002. “Economic Growth, Civil Wars, and Spatial Spillovers.” Journal of Conﬂict Resolution 46:91–110. 8.1

Muth´en, Bengt. 1984. “A General Structural Equation Model with Dichotomous, Ordered Categorical, and Continuous Latent Variable Indicators.” Psychometrika 49:115–132. 5.2

Muthén, Bengt and Linda Muthén. 2004. “Mplus User’s Guide.” Los Angeles, CA: Muthén and Muthén. 5.2

Nettl, J.P. 1968. “The State as a Conceptual Variable.” World Politics 20(4):559–592. 12

Newsom, David D. 1996. “Foreign Policy and Academia.” Foreign Policy 101:52–64. 9

Nordlinger, Eric. 1987. “Taking the State Seriously”. In Understanding Political Develop- ment, ed. Myron Weiner and Samuel P. Huntington. Boston, MA: Little, Brown and Company pp. 353–390. 19, 32, 20

North, Douglas. 1990. Institutions, Institutional Change and Economic Performance. Cambridge: Cambridge University Press. 28

North, Douglass and Barry Weingast. 1989. “Constitutions and Commitment: The Evolution of Institutions Governing Public Choice in 17th Century England.” Journal of Economic History 49:803–832. 2.3

Organski, AFK and Jacek Kugler. 1980. The War Ledger. Chicago: University of Chicago Press. 8.1, 8.4

Plumper, Thomas and Vera Troeger. 2007. “Eﬃcient Estimation of Time-Invariant and Rarely Changing Variables in Finite Sample Panel Analyses with Unit Fixed Eﬀects.” Political Analysis 15(2):124. 8

Poe, Steven C. and C. Neil Tate. 1994. “Repression of Human Rights and Personal Integrity in the 1980’s: A Global Analysis.” American Political Science Review 88:853–872. 3.3.1, 27, 4.2.2, 26, 4.3.1, 9

164 Poe, Steven C., C. Neil Tate and Linda Camp Keith. 1999. “Repression of the Human Right to Personal Integrity Revisited: A Global Cross-National Study Covering the Years 19761993.” International Studies Quarterly 43:291–313. 3.3.1, 26, 4.3.1, 9

Posen, Barry. 1993. The Security Dilemma and Ethnic Conﬂict. In Ethnic Conﬂict and International Security, ed. M. E. Brown. Princeton: N. J.: Princeton University Press. 1

Posner, Daniel. 2004. “The Political Salience of Cultural Diﬀerence: Why Chewas and Tumbukas Are Allies in Zambia and Adversaries in Malawi.” American Political Science Review 98(04):529–545. 9

Putnam, Robert. 1988. “The Logic of Two-Level Games.” International Organization 42((3)):427–460. 8.2

Ragin, Charles C. 2000. Fuzzy-Set Social Science. Chicago: University of Chicago Press. 29

Ray, James Lee. 2003. “Explaining Interstate Conﬂict and War: What Should be Controlled For?” Conﬂict Management and Peace Science 20:1–31. 4.2.3, 5.1, 5.5

Ray, James Lee. 2005. “Constructing Multivariate Analyses (of Dangerous Dyads).” Conﬂict Management and Peace Science 22:277–292. 4.2.3

Raykov, Tenko and George Marcoulides. 2006. A First Course in Structural Equation Modeling. London: Lawrence Erlbaum Associates. 29, 5.1, 2, 10

Regan, Patrick M. 2000. Civil Wars and Foreign Powers: Outside Intervention in Intrastate Conﬂict. Univ of Michigan Press. 7

Rivers, Douglas and Quan H. Vuong. 1988. “Limited Information Estimators and Exogeneity Tests for Simultaneous Probit Models.” Journal of Econometrics 39:347–366. 4.3

Rosenbaum, Paul R. and Donald B. Rubin. 1983. “The Central Role of the Propensity Score in Observational Studies for Causal Eﬀects.” Biometrica 24:295–313. 6.1, 6.1.1, 6.1.3

Ross, Michael. 2004. “How do Natural Resources Inﬂuence Civil War? Evidence from 13 Cases.” International Organization 58:35–67. 7

Ross, Michael. 2006. “A Closer Look at Oil, Diamonds, and Civil War.” Annual Review of Political Science 9:265–300. 7, 4.2.2

Royston, Patrick. 2005. “Multiple Imputation of Missing Values: Update of Ice.” Stata Journal 5:527–536. 23

Sala-I-Martin, Xavier X. 1997. “I Just Ran Two Million Regressions.” The American Economic Review 87(2):178–183. 6.1.2

Salehyan, Idean and Kristian Skrede Gleditsch. 2006. “Refugees and the Spread of Civil War.” International Organization 60:335–366. 8.1

165 Sambanis, Nicholas. 2001. “Do Ethnic and Nonethnic Civil Wars Have the Same Causes?: A Theoretical and Empirical Inquiry (Part 1).” Journal of Conﬂict Resolution 45(3):259. 1

Sambanis, Nicholas. 2002. “A Review of Recent Advances and Future Directions in the Quantitative Literature on Civil War.” Defence and Peace Economics 13:215–243. 4.2.2

Sambanis, Nicholas. 2004. “What is Civil War? Conceptual and Empirical Complexities of an Operational Deﬁnition.” Journal of Conﬂict Resolution 48:814–858. 2.1, 8, 26, 4.2.2

Sartori, Giovanni. 1970. “Concept Misinformation in Comparative Politics.” American Political Science Review 64:1033–1053. 2.4

Shafer, D. Michael. 1988. “The Unlearned Lessons of Counterinsurgency.” Political Science Quarterly 103(1):57–80. 7.2

Shellman, Stephen. 2004. “Time Series Intervals and Statistical Inference: The Eﬀects of Temporal Aggregation on Event Data Analysis.” Political Analysis 12(1):97–104. 9

Shellman, Stephen. 2006. “Leaders & Their Motivations: Explaining Government-Dissident Conﬂict-Cooperation Processes.” Conﬂict Management & Peace Science 23:73–90. 3.3

Signorino, Curtis S. and Jeﬀrey M. Ritter. 1999. “Tau-b or not Tau-b: Measuring the Similarity of Foreign Policy Positions.” International Studies Quarterly 43:115–144. 8.5

Singer, J. David. 1972. “The Correlates of War Project: Interim Report and Rationale.” World Politics 24(2):243–270. 1

Singer, J. David and Melvin Small. 1994. “Correlates of War Project: International and Civil War Data, 1816-1992, data ﬁle, Inter-University Consortium for Political and Social Research.”. Ann Arbor, MI. 3.3

Singer, J. David, Stuart Bremer and John Stuckey. 1972. “Capability Distribution, Un- certainty, and Major Power War, 1820-1965”. In Peace, War, and Numbers, ed. Bruce Russett. Beverly Hills: Sage pp. 19–48. 8.5

Skocpol, Theda. 1979. States and Social Revolution. New York: Cambridge University Press. 1

Skocpol, Theda. 1985. “Bringing the State Back in”. In Bringing the State Back in, ed. Peter B. Evans, David Rueschmeyer and Theda Skocpol. Cambridge: Cambridge University Press chapter 1, pp. 3–43. 2.2, 9

Small, Melvin and J. David Singer. 1982. Resort to Arms: International and Civil War Data, 1816-1980. Beverly Hills: Sage. 3.3

Snyder, Richard. 2006. “Does Lootable Wealth Breed Disorder?: A Political Economy of Extraction Framework.” Comparative Political Studies 39:943–968. 7

166 Snyder, Richard and Rhavi Bhavnani. 2005. “Diamonds, Blood, and Taxes: A Revenue- Centered Framework for Explaining Political Order.” Journal of Conflict Resolution 49:563–597. 7 Starr, Harvey. 1994. “Revolution and War: Rethinking the Linkage between Internal and External Conflict.” Political Research Quarterly 47(2):481–507. 8.2 Stepan, Alfred. 1979. The State and Society: Peru in Comparative Perspective. Princeton: Princeton University Press. 2.2 Thies, Cameron G. 2006. “Public Violence and State Building in Central America.” Comparative Political Studies 39:1263–1282. 8.3 Thies, Cameron G. 2007. “The Political Economy of State Building in Sub-Saharan Africa.” Journal of Politics 69:716–731. 8.3 Thyne, Clayton. 2006. “Cheap Signals with Costly Consequences: The Effect of Interstate Relations on Civil War.” Journal of Conflict Resolution 50(6):937. 1 Tilly, Charles. 1985. “War Making and State Making as Organized Crime”. In Bringing the State Back in, ed. Peter B. Evans, David Rueschmeyer and Theda Skocpol. Cambridge: Cambridge University Press chapter 5, pp. 169–191. 8.1, 8.3 Tilly, Charles. 1995. “To Explain Political Processes.” The American Journal of Sociology 100:1594–1610. 3.3 Tilly, Charles. 2003. The Politics of Collective Violence. Cambridge: Cambridge University Press. 2.2 Trujillo, A.J., J.E. Portillo and J.A. Vernon. 2005. “The Impact of Subsidized Health Insurance for the Poor: Evaluating the Colombian Experience Using Propensity Score Matching.” International Journal of Health Care Finance and Economics 5(3):211–239. 6.1, 2 Walt, Stephen. 2005. “The Relationship between Theory and Policy in International Relations.” Annual Review of Political Science 8:23–48. 9 Waltz, Kenneth. 1979. Theory of International Politics. Boston, Mass.: McGraw-Hill. 8.1, 8.4 Ward, Mike and Kristin Bakke. 2005. “Predicting Civil Conflicts: On the Utility of Empirical Research.” Paper presented at the Conference on Disaggregating the Study of Civil War and Transnational Violence, University of California Institute of Global Conflict and Cooperation, San Diego, CA. 4.4, 6.1, 6.1.1, 6.3 West, Kenneth. 1986. “Full Versus Limited Information Estimation of a Rational Expecta- tions Model: Some Numerical Comparisons.” NBER Working Paper No. T0054. 5.2 Wooldridge, Jeffrey. 2000. Introductory Econometrics. South-Western College: Cincinnati, OH. 25

167 BIOGRAPHICAL SKETCH

Joseph K. Young

Joseph K. Young received his Bachelor of Arts degree in Economics and International Relations from Stetson University in 1998. In 2003, he completed a Master of Arts degree in Political Science from Ohio University. In the fall of 2008, Joseph joins the faculty at Southern Illinois University and will receive a joint appointment in Political Science and Criminology.

168