PART I
Double Take
CHAPTER 1
Democracy, Ethnicity, Religion, and Civil War
Endogeneity Bias
Great scientific discoveries have been made by men seeking to verify quite erroneous theories about the nature of things.
ALDOUS HUXLEY, âWordsworth in the Tropicsâ
Although most existing empirical studies acknowledge that some of the variables included in civil war models are endogenous, they nevertheless rely on a single equation logit analysis that assumes the exogeneity of all independent variables.1 However, there is much literature to show how such a reliance leads to the reporting of biased estimates (Miguel, Satyanath, and Sergenti 2004; Bueno de Mesquita 2005; Sobek 2010; see also James, Solberg, and Wolfson 1999; DeRouen 2000; DeRouen and Sobek 2004; McLean and Whang 2010; Thies 2010; Whang 2010). This study, then, offers a methodological improvement to the existing statistical models by focusing on the most widely cited and replicated civil war model, namely, that built by Fearon and Laitin (2003); in the process, this study challenges the main findings of their civil war research.
Before the publication of Fearon and Laitinâs (2003) work, it was widely accepted by scholars, policy makers, and journalists that countries with more ethnically or religiously diverse populations have a greater risk for civil war, especially in the presence of political discrimination (e.g., Gellner 1983; Huntington 1996). At first glance, Fearon and Laitinâs (2003, 75) single equation logit analysis of 161 countries from 1945 to 1999 appears to have successfully refuted this commonly held belief; their findings indicate that when per capita income and population are considered alongside other conflict-related factors, the main causes of civil war are other than democracy (used as a proxy for broadly held grievances/discriminations in their study), ethnic difference, or religious difference. Compelling as they are, these results may be spurious due to the fact that the single equation logit analysis overlooks the endogeneity of civil violence in relation to each of the three factors. For example, countries burdened by political grievances are likely to experience higher levels of civil violence; correspondingly, a higher level of civil violence will result in additional political grievances. Therefore, the predictor and the outcome variables are codependent, implying that causality runs in both directions. A single equation logit model fails to capture the nuance of this two-way causal relationship.
Rigorous statistical analysis assumes that each variable on the right-hand side of a single equation exerts an independent effect on the dependent variable; it also assumes that these variables are not correlated with the error term, which means that the outcome variable is not supposed to affect any of the predictors. However, when an endogenous variable is placed on the right-hand side of an equation (e.g., the democracy variable in Fearon and Laitinâs single equation of civil war), it becomes correlated with the error term, thereby violating a fundamental assumption of statistical analysis and subjecting the findings to endogeneity bias. By distorting estimated results, endogeneity bias makes it impossible for researchers to accurately conduct significance tests and draw correct statistical inferences. A simultaneous equations model in which two equations of endogenous variables are jointly estimated for robust results is typically recommended as a remedial measure (see James, Solberg, and Wolfson 1999; DeRouen 2000; Greene 2003; DeRouen and Sobek 2004; McLean and Whang 2010; Whang 2010).
Civil war researchers tend to ignore the potential effects of endogeneity bias in their empirical analyses for two reasons. First, due to the dichotomous nature of one of the endogenous variablesâthe onset of civil warâa standard simultaneous equations estimator designed for two continuous endogenous variables does not apply. Second, the identification of valid instrumental variables is a challenging task in simultaneous equations modeling. Rather than deal directly with these problems, existing studies of civil war either ignore them or rely on ad hoc remedies such as the practice of taking lags of right-hand-side variables and using observations at five-year intervals. As a result, the estimated results are biased at best and totally inaccurate at worst (Bueno de Mesquita 2005; Hegre and Sambanis 2006; Sobek 2010; Whang 2010).
This study reexamines Fearon and Laitinâs civil war model in an attempt to offer a methodological improvementâone that takes into consideration the endogenous relationship between the onset of civil war and democracy, ethnicity, or religion. This study postulates that democracy, ethnic fragmentation, or religious division is affected by internal political conflict and that the onset of internal political conflict is likewise affected by each of these three factors. To test the hypothesis, this study draws on Keshkâs (2003) two-stage probit least squares model,2 which provides the necessary methodological correction to Fearon and Laitinâs single equation logit analysis; that is, the Keshk estimator is designed to solve the endogeneity problem even in cases where one endogenous variable is continuous and the other dichotomous. After applying Keshkâs simultaneous equations model, this study uncovers evidence that, all other things being equal, well-established democracies and religiously diverse countries experience more civil wars, while ethnic diversity indicates that a country is less likely to encounter this type of violence. These new findings indicate that Fearon and Laitinâs results are the statistical artifacts of a biased single equation estimation. This study suggests that to better understand the roles played by democracy, ethnicity, and religion in the onset of civil war, future research must formally address endogeneity bias in the effort to ensure accurate statistical inferences and make appropriate policy recommendations.
BUILDING A TWO-STAGE SIMULTANEOUS EQUATIONS MODEL OF CIVIL WAR
The first subsection briefly provides a rationale for the existence of two-way causality between civil war and grievances, thus serving as theoretical justification for this methodological innovation. The second subsection outlines the process necessary to test the two-way causality with a two-stage simultaneous equations model. Finally, the third subsection addresses four potential concerns that arise with simultaneous equations modeling.
Why Two-Way Causality?
Fearon and Laitinâs (2003) study finds no evidence that large cultural divisions or broadly held grievances are associated with higher risks of civil war; these findings are based on their introduction of control variables such as per capita income and population into a single equation logit analysis. However, they do note that âintense grievances are produced by civil warâindeed, this is often a central objective of rebel strategyâ (88; emphasis in the original).3 This statement implies the existence of a two-way causal relationship between civil war and democracy (i.e., a proxy for grievances) in the sense that while countries that do not foster democratic governance tend to be more prone to civil war, civil war also negatively impacts the process of democratic governance in those countries by exacerbating the grievances held by minority groups that are not upwardly mobile.
This study attempts to conceptualize a reciprocal relationship between civil war and democratic grievances. Noting that Fearon and Laitinâs study has already modeled a causal relationship that moves from grievances to civil war, the new model pays closer attention to the opposite causal relationship wherein civil war may contribute to further grievances. The outbreak of civil war usually results in a great number of military and civilian casualties, as well as enormous costs in property damage; civil wars often devastate national economies, leaving factories and bridges destroyed, livestock and natural resources pillaged, and the machinery of manufacture damaged beyond repair; following such economic devastation, we expect to see a proliferation of private and public grievances for which citizens demand compensation. Furthermore, because participants of civil violence value the exigencies of war over civil liberties, political rights, and the rule of law, the outbreak of civil war may severely curtail the effectiveness of democratic channels of conflict resolution, in which case the growing numbers of grievances find limited opportunity for redress through institutional outlets (Hurwitz 2008; Choi 2010b).
The attenuation of democratic governance emboldens dominant groups to protect their vested interests, potentially limiting or depriving ethnic, religious, and ideological minority groupsâ rights. Moreover, the death or displacement of a large number of ethnic or religious minority group members may drastically change the demographic composition of the nation in question. As a result, new political and social grievances or tensions related to issues of property, employment, wealth, and political power may be engendered within conflict-prone countries (Huntington 1996; Fox 2004; Bueno de Mesquita 2005). This line of reasoning is consistent with Kalyvasâs (2007, 430) observation that âthe [civil] war itself aggregates all kinds of cleavages from the most ideological to the most local.â This theoretical rationale undoubtedly warrants examination of the effect of the onset of civil war on grievances, which is a missing causal direction in Fearon and Laitinâs single equation logit analysis. In other words, the codependence of the outcome variable and each predictor warrants the examination undertaken by this study of the possibility for a two-way causal relationship between the onset of civil war and each of the three causal factors identified in Fearon and Laitinâs study.
How to Test the Two-Way Causality?
By replicating Fearon and Laitinâs (2003, 84) results from Models 1 and 2 in Table 1.1,4 this study explores the empirical implications of two-way causality with respect to political democracy and then applies the same method, in sequence, to ethnicity and religion. It is important to note that the analysis in this study is conducted under the implicit assumption that only one of the three variables of interest (i.e., democracy, ethnicity, and religion) is endogenous to the onset of civil war in each of the simultaneous equations models that is presented below, an assumption that is made due to the limitations of existing statistical software.5 This issue will require further examination and will be revisited at the end of the empirical results section.
For simultaneous equations modeling, this study turns to the two-stage probit least squares method that was developed by Keshk (2003) and has been applied in several existing studies (e.g., Keshk, Pollins, and Reuveny 2004; Hegre, Oneal, and Russett 2010; Thies 2010; Whang 2010). At the first stage, this study assumes democracy to be a function of the onset of civil war, per capita income, national material capabilities, and Islam. Because there is no formal model that predicts key determinants of democracy on theoretical grounds, the choice of these variables was made after reviewing existing empirical studies that, as a whole, find a causal connection between democracy and each of the four factors listed above; of the four, national material capabilities and Islam are included as instrumental variables.
In a system of M simultaneous equations, any one equation is identified if the number of exogenous variables excluded from that equation is greater than or equal to the total number of endogenous variables in that equation less one. Gujarati (2003, 753) stresses that âthe order condition is usually sufficient to ensure identifiability.â By the order condition, the democracy equation (and the civil war onset equation) is identified. The validity tests for the two instrumental variables are addressed in the empirical results section. Because finding valid instrumental variables is a difficult undertaking, and because the purpose of this study is not to explain as much variance in democracy as possible but to account for endogeneity, other control factors that might be included are not included in this research project.6 The inclusion of variables such as standard of living and education would mean that the empirical results reported below are not comparable to Fearon and Laitinâs study, due to shorter data points and a smaller sample size.7
The dependent variable, democracy, at the first stage is obtained from the original data collection of Fearon and Laitin. Their study uses the Polity IV data to measure the level of democratic governance (see Marshall and Jaggers 2007). Polity provides an eleven-point additive score for both democracies and autocracies (each score ranging from 0 to 10). Subtracting the autocracy score from the democracy score gives an overall polity score that ranges from full democracy (+10) to full autocracy (â10).8
As noted, this study expects to find that the onset of civil war is causally related to a deterioration in the quality of democracy and to an intensification of public and private grievances.9 It is easy to imagine how civil violence would be detrimental to democratic political systems. This is because civil conflict drastically weakens a governmentâs ability to protect the political rights and civil liberties of its citizenry, and because the challenges posed by active minority groups may disrupt democratic election processes and socioeconomic safety nets. For example, civil wars waged in the name of national security may end up breaking down the democratic tradition of the rule of law, thereby rendering the public less likely to seek resolution of political and economic grievances through fair and impartial justice systems (Hurwitz 2008; Choi 2010b). As a result, civil wars exacerbate, rather than alleviate, grievances. For this analysis, the onset of civil war variable is coded as 1 when a civil war occurs and as 0 otherwise; the relevant data are derived from Fearon and Laitinâs study.
It is well known that ...