Although logistical and power-related conditions—such as low state-level per capita income, weak state institutions, and peripheral and inaccessible territory—enjoy near-consensus support as explanations of civil war onset, most of the contemporary literature regards explanations rooted in political and economic grievances with suspicion (Blattman and Miguel Reference Blattman and Miguel2010). In fact, the debate over the status of grievances in such explanations dates back at least to the 1960s, with the introduction of relative deprivation theory. Inspired by psychological theories of conflict, Gurr (Reference Gurr1970) and his colleagues argued that economic and other types of inequality increase the risk of internal strife through frustrated expectations. In contrast, today's most influential quantitative studies of civil war give short shrift to grievance-based accounts, based on reports that unequal individual wealth distributions have no statistically distinguishable relationship to internal conflict (e.g., Collier and Hoeffler Reference Collier and Hoeffler2004; Fearon and Laitin Reference Fearon and Laitin2003).

Yet, despite these alleged nonfindings, the debate over grievances is far from dead. Indeed, inequality continues to occupy a prominent place in the qualitative literature on civil wars and has repeatedly been linked to conflict processes (Sambanis Reference Sambanis, Collier and Sambanis2005, 323; Stewart Reference Stewart2008b; Wood Reference Wood2003). Moreover, in the last few years, some quantitative studies have started to appear that argue that the current literature's failure to connect distributional asymmetries with conflict behavior may actually be due to inappropriate conceptualization and imperfect measurements, rather than reflecting a fundamental absence of any causal effect (Østby Reference Østby2008b; see also Stewart Reference Stewart2008b).

Also relying on quantitative evidence, we join these recent contributions in shifting the explanatory focus from individualist to group-level accounts of inequality and conflict. Because formidable problems of data availability associated with the uneven coverage and comparability of surveys have stood in the way of assessing such “horizontal inequalities” (HIs), most scholars have had to content themselves with selective case studies or statistical samples restricted to particular world regions.

To overcome these difficulties, we combine our newly geocoded data on politically relevant ethnic groups’ settlement areas with Nordhaus's (Reference Nordhaus2006) spatial wealth measures, both with global coverage. Based on this novel strategy, we present the first truly worldwide comparison of horizontal inequality and ethnonationalist civil wars. Controlling for political power access, we show that both advanced and backward ethnic groups are more likely to experience such conflict than groups whose wealth lies closer to the national average. Moreover, in agreement with a broad conception of horizontal inequalities, we find that both political and economic inequalities contribute to civil war. Extensive sensitivity analysis confirms that our findings are robust to various model specifications, alternative inequality measures, and sample definitions.

We proceed as follows: First, we review the extensive literature on inequality and conflict before narrowing down the scope to our own theoretical framework, which connects HIs with collective violence through grievance-based mechanisms. Based on this approach, we derive our main hypotheses on the effect of HIs on civil war onset. To test the hypotheses, we then introduce the datasets and describe our spatial method of wealth estimation in detail, including how to use the contours of the ethnic groups’ settlement areas as “cookie cutters,” which allows us to extract the relevant wealth estimates from the spatial wealth map. After specifying the variables used in our empirical analysis, we then expose the hypotheses to systematic tests in the main models, followed by another section that assesses the robustness of our findings. Finally, concluding at least tentatively that both economic and political inequality at the group level increase the risk of ethnonationalist civil war, we argue that the civil war literature's tendency to downplay the importance of grievances as a source of internal conflict is both premature and misguided.

INEQUALITY IN THE CIVIL WAR LITERATURE

Intimately related to issues of power and wealth distribution in society, inequality plays a central role in classical theories of conflict. In an influential article, Davies (Reference Davies1962) argued that revolutions were motivated by frustration resulting from an evolving gap between individual aspirations and actual economic status. Also adopting an explicitly psychological perspective, Gurr's (Reference Gurr1970) well-known theory of relative deprivation characterizes various types of collective violence as reactions to frustrations stemming from unfulfilled aspirations, usually related to material well-being (see review in Brush Reference Brush1996). Such a perspective differs radically from earlier sociological theories of mob behavior that explained collective violence as a societal pathology (e.g., Le Bon Reference Le Bon1913). Instead, relative deprivation theorists argue that individuals’ widespread discontent with their social situation triggers conflict, especially where modernization fuels a “revolution of rising expectations” (Davies Reference Davies1962).

Although related indirectly to inequality through this psychological mechanism, relative deprivation theory does not explicitly focus on interpersonal or intergroup wealth comparisons (Hogg and Abrams Reference Hogg and Abrams1988; Regan and Norton Reference Regan and Norton2005; cf. Gurr and Duvall Reference Gurr and Duvall1973). Other theories adopt a structural perspective, linking various types of inequality explicitly to structural imbalances in society, such as uneven income or land distribution (Acemoğlu and Robinson Reference Acemoğlu and Robinson2005; Muller Reference Muller1985; Muller and Seligson Reference Muller and Seligson1987; Russett Reference Russett1964). Partly under the inspiration of Marxist principles, the literature on peasant rebellions explains violent collective action as a response to unequal wealth allocation (Moore Reference Moore1966; Scott Reference Scott1976). Frustrated with their lot, the peasant masses and other underprivileged groups are expected to take up arms as a way to seize power and redistribute wealth in their favor.

Relative deprivation theory remains perhaps the most prominent explanation that connects grievances with conflict, but has a very mixed record as regards empirical evidence (Brush Reference Brush1996; Oberschall Reference Oberschall1978). Early on, the theory attracted criticism from Snyder and Tilly (Reference Snyder and Tilly1972), who argued that opportunity-based mobilization rather than grievances causes internal conflict and revolutions. Contending that all societies contain a number of aggrieved and frustrated individuals, they did not think “there is any general connection between collective violence and hardship such that an observer could predict one from the other” (Tilly 1972, 520; see also Skocpol Reference Skocpol1979; Tilly Reference Tilly1978). Along similar lines, a series of studies challenged the results pertaining to income inequality, which was usually seen as closely connected to the notion of relative deprivation (see, e.g., Weede Reference Weede1987). By the end of the 1980s, the debate remained unresolved, with virtually all possible causal connections—negative, positive, curvilinear, or none—being represented in the literature (Lichbach Reference Lichbach1989).

As the end of the Cold War brought with it a new wave of conflict, most of which was ethnonationally motivated, Gurr (Reference Gurr1993, Reference Gurr2000a, Reference Gurr2000b) extended his previous theory and began to study ethnic minorities’ reactions to state-imposed disadvantages and discrimination. In agreement with Horowitz’ (Reference Horowitz1985) seminal study of ethnic groups in conflict, Gurr found that ethnic grievances contributed indirectly to collective violence through ethnic mobilization.

In contrast, the contemporary civil war literature has been pioneered by scholars who take issue with such reasoning. Positioning themselves explicitly against grievance-based theories in political science and sociology, Collier and Hoeffler (Reference Collier and Hoeffler2004) follow in the footsteps of earlier critics of relative deprivation (although without referring to them explicitly). Very much as Snyder and Tilly had done three decades earlier, Collier and Hoeffler (Reference Collier and Hoeffler2004, 564) point to the ubiquity of frustration around the world, asserting that this fact deprives the theory of explanatory value: “Misperceptions of grievances may be very common: All societies may have groups with exaggerated grievances. In this case, as with greed-rebellion, motive would not explain the incidence of rebellion.” Explicitly classifying inequalities as grievance-related indicators, these authors rely on the Gini coefficient to measure the income distribution among individuals. Having found no statistical effect for this and other hardship proxies, Collier and Hoeffler feel vindicated in their wholesale rejection of grievances and inequality as causes of civil war. In his best-selling book The Bottom Billion, Collier (Reference Collier2007, 18) confirms these doubts:

In another influential study that stresses political and institutional causes of civil war, Fearon and Laitin (Reference Fearon and Laitin2003) present findings that cast doubt on ethnic and political grievances as explanations of conflict onset. Like Collier and Hoeffler, Fearon and Laitin rely on a series of individual-level statistical proxies, including the Gini coefficient, which provide no evidence of economic inequality increasing the risk of conflict. Summing up these and other studies, Laitin (Reference Laitin2007, 23) concludes that

Despite these negative findings, a number of scholars find the rejection of grievances and inequalities premature. Although the poor quality of distributional economic data within and across countries should perhaps be reason for pause by itself,Footnote ² the main reason that grievance-based arguments cannot be so easily rejected is that these studies largely miss their theoretical target. Insisting that conflict-inducing inequality cannot be reduced to household-level measures of income distribution such as the Gini coefficient, Cramer (Reference Cramer2003) calls for an alternative that is explicitly relational and theoretically grounded.Footnote ³ In the concluding chapter of a two-volume compilation of case studies testing Collier and Hoeffler's (Reference Collier and Hoeffler2004) model, Sambanis (Reference Sambanis, Collier and Sambanis2005) draws the same inference. Noting that there is a major discrepancy between the quantitative nonfinding and the repeated references to inequality in the case studies, Sambanis (Reference Sambanis, Collier and Sambanis2005, 324) considers a number of explanations, including problems relating to interpretation and sampling of case evidence, as well as the fundamental issue of aggregation level:

A more promising way to capture the link between uneven wealth distributions and conflict has been proposed by Stewart (Reference Stewart2008b) and her colleagues, who contrast vertical, or individual-level inequalities, with horizontal inequalities. Defining the latter as “inequalities in economic, social or political dimensions or cultural status between culturally defined groups,” Stewart (Reference Stewart and Houndmills2008a, 3) argues that to a large extent, scholars have failed to find evidence of inequality's war-causing effect because of their reliance on individualist, rather than group-based, measures of income and power differences:

Following the lead of Horowitz and Gurr, Stewart (Reference Stewart and Houndmills2008a) conceptualizes horizontal inequality broadly by considering political, economic, social, and cultural dimensions explicitly. Political HIs entail blocked or limited access to central decision-making authority within the state. The economic dimension taps the distribution of wealth among households. Social HI measures primarily groups’ uneven social access, for example, in terms of education and societal status. Finally, the cultural aspect captures group-level inequalities with respect to cultural policies and symbols, including national holidays and religious rights.

Recognizing the difficulties of measuring HIs, Stewart's team has so far primarily relied on case studies rather than large-N comparisons. The picture that emerges from this research suggests that both disadvantaged and advanced groups have a higher likelihood of getting involved in internal conflict than groups closer to the country average (Stewart Reference Stewart2008b). Yet some quantitative researchers have attempted to generalize from the case studies to a larger set of countries. In a pioneering statistical test, Barrows (Reference Barrows1976) detected an influence of group-level differences on conflict in sub-Saharan Africa. Relying on household surveys conducted in 39 developing countries, Østby (Reference Østby2008b) finds evidence that social horizontal inequality causes civil war, although the economic dimension appears to be weaker (see also Østby Reference Østby and Houndmills2008a). In a follow-up study based on geocoded conflict and survey data from sub-Saharan Africa, Østby, Nordås, and Rød (Reference Østby, Nordås and Rød2009) reach firmer conclusions, showing that both economic and social group-level differences are likely drivers of conflict behavior.

THEORIZING HORIZONTAL INEQUALITIES

We now turn to our own account of inequality and conflict. The starting point of our approach to ethnonationalist warfare is the realization that ethnic groups find themselves in radically different situations for various historical reasons. Whereas some ethnic groups came out on top of the geopolitical game, others were conquered early on, and therefore lost out in the competition for wealth and influence. Moreover, the uneven spread of nationalism delayed mass-level political mobilization in many parts of the world, thus creating differences in both economic and political development that were often exploited by alien rulers (Gellner Reference Gellner1964). As argued by Tilly (Reference Tilly1999), nationalism can be thought of as a case of “categorical inequality” because

Adopting Stewart's definition of HI, we focus on the political and economic dimensions of HI.Footnote ⁴ Although these two types of inequality are often likely to be related, we follow Stewart in treating the distribution of power and of wealth as conceptually separate components. Previous quantitative research has examined the role of political exclusion of ethnic groups on the risk of civil war, but has not considered explicitly the economic aspect of exclusion and the relationship between the two (see e.g., Cederman, Buhaug, and Rød Reference Cederman, Buhaug and Rød2009; Cederman and Girardin Reference Cederman and Girardin2007; Cederman, Wimmer, and Min Reference Cederman, Wimmer and Min2010; though see Gurr Reference Gurr1993, Reference Gurr2000b). By contrast, some critics of the role of ethnicity, such as Woodward (Reference Woodward1995), argue that alleged ethnic conflicts are really driven by underlying economic inequalities that lead ethnic identities to become politicized. A strong version of this argument would hold that ethnic political inequalities are irrelevant in the absence of economic inequalities.

Building on our previous work, we view HIs as structural asymmetries that make ethnonationalist civil war more likely and adopt an indirect research strategy that explains the effect of inequality by postulating a set of causal mechanisms. To close the gap between the structural background conditions and behavioral patterns, we propose intermediate analytical steps.Footnote ⁵ First, we postulate that objective political and economic asymmetries can be transformed into grievances through a process of group comparison driven by collective emotions. Second, we argue that such grievances trigger violent collective action through a process of group mobilization.

DERIVING TESTABLE HYPOTHESES ON HORIZONTAL INEQUALITIES AND CIVIL WAR

Having postulated our causal mechanisms, we now return to the macro-level to perform the actual empirical analysis. It is in principle possible to measure directly grievances (Petersen Reference Petersen2002) and to trace mobilization processes (e.g., Beissinger Reference Beissinger2002), but such detailed analysis is beyond the scope of the current study. As will become clear later, collecting and evaluating structural data on economic HIs is a major challenge in its own right.

What are the observable implications of our analytical framework? The first, and most obvious, hypothesis expects a positive effect of economic HIs on civil war onset. If the causal chain operates as we have postulated, there should be a statistically discernible signal indicating that ethnic groups with GDP per capita far from a country's average have a higher risk of experiencing conflict:

However, as we have seen, an uneven wealth distribution is not the only possible type of structural asymmetry. Drawing on Stewart's multidimensional conceptualization of HIs, we hypothesize that both economic and political HIs contribute jointly to the outbreak of civil war. Even controlling for political HIs, such as groups’ exclusion from political power, income inequalities among ethnic groups should increase the risk of civil war. These theoretical expectations dovetail with Stewart's (Reference Stewart and Houndmills2008a, 18) hypothesis that “political mobilization is especially likely when there are consistent HIs, that is both political and economic HIs run in the same direction.” Based on statistical evidence from sub-Saharan Africa, Østby (Reference Østby2008b) finds support for a strong effect of interaction between interregional asset inequality and political exclusion. Case studies of Côte d'Ivoire (Langer Reference Langer2005) and Nepal (Murshed and Gates Reference Murshed and Gates2005) confirm this finding (see also Stewart, Brown, and Langer Reference Stewart, Brown, Langer and Houndmills2008, 289–90; cf. Hegre, Østby, and Raleigh Reference Hegre, Østby and Raleigh2009).Footnote ¹⁰ Our second hypothesis summarizes these theoretical expectations:

So far, we have not differentiated between advanced and backward groups’ conflict-proneness. As Horowitz (Reference Horowitz1985) explains, arguments can be advanced for both types of economic HIs leading to a higher risk of conflict. Poorer groups, especially those residing in backward and peripheral regions, often desire to break away from the cores of their countries regardless of the cost, because they perceive themselves to be systematically disadvantaged compared to their wealthier compatriots in terms of economic development and distribution of public goods. Perceptions of disadvantage also characterize members of some relatively wealthy groups, especially if they feel that state-level redistribution denies them the fruits of their success: “Advantaged regions usually generate more income and contribute more revenue to the treasury of the undivided state than they receive. They believe that they are subsidizing poorer regions” (Horowitz Reference Horowitz1985, 249–50). Because these groups have more to lose, and are sometimes demographically represented outside their original settlement areas, however, they can be more cost-sensitive as regards secession, but such cases do occur, as illustrated by Slovenia and Croatia (cf. Gourevitch Reference Gourevitch1979). Nevertheless, there is no reason to assume that the effect of group inequality is perfectly symmetric around relative equality. Remaining agnostic as to the relative frequency of HIs in either direction, we therefore submit these arguments to separate tests by dividing H1 into two subhypotheses:

GLOBAL DATA ON HORIZONTAL INEQUALITIES AND OTHER DIMENSIONS

Our theoretical expectations must now be confronted with empirical evidence. As we have seen, data availability constitutes a major stumbling block in studies of inequality and conflict. So far, virtually all existing statistical studies of HIs have used survey data on economic welfare by households, as collected by the Demographic and Health Surveys (DHS) project.Footnote ¹¹ Although these data offer a relatively direct measure of well-being, and are therefore useful as grievance indicators, the information source and survey approach are associated with a number of practical limitations. The DHS project is limited to a selection primarily of developing countries, and information on ethnic affiliation or the geographic location of households is only available in some of the surveys.Footnote ¹² Although the surveys are constructed to be nationally representative, the number of responses for an ethnic group or location may often be very low, and there is no guarantee that the samples will be representative. Finally, survey data are subject to a host of potential response biases, both conscious and unconscious. For example, aggregate responses from surveys may well be systematically biased against finding evidence of inequality if poorer individuals overstate their assets and richer individuals consistently understate theirs. In sum, survey data may be helpful for many purposes, but the DHS data do not provide a plausible alternative for evaluating the role of horizontal inequalities on a global basis.

Given these difficulties, it makes sense to consider spatial datasets as an alternative to survey-based methods. In fact, the only broadly available cross-national data source on variation in wealth within countries is the G-Econ data, developed by Nordhaus (Reference Nordhaus2006; see also Nordhaus and Chen Reference Nordhaus and Chen2009).Footnote ¹³ The G-Econ dataset tries to assemble the best available data on local economic activity within countries for geographical grid cells, and convert these to comparable figures in purchasing power parity to allow meaningful comparisons. The resolution of the spatially explicit data set is 1° grid cells. The data are constructed from a variety of sources, including regional gross product data for the lowest available political subdivision, estimates of regional income by industry, and estimates of rural population and agricultural income. The specific methodologies differ by countries and data availability (see Nordhaus et al. Reference Nordhaus, Azam, Corderi, Hood, Victor, Mohammed, Miltner and Weiss2006 for a detailed discussion). The database has global coverage, but the temporal scope is limited to a single year, 1990.Footnote ¹⁴ We therefore restrict our analysis to the post–Cold War period, although we present supplementary results extending back to 1946 in the sensitivity analysis that follows. Because it is well known that relative inequality, as opposed to absolute wealth, is characterized by considerable inertia, these assumptions would seem plausible (Stewart and Langer Reference Stewart, Langer and Houndmills2008; Tilly Reference Tilly1999).Footnote ¹⁵

Despite their relatively broad coverage, there are a number of disadvantages to the Nordhaus data for testing propositions on HIs. Any measure of the value of economic production is strictly speaking a “flow” measure and hence an imperfect proxy for the “stock” of wealth, although this criticism obviously applies with equal force to national-level productivity measures. Because the quality varies considerably across countries, the data are likely to understate the extent of inequality in countries with poor data coverage. Indeed, in some countries the official data may be of such poor quality that the variance is suppressed and accuracies of survey reports may be questionable. We will return to these issues in the section on sensitivity analyses.

Based on the G-Econ data, Buhaug et al. (n.d.) present the first global results on the relationship between spatial inequalities and civil war violence. However, their research design focuses on local measures of inequality across geographic grid cells and the specific locations where conflict first breaks out, and does not capture group-level participation or wealth differences. Another useful approach estimates the wealth of regional subunits of states (Sambanis and Milanovic Reference Sambanis and Milanovic2009). A more direct assessment of HIs requires geocoded data on ethnic groups, and in view of H2, also information about their access to executive power. Fortunately, the Ethnic Power Relations dataset (EPR), together with its recent geocoded extension, GeoEPR, fulfills these requirements.

The EPR dataset identifies all politically relevant ethnic groups around the world and measures how access to state power differs among them in all years from 1946 to 2005 (Cederman, Wimmer, and Min Reference Cederman, Wimmer and Min2010). Based on an online expert survey, the sample includes 733 politically relevant ethnic groups in 155 sovereign states.Footnote ¹⁶ The coding rules define as politically relevant all ethnic groups for which at least one political organization exists that promotes an ethnically oriented agenda in the national political arena, or ethnic groups that are subject to political discrimination. This dataset improves significantly on previous efforts to code ethnic groups’ access to power, such as the Minorities at Risk (MAR) dataset (Gurr Reference Gurr1993), which restricts the sample to mobilized and/or discriminated-against minorities and thus largely overlooks the ethnopolitical constellation of power at the center, and Cederman and Girardin (Reference Cederman and Girardin2007), who rely on preliminary, static measures of the political status of ethnic groups and limit their sample to Eurasia and North Africa.

Because the politically relevant groups and their access to political power may change over time, the EPR dataset provides separate coding for subperiods from 1946 to 2005. For each such time period, the demographic size and access to power enjoyed by representatives of an ethnic group are specified. Focusing on executive power only, i.e., representation in the presidency, the cabinet, and senior posts in the administration, including the army, the coding rules categorize all politically relevant ethnic groups according to whether (1) their representatives enjoyed absolute power through monopoly or a dominant position in the executive branch,Footnote ¹⁷ (2) they shared power with other groups in a junior or senior role,Footnote ¹⁸ or (3) they were excluded altogether from executive decision making but enjoyed regional or separatist autonomy, or were powerless or discriminated against.Footnote ¹⁹ In our analysis that follows, we drop category 1, because according to our conflict coding, dominant and monopoly groups cannot by definition stage rebellions against themselves, and base the dummy variable of exclusion on the difference between categories 2 and 3.

To obtain spatial estimates of economic performance for EPR groups based on the Nordhaus grid, we need information on their settlement areas or regions. Because this overlay operation requires data on the precise extent of these regions rather than a simple textual description, existing datasets such as Minorities at Risk (Gurr Reference Gurr1993) are insufficient. We therefore rely on the recently completed GeoEPR dataset, a comprehensive geocoded version of the EPR groups (Wucherpfennig et al. n.d.). GeoEPR provides two types of information about ethnic groups. First, for each group in EPR, the dataset categorizes the type of settlement pattern, distinguishing between regional, urban, and migrant groups (plus mixed categories). For all groups with regional bases, GeoEPR represents the settlement area of a group as a polygon (or a set of polygons, if there is more than one group region in a country). In contrast to earlier geocoding attempts, GeoEPR also tracks major changes in the settlement pattern of a group over time, including those resulting from ethnic cleansing.

Our analysis requires conflict coding at the group level. Because groups as a whole typically do not participate in conflict (e.g., Kalyvas Reference Kalyvas2006), we use a refined procedure that codes whether a group has links to a rebel organization that was actively involved in fighting. These new data represent an improvement on previous group-level conflict coding used with the EPR data (cf. Cederman, Wimmer, and Min Reference Cederman, Wimmer and Min2010). More precisely, we code our dependent variable as “1” if a rebel organization expresses its political aims (at least partly) in the name of the group and a significant number of members of the group were participating in the conflict. For a full sample of rebel groups and their conflict involvement, we rely on the Non-State Actors dataset (Cunningham, Gleditsch, and Salehyan Reference Cunningham, Gleditsch and Salehyan2009) that identifies the fighting organizations involved in civil wars (according to the Uppsala/PRIO Armed Conflicts Data, see Gleditsch et al. Reference Gleditsch, Wallensteen, Eriksson, Sollenberg and Strand2002). The link between these organizations and our EPR groups is provided by NSA2EPR, a new conflict resource that identifies organizations fighting for, and recruiting from, particular EPR groups. We provide a list of the conflict onset cases in the article.

DERIVING OPERATIONAL MEASURES

The G-Econ data allow deriving ethnic group–specific measures of wealth by overlaying polygons indicating group settlement areas with the cells in the Nordhaus data. Dividing the total sum of the economic production in the settlement area by the group's population size enables us to derive group-specific measures of per capita economic production, which can be compared to either the nationwide per capita product or the per capita product of privileged groups.

A visual illustration helps explain the estimation of group GDP from Nordhaus data. Figure 1 shows the G-Econ estimates as gray cells; darker shadings indicate wealthier cells. The map shows the spatial variation in wealth across the Yugoslav federation in 1990, based on the CShapes dataset for the historical boundaries (Weidmann, Kuse, and Gleditsch Reference Weidmann, Kuse and Gleditsch2010). Relatively wealthy pockets appear primarily in the northwest of the country, in the constituent republics of Slovenia and Croatia. Compared to the other parts of the country, Serbia shows up as a generally poor region.

FIGURE 1.

G-Econ Cells for Yugoslavia, Overlaid with GeoEPR Group Polygons for Slovenes and Albanians

Note: The darker the shading, the wealthier the cells.

Together with the settlement areas of GeoEPR, the G-Econ data can now be used to estimate group wealth spatially. Figure 1 also shows the settlement regions for the Slovenes and the Albanians. Using techniques similar to those pioneered in Cederman, Buhaug, and Rød (Reference Cederman, Buhaug and Rød2009), we derive an indicator of group wealth by summing up the (population-weighted) proportions of the Nordhaus cells covered by a group.Footnote ²⁰ For example, as a result of this procedure, the Slovenes get a high score, because their settlement region is located in the rich parts of Yugoslavia. Figure 2 shows horizontal inequality for Yugoslavia, measured as the ratio of the group's GDP per capita estimate to the average value for the entire country, depicting wealthier groups in darker shades and poorer ones in brighter shades. Slovenes and Croats receive high scores, but the opposite is true for the Albanians in Kosovo, which are among the poorest groups in the country.

FIGURE 2.

Result of Spatial Wealth Estimation for Groups in Yugoslavia

As a further illustration, Figure 3 shows the same information for the Sudan. Unsurprisingly, the southern and western groups, the latter including the Fur, emerge as the most impoverished in that state. Extending the comparison to Myanmar, we also illustrate the limitations of our spatial approach (see Figure 4). Despite considerable wealth discrepancies between peripheral and central areas, the Nordhaus data exhibit very limited variation, because of underlying data quality issues. To the extent that similar measurement problems afflict other countries that experienced conflict, we can expect the effect of inequality to be underestimated by our study.

FIGURE 3.

Result of Spatial Wealth Estimation for Groups in the Sudan

FIGURE 4.

Result of Spatial Wealth Estimation for Groups in Myanmar

As explained by Mancini, Stewart, and Brown (Reference Mancini, Stewart, Brown and Houndmills2008), there are many different ways to operationalize horizontal inequalities, most of which apply to entire countries. In this article, we use two group-level measures of inequality, namely a symmetric logged form (see H1) and an asymmetric, nonlogged form (see H1a, 1b). The former indicator defines inequality as the square of the logarithmized ratio between g, the GDP per capita of the ethnic group, and G, the average GDP per capita of all groups in the country:

This definition captures deviations from the country average symmetrically and is zero for groups at the country average. As a complement to this symmetric indicator, we also measure inequality asymmetrically with two variables that correspond to groups that are poorer and wealthier than the country average, respectively:

This operationalization guarantees that deviations from the country mean are always positive numbers greater than one. For example, a group that is twice as wealthy as the average has low_ratio = 0 and high_ratio = 2, and a group that is three times poorer has low_ratio = 3 and high_ratio = 0. See Table 1, which offers descriptive statistics for these indicators and all other independent variables used in the analysis that follows.Footnote ²¹

TABLE 1.

Summary Statistics for Independent Variables

We use the nonspatial EPR dataset to derive variables capturing political HIs and group sizes. As we have seen earlier, the EPR dataset provides a time-varying indicator for groups’ exclusion from central power. In addition, we measure the group's demographic power balance with the ethnic group(s) in power (EGIP) as its share of the dyadic population.Footnote ²² We use a combination of the linear and squared terms to capture the logic of bargaining theory, according to which intermediate power levels are the most conflict-prone, given that weaker groups stand no chance of prevailing in contests, whereas stronger ones do not necessarily have to resort to violence to get concessions (see Wucherpfennig Reference Wucherpfennig2011).

At the country level, we control for GDP per capita, based on nonspatial, time-varying statistics drawn from the Penn World Tables and World Bank sources, and num_excl, the total number of excluded groups in the country, as defined by EPR. Both measures should have a negative impact on conflict probability. A large number of studies find a negative association between national GDP per capita and civil war onset (see Hegre and Sambanis Reference Hegre and Sambanis2006). According to Walter's (Reference Walter2006) strategic argument, the num_excl variable can be expected to be negatively related to the risk of conflict, because governments facing many ethnic groups fear domino effects and will thus be less willing to make concessions to single groups, as illustrated by Moscow's hard line in dealing with the Chechens’ claims. This firmness can be expected to deter other groups from challenging the government. In addition, we also control for the calendar year, because we anticipate a declining trend in terms of conflict probability during the Cold War, thanks to benign effects in the international environment such as peaceful international norms and institutions (Gurr Reference Gurr2000b). Finally, the models also contain nonparametric corrections for temporal dependence based on the peace_years variable, which measures the number of years a group has lived in peace, as proposed by Beck, Katz, and Tucker (Reference Beck, Katz and Tucker1998).Footnote ²³

EMPIRICAL ANALYSIS

We are now ready to present the results. Given the limited temporal availability of inequality data, we restrict the sample to group years after the Cold War, from 1991 through 2005. All groups represented in GeoEPR are included, except the dispersed ones that cover their respective countries’ entire territory. This leaves us with about 450 groups per year or a total of 6,438 group years, with only 52 conflict onsets (see the supplemental online Appendix available at http://www.journals.cambridge.org/psr2011010). Unless otherwise stated, our analysis therefore relies on rare-events logit models. We compensate for country-level dependencies by estimating clustered standard errors.

Table 2 presents the main results. Our starting point is Model 1, which subjects the inequality hypothesis H1 to a first test based on the lineq2 variable. The result is both substantively and statistically significant, suggesting that groups with wealth levels far from the country average are indeed more likely to experience civil war.Footnote ²⁴ Moreover, the other variables behave as expected. The coefficients of the variables measuring the demographic power balance point in the right direction, but fail to reach statistical significance. At the country level, both GDP per capita and calendar year have strongly negative effects on the probability of conflict, as theoretically expected. The coefficient for the variable capturing the number of ethnic groups is also negative, but nowhere near statistical significance at conventional levels. Finally, the temporal controls do not seem to make much of a difference, but are retained for comparative purposes.

TABLE 2.

Explaining Onset of Group-level Conflict

Notes: Robust, country-clustered standard errors in parentheses.

***p < .001. **p < .01. *p < .05.

To improve the precision of our inequality measures, Model 2 discards ethnic groups with a spatially estimated population less than 500,000. Although this censoring limits the number of group-year observations to 3,967 and the conflict onsets to 42, we prefer to rely on this specification.Footnote ²⁵ Our spatial method becomes unreliable for small population sizes, primarily because of the low resolution of the G-Econ data and the limited precision of the population estimates for tiny groups.Footnote ²⁶ Consequently, the group-size restriction almost triples the inequality coefficient reported in Model 2 without affecting the size of the standard error. Except for this important change, there are few other surprises, except that the coefficient for the num_excl variable now becomes significant.

Having considered H1, we now test H2, which postulates that both economic and political HIs increase the risk of internal conflict. Retaining the size restriction of the previous specification, Model 3 introduces a dummy variable for excluded groups that has a strong and statistically discernible impact on onset likelihood. However, this does not undermine the results with regard to economic inequality. This is an important result that strengthens our confidence that different types of grievances operate together and enables us to rule out the possibility that economic inequality could be an artifact of groups’ access to executive power (and vice versa). Moreover, it is clear that the addition of the exclusion dummy either preserves or increases the effect of the other variables.Footnote ²⁷ We illustrate the effect of inequality in Figure 5, which indicates how the predicted probability of conflict increases as the group's wealth level deviates from the country average in both directions, for a median observation for the post-1990 sample, for excluded and included groups, respectively. The figure reveals that, other things being equal, excluded groups (see the solid curve) are much more likely to experience conflict than included ones (see the dashed curve). However, the increases in risk from greater relative deviations in economic wealth are also substantial, especially for an excluded group, consistent with our argument that both political and economic grievances increase the risk of conflict.

FIGURE 5.

Predicted Effect of Horizontal Inequality on Probability of Civil War (See Model 3 of Table 2)

Note: The solid line corresponds to a median profile for excluded groups; the dashed one to the median profile for included groups. The horizontal axis is on a log scale, with tick marks indicating specific values of the g/G ratio.

So far we have made the simplifying, but questionable, assumption that the effect of inequality is the same for groups below and above the country's average level of wealth. We therefore need to test Hypotheses H1a and H1b separately. As a way to do so, Model 4 relaxes the assumption of a parabolic functional form by relying on a smoothed, spline-based local regression specification with three knots. Even if not perfectly symmetric, the estimated functional form shown in Figure 6 tells us that both relatively poorer and wealthier groups are more likely to experience civil war, thus confirming both H1a and H1b. The error bands are relatively broad but clearly separate from zero, at least for the poorer groups.Footnote ²⁸

FIGURE 6.

Effect of Logged Horizontal Inequality on the Probability of Civil War (See Model 4 of Table 2)

Further increasing our confidence in the separate effects, Model 5 uses the two linear ratio indicators low_ratio and high_ratio while still controlling for political exclusion. The results are strongly positive for both directions of inequality, lending further support to H1a and H1b. With this model specification, the demographic measures of power balance also become significant in the expected direction. Furthermore, the impact of the control variables at the country level also becomes stronger compared to Model 3.

SENSITIVITY ANALYSIS

Because the number of onsets is quite limited, our findings need to be treated with some caution. Therefore, the current section presents the results of several robustness checks. Table 3 introduces four models that evaluate different sensitivity aspects. As noted earlier, the GDP per capita data used in the article include all economic activity, including exploitation of raw material. This in itself does not necessarily undermine a grievance interpretation, especially where disadvantaged groups are deprived of the riches extracted from their homelands (Østby, Nordås, and Rød Reference Østby, Nordås and Rød2009, 307; Stewart, Brown and Langer Reference Stewart, Brown, Langer and Houndmills2008, 346). Examples include the Aceh in Indonesia (Ross Reference Ross, Collier and Sambanis2005) and the Ijaw in Nigeria (Osaghae Reference Osaghae2008). Although our data sources do not enable us to separate a pure grievance effect from the low_ratio and high_ratio variables, Model 6 reports the coefficients of these variables, excluding wealth deriving from oil production, based on data from Nordhaus's dataset. The results tell us that the conflict-proneness of wealthy groups is not primarily driven by oil income. In fact, once oil income has been subtracted, the coefficient of high_ratio increases considerably compared to that in Model 5.Footnote ²⁹

TABLE 3.

Explaining the Onset of Group-level Conflict (Sensitivity Analysis)

Notes: Robust, country-clustered standard errors in parentheses.

***p < .001. **p < .01. *p < .05.

We continue the sensitivity analysis by extending the sample back to 1946. Although we prefer to base our conclusions on the period explicitly following the point of measurement, this backward projection of inequality values is not completely unreasonable, because of the already noted durability of inequality. In any case, this rough robustness test tells us whether the findings should be seen as an artifact of the post–Cold War sample or whether they can be potentially generalized beyond this time period. Extending the sample of Model 3 to the entire post-WWII period, Model 7 reveals that the main inequality result holds in this case as well.Footnote ³⁰

Because the Nordhaus data represent a temporal snapshot, there are also reasons to be concerned that endogeneity could have distorted the results. We therefore discard all groups that were experiencing ongoing conflict in 1990, which could have affected these groups’ relative wealth estimates. The results of this test are presented in Model 8, which indicates that the inequality effect reported in Model 3 cannot be dismissed as being driven by endogenous influences. In fact, the coefficient of lineq2 hardly changes at all and remains discernibly greater than zero, and the other variables behave in a fashion very similar to that in the original model.Footnote ³¹

Finally, to rule out potentially distorting temporal dependencies, we present a static model that uses the group list in 1992 as the basis of comparison, while coding the dependent variable as one if there is any group-level onset from 1992 through 2005.Footnote ³² Again, we find that the inequality result in Model 8 becomes even stronger than in the dynamic models. As would be expected, however, the exclusion dummy loses significance, because this measure is sensitive to shifts of groups’ power access that cannot be captured by the static analysis. We find no positive result for the power balance, but the two remaining country-level variables retain a statistically significant negative effect.

Obviously, this set of sensitivity tests does not exhaust all possibilities. We refer the reader to our supplemental online Appendix for detailed results, but report here on the most important findings. For example, disaggregation of the dependent variable into territorial and governmental conflict, based on the contested incompatibility, shows that it is really the former type that is influenced by our inequality indicators. This is to be expected, because our method is explicitly geography-dependent, and it therefore cannot be excluded that HIs trigger governmental onsets. However, confirming this would require complementing the current measurements with nonspatial information, for example, from surveys.Footnote ³³

We also conducted several sensitivity tests by removing extreme values, both among the (resource-)richest groups, such as the Ijaw in Nigeria and Arabs in Iran, and among the relatively poorest groups, such as the Chechens in Russia. Even if discarding these groups separately or together weakens the results, the inequality effect remains statistically significant and strong. We also ran models controlling for world regions, which further confirm the robustness of our findings. Furthermore, another test that entails replacing the dummy variable for excluded groups with the full set of EPR categories also makes little difference to the impact of inequality.