A1. Conflict over Ownership of the State

The state is at the very center of nationalist conflict. In a competition for state goods, ownership of the state is the ultimate prize (Wimmer 2002). In addition to securing material resources, the winner acquires crucial symbolic powers and international recognition. Domestically, the state's importance grew as its role became increasingly intrusive in citizens' lives. As state control has increased, opportunities have arisen to favor a specific ethnic group or groups through an ethnicized bureaucracy in terms of public schooling, language laws, and religious regulations. In this way, ethnicity started to serve as a formidable instrument of social and political exclusion.

A2. Conflict at the Level of Entire Groups

In a pioneering contribution to the theory of nationalism, Gellner (1964) explained how individuals in the modern world identify themselves, and classify others, according to cultural categories. This stands in stark contrast to the pre-nationalist world, in which direct interpersonal connections were paramount. Categorization makes it possible to conceive of large groups as “imagined communities,” a prerequisite for nationalism (Anderson 1991). In general, nationalist groups derive their cohesion from collective-level identities that resonate deeply with the emotional needs of modern citizens. Conversely, nationalism can quickly implode multi-ethnic states, such as the former Yugoslavia, once ethnic mobilization starts to pull the component groups apart. All this supports the hypothesis that “competition for state resources is seen as a matter concerning not just individuals or associations of shared interests but rather whole ethnic groups” (Wimmer 2002, 103).

Still, establishing that internal conflict is over state ownership, and that the main competitors in such conflict are ethnonational groups, does not mean that we have explained why violence actually ensues in particular cases. Ultimately, this task calls for a complete theory of collective action that explains how the groups in question mobilize politically and how their mobilization triggers violence.

Mobilization depends on availability of collective identities, shared motivations and opportunities for collective action (Gurr 2000). Collective identities constitute the boundaries of the group in terms of ethnically defined membership and thus form the basis for the articulation of common interests. Shared motivations refer to the group members' incentives to participate in collective action endeavors. In general, perceived grievances, whether socioeconomic (Gellner 1983) or political (Wimmer 2002), could contribute to motivational articulation. Finally, action opportunities hinge on the power balance among ethnic groups, which in itself is a complex function of demographic, technological, geographic conditions, including the influence of international factors (Tilly 1978; Gurr 2000).

These key factors should be seen as necessary, rather than sufficient, conditions of collective action. In the following, however, we will focus primarily on ethnopolitical opportunity structures, while assuming that politically relevant group identities already exist and that all groups possess the motivation to engage in collective action. This does not mean that we believe that identities are primordially given or that the process of identity formation has nothing to do with the outbreak of conflict. Nor should it be concluded that the specific level of ethnopolitical grievance is unrelated to the frequency or intensity of violence. Rather than advancing a complete theory of nationalist mobilization and violence, our modest goal is to develop a simplified model that associates ethno-political configurations with civil-war outcomes.

It goes without saying that our mechanisms can only be partial and should be seen as the starting point for future theory building. In brief, we assert that escalation to violence is more likely where the dominant group is a demographic minority. The more demographically significant ethnic groups are excluded from state power, the more likely it is that there will be violent attempts at overcoming such imbalances. If the principle of ethnonational representativity is violated, we postulate a higher likelihood of grievance along ethnic lines, as well as a higher potential for escalation to political violence. In order to test this hypothesis, we need to measure the relative degree to which ethnic groups are included or excluded from state power.

The N* Index: A Model of Ethnonationalist Civil Wars

The general reasoning of the preceding section leaves many gaps open that will need to be filled before proceeding to empirical analysis. Our two main tasks are to identify the ethno-political configuration C* and the mechanism M* that make up the core of the model. We will consider each task in turn.

A Starlike Ethnic Configuration C*

Based on our assumption of the state's central role in ethno-nationalist conflict, we decided to reject the symmetric interaction topologies implied by the ELF index. Instead, drawing on assumptions A1 and A2, we postulate a starlike configuration with the ethnic group(s) in power (EGIP) at the center, surrounded by peripheral groups. This implies that the governmental group (or coalition) interacts with each nongovernmental group, and that the latter groups do not interact with each other. This configuration adds up to a starlike interaction topology.

It is now straightforward to compute the conflict propensity of the entire system:

where p(i) is the probability of dyadic conflict erupting between the EGIP and the excluded group i. Here we are assuming that the relational probability of conflict p(i) does not depend on the conflict propensity of the other dyads. A more realistic model could capture interactions among the provinces due to demonstration effects or other types of conflict contagion.

Note that this center-periphery logic resembles Fearon and Laitin's (2003) insurgency model, although their focus is of course entirely nonethnic. In contrast, Rokkan (1999) proposes a richer representation of how peripheries interact with central states, by accounting not only for economic and geographic factors but also for a cultural dimension (see also Cederman forthcoming).

An Opportunity-Based Mechanism M*

To complete our simple model, we need to specify what happens in each dyad, which in technical terms boils down to specifying the function p(i). Let us assume that the n ethnic groups can be ordered as a vector {s₀, s₁, s₂…s_n−1}, where s₀ denotes the size of the EGIP and the remaining entries the non-governmental groups is whatever order. Following the logic developed in the last section, we assume that conflict is likely to occur wherever demographically important groups are excluded from power. The larger these peripheral groups are, the more likely it is that they will successfully challenge the center.

It seems reasonable to select a logistical contest success function to formalize this logic. In a seminal article, Tullock (1967) proposed a success contest function on ratio-form. According to Hirshleifer (2001, 94), however, logistical contest success functions are better suited to military contests under uncertain and unconventional conditions, as is usually the case in civil wars. Thus we define the probability of dyadic conflict with nongovernmental group i as

where r(i)=s_i/(s_i+s₀) is group i's share of the total dyadic population, r is a threshold value and k a slope parameter. It should be noted that our functional form differs slightly from the standard version of the logistical conflict success function because it features the total dyadic resources in the denominator rather than merely those of the opponent. We choose this format since it seems more intuitive without changing our main findings.

The threshold parameter r stipulates at what demographic balance the odds are even for a challenge, that is, where p(i) = 0.5. Parameter k controls how steeply the curve slopes, and can be interpreted as a measure of how decisive combat is (Hirshleifer 2001). For k = 0 it is entirely flat and as k goes to infinity, it becomes a perfect step function. The curve in Figure 1 illustrates the functional form for parameters r = 0.5 and k = 5.

The Dyadic Probability of Conflict p(i) between Marginalized Group i and an Ethnic Group in Power

Note that although the mechanism seems to be entirely one-directional, this impression is misleading. In fact, no assumption has been made as to which side starts the fight. It may well be that the government preempts what it perceives as a threat by a demographically sizeable, but politically marginalized, group. Moreover, even if the peripheral actor takes action first, the most extreme instances of ethnic violence have almost all been perpetrated by states in response to insurgencies (Mann 2005, 7).

Obviously, beyond the initial demographic approximation of power, there are countless external factors that could enter this equation, including the geographic reach of the state as well as help to ethnic kin from third countries. These factors require additional coding and will therefore not be considered in this paper. Let us now define our indicator N* as C*(M*), a label that reflects the underlying model of ethnonational politics based on a star-like ethnic configuration C*:

To get a better feeling for how N* behaves in comparison to the conventional measures, Table 1 lists the values of ELF and N*(0.5, 5) for a number of sample groups. It immediately becomes clear that the new index is asymmetric with respect to the order of groups. We always list the EGIP first (and to clarify matters, it is highlighted in boldface). Therefore, whereas {0.7 0.3} yields 0.072, the reversed order {0.3 0.7} has a high probability of conflict: 0.843. This is so because if the EGIP controls 0.7 of the resources, it is very hard to unseat, whereas a minority government of 0.3 would constantly be threatened. In contrast, the ELF does not change with the order of group listings. As more groups are successively added, the ELF increases due to growing fragmentation. In our model, however, the N* goes up as long as the dominant group's share of the total population declines.

Empirical Calibration of the N* Index

It is now time to turn to the empirical performance of the new measure. The main obstacle to coding N* is the need to specify the ethnic group(s) in power (EGIP) for each state. For operational purposes, we consider a group, or a coalition of groups, to be in power if their leaders serve (at least intermittently) in senior governmental positions, especially within the cabinet. Beyond the ethnic background of a country's leading politicians, specific institutional arrangements, such as different types of power sharing and consociationalism, may also be indicators of power inclusion. Mere regional autonomy without significant input into cabinet-level governmental decision-making, on the other hand, is clearly not sufficient to warrant status as EGIP. Thus, in cases where the group in question is systematically excluded from power it is judged to be a marginalized ethnic group (MEG).

It goes without saying that this definition is hard to operationalize (see the Appendix). One source of potential confusion relates to the exact nature of governmental involvement. For example, because of elite assimilation, our coding of Ukrainians as being marginalized within the USSR can be disputed. Another difficulty relates to the possibility that different parts of the ethnic group might be included while other, possibly more radical branches, are excluded.The contemporary U.S. counterinsurgency strategy in Iraq aims to split the Sunni resistance in this way. There is also a potentially treacherous temporal dimension: one group may be EGIP during one period of a country's history but an MEG during another. To be sure, the Sunni dominance in Iraq did not necessarily extend to the entire sample period, especially not before 1958. Moreover, from 1963 to 1970 the Syrian EGIP was Druze-Alwite dominated, and so forth. These are important shifts that would have to await a dynamic coding of N*.

Ultimately, it can be debated to what extent it is possible to identify clearly distinct ethnic groups at all. In the long run, at least, ethnic and national identifications are endogenous to other political processes, including state formation and conflict behavior (Laitin and Posner 2001). By applying the measure to the post-1945 period, however, we are assuming that group identities are, at least in the majority of the cases studied, clearly marked and relatively stable.

In a first cut, we focus on Eurasia and North Africa. Within this subset of the population, we believe that most of the cases can be coded in a relatively straightforward manner. Because of its notoriously hard-to-grasp, fluid identities, sub-Saharan Africa was judged to be too difficult to code with respect to EGIP without extensive additional research. In future work, we plan to extend the coding to this and to other parts of the world, but extensive area expertise will be needed to get this data-gathering exercise right (cf. Posner 2004).

Our coding effort yielded observations for 88 countries, as indicated in the Appendix. We built on Fearon's (2003) list of ethnic groups and selected the EGIPs. Wherever more than one group was found to belong to the EGIP, the strength of the entire coalition was summed and kept as the first entry in the group vector. We compared our EGIP partitioning with an independent coder. To determine if a group was an EGIP, we relied on Heger and Salehyan's (2005) data set of leaders' ethnic affiliations that is based on information from a leader data collected by Goemans et al. (2004). Because these data are limited to conflict periods, additional information was extracted from the Minorities At Risk (MAR) database and the CIA World Factbook.

This empirical procedure generated an N* index that is radically different from the ELF (correlation-coefficient 0.42). This is confirmed by Figure 2, which plots N*(0.5, 5) based on real-world observations as a function of the ELF. If N* were identical to these other indices, the scatter plots would have formed a diagonal line from the lower left to the upper right. Instead, in both cases, the N* remains very low for most observations with low to medium ELF. For high ones, however, the variance is very high. Thus, there is no simple one-to-one correspondence between the N* and fractionalization.

A Scatter Plot of N* Against the ELF Index for All Countries in the Sample

Regression Analysis

This section evaluates the N* index's impact on conflict behavior. We do so within the context of Fearon and Laitin's (2003) econometric model that focuses on the onset of civil wars during the post-WWII period. This model is based on a pooled time series that contains country-year observations coded as a one if a civil war started within that observation and as a zero for all other cases. We use exactly the same list of independent variables as Fearon and Laitin include in their Model 1 except for the religious fractionalization index, which is dropped due to its marginal theoretical relevance to this paper. We did not find any model specification in which it differed noticeably from the ELF. Without exception, it remained as insignificant as that index. Thus the list includes lagged ongoing wars, per capita income, population, terrain, territorial contiguity, oil exports, recent independence, political instability, democracy, and the ELF (see Table 2). For details about the operationalization of these variables, the reader is referred to Fearon and Laitin (2003).

Relying on logit analysis, Table 3 contains the results of our first set of estimations.⁵

Although the threshold r seems quite reasonably located at 0.5, it would be desirable to estimate its exact location. Fortunately, it is possible to do so by dynamically recomputing a new N*(r,k) vector each time the log-likelihood function is called in the course of the estimation. In this case, all dyadic probabilities p(i), which depend on parameter r and the group sizes, are recalculated as a part of the log-likelihood function.⁷

Finally, in order to check the robustness of our parameter settings, we check whether a much simpler model yields convincing results. Model 4 introduces a dummy variable called “Minority in Power,” or MINIP for short, that is one if the EGIP is in demographic minority compared to the summed peripheral population and zero otherwise. The right-most column of Table 3 indicates that this simplified specification has a considerable impact that almost reaches significance at the 0.05 level.

It is quite remarkable that such crude models of ethno-nationalist politics generate strong results. All estimations reported in Table 3 are based on a dependent variable that treats all civil wars as one large category. Still, there are good reasons to believe that such an inclusive assumption ignores considerable unit heterogeneity, especially since it lumps together ethnonationalist civil wars with coups and other nonethnic conflicts (Sambanis 2004a). Fortunately, Fearon and Laitin's (2003) data set includes a variable that separates ethnic wars from other types of internal conflict. If our model of ethnonationalist conflict is correct, it should perform especially well for this dependent variable. After all, such a coding corresponds directly to the causal logic advanced in this paper.

To find out if this intuition can be empirically confirmed, we reran all models of Table 3 for the alternative dependent variable (see Table 4). As expected, Model 5 shows that the ELF index remains totally insignificant for the new specification. In contrast, our new N* measure performs extremely well. As suggested by Model 6, a fixed index N*(0.5, 5) yields a very high level of significance at p = 0.001 that equals or even surpasses the materialist variables per capita income and oil exports in this respect. As before, the democracy indicator is significant and the variables associated with prior war, population size, and recent independence continue to have a strong impact.

Model 7 presents a dynamic estimation of r, similar to that shown by Model 3. In this case, the parameter comes out somewhat lower (r = 0.421) than in the previous model. As expected, the estimation is even more significant for ethnic wars than for all civil wars. The model seems to fit the data remarkably well for these cases. Indeed, the statistical results of N* confirm that the postulated causal mechanisms do a good job of summarizing the center-periphery relationship between EGIPs and MEGs. In addition, Model 8 reveals that the simple MINIP indicator now becomes significant, thus suggesting that our fundamental causal insight does not hinge on the exact functional form of the N* index.⁸

What is the substantive effect of our new indicator? Figure 3 plots the predicted annual probability of ethnic war as a function of the N* values. The graph displays three curves, all based on the estimates of Model 5. The middle one represents the conflict propensity of a typical state where all other independent variables are held at their means. If N*=0, conflict is virtually excluded. As the indicator increases, however, the risk of ethnic war grows steadily up to about 0.02 for N*=1. The two remaining curves correspond to two subsamples holding states with GDP values below and above the mean. Whereas the effect of N* is just below that of the overall sample for the wealthy cases, the poor ones demonstrate a much stronger effect. In these cases, the conflict propensity increases to almost 0.3 for N*=1.

The Substantive Effect of N* on the Probability of Ethnic Civil Wars

Of course, it would be premature to draw any definitive conclusions for the entire world. In this paper, we have limited ourselves to slightly more than half of the global sample studied by Fearon and Laitin. An extension of the N* index to the non-Eurasian cases will require careful empirical coding of EGIP. Moreover, it will be necessary to study the robustness of the operationalization proposed in the Appendix even for the cases at hand. Such a coding project should also pay more attention to diachronic differences in the status of EGIPs than we have been able to do in this paper.

A particularly exciting extension of the current paper would attempt to derive group-specific measures beyond demographic size. Based on the agent-based model introduced in Cederman (forthcoming), we expect the location of the ethnic group to be very important. Our current research, which is based on Geographic Information Systems (GIS), confirms the findings of this article at the level of geo-coded center-periphery dyads (Buhaug, Cederman and Rød 2006). In reality, even small groups can be very powerful in asymmetric conflicts, thanks to their ability to take advantage of difficult terrain far away from the capital. It would also be interesting to go beyond the primarily opportunity-driven logic of the present model by separating those cases that feature actual ethnonationalist escalation from structural situations that are conducive to conflict.