The conventional approach to the quantitative study of civil war is to compare observations of civil war onset to a heterogeneous control group that combines cases of actual peace with cases of nonviolent conflict of varying type and intensity. This approach has identified some robust correlates of civil war, but it ignores the question of conflict escalation. Civil wars almost always grow from nonviolent claims expressed intra- or extra-institutionally,Footnote 1 so to understand civil war we must really understand why nonviolent conflicts escalate. A key insight from the literature on contentious politics is that we cannot simply assume that violent and nonviolent conflicts have different causes.Footnote 2 To effectively test civil war theories, we need to study the process of conflict escalation from nonviolent claims to violence.
We address this gap in the literature by analyzing the role of ethnic grievances in the process of conflict escalation. Recent studies have established that grievances increase the risk of ethnic war.Footnote 3 However, these studies cannot tell us whether grievances have this effect because they trigger the emergence of nonviolent claims, or because they affect the likelihood that such claims, once formed, will turn violent. We are aware of only two studies that explicitly consider the effects of ethnic grievances on conflict escalation and they point to mutually contradictory conclusions.Footnote 4 Thus, the exact role of grievances in conflict processes remains an open question.
We focus our analysis on conflicts over self-determination. Self-determination (or separatist) conflicts revolve around disagreements over ethnic self-rule. While fundamentally domestic in nature, separatist conflicts can have important consequences for the international system. Separatist claims led to the formation of more than twenty new states since the end of the Cold War, as well as several de facto states and officially sanctioned autonomy regimes. Separatism can sow the seeds for major inter-state disputes, as it has in Sudan, Kashmir, and eastern Ukraine, and it accounts for more than a third of all civil wars fought since 1945. However, according to a novel data source that we describe in more detail later, 90 percent of all separatist movements emerged as nonviolent—and two thirds of all separatist movements never turned to violence. We explore whether ethnic grievances are associated with the emergence of nonviolent claims for greater autonomy or statehood and with the escalation of those claims to violence.
Previous literature has analyzed the effect of ethnic grievances on both separatist and center-seeking wars. An advantage of our narrower focus is that we can theorize the role of grievances among a set of cases with greater causal homogeneity and, therefore, with greater specificity. We focus on two sources of grievances that are widely seen as pertinent in the context of separatism: the exclusion of ethnic groups from state power and losses of territorial autonomy, which can arise for both domestic reasons (e.g., state consolidation or assimilation) and because of international dynamics (e.g., annexation, conquest, or border changes). Combining insights from different theoretical traditions, we argue that while exclusion and lost autonomy may fuel both the emergence of nonviolent separatist claims and their escalation to violence, they are not equally relevant at each stage of conflict escalation, in part because they create different incentives for mobilization and opportunities for redress.
In keeping with a processual conception of conflict, our empirical analysis involves two steps. In the first, we estimate the association between grievances (exclusion and lost autonomy) and the onset of nonviolent separatist claims, broadly defined to include both extra-institutional protest and institutional mobilization. In the second step, we establish the association of the same two grievance factors with escalation to separatist civil war conditional on a prior nonviolent separatist claim. Our analysis combines recently introduced, group-level data on violent and nonviolent claims for self-determinationFootnote 5 with data on political exclusionFootnote 6 and new data on recent and historical autonomy loss for more than 750 ethnic groups around the world.
We find empirical support for our argument that not all types of ethnic grievance are equally relevant at different stages of separatist conflict escalation. According to our results, autonomy loss has a strong and highly robust association with the emergence of nonviolent separatist claims while exclusion does not. We suggest that this could be because exclusion is more likely to motivate efforts to regain inclusion in the central government rather than mobilization for territorial autonomy or secession. However, exclusion from the central government is robustly associated with the escalation of nonviolent separatist claims to violence. We argue that this is in part because lack of government representation reduces the effectiveness of pursuing separatist claims nonviolently. Finally, our results suggest an association between autonomy loss and violent escalation; however, that applies to only recent autonomy revocations and not to historical cases of autonomy loss. A possible reason is that distant memories from the past do not inspire the same degree of resentment.
These new results contradict claims by opportunity theorists who see no significant role for grievances in the escalation of nonviolent claims to violence. However, they also suggest that sharp distinctions between opportunity and grievance factors may be misguided since exclusion could be related to the escalation of nonviolent separatist claims via both affective mechanisms highlighting the role of state illegitimacy and unfairness and by shaping the opportunity structure. This points to the intertwined logics of grievances and opportunities, which are complementary, rather than competing explanations for conflict.
Related Literature and Approach
The Classic Debate
Classic studies by Gurr, Horowitz, and others see grievances as a direct cause of both nonviolent contention and rebellion.Footnote 7 However, that view has been questioned by the opportunity school, which argues that while grievances may be necessary for the formation of nonviolent claims and social movements, they are too ubiquitous to explain why some dissident groups resort to violence and others do not. For Tilly and other opportunity theorists, it is not grievances that are the key to understanding why nonviolent conflicts escalate, but the political opportunity structure—constraints and costs to violent mobilization.Footnote 8 These arguments are echoed in political economy approaches, such as those by Fearon and Laitin or Collier and Hoeffler.Footnote 9
Recent studies have challenged the primacy of opportunity over grievances as an explanation of rebellion. Using new data coded at the ethnic group level, Cederman, Wimmer, and Min show that political exclusion is strongly correlated with ethnic war onset.Footnote 10 Other studies using group-level data suggest that losses of territorial autonomyFootnote 11 or wealth differences between ethnic groupsFootnote 12 are associated with ethnic war.
While these studies changed the debate on the role of grievances in civil war, a limitation is that they do not account for prior nonviolent mobilization. They cannot, therefore, establish whether grievances are directly related to civil war onset or indirectly, through their effect on nonviolent mobilization. A series of recent studies have linked political exclusion, lost autonomy, and other grievance factors with the emergence of extra-institutional protest campaignsFootnote 13 and the occurrence of self-determination claims.Footnote 14 Overall, it remains unclear whether grievances affect only the emergence of nonviolent claims (as opportunity theory predicts) or both the emergence of nonviolent claims and their escalation to violence (as grievance theory predicts).
Approach and Antecedents
To improve our understanding of grievances’ role in civil war processes, we need to shift to a more processual understanding of intra-state conflict. We do so by using a two-step approach that first considers the role of ethnic grievances in the emergence of nonviolent separatist claims and, in a second step, their role in the escalation of nonviolent claims to separatist war. This approach allows us to explore the role of ethnic grievances at different stages in the process of conflict escalation.
Our focus on escalation has antecedents in a small number of quantitative studies.Footnote 15 Directly relevant is a recent article by Lindemann and Wimmer that investigates the conditions under which fifty-eight ethnic groups with high propensities for rebellion resort to arms.Footnote 16 Their findings suggest that both grievances (resulting from indiscriminate state violence) and opportunities (in the form of refuge areas) matter for conflict escalation. Our study complements their findings in two main ways. First, we consider the association between violent outbreaks and nonviolent forms of indiscriminate repression, specifically revocations of autonomy and lack of representation at the center. Second, we use a two-step approach that allows us to explicitly distinguish the effects of grievances on nonviolent claims and their escalation in the same framework.
Two-step models similar to ours are common in the literature on inter-state war, but we are aware of only two prior studies employing this approach in the literature on civil war.Footnote 17 The most directly relevant study is by Bartusevičius and Gleditsch, who also use a two-step approach to investigate the role of ethnic exclusion and discrimination in intra-state conflict. Their findings suggest that ethnic exclusion/discrimination is positively related to the emergence of “incompatibilities” between the state and domestic challengers, but not with the escalation of incompatibilities to violence, thus adding to skepticism regarding the role of grievances in violent rebellion. But while we agree with Bartusevičius and Gleditsch about the value of two-step approaches, we believe that their study design suffers from limitations that make us question their findings.
First, Bartusevičius and Gleditsch simultaneously analyze both ethnic and non-ethnic conflicts and they conduct all analyses at the country level. However, there is no reason to expect that ethnic grievances should be related to non-ethnic conflict escalation and aggregating all group-level data to the country level implies that the effects of group-specific grievances cannot be adequately captured. Group-level data, as we use, are more appropriate to study group-specific escalation.
Second, Bartusevičius and Gleditsch argue that civil wars are likely to emerge from only extra-institutional nonviolent mobilization, such as demonstrations, strikes, or civil disobedience. They suggest that as a result, institutional nonviolent mobilization can be disregarded in the study of conflict escalation. We disagree with this view. While most civil wars have roots in some kind of nonviolent conflict, the buildup to civil wars does not necessarily progress linearly from institutional to extra-institutional contention to war.Footnote 18 There can be direct transitions from institutional contention to war, such as when militant and extremist groups capture party politics.Footnote 19 According to a recent data collection, only 26 percent of separatist violence onsets were preceded by one or more out of five common forms of extra-institutional mobilization in the previous year (40 percent when considering the three years preceding violence onsets).Footnote 20 While a narrow focus on extra-institutional contention may be consistent with recent studies of nonviolent strategies of resistance,Footnote 21 it is not necessarily the right approach when it comes to the study of conflict escalation because it results in omitting the nonviolent formative stages of some civil wars. Therefore, in our analysis we choose to rely on a broader definition of nonviolent conflict that includes any kind of organized nonviolent claim making, including institutional forms.
Finally, Bartusevičius and Gleditsch draw their data on extra-institutional protest from the CONIAS data set.Footnote 22 This compounds the problems we described because CONIAS covers protests only if they cross an (ambiguously defined) intensity threshold. In particular, CONIAS includes protests only if they are rejected by the state as “unacceptable,” which makes it even more likely that the nonviolent formative stages of civil wars are missing. Two thirds of the conflicts included in CONIAS are violent from the startFootnote 23 despite clear evidence of prior nonviolent mobilization in many cases.Footnote 24 Furthermore, by dropping “acceptable” forms of protest, CONIAS selects out protest campaigns that are unlikely to turn violent—especially in democracies, where demonstrations and strikes are widely accepted means of claim making—and this can lead to bias. To avoid these problems, we rely on an alternative source of data on nonviolent separatist claims with improved coverage of the nonviolent formative stages of separatist wars.
Theory
We now proceed to develop a new theory of the role of political exclusion and lost autonomy in separatist conflict processes. We define political exclusion as the lack of meaningful representation of an ethnic group in a state's governing coalition. Lost autonomy refers to ethnic groups that have a diminished degree of territorial self-rule compared to the past, such as when a group used to control an independent state that was annexed by another state and therefore no longer exists, or when the state revoked an earlier autonomy arrangement. Consistent with grievance theory, we expect that political exclusion and lost autonomy matter at both conflict stages. However, we argue that exclusion has a weaker relation with the emergence of nonviolent separatism because it generates different mobilization incentives and opportunities for redress. Moreover, we suggest that affective mechanisms are not the only possible link between grievances and violent escalation.
Political Exclusion, Lost Autonomy, and Nonviolent Claims for Self-Determination
Existing theories point to two different mechanisms linking political exclusion and lost autonomy to the emergence of nonviolent separatist claims. First, they can both lead to a collective interest in increasing ethnic self-determination. In the case of exclusion, this is because exclusion violates a core principle of political legitimacy in the modern era—rule by co-ethnics—and because it can generate economic inequality and material deprivation.Footnote 25 Lost autonomy, in turn, can stoke resentment about the group's diminished social status and incentivize efforts to restore the group's former power.Footnote 26
Second, ethnic grievances generated by exclusion or lost autonomy can alleviate collective action problems.Footnote 27 Both exclusion and lost autonomy can be perceived as a form of nonviolent, indiscriminate repression targeting an ethnic group. The indiscriminate nature of such repression should increase ethnic solidarityFootnote 28 and, by triggering emotions such as fear and resentment, increase the willingness to resist.Footnote 29
These mechanisms can provide ethnic groups with a motive and increase their ability to pursue nonviolent separatist claims. However, we argue that autonomy loss should have a stronger association with nonviolent separatist claims than exclusion. This is because aggrieved groups do not only face a choice between no action and mobilizing for territorial self-determination. Ethnic groups can also mobilize for inclusion at the center. And because representation at the center is likely to reduce political and economic dominance by “ethnic others,” mobilizing for inclusion can be an equally if not more attractive goal for excluded groups. By contrast, regaining territorial self-rule clearly constitutes the most direct form of redress for groups who have lost territorial autonomy. Therefore, we expect that autonomy losses are primary motives in the nonviolent pursuit of self-determination, whereas exclusion will have a weaker association.
H1 Both political exclusion and lost autonomy are associated with a higher risk of nonviolent separatist claim onset, but the association between exclusion and nonviolent separatist claim onset is weaker.
So far we have taken a static view of grievances, but timing is likely to matter. Grievance theory suggests that the more recent grievances due to state policy are, the more intensely felt is the frustration and motivation for collective action.Footnote 30 In particular, recent retractions of autonomy are more likely to generate resentment about unfair treatment by the state and push groups to “reverse the reversal.”Footnote 31 Recent autonomy retractions should therefore be especially likely to increase the onset of nonviolent separatist claims.
However, analogously to our previous argument, this does not necessarily extend to recent loss of representation at the center. In fact, recent exclusion could plausibly have no effect at all. The most likely response to a loss of status is a desire to regain the same status. Furthermore, groups that until recently have formed part of a state's governing coalition are likely to have a relatively high degree of attachment to the state or nation, at least when compared to groups that have always been excluded. While recent exclusion is likely to motivate collective action, such action is therefore likely to be directed at the regaining of representation at the center. Accordingly, we expect that recent exclusion has either a weak or no association with the onset of nonviolent separatist claims.
H2 Recent autonomy loss is associated with a higher risk of nonviolent separatist claim onset, whereas recent exclusion has a weak or no association.
Escalation to Violence
Given social norms against the use of violence and considering the high costs of violent conflict, groups seeking self-determination are likely to make nonviolent claims initially. However, as Gurr and other grievance theorists have long maintained, grievances can increase the risk that nonviolent claims escalate to violence.Footnote 32 In part, that is because perceptions of unfair treatment by the state increase the plausibility, justifiability, and diffusion of the idea that the state needs to be violently “smashed” and reorganized.Footnote 33 Moreover, perceptions of unfair treatment increase the willingness of group members to participate in risky actions and rebellion.Footnote 34
In contrast to our argument about the emergence of nonviolent claims, we expect both exclusion and lost autonomy to have strong effects on escalation to separatist violence. The fact that groups have articulated an interest in greater self-rule and have begun to mobilize means that they can overcome some obstacles to collective action; and affective mechanisms triggered by either exclusion or autonomy loss should make it more likely that nonviolent claims escalate to violence. However, affective mechanisms are not the only possible link with violent escalation. We argue that political exclusion may also be connected to violent escalation because it limits the opportunity to pursue claims nonviolently. The tactical choices of dissident groups are at least in part based on rational cost-benefit evaluations.Footnote 35 Therefore, factors such as state capacity that are commonly associated with opportunity models should shape the decision to escalate.Footnote 36 But so should exclusion, which reduces access to institutional channels for claim making and thereby nonviolent strategies’ effectiveness. More generally, persistent grievances, including those caused by autonomy losses, demonstrate to nonviolent movements the futility of nonviolent tactics.
The bargaining model of war provides further support for the idea that both exclusion and autonomy revocations make it more likely that nonviolent separatist claims escalate to war. The bargaining model highlights the role of information asymmetries and commitment problems that impede the peaceful resolution of conflicts.Footnote 37 Exclusion or revocations of autonomy serve as reminders that the state cannot be trusted to uphold a settlement, thereby magnifying the commitment problems from the perspective of groups challenging the state.Footnote 38 Grievances arising from long-standing exclusion or autonomy loss could also increase perceptions of issue indivisibility. Highly aggrieved groups are more likely to make maximalist claims, such as claims for outright secessionFootnote 39—and territory is much harder to divide than sovereignty.Footnote 40
H3 Both political exclusion and lost autonomy are associated with increased risk that nonviolent separatist claims escalate to violence.
The preceding discussion suggests that the risk of conflict escalation is highest if grievances are recently imposed. Recent status downgrades—such as losing representation at the center or losing autonomy—are especially likely to stoke ethnic violence arising from resentment and desire for revenge. At the same time, commitment problems are magnified if the state has recently moved to curtail a group's rights. Moreover, since an active separatist claim signals a diminished attachment to the state/nation, a recent loss of power at the center will further diminish national identification and fuel the risk of violent escalation.
H4 Both recent exclusion and recent autonomy loss are associated with increased risk that nonviolent separatist claims escalate to violence.
Data
Self-Determination Claims
We use data from the recently introduced self-determination movements (SDM) data set, which codes all SDMs from 1945 to 2012.Footnote 41 SDMs are defined as movements constituted by one or more organizations that are connected to an ethnic group making claims for territorially defined self-rule. SDM data include a broad range of claims ranging from limited internal autonomy demands (e.g., Mayas in Mexico) to demands for national independence (e.g., Scots in the UK) or merger with another state (e.g., Serbs in Bosnia). There must be evidence of organized political mobilization for a movement to be included in SDM. Mobilization may be violent or nonviolent, extra-institutional, or part of conventional politics. SDM codes an end to a movement if a group ceases to make public claims or the group secedes.
For each year of activity SDM codes whether there was violent separatist conflict over self-rule, defined as lethal conflict with casualties on both sides. Both major wars and low-intensity wars are included. SDM draws its data on separatist war from several sources, including the UCDP data set,Footnote 42 Doyle and Sambanis,Footnote 43 and Minorities at Risk (MAR).Footnote 44 The left panel of Figure 1 gives annual counts of the number of violent and nonviolent SDMs.
Annual frequencies of violent and nonviolent self-determination claims
SDM significantly improves coverage of separatist claims relative to previously available sources, especially when it comes to nonviolent claims and the nonviolent formative stages of separatist wars.Footnote 45 Overall, SDM identifies 464 self-determination movements in 120 countriesFootnote 46—or around three times as many separatist conflicts as CONIAS or the well-known data set by Cunningham during similar time frames.Footnote 47 Two thirds of the separatist conflicts in SDM never became violent and only 10 percent of the separatist conflicts were violent in their first year. By comparison, half of the separatist conflicts that Cunningham identified became violent and a quarter were violent from the start. In the CONIAS data set, 80 percent of the separatist conflicts were violent and almost half were violent in their first year (suggesting that CONIAS used violence outbreaks as an indication that protests were “unacceptable” to the government, according to our earlier discussion).
EPR
We merge the SDM data on separatist claims with group-level data on political exclusion from the Ethnic Power Relations (EPR) data set, version 2014.Footnote 48 EPR is less likely than MAR to overrepresent groups that are discriminated against by the state and merging with EPR allows us to engage with previous studies on political exclusion. However, our reliance on EPR also has costs since by anchoring our study on EPR we lose many of the groups in SDM. In particular, EPR includes only groups defined over race, language, or religion. Similarly to MAR and other extant sources of data on separatism, the SDM data set in addition considers regionally defined groups (e.g., the Lombards in Italy). Given EPR's narrower definition of ethnicity, these cannot be included here. EPR also does not code groups in overseas territories and provides no data for 1945. Overall, we are able to match 289 of the 464 SDMs to EPR groups, or 62 percent.Footnote 49 The SDMs that can be linked to EPR are somewhat more likely to have engaged in separatist violence, but a majority (55%) never used violence, and in only twenty-two cases were self-rule claims violent from the start. The right panel in Figure 1 gives annual breakdowns of the number of violent and nonviolent SDMs in EPR.
Dependent Variables
We analyze two binary dependent variables. The first captures the onset of a nonviolent claim for self-determination, coded 1 in the first year an organization made a separatist claim on behalf of an ethnic group and 0 otherwise while dropping the twenty-two cases of SDMs that start violent. All group-year observations with an ongoing separatist claim after the first year are dropped. Overall, there are 192 nonviolent separatist claim onsets in our data. A total of thirteen groups have two onsets due to discontinuous separatist activity. No group has more than two.
The second dependent variable captures conflict escalation, coded 1 if we observe a transition from a nonviolent separatist claim to separatist violence and 0 otherwise. All observations without a prior nonviolent separatist claim are dropped, including the twenty-two cases of SDMs that start violent, as are observations with ongoing armed conflict. We code 159 cases of conflict escalation; seventy-seven are “first-time” escalations, while the other eighty-two represent cases of conflict recurrence in the same state-group dyad.
Main Explanatory Variables
Our main explanatory variables are political exclusion and lost autonomy. We use data on exclusion from the EPR data set, which measures exclusion as a binary variable indicating whether a group has (0) or does not have (1) representation in the national executive at the beginning of a calendar year.Footnote 50
We provide new data on autonomy loss. Many previous studies have drawn data on lost autonomy from MAR,Footnote 51 but MAR covers only a fraction of the groups in EPR (250 out of EPR's 800 groups). We revised and expanded the MAR data on lost autonomy to include all EPR groups in our analysis. Similarly to MAR, lost autonomy is coded equal to 1 under three scenarios: if a group used to control an independent state that was annexed, invaded, or no longer exists for any other reason (e.g., the Estonians in the former Soviet Union); if a change of borders leads to groups being stranded outside of their home state (e.g., Russians in Ukraine after 1991); and if a group had, but lost, significant internal autonomy within a larger state (e.g., the Kosovar Albanians in Serbia after 1989). For each scenario we code autonomy loss since 1800. We drew on a broad array of sources for the coding of lost autonomy, including several encyclopedias focused on ethnic and separatist groups, Encyclopedia Britannica, the country studies series of the Library of Congress, EPR's regional autonomy indicator, MAR, and various case-specific sources.Footnote 52
Because it captures more than 200 years of ethnic group histories, this measure of autonomy loss is fairly static. To test our hypotheses about the short-term implications of losses of autonomy and losing representation, we code two additional binary variables measuring, respectively, whether groups lost representation at the center (recent exclusion) or autonomy (recent autonomy loss) during the previous two years.
Controls
We control for a large number of variables that have been associated with separatist war in previous studies. Group-level controls include regional concentration; relative group size; cross-border separatist kin groups; regional autonomy; presence of hydrocarbon reserves (oil/natural gas resources); mountainous terrain; and noncontiguity to the main body of the country. The latter three are specific to ethnic settlement areas and therefore available only for regionally concentrated groups. Country-level controls include constant gdp per capita (in logs); total population size (in logs); democracy score; federal institutions; and the total number of politically relevant ethnic groups. Systemic conditions that might influence separatism are captured by a binary indicator for the cold war. The online appendix provides information on data sources and summary statistics.
Results
Nonviolent Separatist Claim Onset
We start by analyzing the effects of political exclusion and lost autonomy on the onset of nonviolent separatist claims between 1946 and 2012. Table 1 reports descriptive statistics and Table 2 shows a series of regression models. We drop all groups that dominate the executive branch of government without sharing power with any other groups (e.g., Turks in Turkey) because these groups almost by definition make no separatist claims against the state that they control. To account for time dependence, all regression models include cubic polynomials counting the number of years since the beginning of the sampling period or since the last time a group made a claim.Footnote 53 We estimate both logit regressions with region fixed effects (odd model numbers) and ordinary least square regressions with country fixed effects (even model numbers). Standard errors are clustered by country. The unit of analysis is the country-group-year.
Nonviolent separatist claim onset propensity by exclusion and lost autonomy
Regression models explaining the onset of nonviolent separatist claims
Notes: All models include a constant (not shown). Standard errors clustered by country in parentheses. + p < .10; * p < .05; ** p < .01; *** p < .001.
Consistent with H1, we find a positive, statistically significant, and robust association between lost autonomy and the onset of nonviolent separatist claims.Footnote 54 According to model 1 in Table 2, groups that have experienced a loss of autonomy since 1800 are 0.9 percentage points more likely (0.4 vs 1.3%) to start making a nonviolent separatist claim (p < 0.01).Footnote 55 For comparison, regional concentration, which many consider a necessary condition for separatism, increases the probability of a nonviolent separatist claim onset in the same model by 0.8 percentage points (from 0.2 to 1%). Cross-border separatist kin—another frequently cited factor conducive to separatism—leads to an increase of 0.5 percentage points (from 0.8 to 1.3%).
Model 1 also suggests a positive correlation between political exclusion and the onset of nonviolent separatist claims. However, this correlation is both smaller and less robust.Footnote 56 As model 2 shows, the association of exclusion with nonviolent separatist claim onset misses conventional levels of statistical significance when country fixed effects are included (p = 0.09). By contrast, the effect of autonomy loss since 1800 increases in both size (+1.3 percentage points) and statistical significance (p < 0.001) when accounting for unobserved country-level heterogeneity. Similar conclusions are reached when we restrict the sample to regionally concentrated groups (see models 3 and 4). Regional concentration comes close to a necessary condition for separatismFootnote 57 and restricting the sample to concentrated groups allows us to include three additional controls that are specific to ethnic settlement areas: the presence of hydrocarbon reserves, noncontiguity, and mountainous terrain.Footnote 58
Models 5 to 8 in Table 2 re-estimate the same suit of models while replacing the static versions of exclusion and lost autonomy with our variables measuring recent exclusion and recent autonomy loss within the previous two years. In line with H2, model 5 suggests that a recent autonomy revocation is associated with a six-percentage-point increase in the probability of a nonviolent separatist claim onset in model 5 (from 0.8 to 7%, p < 0.001). This suggests that ethnic groups that have recently experienced a loss of autonomy are almost 800 percent more likely to make nonviolent claims for self rule. This result is robust to the inclusion of country fixed effects (model 6) and to the addition of numerous controls (models 7 and 8). Meanwhile, we find no evidence to suggest that recent loss of representation would affect the probability of nonviolent separatist claim onset—a conclusion already suggested by a simple χ 2-test (p = 0.39).Footnote 59
We report additional robustness checks in section 8 of the online appendix, including models with only region or country fixed effects and no other controls, models with a large battery of additional controls, a formal sensitivity analysis to assess sensitivity to hidden bias, changing the threshold used to code historical losses of autonomy from 1800 to 1900, and using different temporal cut-offs for the recent exclusion and recent autonomy loss variables. Overall, these additional checks suggest that both historical and recent autonomy loss have a highly robust, positive association with nonviolent separatist claim onset whereas political exclusion at the center does not. The formal sensitivity analysis provides additional evidence that the correlation of exclusion with nonviolent claim onset is sensitive to violations of the exogeneity assumption. Moreover, we find that the coefficient of exclusion (but not autonomy losses) is sensitive to dropping influential countries with large numbers of nonviolent separatist claim onsets, such as Russia and the former Soviet Union. Across a large number of specification and measurement choices, recent loss of representation almost never has a statistically significant association with nonviolent claim onset.
Summing up the results thus far, we find strong evidence that autonomy loss and especially recent autonomy revocations are correlated with the onset of nonviolent separatist claims. Meanwhile, political exclusion has a weaker association with nonviolent separatist claim onset that is not robust, whereas recent exclusion is clearly uncorrelated. These results could suggest that exclusion (especially if recent) is more likely to lead to mobilization aimed at reinstating the group's representation at the center rather than a push for self-determination. We posit this as a hypothesis in need of further testing, since a direct test would require the collection of new group-level data on nonviolent claims for more inclusion in central government. We cannot, therefore, rule out that exclusion is simply a weaker type of grievance around which to mobilize nonviolently.
Conflict Escalation
Most SDMs in our data remain nonviolent and those that do escalate are, on average, preceded by nine years of nonviolent claim making before the first outbreak of violence. We now explore whether exclusion and lost autonomy are associated with the violent escalation of SDMs between 1946 and 2012. The unit of analysis remains the country-group-year, but all analyses are now conditional on prior nonviolent separatist claims. The dependent variable is conflict escalation, defined as a transition from nonviolent separatist claims to separatist war. Table 3 shows descriptive statistics for each type of grievance.
Escalation propensity by exclusion and lost autonomy
Table 4 reports the regression results. We show separate models for first-time escalations (dropping all observations after the first incidence of violence) and all escalations (including cases of war recurrence). All regressions include controls for time dependence (cubic polynomials of the number of years since the group first made a nonviolent separatist claim or, where applicable, since the last spell of separatist war). As before, we estimate both logit models with region fixed effects (odd model numbers) and ordinary least squares regressions with country fixed effects (even model numbers). Because the groups that make separatist claims are almost all regionally concentrated and this allows us to include the full set of controls, all regression models restrict the sample to concentrated groups (results are similar when all groups are included; see the online appendix). Standard errors are clustered at the country level.
Regression models explaining the escalation of nonviolent separatist claims to separatist war
Notes: All models include a constant (not shown). Standard errors clustered by country in parentheses. + p < .10; * p < .05; ** p < .01; *** p < .001.
In line with H3, we find evidence that nonviolent claims are more likely to escalate to separatist war if groups are excluded from power. According to model 1, exclusion increases the risk of first-time escalation by 1.5 percentage points (from 0.7 to 2.2%, p < 0.01). If we include cases of war recurrence in the analysis (see model 5), exclusion increases escalation risk by 1.2 percentage points (from 1.6 to 2.8%, p < 0.05). The magnitude of these changes is similar to the association of gdp per capita with escalation;Footnote 60 GDP per capita is generally considered the strongest predictor of civil war onset. Models 2 and 6 suggest that exclusion remains positively associated with both first-time escalation (p < 0.05) and all escalations (p = 0.06) when country fixed effects are included.
Table 3 shows that escalations are around twice as likely among groups that have lost autonomy. However, this correlation misses conventional levels of statistical significance after regression adjustment (see Table 4). A possible reason is that this variable captures historical losses of autonomy going back decades or longer. Resentment and other effects of autonomy loss may dissipate over time. In line with H4, we find much larger and statistically significant associations when looking at more recent autonomy losses. This applies especially when we include repeated escalations. According to model 7, a recent autonomy revocation increases the risk of violent escalation by a massive ten percentage points (from 2 to 12%, p < 0.001). If we drop cases of war recurrence, the association with escalation is smaller (+3 percentage points according to model 3, p < 0.05) and misses conventional levels of statistical significance when country fixed effects are included (model 4). Although we cannot make any causal claims with this analysis, these patterns could suggest that the logic of conflict escalation is different in conflicts that have already turned violent and that revocations of autonomy are then especially damaging. The violence and protests in the aftermath of the recent scrapping of Kashmir's special autonomy arrangement offer a case in point. However, the small number of cases constitute a significant limitation and we point out that the unadjusted escalation risk after a recent autonomy downgrade is not too different in the first-time escalation sample (see Table 3).
Finally, Table 3 also points to a possible relation between recent exclusion and conflict escalation. However, there are only thirty-two instances of groups losing representation at the center during an active nonviolent separatist claim and in only three cases do we see a violent escalation. In regression models, we generally find no statistically significant association but in light of the small number of cases, we cannot fully explore the connection between recent exclusion and escalation.
Additional results reported in section 9 of the online appendix suggest that the effects of recent autonomy loss (when including cases of war recurrence) and exclusion survive a large number of robustness checks, model specification changes to add/drop controls, models using different temporal cut-offs to code recent autonomy loss, and models dropping influential countries with multiple instances of conflict escalation.Footnote 61 We get similar results when using data on separatist armed conflict from a different source (UCDP) and, according to a formal sensitivity analysis, the effects of these variables are relatively robust to unobserved confounders.Footnote 62 We also find evidence that exclusion has a pronounced effect on conflict escalation if groups are not only powerless but also actively discriminated against by the state. This is consistent with our theoretical framework, given that discrimination is likely to increase resentments against the state and commitment problems.Footnote 63
Beyond Exclusion and Lost Autonomy
The regression models we reported control for many common predictors of separatist war onset. Therefore, our results also allow us to shed light on the ability of variables other than exclusion and lost autonomy to account for the escalation of nonviolent separatist claims. First, though, it is worth noting that several of our controls have more or less robust associations with the onset of nonviolent separatist claims—this is true notably for regional concentration, territorial noncontiguity, country-level population size, and gdp per capita. However, we find that many of the variables that are purported to measure “opportunity” for insurgency (e.g., noncontiguity, mountainous terrain, and country population)Footnote 64 have no robust association with the escalation of nonviolent separatist claims to violence. Similarly, resource wealth (hydrocarbon reserves) in the territory occupied by separatist groups, relative group size, as well as regional concentration, the size of a government's military, and occurrences of civil war in neighboring countries (see Tables S11 and S15 in the online appendix) have no robust association with the escalation of nonviolent separatist claims.
An important conclusion that emerges from our analysis is that many of the variables that have been thought to explain the outbreak of separatist war are in fact capturing conditions that are conducive to the emergence of nonviolent separatist claims and cannot explain why nonviolent separatist claims escalate to violence. An exception is country-level gdp per capita, which has a positive and significant association with the onset of nonviolent separatist claims and a negative and significant correlation with violent escalation. Income is therefore one of the few covariates that increase separatist war models’ specificity beyond exclusion and recent autonomy loss. Additional results reported in the online appendix suggest that the proximity of an ethnic group to international land borders can also increase the risk of escalation. This is consistent with arguments about the difficulty of state building in peripheral areas and with previous results on the destabilizing effect of cross-border groups and cross-border sanctuaries.Footnote 65
Conclusion
Patterns of conflict escalation have been underexplored in the literature on civil war. We made use of novel data and a two-step approach to explore the role of ethnic grievances in separatist conflict processes. While our analysis cannot identify causal effects of political exclusion or lost autonomy, the two-step approach improves over the conventional way of modeling civil war onset and produces valuable new insights.
One new insight is that ethnic grievances matter for both the onset of nonviolent separatist claims and the escalation of such claims to violence; however, different types of grievances matter more at different stages of the escalation process. On the one hand, we find that while political exclusion is robustly associated with the escalation of separatist conflicts to violence, exclusion has no robust association with the emergence of nonviolent separatist claims. A possible explanation that could be usefully explored further using qualitative methods is that excluded groups often choose to mobilize for representation at the center rather than pursue territorial self-determination and that this might reflect a higher attachment to the nation for groups that have had some prior experience of inclusion. On the other hand, our results suggest that whereas both recent and more historic autonomy losses increase the probability that groups start to make nonviolent separatist claims, only recent autonomy revocations affect the escalation risk.
Taken as a whole, our analysis contradicts claims by opportunity theorists that grievances are too ubiquitous to explain why conflicts escalate from nonviolent claims to violence while lending support to grievance theory as articulated previously by Gurr, Horowitz, Cederman, and Wimmer. However, we add a more nuanced perspective to this literature that puts grievances front and center while also supporting opportunity-cost theories of mobilization and rebellion. Political exclusion is not simply a measure of grievance. Reduced access to the state implies diminished opportunities to address grievances nonviolently and, in turn, limited opportunities for the nonviolent adjudication of disputes increase the risk of war. More generally, consistent with recent studies that merge grievance- and process-based theoriesFootnote 66 we find that factors associated with both grievance and opportunity models (especially country wealth and proximity of ethnic groups to borders) are associated with the escalation of separatist conflicts from nonviolent claims to violence.
That said, we also found that many other variables that are thought to be “determinants” of civil war cannot in fact distinguish between violent and nonviolent separatist claims. While models of civil and separatist war tout their specificity and predictive accuracy, many of their key explanatory variables seem to explain separatist claims generally, rather than separatist war per se. Our results suggest that the extant literature on civil war may be overly confident about its ability to identify the causes of war onset, which could explain the predictive failures that have been identified by scholars in the forecasting literature.Footnote 67
Our analysis points to several avenues for future research. First, future work could extend our focus on separatist conflict processes and collect analogous data on ethnic claims for representation at the center. That would allow similar two-step tests of the role of ethnic grievances and other factors in ethnic conflicts related to control over the center, including a direct test of one of our key predictions—that exclusion has a weaker association with the emergence of nonviolent separatist claims because many excluded groups instead mobilize for inclusion.
Second, we employed a broad understanding of nonviolent separatist claims that includes both institutional and extra-institutional mobilization. While this allowed us to more reliably cover the nonviolent formative stages of separatist war, it could also be instructive for future research to disentangle escalation patterns following different forms of nonviolent claim making. Of interest would be to explore nonviolent escalation from conventional claim making to extra-institutional protest.
Finally, even though the number of cases we had to work with was small, our finding that recent autonomy revocations make escalation to separatist war more likely suggests that a cognitive shift away from structural indicators to more fine-grained data and dynamic models could prove the key to increasing the specificity of civil war models and improving out-of-sample predictions.Footnote 68 A promising avenue for further research would be to collect detailed event data on both violent and nonviolent government responses to groups making claims for self-determination.
Data Availability Statement
Replication files for this research note may be found at <https://doi.org/10.7910/DVN/EI1JMG>.
Supplementary Material
Supplementary material for this research note is available at <https://doi.org/10.1017/S0020818320000557>.
Acknowledgments
We thank Brendan O'Leary, Lars-Erik Cederman, Bridget Coggins, Kathleen Cunningham, Ed Mansfield, David Siroky, and participants of panels at the 2018 Annual Meeting of the American Political Science Association and the 2018 Conference of the Swiss Political Science Association, participants in a 2018 seminar at ETH Zurich, as well as two anonymous reviewers and the editors of this journal for helpful comments. We owe special thanks to Andreas Schädel for his collaboration on the SDM data set.
