email icon Email this citation


Fearful, but not MAD — Modeling the Security Dilemma of Conventional Conflict

Erik Melander
Uppsala University

International Studies Association

March 1998

Abstract

This paper presents a simple and straightforward model of the security dilemma in settings in which the parties lack secure second–strike capabilities with weapons of mass destruction. The model includes first–strike advantages and incomplete information as to the antagonist's preference ordering. Imperfect information is used to simulate mutual fear of surprise attack.

An important results is that moderate or "dovish" actors may rationally initiate preemptive attacks under a wide range of conditions.

The central propositions suggested by Robert Jervis in his seminal 1978 discussion on the security dilemma are in this paper subsumed under a logically coherent game theoretic model.

The results hold for a number of reasonable variants of the possible types of actors involved.

Introduction and Purpose

The application of the security dilemma theory, originally developed in the context of international relations, to interethnic relations within states has started a new vital branch within the scholarly study of the causes of ethnic cleansing and warfare (Kaufman 1994/95, 1996a, 1996b, 1996c, 1996d, Kaufmann 1996, Lake and Rothchild 1996, Posen 1993, Van Evera 1994). The security dilemma consists in that defensively motivated actors, through efforts to enhance their own security, may provoke fear and countermeasures, with the result that less rather than more security is obtained in the end (Herz 1950, Jervis 1978, Buzan 1991). Accelerating negative spirals of such countermeasures may lead to preemptive wars that "no one wants."

So far the application of the security dilemma logic to ethnic conflict has mainly given rise to theoretical work illustrated with a few examples. Only a handful of case studies have been identified which explicitly aim to test the security dilemma logic in the context of ethnic conflict (Kaufman (1994/95, 1996a, b, c, d; Posen 1993). I have been unable to find any study of this phenomenon which subjects a large number of systematically selected cases to statistical analysis. The objective of the present study is to make a first contribution to the process of filling in this lacunae.

A central assumption underlying the application of the security dilemma theory to ethnic conflict is that anarchy may arise within states between hostile or distrustful ethnic groups when the ability or willingness of the state to provide protection is being questioned. It can be argued that this is especially likely to happen when societies undergo political transitions, i.e. when new states emerge, and when states democratize or move in the direction of increased autocracy. Such political transitions create uncertainty, which easily can evolve into the fear that feeds the security dilemma. Indeed, recent empirical studies indicate that political transitions are perilous for ethnically divided societies (Gurr 1994, Ellingsen 1995, Jakobsen 1995). At the same time, most ethnic conflicts in societies undergoing political transition do not escalate to high levels of violence. This pattern presents an intriguing puzzle of high substantial importance. The more precise purpose of this study is to perform a test of the extent to which the security dilemma theory can explain this variation in the intensity of ethnic conflicts within states undergoing political transition. This focus on the security dilemma and political transition sets this study apart from previous quantitative studies of the causes of ethnic violence.

The dependent variable, labeled Restraint Breakdown, is an indicator of whether large scale ethnic violence erupted or not in a particular ethnic conflict dyad during the period of study. 1 The six independent variables have been chosen so as to cover three central aspects of the security dilemma, namely status quo satisfaction, first–strike advantages, and fear. The explanatory power of these independent variables are evaluated using a set of 93 ethnic conflicts during the years 1990–94. This set represents all ethnic conflicts during the period of study which were situated in a context of political transition and did not produce high levels of ethnic violence during the 1980s. The conflicts and many of the codings are derived from the files of the Minorities at Risk Phase III (MaR) project which was developed under the leadership of Ted Gurr. 2

Theory: The Security Dilemma in Ethnic Conflict

The concept of the security dilemma is usually attributed to the neorealist paradigm. One central thesis within neorealism is that the magnitude of anarchy varies within the system of states (Buzan 1991). The basic definition of anarchy is the absence of government. The consequence of anarchy is a competitive, self–help, political environment. Only in anarchy can the security dilemma operate. The central defining feature of states is sovereignty, which means the rejection of overarching government (Buzan 1991: 146f). Ideally, within states government — as opposed to anarchy — should reign. But the degree of anarchy within states is not necessarily nil, but may rather vary to a significant extent.

Following Posen (1993), several scholars have drawn attention to the possibility of anarchy emerging in relations between ethnic groups within a state. When the state becomes unable or unwilling to provide physical security, and does no longer uphold legitimate control over the use of force, ethnic groups may organize themselves militarily for self–defense or for settling scores with other ethnic groups. Similarly, if the state security organs themselves execute attacks against members of a certain ethnic group, it becomes clear to the targeted group that it can no longer rely on the government for physical protection. This problem of emerging anarchy may also arise when the actors believe that the state is unwilling or unable to provide protection.

The possibility of anarchy emerging within states is a general condition which makes the security dilemma between ethnic groups conceivable. A more specific precondition of any security dilemma is that in case of confrontation it is better to move first then to wait for the adversary to make his or her move. This is to say that there must be first–strike advantages associated with the available actions. First–strike advantage refers to a benefit which might be secured in a conflict situation by acting before one's adversary, or at least simultaneously with the adversary so as to deny him the benefits of an un–matched first strike. If there were no such first–strike — or first–move — advantages, a defensively motivated actor could never lose anything from waiting and observing the adversary's move before the decision is taken whether to take confrontational action or not. There would thus be no incentive to undertake any efforts to gain security until — and if — an adversary revealed hostile intentions, and a defensively motivated actor would not be pushed to unwillingly undermine others' security. First–strike advantages in ethnic conflicts will be more thoroughly discussed in section 2.2.

A second prerequisite for a security dilemma is that the means to gain security undermine others' security. The prime example is the preemptive attack: one actor launches an attack in order to reduce the potential threat from another actor, and this attack in itself certainly constitutes a very grave threat to the targeted actor's security. The Israeli attack on Egyptian airfields in the opening of the 1967 war is the most frequently cited example of a preemptive attack in the international context (Levy, 1987: 90). But also more commonly used means to gain security — armaments and military preparations — typically undermine others' security. This is so since it is difficult or impossible to distinguish between offensive and defensive armaments and preparations (Levy 1984; Posen 1993, 28f; see also Lynn–Jones 1995 and Shimshoni 1991). If there were purely defensive means to security available, defensively motivated actors could in principle acquire only these means and take these actions, thereby negating the security dilemma.

Ethnic groups in severe conflict, Posen (1993) observes, tend to organize technologically relatively unsophisticated armed forces, mainly based on infantry. The offensive potential of such forces against other, similar forces, is to a large extent determined by the commitment and quantity of the troops. Provided that the combatants are highly motivated, presumably by ethnic sentiment, such forces can possess considerable offensive potential. Thus, ethnic groups in conflict tend to organize armed forces which cannot be categorized as either defensive or offensive, and which may pose formidable threats to neighboring ethnic groups. Posen notes that the principal way of strengthening the cohesion of these groups' militaries — the promotion of nationalist sentiment — is likely to be seen as very threatening to neighboring groups, even if this nationalism is spurred by defensive considerations (Posen 1993: 29f).

The third prerequisite of a genuine security dilemma may perhaps seem trivial, but deserves to be spelled out nevertheless: there must exist enough uncertainty about adversaries' intentions that a hostile move neither can be excluded, nor is guaranteed. Directly observing others' preferences is impossible. There is therefore always some uncertainty regarding whether others' intentions are good or evil. What is possible for conflict actors is to form estimates based on observed behavior and signals. Scholars have drawn attention to the role of past conflict behavior — ethnic cleansing, war, etc. — and present signals in the form of propaganda and threats, in forming people's fear of — or trust in — members of other ethnic groups (Fearon 1994: 28f; Horowitz, 1985: passim; Lake and Rothchild 1996: passim; Kaufman 1996a: 150f; 1996b: 13, 1996c: 109f; Kaufmann 1996: 147, 154; Posen 1993: 107; Van Evera 1994: passim).

Given that the preconditions of the security dilemma are fulfilled it becomes interesting to ask under what conditions the security dilemma becomes more acute, i.e. more likely to override the actors' initial restraint on the use of force? Building on the game model "Stag Hunt", Jervis (1978) made use of informal reasoning, or what O'Neill calls "proto–game theory" (1988, cited in Jervis 1988), to derive a number of propositions about the security dilemma. Jervis suggested that the cooperative outcome (i.e. that restraint on the use of force to regulate conflict prevails) is more likely [1] the greater the gains of mutual cooperation; [2] the smaller the costs an actor incurs if it cooperates and the other does not; [3] the smaller the gains from not cooperating when the adversary tries to; [4] the greater the costs of mutual noncooperation; and [5] the greater each side's expectation that the other will cooperate. 3 Factor [1] can be equated with the utility to the conflict parties of the present status quo. Factors [2] and [3] are captured by the magnitude of the first–strike advantages. Factor [4] would be the cost of mutual confrontation and factor [5] would correspond to the amount of fear or trust between the conflict parties. The independent variables selected for this study are intended to capture aspects of these exacerbating or mitigating factors: 4 the value of the status quo is represented by the variable Status Quo Satisfaction; the magnitude of first–strike advantages by Ethnoterritorial Dominance, and Ethnic Exclaves; the level of fear by Double Minorities, Democracy and Democratization.

The Status Quo: Infant Mortality and Grievances

According to the theory of the security dilemma, the more the present situation is valued, the more wary the actors become of upsetting the status quo by initiating preemptive action. Thus, the value of the status quo can be expected to be negatively related to the risk of escalation in a security dilemma. Utilizing various objective indexes of standards of living and quality of life, as well as human rights indexes, is one possible way of operationalizing this proposition. In this study infant mortality, the ratio of reported deaths of infants under one year old per thousand live births, will first be used as a rough measure of the utility of the status quo. The magnitude of infant mortality was found to be positively related to ethnic war in a recent large–n study (Esty et al 1997). 5

An alternative approach to operationalizing the utility of the status quo is to look for grievances publicly voiced by members of the ethnic group in question. MaR collected and coded data on a number of such indicators, which have been used in several studies on ethnic mobilization, protests and violence. Dudley and Miller (1998) found positive relationships between grievances about autonomy and rebellion, as well as between grievances about political rights and rebellion (economic grievances and grievances about cultural rights were not significant, however). For the purposes of this study, data about autonomy grievances and political grievances will be used to construct three alternative indicators for Status Quo Satisfaction.

The theoretical expectation is that High Infant Mortality, Separatist Grievances, Autonomy Grievances, and Political Grievances are positively related to Restraint Breakdown.

First–strike Advantages: Ethnoterritorial Dominance and Exclaves

A recurring proposition in the literature on the security dilemma in ethnic conflict is that ethnically more diverse areas are more susceptible to the escalatory logic of the security dilemma than are areas in which each ethnic group is dominant in demographical terms. (Kaufman 1994/95: 293f, 1996a: 151, 1996b: 14, Kaufmann 1996: 148, Lake and Rothchild 1996: 53, Posen 1993: 109, Van Evera 1994: 19) Two ethnically very heterogeneous countries afflicted by ethnic conflict may differ completely with regard to the extent that the subregions in this sense are characterized by ethnoterritorial diversity or dominance. 6 For example, in Czechoslovakia before the split in 1993, Czechs and Slovaks to a high degree lived separated in ethnically relatively homogenous parts of the state, each part being clearly demographically dominated by respective group. In contrast, in Bosnia–Herzegovina before the outbreak of war in 1992, Bosniaks (Bosnian Muslims), Serbs and Croats lived mostly intermingled with each other in ethnically very diverse areas. When different ethnic groups live interspersed in the same territory like this, each side is extremely vulnerable to attacks from the other, since prospective assailants have immediate access to more or less defenseless targets in other ethnic groups' civilian population. Population defense becomes very difficult under these circumstances. As the experience from guerrilla warfare testify, it is extremely hard to win military control over a territory if it is contested by a guerrilla movement supported by a substantial fraction of the population. Therefore each side may perceive an incentive to drive away the other side's population in order to render a territorial defense possible, and thereby alleviate this situation of vulnerability. Ethnic cleansing and even genocide can possibly be analyzed as a form of preemption (Kaufman 1996a: 151, 1996b: 14, Kaufmann 1996: 148, Posen 1993: 109, Van Evera 1994: 17f). The dichotomous independent variable Ethnoterritorial Dominance captures whether an ethnic conflict is situated in a context of ethnoterritorial diversity or dominance in this demographical sense, and the theoretical expectation is that Ethnoterritorial Dominance is negatively related to Restraint Breakdown.

A second form of geographically induced first–strike advantage concerns Ethnic Exclaves, or "islands of one group's population ... stranded in a sea of another" (Posen 1993, p108). An example would be the predominantly Armenian–inhabited Nagorno–Karabakh region in Azerbaijan, which became the stage for a war between Armenians and Azeris as the Soviet Union dissolved. Nagorno–Karabakh was separated from the territories of Armenia proper by a strip of land inhabited mainly by Azeris and Kurds. A scholarly view is that such islands of settlement often are very vulnerable to encirclement, and siege or attack. This is so because the defender is compartmentalized in isolated pockets, which cannot easily support each other, whereas the attacker can mass forces against one pocket at a time. A basic principle of offensive military action is to concentrate superior forces at a point in the defender's positions so as to overwhelm the defending forces in a bit at a time (O'Sullivan and Miller 1983: 49; see also Kaufmann 1996: 148). Fellow ethnics in the wider region may therefore perceive a need to come to their exposed brethren's help. The defense of such exclaves, or irredentas, can be reinforced from other, less vulnerable settlements or, alternatively, the exclaves can be redeemed by offensives breaking the encirclement. There are clear incentives to act preemptively in such situations; a resolute aggressor would like to impose his or her will regarding the enclaves while they are still more vulnerable, and a defender would like to forestall any attacks by way of securing the exclaves before the aggressor moves (Kaufman 1996a: 151, Posen 1993, Van Evera 1994: 18f). The dichotomous independent variable Ethnic Exclaves reflects whether or not an ethnic conflict is situated in a context of such ethnic exclaves. 7 The expectation is that Ethnic Exclaves are positively related to Restraint Breakdown.

Fear: Double Minorities, Democracy and Democratization

A third geographical aspect of the ethnic security dilemma has to do with "double–minority situations" in which the minority ethnic group in the region of conflict make up the majority in the wider region. An example is the situation in Croatia before the outbreak of war in 1991. The Serbs constituted a majority in Croatian Krajina, but a minority in Croatia as a whole. Conversely, the Croats were the majority in Croatia, but a minority in relation to the Serbs if the whole of Yugoslavia was the perspective used. When this kind of situation is at hand both antagonistic groups in an ethnic dispute may simultaneously harbor the fears associated with being outnumbered, depending of the frame of reference being used. (Horowitz 1991: 455, see also Kaufman 1996d: 114) 8 Double minorities would then fuel the fears which drive the security dilemma. The forth independent variable, Double Minorities, indicates whether the conflict is situated in a context of double minorities in this sense, and the expectation is that Double Minorities are positively related to Restraint Breakdown.

A fifth independent variable to be included is the level of democracy. A previous study found a negative relationship between democracy and violent ethnic conflict (Gurr 1993b). It can be argued that the availability of peaceful mechanisms for regulating conflict and the political norms associated with democracy promote trust between ethnic groups in conflict, and thus reduces the risk that a security dilemma will escalate. This reasoning is basically analogous to the so called cultural explanation within IR theory for the separate peace among democracies (see for example Russet 1993). Thus, Democracy ought to be negatively related to Restraint Breakdown.

Whereas a context of democracy is expected to mitigate ethnic conflict, the sixth independent variable, Democratization, is expected to exacerbate the security dilemma between ethnic groups in conflict. The rationale behind this is that democratization often opens up for increased political participation faster than institutions, such as political parties committed to the democratic process and a free media, develop that can channel the participation and provide antidotes to chauvinistic appeals. (Gurr 1993b, Huntington 1968, Snyder 1993, Van Evera 1991) If this argument is correct democratization may contribute to a climate of interethnic fear, and therefore the expectation is that Democratization is positively related to Restraint Breakdown.

The Domain of Application

It can be argued that the security dilemma theory overstates the degree of mistrust among ethnic groups if the logic is indiscriminately applied to all interethnic relations (Cederman 1996). As pointed out among others by Fearon and Laitin (1996), most ethnic groups have basically peaceful and trustful relations with each other most of the time. It might therefore seem somewhat farfetched to assume that a security dilemma may affect, for example, the present–day relationship between the Finnish majority and the Swedish minority in Finland. The theoretical domain of application of the security dilemma therefore needs to be specified. As previously mentioned uncertainty with regard to what other actors are up to and their ultimate goals is a central cog in the logic of the security dilemma. The less predictable the situation, the more likely are worst–case scenarios and fear — "the passion of our nature which excites us to provide for our security, on the approach of evil" (Webster's New Twentieth Century Dictionary, 1978) — to take hold.

Abrupt changes in the authoritative structures in a society are likely to induce such uncertainty. New rules for decision making and new relationships between actors need to be developed. Often untested regimes with doubtful or at least not substantiated democratic convictions assume power.

Furthermore, it seems that there must exist some minimal level of prior distrust or hostility for a security dilemma to arise between two ethnic groups. Given the large number of ethnic groups in the world, possibly up to 5.000 according to some sources (Nietschmann 1987, cited in Gurr 1993a, p5), and the relatively scarcity of ethnic conflicts, it can be inferred that ethnicity only rarely is a component of serious social conflict. The security dilemma theory does not purport to explain why low–level hostility and distrust arise in some interethnic relations and not others; that is the domain of other theories. Rather, the security dilemma theory aims to explain when restraint breaks down in already existing conflicts between ethnic groups. Against this background, the security dilemma theory will in this study be applied to low–level ethnic conflicts in states undergoing political transition.

Data and Operationalizations

To my knowledge, only one large–n data set covering ethnic conflicts exists; the Minorities at Risk Project Phase III (MaR). This dataset will therefore be used as a point of departure for the quantitative analysis. The Polity III dataset will be used to identify cases of political transition. The time period studied will be 1990–1994. Only for the these years are there yearly codings in both datasets. Although data availability thus has been an inescapable underlying constraint, the 1990–94 time period is actually excellent from a theoretical perspective. It covers the end of the cold war and the wave of political transitions that accompanied it. A hail of new states splintered off established states in a manner unprecedented during the Cold War. The number of ongoing wars in general in the world peaked in 1992 (Wallensteen and Sollenberg 1995). When ethnic conflicts are concerned, the global number of new outbreaks of serious violence jumped from four new ethnorebellions in 1989 to 13 new rebellions in 1990 and then started to decline again in the following years (Gurr 1996). Given the emphasizes in the theory on uncertainty, the security dilemma should therefore be highly relevant for this unstable and fluid period. The high number of escalations of ethnic conflicts is also an advantage in that the variability in the dependent variable, Restraint Breakdown, is maximized. This is important since restraint breakdown is a rare phenomenon, as are all forms of extremely violent conflicts.

Thus, we are to select for study all those ethnic conflicts which were situated in a context of a political transition. The ethnic conflicts will be selected from the MaR dataset, which includes groups that meet one or both of two main criteria:

"The group collectively suffers, or benefits from, systematic discriminatory treatment vis a vis other groups in a society.

The group is the basis for political mobilization and action in defense or promotion of its self–defined interests." (Gurr and Haxton 1996, 2) 9

Examples of ethnic groups that are not included in the MaR and consequently not eligible for the selection process are the Swedes of Finland and Walloons of Belgium. The operational definition of political transition will be presented in section 4.1. Examples of ethnic groups included in the MaR but which will not be selected for this study because no political transition occurred in their country are the Sardinians of Italy and the Turkmen of China.

A third selection criteria is that the conflicts studied have not recently gone through a phase of restraint breakdown. This is important since the likelihood of the occurrence of restraint breakdown probably is related to the amount of time passed since the latest instance of restraint breakdown, if any, in that particular conflict. This is part of a general problem of using observations that follow one another in time (time series data); there is a risk that the observations are not independent from each other. In other words, the probability of a conflict in a dyadic relationship between a state and an ethnic group or between two ethnic groups is likely to be dependent on the conflict history of that dyad. In the context of international relations, it has been found that the probability in a certain year of a militarized dispute in a state dyad is almost 25% if there was a militarized dispute in the immediately preceding year. The second year after a dispute the risk of a new dispute falls to about 5%, and the third year the risk drops to 2%, where it remains the following years until another dispute occurs (Beck, Katz and Tucker 1997). Therefore, to minimize the impact of the prior conflict history on the sample, cases in which restraint breakdown (to be operationally defined in section 4.2.) was reported in the 1980s will be excluded. This means that all the admitted cases have been free from restraint breakdown for at least ten years. Examples of groups excluded because of this criterion are the Indigenous Peoples of Nicaragua and the Pashtuns of Pakistan.

One selection criteria remains to be presented before we turn to the specific operational procedures. In the literature on the security dilemma in ethnic conflict it has been argued that ethnic groups that are predominantly urban cannot be expected to face the same kinds of geographically induced preemptive incentives as other groups do. This is so, the argument goes, since urban groups lack a territorial base (Kaufman 1996b: 14, see also Kaufman 1996d: 114). As a consequence, the prospect of preemptively striking at an adversary group in order to capture and secure a defensible piece of territory is not a realistic option for urban groups. For theoretical reasons the kind of security dilemma dealt with in this study would therefore not apply to urban groups that lack a territorial base. Against this background the ethnic groups which are characterized as predominantly urban in the MaR will be excluded. Examples of such excluded groups are the Chinese in Panama and the Russians of Azerbaijan.

Selecting the sample

In this section I will explain the exact operational procedures according to which the sample was constructed. The first selection criteria is really simple: only ethnic groups which are included in MaR are considered. The total number of ethnic groups in MaR is 268.

For the coding of political transition the Polity III dataset is used. The first way of defining political transition is regime change. Polity III characterizes regimes according to two dimensions: the degree of democracy and the degree of autocracy. For details about these variables and codings, the reader is referred to Jaggers and Gurr (1995). 10 Here, political transition will be defined as a change in either the democracy or autocracy score with two or more. The political transition period will be considered to start with the year in which either score differs with two or more from the previous year, and then the transition will be considered to last the following five years. For example, Nicaragua was coded for score 2 on the autocracy variable and 1 for democracy in 1988. In 1989 the scores were 0 and 6 respectively. Nicaragua is consequently considered to be in political transition during the years 1989–1994. Since the five years following the initial transition are admissible for inclusion, and we are to study the period 1990–94, we need to start the search for political transitions in 1985 and check every year up to and including 1994.

Polity III also registers other abrupt changes to the regime characteristics of a state. For the purposes of this study, the following events as coded in Polity III will also be considered as instances of political transition. The year of the event and the five following years will be the period of transition:

Transition

Period during which new institutions are planned, legally constituted, and put into effect. Democratic and quasi–democratic polities are particularly likely to be so established, in a procedure involving constitutional conventions and referenda. (Jaggers and Gurr 1996, 5)

A final indicator of political transition is the break up of an existing state. The year of the dissolution and the five consecutive years will be considered as political transition for the new state(s) as well as the rump state.

Now things start to become a little complicated. From the list of politicized ethnic groups in MaR the next step is to construct conflict dyads, since the unit of analysis will be ethnic conflict dyads. In MaR groups are coded for two dependent variables representing violent ethnic conflict: Rebellion and Communal Conflict. The former variable signifies action against "those who claim to exercise authority over the group" (Haxton and Gurr 1996, p95), which usually is the government of the state in which the group in question lives. Communal Conflict reflects "open hostilities between the minority and other communal groups, acting privately" (Haxton and Gurr 1996, p94). In order to construct the dyads I have to decide whether a certain ethnic group is in conflict mainly with the government (Rebellion), mainly with another ethnic group (Communal Conflict), or both but in separate conflict dyads. As a guidance for this coding I look at the magnitudes of the Rebellion and Communal Conflict variables, and at a variable which names the group's main adversary communal group, if any, in Communal Conflict. Other information in the profiles of each group which accompanies MaR and other sources have also been used for several cases.

I have been restrictive in allowing that a group is involved in more than one conflict dyad. The reason is that I want to avoid inflating the number of observations. Rather I have tried to identify the main conflict for each group. A good example of an exception to this ambition is the situation in Bosnia–Herzegovina where I have constructed three main dyads; "Bosnian Serbs versus the government of Bosnia–Herzegovina", "Bosnian Serbs versus Bosnian Croats", and "Bosnian Croats versus the government of Bosnia–Herzegovina." In most cases it is of no importance for the purposes of this study whether a dyad is defined as "Group A versus the government of..." or "Group A versus 'The Majority Ethnic Group of state...". An example of the former case is "Slovaks versus the government of the Czech Republic," and an example of the latter type of dyad is "Roma of Hungary versus Hungarians".

Operationalizing Restraint Breakdown

The next step is to find all instances of Restraint Breakdown, to be defined precisely within short, in these dyads. If restraint breakdown occurred at anytime during the 1980s, that dyad is excluded for reasons which were explained in the previous section. If an instance of restraint breakdown is identified during the period of study, 1990–94, the next step is to check whether this event preceded the political transition or not. If it did, the dyad is excluded. For example, the Tuareg of Mali initiated a guerrilla war against the government in 1990. Mali is coded for regime change in 1991, i.e., after the outbreak of ethnic warfare. Consequently the dyad "Tuareg of Mali versus the government of Mali" is excluded. When restraint breakdown occurs in an admissible conflict dyad in a given year during the period of study (1990–94), that year is included and counted as an instance of change in the dependent variable. The case is then excluded from the remaining period of study, for the same reasons as referred to previously. For example, in the dyad "Abkhazians of Georgia versus the government of Georgia" an extremely violent civil war started in 1992, and consequently the years 1993 and 1994 are not included for this particular dyad.

Before we can construct the list of cases we then need a definition of Restraint Breakdown. The main theme of the security dilemma theory as applied to ethnic conflicts is the possibility of preemptive attacks and waves of ethnic cleansing directed against government representatives and/or other ethnic groups. This theme dictates that restraint breakdown must be conceived of as very drastic and violent events. The security dilemma theory is not up to explaining, for example, such events as when a few Welsh nationalists blow up a phone boot as a protest against the British government. Rather, the theory is concerned with massive outbreaks of violence and coercion involving large segments of ethnic groups, as in Bosnia–Herzegovina in 1992.

The MaR variables Rebellion and Communal Conflict will be used to code the dichotomous dependent variable Restraint Breakdown. The MaR–variable Rebellion captures a spectrum of organized ethnic violence from political banditry to protracted civil war. Communal Conflict covers violent conflict between ethnic groups, ranging from harassment to communal warfare. Clearly, the events corresponding to Restraint Breakdown must be looked for among the upper levels of both Rebellion and Communal Conflict.

In Figure 1 the definitions for the different levels of Rebellion from the MaR codesheet (Haxton and Gurr 1996, p76) are presented. A cutoff point is inserted for the purposes of this study between levels 4 and 5; rebellions at level 5 and above are here coded as instances of Restraint Breakdown.

Figure1. Restraint Breakdown in Rebellions
Rebellion
0 = none reported
1 = political banditry, sporadic terrorism
2 = campaigns of terrorism
3 = local rebellions: armed attempts to seize power in a locale. If they prove to be the opening round in what becomes a protracted guerrilla or civil war during the year coded, code the latter rather than local rebellion. Code declarations of independence by a minority–controlled regional government here.
4 = small–scale guerrilla activity
– – – Restraint Breakdown – – –
5 = intermediate–scale guerrilla activity
6 = large–scale guerrilla activity
7 = protracted civil war, fought by rebel military w/ base areas

The choice of a cutoff point is necessarily somewhat arbitrary as it can not be deduced from the theory. On the one hand, we want to code as Restraint Breakdowns only those instances which are drastic enough to be in line with the theory. On the other hand, if we are too restrictive, there is a clear risk that we will end up with an extremely limited variation in the dependent variable, since such drastic escalations in ethnic conflicts are very rare phenomena. 11

Now we turn to Communal Conflict. The highest level of this variable (which is a score of 6) clearly seems to indicate Restraint Breakdown in the sense of this study: "6 = communal warfare (protracted, large–scale intercommunal violence)" (Haxton and Gurr 1996: 77). However, when we look at the preceding level of Communal Conflict the issue is less clear: "5 = communal rioting, armed attacks" (Haxton and Gurr 1996, p77). It seems possible that this coding incorporates some very different phenomena. When a number of actual cases coded as "communal rioting, armed attacks" are examined, this becomes clearer. For example, in 1992 an extremely violent communal conflict erupted between Ingush and (North) Ossetians, two small Caucasian groups which together number about 600.000 individuals. These events left around 600 people dead, produced approximately 60.000 refugees and almost devastated the Prigorodnyi region of Russia. Military weapons such as assault rifles and grenade launchers were used by organized, durable para–military forces on both sides. This conflict is coded for level 5 of communal conflict, the same score as, for example, the racial riots in 1991 in the USA and the violence involving Roma in Spain in 1991.

Against this background, it becomes necessary to go through all cases coded as Communal Conflict 5 in order to separate instances of Restraint Breakdown from other kinds of events. The criterion used for this process was whether at least 500 people were killed in the events in question. Two additional cases of Restraint Breakdown which made it to the final sample were identified in this manner: the 1992 war between Ingush and Ossetians, and the communal violence and ethnic cleansing in Kenya the same year.

To recapitulate, Restraint Breakdown is defined as levels 5, 6 or 7 on the MaR variable Rebellion or level 6 of the MaR variable Communal Conflict or level 5 of Communal Conflict combined with at least 500 fatalities.

Now we have everything we need to decide whether a particular observation should be included or not in the sample to be studied. The selection criteria presented above delimit a sample of 93 ethnic conflict dyads, involving a total of 391 years of observation during the 1990–94 period. The average number of years per dyad is thus about four.

Operationalizing the Independent Variables

First the four alternative indicators for Status Quo Satisfaction will be operationalized. The first indicator, High Infant Mortality, is the ratio for the previous year of reported deaths of infants under one year old per thousand live births as given by Social Indicators of Development by the International Bank for Reconstruction and Development/The World bank. 12 The value of the preceding year is used in order not to confound the possible causes of Restraint Breakdown with its consequences. 13

The second indicator, Autonomy Grievances, is coded 1 for the group in question if in MaR any concern for any form of autonomy issue was coded as an "issue important for most of the group" or "issue significant but its relative importance cannot be judged" anytime during the period of observation, but before any Restraint Breakdown; otherwise this indicator was coded 0. If autonomy grievances were recorded for a year in which Restraint Breakdown occurred, it was checked with other sources that these grievances were not expressed only after the outbreak of large–scale violence. 14 An alternative indicator of grievances about autonomy, called Separatist Grievances, was also constructed. This indicator was only assigned the value 1 if "union with kindred group elsewhere" or "political independence" was the issue. A similar indicators was constructed for Political Grievances, but was not significant in any of the tests. 15

The next independent variable to be operationalized is Ethnoterritorial Dominance. This raises a number of complex questions about what the relevant regional unit is. Whether a locale can be considered to be ethnically diverse or demographically dominated by one ethnic group may for example depend on whether the village level or the district level is considered. The individual villages may be homogenous but interspersed with villages dominated by the other ethnic group, so that the district is heterogeneous. This pattern was, for example, common in many parts of Bosnia–Herzegovina before the war 1992–95.

A thorough discussion of similar problems can be found in for example Poole and Doherty (1996). There is no universal answer to these tricky questions, whence a pragmatic approach will be taken for the purposes of this study. It seems to be in line with the security dilemma theory to consider the highest possible level of aggregation as the theory deals with outbreaks of wars and waves of ethnic cleansing that involve large segments of the population. When judging whether ethnoterritorial dominance is at hand or not, a territorial unit which encompasses the overwhelming majority of the smaller ethnic group in the conflict dyad will be pictured. If the settlements of this group are very concentrated this territorial unit will be small, if the ethnic minority is evenly dispersed over the whole country the territorial unit will more or less coincide with the borders of the state. The next step is to ask what proportions of the population respective ethnic group makes up within this imagined territorial unit. It is important to realize that the smaller ethnic group, the "minority" of the MaR, may constitute the local majority within its region of settlement.

A number of examples will hopefully help to clarify this idea. Consider the Adzhars of Georgia, who live concentrated in the southernmost part of this Caucasian republic. Within this small area the Adzhars make up almost 83% of the population. Intuitively, the settlement patterns of the Adzhars and the Georgian majority are thus not intermingled to any high degree. It seems reasonable to conclude that Ethnoterritorial Dominance characterizes the dyad "Adzhars of Georgia versus the government of Georgia". In contrast, the Russians of Latvia are interspersed over the whole of this Baltic country. They make up about 34% of the total population of Latvia, with the titular nation of Latvians composing a slight absolute majority at 55%. Reasonably, the degree of Ethnoterritorial Dominance ought to be much lower in the Russian–Latvian dyad, than in the Adzhar–Georgian dyad. A third example is the Sandzak Muslims of Yugoslavia. Their settlements are concentrated to a very small part of Yugoslavia — the Sandzak. But in Sandzak, about half of the population is Serb. Thus, although the Sandzak Muslims are concentrated to a small region, the degree of ethnoterritorial dominance is low in this dyad

Now this intuitively notion will be specified for the dichotomous independent variable Ethnoterritorial Dominance: consider the territorial unit demarcated by the area of settlement of the smaller ethnic group in the conflict dyad; does the largest ethnic group within this particular region make up more than 70% of the total population of that region? If so, Ethnoterritorial Dominance is coded 1; otherwise it is coded 0. Note that "the largest ethnic group within this particular region" and the "smaller ethnic group in the conflict dyad" often is one and the same group, but that this does not have to be the case. The example with the Adzhars illustrates the former case. The Hungarians of Slovakia illustrates the latter case. According to MaR they live concentrated in one region of Slovakia, but still only make up 12% of the population of this region.

To code this variable material from MaR has been used as a point of departure, but other sources have been extensively used to make the data as complete as possible. It should be noted right away, however, that it is quite hard to obtain the necessary information to code this variable; data is missing for 22 out of the 93 dyads.

The third independent variable reflects whether a so called double minority situation exists. To decide this, the first step is to check whether the smaller group in the conflict dyad is a minority or majority within the territorial unit demarcated by its area of settlement. If the group is a majority within this territory it follows that it has to be a minority in the country as a whole (otherwise it would not be the smaller group in the dyad). If the smaller group in this sense is a majority in its region but a minority in the state as a whole we thus have a double minority situation and the variable Double Minorities is coded 1. A good example of this would again be the Adzhars of Georgia. If, on the other hand, the group in question is the minority in its area of settlement, it is still possible that we are faced with a situation of double minorities. This is so since this ethnic minority, although it is a minority both in its home region and in the state as a whole, might have ethnic brethren living in neighboring countries. If so, it is checked whether the group in question still is the minority if the total population of both states are considered. For example, the Hungarians of Slovakia is a minority within the region of Slovakia which borders on Hungary. Naturally, the Hungarians are also a minority in Slovakia as a whole. However, if the total balance between Hungarians and Slovaks in the joint population of the two countries is considered, the Hungarians are the majority, and the Slovaks the minority. Double Minorities for the dyad "Hungarians of Slovakia versus the government of Slovakia" is consequently assigned the value 1. An example of a conflict dyad in which the variable Double Minorities takes the value 0 is "Crimean Tartars versus Crimean Russians." Also for this variable missing data is a problem.

The next variable to be operationalized is the presence of ethnic exclaves in the dyad. This is coded in a very simple fashion. If an ethnic group has ethnic brethren living in a neighboring state, it is checked whether a territorial corridor of areas inhabited by ethnic brethren can be traced which connects the main settlements of the group with its brethren in the neighboring state. If it is impossible to trace such a corridor the variable Ethnic exclaves is assigned the value 1; the group lives in an exclave surrounded by strips of territory inhabited by other groups which separates the exclave from its ethnic brethren. An example is the Turkish minority of Bulgaria which is thus separated by areas dominated by ethnic Bulgarian settlements from Turkey. If, on the other hand, such a territorial connection joins the areas of settlement of the minority with those of its brethren, no exclave is at hand, and Ethnic exclaves consequently assigned the value 0. An example would be the previously mentioned Hungarian minority of Slovakia; since this minority lives on the border to Hungary proper it has a direct territorial connection to its ethnic brethren.

It remains to explain the operationalizations for the variables Democracy Level and Democratization. This will be relatively straightforward. Data from Polity III was used. Following Jaggers and Gurr (1995), the two Polity III variables Institutional Democracy and Institutional Autocracy was utilized to determine whether a country is a "coherent democracy" or not: if subtracting a state's autocracy score from its democracy score yields a value >=7, the state was considered to be a coherent democracy that year. If the status in this regard changed for a dyad under the period of study, but no Restraint Breakdown occurred, the status which was recorded in Polity III for the largest number of years was coded for the whole observation period. If the status differed from the immediately preceding year, and Restraint Breakdown occurred, it was checked with other sources whether the change took place before the outbreak of large–scale violence. If so the new democracy status just preceding the outbreak of violence was recorded. If the change in democracy status took place simultaneously with or after the outbreak of violence, the previous status was retained.

Similarly, for each dyad it was examined whether the last change in the Institutional Democracy–Institutional Autocracy score was positive, in which case the value 1 is coded for Democratization, or negative, indicating a move in the direction of less democracy, and as a consequence the coding 0 for the Democratization variable. If several changes in different directions took place, the last change immediately preceding the outbreak of large–scale violence was used if Restraint Breakdown occurred. If Restraint Breakdown did not occur, the number of changes during the period of observation was counted, and the value 1 was assigned to the Democratization variable if the positive changes were the most numerous. If there were more negative changes than positive ones, the value 0 was assigned to the Democratization variable.

Results

This section opens with a descriptive presentation of the data. Restraint Breakdown occurred only in twelve dyads, i.e. in 13% of the 93 studied conflicts. There were no cases of missing data on the dependent variable. In Table 1 the frequencies are given for the dichotomous independent variables.

Table 1. Frequencies for the Independent Variables
  Variable   Frequency (total=93)  Number of Restraint Breakdowns(total=12)
Separatist Grievances    
    0=No Separatist Grievances
75 3
    1=Separatist Grievances
18 9
Autonomy Grievances    
    0=No Autonomy Grievances
46 3
    1=Autonomy Grievances
47 9
Ethnoterritorial Dominance    
    0=Diversity
33 12
    1=Dominance
38 0
Missing data  22  0
 Ethnic Exclaves    
    0=No Ethnic Exclave(s)
73 10
    1=Ethnic Exclave(s)
12 2
Missing data 8 0
Double Minorities    
    0=No Double Minorities
26 2
    1=Double Minorities present
47 8
Missing data 20 2
Democracy level    
    0=No coherent democracy
51 0
    1=Coherent Democracy
42 2
Democratization    
    0=No Democratization
7 0
    1=Democratization
86 12

As a first cut, we can notice that there are two strong and extremely significant bivariate correlations, and two weak but moderately respectively weakly significant, bivariate correlations in the directions predicted by the theory – see Table 2.

Table 2. Bivariate Correlations with Restraint Breakdown
Independent variable Pearson's R Chi–square significance
Separatist Grievances 0,54212 0,00000
Autonomy Grievances 0,18832 0.06935
Ethnoterritorial Dominance –0,45738 0,00012
Democracy Level –0,22039 0,0356

Perhaps the reader is now already impatient at descriptives and bivariate correlations, and immediately wants to see the results from the multivariate logistic regressions. There is, however, one problem that has to be dealt with first: we have two "zero cells". In other words, the dependent variable is invariant for one value of each of two independent categorical variables. There are no instances of Restraint Breakdown when there is Ethnoterritorial Dominance, and all Restraint Breakdowns occurred in Democratizing countries. This will cause severe efficiency problems if we try to run logistic regression (Menard 1995, 68f). The latter zero cell is less of a problem because there is probably no significant relationship between Democratization and Restraint Breakdown in these data anyway (bivariate chi–square significance 0.28961). The problem is potentially much more serious when Ethnoterritorial Dominance is concerned. The bivariate frequencies look very interesting and suggest a strong and significant relationship between Ethnoterritorial Dominance and the absence of Restraint Breakdown. Inductively speaking, and judging from these data, it looks in fact as if Ethnoterritorial Dominance is a sufficient condition for restraint on the massive use of violence to prevail.

A possible solution for this dilemma is to find a way to recode the data so that the zero cell associated with Ethnoterritorial Dominance disappears. If we, for the purposes of the logistic regression, lower the threshold required for the largest ethnic group from >70% to >67% we encounter one case of Ethnoterritorial Dominance among those cases which ended in Restraint Breakdown. Similarly, one dyad which did not result in Restraint Breakdown is also recoded to Ethnoterritorial Dominance when this modified coding rule is applied.

After this recoding we can use logistic regression for multivariate analyses. We then ought to bear in mind that the effect and significance of the relationship between Ethnoterritorial Dominance and Restraint Breakdown were more pronounced with the original coding. 16 In Table 3 the results are presented from the regression with all six independent variables included, and Separatist Grievances used as the indicator for Status Quo Satisfaction. 17

Table 3: Multiple Logistic Regression
N 62 (31 cases rejected because of missing data)
–2 Log Likelihood 54,784            
Goodness of Fit 23,068            
0,387            
   Chi–Sq. df Sign.        
Model Chi–Square 33,603 6 1 0,0000      
               
Variable B S.E. Wald df Sig R Exp(B)
Separatist Grievances 1,8785 0,6684 7,8983 1 0,0049 0,3281 6,5435
Ethnoterritorial Dom. –1,698 0,8418 4,0710 1 0,0436 –0,1944 0,1830
Ethnic Exclaves –0.6178 0,8660 0,5089 1 0,4756 0,0000 0,5392
Double Minorities 0,7574 0,7867 0,9270 1 0,32356 0,0000 2,1328
Democracy Level –1,7546 0,8285 4,4857 1 0,0342 –0,2130 0,1730
Democratization 2,3852 28,1782 0,0072 1 0,9325 0,0000 10,8610

First we can notice that for the model as a whole the goodness–of–fit as measured by (Menard 1995, 22f) is good, and the significance is extremely good. Three independent variables are significant at 0.05–level or better and show the predicted signs and substantial effects.

Before these results are further discussed, a few words about missing data are required. At a first glance this is a potentially very serious problem. For one of the most interesting variables, Ethnoterritorial Dominance, data is missing in 22 cases. In the regression presented above 33 cases were excluded because data is missing on different variables. However, data is not missing in any of the twelve cases which resulted in Restraint Breakdown, except for the independent variable Double Minorities (2 cases of missing data when Restraint Breakdown occurred). An alternative test is therefore made with this variable removed. The number of missing cases are then down to 26 and the results are stronger than in the original equation. If only the three significant variables are included we arrive at 22 missing cases and even better results. Finally, I tested what would happen if all the missing cases were recoded as instances of ethnoterritorial diversity (i.e. Ethnoterritorial Dominance = 0), which is very unlikely and the toughest possible case for the hypothesis tested. Ethnoterritorial Dominance was still weakly significant (p=0.0691 at worst) in all the tested variants of the regression equation (the number and combinations of other independent variables varied). The other two significant variables, Democracy Level and Separatist Grievances, stayed significant in all tests. I therefore conclude that the problem with missing data is of minor consequence for the results of this study.

The substantial effects of respective independent variable can be read from the logistic regression coefficients B, the partial correlations R (see Norusis p5f), or the odds ratios Exp(B) ("e raised to the power Bj is the factor by which the odds change when the jth independent variable increases by one unit", Norusis, p6). However, it might also be interesting to see the predicted probabilities of Restraint Breakdown based on the model, since this gives an intuitively easily comprehensible picture of the effects. In Table 4 these probabilities are presented for the different possible combinations of the three significant dichotomous independent variables. The model is obtained from the regression including these three variables as independent. The combinations of variable values are ordered from highest risk of Restraint Breakdown to lowest.

Table 4. Predicted Probabilities Based on the Statistical Model
Ethnoterritorial Dominance Separatist Grievances Coherent Democracy Odds p
No Yes No 0,469 0,32
Yes Yes No 0,0933 0,0853
No Yes Yes 0,0872 0,0802
No No No 0,0826 0,0762
Yes Yes Yes 0,0174 0,0171
Yes No No 0,0164 0,0162
No No Yes 0,0154 0,0152
Yes No Yes 0,0031 0,0031

From this table we can see that the predicted risk of Restraint Breakdown is extremely low, 3°/oo, when the conditions are the most favorable, i.e. when there is Ethnoterritorial Dominance, no Separatist Grievances, and when there is Coherent Democracy. If one of these favorable conditions change, the predicted risk is almost 2%. If a second favorable condition drops out the risk increases to around 8%, and when the situation is the least favorable the risk is up to 32%.

The dampening effect of Ethnoterritorial Dominance on the risk for large–scale ethnic violence is especially noteworthy, as it has not been demonstrated in any previous study of which I am aware. The Exp(B) values of Table 3 tell us that Ethnoterritorial Dominance reduces the odds of Restraint Breakdown by 82% compared to ethnoterritorial diversity, when all other independent variables are held constant. If it would have been possible to keep the original operationalization also in the logistic regression, the effect would have been stronger. However, we have to keep in mind that large scale ethnic violence is a very rare phenomenon, whereas ethnoterritorial diversity is common. Judging from the data of this study, nearly half of the ethnic conflicts are situated in a context of ethnoterritorial diversity. It follows that Ethnoterritorial Dominance only can account for a limited part of the variation in Restraint Breakdown.

The likewise dampening effect of Democracy is in line with the findings of Gurr (1993). In the logistic regression the effect is almost identical to that of Ethnoterritorial Dominance (i.e. after the threshold was lowered from >70% to >67% for Ethnoterritorial Dominance).

The aggravating effect of Separatist Grievances is paralleled in the effects uncovered for grievances about autonomy in Dudley and Miller (1998). The effect is quite strong which perhaps is not very surprising. It might seem somewhat trivial that there is a clear relationship between separatism and the outbreak of large–scale ethnic violence. But this result may still be interesting for prediction purposes (early warning). Furthermore, the inclusion of this variable may be viewed rather as a statistical control; even when the effect of this variable has been taken into account, Ethnoterritorial Dominance and Democracy are still significant. It is an interesting question for further research why the other indicators of Status Quo Satisfaction were not significant. It is too early to judge whether the problem lies with the security dilemma theory or the indicators.

That Ethnic Exclaves and Democratization were non–significant might well depend on the very limited variation in these variables. The sample of this study did not allow for a good test, and further research may turn out interesting results also for these variables. But it is also possible that the assumptions that Ethnic Exclaves create first–strike advantages and that Democratization promotes uncertainty and fear simply are wrong. The Double Minority variable is plagued with missing data, and if there is a significant effect waiting to be uncovered, this problem has to be addressed. On the whole, data availability is a problem and the further study of the security dilemma in ethnic conflicts requires more and better data.

Implications

What are the theoretical inferences from this study for the theory of the security dilemma? In the tests three independent variables, Separatist Grievances, Ethnoterritorial Dominance, and Democracy, were significant and had substantial effects. Each of these three variables covers one of the three central dimensions of the security dilemma; Status Quo Satisfaction, First–strike Advantages, and Fear, respectively. This corroborates the confidence in the basic soundness of the theory and its relevance for the study of ethnic conflict. However, the study undertaken here is admittedly somewhat crude, and more refined quantitative tests are needed, preferably in combination with in–depth case studies. Reasonably there must be other causal paths to the escalation of ethnic conflict than the security dilemma logic. Equifinality — that one phenomenon may be caused by several different causes — is probably common in the world outside parsimonious theorizing. Vasques argues convincingly that there are several causal paths leading to international war (1993, 48f), and it would be surprising if the matters were much different when ethnic warfare and cleansing is concerned. There is probably still an enormous amount of research left to do before we can untangle this causal web. The study of the security dilemma in ethnic conflict has only just started.

When it comes to practical implications of this study, it might be that the distinctive pattern uncovered here is practically useful for early warning purposes. Can any tentative policy advice be based on the results of this study? The relationship between ethnoterritorial diversity and an increased risk of serious ethnic conflict has been suspected but not demonstrated previously. It is important, though, to remember that in this study the majority (64%) of the ethnic conflicts situated in a context of ethnoterritorial diversity underwent political transition without the outbreak of large–scale violence. It might still appear that these results render support to the argument that societies characterized by ethnoterritorial diversity are inherently unstable and that the separation of ethnic groups and the exchange of minorities therefore is desirable, preferably before serious ethnic violence erupts. I think that would be a dangerously flawed conclusion. It is hard to see how such exchanges could be executed in such a way that most of the people affected would perceive them as fair and agree to leave their homes voluntarily. On the contrary it is likely that uprooted and dissatisfied members of the resettled groups would form revanchist organizations which would ensure that the hostility between the separated groups would not subside for decades or even centuries. The experiences from the population exchanges between Greece and Turkey in the 1920s and between India and Pakistan in 1947 suggest that high levels of hostility will persist between the separated peoples. India and Pakistan have fought several wars, and it seems likely that Greece and Turkey would have done the same if it was not for the shared NATO–membership and the restraining influence from the USA during most of the time since the exchanges.

Instead of separation a much more practical and in my view ethical approach is the twofold way of democratization and the accommodation of ethnic grievances. If coherent democracy can be achieved and ethnic grievances accommodated before they reach the level of separatism, the risk of large–scale ethnic violence is reduced approximately by a factor of 20 according to the results of this study. In this connection it is interesting to note that autonomy grievances and political grievances were not significant in the multivariate analyses. This could be an indication that these kinds of grievances can be accommodated sufficiently to prevent them from causing large scale violence. It would be especially important to concentrate such efforts in areas of ethnoterritorial diversity. The questions of how democracy can be promoted in multiethnic societies and how ethnic grievances can be accommodated fall outside the scope of this study, but are subject to intensive research by many other scholars.

Bibliography

Notes

Note 1: I use the term Restraint Breakdown to emphasize that I do not just look at ethnic rebellions directed against the state, but that my dependent variable also includes fighting between ethnic groups. This is yet another difference from most previous quantitative studies. Also, I think the term Restraint Breakdown fits the security dilemma theory well. Back.

Note 2: For information on this project see http://www.bsos.umd.edu/cidcm/mar. Back.

Note 3: These propositions have been subsumed under a logically coherent game theoretic model of the security dilemma in conventional conflict (Melander 1997). Back.

Note 4: The exception is the cost for mutual confrontation, for which I have not been able to come up with a useful indicator yet. Back.

Note 5: Another recent study found a negative relationship between per capita wealth and ethnic rebellion (Dudley and Miller 1998). It is highly likely that low infant mortality is related to high per capita income. Back.

Note 6: Following Poole and Boal (1973: 22-23) and Boal (1976: 51), Poole and Doherty (1996: 19f) advocate that the concept of "dominance" is used to express "the deviation of a social composition from the situation in which each social group is found in equal numbers — i.e. in the 50% situation in the two group case." Back.

Note 7: The geographic location of minority settlements acquires an importance also in the wider strategic context. As Kaufmann notes (1996: 163), concentrations near disputed borders and astride critical communication routes are especially likely to spark conflict. Any first-strike advantages in such situations are magnified by the strategic value of the territory. An illustration could be the outbreak of the war in Bosnia in 1992; the fighting started in February in Posavina, the corridor along the border to Croatia linking Bosnian Serb areas in western Bosnia-Herzegovina with Serbia proper (Glenny1994: 184, Silber and Little, 1995: 243). Had the predominantly Bosnian Croat population managed to take control over Posavina, the Bosnian Serb position would have been precarious. The Bosnian Serb areas of western Bosnia-Herzegovina was a vulnerable exclave which was threatened to be cut off from crucial sources of support in Serbia proper. A reasonable hypothesis is therefore that the Bosnian Croats and Bosnian Serbs preempted each other in Posavina.

A factor related to political geography is the presence of military depots which can be expropriated by the side acting first. According to Posen (1993), the numerous depots belonging to the territorial defense were important sources of temptation and worry at the outset of the war in Croatia in 1991. Back.

Note 8: Horowitz writes about this as a problem of "external affinity " but he seems to limit this concept to cases in which the minority in one country has ethnic brethren in another country. Back.

Note 9: For a more detailed description of the selection criteria of MaR III see Gurr and Haxton 1996. Back.

Note 10: Information about this dataset can also be obtained from http://garnet.acns.fsu.edu/~whmoore/polity/polity.html. Back.

Note 11: In order to examine the robustness of the results, the cutoff point between levels 4 and 5 of Rebellion was then lowered to between level 3 and 4 in an alternative test. It turned out that this had no effect since no new cases of Restraint Breakdown qualified to the final sample. Back.

Note 12: Downloaded from http://www.ciesin.org/IC/wbank/sid-home.html. Back.

Note 13: In case of missing data for a given year, the average between the last previous and next immediately following number was used, except in case in which Restraint Breakdown occurred. Then only the last previous value was used. Back.

Note 14: MaR variables AUTREG11-AUTRIG55. The types of issues are "General concern for autonomy," "Union with kindred groups", "Political independence", Widespread autonomy", and "Limited autonomy." Back.

Note 15: MaR variables POLRIG11-POLRIG65. The issues are "Diffuse political grievances", "Communal Rights", "Participation in decision-making", "Equal civil rights", "Change is policy", and "Other political grievances". Back.

Note 16: Intuitively, it seems to me that already 70% is a fairly low threshold for the concept of dominance. It would have been nice to raise the threshold instead as an alternative test, but this is not a practical option with these data. Back.

Note 17: None of the other alternative indicators of Status Quo Satisfaction (Infant Mortality, Autonomy Grievances, and Political Grievances) were significant. The results for the other independent variables were very similar to those of Table 3 across these four regressions. Regressions were also run with interaction terms included, but none of the interaction terms were significant. Back.