SEPARATION OR INCLUSION?
TESTING HYPOTHESES ON THE END OF ETHNIC CONFLICT

R. William Ayres

Introduction

Since the rediscovery of "ethnic conflict" as a key phenomenon in international relations, the question of how such conflicts end has become an urgent one for both scholars and policy makers. We know that some long-running conflicts have been extremely resistant to peaceful settlement-in Palestine and Kashmir, for example. We also know that some "interminable" conflicts have made real progress towards peaceful resolution (e.g. Northern Ireland, Palestine, Bosnia), but that the future of these arrangements is unclear and a return to conflict could happen. Finally, the end of the Cold War produced a new and frightening set of conflicts-in the former Yugoslavia, for example-that have reminded us of the existence and tragedy of "ethnic cleansing" and the brutality of nationalist conflict.

All of this has brought the study of ethnic and nationalist conflict to the fore, and has raised a key question in the minds of both scholars and policy makers: How do violent nationalist conflicts end? Can they be peacefully resolved at all, or are the participants "doomed" to cycles of violence and revenge? On this second question, we do have some evidence (Gurr, 1994; Heraclides, 1997; Ayres, 2000)-violent ethnic conflicts can and do resolve, in various ways. The first question-how-is now more pressing. What do we know-or what can we discover-about the various ways in which these kinds of conflicts can be ended?

One possible answer to this question is that violent ethnic conflict by its very nature leaves only one possible solution: the permanent separation of warring groups (Kaufman, 1996). Violence, in this argument, makes any solution in which the groups still have access to each other untenable. This logic could have very serious policy consequences, both for the participants in conflict and for those outside who attempt to bring about peaceful resolution. It would suggest, for instance, that a "federalist" solution of one multiethnic Bosnia is bound to fail; it would also lend strength to calls by Israel, for example, for "defensible borders". This paper's initial puzzle, therefore, is to empirically explore the link between levels of violence and "possible and impossible" resolutions, using a different set of cases than are considered by Kaufman's original arguments.¹

The necessity of separation as a result of violence is not the only logic regarding the resolution of conflict, however. Others have suggested that conflict endings are reliant on the intervention of outside mediators (Walter, 1997; Zartman and Touval, 1996; Crocker, 1996; Regan, 1996), or the depth of hostile perceptions held by the two sides (Ayres, 1997; Stein, 1996; Kelman, 1987; Kaufman, 1996 also includes this as part of his argument), or the balance of military power between them (Posen, 1993). This suggests a multiplicity of possible explanations for the same outcome-types of conflict resolution. This paper will examine these lines of reasoning empirically, by comparing the characteristics of conflicts and types of settlements reached across some 48 violent nationalist conflicts from 1945-1996. The next section will define the dependent variable by laying out what we mean by "resolution", and explore the logic of the alternative explanations suggested above to translate them into testable hypotheses. Data and methods will then be introduced, and results of a series of correlative tests presented. The final section will conclude with a discussion of what these results mean for the different explanations presented here, and how they might guide future research inquiries.

How Do They End? Measures of "Resolution"

If we are to test these general notions-that violence, or third parties, or perceptions, or military power, condition how conflicts are resolved-we must have some hypotheses that we can test on real-world cases. For this, we need to specify the dependent variable we are explaining: what do we mean by "conflict ending"? There are two ways in which to approach the question of "how conflicts resolve": the results of conflict, and the length of time it takes to reach those results. Addressing what kinds of outcomes end conflicts has been the focus of some previous studies in the arena of civil wars (Mason & Fett, 1996; Licklider, 1995; Wagner, 1993); scholars have looked at whether negotiated or military victories are more likely to last, but these efforts have not looked at the specific terms of the settlement. This is the core of Kaufman's (1996) logic-not only that violence matters, but that in particular it determines which terms of settlement will provide viable solutions and which will not. In particular, he argues that high levels of violence necessitate a particular kind of ending-separation into defensible enclaves-in order to be "possible". Therefore, it makes sense to look to the terms of conflict ending-are the parties separated or not--as an important dependent variable.

It is also useful to ask not only what outcome a resolution produces, but how long it takes to get there. The "conventional wisdom" (at least, as embodied in popular commentary) suggests that violent ethnic conflicts are "intractable"-that high levels of violence, or the depth of historical hatred, make peaceful endings difficult if not impossible. This is also implicit in the logic of Kaufman's argument-unless the "possible" solution of partition is achieved, conflicts will continue for a long time. Similarly, proponents of other explanations also purport to explain conflict length-by reference to how long it takes third parties to get involved, or balances or imbalances in power, or the parties' ability to change perceptions of each other.

Either of these two approaches could be defended as the "outcome" of efforts to resolve conflict; for the purposes of this study we will use both. The next section will explore the logic of each of the four alternative explanations outlined above in light of these different ways of measuring resolution, and bring out testable hypotheses for each.

From Theories to Hypotheses

These two approaches to measuring "outcome", combined with the four different explanations offered above (levels of violence; differences in power; strength of stereotyping; third party involvement), yield four different sets of hypotheses-one set of expectations for the effects on various kinds of outcomes for each independent variable. In addition, since violence itself can be considered a dependent as well as independent variable, the following discussion will include expectations about how the other three independent variables might be expected to affect the level of violence-providing a preliminary look at expectations regarding interaction effects among some of these differing explanations.² This section thus has four sub-sections, each laying out, in hypothesis form, the expected effects of that variable on the terms of outcome and length of conflict.

Levels of Violence

Kaufman's (1996: 137) logic argues that sufficiently high levels of violence necessitate the separation of the warring parties "into defensible enclaves". The corollary to this is that if a settlement is achieved which does not create such separation, it will not actually end the conflict. Hence, the testable version of this hypothesis is as follows:

• Hypothesis 1: High levels of violence should correlate with separation endings; low levels of violence should have a greater chance of producing an inclusion ending.

In addition to predicting to type of conflict ending, the level of violence may also affect how long it takes conflicts to end. This is inherent in the logic of "possible and impossible solutions"; if a particular ending is "impossible" or unsuited to a particular conflict, it will presumably not end that conflict. Thus, following Kaufman's original logic that levels of violence are the ultimate fuel to the cycle of conflict, and the conventional wisdom that bloody ethnic conflicts are particularly intractable, we get:

• Hypothesis 2: Greater levels of violence will be correlated with longer conflicts.

Balance of Capabilities

Level of violence is only one possible explanation, of course. Some might suggest that it is not the level of violence per se, but the balance of capabilities between the parties that creates the "possible or impossible" solutions. Since nationalist conflicts are nearly always between an existing state and a nationalist group, we would expect different outcomes depending on which side has the advantage. Where capabilities are relatively balanced, separation might be necessary, because the nationalist group can continue pressing its demands against the state until the state lets them go. However, where capabilities are strongly in favor of the state actor, the state is presumably capable of enforcing its will on the rebelling group, even to the point of forced ethnic inclusion.³ This leads to the following expectation:

• Hypothesis 3: When capabilities are balanced between the two sides, separation is the more likely outcome, but when capabilities are unbalanced in favor of the state, inclusion is more likely.

It might also be suggested that the length of ethnic conflicts-as with the length of most interstate wars-hinges more on the balance of capabilities between the two than on how much damage they cause. This reflects a very conventional Realist notion: conflict with evenly-matched actors will likely last for longer than those where one side has an advantage, presumably because the more powerful side will press its advantage to victory. Thus:

• Hypothesis 4: The balance of capabilities will explain conflict length; more evenly-matched conflicts are expected to last longer, while unbalanced ones should take less time to end.

We can extend this logic one step further. If the balance of capabilities determines how long a conflict lasts, it may also be the prime determinant of the level of violence, suggesting that the argument that treats level of violence as an independent variable is really focusing on an intervening one. This argument would suggest an additional hypothesis:

• Hypothesis 5: The balance of capabilities will determine the level of violence; greater imbalances will mean lower overall levels of violence (as the more capable side will win quickly and/or decisively), while parties more evenly balanced may experience higher overall levels of violence.

Strength of Stereotypes

There is also the possibility that level of violence (Hypotheses 1 and 2 above) does not exert a direct effect on outcome, but instead influences conflict outcomes indirectly through intermediate causal mechanisms. One such mechanism that political psychologists have recognized as important is the strength of images and stereotypes (Ayres, 1997; Stein, 1996; Kelman, 1987). Kaufman also acknowledges the importance of "harden[ed] ethnic identities" (Kaufman, 1996: 137) in arguing for the importance of violence as a determining factor. It is possible, then, that outcome depends not on violence per se, or on capabilities, but on the extremity of perceptions:

• Hypothesis 6: Strength of stereotyping by the conflicting sides of each other will determine outcomes; high levels of stereotyping will correlate with separation endings, while lower levels of stereotyping will have a greater chance to produce inclusion endings.

This same logic suggests that perceptions and stereotyping should be expected to influence the length of conflict. Part of the "intractability" argument sketched above flows through the logic of images and stereotypes; the stronger and more deeply held the two sides' perceptions of each other, the less likely they are to agree to stop fighting and the more difficult the conflict will be to resolve. Thus:

• Hypothesis 7: Stronger stereotypes by conflict participants will correlate with longer conflicts.

Finally, it is possible that the level of violence, rather than being a function of the relative capabilities of the two sides, is determined by their mutual perceptions of each other. If stereotypes matter, we should expect that stronger ones will provide both cause and justification for greater levels of violence against the stereotyped group. This, we should expect:

• Hypothesis 8: Stronger stereotypes will be correlated with higher levels of violence.

Third Party Involvement

Finally, some scholars and policy makers suggest that what third parties do matters in determining conflict outcomes. This has certainly been the logic behind much US (and UN) foreign policy towards areas torn by ethnic secessionist strife-that careful mediation can (in some cases at least) bring warring parties to live peacefully with each other. On the other hand, one would expect that third parties which get involved for the purpose of supporting one side against the other would only exacerbate the existing problems and make separation more likely.⁴ This leads to separate expectations for mediators and interveners, which can be expressed as one hypothesis:

• Hypothesis 9: Involvement by mediators will be correlated with inclusion, while involvement by interveners will be correlated with separation.

The logic of third-party intervention (as a mediator or biased ally) also suggests that what third parties do makes a difference in how long conflicts last. Many mediation and peace-making missions are undertaken with the explicit hope of cutting short a conflict that is otherwise expected to last into the indefinite future.⁵ On the other hand, intervention by third parties to aid one side in the fighting may only prolong the conflict by providing additional sources of arms and support. Therefore:

• Hypothesis 10: Where third parties are involved as mediators, conflicts should be shorter, while third party interveners should be associated with longer conflicts.

The same argument that leads third party mediators to intervene to shorten conflicts also suggests that their involvement may help to lessen the damage caused-to lower the level of violence the conflicting parties inflict on each other. Conversely, parties that are intervening as allies of one side or the other will likely only exacerbate the level of violence by allowing for escalation or more staying power on the battlefield. Thus, we can expect that:

• Hypothesis 11: Involvement by third party mediators will be correlated with lower levels of violence, while third party interveners will be correlated with greater levels of violence.

Data and Measures

In order to test these hypotheses, we need some set of cases upon which to test them, measures for each of the named variables, and appropriate methods for examining the hypothesized correlations. This section is divided into two parts: the first will address the question of which cases to use, and the second will introduce measures for the variables named in the hypotheses.

Selecting Cases

Before we can test hypotheses or measure variables, we need to know what set of cases are appropriate for such measurement and testing. The literature on "ethnic conflict" is fairly vague on this point; there are few if any widely accepted definitions of what constitutes an "ethnic conflict", a "nationalist conflict", or an "ethnic war". Part of this difficulty stems from the behavior of the parties in conflicts themselves, who often have important political interests in how they define their struggles and how those struggles are perceived by the outside world. Part of the fault also lies with us as scholars; although there have been some debates over what "nationalism" means (Haas, 1986), few have attempted to create a generalizable definition for the term "ethnic conflict", and some reject the term altogether.⁶

I do not propose a definition to which we should all adhere, but I obviously need a working definition for the purposes of this research. This paper will test the above hypotheses on a group of Violent, Intrastate Nationalist conflicts (VINC). These are defined as conflicts between nations-groups of people who give their primary identity to the group and who think their group can and should have a sovereign state (Cottam, 1984)-which take place within states-entities with governmental institutions which hold sovereignty over a definable territory-and which have a history of violence -organized efforts by either side to kill members of the other or destroy values important to them (Ayres, 2000).⁷

In order to study conflict endings and length, we also need definitions of starting and ending points. Conflicts start when a group raises a nationalist demand for statehood and either or both parties actively attempt to deal with that conflict (by fighting, or discussing, or some combination of means), at any time between 1945 and 1996; start and ending dates are coded by month and year. Violence must be involved at some point in the conflict, but not necessarily at all times; prior violence both creates a history which indicates that future violence is possible, and carries forward many of the psychological effects of violence to future conflict interactions. Conflicts end when both sides are no longer either fighting or talking with each other about what the solution to the conflict should be.⁸ Hence, any one conflict dyad or group-say, Kurds in Iraq-could have a number of cases within it, separated by periods during which the parties are essentially doing nothing despite their continued differences.⁹ Individual cases within conflict groupings (such as the Kurds and Iraq) are known as "conflict episodes", to separate each from the broader issue of the "conflict" between any two actors (Ayres, 2000).

Using these definitions, I have put together a list of 77 conflict episodes from 1945 through 1996. Of these, 48 ended prior to the end of 1996; since this study is about how conflicts end, it is this latter group of 48 which will be used to test the above hypotheses.¹⁰

Measuring Concepts

There are six important variables that need to be measured to test the hypotheses laid out above. The first two are dependent variables; we need a measure of conflict outcome (separation or inclusion) and a measurement of conflict length. The latter is easily obtained by comparing the starting and ending dates for each conflict episode (see Appendix A); as these are listed by month, conflict length is measured in months.

To measure the former, each conflict episode was coded dichotomously, based on whether its ending produced a situation in which groups were demographically separated or not. The core criterion for this coding was whether, after the conflict ended, members of either side were able to attack members of the other with impunity, or to otherwise engage in easily performed violence.¹¹ These codings are also presented in Appendix A.

The sets of hypotheses outlined above also require the coding of four independent variables: level of violence; balance of capabilities; strength of stereotyping and perception; and third party involvement. Each requires a somewhat different coding approach, and (across all four) multiple data sources.

Level of violence can be coded two different ways: number of casualties or deaths (adjusted for size of total population), and severity of fighting (from occasional political banditry through various levels of terrorism and guerrilla warfare to full-scale civil war). Both are useful approaches; one assesses the actual amount of damage done, while the other measures the amount of disruption a given population will likely experience based on the behavior of the conflicting sides. This paper adopts both approaches, and uses two separate measures in testing the hypotheses involving levels of violence. Data were gathered on casualty estimates for as many cases as were available (a total of 30 out of 48 cases), from multiple sources (listed in References). These raw estimates are used as one measure of level of violence; they are also divided by total population to get estimates of conflict deaths per 1000 of population.¹² For a behavioral measure of violence, this paper adopts the scale used in Ted Gurr's (1999) Minorities At Risk (MAR) data set; it is an eight-point scale, ranging from 0 (no violence reported) to 7 (protracted civil war). The MAR codings were used where these match to the case list here; additional codings were added by the author, based on case data materials, resulting in codings for all 48 cases.

Balance of capabilities is likewise a complex concept, containing a number of possible important factors. Ideally, we want some measure that captures the relative balance of control between the two actors-that is, the ability of one side to force outcomes on the other. Such measurement must, of course, be without reference to the actual outcomes in question; to do otherwise is simply to code tautologically. While there are a large number of potential factors that influence capability to control (terrain, morale, training and equipment), a few factors should give us a rough estimate of this balance. In intrastate nationalist conflicts of the kind to be examined here, we should expect that population, number of men under arms, and strength of outside allies will all play a role in determining the overall balance of capabilities.

The first of these, population size, is coded for each case in thousands of people. For sub-state nationalist groups, this is simply an estimate of the group's population; for state actors, it is an estimate of the state's population minus the nationalist group, since the state presumably cannot mobilize a nation against itself.¹³ Population is a rough but usable measure of capabilities; it is reasonable to expect that very small nationalist minorities will have less ability to control outcomes by force than those which make up a much larger proportion of the population of the state in which they reside, although this is not always the case.¹⁴ For both states and non-state groups, the highest estimate of population for the time period of a conflict episode is used. Data for these estimates are taken primarily from the World Factbook, published by the CIA; Ted Gurr's Minorities at Risk data; and the US Census Bureau's data on world population.

The second measure, men under arms, is also coded as the highest estimate for each actor for the time period of the conflict episode. For states, this is the total of their armed forces; for nationalist groups, it may be a number of men in an army (for those with organized military arms), or a number of guerrilla fighters, or it may (for particularly disorganized nationalist movements) be none at all. Measuring troops provides a basic measure of military strength and ability to conduct operations on a battlefield-a staple of control efforts. In addition, although counting men is a crude measure for many conventional warfare situations (where measures of the quantity and quality of military hardware may be better predictors of success), it provides the best comparable measure across conflicts which range from conventional warfare through jungle guerrilla tactics to the on-and-off fighting of urban terrorism. Data for states are easily obtained from the Military Balance, published by the International Institute for Strategic Studies, and from the Correlates of War Capabilities data set. Data for non-state nationalist groups is also available, in recent years, from the Military Balance; coding for earlier conflicts must rely primarily on secondary sources and press reports, particularly Reuters World News Service. Troop estimates for the nationalist group actor were unavailable for 3 out of the 48 cases, restricting this variable (and the composite balance of capabilities variable, below) to an N of 45.

Finally, a measure of strength of allies is included, where allies are those other actors which are actively engaged in helping an actor to fight or otherwise attempt to unilaterally control the conflict outcome. Measuring the strength of allies provides another significant input into the capabilities equation. In many cases, minorities with little military strength and/or population can have their causes greatly assisted by a strong ally willing to lend them assistance. To measure this this, both actors receive a single code, based upon the following rankings:

1. No allies
2. Regional Non-State Group
3. Regional State
4. Regional Power/Hegemon
5. Great Power

Each actor receives the code corresponding to its most powerful ally; for example, during its war with the Ethiopian government, the Eritrean People's Liberation Front received assistance from the Tigrayans (a Regional Non-State Group), and from Syria and Saudi Arabia (Regional States, but not powerful or close enough to be Regional Hegemons), so it is coded as a 2. Data for coding ally strength for both actors comes mainly from primary news accounts and secondary case histories.

Given these three factors of capability, we need some way to combine them. This is done by creating standardized scores for each measure, so that they can be combined in a manner that makes mathematical sense. Data for troops, population and ally strength for all actors are thrown into three groups, from which means and standard deviations for the population of cases are calculated. In the case of troops and population, the raw data are converted using the natural log (ln) function, to control for the extremely skewed nature of the data.¹⁵ Standardized (or z) scores are then calculated for each actor for each of the three components. These z scores are added across all three capabilities components (population, troops and allies), giving a combined score for each actor. These are then subtracted to create a combined balance of capabilities score for each case. Non-state actor scores (usually the lower of the two) are subtracted from state actor scores, creating a scale where higher positive numbers represent a greater capabilities advantage for states; higher negative numbers represent greater advantage for non-state nationalist groups, and numbers near 0 represent cases of near-balance of capabilities. Because of the lack of troop data for a few non-state actors, the total N for this combined score is 45.

Strength of stereotyping-the third independent variable-was measured by coding elite statements for extremity, using the framework of Foreign Policy Images developed by Richard Cottam (1977) and Richard Herrmann (1988; Herrmann and Fischerkeller, 1995). The intent was not to determine which particular images (Enemy, Imperialist, Colony) elites were using, but how extreme those images were. Elite statements were content-analyzed for image elements, and coded for the extremity of the image being used on a three-point scale:

1. Little to No Stereotyping Evident
2. Moderate Stereotyping
3. Extreme Stereotyping

A "total stereotyping score" for each case is calculated by adding the level of stereotyping of each side, resulting in a scale from 2 to 6. Data for this coding was gathered primarily from Foreign Broadcast Information Service translations and world press sources, particularly the BBC. As with the coding of casualties, data availability limits the measurement of stereotypes to a subset of 23 of the 48 cases.
Finally, we need a measure of third party involvement. It was suggested above that the primary dimension here is how third parties involve themselves -- as mediators (seeking to broker a peaceful solution between conflicting parties) or as allies and interveners (seeking to advance the interests of one side against the other, or to support one side's position at the other's expense). This leads to a scheme of four nominal categories:

1. No Third Party Involvement
2. Third Parties Involved as Allies/Interveners
3. Third Parties Involved as Mediators
4. Third Parties Involved as Both Allies and Mediators

This coding was applied to all 48 conflict episodes; data for these judgments came primarily from news and press coverage (for more recent cases) and secondary history sources.

Methods and Results

This section is divided into two parts-methods and results. The first briefly lays out the approach and methods used to test the hypotheses on the described data. The second then lays out and briefly discusses, in a series of sections the results of these tests; these sections are grouped by independent variable (level of violence, balance of capabilities, strength of stereotyping, third party involvement).

Methods of Testing

These measurements leave us with a variety of different types of variables, requiring different testing strategies. As a first cut at examining the hypotheses above, this paper will limit itself to bivariate testing of each of the hypothesized relationships. This approach is warranted for two reasons: the differing types of independent variables (interval, ordinal, nominal), and the small N of cases for which all of the independent variables are available. Taking the bivariate approach leads to three different testing methods:

• For tests with a dichotomous DV (separation/inclusion) and an interval or ordinal IV, a difference-of-means test is used. This same method is applied (in reverse) to test a nominal IV (third party involvement) and interval DV (length).
• For tests with an interval DV (length) and interval or ordinal IVs, bivariate Pearson and Spearman coefficients (respectively) are used.¹⁶
• For the one test with a dichotomous DV (separation/inclusion) and nominal IV (third party involvement), a cross-tab is used, with Chi-square used to test the significance of the relationship.

Results: Level of Violence

• Hypothesis 1: Levels of Violence and Outcome Type

Since we have three measures of the level of violence (deaths, deaths per 1000 population, and a behavioral scale), there are three separate difference-of-means tests to examine this hypothesis. Results are as follows:

These results do seem to support the prediction of Hypothesis 1-that conflicts which end in separation are more likely to have experienced higher levels of violence than conflicts which end in non-separation outcomes. This seems to be true only for the behavioral measure of violence, however¹⁷ ; the hypothesis receives weak confirmation (one-tailed significance < .1) using the casualties per 1000 data, and none whatsoever using raw casualties alone. Moreover, if only one case (the Bosnia conflict) is removed from the data set, the borderline significance between casualties per 1000 and outcome disappears (sig. F = .772, and the ordering of group means is reversed). These results suggest that Kaufman's basic premise-that violence matters-is correct, but that what may matter most is not how much damage the two sides do but how they act towards each other. This has potentially significant consequences for how we go about studying these kinds of conflicts, and what sorts of measures we choose to adopt for future research.

• Hypothesis 2: Levels of Violence and Conflict Length

As with Hypothesis 1, we have three different bivariate tests comparing measures of level of violence to conflict length. The results of those tests are as follows:

The logic outlined above suggests that more violent conflicts ought to be expected to last longer, if violence is connected either with difficulty of settlement or with "intractability". These results, however, show no significant relationship between any of the measures of violence and length of conflict -- not even between length and total casualties, which we might have expected based on the simple logic that given more time, groups can do more damage to each other. Whatever influence level of violence (behavioral or actual) may have on the outcomes of conflict, it appears to play no role in how long it takes to reach that outcome.

These results are also an indication of the relatively wide range of levels of violence within the phenomenon of violent nationalist conflict. It is clear that some conflicts can go on for a very long time and either cause relatively few casualties (e.g. the Basques in Spain, still ongoing at the end of 1996) or a great many over time (e.g. the Eritreans in Ethiopia). This points to the need to be careful in generalizing across this category of conflict, and may suggest that some important sub-types of conflict exist within this broader category.

Results: Balance of Capabilities

• Hypothesis 3: Balance of Capabilities and Outcome Type

Since we only have one combined measure for the balance of capabilities, only one difference-of-means test is required. Results of this test are as follows:

Here we see weak support (one-tailed sig. < .1), in the expected direction, for the logic of Hypothesis 3. Conflicts that end in separation are characterized by a more even balance of capabilities (numbers closer to 0) than those that end in non-separation situations. There is not an especially strong relationship here, however, suggesting that while capabilities may matter, they are clearly not a decisive factor-that states which have an advantage may still "lose" their separatist populations. This is consonant with observations of individual nationalist conflicts; despite being out manned, outgunned and outnumbered, many nationalist movements have managed to fight on for very long periods of time without being reincorporated into their host states, and some (Chechnya, for example) have even won separation against a numerically superior state.¹⁹

• Hypothesis 4: Balance of Capabilities and Conflict Length

This relationship is tested with a single bivariate correlation, with results as follows:

The result indicated here shows a relationship opposite the one hypothesized. It suggests (at a very weak level of significance) that conflicts with a greater imbalance of capabilities (in the state's favor) last longer than those more evenly matched. This finding-which should most fairly be treated as a non-finding, indicating support for neither direction-suggests again that traditional measures of capability may not be relevant to the ways in which these kinds of nationalist conflict are conducted. It also suggests that expectations about a particular nationalist group's chances to hold out against a state which are based on such traditional factors are not likely to be accurate; we clearly need a deeper understanding of the factors that enable some conflicts to last longer than others, outside common conceptions of the bases of power.

• Hypothesis 5: Balance of Capabilities and Level of Violence

This hypothesis is tested by correlating the balance of capabilities score with the three levels of violence measures. Results are as follows:

These data show only weak support for the logic of this hypothesis-that greater imbalances of power should reduce the level of violence, as the more advantaged side can win more quickly. This is not supported at all for the raw casualties data; balance of capabilities apparently has little to do with how many people overall are killed. Likewise, it does not seem to affect the kind of fighting that happens; an overwhelmingly preponderant state is no less likely to keep nationalist groups from escalating to large-scale guerrilla activity or protracted civil war than one more evenly matched with its opponents. This suggests, for those studying the beginnings and process of these kinds of conflicts, that calculations of capability-at least as measured in traditional fashions-may not play very much of a role (Ayres and Saideman, 1999).

These results do support the hypothesis for proportional casualties below the .1 level (one-tailed). This is important because the proportional casualties measure (of the three measures of level of violence) best represents the logic of the hypothesis. There are clearly many factors which can contribute to overall levels of casualties (among them population sizes, weapons available, length of conflict, and so on); likewise, the behavioral level of violence is in large part a result of strategic calculations made by both sides (particularly the nationalist group seeking to secede), which may in turn be swayed by a host of political factors besides raw capabilities. But we should expect that if the capabilities of the two sides affect the amount of violence committed, they might well do so in terms of proportional casualties-how much damage is done given the size of the groups fighting. The hypothesis as stated suggested that when there is a power imbalance (invariably in favor of the host state), damage should be less because the state would use its might to quell the conflict. This result supports this, although the results of these tests and those under Hypothesis 4 rule out two mechanisms by which this could occur: length of conflict (which is not correlated to the balance of capabilities; see above) and intensity of fighting (as measured by the behavioral scale of violence, listed here). This suggests that there is some other means by which states can use a capabilities advantage to limit damage-perhaps by protecting their own populations and troops against armed nationalists. How exactly this works is left for future research.

Results: Strength of Stereotyping

• Hypothesis 6: Strength of Stereotyping and Outcome Type

This hypothesis represents the logic that the important determinant of how conflicts end is not the "objective reality" of how much damage the parties take or what tools of control they possess, but how they perceive each other. Results of a difference-of-means test between outcome type and strength of stereotyping are as follows:

While the observed significance level here might suggest very weak support for Hypothesis 3, the direction of the means difference is opposite that of the predicted relationship. To the extent that we could attribute borderline significance to this finding (a tentative prospect at best, given the very small N involved), it would suggest that separation outcomes are actually associated with lower levels of stereotyping than non-separation outcomes. However, even this relationship disappears when examined with other methods (a standard cross-tab Chi-square produces a significance above .3, while a two-by-two difference-of-medians table is scarcely distinguishable from chance). The best conclusion that can be drawn is that, for this subset of cases, the strength of stereotyping does not seem to matter much in what kind of outcome a conflict reaches.

• Hypothesis 7: Strength of Stereotyping and Conflict Length

As with the testing of Hypothesis 4 (balance of capabilities and conflict length), this hypothesis is tested with a single correlation, with results as follows:

There is no support here for a significant relationship between the strength of stereotyping and length of conflict. This may seem counter-intuitive to those who insist that perception is an important part of conflict resolution; however, we also know that perceptions and stereotypes can change (Ayres, 1997), sometimes rapidly. This in turn suggests the need for further research into the antecedents of such change, and cautions us not to assume that it will take longer to change extreme stereotypes than it does moderate ones.

• Hypothesis 8: Strength of Stereotyping and Level of Violence

Finally, to test potential effects of the strength of stereotyping on the level of violence, we generate a series of three correlations, one for each of the violence measures:

These results also show very little impact from the stereotyping variable. Strength of stereotyping has no significant correlation with proportional deaths or intensity of fighting, and has only borderline significance with total casualties (as the coefficient runs in the expected direction, one-tailed sig. < .1). This may be a function of the relatively small number of cases for which both stereotyping and casualties data are available, or it may be that it is important to understand the kind and content of stereotypes as well as their intensity (something not measured here).

Another argument is that intergroup perceptions are epiphenomenal-that they are caused by high levels of violence, not the other way around. Kaufman (1996) suggests this logic, as do many popular commentaries on ethnic conflict which use tales of horrific killings to explain to outsiders how deep-seeded hatreds are created. These data do not support this logic either; if violence produces extremity of perception, we should expect proportional deaths to do so at least as strongly as raw death totals (since proportionality helps define the "impact" of violence on a given society). There is clearly a need for more data here, but these initial findings suggest that the various hypothesized links between violence and perception may not work the way we think they do.

Results: Third Party Involvement

• Hypothesis 9: Third Party Involvement and Outcome Type

In testing this relationship, we cannot use the difference-of-means (or difference-of-medians) approach, since the 4 categories of third party involvement have no ordinality. The best that can be created, therefore, is a cross-tab table, which appears as follows:

The relationship between these two is significant (Chi-square significance = .039), and the distribution in the table suggests that there is a fairly clear pattern: where mediators get involved (either with or without intervener involvement), separation outcomes are more likely than when they are not. This is a finding that contradicts the logic set out under Hypothesis 9 above; efforts by mediators seem to make separation, not inclusion, more likely. Also of interest, the contribution of allies (third parties seeking to intervene on one side or the other of a conflict) does not make non-separation outcomes any more likely than if no third party involvement takes place at all. These findings suggest that what third parties do can matter-particularly if they are trying to make peace-but that their mere presence and general intention may not be enough to explain their effectiveness or their impact on conflict outcome.

• Hypothesis 10: Third Party Involvement and Conflict Length

Because the Third Party Involvement measure is nominal, not ordinal, this hypotheses requires a return to the difference-of-means test. Mean length of conflicts were compared across the categories of third party involvement; results are as follows:

These results show that third party involvement does have a significant relationship with length of conflict. Unlike the results for Hypothesis 9, these results confirm the logic of third party intervention suggested in the hypothesis section above: mediators tend to be associated with the shortest conflicts, interveners with the longest, and the mixed and no-involvement conditions are somewhere in between. This supports the policy contention that impartial mediation can help bring conflicts to a close sooner than they otherwise would, and also suggests that intervention by outside powers may only lengthen and exacerbate the problem.²⁴

Hypothesis 11: Third Party Involvement and Level of Violence

Because of the nominal nature of the third party involvement measure, to test this hypothesis we must revert to a difference-of-means test procedure. As there are three measures of the level of violence, this produces three sets of results:

These data provide fairly strong support for a relationship between third party involvement and level of violence, in the manner predicted by the hypothesis. Involvement by mediators alone is consistently correlated with the lower (if not lowest) levels of violence across all three measures (two of which are significant), while the involvement of interveners is consistently related to higher levels of violence. Curiously, the combination of both mediators and allies correlates with the highest levels of violence; conflicts in this group have the highest means scores across all three measures. This suggests that what third parties do does matter, and provides some support for the notion that mediation by itself can be a way of ameliorating the effects of conflict.

There is an interesting difference among these results: in the behavioral violence intensity data, those conflicts which have no third party involvement at all are on average less violent than those where mediators get involved-a comparison which is reversed in the proportional casualties results. This may be attributable to a problem of endogeneity: what third parties do can be expected to affect the level of violence in a conflict, but it is also possible that third party involvement may be affected by how violent a conflict is. While these results cannot prove directionality one way or another, they are suggestive of a logic worthy of further study, as follows. Third parties may use intensity of fighting (which is knowable at any point in the midst of a conflict) as a cue to involvement. Mediators and interveners alike may be more likely to involve themselves in more intense conflicts-the former because they fear the consequences of fighting gone rampant (including the possibilities of diffusion and contagion to neighboring states), the latter because they may have more cause to fear that whichever side they support is losing. Hence, conflicts with no third party involvement are likely to be the ones with the lowest intensity levels. On the other hand, casualties (total or proportional) are not knowable until the end of a conflict, and thus (as measured here) do not serve a useful cueing function to potential third parties; however, once those third parties get involved, we should expect them to impact the course of the conflict in precisely the ways outlined in the proportional casualties data above-mediation ameliorates conflict, while allied intervention exacerbates it. These results could therefore be construed as supporting relationships in both directions, suggesting that intensity is a cueing mechanism and casualties an outcome. Further exploration of this logic is left for future research.

Conclusions

These results lead to two different sets of conclusions: what they say about the theories and variables examined, and what they suggest for future research. Because of the relatively small N and lack of multivariate testing, we must consider these conclusions provisional at best, though they do shed some light for use in future efforts.

Results for the various tests involving level of violence as an independent variable do not show very much support for the notion that violence, in and of itself, determines what kinds of conflict outcomes you get. The only significant relationship discovered was between behavioral intensity of violence and separation/non-separation outcomes; here, greater levels of fighting are correlated with endings in which the parties are separated. The broader argument that violence is what fundamentally matters, however, does not seem to be well supported.

If level of violence received little support as an independent variable, balance of capabilities received almost none. Knowing the relative sizes, troop strengths, and strengths of allies of the conflicting parties does not help to predict how long a conflict will last, or whether it is likely to end in separation of the two sides or not. There was some evidence that the balance of capabilities has some impact on proportional casualties, but otherwise seems to have no relationship with the outcomes of conflict. This suggests that theories of conflict that emphasize capabilities, as many studies of interstate war do (e.g. Singer and Small 1982), may not be as applicable to this particular class of nationalist, intrastate conflicts.

Approaches that emphasize the importance of intergroup perceptions and stereotypes likewise received little support in these data. Intensity of stereotyping is correlated in these cases neither with outcome type, nor with length of conflict, nor with level of violence. While it may be that perceptions and stereotypes do play a role, it is obviously a more complex one than is captured by the (admittedly simple) measure used here; what effects there are may lie more in the content of perceptions, or their change over time, than in sheer intensity. It must be emphasized that these results are particularly tentative, given the small N (23) involved.

Finally, these results do show that the involvement of third parties plays a significant role in the conduct and outcome of nationalist intrastate conflicts. Involvement by third parties attempting to mediate a resolution makes separation more likely, and is associated with shorter conflicts; mediator-only conditions are also associated with lower levels of casualties. The influence of third parties allied with one of the primary actors, on the other hand, consistently exacerbates conflict, leading to longer, deadlier conflicts. This provides support for both scholars and policy makers who argue that what third parties do matters; that mediation or other attempts to arbitrate a settlement provide beneficial effects; and that biased intervention by outside powers generally only makes things worse.

These results also leave open a number of important questions and suggest further paths of inquiry. The first and obvious next step is to incorporate these variables into a multivariate model which measures the effects of each while controlling for the others. This is made difficult by the variety of types of variables included in this study, particularly the nominal-category third party involvement variable. Based on the results above, there are clearly important potential interactions among IVs; it also appears that there may be some ordinality to the way in which third parties impact conflict outcomes. If this relationship is explored further, it may be possible to construct an ordinal-type third party variable which would allow for inclusion in a multivariate model.²⁶

The strength of the third party results also suggests the importance of further study of this relationship on its own merits. If mediators truly are able to ameliorate conflicts and end them with fewer deaths, we would like very much to know why. Another logical next step, therefore, would be to categorize third party mediation strategies and correlate those with various types of outcomes, to examine the impact of what third parties actually do rather than their mere presence. Where the number of cases is too small to use statistical tests (only seven of the cases in this data set involve purely mediator involvement), rigorous comparative case studies may prove useful as well.
Finally, there are a host of timing issues which, because this data set does not include time-series data, are left open to further study. Does it matter when third parties become involved? Would perception play a role if we could look at points of change? It is likely that violent behavior and casualties are not spread evenly over the course of many conflicts; do "clusters" of violence matter, and if so, how? To truly understand the process of ethnic conflicts, and how they end, we need time-series data on these and a variety of other dimensions-a very large undertaking, but a necessary one for answering the important questions about ethnic conflict.

Acknowledgments

The author would like to thank Peter Radcliffe, Donald Sylvan, Andrew Enterline, Volker Franke and Richard Herrmann for their help and advice, and Roy Licklider and David Mason for their suggestions and assistance in locating data. All errors are the sole responsibility of the author.

Appendix A

Notes

¹ Kaufman explores some of his claims using a set of 27 cases from Ted Gurr's Minorities at Risk data set; this study seeks to use a larger and somewhat different set of cases to explore the same logic.

² These last hypotheses are offered to guide this as well as future research; because of variable type and data availability problems, multivariate analysis, which might pull out these same relationships, is not available for the tests being done here. See the methods section, below, for a more complete discussion.

³ One might ask what we should expect if the nationalist group is substantially stronger than its state enemy. In practical terms, this is a moot issue; as we shall see in the measurement section below, in ethnic conflicts between nationalist secessionist groups and states seeking to prevent secession, the real-world range of capabilities runs from relatively balanced to unbalanced in favor of the state. This will be discussed further below in the measurement and conclusions sections.

⁴ An alternative argument would suggest that, since interveners alter the balance of capabilities between sides, they make inclusion more likely if they line up against the secessionists and in favor of the state. This possibility will be taken into account in the measurement of capabilities, and is thus at least partially accounted for in the testing of Hypothesis 2.

⁵ Much of the official rhetoric about the Kosovo problem of 1999, and potential attempts to mediate a settlement, made this argument explicitly. (See for example Erlanger 1999.)

⁶ This last is an argument I heard most forcefully expressed by J. David Singer at the Annual Meeting of the Peace Science Society, International, October 16-18, 1998 in Brunswick, NJ.

⁷ This definition is similar, but not identical, to Gurr's definition of "ethnonationalist" conflicts (Gurr 1993); Gurr's data set provided the cases for some of Kaufman's own testing. The definition proposed here allows for a more expansive list both in time (up through 1996) and in levels of violence than previous work.

⁸ This latter criterion excludes talks on implementation. Hence, for instance, the 1991 agreement between Morocco and the Saharawis marks the "end" of that conflict, because after its signing, both sides ceased both fighting and arguing over how the conflict should be settled (having agreed on a referendum as the principle of solution).

⁹ The time period for these behaviors-how long do parties have to refrain from fighting or talking before a conflict episode is "over"--is set at a 12-month interval. Hence, parties must refrain from either fighting or discussing the conflict for a consecutive 12 month period to be coded as a separation between two individual case episodes; lapses of activity lasting for shorter periods (three-month cease-fires which end, for instance) are considered as part of one episode.

¹⁰ A list of conflict episodes, along with starting and ending dates and selected variables, is presented in Appendix A.

¹¹ Note that this criterion does not include the inability of one side to assault the other through massive conventional force ... that is, to make war. This would be an almost impossible feat, as war is nearly always possible given sufficient time to build an army. The point here is rather whether violence on a more interpersonal level is possible-whether, absent an ongoing conventional war, members of one group in their normal lives are subject to attack by members of the other. This coding is intended to adhere to Kaufman's understanding of the necessity of partition (see Kaufman, 1996).

¹² In some instances, casualty estimates were only available for a group of cases-for example, the Burmese civil war from 1948 on (within which there are a number of ethnic groups fighting-Mons, Shans, and Karens, among others) In these cases, the total estimate figure was divided proportionally among groups according to their relative troop strength, on the assumption that groups with larger militaries would a) do more damage, and b) be subject to more intense attacks by the opposing government, since they represented a bigger threat. Where further division was needed into multiple cases over time, it was assumed that deaths per month would remain constant over the multiple cases. These calculations were performed for the first 2 Iraqi Kurd cases, and for the Burmese civil war.

¹³ For states which fight multiple national movements simultaneously, the state's population is calculated for each dyad as the total population of the state minus the populations of all minority nations involved, for the same reason-that these are populations which, because they are in conflict with the state, cannot be mobilized by it, even against other nations besides themselves.

¹⁴ Population ratio is not always a very good predictor of ability to control outcomes; blacks in South Africa, despite having a substantial advantage in numbers, were unable to force an end to apartheid for many years. This points to the need for other sources of measurement.

¹⁵ When grouped together across actors, both the population and troops figures are dominated mathematically by the few cases involving China and the USSR/Russia, which in turn distorts standardized scores created on the raw data alone and eliminates most of the variance at the lower end of both distributions, particularly among non-state nationalist groups. In substantive terms, the process of taking standardized scores on the raw, untransformed data means that nationalist groups with military forces as few as 500 look nearly identical to groups or states with 10,000 troops. Using the logarithmic transformation compresses this distortion and preserves the variance among the vast bulk of cases in the low-to-middle end of both scales.

¹⁶ Separate checks with difference-of-medians tests will also be used for ordinal IVs; as these provide essentially the same results, the difference-of-means test results are presented for consistency.

¹⁷ A separate difference-of-medians test, dividing the ordinal behavior scale into 2 groups above and below the median, produced similar results, with a Chi-square significance = .065.

¹⁸ Coefficients marked with a * are Pearson correlation coefficients; those with † are Spearman correlation coefficients.

¹⁹ Another potential conclusion is that the factors used to measure capability here-men under arms, population, and allies-are not the decisive ones in nationalist struggles. Additional factors, such as a measure of commitment, or ability to mobilize existing resources, may be helpful in clarifying the relationships between power and outcome; these are left for future research.

²⁰ Coefficients marked with a * are Pearson correlation coefficients.

²¹ Coefficients marked with a † are Spearman correlation coefficients.

²² The very low N here is the product of cross-tabulating two variables with missing data in this dataset: stereotyping (N=23) and deaths (N=30).

²³ Length (and therefore mean length) is measured in months.

²⁴ As will be shown below, interveners are also associated with higher levels of violence and greater proportional casualties.

²⁵ Mean of 8-point (0-7) scale of violence. A difference-of-medians test was also performed on these same data, which confirmed a relationship significant below the .05 level with the same ordering of third-party categories.

²⁶ Another possibility would be to break out the different third party intervention categories as dichotomous dummy variables. This approach would overlook any actual ordinality within the concept, but might be a useful next step when more data are available.

References

Ayres, R. William. 2000. "A World Flying Apart- Violent Nationalist Conflicts and the End of the Cold War." Journal of Peace Research, Vol. 37, No. 1, forthcoming.

Ayres, R. William and Steve Saideman. 1999. "Reuniting, When Does It Feel So Good- The Causes of Irredentist Movements." Paper presented at the Annual Meeting of the International Studies Association, February 16-21, Washington, DC.

Ayres, R. William. 1997. "Mediating International Conflicts: Is Image Change Necessary?" Journal of Peace Research, Vol. 34, No. 4, November, pp. 431-447.

van Bruinessen, Martin. 1994. "Genocide in Kurdistan?" In George Andreopoulos, ed., Genocide: Conceptual and Historical Dimensions. Philadelphia: University of Pennsylvania Press.

Cottam, Richard. 1984. "Nationalism in the Middle East: A Behavioral Approach." In Said Amir Arjomand, ed., From Nationalism to Revolutionary Islam. Albany: SUNY Press.

Cottam, Richard, 1977. Foreign Policy Motivation: A General Theory and a Case Study. Pittsburgh: University of Pittsburgh Press.

Crocker, Chester. 1996. "The Varieties of Intervention: Conditions for Success." In Chester Crocker & Fen Osler Hampson, eds.,

Managing Global Chaos: Sources of and Responses to International Conflict. Washington, DC: United States Institute of Peace Press.

Erlanger, Steven. 1999. "US Presses Kosovo Rebels as Violence Claims More Civilians," New York Times, January 26, p.A4.

Gurr, Ted Robert. 1999. "Minorities at Risk" Dataset, www.bsos.umd.edu:80/cidcm/mar/.

Gurr, Ted Robert. 1994. "Peoples Against States: Ethnopolitical Conflict and the Changing World System." International Studies Quarterly, Vol. 38, pp. 347-377.

Gurr, Ted Robert. 1993. Minorities at Risk: A Global View of Ethnopolitical Conflicts. Washington, DC: United States Institute for Peace Press.

Haas, Ernst. 1986. "What is Nationalism and Why Should We Study It?" International Organization, Vol. 40, No. 3, pp. 707-744.

Heraclides, Alexis. 1997. "The Ending of Unending Conflicts: Separatist Wars." Millennium, Vol. 26, No. 3, pp. 679-707.

Herrmann, Richard, 1988. "The Empirical Challenge of the Cognitive Revolution." International Studies Quarterly, Vol. 32, pp. 175-203.

Herrmann, Richard and Michael Fischerkeller, 1995. "Beyond the Enemy Image and Spiral Model: Cognitive-Strategic Research After the Cold War." International Organization, Vol. 39, No. 3, pp. 415-450.

Kaufman, Chaim. 1996. "Possible and Impossible Solutions to Ethnic Civil Wars." International Security, Vol. 20, No. 4, pp. 136-175.

Kelman, Herbert. 1987. "The Political Psychology of the Israeli-Palestinian Conflict." Political Psychology, Vol. 8, No. 3, pp. 347-364.

Licklider, Roy. 1998. "Civil War Termination 2.1" Dataset, www.rci.rutgers.edu/~licklide.

Licklider, Roy. 1995. "The Consequences of Negotiated Settlements in Civil Wars, 1945-1993." American Political Science Review, Vol. 89, No. 3, pp. 681-690.

Mason, T. David and Patrick Fett. 1996. "How Civil Wars End: A Rational Choice Approach." Journal of Conflict Resolution, Vol. 40, No. 4, pp. 546-568.

Posen, Barry. 1993. "The Security Dilemma and Ethnic Conflict." Survival, Vol. 35, pp. 27-47.

Regan, Patrick. 1996. "Conditions of Successful Third-Party Intervention in Intrastate Conflicts." Journal of Conflict Resolution, Vol. 40, No. 2, pp. 336-359.

Singer, J. David and Melvin Small. 1982. Resort to Arms: International and Civil Wars, 1816-1980. London: Sage Publications.

Stein, Janice Gross. 1996. "Image, Identity, and Conflict Resolution." In Chester Crocker & Fen Osler Hampson, eds., Managing Global Chaos: Sources of and Responses to International Conflict. Washington, DC: United States Institute of Peace Press.

Wagner, Robert Harrison. 1993. "The Causes of Peace." In Roy Licklider, ed., Stopping the Killing. New York: New York University Press.

Walter, Barbara. 1997. "The Critical Barrier to Civil War Settlement."International Organization, Vol. 51, No. 3, pp. 335-64.
Wire Service Stories, LEXIS-NEXIS Search; includes Reuters, Agence-France Presse, BBC.

Zartman, I. William & Saadia Touval. 1996. "International Mediation in the Post-Cold War Era." In Chester Crocker & Fen Osler Hampson, eds., Managing Global Chaos: Sources of and Responses to International Conflict. Washington, DC: United States Institute of Peace Press.