ETHNOLINGUISTIC FRACTIONALIZATION AND OTHER MEASURES OF ETHNIC DEMOGRAPHY

We agree that it is natural to wonder whether political dominance of an ethnic minority raises civil war risks. But we also believe that it is interesting and important to ask, as we did in Fearon and Laitin (2003a), whether civil war is more likely in more ethnically diverse countries. It is important because it is commonly supposed to be so in the media and by politicians, and because much work by political scientists and sociologists on ethnic politics has presumed or argued that “plural societies” are more prone to intense internal conflict (e.g., Horowitz 1985; Rabushka and Shepsle 1972). Gellner (1983) himself observes that the nationalist principle is violated when there are multiple “potential nations” within state boundaries. He says that multiple groups imply that “a territorial political unit can only become ethnically homogeneous, in such cases, if it either kills, or expels, or assimilates all nonnationals. Their unwillingness to suffer such fates may make the peaceful implementation of the nationalist principle difficult” (2). This logic suggests that the more ethnic groups, the more unsatisfied and conflicting nationalist sentiments, and the greater the risk of internal conflict.

An ethnolinguistic fractionalization (ELF) index is just one way of measuring ethnic diversity. It has advantages over, say, counting groups or using the size of the largest group. ELF increases with the number of groups holding the distribution of sizes constant, and it increases as one changes the distribution toward greater equality among the groups. Both are desirable properties in a measure of diversity. If one simply counts the number of groups, then a country with group shares (.97, .01, .01, .01) looks as diverse as a country with (.25, .25, .25, .25), which seems wrong relative to intuitions about what ethnic diversity means. (The ELF scores are .06 and .75, respectively.)

But there are other defensible measures of ethnic or religious diversity at the country level, including the size of the largest group, a count of the number of groups above some threshold, or the “effective number of groups” (1/(1-ELF)). One can also make arguments about specific size distributions, such as the idea that a country with a majority ethnic group and a large minority group may be particularly prone to violence (e.g., Horowitz 1985, 36–41). We considered these several measures and others for the analysis in Fearon and Laitin (2003a; see also Fearon and Laitin 2003b). In addition we used nonparametric methods in case the linearity assumptions of the regression model might be obscuring some less linear pattern. We did not view ELF as representing any particular argument about why ethnic diversity would cause greater civil war risk—much less the specific “individualist” argument proposed by CG—but rather as a measure helpful for empirically assessing any argument that proposes a positive relationship between ethnic diversity and civil war risk, “individualist” or not.

In their conclusion, CG assert that by using ELF, we “tend to assume that violence is primarily a reflection of individual, as opposed to group-level dynamics” (2007, 182). It seems a bit unfair for CG to make up an “individualist” story to rationalize ELF and then attribute it to us, simply because we use the measure! In any event, if a “group-level” rationale for ELF is wanted, it is not difficult to construct one. Suppose that major conflicts occur according to a Poisson process, and that the rate at which they emerge between any two distinct groups i and jis proportional to p_ip_j, where these are their respective population proportions. (Thus, violent conflict is assumed more likely the larger the two groups and the more equal they are.) Then the total rate is proportional to

ELF, since ELF

. No assumption about individual-level interactions driving conflict is needed here; the groups may be conceived as unitary actors, as in CG's arguments for N*.

ETHNIC GROUPS AND CONTROL OF THE STATE

The main obstacle to answering the question posed at the outset is how to code “control” of the state by different ethnic groups across a large number of countries. CG say that they “consider a group, or a coalition of groups, to be in power if their leaders serve (at least intermittently) in senior governmental positions, especially within the cabinet” (2007, 178). It is not clear whether they examined ethnicities of cabinet members, however, or indeed just what the operational procedure was. Their paragraph on coding says that “To determine if a group was an EGIP, we relied on Heger and Salehyan's (2005) dataset of leader's ethnic affiliations, based on information from a leader data [set] collected by Goemans, Chiozza, Gleditsch, and Choung (2005)” (2007, 178), taking additional data from the Minorities at Risk project and the CIA World Fact book.

Whatever their exact procedures and rules, in practice CG's choice for the “Ethnic Group in Power” (EGIP) is the plurality ethnic group in a large majority of the Eurasia and North Africa countries in their sample—74 out of 89, or 83%. Only four countries are coded as having a minority EGIP (Iraq, Jordan, United Arab Emirates, and Lebanon). The biggest departure of EGIP from the plurality group share comes from CG's decision to code 11 countries as having a coalition of EGIPs, which they treat as equivalent to a unified ethnic group with population share equal to the sum of the component groups' shares (Belarus, Belgium, Czechoslovakia, Finland, Pakistan, Spain, Switzerland, Syria, United Kingdom, Ukraine, and Yugoslavia). For N*, this decision makes such cases equivalent to nearly homogenous countries in which the plurality group controls the state. N* assigns these cases essentially zero probability of conflict.

This coding procedure folds together two issues that might better be kept separate. First, are countries with a minority EGIP more civil war prone? This is the question most directly implied by Gellner's and CG's theoretical arguments. Unfortunately, because there are only four such cases in their sample (by CG's coding), there is probably not enough evidence here for establishing any robust empirical tendency or pattern.

Second, are multiethnic countries that succeed in sharing state power among groups at lower risk of civil war? It would not be too surprising if the answer were “yes,” because this is not far from asking “is civil war less likely in a country in which the major ethnic groups have demonstrated an ability to get along?”²

Using CG's codings, civil wars broke out in 2.60% (5/192) of the country years with a minority EGIP, versus 1.80% (51/2837) for plurality EGIPs and .99% (5/503) for coalition EGIP cases. Although this pattern is consistent with the hypotheses that minority rule raises civil war risk, and that countries with some measure of ethnic power-sharing are less civil war prone, there is not much evidence here. A chi-square test fails to reject the null hypothesis of no association (p=.28). Again, there are only four minority EGIP countries in CG's sample, two of which had two civil wars (Lebanon and Iraq), and one (Jordan) which had one, using Fearon and Laitin's (2003a) civil war codings.

Table 1 reexamines CG's logit analysis of the country-year data for their Eurasia and North African subsample. Model 1 replicates their Table 3 (Model 2), which finds a positive effect of their N* measure in the Eurasia/North Africa sample.³

Model 4 drops both minority and coalition EGIP countries, allowing us to distinguish between the effect of N* that comes from it being (essentially) a nonlinear transformation of the plurality group share and the effect of N* coming from the fact that in some countries the plurality group does not control the state. Note that N*'s estimated coefficient is now essentially zero (adding ELF sends the estimate for N* to −1.36). So there appears to be little to choose from between N* and ELF when we consider only countries with plurality EGIPs.

Model 5 drops N* and instead uses dummies for whether the country had a minority EGIP and whether the country had a coalition EGIP as coded by CG (the omitted category is countries with a plurality EGIP). It shows a positive effect on civil war risk for the four minority EGIPs, and a negative but statistically insignificant effect for the 11 coalition EGIPs.

Overall, it is clear that the “punch” of N* comes entirely from its marking off the four minority EGIP countries, three of which are coded as having had civil wars in this period (the bivariate correlation between N* and a dummy marking these four countries is quite high, at .89). It is a bit disconcerting that including either Iraq or Lebanon, while omitting the other three minority EGIP countries, suffices to produce “statistical significance” for N* in this sample. Worse, Model 6 shows that when we recode Syria as having a minority EGIP, which seems at least as plausible as CG's coding,⁴

In our view this evidence provides no support for the N* formulation per se, and weak support for the notion that states dominated by ethnic minorities have higher civil war risk. The evidence is weak because it comes from the experience of just four countries, and Lebanon and Iraq in particular. Prior to its political disintegration, Lebanon could have been coded as a country with a coalition EGIP quite easily. Iraq has two civil wars coded in Fearon and Laitin's (2003a) data, neither of which involved the majority Shias rebelling against the dominant minority Sunni regime (although arguably there should be such a civil war coded after the first Gulf War). The Kurds rebelled under Sunni minority rule in Iraq, but they are a minority as well. Further, Kurds have rebelled in neighboring Turkey and Iran where plurality ethnic groups control government. They have not rebelled in Syria where there is, arguably, a minority EGIP. The 1971 war in Jordan might be linked with Palestinian unhappiness with the Hashemite regime's dominance, although regional political struggles were also very important. So the evidence is suggestive but certainly not a robust generalization.

LEADER ETHNICITY AND CIVIL WAR RISK IN A LARGER SET OF COUNTRIES

Most political scientists would probably say that an ethnic group “controls” or “dominates” a state or government to the extent that its members monopolize major government offices and also determine significant policies. In principle one could try to operationalize this concept by coding the ethnic composition of cabinets or all state offices for a large sample of countries. In practice this would be a difficult enterprise, especially for highly diverse subSaharan Africa. Moreover, because the real power of different ministries and other offices can vary enormously across countries and time, one would face difficult aggregation problems even if the data were available.

A compelling alternative is to use the ethnicity of the head of state as an indicator of ethnic influence or control. It has some obvious problems, but these can be managed and the compensating advantages are great.

The main problem is illustrated by the case of Canada, which has had three French Quebecois as prime ministers since 1945. To say that the French dominated or controlled the Canadian state in these periods seems wrong. The same issue arises in a number of other cases, mainly in Europe: for example, a Scottish PM in Britain (Douglas-Home, 1963), the Galician Franco in Spain, the Georgian Stalin in the USSR, Slovak General Secretaries in Czechoslovakia, and rotating presidencies in Switzerland and post-Tito communist Yugoslavia. However, if we are trying to tap the idea of “state control” that CG and other analysts of ethnic politics seem to have in mind, we can drop such countries from the sample, or recode them on an ad hoc basis as “majority dominated.”

There are several significant advantages to using leader ethnicity as an indicator. First, we have been able to code it for more than 160 countries for the period 1945 to 1999. Second, problematic cases of the type just mentioned appear to be unusual. In general, leader ethnicity does an excellent job of picking out the group locally regarded as politically most powerful. In sub-Saharan Africa, for example, it is often thought that the president's ethnic group is the most favored and politically dominant, even in cases where many government bureaucrats come from a different group. Comparing this indicator with CG's EGIP codings for the countries in their Eurasia/North Africa sample, we find very few differences. Of the 835 leaders in the countries of CG's sample, less than 5% come from an ethnic group different from a group CG identified as “in power” in that country. This suggests that the indicator is worth considering for the larger sample beyond these regions.

A third important advantage of using leader ethnicity is that the measure varies over time within a good number of countries, allowing us to examine whether civil wars tend to begin with transitions to minority leaders. By contrast, CG's codings assume that the “ethnic group in power” never changes, which is implausible for a number of their countries and would be an untenable assumption for sub-Saharan Africa. A fourth advantage is that although the real significance of different ministries varies across countries and over time, the question of who gets to be the head of state is always critical.⁵

Using Goemans et al.'s (2005) list of heads of state and the ethnic categories from Fearon (2003), we coded leader ethnicity for 161 countries for the period from 1945 to 1999. Table 2 provides summary statistics on the prevalence of leaders from minority ethnic groups. Worldwide, between 20% and 30% of heads of state have been from an ethnic group other than the plurality group for the period from 1945 to 1999, depending on whether we treat “white” and “mestizo” as the same or different groups in a number of Latin American countries.⁶

Table 2 also compares rates of civil war outbreak by whether the leader was from the plurality group or an ethnic minority. Overall, civil war has broken out at a slightly higher rate when the leader came from an ethnic minority (2.05% or 2.27% vs. 1.5% per year, depending on how we code the Latin America cases). This is also true within regions except for Western and Eastern Europe and the former Soviet Union. However, using either a chi-squared or a simple t test on differences of means, the raw difference for the whole sample is not statistically significant under the “white different from mestizo” coding, or just barely so under the “white same as mestizo” coding (at p=.05).

Because ethnic minority leaders are clearly more common in poorer countries—as seen in the higher proportions in sub-Saharan Africa—bivariate comparisons may be misleading. Table 3 displays the results of a multivariate logit analysis using the core specification and data from Fearon and Laitin 2003a, but adding a dummy variable marking country years in which the leader came from a group other than the plurality group. The first model uses the coding with “white different from mestizo” for Latin America; the second uses “white same as mestizo.” For the third model, in addition to using “white same as mestizo,” we recode the countries Canada, Belgium, Spain, Switzerland, Czechoslovakia, Yugoslavia, and the USSR so that they have no minority leaders, in order to see if some implausible attributions of minority dominance in these cases are affecting the results.⁷

Having a leader from an ethnic minority is positively associated with civil war onset in Models 1–3, but the estimates are small in substantive terms in addition to being quite uncertain. In Models 1 and 2 they translate to about 34% greater annual odds of civil war outbreak in minority-led country years, which is a bit more than half the impact of anocracy, the next most “important” dichotomous variable in the model. In Model 3, with the most favorable coding for minority leader, odds of civil war outbreak are estimated to be 44% greater per year under minority leaders, just barely significant at the 10% level. The estimates are also fragile in the sense that adding region dummies reduces their size and significance considerably (while leaving the other variables relatively unaffected).

Perhaps the simple indicator for whether the leader is from a minority group is too crude a measure of plurality group nationalist resentment. We considered alternative measures, such as the difference between plurality and leader group sizes, the raw size of the leader's group, and N* calculated using our leader ethnicity data for the EGIP. The results are always worse for these variables than for the dummy variable indicating minority leader.

For countries that had a civil war and controlling for unmeasured country-specific factors (Table 3, Model 4), the estimated coefficient for ethnic minority rule is much larger. Substantively, it implies that the odds of civil war onset were about 2.75 times greater each year under the minority leader, other things equal. The estimate is again fairly uncertain, and the results are rather unstable (depending on specification). This is typical of fixed-effects logit models for rare events; there simply isn't much within-country variation in the dependent variable available to identify effects. In the case at hand, we find no association at all for ethnic minority rule within sub-Saharan Africa or Latin America using the fixed-effects approach, and a strong association in Eurasia and North Africa that depends heavily on a small number of cases.⁸

What about “ethnic” civil wars? The pattern proves to be similar: hardly any association between minority rule and ethnic civil war within Africa or Latin America, and a somewhat stronger association outside of these regions that depends on a small number of cases (for the regression results and discussion, see Fearon, Kasara, and Laitin 2006).