One Ukraine or many? regionalism in Ukraine and its political consequences

Click to increase image sizeClick to decrease image size

Notes

He would probably add to this list a third point that he makes repeatedly in the paper: that the strength of the regional effect varies by the nature of the issue in question. We would concede this point to an extent; after all, the Southern culture in the United States will be reflected more strongly in certain issues than on others. But we also demonstrate strong regional effects across different kinds of social and political attitudes and political behaviors.

In one‐way analysis of variance (ANOVA) estimations, in‐group variation is often larger than between‐group variation; this fact alone does not invalidate the estimation of between‐group differences (such as between regions), although it would tend to indicate a fair amount of noise in the model. Our thanks to John McAdams for bringing this point to our attention.

Katchanovski (2001) employs the east Ukraine versus west Ukraine approach, although he includes more oblasts in the east than we do in our two‐region models below. In a study of voting behavior, Roper and Fesnic (2003) also present a two‐region approach, concentrating their analysis on differences between Galicia and the rest of the country. See also Zimmerman (1998) for a three‐region approach.

The best discussion of these various features across Ukraine can be found in Birch (2000). While she ultimately chooses to examine five regions, her discussion of history and economic development is consistent in many ways with the eight‐region approach we examine in this study.

Barrington (e.g. 2002; 1997) has employed a nine‐region framework in studies of mass attitudes in Ukraine.

Of the 1,606 respondents 249 came from the east region.

Both of these oblasts are well above the national average for population density, with Donetsk being the most densely populated oblast in the country.

Arel (1992) also points out that the two oblasts of Donetsk and Luhansk were treated as a separate “land” in proposals floated for a federal system in Ukraine in 1991.

Two hundred eighty of the respondents to the survey are from the east‐central region.

After Donetsk; see Hesli, 1995, Table 2.

Three hundred seventy‐three of the survey respondents are from the north‐central region.

Arel (1992) also mentions that these three oblasts were considered a single region (or “land”) in one of the federalization plans floated in 1992, and they form one region discussed by Melnyk (1992). One hundred forty‐four of the survey respondents are from the south region.

While the Donetsk and Luhansk provinces are over 40% Russian, the percentage is less than half that in Kherson and Mikolaiv (Solchanyk, 1994). But, as Stebelsky (1997) points out, non‐Soviet Russian place names are more common in the south than in most of the east. Melvin (1995) argues that for historical reasons ethnic relations between the Ukrainians and Russians in the south were also less polarized than in other parts of the country.

Krym is often included in the south. Seventy‐three of the survey respondents in our analysis are from Krym.

While support in all other oblasts was over 80%, in Krym only 54% supported independence.

The west‐central region includes 238 of the survey respondents.

Volyn and Rivne were re‐annexed into Poland from 1919 to 1939 (Arel, 1992).

Seventy‐two of the survey respondents are from the southwest region.

Zakarpatia was controlled by Hungary for 500 years and then was briefly part of Czechoslovakia; Chernivtsi was under Austro‐Hungarian rule from the 1700s until World War I and then under the control of Romania until World War II (Arel, 1992).

The west has 177 of the survey respondents.

The survey was the November 1998 Kiev International Institute of Sociology (KIIS) National Omnibus Survey. There were 1,606 respondents. Additional details about the survey can be found in the Appendix.

The variables examined in the statistical analyses of the survey data are all demographic factors. We have chosen to focus on demographic factors alone rather than mixing together demographic and attitudinal variables. This is in part because of the difficulties of analyzing demographic and attitudinal variables through estimating single‐equation OLS models (see Barrington and Herron, 2001). But it makes particular sense in this case, since the variation we are primarily concerned with—region—is a demographic variable. While some attitudinal variables may play a role in shaping views of the ethnic other, we assume in this study that they fall at a later stage of the causal process and are themselves affected by the demographic variables. As such, our demographic‐only model is a reduced form equation of a larger causal structure.

Those who would expect proximity to lead to hostile views of the ethnic other emphasize the importance of contact in feeding negative stereotypes and concerns. Those who see proximity as a factor in positive stereotyping stress that the more familiar one is with members of the ethnic other, the more difficult it becomes to hold stereotypes. But existing studies (e.g. Ray, 1983) indicate this is not the case with low‐level, informal contact.

Demographers would probably not agree with this label. While acknowledging that the variables include a wide‐ranging set of socioeconomic and demographic features, we use “demographic” to capture the difference between the variables analyzed in this section and attitudinal variables commonly included in statistical analyses of mass attitudes.

Bremmer (1994) points out something often overlooked as a part of this cross‐cutting cleavage, that Russians in the far west of Ukraine have a high (self‐reported) level of ability in the Ukrainian language.

Table 2 also includes “model fit” statistics. While one should generally not put great stock on R ² as an indication of whether or not a particular model is appropriate (see King, 1986), the use of the statistic as a way to compare alternative models of effects on the same dependent variable using the same sample is appropriate. In this case, the R ² statistic—and, more important, the adjusted R ² statistic, which takes into account the number of variables in the model—increases as one moves from the two‐region to the four‐region to the eight‐region model. In the traditional, though not completely proper, way of interpreting R ², one could say that a “greater portion of the variance” of the dependent variable is explained by the eight‐region model than by the alternative models. Note also that the standard error (SE) and F statistic for the equations also decrease as one moves from the two‐region to the four‐region to the eight‐region model. These are, again, signs of improved model fit.

In an earlier version of this article, we were criticized by reviewers for excluding the west region in the four‐ and eight‐region analyses. While the pattern of which region is more or less likely to see the ethnic other in positive terms is the same regardless of which region would be excluded, we have excluded the south from these analyses in this version of the article.

And generally much larger size of the coefficients compared with those of the other dummy variable terms.

The model fit statistics in the eight‐region model also again support its use over the others, with adjusted R ² and standard errors all indicating the better model fit of the four‐region model over the two‐region model and of the eight‐region model over the four‐region model.

The importance of the coefficient was another story. The 0.066 coefficient on the age variable in the eight‐region model, for example, means that a respondent 25 years older than another would support the government only 1.65 more on the support scale. Compare this with the difference from speaking Russian, living in a large cities or living in almost any of the regions other than the west.

Again, the west region was excluded in the two‐region model, while the south serves as the comparison region in the four‐ and eight‐region models.

This effect has been demonstrated in other mixed electoral systems (Cox and Schoppa, 2002; Herron and Nishikawa, 2001).

We focus on these measures of electoral effects because they vary at the district level. Data were not available at the district level for other variables that we considered including in the model.

Because the dependent variable cannot take on values less than 0 or greater than 100, we also conducted a tobit analysis for each model of parliamentary and presidential elections. The results of the OLS and tobit analyses were virtually identical in each case, so we report the OLS results here.

When the four‐region model is tested using tobit, SMD is significant at the 0.05 level.

Overall model performance improves across the regional definitions; the eight‐region model provides a better fit than the two‐ or four‐region approach and facilitates a more nuanced interpretation of regional effect. The coefficient for Krym is significant at the 0.10 level in the eight‐region model. The interactive term is dropped because the KPU fielded candidates in all districts in the region. In addition the interactive term for the southwest is significant at the 0.10 level.

Counter‐intuitively, KPU performance seems to be negatively affected by candidate placement in the eastern and western districts. The former is due, in part, to limited variation in nomination patterns in the east; the KPU placed candidates in all but four of the 35 districts in the east region.

It is possible that multicollinearity introduced by the interactive terms has caused Type II error—failure to reject the null when it should be rejected. Nevertheless, we find that the interaction between region and candidate placement is significant in at least some cases. This suggests that region exerts both a direct effect on performance and an indirect effect through candidate placement. For more on interactive terms, see Friedrich (1982).

The interactive term for southwest was excluded in the eight region analysis because Rukh nominated candidates in every district in that region.

SMD placement does not present an independent effect in part because there is limited variation on the variable; Rukh ran candidates in most district races.

The coefficients for east and east‐central are negative, but east‐central is significant only at the 0.10 level.

Higher levels of turnout among pro‐Kuchma voters in certain regions could also reflect ballot box stuffing (Herron and Johnson, 2003).

Data available at the Central Electoral Commission website (<http://195.230.157.53/pls/vp1/webproc0>).

The left includes the KPU, Agrarian Party of Ukraine, Progressive Socialist Party and Socialist/Peasant Bloc.

As in our analysis of parliamentary results, we focus on variables that vary at the district level.

The overall model fit is lower with the eight‐region model than the four‐region model; an increase in the number of regions permits a more detailed interpretation of regional effects, but the overall model is better predicted with fewer macro‐regions.

In the eight‐region model of round 1, the coefficients for east‐central and west‐central are significant at the 0.10 level, reinforcing variation in western support as well as Kuchma's relatively weak eastern support.

In fact, we might find substantial heterogeneity within and across regions, depending on the question that is asked. Does this mean that the regional effect is bogus? Not necessarily. Interests across regions can intersect on certain questions but not on others. Scholars of U.S. politics consider the South to be an important region and sometimes control for regional effects in their analyses of political attitudes and behavior. But, while the South may be different than some regions on some questions (i.e. abortion or prayer in schools), it may be indistinguishable from other regions on other questions. Our findings in Ukraine suggest that regional differences matter, but the nature of the question influences how regions matter.