E.2.b Ponderación de los intereses en juego

Two dependent variables are estimated by region. First, a turnout test is used to estimate overall levels of manipulation (Myagkov et al., 2009). Second, a vote-share test designed by Rozenas (2017) is used to estimate the number of precincts in each region that may have been subject to falsification. To estimate the overall degree of manipulation, I conduct turnout tests on the results for the major pro-eastern party in the election. These tests assume that, in a clean election, a party’s absolute vote-share (the party’s number of votes divided by the number of registered voters) should increase consistently across precincts within a territory, as precinct turnout increases. This approach assumes homogeneous precincts, an assumption which is bolstered by testing precincts within regions or through the use of relevant socioeconomic control variables to weed out potentially confounding relationships (Deckert, 2013). I employ both approaches here.

First, I use multilevel models to estimate the relationship between turnout and absolute vote- share for both parties, for each region-year, along with standard errors for those coefficients. The coefficients generated at this first stage are then used as dependent variables in a second-stage model. Turnout coefficients in the Ukrainian case appear systematically different from those in the Russian case, discussed in a previous chapter. While the ruling United Russia party regularly produces large turnout coefficients, suggesting that the party draws relatively more votes from the population as turnout increases—likely illicitly—major parties in Ukraine frequently exhibit negative turnout coefficients. A negative coefficient indicates that a party systematically loses absolute vote-share as

turnout increases; that is, it draws fewer votes in absolute terms when turnout is high than when turnout is low—a suspicious outcome, though one that is of course not impossible under a clean election. A notional example helps clarify this.

A party’s absolute vote-share per precinct is the number of votes it receives divided by the total number of registered voters in that precinct. Turnout is the total number of votes cast divided by the number of registered voters. If there are 1,000 registered voters in each precinct, a precinct with ten percent turnout will have 100 total voters, and a precinct with ninety percent turnout will have 900. Assuming that a party has an approval rating of about forty percent across all precincts, under a clean election the party will receive about forty votes in the low-turnout district, and about 360 in the high turnout one. The corresponding turnout coefficient, represented by the slope of the line between the two points, will be about 0.4. For the party to produce a negative turnout coefficient, significant deflation in vote-share must occur; in other words, the number of votes it receives in high-turnout regions must be exceedingly low. In the current example, if the party wins forty percent of votes in the low-turnout precinct and one percent of the votes in the high turnout precinct (i.e. 9 out of 900 votes), it will produce a small negative turnout coefficient of -0.04. In other words, to earn even a small negative coefficient a party must suffer dramatic losses in absolute vote-share as turnout increases. Here, the party wins 40 out of 100 votes in the low-turnout precinct, and only 9 votes out of 900 when turnout is high.4

Negative turnout coefficients for Ukraine’s major parties are not small, as Figure 4.1 illustrates. Large numbers of precincts show turnout coefficients that are not only negative, but substantially so. Such an outcome is highly unlikely to occur in a clean election, as discussed above, but a small positive turnout coefficient cannot be easily regarded as suspicious—a party might simply have a very low latent level of support in a region. As a conservative indicator of possible deflationary manipulation, then, I code a dummy variable that takes on a value of one for regions with a negative turnout coefficient. This binary variable, deflationary manipulation, is used as a dependent variable in logit models.

4

The alternative source of a negative coefficient, in which the party’s vote-share in the low-turnout precinct is highly inflated, is not mathematically possible. In the running example, assuming that the party wins forty percent of votes in the high-turnout precinct, it would have to win 390 votes in the low-turnout precinct (where only 100 votes are cast) in order to produce a small negative turnout coefficient.

Figure 4.1: Histogram of turnout coefficients for major pro-east parties (2002-2014)

The vote-share test, used for the second dependent variable, is conducted as follows. For each election year, the precinct-level results are broken down by region. Within each region-year, the results are broken out by party. I examine the results for the primary pro-eastern party for the election year. The Rozenas test is then conducted on the results for each party, by region, according the method described below.

In brief, when the distribution of a party’s vote-shares by precinct is plotted, the distribution will exhibit random peaks that represent a clustering of precincts around particular vote-share values (see Figure 4.2). By random chance, for example, a large number of precincts might report vote-shares of 32 percent for the party, resulting in a peak at that value. In an election without falsification, these peaks should be randomly distributed based on voters’ stochastic decisions to turn out and cast their ballots one way or another. However, a particular kind of falsification will distort these random patterns and produce statistically unlikely vote-share clusters (possibly, but not necessarily, at values divisible by 5—55%, 60%, and so on). The test compares the ‘spikes’ in the actual vote distribution to those produced randomly through bootstrapped samples of the observed precincts. By comparing the real distribution with the hypothetical distributions, it is possible to estimate the number of precincts per region where fraud is suspected.

Figure 4.2: Vote-share density for the Party of Regions

Such non-random clusters can result from falsification due to human propensity to target whole numbers when making up figures, but are more likely to reflect agents’ fulfillment of direct specifications from their principals (Rozenas, 2017). Consequently, while this election forensic technique is only an indicator of falsification, it is especially well-suited to answering the question at hand since it allows for a test of how much agents engage in the kinds of behavior that can signal loyalty to particular patrons. The measure is used to produce a count of suspicious precincts, which is used as the dependent variable in Poisson models.