What happened in Ankara 2009 Elections?

Note: To read an extended summary of the two analyses of the local elections (this article and the next one) in Turkish please see here.

Several people have been analyzing different aspects of the preliminary results from Turkish local elections collected through the Yerel Seçim 2014 Sandık Takip Sistemi (the ballot box tracking system), an initiative of CHP, the main opposition party in the parliament.

Erik Meyersson, an economist at Stockholm School of Economics, posted a few articles analyzing this data. His first article reports preliminary evidence from Ankara and Istanbul Mayor elections indicating a positive relationship between invalid vote share and AKP-CHP vote margin at the ballot box level. The magnitude of the relationship was smaller when district- or voting station-specific factors were netted out, but it remained statistically significant. While definitely not conclusive, these results might be a sign of electoral manipulation whereby a disproportionately greater [lower] share of votes cast for CHP [AKP] were invalidated (either blatantly or through indirect methods).

One may wonder if the relationship between AKP-CHP vote share difference and the invalid ballot share was also there back in 2009. If the answer is no, then what we have in 2014 is certainly more likely to reflect voting manipulation, because it is harder to come up with a structural explanation of the documented relationship that would hold today but not 5 years ago.However, AKP has been in power for more than 10 years now, and there is no reason to doubt that incentives and capabilities to manipulate back in 2009 were dramatically lower than it was in the last elections. Then, one should expect a similarly significant relationship in the previous local elections. To see what the answer is, I repeat Erik's analysis of the 2014 elections using 2009 official results for Ankara.

First, I will focus on election results for district municipalities --as opposed to metropolitan area mayor. The figure below plots the unconditional relationship between AKP-CHP vote share difference (akpchpdiff) and the share of invalid votes (invshr) at the ballot box level. The results indicate a positive and significant association, just like in the last elections !

The slope coefficient (4.46) is somewhat smaller compared to what Erik has found (5.86). To see whether the relationship remains qualitatively intact after we account for district level unobserved factors (like socio-demographic characteristics) that might influence the vote shares, I add dummy variables for each district. Hence, the figure below simply shows the relationship between the residual variations in akpchpdiff and invshr respectively after netting out the association between cross-district variations in these variables.

Results are equally significant albeit the size of the estimated coefficient (2.46) is about half the unconditional coefficient. This indicates that indeed some district specific factors (including but not limited to incentives to steal votes) were partly (but not fully) responsible for AKP's electoral advantage vis a vis CHP as well as the share of invalid ballots. I should note that the magnitude of my estimate is strikingly close to the estimate (2.72) Erik reported in his corresponding specification (see here).

Finally, we can go one step further and control for voting station fixed effects. To be fair, that the relationship may completely disappear at this stage does not say much about the relevance of voting manipulation. On average there were roughly 14 ballot boxes in a given voting station, as opposed to about 415 ballot boxes per district. Thus, there may not be enough variation in our variables of interest within each voting station to obtain a sufficiently precise estimate even if there is a systematic relationship between vote shares and invalid ballot shares. Or it may very well be the case that the feasibility of and incentives for manipulation are quite unlikely to differ much across ballot boxes in a given voting station.

The figure below plots the same relationship, this time conditional on voting station fixed effects.

Not surprisingly the relationship is much lower in magnitude. However, it is positive and very precisely estimated. Once again, the coefficient (0.39) is very similar to what Erik found in his analysis (0.40). At this point, I should add that the 2014 data Erik has been using belongs to metropolitan mayor elections while the one I have used in this analysis come from district municipality elections. While, voting behavior in the two types of elections should be highly correlated, it is still interesting that the factors which are driving these results (whatever they are) seem to have persisted throughout last 5 years. To what extent this remarkable persistence between 2009 and 2014 elections can be attributed to persistence in potential electoral fraud is a big question that begs an answer. Perhaps more on that later on.