The 2015 Turkish Parliamentary elections went, perhaps much to our surprise, relatively smoothly. The previously governing AK Parti (AKP) was denied for the first time in its history an absolute majority. The AKP received 41% of the votes (they achieved 49% in the previous parliamentary elections and 52% in the last presidential election). This won the AKP 258 of 550 available seats, 18 short of a majority. The secular opposition CHP retained its 25%, winning 132 seats. The right-wing MHP secured 80 seats with 16% of the votes. The biggest success, however, belongs to the “Kurdish” HDP, which has for the first time exceeded the national threshold of 10% with a comfortable 13%. This gave HDP 80 seats. Preliminary analyses suggest that the HDP carved most of its votes from the part of the Kurdish electorate which previously voted for the AKP, and that an additional 1.5% came from liberal Turks.
These results mark a significant blow to Mr. Erdogan’s quest for a constitutional reform which would make him the first executive President of Turkey with very generous powers. Yet it still remains unclear what will happen next. The HDP pledged before the elections that they would not form a coalition with the AKP. An AKP-MHP or an AKP-CHP coalition seems a more likely outcome. Alternatively, if a government cannot be formed within 45 days, Mr. Erdogan can call for an early election. This would seem to be the preferred outcome for Mr. Erdogan, as it would give the AKP a second chance to try winning a majority. A tough interim period without a government may persuade the electorate that a strong AKP government is better than chaos. All in all, as a Turkish idiom puts it, “showing death persuades one to settle for malaria.” Polls, however, report conflicting predictions on what the likely outcome of an early election would be. What does seem clear is that the performance of Mr. Erdogan and of the four biggest parties in the near future will determine whether the AKP will win a majority in an early election, should there be one.
The reasons behind the drop in popular support for the AKP deserves a comprehensive treatment which might be the topic of a future post. In this post, however, I will go back in time and present some curious statistical anomalies in two previous Turkish elections: the 2004 mayoral elections – the first mayoral elections of the AKP – and the 2014 Presidential election, the last election before the 2015 general elections. These two elections are of different types, and hence taken together draw a broad picture of the types of elections held in Turkey. The anomalies found in these two cases are also of different sorts. Yet, in both cases, a single party seems to benefit from these anomalies.
2004 Mayoral Elections
The regression discontinuity (RD) design in electoral studies relies on the assumption that the outcomes of extremely close elections are approximately random with equal chance of winning and losing. This ensures that incumbency in such close races is assigned randomly, which implies a natural experiment allowing for causal inference. Scholars have provided formal-theoretical accounts of why–apart from tiny disparities due to the curvature of the forcing variable (Snyder, Folke & Hirano, 2014)–close popular elections should indeed be random (e.g., Imbens & Lemieux, 2008; Eggers et al., 2014). In a nutshell, this is because no candidate can have full control over the results in popular elections, and in extremely close races arbitrary noise in the election system should overcome the few decisive votes that determine the result. Eggers et al. (2014) show after analysing 40.000 close elections in several countries that the outcomes of almost all close popular elections are random.
In the 2004 Turkish first-past-the-post mayoral elections, however, the outcomes of extremely close races were not random. The 2004 elections were run for 3209 municipal and 16 metropolitan mayoral seats. As I have shown in detail elsewhere (Aksoy, 2015) the AKP had a clear advantage in razor sharp races. Figure 1 plots the frequency distribution of the AKP winning/losing margin, which is simply the vote share of the AKP in a mayoral district minus the vote share of the largest of the remaining parties. As Figure 1 shows, the chance that the AKP wins an extremely close race is much higher than the chance that the AKP loses an extremely close race. A more formal McCrary (2008) test shows that the jump around the winning threshold of zero in Figure 1 is statistically highly significant (Discontinuity estimate = 0.23, Standard Error = 0.08).
There are additional anomalies that accompany this discontinuity around the winning threshold. For instance, there is a statistically significant drop in election turnout and a significant drop in the vote share of the largest contending party in districts in which the AKP barely won compared to the districts in which the AKP barely lost. The parties that disproportionately lost these razor-sharp races to the AKP were exclusively right wing and/or Islamic parties – that is the ideological competitors of the AKP. These parties were the Islamic Saadet Partisi, and the right-wing MHP and DYP. The secular CHP were not particularly advantaged or disadvantaged in close races. No party except the AKP had an advantage in close races in 2004.
Note that the 2004 mayoral elections were the AKP’s first mayoral elections. Hence the AKP did not have an obvious incumbent party advantage. Furthermore, a comparison of the 2004 results with the 2009 mayoral elections shows that there is a statistically significant incumbent party disadvantage for the AKP. In brief, the anomaly in Figure 1 cannot be explained by an incumbent party advantage. A detailed description of these and additional anomalies in the 2004 mayoral elections and a discussion of potential explanations can be found in Aksoy (2015). It suffices here to say that there seems to be no easy and fair explanation.
Figure 1. AK Parti winning/losing margin (AKP vote share minus the vote share of the largest of the remaining parties) in 2004 mayoral elections. Every observation is a mayoral seat.
2014 Turkish Presidential Election
For the first time in Turkish history, the president was elected by popular vote in 2014. Three candidates – Mr. Erdogan of the AKP, Mr. Ihsanoglu of the CHP, and Mr. Demirtas of the HDP – participated in these elections. In the first of the two possible rounds Mr. Erdogan gained 51% of the votes and won the election. If no candidate had gained more than 50% in the first round, there would have been a second first-past-the-post round between the two most popular candidates.
Erik Meyersson of the Stockholm School of Economics has discovered a number of anomalies in the 2014 Turkish Mayoral elections which he describes, e.g., . What he discovered was a systematic positive relationship between the proportion of votes deemed invalid in a ballot box and the vote share of the AKP in a ballot box and a systematic negative relationship between the proportion of invalid votes and the vote share of the opposition. Moreover, which opposition party’s vote share was negatively associated with the proportion of invalid votes depended on the district. In districts in which the biggest competitor of the AKP was the CHP, it was the CHP vote share, and in districts in which the biggest competitor of the AKP was the MHP, it was the MHP vote share that decreased significantly with the proportion of invalid votes.
Following Meyersson’s analysis, we will now look at how the absolute number of votes for the three candidates in the 2014 presidential election is associated with the absolute number of votes deemed invalid in a ballot box. Of course, in such a regression specification, one should also account for the number of total votes cast in a ballot-box. This is because the number of invalid votes is highly correlated with the number of total votes in a ballot box, which, in turn, is obviously correlated with the number of votes cast for any candidate. Also, some unobserved characteristics of a ballot box, such as the average education of the voters, can influence positively both the likelihood of voting for a specific party and the likelihood of casting a vote improperly which would later on be deemed invalid. Including polling station fixed-effects in the regression model should solve this unobserved heterogeneity problem to a great extent. When polling station fixed-effects are included, all variables at the station level, such as average education, income, etc. in a neighbourhood, is controlled.
Table 1 shows these regression models. The table shows that for a single vote deemed invalid, there is a 0.84 decrease in Ihsanoglu votes, controlling for everything at the polling station level and the number of votes cast in a ballot box. Similarly, for a single vote deemed invalid, there is a 0.21 decrease in Demirtas votes. However, the effect of the number of invalid votes on the number of Erdogan’s votes is nil. Negative effects for other candidates on the one hand, and a nil effect for Erdogan votes on the other is telling. It shows firstly that there is no inherent tendency for Erdogan voters to cast a vote which would be deemed invalid afterwards. Such a tendency, if existed, would explain much of the Meyersson’s findings. The results in Table 1 are consistent with the hypothesis that there are systematic differences across candidates with respect to the likelihood that a vote cast for the candidate is deemed invalid.
Table 1. Regression models predicting the absolute number of votes for candidates, polling station fixed-effects are included in the models
Figure 2 below plots visually the same relationship shown in Table 1 between the number of votes cast for the three candidates and the number of invalid ballots in a ballot box.
Figure 2. The relationship between the number of votes for a candidate and the number of votes deemed invalid in a ballot box in the 2014 presidential election.
We could calculate what effect, if indeed existed, a differential discarding of non-Erdogan votes had on the aggregate outcome. The total number of invalid votes excluding prisons and abroad (733.464) corresponds to about 1.8% of all votes in the 2014 presidential election. These invalid votes yield no change in Erdogan votes, but are associated with a combined overall drop of 1.8% in Ihsanoglu and Demistas votes (the sum of all coefficients in Table 1 should add up to -1). This means that if there was no invalid vote, we would expect a drop of 1.8% in Erdogan’s overall vote share. This would be enough to ensure a second round (Erdogan’s first round vote share was 51.18%). If invalid votes were distributed to candidates proportionally to their vote shares, about half of these invalid votes would be Erdogan votes. This would still ensure a second round. A second round, had opposition was able to force, would mark a significant symbolic success for the opposition.
I leave it to the reader to find explanations for these anomalies. Perhaps the commentary section below might help start a discussion. I must add that the anomalies found in Turkish elections are not as extreme as those found in countries such as Russia and Uganda (see e.g., Klimek et al., 2012). Rather, they seem to fit the mold of what Turks call a “provincial Jinn” (Taşra Cini) would typically do. The expression is used to describe a person who otherwise seems righteous but when he has the chance displays small opportunistic behaviours here and there. The anomalies described above seem to be in the form of small “kicks” in favour of a particular party in critical districts and, depending on whom the vote is for, a slightly more or less lenient interpretation of what counts as an invalid ballot.
A valid question, then, is the following: if there have been such anomalies, why did the AKP lose the majority in the 2015 parliamentary elections? There are two possible answers. Firstly, these anomalies might have been prevented in 2015. This may likely be the case as in 2015 there was a massive civil awareness for the integrity of the election process. As an example, in 2015 Oy ve Otesi , a civil non-partisan initiative, has sent about 40,000 observers to poll stations in 45 cities and 162 districts. Such initiatives likely minimised anomalies. Alternatively, these anomalies may still have been present, but their presence was simply not sufficient to tip the balance. A detailed analysis of the 2015 election results will show whether this was the case. The official 2015 election data have just been published by TurkStat and if there is demand, I can present an analysis in a next post.
The discussion so far shows that the fall of the AKP in the 2015 parliamentary elections came despite a number of anomalies that have previously favoured the AKP. This indicates further how significant the fall was.