Since the referendum last Sunday, there has been much discussion over the relatively high support for the YES-vote in Turkey’s predominantly-Kurdish Eastern part of the country. To give a quick visualization of it, the below graph on the left-hand side shows the difference between the number of YES votes in the referendum and the number of AKP votes in the last (November 2015) parliamentary election, by geographic (NUTS-1) regions. (These numbers are not official data so maybe updated slightly).

The largest increase can be observed in Southeast Anatolia, but Central East and Northeast Anatolia are also among the top regions where the AKP(‘s amendments) gained the largest increase in votes. These three Anatolian regions added circa 570,000 votes (the smaller number reported in my previous post is from narrower subset of provinces) corresponding to ca 38% of the added votes nationally. This is especially noteworthy given that this area corresponds to just under 15% of the electoral population.

Continue reading →

There are a number of striking features of Sunday’s referendum in Turkey, which resulted in a narrow victory (51.4% being the official “YES” vote share at the time of writing this blog post) for the constitutional amendment proposed by the government.

There are, however, numerous challenges to the result, both with regards to the environment in which it was carried out as well as more direct accusations of election-day irregularities. Going through the election data is probably going to keep many people busy for the next couple of days weeks months. For now, I wanted to highlight one particular phenomenon which struck me as interesting.

At a first glance, when plotting the vote share distribution of the YES vote at the ballot box level was that the “Never-AKPers” were gone. These Never-AKPers are voting groups that have extremely few or literally zero votes for the AKP. In past elections, they have been quite noteworthy. The below illustrates this by plotting the distribution of the AKP vote shares for the 2011, June 2015, and November 2015 parliamentary elections and in the bottommost graph the YES vote share in Sunday’s referendum.

Essentially, the fat left tail (i.e. the ballot boxes with next to zero AKP votes), which stands out in the previous three elections is almost entirely gone in 2017. What happened? Did the areas where the AKP was the least popular suddenly have a change of heart?

Continue reading →

There are several interesting phenomena evident in the last Turkish elections. Whereas my previous post dealt with a couple of these, in this one I want to focus more on a specific region of Turkey, namely the Southeastern provinces, sometimes also referred to as the Kurdish provinces. (There’s a discussion to be had about what distinguishes Kurdish- from non-Kurdish provinces, but for the below analysis, I will semi-arbitrarily include Diyarbakir, Sanliurfa, Bitlis, Bingol, Mardin, Mus, Agri, Igdir, Batman, Sirnak, Siirt, Van, and Tunceli in the group of Kurdish provinces).

As I wrote back in June, this region saw a significant voter swing from the AKP to HDP, going a long way to helping the latter over the ten percent threshold required for participation in the Turkish parliament. In the most recent election, some of these voters appear to have switched back. The upper panel of the below graph shows differences in the AKP vote share between both November 2015 and June 2015, as well as June 2015 and June 2011, for the Kurdish and Other provinces. In June 2015, the vote shift away from AKP for the region was almost twice as large in relative terms as that in the rest of Turkey, whereas in November 2015, the relative swing in favor of AKP was larger than for the rest of the country.

Another distinguishing factor of vote patterns in the Kurdish provinces is not only that the Kurdish provinces exhibit larger changes in turnout from election to election (see the bottom panel of the below figure) but also that whereas in the most recent election the rest of Turkey experienced an increase in turnout compared to June, the Kurdish provinces experienced a decrease. Still, even though the average change in turnout was negative, there is still plenty of variation (which can be seen from the red curves, compared to the blue curves, being “shorter and fatter”), and a significant number of the neighborhoods in the latter provinces experienced significant positive increase in turnout.

Undoubtedly, one possible explanation for the decrease in turnout and the increase in AKP vote share could be the ongoing conflict between the Turkish state and the PKK.  The decrease in turnout could be due to safety concerns in conflict-affected areas, which might be more likely to affect pro-HDP neighborhoods, and might also be correlated with a shift to the AKP as a means to achieve stability for some.

In this case, we would expect an overall negative relationship between the change in turnout and the change in AKP vote shares between the November and June elections. As I showed back in June, the large increases then observed in HDP vote shares in provinces like Agri, Igdir, Mus, Batman, Diyarbakir etc, were also accompanied by substantial increases in turnout.

This relationship, however, changed in the November elections, as can be seen in the below graph which plots (local averages of equally-sized bins for) differences in party vote shares against differences in turnout between November and June.

The main thing to note is the apparent “kink” in the relationship for the AKP and HDP (and to a certain extent in the MHP) graphs. In both of these graphs, the slope of the line appears to change sharply at the zero turnout change point. For AKP, the party appears to benefit not only the larger the fall in turnout is, but it also larger the increase in turnout is. (The HDP graph has the almost exact inverse relationship, which is unsurprising since in the Kurdish provinces votes tend to go either to AKP or HDP and so a positive change in the former implies a negative change in the latter.)

Before I go any further with this, an important question to ask is why this might matter. It matters because this kink is not to be seen in the previous election, nor is to be seen for the rest of Turkey in either the current or past election. In order to demonstrate this, I below compare differences in AKP vote shares and differences in turnout for Kurdish versus the Rest of Turkey, for the November and the June election respectively. Also, to make the analysis somewhat more robust I use the residuals from regressing the difference in party vote shares on lagged levels of turnout, AKP vote share, and the log number of registered voters in a neighborhood. As before, each dot represents the local average of equally-sized bins.

Here, it is clear how the slope changes only for the Kurdish provinces, and how it is the neighborhoods with positive turnout change that exhibit a peculiar slope; in November they exhibit a positive relationship with the change in AKP vote shares, whereas in June it was negative.

One possibility could be that this reversal of the slope does not represent a kink but rather some non-linearity that may have come about for a number of reasons that are unknown, yet completely innocuous. Moreover, the non-Kurdish provinces suggest that a non-linear relationship per se is not necessarily something strange. So how do we know this is actually a real “abnormal” kink and not just some “normal” non-linearity.

Fortunately, we can test for the presence of a kink at zero turnout change using Regression Discontinuity methods, which have relatively recently been extended to allow for discontinuous changes in the first derivative of the outcome (rather than the level of it). This can also easily be implemented with local linear RD models with optimal bandwidths using the methods and tools developed by Calonico, Cattaneo, and Titiunik (hereby CCT, see here, and here).

Using RD shifts the focus entirely on the cutoff, which in this case is the zero turnout change, and allows us to answer the question of whether an infinitesimal change in the “running variable” (here the change in turnout) results in a discontinuous change in the first derivative of the outcome variable.

(Note that whereas RD designs are usually employed to infer causal effects, here I’m merely interested in whether the first-derivative changes in a statistically significant way.)

Below, I show a simple RD kink plot with data in bins and a quartic polynomial fit on each side of the zero turnout change cutoff.

The slope appears to change discontinuously at the threshold, and formal testing using CCT’s rdrobust command in Stata results in a statistically significant estimate of 2.6 (the left-hand slope is -0.26 and the right-hand slope is 2.35). On the right-hand side of the cutoff, this represents a large effect, as a one percentage point increase in the difference in turnout corresponds roughly to a 2.4 percentage point increase in the difference in AKP vote share.

Sure, this may seem odd, but why would this necessarily imply anything irregular? And is there a way to link this analysis with that in my previous post?

The previous post focused on the last digit distributions of party counts, and one of the main findings in that last digit distributions changed character in the Kurdish provinces between the November and June elections. One interpretation (although there may be other ones too) of this is not just that these last digits are from different distributions, but also that this implies manipulation.

So a logical follow-up related to the above analysis is whether we can observe a discontinuous jump in the distribution of the last digit at around zero turnout change. Similar to the outcome in the past post, I use the mean of the last digit, with the difference here being that the outcome variable is the neighborhood-level average of the party vote count’s last digit across all ballot boxes within that neighborhood. And as this mean should arguably be flat, I look for the existence of jumps in the level as opposed to the first derivative as above.

Below, I illustrate the RD design for discontinuous jumps in the last digit for AKP and the HDP’s vote counts for November and June respectively. Under each graph, I also show the RD estimate and its associated standard error and optimal bandwidth used in the estimation (again using the rdrobust command).

In this case, and as opposed to that in the previous post, the test is not whether the last digit distribution changes across time but whether it changes at zero turnout change in each of the November and June elections respectively. The stipulated assumption here is therefore that, absent manipulation, we should not observe any discontinuous jumps at the cutoff around zero turnout change.

As can be seen form the above figures, the AKP last digit exhibits a statistically significant jump around the cutoff in the November election, with an estimate of around -0.7. which corresponds to roughly an 18 percent change relative to the mean just to the left of the cutoff. Again, this is a rather significant effect even in terms of magnitude.

In the June elections, the AKP last digit also exhibits a negative RD estimate although it’s somewhat smaller in magnitude. As for the HDP, there is no statistically significant jump at the cutoff and estimates tend to be much closer to zero. Consequently, to the extent that a discontinuous jump here implies manipulation of votes, it’s only occurring for the AKP’s votes.

One mechanism which could drive the above result that comes to mind is if the AKP, for a subset of neighborhoods, was able to “add” AKP voters in such significant numbers so as that all of the neighborhoods in which this occurred resulted in positive turnout relative to the June election. If, in turn, these were manipulated votes and their numbers decided by an individual as opposed to a random process, this could explain the jump in the last digit distribution at the cutoff.

Nonetheless, as with the previous statistical analysis of Turkish election I’ve done on this blog, the by now usual disclaimer applies: these statistical relationships could be considered consistent with some form of voter manipulation, but they are by themselves not necessarily proof of it. And just because I have not been able to imagine any more innocuous explanations for these kinks and jumps doesn’t mean they don’t exist.

Sunday’s elections in Turkey were a landslide for the ruling AKP. Its vote share rose nearly 9 percentage points from what it received in June. One interpretation is that AKP’s political strategy since its summer defeat has paid off, a chilling evaluation of one that has at times seemed both divisive and violent, not to mention authoritarian.

As in last elections, much of the change in voting seems to have occurred among nationalist as well as Kurdish voters, with this election seeing a difference of priority among them. Whereas June’s election was HDP’s to win, this one appears to have been to a large extent the nationalist MHP’s to lose. As the below figure shows, plotting the difference in vote share between November and June, the AKP’s gain appears to come predominantly at the expense of MHP. In some other cases, the vote swing seems to be driven by voters in Kurdish provinces leaving HDP for AKP (likely the poor and pious I have discussed in this blog before).

Part of the story could be explained by turnout. After all, several provinces show significant changes in turnout compared to the June elections. Several Kurdish provinces like Agri, Batman, Hakkari show substantial reductions in turnout, likely a result of the ongoing conflict between the PKK and the Turkish state.

Election night was particularly embarrassing to Turkish pollsters who in unison (almost, at least) were predicting a repeat of the June elections. In fact, using the mean and standard deviations of this sample of pollsters, predictions were off by an incredible 4.9 standard deviations.

Continue reading →

(Note: This post has been updated to 1) add Istanbul to the provinces where Kurdish parties were active in 2011, 2) adding the other large cities Ankara, Izmir, Bursa, and Adana to the neighborhood-level analysis, and 3) adding graph of partial correlation between Refah vote share in 1995 and HDP vote swing in 2015.)

Given yesterday’s tumultuous election in Turkey, I thought it might be a good idea illustrate using the data available what just happened.

There’s the obvious: AKP lost about 10 percentage points of its vote share, and the Kurdish-and-what-not HDP received around 13 percentage points, pushing it above the ten percent threshold, allowing it to take seats in parliament – as far as I know, the first time a political party with such a clear pro-Kurdish constituency has done so. This means AKP’s seats in parliament fell from the 327 it won in 2011 down to 258.

Then there’s what it all means, which there’s no way I can discuss in one post. Instead I want to focus specifically here on the HDP and what kind of electorate brought it above the ten percent threshold.

Some herald Demirtaş. the HDP, and its electoral success as the comeback of the left or liberalism (here and here), noting amongst others his supportive stance toward the LGBT community as well as his background as a human rights lawyer. It is not for nothing that many refer to him as “Kurdish Obama”.

A following question is then to what extent HDP’s electoral success is a manifestation of the voting power of progressives and liberals in Turkey?

Despite talk of “borrowed votes”, i.e. strategic voting by (I assume) predominantly traditional CHP supporters, an initial look at the election suggests that what pushed HDP into parliament was a shift among traditional right-wing voters – the socially conservative Kurdish communities in the East and some living in the large cities who abandoned the AKP for the HDP.

Continue reading →

I recently (finally) had the chance to take a look at the Turkish presidential elections data from August 2014, with the aim of comparing some of the measures of election irregularities I documented from the earlier local elections in March of the same year. The analysis from those elections were essentially a snapshot of potential irregularities with little to say about how this evolved over time.

What I focused on last year was the rather peculiar degree with which the share of invalid ballots appeared to be positive correlated, at the ballot-box level within voting stations, with AKP vote shares (or rather the vote share difference between the AKP and the largest opposition party). Now, with the presidential elections data, I can compare how this correlation differs between that election and the local elections. As for the interpretation of the correlations, here’s what I wrote in April last year:

“An obvious question mark in the analysis is to what extent any correlation represents systematic mistakes, not systematic fraud. Suppose the AKP has a higher support among the illiterate who are more likely to make mistakes when voting. In this case, we would not be surprised if there was a correlation between invalid ballots and AKP support. An explanation would be that those more likely vote for the AKP are also those more likely to make voting mistakes and have their ballots declared invalid. Given the large difference across districts in the large cities in Ankara and Istanbul, one can easily imagine this as a plausible explanation for the simple unconditional correlations.

It is here that the fixed effects used in the previous analysis becomes crucial, i.e.  including fixed effects (FEs) to regressions of vote shares on invalid ballots control for all factors that vary across the FEs. Adding FEs for districts (Ilce) means we’re only looking at variation across ballot boxes within districts, whereas adding FEs for voting station means only looking at variation across ballot boxes within voting stations.

When doing this, although voters going to the same station to vote may still differ along several characteristics, it is much more difficult to argue that this systematically affect their likelihood of making mistakes in voting. The strength of the FEs is thus not that they control for everything, but that they reduce these differences to the point where it is less likely that the remaining differences represent an competing explanation for the correlation.

Furthermore, to the extent that this represent fraud, one would expect the relationship between AKP’s vote share and invalid ballots to be stronger in races with significant competition and less likely in races where the AKP was safe. (For the obvious reason that there is little return to engaging in fraud in races where you’re expected to win without fraud.)”

Below I show such correlations between the AKP’s vote share and the invalid share of ballots, with red representing the August presidential elections and blue representing the local elections. These are essentially the same kind of graphs as the first set of ones I posted here, the main difference being that I now from the start subtract the voting station means from the ballot-box level data. Another difference is that instead of plotting the raw voting data (which, with large amounts of data, makes graphs overly crowded and visually less attractive) I’m plotting equal-sized binned means – essentially a scatterplot of grouped data, with groups made up of equally large number of observations (see here for an easy way to implement such grouped scatterplots in Stata).

Continue reading →

Turkey’s heading for another election, arguably one of its most important so far, as made clear by the unusually vitriolic statements by leading government members (and opposition candidates too, I gather, if the Turkish media reported on them to the same extent), seemingly stumbling over each other to say anything that would make them noticed.

Much ink gets spilled over the main issue at hand this time around: whether the AKP will get the super-majorities required to either unilaterally write a new constitution, or to push one through a referendum. A particular trait of the current constitutional debate is President Erdoğan’s desire for a stronger presidency in Turkey (and one specifically without the checks and balances of, for example, the US). The focus on this issue is quite understandable, as it could change the formal rules of government in the country quite substantially.

A concern, however, is that a too singular focus on Erdoğan and his desire for a presidency draws attention away from what the problem of the current institutional setting really is. And that isn’t parliamentarism as a system of government, as Erdoğan would like his constituency to believe, but the poverty of Turkey’s democracy overall. And while Erdoğan and his AKP government may very well have improved the living standards of its constituency, they have so far failed abysmally in extending the same courtesy to the country’s democratic system.

This post is about one particular segment of Turkey’s institutions: its elections.

Last year, just after the local elections, I wrote a series of blog posts (1, 2, 3) suggesting all was not well with how votes were counted. At that time, although members of the opposition had been complaining about this for a while, this appeared to hit a raw nerve. Within a week of my initial post, my blog had been visiting more than 70,000 times, and the various posts had been cited or linked to in FT, Washington Post, The National Interest, Jadaliyya, Christian Science Monitor, Bloomberg View, Dagens Nyheter, Cumhuriyet (1 2, 3), and Radikal. In the comments section of my blog, pro-government trolls were threatening the election NGO that had shared their data with me, and the “newspaper” Takvim generously nominated me a member of the “banker lobby” together what they perceived to be my main co-belligerents: an Australian media mogule, referred to as “the Jewish Baron”, and a Swedish tycoon, dead since 1938. An unusual week, to say the least.

Continue reading →

The troubles and irregularities in Ankara’s mayoral election keep piling up.

A recent article in McClathyDC has many interesting points. It notes that around a quarter of all ballot boxes arrived at counting centers without an official stamp from the election board:

The unstamped ballot box tallies alone raise serious questions about the outcome. The summaries represented more than 713,000 votes – nearly a quarter of the 3.3 million votes cast, the CHP said. Incumbent Mayor Melih Gokcek, a stalwart in Erdogan’s Justice and Development Party, or AKP, defeated CHP candidate Mansur Yavas by about 31,000 votes.

The case of the missing stamps are not isolated events:

Tally sheets that lacked the official stamp were a major irregularity. Other means were more subtle. An example was the redistricting decided by the AKP-ruled Parliament in a late-night vote in 2012, which moved conservative rural municipalities into CHP-leaning districts in the city of Ankara for the mayoral election. The CHP also charged there had been at least 53,000 duplicate registrations; that out of 124,000 votes that were voided, 100,000 had no reason stated; and some 450 tally forms had their contents incorrectly entered in the official database. The total the CHP says it lost is at least 120,000 votes, possibly double that.

As the counting went into the night of March 30-31, top government officials are reported to have descended on the counting center in one of the most crucial districts, Yenimahalle, and staying in one instance for hours.

Continue reading →

(Note: This post may be updated)

My previous post documented a peculiar relationship between the share of invalid ballots and higher voting share of the ruling AKP government in last week’s local elections in Turkey. In both Ankara and Istanbul, this relationship was robust to across-district, even across-voting-station, differences. Thus even within a single voting station, like a primary school etc, invalid ballots appear to drive up the vote share for the AKP. Moreover, the relationship appeared to particularly bias votes toward the AKP in areas that tend to have more support for the opposition.

By now, that post has received considerable attention (see here, here, here and here), raising the question of whether these elections included non-trivial degrees of voting fraud. I’ve from the start been deliberately careful in not labeling this as fraud, partly due to the preliminary nature of the analysis, and partly because statistical anomalies remain so only until they receive adequate explanations.

As additional data has come in, thru the CHP-STS data collection system, and thanks to Twitter user @erenyanik for sending the data my way early on, I’ve had some time to replicate this analysis for additional cities. Continue reading →