Ballots, Blocks, and Bullets (and Bulldozers) – The Vote in Eastern Turkey

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.

Another striking aspect of this graph is that, whereas most non-Eastern regions experienced increases in the number of total votes (and a somewhat smaller increase in AKP votes), the experience in the East was quite different. Here, the number of votes either declined overall or were marginal in comparison to the added votes for the AKP. Thus, AKP’s proposed constitutional amendments gained additional votes in a region where the number of total votes declined.

Within the Eastern regions, in the right-hand graph, Diyarbakir and Sanliurfa added the largest amounts of new votes for the AKP, followed by Van and Mardin. This by itself is unsurprising, as these are also (bar Gaziantep) the most populous provinces in the region. Yet even if one further includes Gaziantep, these account for just over half of all added AKP votes from the East. The rest is from the more peripheral provinces, like Sirnak, Agri, Hakkari etc.

A number of factors have been brought forth to try to explain this.

The government position is essentially that the pro-AKP swing represent changes in inhabitants’ political preferences, and that by “making a very clear distinction between the Kurds and the PKK, President Erdoğan and the government have won the Kurdish confidence again.”

The major opposition party in the region contends that these dynamics are an outcome of the government’s actions to curtail its ability to campaign in the region:

“the government embarked on a relentless campaign against the country’s largest pro-Kurdish bloc, the Peoples’ Democratic Party (HDP) and its sister party, the Democratic Regions Party (DBP). At least 10,000 Kurdish lawmakers, officials and supporters remain behind bars over thinly supported terror links. They include the HDP’s co-chairs, Selahattin Demirtas and Figen Yuksekdag, and 12 fellow HDP parliamentarians, one of whom was arrested after the referendum. DBP’s Co-chair Sebahat Tuncel is also in jail together with 85 mayors elected on the party’s ticket in the southeast, where 83 DBP-run municipalities have been taken over by government administrators.”

As such, voters might perceive the HDP to have fewer means to end the fighting, perhaps believing it is only the government that can stop it. As Bloomberg reported a village elder saying “the HDP can do nothing for us now, and we will give our votes to whomever can help us.

This could also have implications for the practice of “block voting” , where local elites determine how their constituency votes.

The fighting has further resulted in extensive internal displacement of large numbers of voters following heavy fighting and the destruction of significant parts of cities like Sur in Diyarbakir, Nusaybin, Yuksekova and Cizre, resulting in dislocation of “people who used to provide 90% of the HDP votes in those cities”.

It is also possible that some voter segments who might not normally vote for either the AKP or the HDP, such as the Huda-Par movement, voted for the AKP(‘s proposals) in this election.

Finally, the security presences in the region has arguably increased in the past two years, and so part of the increase in AKP votes could be due to “security personnel who cast their ballots in situ.”

First of all, it is clear that the voting dynamics of the East are very different from that in the rest of the country. Similar to my previous post, the below graph plot histograms of the YES share vote distribution in 2017 against the AKP vote share distribution in November 2015, all at the ballot box level.


Figure 2

Whereas in the part of Turkey outside the (C/N/S) Eastern Anatolia region the rightwards push in the mass appears to be rather smooth, in the East the main changes have occurred at the edges of the distribution. The prevalence of the Never-AKPers, who vote close-to or never for the AKP, that have existed going back at least the last four elections, have disappeared and have instead been replaced by a significant share of Always-AKPers, who tend to vote homogeneously for the AKP.

Yet, as the below graph shows this doesn’t seem to be block votes moving from one edge of the distribution to the other, as the disappearance of the Never-AKPers (Diyarbakir, Hakkari, Mardin, Mus, Sirnak, Tunceli, Van, Sirnak, and Van) is observed in different provinces than where the appearance of the Always-AKPers occur (Erzurum and Sanliurfa).

Figure 3

This graph shows the difference in the frequency of the YES vote share  in the referendum and the AKP vote share in November 2015. At the bottom of each panel, with a compressed scale, is the frequency of the 2017 YES vote share.

Block voting could still occur, it just doesn’t seem offer an explanation for the transfer of mass between the edges of the distribution, as these seem to be driven by different processes in different provinces

As for the Always-AKPers, in Sanliurfa this could potentially have to do with increases in security forces there. But has there been similar increases in security forces in Erzurum too? (Not to my knowledge). Also, it is likely that security forces have become more prevalent in the other southeastern provinces too, but this phenomenon does not show up there. In particularly Sanliurfa, the magnitude of these Always-AKPers is large. Boxes with between 99-100 % or more voting for the AKP increased by 246 boxes, and given the province’s average number of valid votes of 253 per box, this translates into 62,238 votes, or roughly 77 percent of the added votes the AKP gained compared to the previous election (80,453) in the province. In Erzurum, there were 98 more ballot boxes (compared to the previous election) with more than 99% AKP votes. Given an average number of valid votes per box of 192.7, this implies 18,885 votes affected, roughly 75% of the votes the AKP gained compared to the previous election (25,232 added votes).

The disappearance of the Never-AKPers affects a larger area in the East. The change in frequency for the previously mentioned provinces (Diyarbakir, Hakkari, Mardin, Mus, Sirnak, Tunceli, Van, Sirnak, and Van) with sharp drops in frequency for the 0-1% bin AKP vote share was 896 ballot boxes, and given an average votes per box of 212, this affects roughly 189,952 voters, or around  72% of the additional votes AKP gained (264,064) compared to the last election in these provinces.

One thing should be clear at this point: whatever factor is driving the Never-AKPers and the Always-AKPers, they have a significant bearing on the magnitude of the change in AKP/YES votes between the current and previous election.

Is the Disappearance of Never-AKPers driven by Conflict-Driven Displacement?

Whereas I have a hard time explaining the appearance of the Always-AKPers in two provinces, the broader disappearances of the Never-AKPers may have something to do with the conflict between the Turkish state and the PKK, particularly with regards to the significant internal displacement that has followed as a result.

It is arguably safe to assume that the neighborhoods in Eastern Turkey least likely to vote for the AKP are also those that have experienced the highest conflict intensity so far. For example, it should come as no surprise that most neighborhoods in Nusaybin and Cizre — where whole quarters of the cities have been razed — have tended to vote almost exclusively for the pro-Kurdish HDP (in November 2015 the vote shares were circa 9 and 5 % respectively).

Consequently, is there a link between initial political preferences (which are also the places that  and change in voter turnout?

The below graph plots the correlation between the total votes cast in 2017 at the neighborhood-level (muhtarlık) against the previous election’s AKP vote share. Each orange circle represents a local mean in 2-percent bins with local quadratic smoothers with associated 95% confidence intervals overlaid in blue, by super-region in columns and by election in rows. (I’ve labeled the regions of Central/North/South Eastern Turkey simply as “Eastern Turkey”.)

Figure 4

The relationship between the AKP vote share in the past election and the percentage change in total votes cast (between current and past election) was roughly similar in the East and non-East up until 2017. In 2017, the relationship looks quite different comparing the two regions. For the non-East, except for the bottom three percentiles (orange circles) of the AKP vote share, the relationship is roughly linear and negative. Yet in the East, it’s positive at lower levels of the previous election’s AKP vote share and flat at higher levels. Whereas, in the non-Eastern part of Turkey, higher initial-AKP-support areas saw larger turnout declines, in the Eastern part the lower initial-AKP-support areas saw larger declines, but only for neighborhoods where the AKP was getting a minority share of the vote. This non-linearity in the relationship in the East suggests there is some process that affected the low-AKP-voting areas in the East to a greater extent than higher-AKP-voting areas in that same region.

Many of the neighborhoods with both the lowest initial AKP support as well as the largest turnout declines appear to be among those who have also experienced large-scale fatalities, crippling 24-hour curfews, as well as material destruction of the urban space. The disappearance of the Never-AKPers, and the increase in the number of AKP votes in this area, thus appears intimately linked to the forces that drove people out of their homes, or whatever was left of them. The extent to which this is driven by forced out-migration, or the risk of future out-migration, is however beyond the scope of this post.


The Curious Case of the Vanishing Never-AKPers in Southeastern Turkey

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?

These Never-AKPers tend to be, with some exceptions of course, voters in the predominantly Kurdish southeastern region of Turkey. And so, in these areas, there ought to have been a vote swing from those previously voting for the Kurdish party HDP in past elections to a YES vote for the AKP government’s proposed constitution (which the HDP has stringently opposed) in the current election. Interestingly, the AKP may have gained as much as 450,000 (preliminary and subject to possible change) votes in the Southeast compared to November 2015, corresponding to roughly 10% of votes of the entire region. Especially since turnout was significantly lower this time around, by my estimates by some 150,000 (preliminary and subject to possible change).

The degree to which there has been a vote swing in the southeastern areas to the AKP can be seen in the below boxplot graph, which plots the distribution of the district-level (ilce) log difference between the number of votes received by AKP (or the YES side) in any of the  June 2015 November 2015, or 2017 elections  and the number of AKP votes in the previous election.

The graph shows AKP losing votes in the June 2015 election, especially in the Southeastern region, to then experience a sharper-than-the rest swing back in November that same year. In the last 2017 referendum election, this region is again pushing further toward the AKP(‘s constitution), even as the rest of Turkey experiences somewhat of a reversion to the mean. For the vote swing in the Southeast in 2017, the distribution is also clearly skewed towards positive an AKP-swing, partly driven by a number of outliers.

Two things are odd with this: the first is that the Southeast region experiences a roughly equally large (if you care about the median) or a larger (if you care about the mean) pro-AKP vote swing than in November 2015, just as the rest of Turkey experiences the opposite. The second is that the distribution of the AKP vote swing is so much more skewed toward the AKP than any group in any of the previous elections. (In fact, there are only observations in the “outer fence” on top, which makes it stand out among the other distributions.)

There are no commensurate swings in either registered voters or votes cast, as shown in the middle- and right-hand graphs.

This doesn’t have to be anything peculiar or strange if the pro-AKP swing in the Southeast is coming from “latently” pro-AKP areas (areas that were already pro-AKP in the region. Nor does the above graph by itself offer any link to the first graph I showed of the disappeared Never-AKPers.)

However, it turns out this is not the case. In fact, the largest pro-AKP swings appear to come from the least pro-AKP areas in the region. This can be seen in the below graph, which plots, again, the log pro-AKP vote swing against the log number of AKP votes in the past election for both the Southeastern districts (red) and the rest of Turkey (blue).

Note two things about the left-hand and middle graphs: the largest swings away and back to the AKP occurs mostly among the mid-range of AKP voters in the southeast (in red), not the very least AKP-supporting districts. The districts that were already not at all voting for the AKP in the previous elections tend to remain this way. That is, until last Sunday. Here, almost the entire lower end of the distribution is experiencing some of the largest pro-AKP vote swings in the entire distribution. This was unexpected.

But it gets even more surprising when one investigates which provinces and districts are undergoing this political change of hearts. In November 2015, it was predominantly the provinces of Agri, Igdir, Mus, Sanliurfa, Van,  and even a few districts of Diyarbakir that swung the most toward the AKP. In 2017, however, the biggest pro-AKP changes occurred in Tunceli, Hakkari, Sirnak, and Diyarbakir. The three former provinces here are among the worst affected by the war between the PKK and the Turkish state.

It gets still a bit stranger once we look at which districts these are. The Diyarbakir district at the 1.5 point in the right-hand graphs is none other than Lice, which is as much a Never-AKPer as you’re ever going to find. Among the other districts topping the pro-AKP swing distribution are Cizre, Yuksekova, two heavily damaged districts from the military conflict, and Uludere, the scene of the infamous Roboski strike.

The voters in the strongholds of the HDP, where sympathies for the PKK and antipathy toward the Turkish state is among the highest in the country, have truly chosen a peculiar time to switch to the AKP. What could possibly explain how such strongholds of the pro-Kurdish movement would have waited until 2017 to start voting for the AKP?


One possibility could be that the composition of voters has significantly changed in some Kurdish areas since November 2015. For example, if there’s been an increased inflow of security personnel who register to vote in these areas, that could be one explanation.

At an aggregate level, there doesn’t seem to have been any larger positive changes in either votes cast, as can be seen below, but there does seem to have been an uptick in registered voters in 2017 compared to the previous election.

Most of the Kurdish areas saw decreases in votes cast, and no or small increases in registered voters. This is, in a way, also somewhat interesting. Thus, either any inflow of security personnel was too small to really effect the voting population, or the inflow of security personnel would have to have been evenly matched with the outflow of local residents fleeing the conflict. This would then beg the question of how this group of displaced citizens were given the change to vote elsewhere.

Is the Never-AKPer phenomenon by itself a sign of an irregularity?

Above, I make make no judgement as to whether the initial existence of the Never-AKPers is by itself a problem for the elections or not. Perhaps these are simply an outcome of very homogenous political preferences in some areas. Or perhaps not, in which case there’s a Devil’s Advocate kind of argument to be made:

(I am also grateful to a number of midnight-oil-burning DMs from Alex)

Alternatively, could a switch in allegiance of the local elites, effectively tribe leaders, explain the phenomenon? As I wrote just after the June 2015 elections, the HDP’s success was to a large extent brought about by it successfully gaining the support of non-core voters among more religiously conservative Kurds. At the time, there was some reporting on HDP’s specific campaigns to win over large tribal communities in Kurdish Turkey (see here, here, and here). This could have resulted in a vote swing of so-called “block votes” , whereby “persuading only the tribal chief, community leaders secure the backing of thousands of people.”  In the November 2015 elections, there was a significant reversal in this non-core group of Kurds, most likely as the government managed to regain their support.

But before getting into more details on this, let’s take another look at the data. In this case, I’ve split the data into two groups, one including ballot boxes belonging to provinces in Turkey’s Southeastern region, and the second one including all others. The below graph is the same as the first one above, but split along these two subgroups

Although there appears to be some Never-AKPers even outside of the Southeastern provinces, I will focus mostly on the Southeast where they tend to be much more common. The left-hand side panel shows that 10-15 percent of all ballot boxes had little or no AKP votes in the June and November elections, with a smaller share in 2011 and, as previously documented, very few such cases in 2017.

Which provinces are driving this result can be gleaned from the below graph which plots the same kind of plot but for each province (in the Southeast).

In several cases (Agri, Mus, Siirt, Sirnak, Van etc) the high densities at the very lowest part of the distribution seem to occur almost exclusively in the two 2015 elections. In the June 2015 elections, for example, Hakkari saw roughly half its ballot boxes having little or no AKP votes, and the corresponding percentages for Agri was 25%, 15% in Diyarbakir, 20% in Mardin, 40% in Mus and Sirnak, 60% in Tunceli, and 20% in Van.

If this phenomenon was due to pro-HDP irregularities, occurring under an absence of state control, one would expect these to disappear in the November 2015 election, as the state security forces reasserted control over the region. Although this seems to have occurred in some provinces like Agri, Bingol, and even to some extent Siirt, in most others it continued. Similarly, if the very high densities of low-AKP-voting areas were a result of of block voting via traditional tribes, then shouldn’t the high densities at the lower end of the distribution have disappeared entirely in November?

In 2017, however, the Never-AKP phenomenon is entirely gone. Curiously, both in Bingol, Sanliurfa, and even to a degree in Bitlis and Siirt, we now observe the opposite, high densities at the very top of the distribution. And so, under the assumption that Never-AKP ballot boxes were an outcome of some form of anti-AKP irregularities under state absence, this would then imply that the reassertion of state control lead to Always-AKPers, which ought to then constitute pro-AKP irregularities. But the incongruence between the timing of state reassertion of control and the Never-AKPers lingering on even afterwards suggest tribal block voting and irregularities with a pro-HDP bias are unlikely to be the main driver of why the Never-AKPers disappeared.

Overall, however, I suspect there are multiple factors at work. It would be naive to claim that there couldn’t have existed any irregularities benefiting the HDP in previous elections, nor that local elites and block votes don’t matter at all. And if the AKP has been able to stop or harness such practices, it would then be one more institutional feature of Turkey that it has successfully been able to turn to its advantage. (In particular, the distribution for Sanliurfa in 2017 looks rather concerning.) The Southeast remains its poorest region, and if there was any area in Turkey where electoral fraud would be most likely to go unnoticed, it would most likely be here.

Meanwhile, I don’t think one should underestimate factors I have not discussed very much here, such as the internally displacement of large numbers of the population or the possibility that local residents might simply vote for the constitution in the hope that the political instability would end. There is also the issue of why there are Never-AKPers outside of the Southeast. These factors, however, would require a post of their own.







Turkey’s purge is accentuating the gender bias in labor markets and increasing the skills gap

A central challenge for the Turkish economy in avoiding falling into a middle-income trap is to achievea shift away from (what the World Bank has labeled) a “know-who” toward a “know-how” economy. The latter implies not just investments in education but also putting a more educated workforce to productive use.

The recently released employment statistics from the Turkish statistical institute showed two worrisome features of how unemployment has evolved over the last year, beyond the headline 1.9 percentage point jump in the unemployment rate from 10.8% in 2015 to 12.7% in 2016.

1) The increase in the unemployment rate is to a significant degree driven by individuals with higher education degrees, a group that experienced a 3 percentage point increase in the unemployment rate (see graph below). In contrast, less educated brackets show more modest upticks in unemployment. Arguably, this is different from a standard economic downturn which would normally affect blue-collar workers to a similar, if not larger, extent than white-collar workers. (Interestingly,  there’s also a 4.2 percentage point increase among vocational high school graduates.)

2) The unemployment rate change between 2015 and 2015 among women is twice the magnitude to that for men: 2.8 versus 1.4 percentage points respectively.  Moreover, this is mostly driven by higher-educated women with at least a high-school degree, and changes in the unemployment rate varying between 3.8-4.8 percentage points.

Continue reading

Asset prices and Turkey’s revised GDP growth

A curious aspect of the new GDP series (which I’ve covered here, here, and here) coming out of Turkey is the divergence between GDP growth and the stock market. Whereas official GDP statistics shows higher growth during the 2010-2015 period than the preceding six years 2004-2009, the opposite is true for the Turkish stock market.

This can be seen from the data series published by Turkstat called “The Rates of Financial Investment Real Profits’, which tracks the real returns of not just an investment in the flagship BIST100 equity index of the Borsa Istanbul, but also investments in US dollars, the Euro, Gold, Government debt, as well as bank deposits.

The below graph shows real (inflation-adjusted) return indices for all of these asset classes between January 2004 to January 2017, with the base month set to January 2010.


Whereas in the per-2010 period, equities was the asset class that yielded the highest return (the blue line has the lowest starting point among the series and thus the highest increase up until the base month in January 2010), in the post-2010 period it had the lowest return (the blue line has the lowest end point of all the series). Continue reading

Will the Real Real GDP in Turkey please stand up?

In the previous two posts on Turkey’s revised GDP statistics I showed 1) how revised GDP growth numbers diverge from the older version after 2009, 2) the importance of construction investments,  and 3) how standard policy variables failed to explain the revised growth figures.

Especially the previous post used cross-sectional data to show that global relationships between policy variables failed to predict Turkish GDP growth. In this post I will take a more direct approach at trying to validate the new official GDP statistics using time-series variation in Turkey. To the extent that Turkey’s revised GDP numbers reflect variation in economic output, there ought to be alternative measures of real output varying in a similar fashion.

One option would be to compare Turkey’s GDP growth with that of a standard leading indicator like industrial production (IP) or retail sales, all three which are published by the Turkish Statistical Institute (Turkstat). Yet the point of the revision appears to have been to improve statistical collection capabilities by using expanded administrative record of firm activity and so on. In this case, any difference between such a series and GDP data could simply reflect differences between recently revised statistics and soon-to-be revised statistics. It is therefore preferable to seek measures of economic output that are less sensitive to changes in Turkstat’s statistical capacity.

The recent debate over the quality of GDP statistics in China is the starting point here. Leaked Wikileaks cables from 2007 showed the now premier of the state council suggesting three alternative measures of economic output instead of the official GDP data: electricity, rail freight volume, and bank loans. Below I make some adjustments to this in order to incorporate factors more specific to the Turkish growth model.

Continue reading

Is New Turkey’s Growth Model From Outer Space?

The previous post focused the extent to which Turkey’s revised GDP data changed the recent history of economic growth in Turkey. A particularly striking fact of the new series is how much higher the growth rate in GDP in Turkey has been ever since the global financial crisis in 2008/2009. Of interest is then also how this changes Turkey’s economic performance in a comparative sense internationally, both in terms of economic growth as well as key economic indicators.

In this blog post, I take as the basis of economic performance the change in real GDP per capita obtained from the most recent October 2016 World Economic Outlook (WEO) from the IMF. For most of it I will show how this measure of economic growth – in two periods of six-year-averages for 2004-2009 and 2010-2015 respectively – correlates with a selected number of key economic indicators, most of which are included in the WEO database, and how much of Turkey’s (and other countries’) growth can be explained by a relatively simple regression model including a number of indicators of interest.

These indicators are: the natural logarithm of the average GDP per capita during the preceding six-year period, the natural logarithm of the average of population size during the preceding six-year period, the average growth rate in GDP per capita during the preceding six-year period, the current account balance as a % of GDP, the CPI inflation rate, the investment rate as a % of GDP, government debt as a % of GDP, and the unemployment rate. In addition, I also draw on the World Development Indicators database from the World Bank for labor force participation rate, domestic credit  to the private sector as a % of GDP, the age-dependency ratio, the urbanization rate, and from the IMF’s Balance of Payments database I also add the net international investment position (NIIP) as a % of GDP. For the investment rate, private credit, urbanization, and the NIIP-to-GDP measures, I also include a change variable with each measured as the change between the average during one six-year period and the corresponding average of the preceding six-year period. The IMF WEO indicators are quite standard and hopefully require little introduction. Most of added variables from the WDI are a bit Turkey-specific as they will show Turkey’s comparatively low labor force participation rate, the rapid growth of private credit in the economy, and the change in urbanization serving as a proxy for factor that could drive some of the large construction investments apparent in the new revised GDP series for Turkey. Continue reading

Constructing growth in New Turkey

The Turkish Statistical Institute recently released a revision to its GDP series (here and here), with some noteworthy consequences. Not only did the new series produce an upward revision of the level of GDP by around 20 percent (for GDP in 2015), but equally striking is the upward revision in the real growth rate of GDP after 2009 by an average of 1.8 % per year. The quarterly data is plotted below for new and old GDP and GDP growth rates (year-on-year) respectively.

The new statistics revision, taken at face value, arguably boost “the president’s economic arguments”, putting Turkey’s economy in a kinder light than previously thought. The timing is auspicious, as the government will likely try to revamp the constitution during the coming year.

The changes to Turkey’s GDP and growth rates are very large ones not only from an absolute perspective, but especially so in comparison with other cases of ESA (European standards) or SNA (United Nations standards) revisions. In OECD countries, such revisions have tended to have much smaller impacts on the levels of GDP and, at least on average, close-to zero effects on GDP growth. In most of these other cases, large revisions tend to be driven mostly by the changes in standards themselves, although in some of them, wider changes in how statistics are collected were more dominant. Continue reading