Tuesday, April 28, 2015

Predicting Turkey's New Parliament

In my last post, I offered a basic predictive model of the AKP vote share in the upcoming parliamentary elections. Based on a limited model (albeit one with a pretty strong fit), the results suggest that AKP should obtain between 42.5% to 43.2% of the vote. Taking the average, 42.85%, suggests that AKP will see about a 7 percentage point decline in its vote share relative to the 2011 parliamentary elections.

The next step is to turn this vote prediction into a seat share prediction. After all, the election outcome is not really in doubt. The truly important question in these upcoming elections is whether AKP will hit its target of 330 seats (or, less likely, 367) that would allow it to propose a series of constitutional amendments for popular referendum in order to create its desired 'presidential' system.

The answer to that question hinges on several factors besides AKP's vote share:
  • Will Turkey's fourth party, HDP, surpass the 10% election threshold or not?
  • How will the vote for the opposition parties be distributed? (i.e., which parties will gain the most from AKP's losses?)
  • Will AKP actually perform as poorly as this predictive model suggests?
There are two important facts about Turkey's electoral system, both of which are more likely to benefit AKP than the opposition parties (it should be pointed out that both of these factors predate AKP coming to power; while AKP obviously have no incentive to change these rules that benefit it, they cannot be accused of creating these rules to benefit themselves):
  • The unusually high 10% national election threshold. If HDP stays below 10%, then AKP will be the party to gain most of the seats that it fails to win in Southeastern Turkey.
  • The prevalence of many small electoral districts (with district magnitudes of 5 or less), which tend to produce disproportional results. AKP tend to be strongest in these smaller districts, allowing it to win a higher proportion of the seats than its vote share would normally allow.
Projecting the New Parliament
Projecting the composition of the new parliament requires estimating the distribution of seats across all 85 electoral districts in Turkey. Importantly, it requires making some assumptions about the distribution of votes across Turkey's diverse electoral landscape. The first, crucial, assumption is whether HDP get 10% of the vote or not. This fact alone will change AKP's seat share by 40-50. Second, how will the 7 percentage point vote loss for AKP distribute to opposition parties? Will it be evenly distributed, will certain parties benefit, or will there be specific regional patterns?

I will generate a few different predictions using various assumptions.

The first is what I term a "uniform swing" model. In this model, I assume that vote gains among the opposition parties are shared uniformly--between the three parties (i.e., each party will gain one-third of AKP's lost votes) and across all districts (i.e., AKP will lose the same percentage of votes across all districts). To put it differently, AKP is expected to lose about 14% of its vote relative to 2011 (i.e., 7 percentage points of nearly 50% that it received in 2011). This model assumes that each party will gain an equal share of that AKP vote loss. This model puts HDP perilously close to the 10% threshold, so I construct a scenario where HDP is below 10% and where HDP is above 10%. I call these, respectively, the Uniform Swing, HDP<10% and Uniform Swing, HDP>10% models,

The second scenario I consider relaxes one assumption of the uniform swing model--that the three opposition parties will gain the same of lost AKP votes everywhere. Turkey generally has a regional pattern of party conflict with only AKP being a truly 'national' party: CHP are strongest in the western and coastal provinces, MHP are strongest in central and northeast Anatolia, and HDP are strongest in the southeast. Given this, it does not make sense to assume that all 3 opposition parties will gain equally throughout the country. Why would CHP, for example, gain one-third of AKP's lost vote in the southeast, where it has virtually no support? Instead, it is more likely that the most popular opposition party in each district (i.e., HDP in the southeast) would gain the most votes. In this "local swing" model, I allocate half of the gain in votes to the largest opposition party in that district and split the remaining half between the other two opposition parties. This should produce a scenario that is slightly less friendly to AKP (because its largest rival in each district will gain the most). I call this the Local Swings model.

I create a fourth model that uses an entirely different approach. My predictive model is based on the economy, but it could be wrong. One could point out that Turkish voters have recently given us strong evidence about their support for different parties. Much of this election campaign will center on whether Turkish voters want to empower Erdogan and the AKP to create a 'presidential' system. We already have one reasonably useful set of evidence about Turkish voters' preferences towards Erdogan: last summer's presidential election. In this model, I use the district-level results from the presidential election to predict this year's parliamentary election results. Because Ekmeleddin Ihsanoglu ran as a joint CHP/MHP candidate, I apportion his vote within each district according to the relative share of the vote received by each party during the 2011 parliamentary elections. In other words, if CHP received 60% of the combined CHP/MHP in 2011, then it receives 60% of Ihsanoglu's vote here. (This gets slightly more complicated in the three largest cities, which have multiple districts, but the approach is basically the same). As we will see, this approach produces different results. Given that Erdogan received about 51% of the vote in the presidential elections, this should produce a stronger prediction for AKP. I call this the Presidential Vote model.

Finally, I consider the possibility that the main beneficiaries of AKP vote loss will be MHP and HDP. Returning to the economic-based prediction that AKP will get about 42.8%, this model assumes that CHP will not gain votes anywhere compared to its performance in 2011. Instead, the vote gains will be split between MHP and HDP on a 75/25 ratio, depending on the district. In districts where MHP was the 2nd party in 2011, it gains 75% of those lost AKP votes. In addition, it also gains 75% of the lost votes in districts where HDP gained less than 1% of the vote in 2011. In districts where CHP was the 2nd party and HDP got more than 1%, then HDP gains 75% of those lost AKP votes. I call this the MHP/HDP Surge model.

Here are the predictions:

Uniform Swing, HDP<10% 322 158 70 0
Uniform Swing, HDP>10% 283 148 68 51
Local Swings 274 155 70 51
Presidential Vote 321 129 46 54
MHP/HDP Surge 271 136 82 61

There are a few notable points from the predictions:
  • The most favorable scenarios for AKP (the HDP<10%, and presidential vote models) each show AKP still falling short of the 330 seats necessary to trigger a constitutional referendum.
  • The least favorable scenarios (Local Swings and MHP/HDP Surge) show AKP falling just shy of a parliamentary majority (by 2 and 5 seats, respectively).
  • The Presidential Vote model shows HDP outperforming MHP, which is quite a surprise!
Most importantly, these results show that AKP will comfortably win (no great surprise there!), but that it is likely to see its share of parliamentary seats decline from 2011. Taking the presidential vote results as an outlier, these results also highlight that the most important question in this election will be whether HDP get 10% or not. If they do, then AKP will end up with less than 300 seats, and possibly less than a majority. If HDP fall short of 10%, then AKP have a chance of 330 seats (though this seems unlikely).

This fact creates an interesting dilemma for CHP supporters. Given that CHP will gain a fairly safe vote and seat total, would it be wiser to vote HDP in order to ensure that the party gain 10% and thus weaken AKP?

This fact also creates a clear prediction about the targets of possible electoral fraud by the governing AKP. We saw possible evidence of AKP electoral fraud in the closely-contested Ankara mayoral election. Here, electoral fraud will most likely be directed at HDP. This could occur in its areas of strongest support. But, given that the election threshold is national, it is also possible that AKP could seek to minimize HDP votes in the major cities and other parts of the country where it may be less susceptible to detection.

I may continue to do some further simulations as time permits. Comments welcome! Please tweet to me if you have suggestions!

Tuesday, April 7, 2015

Predicting Turkey's 2015 Parliamentary Elections

Last spring, I attempted a basic prediction model of Turkey's local elections, and it did pretty well. The model, including only a measure of GDP growth during the election year and the previous year, projected the governing Justice and Development Party (AKP) to obtain about 43.5% of the national vote. In reality, AKP got 43.9%. Not bad.

This time around, I am going to attempt something slightly more ambitious. In addition to predicting the AKP vote share, I want to predict the distribution of seats in the parliament based on some different possible scenarios.

In many respects, the real question of interest is not whether AKP will win the election. It is whether AKP will gain a large enough share of the seats to push through constitutional amendments to create President Recep Tayyip Erdogan's desired "presidential" system. With 330 votes in Turkey's 550-seat parliament, the AKP government could put these amendments to a national referendum. With 367 votes, they could simply pass the amendments without a referendum.

After predicting the AKP vote share, I'll use that to estimate how the seat totals in the parliament could turn out based on some different possible divisions of the vote.

Predicting the 2015 AKP Vote

For more background, read my post from last year. Predicting election results in Turkey is complicated by the instability of the party system over time. Pooling local and general elections, there are only five data points one can use. In addition, the ongoing consolidation of the party system during the 2003-7 period means that both AKP and the opposition CHP and MHP increased their votes during this time as they drew away the supporters of the defunct parties of the 1990s.

Here are the details of the model (it's very simple). I use the per capita annual GDP growth rate, taken from the World Development Indicators. I tried out some different measures. The one that fit the data best is simply an average of the election year growth rate and the previous year's growth rate. This makes sense, as economic voting is primarily retrospective, and elections often take place mid-year. In other words, the state of the economy during the past 12-18 months has the most significant effect on the AKP vote total. Anything before that is too far in the past to matter greatly, and the future can't really affect vote decisions at the time of the election. Note that this is a small change from last year's projection, where a model using the election year growth rate and a 50% weighted measure of the previous year's growth rate. The results are very similar with these or other possible combinations.

So here is a first model, including all elections between 2004 (the first local elections after AKP came to power) and 2014:

Based on this, the model would project a vote share of about 42.5% for AKP in this year's elections. That is based on last year's growth rate of 2.6% and this year's projected growth of 2.3%. If one has a different expectation about this year's economic growth, one can easily generate a different prediction accordingly.

The reader might notice that the line fits the (five) data points pretty well, except for one outlier. Interesting, that outlier is the 2004 local election, when AKP underperformed relative to what the model would predict. Why? This is likely due to the fact that AKP were still consolidating their electoral support in what was only their second election as a party. If we estimate the model without that election (reducing the available data to a ridiculously low four data points), it looks like this:

Now the line has a ridiculously strong fit, with an almost perfect R-Squared value. This changes the prediction slightly, producing a predicted vote of about 43.2% for AKP. If one averages these two predictions together, one gets a predicted vote of about 42.85%, which is roughly 7 percentage points less than what AKP got in the last parliamentary elections in 2011. One question that this leaves unanswered is which parties will pick up those lost votes. More on that later...

Up next: I will describe how I project the distribution of seats in parliament and show some different simulated results.