• Gábor Tóka , Raluca Toma

We started this investigation with two main questions:

  • Is there anything seemingly aberrant about the temporary votes cast in the 2014 election, something that might point to ballot-stuffing?
  • Does the data support the idea that either of the two main candidates benefitted, on a massive scale, from the temporary vote, via fraud?

The first searches failed to turn up any suspicious patterns.


Recap:

  • In the first round, Ponta seemed to do no better or worse in precincts with a greater proportion of supplementary list voters. In the second round, he did somewhat worse in such precincts nationwide.
  • Iohannis, by contrast, tended to perform better in the polling stations with a greater share of temporary voters — in both rounds.
  • Still, this is hardly an obvious sign of fraud, or lack thereof. So far we have seen no evidence that temporary list voters behaved differently from permanent list voters.

There is a better, multi-dimensional way to look at things, though. Taking off from the last test, where we looked at the evolution of voting patterns between the first round and the second round, we can further investigate the matter through a multi-variate regression based on the first and the second round turnout and the performance of all candidates.

In this model, we use the number of votes Iohannis or Ponta obtained in the second round as our dependent variable (meaning the outcome that changes depending on various factors). The changes in turnout on the temporary and permanent lists will constitute two distinct independent variables (meaning the factors whose impact on the dependent variable the model measures). This will allow us to better see what their individual contribution is. Additional independent variables are the number of votes for the candidates who participated in the first round — not just Iohannis and Ponta, but also people like Monica Macovei, Elena Udrea, Calin Popescu-Tariceanu et al.  We build an additive model that predicts the votes in the second round in each of the 18,553 polling stations on the basis of how much votes various candidates received at these places two weeks earlier.

By introducing this array of independent variables, we should be able to see how much each of them contributed to the second round final results. We will be particularly interested in the effect of temporary list turnout.

The Iohannis vote

The results of the first regression, whose dependent variable is the Iohannis vote in the second round, are illustrated in the next graph. The vertical axis represents the number of votes Iohannis got in each precinct in the November 16 runoff. The horizontal axis represents the effect of the independent variables.

figure-8-en

Taken together, the variables have an extremely good predicting power for the results of the vote, as the concentration of dots along the diagonal line shows.

Although the model is deliberately simplistic in that it assumes that first round votes are converted into second round votes in exactly the same way all over the country, it still explains 98% of all variation across the polling stations in the second round Iohannis votes.

The effect of the change in temporary list turnout is pretty large. However, so is the impact of any change in permanent list turnout. The model estimates that for every 100 new persons showing up to vote on the supplementary list in the second round– that is, persons who did not vote on the supplementary list in the first round — 78 of them voted for Iohannis. Conversely, for every 100 people who voted on the supplementary list in the first round but did not do so in the second round, Iohannis lost 78 votes.  The same logic goes for turnout on the permanent list: 100 extra people on November 16 meant 73 extra votes; 100 less people, 73 fewer votes.

As for the other independent variables, the model estimates that in a typical precinct, all first-round Iohannis, Udrea and Macovei voters went to Iohannis in the second round. By contrast, only about 22% of people who voted for Tariceanu then supported Iohannis in the runoff. The remaining estimates can be read from Table 1 and the figure above. All seem to be consistent with the behavior we might have expected a priori as well as the results of exit polls. [1]

Table 1: The predictive model underlying Figure 8

tab1

You may have noticed a few names of counties on the graph. Those highlight some of the regression’s „outliers” — places that didn’t constitute a very good fit for the model. We will address their potential significance in the next section. First, though, let us turn to the regression for the Ponta vote.

The Ponta vote

This model, too, comes up with a pretty accurate prediction of the second round vote.

As far as the turnout on the temporary and permanent lists goes, its estimated impact is much smaller than it was in the case of Iohannis. The model would indicate that for every extra 100 persons who voted on the supplementary list in the second round, 25 of them opted for Ponta. For every 100 persons who did not turn up for the second round, Ponta lost 25 votes. The difference between the effect of any changes in turnout between the permanent and temporary lists is not significant.

Overall, the model estimates that almost 70% of those who voted for Tariceanu in the first round opted for Ponta in the runoff. None of the people who voted for Macovei or Udrea preferred Ponta in the second round, according to this rough estimate.

figure-9-en

Table 2 below lists all the variables introduced into the model and their corresponding coefficients.

Table 2: The predictive model underlying Figure 9

tab21

Before we turn to the matter of the outliers, let us assess what this regression adds to our investigation of the main questions. Based on polls taken both before the election and on the day of the runoff, we already knew that most, if not all, Macovei voters would go to Iohannis. We also would have expected most, but not all, Tariceanu voters to opt for Ponta, whom Tariceanu endorsed. Our normal expectations about the voting behavior of supporters of other candidates were also by and large confirmed. Did we find out anything new about the supplementary list vote, though?

According to our model, any changes in turnout between the two rounds, be they on the supplementary or permanent lists, had a much greater effect on the Iohannis vote than on the Ponta vote.

Crucially, however, in terms of how they affected each candidate’s vote share, there were hardly any differences between the impact of turnout on the supplementary list and that of turnout on the permanent list.

What would a suspicious situation have looked like? If we saw hints that the new voters from the temporary list divided their vote very differently between Iohannis and Ponta, compared to the behavior of the new voters on the permanent list, then that might indeed raise some questions. Coupled with suspicions about fraud being much more prevalent among temporary voters, such a finding might allow for the possibility that any irregularities benefitted one candidate over another. However, what we now see is that even if all new temporary votes were fraudulent, the way they split between Ponta and Iohannis is pretty much the same as in the case of the new permanent list voters.

In the model for Iohannis, there is a five, and in the model for Ponta, a one percent estimated difference between the votes of supplementary list and permanent list voters. Since the two figures should, in reality, mirror each other, our best guess – obtained by averaging one and five – is that Ponta may have got around three percent more (and Iohannis three percent less) of the new temporary list votes than the new permanent votes in the second round. A trivial difference like this may find an explanation in any of the presumably many, though so far largely unknown, differences between the social and political profile of temporary list and other voters that exist in Romania.

In other words, again, there is no sign of a smoking gun that the temporary list vote is vastly different from the permanent list vote. According to our findings based on the data from all precincts, both groups split roughly the same way in terms of who they voted for in the second round.

Read The Romanian riddle of the supplementary list vote

Read Part 1. Dissecting the supplementary list vote. Suspicions versus facts

Read Part 3. Surprising results. Red flags or red herrings?

Read Part 4. Bonus feature: tell me what you want to prove and I tell you what to look at

Read Conclusions. Ballot stuffing – from suspicion to prevention


[1] You may wonder (1) why we collapsed the voters of some candidates together, as e.g. Funar with Tudor, and (2) how is it that some coefficients are exactly 0 and 1 while none are negative or larger than one. The reason for (1) is that there seemed little reason to distinguish here between these candidates while collapsing them greatly simplifies the presentation. The reason for (2) is that we explicitly constrained the model estimates so that none can be higher than one and none can be lower than zero. Without these constraints, our models would for instance suggest that either candidate in the second round got somewhat more than 100% of his votes from the first round, which is of course impossible but presumably accurately reflects that in their strongholds both candidates got more of the votes from all sorts of voters than our unified estimate for the whole country suggests. As we will see later, these artificial estimates are however useful for identifying precincts where something „out of the ordinary” took place with regard to the temporary votes and may deserve closer scrutiny.