Sunday, December 17, 2017

Statistics and fraud

We now have two separate statistical analyses of Honduran presidential voting data, and both conclude the same thing: there was something fishy that happened in the middle of vote counting.

To recap: The Economist looked at the percentages of votes that went to the two main candidates before and after the long break by the TSE in posting vote counts.

Before the break, with about 57% of the vote counted, Salvador Nasralla had a 5% lead. After that, the lead steadily declined to the final apparent margin of just over 50,000 votes.

The Economist specifically asked whether the explanation offered by the Partido Nacional-- which claimed that the later votes came from more rural locations more likely to favor their candidate-- could account for the shift. They compared vote counts before and after the break in counting within each municipio.

Rather than being differences between urban and rural places, their analysis compared vote counts within each locality. They found an average shift of 3.8% within the same locality. It didn't matter if the municipio was rural or urban-- they all shifted the same way.

One known difference: votes tallied after the break included a large number that were not scanned and transmitted from the polling places on the day of the election. Instead, these were trucked up to Tegucigalpa and scanned there. Much of the discussion about vote counting has centered on the treatment of these vote tallies, including concerns about some arriving in open, unsecured packages, and the rumor that some were scanned in a hotel (and thus potentially could have had substitutes).

The Economist also drew attention to the unusually high voter turnout reported in the late-counted votes, in particular, from three largely rural departments. This, they note, could reflect a better get-out-the-vote operation-- or ballot box stuffing. Here it is worth remembering that the sign-off on vote tallies is done by credentialed party members, and there has been reported fraud and sale of credentials by smaller parties, in 2013 and 2017,

There matters stood until the release by the OAS today of a report by Georgetown University Professor Irfan Nooruddin. His analysis identifies a point when 68% of the votes were counted where, across different regions, both the turnout level and support for the Partido Nacional increases sharply. Either of these would be unusual; both are very unusual.

Nooruddin uses the reported data to do something that the actual vote counting never did: he simulates what vote counts would have looked like if results had posted randomly. This has been a key problem throughout the process: it is unclear what order the TSE used in its vote counting; it was not statistically random nor selected to be a representative sample. The OAS in its initial report noted that the TSE shifted from counting as votes came in to some unexplained selection process. Nooruddin helps us see if the election would have been less confusing if the voting tallies were counted randomly.

The conclusion of this part of the analysis is that if the votes were accurate, and were counted randomly, the pattern seen could have happened, and not result from tampering.

Nooruddin doesn't stop there-- as he notes, this part of the analysis is only worthwhile if the vote counts were accurate. He continues with tests of this assumption, and finds that the differences between early counted vote tallies and later ones "are large and suspicious".

Every department showed the same pattern of early lead for the Alianza followed by a change in pattern. As in the analysis by The Economist, the universality of this pattern is not easily explained by innocent factors. There is nothing about early vs. late vote tallies that would account for this.

It is as if there were two elections being counted, with precincts in every department changing the same way.

The only way we can imagine to have this result would be if for some reason the TSE did a preliminary sort of actas and deferred counting those most favorable to the National Party until last. Needless to say, that makes no sense.

Nooruddin points out that the shift in turnout in the later-counted tallies would be expected less than one in one thousand times-- statistically a significant difference. He presents an in-depth analysis of the Department of La Paz that shows that even in a department that favored Hernández throughout counting, and has a higher-than-average turnout rate, the later vote tallies increased in both reported turnout and voting for the Partido Nacional. The turnout increase is statistically likely to occur only one time in one thousand. Nooruddin concludes "such a sharp increase in turnout in the same department is unusual".

He writes that these findings are "consistent with a hypothesis of tampering with the vote tallies that were counted last".

So what could have happened?

One way to produce such an effect is good old fashioned ballot box stuffing-- reporting more votes than actually took place, and attributing the extra votes to a preferred candidate. Once the acta was signed, no one went back to double check the voter rolls or ballots. As long as the math on the tallies added up, you could have a voting pool in whatever form you like. This might well be correlated with places where votes weren't transmitted online the day of the election, as the ballot stuffing could happen at many points.

A lot of anxiety around these late vote tallies revolves around whether fake actas were substituted on the way to Tegucigalpa, or even fake images of actas in the TSE database. These, again, would work, and would not produce any contradiction unless the full ballot box was opened and recounted.

Both statistical analyses allow for the possibility this was just a really unusual way votes came in and were counted. But in Honduras, there is little trust in the system and unusual has already translated into illegitimate.

What happens next will determine whether Hondurans can begin to rebuild trust in democratic processes.

No comments: