November 17, 2022
This is the fifth in a series of blog posts addressing a report by Diego Escobari and Gary Hoover covering the 2019 presidential election in Bolivia. Their conclusions do not hold up to scrutiny, as we observe in our report Nickels Before Dimes. Here, we expand upon various claims and conclusions that Escobari and Hoover make in their paper. Links to other posts: part one, part two, part three, part four, part six, part seven, part eight, and part nine.
In the previous post, we observed the unusual attention drawn to the interruption of the preliminary count. Again, the TREP was not official; its results were never intended to determine the outcome of the election. The TREP was not even necessary, and there has been no preliminary count since 2019.
Originally, the TSE planned to stop publishing TREP results at “8pm on election day with about 80 per cent of results.” It was the European Union Election Expert Mission (EEM) that pressured the TSE to agree to a second reporting “at midnight with around 90 percent of the results.” The EEM’s reasoning was that “if the results were close, 80 percent is insufficient to give an accurate picture and could be misleading.”
Clearly, the EEM understood the potential problem of counting nickels before dimes.
Again, the TSE did present preliminary results with more than 80 percent of polling stations included, but stopped the TREP shortly thereafter — clearly unnerved by the “maximum alert” from EHC (see the previous post).
No matter one’s opinion of the decision to stop the TREP, there are only two possibilities. One is that the partial results were representative, and therefore nobody should care that the TREP did not proceed to completion. Of course, this means that the results at the excluded polling stations were strange. The other possibility is that the partial results were not representative, and therefore nobody should have been surprised that the results at the excluded polling stations differed from those included in the TSE announcement.
The evidence for the latter is overwhelming. For example, polling stations at precincts where voters had opposed the 2016 referendum were significantly overrepresented in the partial preliminary count. Nickels before dimes.
Consider two divisions of the electoral data. First, we may categorize polling stations based on whether they were included in the TSE announcement (“early polling stations”) or were not (“late polling stations”). This is the division made by Escobari and Hoover. A second categorization is by whether or not entire precincts were entirely included/excluded “early/late precincts” or “split precincts” with some, but not all, polling stations included in the announcement. Combined, these two categorizations create four groups of polling stations.
In Table 1, we start to see how biased the announcement really was. Table 1 shows the four groups of polling stations, the number of eligible voters in each group, and the margin in favor of the 2016 referendum. Voters at late polling stations — and especially those from late precincts — favored the 2016 referendum significantly more than those at early polling stations.
Note that though the early and late polling stations from split precincts represent exactly the same precincts, we observe a 10 percentage point increase in support when comparing late (+5.72) to early (-4.33). Within each of these split precincts, the support for the referendum is unchanged: the calculations in Table 1 are based on precinct-level data. However, precincts favoring the referendum were underrepresented in the TSE announcement. To see how this works, consider the following example:
Split precinct A voted against the 2016 referendum by 10 percentage points, while split precinct B voted in favor by 10 percentage points. In 2019, each precinct had 400 valid votes split evenly across four polling stations for 100 votes at each of eight polling stations. The referendum was a tie at the two precincts, combined.
Within either precinct, the voters across all polling stations are assumed to be identical. This means we say the referendum lost 55-45 at each polling station from precinct A, and won 55-45 at each polling station from precinct B. That is, there was a 10-vote margin at every precinct, either in favor or against. In 2019, three polling stations from precinct A were counted early, compared to only one from precinct B. This means that even though the overall vote at these precincts ended in a tie, the early polling stations in 2019 opposed the referendum by 20 votes out of 400 — a margin of 5 percentage points. Likewise, the late polling stations — from the same two precincts — favored the referendum by 5 percentage points.
Importantly, this has nothing to do with which polling stations were selected from each precinct; the polling stations within each precinct are assumed identical. Rather, it has only to do with how many polling stations were selected from each precinct.
We made similar calculations for the split precincts in Table 1. The only reason for the increase in margin from early to late polling stations is the disproportionate number of voters in early polling stations that represented precincts opposed to the referendum.
Inasmuch as the voters at late stations in split precincts were on balance more favorable to the referendum than the voters at early stations in those precincts, the voters at late precincts (again, those not at all included in the announcement) were on average much much more in favor of the referendum. It is simply not true that the TSE announcement represented the electorate on the whole.
But why were the early polling stations not representative? As we discussed in the previous post, unrepresentativeness could be completely benign. Reasons may include delayed reporting due to unequal cellular coverage, elevated dependence on the electoral notaries, or increased frequency of observaciones/conflicting transcriptions due to juror errors.
In this case, the fact that these polling stations happened to favor Morales more strongly is — from a forecasting standpoint — merely an unfortunate correlation. It could be malicious but have no impact on the final result. We saw this in the United States in 2020, with certain states expressly delaying the counting of mail-in ballots that were known to favor Democratic Party candidates. Donald Trump seized on the resulting change in support over the vote count to make frivolous allegations of widespread fraud.
On the other hand, it could be consistent with actual fraud. For example, Morales-heavy polling stations could have been delayed deliberately for the purposes of executing fraud. This is, broadly, one theory espoused by Escobari and Hoover. One major difficulty in trying to make a case for this theory is that there are so many confounding factors, including but not restricted to benign considerations such as disproportionate cellular coverage, or increased dependence on electoral notaries.
The point is that fraud may make sense as, at best, a partial explanation for the unrepresentativeness of the partial results; and to that extent only after every possible benign explanation has been identified and accounted for.
Another, related, question is why were the official results at the late polling stations more favorable to Morales? Again, we must identify and account for every possible benign factor before falling back on fraud as an explanation. But we must likewise account for any malicious ordering of the count as well, even if there is no actual fraud in the results. Suffice it to say, distinguishing between nickels before dimes and actual fraud requires either working hard with extensive data or putting a lot of faith into some very strong assumptions. Escobari and Hoover chose the latter.
This is absolutely critical to keep in mind to understand Escobari and Hoover when they separate polling stations based on whether or not they were included in the TSE announcement. They assert that “fraud is more likely to have occurred” in their “treatment group” of late polling stations, but there is no factual basis for the latter claim. The OAS audit did not point to any fraudulent results entering into the official results.
On the other hand, we know that the later polling stations are more likely to be in the “treatment group” if they were delayed for ordinary reasons such as reduced cellular coverage, reliance on notaries, and the presence of observaciones on — or conflicting transcriptions of — the acta. Likewise, earlier polling stations are more likely to be in the “control group” if they had fewer votes to count.
If they find that the late-reporting polling stations showed elevated support for Morales, they are hard-pressed to say that the late reporting caused the elevated support as opposed to any other cause or causes for delay — benign or malicious — inducing bias in the selection of polling stations included in the TSE announcement. Nickels before dimes.