September 08, 2020
August 20 was a day that shook a small world of scientists that had all but given up — short of legal threats — on getting a glimpse into the data and methods behind the analyses that took down the Bolivian government. Slowed by stonewalling and gaslighting, researchers had managed to re-create some — but not all — of the results presented by the Organization of American States (OAS) in its case against the legitimacy of Evo Morales’s reelection last October 20.
The OAS had alleged, the day after the vote, that the preliminary count contained an “inexplicable change in trend” of the preliminary results — drastically skewing in Morales’s favor. But its claim was dubious to begin with. As early as October 22, we began raising serious questions suggesting the “inexplicable change” was quite predictable.
The OAS would later support its allegation by claiming that the official count also contained a late break for Morales that “cannot be easily explained away” by Morales’s generally rural support specifically because the official count “data do not reflect the time the results were reported to the TSE [Tribunal Supremo Electoral].” This premise is entirely wrong; votes from the main cities were much more likely to be counted early, because the official count required hand delivery (rather than electronic transmission) of electoral materials to TSE offices.
Faulty reasoning aside, the OAS results were irreproducible.
Ten months later came the revelation that some of the OAS’s previously baffling conclusions are explained by a coding error. It had ordered the time stamps on the tally sheets alphabetically rather than chronologically — thus destroying its narrative of a sudden change in the official count.
Unjustly Forced Out
The damage, of course, had already been done. On November 11, 2019, Bolivia’s president, Evo Morales — his term not yet complete — stepped down from the presidency amid allegations of fraud. Decisive was the report from the OAS, which had just presented its preliminary findings in a binding audit of the October 20 election. These findings were not favorable to Morales, questioning his official first-round victory.
Members of the opposition, some of whom had been saying all along that Morales and his Movement Toward Socialism (MAS-IPSP) party would attempt fraud to stay in power, took to the streets in violent protest. Though Morales agreed to annul the election, the head of the military told him he should step down. He resigned and embarked on a dramatic flight from Bolivia to Mexico — barely making it out alive, according to the Mexican foreign minister.
This isn’t the first time the OAS has deployed poor statistics to overturn election results. In 2010, the OAS intervened in Haiti, demanding that the third-place candidate be permitted to participate in a runoff election with the first-place candidate — leaving the runner-up, who just happened to be the only non-right-wing candidate, out in the cold.
To be sure, security surrounding Bolivia’s election was insufficient. At the risk of whataboutism, the same could be said of most any election in the United States. However, the OAS’s findings presented in November were otherwise full of insinuation and short on detail. The OAS presented no actual evidence that even a single vote was altered. Instead, it offered a statistical analysis that statistician Andrew Gelman characterizes as “a joke.”
Convenient Conclusions, Dodgy Assumptions
Even if the approach was dubious and the results were irreproducible, at least it presented something. The OAS press release the day after the vote had presented no such analysis — though it certainly did raise the volume of opposition protests. Amid cries of “fraud,” prominent members of Morales’s party and their families were assaulted or threatened with murder. Jeanine Áñez’s “interim” government — still in power today, having three times delayed new elections— would later cite the OAS reports as its near-exclusive evidence in its campaign to dismantle MAS.
Despite repeated requests, the OAS offered no justification for their claims. This, even though the results seemed in line with pre-election polls. With the preliminary count 84 percent complete, Morales had received more than 45 percent of the valid votes in a nine-way race. Yet under Bolivian electoral law, this would not be enough to win the race outright: only if he had an absolute majority, or a 10-point lead over runner-up Carlos Mesa, would Morales avoid a second-round runoff.
Morales’s lead increased steadily as the preliminary count had progressed. This offered no particular reason to think anything was odd. Rather, tally sheets from areas supportive of Morales have tended to be counted later than tally sheets from areas favorable to the opposition, and that was the case in this election. Imagine first counting votes from Los Angeles, New York City, Chicago, and Washington, DC, before considering rural Texas, Tennessee, or Alabama. Nobody would be surprised that the inclusion of Republican areas in the count would steadily chip away at the early Democratic lead.
Shortly before the announcement of partial, unofficial, preliminary results on the evening of October 20, there was a surge of votes counted from Santa Cruz — a hotbed of anti-Morales sentiment that dented the nearly constant good news for the incumbent president. In any case, the tally sheets that remained to be counted were coming from areas that had already shown, on balance, a strong preference for Morales. When the rapid count paused with 84 percent of votes counted, Morales’s lead over Mesa was only about 8 percentage points. Yet when the preliminary results were next reported, a day later, with another 9 percent of the vote, Morales had (tentatively) a 10-point lead — and his first-round victory.
Was this victory inexplicable — constructed in the darkness — as the OAS suggested? Was it a change of fate, or inevitable? My colleague Jake Johnston quickly put together an analysis of the results in the capital city of Cochabamba, where the swing was particularly visible. Breaking down the results by precinct, Jake showed that there was very little change in Morales’s support before and after the interruption of the count; in large part, Morales performed better, late, in Cochabamba, because precincts more favorable to the incumbent were — for whatever reason — counted later. A more rigorous approach over the entire election suggested by John Newman would show likewise: in the locality of Cochabamba, Morales actually underperformed late when the mix of precincts is taken into account.
So it went for the whole election. Consistently, studies have shown that once the different composition of precincts before and after the interruption is accounted for, Morales’s victory was predictable.
The OAS Final Audit
The OAS analyses have been notable for a steadfast refusal to consider such intra-geographic differences, presuming — contrary to all evidence — that a candidate’s support should be more or less uniform throughout the count.
The OAS final audit report on December 4 was explicitly biased in casting doubt on Morales’s first-round victory. The audit team expressly looked for irregularities on tally sheets that heavily favored Morales — justifying this selective search based on alleged statistical evidence that Morales’s victory was “inexplicable.” However, the statistical evidence presented in the OAS audit was not merely unconvincing; the analysis was completely wrong. The analysis boiled down to two points.
First, the OAS argued that the last 5 percent of the vote in the unofficial, preliminary count showed “a striking upward trend . . . that is quite different.” However, it is not unusual for late swings in a count to prove decisive. As noted in an open letter by economists and statisticians:
It is not uncommon for election results to be skewed by location, which means that results can change depending on when different areas’ votes get counted. No one argued that there was fraud in Louisiana’s 16 November gubernatorial election, when the Democratic candidate John Bel Edwards, pulled out a 2.6% point victory, after being behind all night, because he won 90% of the vote in Orleans county, which came in at the end of the count.
The order in which preliminary results were reported publicly was slightly different than the order in which Irfan Nooruddin (the author of the OAS statistical analyses) considers them. Many tally sheets throughout the count were transcribed but set aside for later approval. Those approved at the end tended to be much more representative of the entire election— even unfavorable to Morales. Regardless, the upward trend in Nooruddin’s data was predictable based on the earlier data, because his late tally sheets came from areas that had overwhelmingly shown strong support for Morales.
Second, Nooruddin argued that this finding is bolstered by the fact that the last 5 percent of the vote in the official count was also inconsistent with the previous 95 percent and again the penultimate 5 percent. Specifically, the OAS presented graphs showing that Morales’s vote share increased sharply after the 95 percent mark, while Mesa’s plummeted.
But these results were completely irreproducible. Until August 19.
For weeks following the release of the OAS report, researchers had tried unsuccessfully to replicate these findings using publicly available data. The OAS ignored requests for its data and explanation of its methods, but there were two red flags suggesting that its analysis suffered from serious errors. First, the OAS concluded its audit with the results of an “internal analysis” that directly contradicted the graphs. That table implied that Morales performed better — not worse — over the penultimate 5 percent than the last 5. Second, the table above it broke down results by select departments (equivalent to states in the United States). None of these numbers seemed to make sense.
This was not a simple difference between public and official, internal data. On May 25, I finally received data direct from the TSE; again, Cochabamba was counted entirely before the 95 percent mark.
The breakthrough came only on August 19. That was when Nooruddin posted his data set to a Harvard repository. Sorting tally sheets by this progress variable, everything looks fine to start with. But the transition from the October 20 to October 21 stands out.
Shortly before Nooruddin claims the count reaches 11 percent, the time stamps jump from 11:59 p.m. to 01:00 a.m. Was it possible there was an hour break? No. There were tally sheets time-stamped at 12:00 a.m. on the 21st, but they were far down the list — just past the 61 percent mark, immediately between those time-stamped 11:59 p.m. and the tally sheets time-stamped 12:00 p.m.
It was clear what Nooruddin had done. His time stamps were formatted as strings — letters and numbers; when he sorted his tally sheets, he did so alphabetically and not chronologically. Nooruddin had each day starting at 01:00 a.m., proceeding to 01:00 p.m., 01:01 a.m., and so forth until 12:59 a.m. and finally concluding at 12:59 p.m.
The OAS’s claims against the legitimacy of the election results originally centered on what it considered to be an inexplicable change in the trend of the votes over time. But Nooruddin’s analyses reflect no real-world understanding of the order in which tally sheets were counted. His “first 95 percent” included tally sheets as late as 10:55 p.m. on October 22, but his “last 5 percent” included tally sheets as early as 12:19 a.m. — hours earlier. Sorting tally sheets chronologically, Nooruddin’s progress variable is all over the place.
The fact is, the OAS still is not used to this kind of scrutiny, since it usually breezes through it with an air of authority. Like others empowered by this false sense of security, the OAS again doubled down on its defense of the study. On Twitter, Gerardo de Icaza — director of OAS election observations — praised Nooruddin as “one of the best electoral statisticians in the world” and insisted falsely that his results held.
Such has been the consistent pattern of the OAS in response to any criticism of their report: when data fails them, they simply ignore the evidence and lash out. When the New York Times arranged for several academics to evaluate the statistical evidence and published their finding that Nooruddin’s presentation was erroneous, secretary general of the OAS Luis Almagro dispatched a wild attack on the NYT. Others with influence inside the Washington, DC, beltway, such as the International Crisis Group and Human Rights Watch’s Americas director, José Miguel Vivanco, have taken notice of Almagro’s increasingly erratic behavior.
The OAS secretary general has caught even the attention of the Washington Office on Latin America, which has long been critical of many of Latin America’s left-leaning governments. Almagro recently undermined the independence of the Inter-American Commission on Human Rights — the OAS’s autonomous human rights arm — by unilaterally standing in the way of the continuation of Paulo Abrão’s term as executive secretary. The Commission has been critical of the de facto regime in Bolivia for its abuses — a regime which likely would not hold power but for the OAS’s intervention in the election.
It is clear the rot is thorough. This was not just a data slip. Not just an indefensible statistical analysis, officially delegitimizing an election. Not just an audit. This was not an objective, scientific investigation into the election, but a way of defending an indefensible analysis cooked up in advance. The OAS under Almagro is now visibly out of control. Its ostensible mission is to support the international order. It could start by dropping the United States’ two-century-old business of meddling in the Western Hemisphere.