It wasn’t supposed to be a walkover, but it just about was one. Jair Bolsonaro, the far-right candidate of the PSL (Social Liberal Party), fell just short of an outright majority in the first round of Brazil’s elections on October 7, winning 46 percent of all valid votes (that is, excluding spoiled and blank ballots). Fernando Haddad of the PT (Workers’ Party) also made October 28’s runoff, finishing a distant second with 29 percent. According to poll aggregator PollingData.com.br, Bolsonaro is a clear favorite, with a 98 percent chance of winning the second round and an advantage of between 6 and 14 percent in the polls.
Polls significantly undershot both Bolsonaro and Haddad’s numbers, with the biggest polling firms putting Bolsonaro around 40 to 41 percent and Haddad around 24 to 25 percent. I had warned before the election that one should take polling figures in Brazil with a grain of salt. So what went wrong? And could these problems happen in other places, too?
Sampling difficulties for Brazilian polls
Polls in Brazil suffer from problems that also afflict other countries, but these problems can be accentuated because of limitations with Brazilian polls. The Brazilian Polling Error Database (BPED) — compiled with Mathieu Turgeon of the University of Western Ontario for a working paper currently under review — shows that the average absolute error for polls conducted just one day before the 2014 presidential election was 4.86 percent (most other countries average between 2 to 2.5 percent).
One reason is that it can be difficult to access certain areas, which make face-to-face surveys more difficult. However, according to Neale El-Dash, the statistician behind PollingData, they are still often better than telephone-based surveys, which have generally been conducted with lists of phone numbers bought from companies without regard to their origins or possible biases.
To compensate, El-Dash says that “almost all surveys before this year were with quotas,” with some even conducted in public places with passersby. These quotas tend to be primarily for sex and age. This could have underestimated the effect of evangelical voters, who are, ceteris paribus, more likely to be poor and live in difficult-to-access areas. Evangelicals also voted en masse for Bolsonaro.
Social desirability bias
When Donald Trump was elected US president in 2016, outperforming poll predictions, some political scientists and commentators proposed that many Trump supporters could have been reluctant to share their preference for him to pollsters. The evidence has been mixed, but the unwillingness to publicly say socially undesirable things has been shown to affect survey responses on a variety of topics, including sex, drug use, and vote buying. Some voters might refuse to acknowledge support of controversial politicians to pollsters while still voting for them.
This looks to have applied to Bolsonaro, whose fondness for saying repugnant things is no secret. A reliance on face-to-face interviews could have produced an especially pronounced social desirability bias with Bolsonaro, a particularly controversial candidate; telling an interviewer to their face that one supports Bolsonaro could be more difficult than doing so on the phone.
The US was also not the only country with two unpopular leading candidates. Both Bolsonaro and Haddad had sky-high rejection rates going into the election, both of them eclipsing 40 percent. Much like Trump captured an anti-Hillary vote and vice versa, both Haddad and Bolsonaro attracted votes by using the imminent threat of their opposite number.
As polls began to indicate that Bolsonaro and Haddad were the two leading candidates, this could have led to last-minute shifts among those who opposed one of them. A certain subset of voters looks to not “waste” their vote on a candidate with no chance of winning — and this is especially salient when voters want to “stop” certain unpopular candidates. This can lead to two candidates pulling away from others, even in multi-round elections, as happened here.
Strategic shifts in elite support
In the US, we often talk about a coattails effect: The support for one high-profile candidate (typically president) can affect the support of candidates running for other positions. In Brazil, this effect exists, too, but it also exists in reverse: Local politicians use their networks to support majoritarian candidates.
Picking a candidate who ends up losing gains nothing for elites; picking one who wins can give them a job and influence. Something that passed under the radar this year, but was possibly very influential, was the fact that the National Congress of Brazil’s rural caucus abandoned the PSDB’s Geraldo Alckmin for Bolsonaro five days before the election. Alckmin duly underperformed expectations by 3 to 4 percent while Bolsonaro outperformed them.
This kind of shift might not have as big an effect as it did in the 1980s, but it is not irrelevant. While campaigning is illegal 24 hours before the election, it still exists — and sometimes with offers of cash as well. This sort of practice rarely influences voters who have already made up their minds, but it still could affect elections when people don’t have preferences for that race.
This time around, 10 to 16 percent either did not respond or did not choose a candidate in the last polls before the first round. Pre-election polling would not have been able to capture this type of movement or catch if it would systematically help (or harm) one candidate in particular.
In short, many of the same problems that have plagued polling in other countries are also present in Brazil. Yet they are likely accentuated by certain factors that are more specific to Brazil. Pollsters elsewhere — particularly in other developing countries — will have to be attuned to these potential pitfalls, or else they could repeat the same mistakes.