No Sure Things: Brazil’s 2018 Election

Three weeks ago, the 2018 Brazilian election seemed like anyone’s ballgame. Jair Bolsonaro, the far-right candidate of the Partido Social Liberal (PSL), was in first place, but it was anyone’s guess how being stabbed at a campaign event on September 6 would affect his numbers. Fernando Haddad of the PT had only been an official presidential candidate for a few days thanks to ex-President Lula’s insistence on running a campaign from his jail cell.

Since then, things have gotten both clearer and murkier. Fernando Haddad and Jair Bolsonaro currently hold double-digit leads over center-left Ciro Gomes of the Partido Democrático Trabalhista (Democratic Labor Partyand center-right Geraldo Alckmin of the Partido da Social Democracia Brasileira (Brazilian Social Democracy Party), meaning that they would both go to the second round on October 28. There, however, they are technically tied according to most polls, and anything could happen; second-round predictions  are hypothetical and volatile, and both candidates have rejection rates that are comfortably larger than their support rates.

On the one hand, Haddad could win by default, as the anti-Bolsonaro vote could crystallize into a pro-Haddad vote. On the other hand, Bolsonaro could win by default, as the anti-PT vote could crystallize into a pro-Bolsonaro vote.

Yet there are a few reasons why one should take the poll results currently in the news with a grain of salt. In fact, we could have a surprise in the first round on Sunday the 7th, never mind the second.

Polls have limited geographical reach 

Polls in Brazil are limited in terms of their sampling. Infrastructure is not up to scratch in every part of Brazil, and polling institutes cut some corners sometimes, often opting not to use face-to-face polling  or not using cell phones, so it ends up that not every part of Brazil actually makes its way into surveys.

To compensate, most major polling firms use quotas for certain demographic characteristics, not truly random samples. The specific formulas they use are generally not publicly available, but this is well-known among the scholarly and polling community. This has some knock-on effects, as I show below.

My own polling database–compiled with Mathieu Turgeon of the University of Brasília in a working paper that is currently under review–the average absolute polling error for presidential elections one day before the election has been 2.96 percentage points since 2002. In 2014, this rose to 4.86 percentage points. Compared to the U.S.’s average of 1.9%, the U.K.’s of 2.0%, Portugal’s of 2.4%, and France’s of 2.0%, this does not flatter Brazil’s polling.

The Supreme Electoral Court wiping out voters registrations

The Supreme Electoral Court (TSE) decided, this year, to require compulsory biometric registry for all voters. This required voters to go to the Electoral Court in their region and register their fingerprints. Unfortunately, as sometimes happens with major bureaucratic changes, the communication of this change was less than ideal, and 3.3 million voters (2.4% of the electorate) never did make this trip.

This left the TSE with a pickle: undermine the effort and let those who didn’t register vote anyway? Or, despite the fact that voting is compulsory, follow through and ban those who did not register from voting.

Unfortunately, the communication of this decision has also been less than ideal, as voters who have had their registration canceled have not been individually notified. In fact, they will only know that they can’t vote after they show up at the polls to try to vote. After all, if one didn’t find out about the initial decision to require fingerprints, it’s also likely that that person won’t know about the subsequent decision unless notified individually. To make matters worse, a rumor on Whatsapp has already circulated claiming that you can vote even without having done the registration if you show up with photo identification.

This could be a big deal. On a normative level, this registry cancellation is not randomly distributed among the population–it disproportionately will affect the poor and the uneducated, or in other words, those who most need political representation. On an empirical level, this could also make a difference; those with their registry canceled would likely not know, and would include themselves as voters to electoral polls (since everyone, in theory, is, unless they plan to justify not voting to the electoral board). If any candidate draws primarily on voters who are more likely to have not registered, this could cause them to underperform in relation to the polls. If this will hurt anyone, it will probably be Haddad because of his strength with poorer citizens.

The shy Bolsonaro vote

When Trump won, 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 is mixed, but social desirability bias has been shown to weigh on polling results for a variety of subjects, including sex, drug use, and vote buying. It is not outside the realm of possibility that it could have applied to a candidate who was shunned in certain circles for controversial comments about blacks, Latinos, women (etc.). As the theory goes, certain types of voter might be happy to share his support of Trump with interviewers, but others are not, and might refuse to acknowledge that to pollsters while still voting for him.

This dynamic could possibly apply to a candidate such as Bolsonaro, whose fondness for crossing lines would even embarrass Trump. This barrier might have broken down as he has gained more attention, but one can never count out the possibility that people in certain environments might not want to openly show support for someone who says all sorts of repugnant things.

The evidence is not quite clear yet, though, so this is one we’ll have to wait and see about.

Vote buying

In Brazil, the exchange of money for votes on the eve of the election is still a common practice. Candidates with more money than scruples (or who would object to that characterization, but are afraid of their competitors of that profile) will commonly go to the streets the night before the election or the day of the election to try to win over undecided voters.

This practice–commonly called the boca de urna (literally, the mouth of the voting machine)–generally entails illicit campaigning. Either the campaign will send associates door to door the night before the election, or it will send its associates into the street the day of the election to intercept voters heading to the polls. These associates (who are generally several degrees away from a candidate herself so as to avoid implicating her directly) find undecided voters in the street or pass by the houses of people they already know. They then hand over a small campaign advertisement with the candidate’s name and number, and those of the candidates she supports, along with a certain amount of money (last election, between R$50-R$100). This practice can also happen without money changing hands, but it is not as effective.

For obvious reasons, it is extremely difficult to measure how common the boca de urna is and how much of an effect it has, but I gave it a shot (as mentioned above). It is more common for proportional elections (federal and state deputy), but it can also have knock-on effects for majoritarian elections (like president) when people don’t have their own preferences for candidates. We find that, when at least 15% of the electorate is undecided the day before a given majoritarian election, polling error increases in conjunction with financial differences between campaigns (this year, the most recent polls are showing between 13% and 17% undecided). The true effect is likely much higher because campaigns almost always leave many resources undeclared to the TSE.

This type of practice requires, above all else, two things: money and political alliances. Big majoritarian campaigns can’t organize this alone because it requires local knowledge–where to go to buy votes without being caught, who to go to, etc. It also requires financial resources–enough to buy the proportional campaigns and local politicians (e.g., mayors and city councilmen) who do the buying, and for them to in turn buy brokers, and for those brokers, in turn to buy votes.

It is likely no coincidence, therefore, that deep-pocketed campaigns like the PSDB’s tend to grow towards the end of the campaign, whereas candidates such as Marina Silva, who have less money and are loathe to ally with other candidates, tend to fall away; Marina won, on average, 2.2% less than predicted in 2014 while the PSDB’s Aécio Neves won an average of 5.4% more. This dynamic looked most likely to help a candidate like the PSDB’s Geraldo Alckmin, and hurt Marina and Bolsonaro, but Bolsonaro’s recent endorsement from the wealthy rural caucus means that this might have shifted in a direction that will not hurt Bolsonaro, and even might help him.

A surprise in the works?

This does not mean that Haddad and Bolsonaro will not make the second round, but Brazil always throws up some surprises in elections. This year, they could be more pronounced. Don’t be surprised if one or more candidates receives a vote total significantly off the latest polling figures.

If we have 5% errors again, that could mean that Ciro Gomes takes Haddad’s place, or that Haddad takes first place. One candidate could limp into the second round demoralized, finding it difficult to negotiate alliances. Or, given the number of hidden factors that could help Bolsonaro, it might also mean that the errors go in the other direction.

In that case, a second round might not even be necessary.

A mágica da Copa

Tatuagens lamentáveisAmizades pouco prováveisBoladas na cara.

Tudo aquilo faz parte do maior espetáculo no planeta Terra–a Copa do Mundo. Todo país consegue participar, colado na TV, independentemente de ser um superpoder futebolístico pretendendo levar o troféu para casa (você não), ou um país cuja confederação de futebol contratou a desgraça de Bruce Arena.

Pois bem, há muitas pesquisas que mostram que os eleitores gostam de cobrar políticos pelo que acontece durante os seus mandatos, mas não necessariamente tomam cuidado para ter certeza que cobram os políticos certos pelas coisas que realmente são a culpa deles. Enfim, se você tem um cargo muito visível quando boas coisas acontecem, você é recompensado. Se você tem um cargo muito visível quando coisas ruins acontecem, você é punido. Não é justo, mas a política também não é (nem fair play). E é assim para os políticos com a Copa do Mundo também.

De qualquer modo, por causa do alinhamento dos calendários, toda eleição geral do Brasil desde 1994 tem coincidido com a Copa do Mundo. Portanto, faz sentido que, em um país como o Brasil, a Copa realmente poderia mudar os destinos de governantes (incumbents).

Por acaso, eu já tinha montado uma base de dados quase-pública sobre as pesquisas brasileiras de intenção de voto (afinal, não quero que vocês roubem os meus dados antes que o artigo seja publicado…) chamada o Brazilian Polling Error Database, por isso resolvi testar essa hipótese. O BPED reúne pesquisas de intenção de voto feitas entre 2002 e 2014 para eleições para presidente, governador, senador, e prefeito no Brasil. Montei para um artigo que escrevi com Mathieu Turgeon da Universidade de Brasília, o que atualmente está nas mãos dos pareceristas em uma revista acadêmica.

Utilizei uma técnica que se chama difference-in-differences, que controla por problemas de seleção com mais eficácia ao longo do tempo do que as regressões estatísticas normais. Isso porque cria um experimento natural no qual quatro grupos de objetos são comparados:

  1. um grupo que recebe um tratamento
  2. um que, durante esse mesmo período, não recebe esse tratamento
  3. um que já recebeu o tratamento antes que foi recebido pelo grupo de tratamento
  4. um que não recebeu o tratamento antes que foi recebido pelo grupo de tratamento

The idea of this empirical strategy is that if the two treated and the two nontreated groups are subject to the same time trends, and if the treatment has had no effect in the pre-treatment period, then an estimate of the “effect” of the treatment in a period in which it is known to have none, can be used to remove the effect of confounding factors to which a comparison of post-treatment outcomes of treated and nontreated may be subject to.

Neste caso, miro pesquisas de intenção de voto e separo as pesquisas em quatro grupos diferentes:

  1. As para governantes que foram feitas antes que a participação brasileira na Copa terminou
  2. As para governantes que foram feitas depois que a participação brasilera na Copa terminou
  3. As para candidatos não-governantes que foram feitas antes que a participação brasileira na Copa terminou
  4. As para candidatos não-governantes que foram feitas depois que a partipação brasileira na Copa terminou

Para as Copas entre 2002 e 2014, as participações do Brasil terminaram nos seguintes dias:

2002: 30 de junho (ganhou)

2006: 1 de julho (quartas)

2010: 2 de julho (quartas)

2014: 8 de julho (semifinais, de um jeito bastante memorável

Deixei fora eleições para prefeito fora porque eleições municipais são realizadas dois anos após eleições gerais. Também deixei fora as eleições para senador em 2002 e 2010 porque os eleitores tiveram que escolher dois candidatos nesses anos e teria sido um pesadelo a analisar. Para deixar tudo mais simples, também deixei fora covariáveis possíveis ao nível de pesquisa (instituto, númeo de eleitores indecisos, etc.) e ao nível macro (estado da economia, etc.).

Conforme era esperado, ser titular melhorou a situação do político depois da Copa de 2002 quando o Cafú ganhou o seu segundo Mundial. A variável chave aqui é o estimador de difference-in-differences–só por ser titular, se ganha 15 pontos de porcentagem a mais. Valeu, Ronaldo Fenômeno!

2002: Rumo ao Penta!

Variáveis Coeficiente (Erro padrão) valor-t (valor-p)
Pós-2002 -1,05(0,82) -1,27(0,20)
Incumbent 13,85(3,84) 3,61(0,00)
Estimador Difference-in-Differences  14,02(4,48) 3,13(0,00)
N 2887

2002Incumbents

Não parecia ser muito importante em 2006, e em 2010, ser incumbent teve um efeito positivo de novo–pode ser que, dado o elenco de 2010, os brasileiros estavam relativamente satisfeitos com uma saída nas quartas? (mais provável que alguma outra variável macro fosse determinante…)

2006: Nem cheirou nem fedeu

Variáveis Coeficiente (Erro Padrão) valor-t (valor-p)
Pós-2006 -1,53(0,69) -2,22(0,03)
Incumbent 40,94(1,20) 34,02(0,00)
Estimador Difference-in-Differences  -2,28(1,75) -1,31(0,19)
N 4210

2006Incumbents

2010: Aí vem o Felipe Melo…

Variáveis Coeficiente (Erro padrão) valor-t (valor-p)
Pós-2010 -2,40(0,82) -1,27(0,20)
Incumbent 20,20(1,54) 13,10(0,00)
Estimador Difference-in-Differences  14,23(2,13) 6,68(0,00)
N 5864

2010Incumbents

Ser governante, porém, piora a sua situação depois da obra-prima de inestabilidade emocional de David Luiz em 2014. Nesse caso, você acaba sendo punido por 10 pontos.

2014: Goool da Alemanha

Variáveis Coeficiente (Erro padrão) Valor-t (valor-p)
Pós-2014 -0,23(0,68) -0,33(0,74)
Incumbent 16,62(1,93) 8,59(0,00)
Estimador Difference-in-Differences  -9,27(2,15) 4,31(0,00)
N 3971

2002Incumbents

Então, o que isso quer dizer para 2018? Nem tanto para os que acompanham a presidência; apesar do presidente atual do Brasil ser capaz de nunca morrer (se não for antigido pelo sol), a taxa de aprovação dele mal pode diminuir e ele não disputará a reeleição. Se você for um senador ou governador titular, porém, melhor começar a torcer com muita emoção que venha o Hexa.

The Magic of the Cup

Balls in the face, politics, and stats

Regrettable tattoos. Unlikely friendships. Balls in the face.

All of these are features of the greatest show on Earth–the World Cup. Every country gets to take part, staying glued to the TV, whether you’re a world soccer superpower looking to bring the trophy home (no, not you), or whether your soccer federation hired Bruce Arena.

Now, there’s a lot of research out there showing that voters like to hold politicians accountable for things that happen on their watch, but aren’t necessarily diligent about making sure they hold the right politicians accountable for the things that are actually their fault. In short, if you hold a highly visible office when good things happen, you tend to get rewarded. If you hold a highly visible office when bad things happen, you tend to get punished. It’s not fair, but neither is politics (or fair play rules).

This means that, in Brazil, the World Cup can be a political event, in the sense, at least, that the everything is about the World Cup while it’s going on. Feel like stealing a car? Do it during a game of the Seleção, as the streets will be deserted. Feel like getting some medical attention? Try not to schedule it during the World Cup (I’ve personally been there–a suspiciously high number of doctors seemed to be “in surgery” and none of the receptionists even pretended to do work).

As a result, it stands to reason that, in a country like Brazil, the World Cup could really affect incumbent politicians’ political fortunes. Or not–maybe no one pays attention to politics during 3-4 months before an election anyhow. In any case, because of how the calendars align, every general election in Brazil since 1994 has coincided with a World Cup.

I coincidentally have an almost-public dataset on Brazilian polls (can’t have y’all stealing my precious data before it’s published, after all) called the Brazilian Polling Error Database, so I decided to run the data to see. The BPED is a collection of polls conducted between 2002 and 2014 for presidential, gubernatorial, senatorial, and mayoral elections in Brazil. I created it for a paper written with Mathieu Turgeon of the University of Brasília, which is currently under review at an academic journal.

I used a technique called a difference-in-differences, which controls for selection problems over time more effectively than normal statistical regressions  by creating a natural experiment in which four groups of objects are compared:

  1. one group that receives a treatment
  2. one that, over the same time period, does not receive that treatment
  3. one that receives the treatment before it is given to the group that does
  4. one that does not receive the treatment before the treatment is administered

The idea of this empirical strategy is that if the two treated and the two nontreated groups are subject to the same time trends, and if the treatment has had no effect in the pre-treatment period, then an estimate of the “effect” of the treatment in a period in which it is known to have none, can be used to remove the effect of confounding factors to which a comparison of post-treatment outcomes of treated and nontreated may be subject to.

In this case, I look at election polls and separate predictions into four different groups:

  1. Those for incumbent candidates before Brazil’s World Cup participation ended
  2. Those for incumbents after Brazil’s World Cup participation ended
  3. Those for non-incumbents before Brazil’s World Cup participation ended
  4. Those for non-incumbents after Brazil’s World Cup participation ended

For the 2002-2014 World Cups, Brazil’s participation ended on the following days:

2002: June 30 (winners)

2006: July 1 (quarterfinals)

2010: July 2 (quarterfinals)

2014: July 8 (semifinals, but not in a good way)

I left mayoral elections out because municipal elections are held in off-years. Senate elections in 2002 and 2010 are also left out because voters have to vote for two candidates for the Senate and that’s a nightmare to analyze. To keep things simple, I left out covariates at both the poll level (polling institute, number of undecided voters, etc.) and the macro level (state of the economy, consumer confidence, etc.).

As one might expect, being an incumbent becomes better after the 2002 World Cup when Cafú gets number 2. The key here is the difference-in-differences estimator–by being an incumbent, you’re essentially getting a boost of 15 percentage points! Thanks Ronaldo Fenômeno!

2002: Penta Glory!

Variables Coefficient (Std. Error) t-value (p-value)
Post-2002 -1.05(0.82) -1.27(0.20)
Incumbent 13.85(3.84) 3.61(0.00)
Difference-in-Differences Estimator 14.02(4.48) 3.13(0.00)
Number of observations 2887

2002Incumbents

It doesn’t seem to matter too much in 2006, and in 2010, being an incumbent does have a positive effect again–maybe, given the side they put out, Brazil was relatively satisfied with a quarterfinal exit? (More likely that there was some other macro-variable at work, of course…)

2006: Meh

Variables Coefficient (Std. Error) t-value (p-value)
Post-2006 -1.53(0.69) -2.22(0.03)
Incumbent 40.94(1.20) 34.02(0.00)
Difference-in-Differences Estimator -2.28(1.75) -1.31(0.19)
Number of observations 4210

 

2006Incumbents

2010: The Felipe Melo Experience

Variables Coefficient (Std. Error) t-value (p-value)
Post-2010 -2.40(0.82) -1.27(0.20)
Incumbent 20.20(1.54) 13.10(0.00)
Difference-in-Differences Estimator 14.23(2.13) 6.68(0.00)
Number of observations 5864

2010Incumbents

Being an incumbent, though, becomes much worse after David Luiz’s 2014 masterclass in emotional imbalance. In this case, you’re getting essentially a 10-point hit.

2014: Tears Everywhere

Variables Coefficient (Std. Error) t-value (p-value)
Post-2014 -0.23(0.68) -0.33(0.74)
Incumbent 16.62(1.93) 8.59(0.00)
Difference-in-Differences Estimator -9.27(2.15) 4.31(0.00)
Number of observations 3971

2002Incumbents

So what does this mean for 2018? Not too much for those following the presidency; although Brazil’s current President might never die (unless he gets hit by direct sunlight), his approval rating can’t really get any lower anyway and he won’t run for re-election. If you’re an incumbent senator or governor, though, you might want to cheer real hard for the Hexa.