Most of the opinion pollsters in the US can claim that they were right in predicting a Biden win in the recent US elections. But perhaps only one pollster emerged from the election with any credibility. While the Trafalgar Group incorrectly predicted a Trump win, they did forecast the narrowness of the election, and identified the underlying reasons for lower polling for Trump, which many of the other pollsters were in denial about.
At the national level, Biden won about 3.5% more votes than Trump, while pollsters predicted a 7–10% advantage. The Electoral College determines the outcome, so swing states were more in the spotlight. Polls predicted a 10% win for Biden in Wisconsin (1% actual), 8% in Michigan (3% actual), and 3% in Florida (with Trump instead winning the state by 3%).
These errors are far beyond the margins of error for sampling. The Economist estimated a 97% chance of a Biden win based on all their modelling, but when less than 1% of the vote could have changed the outcome of the election, a flip of a coin would have been a better estimator of probability. Indeed, the betting markets (where hundreds of millions of dollars were wagered) were at times offering around 50:50 odds – this tells us something about how voters really felt, rather than what they were saying or not saying to pollsters.
The pollsters were also wrong in the 2016 US elections, and they blamed this error on a lack of representation of certain groups in their polls, e.g. non-college-educated voters. They were supposed to have corrected their sampling methods in time for the 2020 elections, but with errors of up to 9%, things are still fundamentally wrong with their polling.
In analysing the new errors in 2020, let us first look at the evidence. All the polling errors were again in one direction, i.e. underestimating the number of people who would vote for Trump. We still need to use research as a means for uncovering why this is the case. A survey conducted by the Cato Institute in 2020 showed that 62% of Americans say the political climate prevents them from saying things that others find offensive. While this attitude spans partisan lines, those with conservative opinions are far more likely to hide their opinions (77%) than liberals (42%).
In the run-up to the election, people were far more inclined to wear a Biden/Harris face mask than a Trump/Pence one (and this was not just down to the keenness for masks in the Biden camp). Showing support for the Republican Party could make you a social pariah, make you lose your job, get you assaulted, or worse; hence, keeping quiet about one’s voting intention can simply be the easiest (and safest) approach to take.
While pollsters are unlikely to assault you for taking part in a survey (quite hard to do over the phone, too), social acceptability bias means that a closet Trump voter could say to the pollster that they will not vote, claim they are undecided, or lie about their voting intentions. Even with online polling, bias creeps in. Online polls rely on people taking the initiative to take the survey. After the appalling first debate, support for Trump in the polls dropped, but analysts feel that this was a reflection of people simply not wanting to take part in the polls when their candidate was doing badly (‘sulking in silence’).
Exit polls are also very useful, as these are based on actual voters. While you may still get shy Trump voters at the polling stations, the 2020 exit polls reveal some startling insights – Trump actually increased his share of votes among Blacks and Latinos (both men and women) and increased his vote with Muslims and LGBT groups. Paradoxically, the only major group in which Trump lost a share of the vote was with white men, implying that the Republican Party needs to reach out to them more in the next election!
And finally, we need to use intuition. Many people feel cowed, talked down to, and bullied by activists, big tech, and media about their political beliefs. One way of safely hitting back against this authoritarianism is via the privacy of the voting booth. This is thought to have been one of the key factors in swinging the narrowly won Brexit vote in the UK in favour of leaving the European Union, when people were being vilified by the establishment for wanting to leave the EU, with accusations of bigotry and stupidity. Insults, such as the infamous “basket of deplorables” comment from Hillary Clinton about Trump voters in 2016, are not a great way to win their vote!
How can we improve polling accuracy going forward?
The so-called ‘culture wars’ between conservatism and the new forms of liberalism have intensified, particularly since 2016, which saw the Brexit vote and the election of Trump. This has accelerated recently with cancel culture (e.g. the de-platforming of conservative speakers), Extinction Rebellion, BLM, and the politicisation of race. The implications are that people will become even more reluctant to express their political beliefs and voting intentions.
Some of the approaches to address this have asked people not only about their voting intentions, but also who they think their neighbours will vote for, and who they think will win the election overall. The ‘blame a neighbour’ approach is sometimes used as a proxy for a person’s own voting intentions – that is, people who are similar to them (by socioeconomic grouping and location). Asking people to predict the winner is similar to asking people to bet, which, as stated earlier, is actually a more accurate ‘poll’ – something undertaken by hundreds of thousands of people, rather than the tens of thousands in the polls.
However, as the public wise up to these oblique questioning methods, they might still feel inhibited about telling the truth. Hence, we might come to rely more on correlations, examining the main concerns voters have alongside what the parties generally stand for. The exit polls in the US, for example, show that the key issues concerning those who voted for Trump were the economy and law and order. Some of the key issues concerning those voting for Biden were racism and climate change.
The list of voters’ concerns is long and complex, but this in itself can help us build predictive models for likely voting intentions, especially for binary choice elections like in the US. We might not even have to ask people how they would vote or even if they would vote at all.