2016 was not a disaster in terms of polling predictions if we take a closer look

Originally Published at Shout Out UK on Friday 25th September.

On Monday, Chris Whitty and Sir Patrick Vallance alarmed the country with warnings of rising Covid-19 cases. On Tuesday, Keir Starmer stressed (again) that he wasn’t Jeremy Corbyn and that Labour ‘love this country as you do’. That evening, the Prime Minister warned the public to follow new public health restrictions to avoid another national lockdown.

The above are examples of people overcompensating to ensure that past mistakes are not repeated. For the PM and his scientists, this means taking concrete action earlier rather than later — in contrast to March. For the Labour leader, it means highlighting certain values that many voters believed to be lacking during last December’s election.

A similar process is presently occurring when analysing the US Election. In 2016, pollsters correctly predicted an overall victory in the popular vote for Hillary Clinton. Yet concerning the electoral college, the polls were found wanting.


Of the thirteen key battleground states, Michigan, Pennsylvania, and Wisconsin were wrongly predicted for Clinton, and in Iowa and Ohio, Trump’s victory was larger than had been predicted: by 7 per cent and 4 per cent respectively.

These fine margins made a huge difference. Without Pennsylvania, Wisconsin, and Michigan, Trump would have fallen twelve electoral college votes short of victory.

Consequently, when discussing today’s numbers, voters and analysts alike, qualify their comments with the adage ‘if the polls are correct’.

Yet it would be unfair to conclude that after their performance in 2016, polling cannot be trusted again. Whilst I do not believe that analysis should be determined by who is up and down in the polls, such numbers do play an important role in understanding the election. More importantly, we should be confident that the mistakes of 2016 will not be repeated again.

Why we should trust the polls

The two main reasons for failure in 2016 are easily fixable.

Encouragingly, Pew Research Group has found ‘no systematic bias’ in American polling towards either of the two major parties. Rather, ‘most evidence exists’ for a ‘late swing in voter preference towards Trump’ and a ‘pervasive failure to adjust for overrepresentation of college graduates’.

Firstly, the late swing. In Michigan, Wisconsin, Pennsylvania, and Florida 11-15 per cent of voters said that they decided whom to vote for in the last week of the campaign. According to exit polls, these voters broke for Trump by nearly 30 points in Wisconsin, 17 in Pennsylvania and Florida, and by 11 in Michigan.

Although a late swing has occurred multiple times since 1948, when Gallup famously predicted Dewey winning the presidency weeks before Harry Truman secured victory, it is less likely today. A feature of the campaign so far has been how static support for Biden has been. In Five ThirtyEight’s Poll of Polls, Biden has sat above 50 per cent since August 10th, with his smallest lead being 3.4 per cent back in April. Also, unique to this election will be the reduced impact any late swing will have. Sixty-one per cent of voters want to cast their ballot before election day.

Secondly, the over-representation of college graduates that led to disproportionately high numbers for Clinton, is something easily rectifiable. For numerous elections, pollsters have adjusted their samples ‘to population benchmarks’ for education, and 2016 was no different. The problem however, was that voting was more strongly ‘related to education’ than had ‘historically been the case’.

In previous elections, a curvilinear pattern had occurred, between voting and education. Reflected in 2012 when both the least educated and the most educated voters supported Obama, whilst those in the middle split evenly.

In 2016 this was not the case. Trump’s 39 per cent lead amongst whites without a college degree was the largest of any candidate since 1980. Whilst polls also underestimated Clinton’s lead with the non-white college-educated population.

Pollsters did not expect the dividing line of education to be as large or as significant as it was. In West Virginia, with the largest white non-college-educated share of all states, Trump’s margin of victory was 14.3 per cent higher than the polls predicted. In Wisconsin, this was 6.1 per cent.

There is little evidence for the ‘Shy Trump’ vote

Something not as fixable would be if voters were shy of informing pollsters that they supported the President. This is an idea based on what happened in the California gubernatorial election of 1982 when Democrat Tom Bradley lost to Republican George Deukmejian, despite having been ahead in the polls. This was supposedly caused by voters being reluctant to inform pollsters that they were not going to vote for a black candidate.

Yet there is little indication of the same tendency in 2016. One of the pieces of evidence to support this claim is polling by the Morning Chronicle which found, ‘while the overall results completing the survey online didn’t boost Trump significantly, he did run stronger among more-educated voters’. For example, in phone interviews, likely voters ‘with a college degree said they supported Clinton by a 21-point margin but online, that margin shrunk to just 7 points’.

This survey however was conducted eight months before the election, and in Pew Research Centre’s 2017 study, ‘there was no significant difference by mode of interview on any of four questions asked directly about Trump’.

The public’s mistrust of voting forecasts

A study conducted by Westwood, Messing, and Lelkes, has found that in the event of election forecasts announcing ‘higher win probabilities, there is a decrease in voting’ for that party or candidate. Whilst they found no evidence for such an impact if a candidate’s vote share decreased in numerical polls. The biggest impact was made by forecasts such as those currently run by the New York Times or the Economist, which report the chances of a candidate’s victory.

The study reported: ‘those who stated that they expect one candidate to win by quite a bit are about two and a half percent less likely to vote than those who believe a race is to be close’.

This is significant when one considers the forecasting that was made in 2016. Clinton’s victory was reported to be between 70 per cent and 99 per cent certain.

Compounding the detrimental effect of these reassuring predictions for the Democrats is the detail that such forecasts were usually predominantly discussed on the more liberal media channels. As many as 30 per cent of Democratic voters expected Clinton to win by a ‘comfortable margin’ — the highest proportion in any US election this century.

Clinton lost by 0.7 per cent in Pennsylvania, 0.2 per cent in Michigan, 0.8 per cent in Wisconsin, and 1.2 per cent in Florida. The study concludes that the aforementioned effects ‘are large enough to meaningfully alter turnout in marginal states’.

The mistakes of 2016 also remind us of a very important fact: polls have been wrong many times before. Five ThirtyEight have found that ‘the average error in all polls conducted in the late stage of campaigns since 1998 is about 6 percentage points’. Such renewed scepticism will mean candidates are less likely to suffer from their voters not turning out because they believe victory is secured.

Not all the polls were wrong

There are essentially three different forms of polling (quantitative, qualitative, and polling based on ‘fundamentals’) — and some performed better than others in 2016. The closest final prediction, for instance, came from Alan Abramowtiz’s ‘Time for Change’ model, which assesses the national vote from a range of fundamentals. Namely, a candidate’s approval rating and the state of the economy. Only in two of the last five elections, has the electoral college swung the opposite way.

Whilst there is cause for more confidence, we shouldn’t be complacent. Problems still remain, particularly concerning state polling.

State and district-level races often see relatively little polling, and many polls differ widely in how they are conducted. A recent report found ‘little evidence’ that supposedly ‘gold-standard polls’, were ‘innately better performing’ than others.

Nevertheless, next time you mention what the polls say, I’d do so with a little extra confidence.