How did the polls get Trump so wrong?
FROM THE BREXIT REFERENDUM TO THE VICTORY OF FRANÇOIS FILLON IN THE FRENCH CENTER-RIGHT PRIMARY (HE HAD BEEN THIRD IN THE RACE BUT WON 44% OF THE VOTES, LEADING BY 20 POINTS), THIS IS FAST BECOMING A REAL PROBLEM.
In the weeks leading up to the US election, polls showed Hillary Clinton enjoying a wide lead over Trump, both in the Electoral College and for the popular vote. However, come November 8th, people around the world saw how a leading candidate could lose so unexpectedly on the electoral map and how easily Democratic-leaning states could flip. Secretary Clinton may have won the popular vote, but the polls were still hugely wide of the mark. Polling companies are scrambling to see how they could have gotten it so wrong. The American Association for Public Opinion Research (AAPOR), America’s biggest pollster association, is pulling up a committee to evaluate their errors. They have already come up with some answers.
They are first suspecting a non-response bias towards Trump supporters. A non-response bias is a polling phenomenon where an entire part of the population does not respond to pollsters when called. This results in an inaccurate sample of voters. Usually, pollsters assume that there is an equal proportion of non-response for both candidates. However, it seems that, as Trump scored high with less educated people (51% among those voters, higher than both McCain and Romney) and as the latter are less likely to respond, a large part of the Trump electorate was undervalued. Pollsters could have revaluated their poling samples but as Obama won those populations with a high margin (more than 50% of people with a high school degree or less) pollsters did not see the bias in favour of Clinton.
The second possible bias is the fact that many Trump supporters may have lied to pollsters. Indeed, as much of the media had started since June 2015 a public ‘Trump-shaming’, many might have preferred simply not to admit that they were going to vote for Trump. This bias is almost impossible to adjust, as we have no way whatsoever, other than the actual results, to evaluate the extent of the bias.
The last possible mistake, and possibly the most significant, is the underestimated and overestimated turnout for different populations. As we saw on Election Day, the overall turnout was globally low with only 58.4% of Americans taking the time to vote compared to 62.2% in 2008. This was even more flagrant in key states, like the very blue Wisconsin, where Clinton led by more than 5 points in the polls, but lost by one point on Election Day. Wisconsin turnout went from 70% in 2012 to 66% this year, the gap being especially present with young voters. They represented 23% of the voting population in 2012 and, even though it is one of the fastest growing age population, makes now only 17% of this year’s voters. This low turnout is difficult to foresee for pollsters when people that say they will probably vote for Clinton, do not vote at all, creating an uncertainty.
Pollsters are now questioning their methods of polling as around the world many other example of voting results being very different from previous polls can be seen. From the Brexit referendum to the victory of François Fillon in the French center-right primary (he had been third in the race but won 44% of the votes, leading by 20 points), this is fast becoming a real problem.
Featured Image Credit- Wikimedia Commons