Election polling has taken a thrashing on both sides of the Atlantic in the wake of what many perceive to be a succession of catastrophic failures. So what has gone wrong?
In Britain, polling problems surfaced at the 2015 General Election when many were expecting an outcome that put a Conservative majority beyond reach. Election night itself was a wince-fest for pollsters as the Conservatives headed steadily towards a House of Commons majority, their prospects—like those of Donald Trump—having been written off by numerous pundits and pollsters alike.
This herding of pundit expectation was part of the problem. Not one ComRes poll put Labour ahead of the Conservatives in the months leading up to the 2015 Election, and our final predictive poll release was headlined ‘Conservatives set to win popular vote’.
To be sure, we underestimated the size of the gap between the two leading parties, but our average error (the difference between actual and predicted vote shares) was 1.3%. Hardly catastrophic.
One main reason for our success at an Election in which many other pollsters were confounded was our Turnout Model. Since the 2010 Election we had been looking carefully at which demographic groups consistently over-estimated their propensity to vote, in order to be able to increase our accuracy.
That Model worked fine in 2015, but in the 2016 EU Referendum it actually increased our predictive error because some of those demographic groups, particularly erstwhile inactive older people, surprised us by actually voting. These are people who often reported that they had not bothered voting in decades—and sometimes never.
The EU Referendum campaign was much more hit-and-miss because as an industry we had no antecedents on which to base our assumptions.
In the event, according to ComRes calculations, of the 170 or so campaign polls, 59% had a Remain lead, 33% had a Leave lead, and 8% had a dead heat. During the course of the campaign it is possible that ‘real’ opinion swung oscillated a lot but it was very clear that people overwhelmingly chose to believe the polls that confirmed their own desires.
This is the irony of life in an information age. We live in an era of near-perfect information access, and yet—like newspaper readership—people tend to access news sources that amplify their own prejudices and reinforce their own desired outcomes. Social media makes it all the worse, echoing incorrect analysis across cyberspace and whipping up a frenzy of misplaced expectation. People tend to believe those polls and analysis that suit their own wishes.
The Trump victory in November should have been much less of a surprise than the EU Referendum result because the national polling was very close to reality. Hillary Clinton’s lead nationally was 2.1% while Real Clear Politics puts the average final poll lead at 3.3%, making the variance a paltry 1.2 percentage points. But, again, alarm bells should ring when it becomes fashionable to expect one outcome on the basis of chattering class desire rather than dispassionate evidence.
Much has been made of the link, if there is one, between the Leave vote and the Trump victory. Without leaping to too many conclusions, there is at least some read-across. This is evidenced by polling by Lord Ashcroft (percentages below), who polled Britain on EU Referendum Day and the US shortly before the Presidential election:
Q. Do you think of each of the following as a force for good, a force for ill, or a mixed blessing?
Broadly speaking, there appears to be a good degree of overlap between conservative views on the factors tested by Lord Ashcroft, and propensity to vote both Leave and Republican. And this, in turn, provides us with some justification for the claim that a sizeable proportion of people in both the UK and US wanted to shout their dissatisfaction with the prevailing liberal consensus.
The assumption that it is always seen as ‘good’ to be a feminist, a liberal, an advocate of green politics and globalisation, is simply untrue. As for immigration, the schism in societal views becomes even more acute and something which no politician can afford to ignore.
The main challenge for pollsters will come in 2020, when the UK is next due to have a General Election, and the US holds its next Presidential election. The only valid way to model predictive election results is by adjusting your assumptions according to the most recent outcome. In the UK, does this mean the 2015 General Election, or does it mean the 2016 EU Referendum? It is impossible to know whether the people who surprised us by voting in June will do so again in 2020.
In the current political climate, glibly hoping that the past is a good indication of the future is about as certain as playing pin the tail on the donkey. If we are not careful, the electorate will make asses of us all.
Andrew Hawkins is chairman of ComRes, a UK-based global research consultancy specialising in reputation, public policy and communications. Reach him on Twitter: @Andrew_ComRes