One of the key bits of evidence on why the polls got it wrong has today popped into the public domain – the British Election Study face to face survey. The data itself is downloadable here if you have SPSS or Stata, and the BES team have written about it here and here. The BES has two elements – an online panel study, going back to the same people before, during and after the election campaign, and a post-election random face-to-face study, allowing comparison with similar samples going back to the 1964 BES. This is the latter part.
The f2f BES poll went into the field just after the election and fieldwork was conducted up until September (proper random face-to-face polls take a very long time). On the question of how people voted in the 2015 election the topline figures were CON 41%, LAB 33%, LDEM 7%, UKIP 11%, GRN 3%. These figures are, of course, still far from perfect – the Conservatives and Labour are both too high, UKIP too low, but the gap between Labour and Conservative – the problem that bedevilled all the pre-election polls, is much closer to reality.
This is a heavy pointer towards the make-up of samples having been a cause of the polling error. If the problems had been caused by people incorrectly reporting their voting intentions (“shy Tories”) or people saying they would when they did not then it is likely that exactly the same problems would have shown up in the British Election Study (indeed, given the interviewer effect those problems could have been worse). The difference between the BES f2f results and the pre-election polls suggests that the error is associated with the thing that makes the BES f2f so different from the pre-election polls – the way it is sampled.
As regular readers will know, most published opinion polls are not actually random. Most online polls are conducted using panels of volunteers, with respondents selected using demographic quotas to model the British public as closely as possible. Telephone polls are quasi-random, since they do at least select randomised numbers to call, but the fact that not everyone has a landline and that the overwhelming majority of people do not answer the call or agree to take part means the end results is not really close to a random sample. The British Election Study was a proper randomised study – it randomly picked consistencies, then addresses within in them, then a person at that address. The interviewer then repeatedly attempted to contact that specific person to take part (in a couple of cases up to 16 times!). The response rate was 56%.
Looking at Jon Mellon’s write up, this ties in well with the idea that polls were not including enough of the sort of people who don’t vote. One of the things that pollsters have flagged up in the investigations of what went wrong is that they found less of a gap in people’s reported likelihood of voting between young and old people than in the past, suggesting polls might no longer be correctly picking up the differential turnout between different social groups. The f2f BES poll did this far better. Another clue is in the comparison between whether people voted, and how difficult it was to get them to participate in the survey – amongst people who the BES managed to contact on their first attempt 77% said they had voted in the election, among those who took six or more goes only 74% voted. A small difference in the bigger scheme of things, but perhaps indicative.
This helps us diagnose the problem at the election – but it still leaves the question of how to solve it. I should pre-empt a couple of wrong conclusions that people will jump to. One is the idea polls should go back to face-to-face – this mixes up mode (whether a poll is done by phone, in person, or online) with sampling (how the people who take part in the poll are selected). The British Election Study poll appears to have got it right because of its sampling (because it was random), not because of its mode (because it was face-to-face). The two do not necessarily go hand-in-hand: when face-to-face polling used to be the norm in the 1980s it wasn’t done using random sampling, it was done using quota sampling. Rather than asking interviewers to contact a specific randomly selected person and to attempt contact time and again, interviewers were given a quota of, say, five middle-aged men, and any old middle-aged men would do.
That, of course, leads to the next obvious question of why don’t pollsters move to genuine random samples? The simple answers there are cost and time. I think most people in market research would agree a proper random sample like the BES is the ideal, but the cost is exponentially higher. This isn’t more expensive in the sense of “well, they should pay a bit if they want better results” type way – it’s more expensive as in a completely difference scale of expense, the difference between a couple of thousand and a couple of hundred thousand. No media outlet could ever justify the cost of a full scale random poll, it’s just not ever going to happen. It’s a shame, I for one would obviously be delighted were I to live in a world where people were willing to pay hundreds of thousands of pounds for polls, but such is life. Things like the BES only exist because of big funding grants from the ESRC (and at some elections that has need to be matched by grants from other charitable trusts).
The public opinion poll industry has always been about a finding a way of measuring public opinion that can combine accuracy with being affordable enough for people to actually buy and speedy enough to react to events, and whatever the solutions that emerge from the 2015 experience will have those same aims. Changing sampling techniques to make them resemble random sampling more could, of course, be one of the routes that companies look at. Or controlling their sampling and weighting in ways to better address shortcomings of the sampling. Or different ways of modelling turnout, like ComRes are looking at. Or something else yet unspeculated. Time will tell.
The other important bit of evidence we are still waiting for is the BES’s voter validation exercise (the large scale comparison of whether poll respondents’ claims on whether they voted or not actually match up against their individual records on the marked electoral register). That will help us understand a lot more about how well or badly the polls measured turnout, and how to predict individual respondents’ likelihood of voting.
Beyond that, the polling inquiry team have a meeting in January to announce their initial findings – we shall see what they come up with.