Jon Mellon and Chris Prosser from the British Election Study team have written a new post and paper on the emerging evidence from the BES data on what went wrong with the polls. Last month they wrote a piece, which I covered here, on some of the potential causes of error they could use the BES data to look at. Now the BES post-election data is out they’ve done so, and come back with some findings.
Firstly, late swing – the BES data finds virtually no net change at all between how people said they would vote pre-election and how they reported having voted after the election. The BES team conclude from this that late swing is unlikely. We’ve now got published re-contact data from the British Election Study, ICM, Opinium, Populus and Survation, only Survation found any obvious evidence of late swing in their re-contact survey.
Secondly, Shy Tories. This is essentially the most difficult potential cause to evidence – if people lie before the election, and lie after the election and we can’t check their actual ballot papers, how do you detect it? You need to look for circumstantial evidence. The BES team have compared levels of Tory support in their polling in different types of area, on the assumption that if people feeling embarrassed to admit voting Tory really was a problem it would be less of an issue in heavily Tory areas than in areas where no one else voted Tory. They did not find this pattern. They also have some experimental data about question order and priming, one suggested solution to the polling error. The first three waves of the BES, conducted back in 2014, randomised where it asked the voting intention question – at the start, or later in the survey. Asking it later in the survey only made a minimal, non-significant change to the Tory vote (while it doesn’t say so in the article, the full paper also makes clear it doesn’t change the Labour vote either!)
Thirdly the BES team looked a bit at sampling, specifically around age, looking at an issue Opinium have already commented on at the BPC inquiry meeting. All pollsters weight by age using various age bands such as 18-24, 25-40, etc, etc. But are people evenly distributed within those bands? Within the top age group, for example, the BES found that people over the age of 80 were underrepresented and people in their early 60s were overrepresented. Whether that makes any different to the results they can’t yet say, as it may be countered by other things like political weighting.
Finally, and most importantly, they wrote about turnout and suggested that people may have been overestimating their likelihood to vote, and that the people who were actually less likely to vote were increasingly skewed towards Labour. Currently this pretty circumstantial evidence – we don’t know if people lied about voting in the general election, but there’s evidence to suggest they might have. For example, a small proportion of people said they had already voted by post before most of the ballot papers had even been sent out, in areas where there were not any local elections this May there was still a chunk of people who reported having voted in their local elections. At the moment, these people who look as if they might be lying disproportionately break to Labour, so would explain some of the error. In their article Jon and Chris instead try modelling people’s likelihood to vote based on their demographics and characteristics of their seat, and that increases the Tory lead by 1.8%. They conclude that turnout, people saying they’ll vote when they won’t, is a major factor behind the error, thought they conclude that it’s one pollsters can probably address quite easily through a better turnout model.
Two caveats though, before you think things are solved. One – at the moment we’re going on indirect evidence of people overstating their likelihood to vote. In the fullness of time the BES are going to do a validation exercise of their data (that is, checking respondents names against the marked electoral register) so they will be able to conclusively prove whether or not there are a significant proportion of people who told pollsters they voted, but lied about it. Secondly, Jon and Chris estimate that getting turnout wrong probably explains about a quarter of the difference between the final polls and the election result, which would still leave us another three-quarters to explain…
Meanwhile there have been two new Scottish polls in the last week – both show the SNP still on course for another landslide at the Holyrood elections next year.
Survation for the Scottish Daily Mail have Holyrood voting intentions of
Constituency: CON 14%, LAB 20%, LDEM 7%, SNP 56%, Others 4%
Regional: CON 12%, LAB 19%, LDEM 8%, SNP 45%, GRN 11%, UKIP 5%
TNS have Holyrood voting intentions of:
Constituency: CON 14%, LAB 20%, LDEM 5%, SNP 60%, Others 2%
Regional: CON 13%, LAB 21%, LDEM 5%, SNP 51%, GRN 7%