ComRes have released their first voting intention poll since the election, and have topline figures of CON 41%, LAB 29%, LDEM 8%, UKIP 10%, GRN 5%. Full details are here.
The ComRes poll also had the first attempt at a methodology change to address the failings of the polls at the 2015 general election – though as ComRes make clear in their explanation this is not ComRes’s final word on the topic, they are continuing their internal review and may make extra changes too.
As with all the pollsters who use political weighting, the initial change is to move from past vote weighting using the 2010 election to past vote weighting using the 2015 election, something that would have been done anyway. The second change is a new model of turnout weighting. This is based on the theory that a cause of the error was people overestimating their likelihood to vote in an uneven way – that is, we all know people overestimate their vote, but ComRes suggest they overestimate it unevenly, that people in some social groups (who happened to support Labour this time) overestimated their likelihood to vote more than other groups, thus skewing the polls.
In the past almost all the pollsters accounted for likelihood to vote using a straightforward system of asking people to rate their likelihood to vote on a scale of 0 to 10, and then either filtering out those people who gave a low score, weighting people according to how likely they said they were to vote, or a combination of the two. ComRes’s new method still filters out people who say they are less than 5/10 likely to vote, but after that bases likelihood to vote weighting on demographics, based upon patterns of turnout at the general election, specifically that there tends to be lower turnout in areas of social deprivation and in areas with a high proportion of social classes DE and low proportions of ABs.
The mechanics of this aren’t completely clear yet (I’ve asked ComRes for some more details which I’ll update later), but essentially it looks as if younger and more working class respondents are assumed to be less likely to vote than they claim they are and weighted downwards accordingly. It means, in effect, that the final headline voting intention figures are made up of 41% AB, 31% C1s, 19% C2, and just 9% DEs, so the effective sample once it’s modelled for the sort of people who actually turn out to vote is far more middle class than the pre-election samples that got it wrong.
The impact of the change is, as you might expect, to produce significantly more Conservative figures. In this particular poll it increased the Conservative lead from eight points to twelve points. In ComRes’s final pre-election poll it would have changed the result from a one point Tory lead to a five point Tory lead, significantly nearer what actually happened.
UPDATE: ComRes have got back to me with some more details of their turnout model. In my original version of this post I’d assumed ComRes were still weighting people according to their 0-10 score, but were adjusting this score based on demographics too. In fact ComRes are now only using the 0-10 score to filter out people who say they are less than 5/10 likely, otherwise the turnout weights are all based on demographics.