FAQ: Likelihood to Vote
Turnout at the last few elections has been only around 60% – so 4 out of 10 people, whatever they told pollsters, didn’t actually bother to vote (theoretically at least, actually the real percentage of people able to vote who do is higher than the quoted turnout. Quoted turnout figures are of all those people on the electoral register, some of whom will actually be dead or incapacitated).
If supporters of different parties have different levels of turnout then this becomes an important consideration in polling, and one that in practice makes a big difference to published figures.
All opinion polls include in their voting intention questions the chance to say “I wouldn’t vote”. In practice only a very small proportion of people, often under 10%, pick this option. Given the actual turnout at elections, it seems obvious that some people are saying they would vote tomorrow when in reality they wouldn’t. The challenge is working out who those people are.
Aside from YouGov, who we will come to at the end of this article, the simplest approach to dealing with turnout is that taken by Ipsos MORI. Historically MORI never used to take account of turnout inbetween elections, taking a purist approach that polls away from elections were only snapshots of opinion, not election predictions. Come pre-election polls however they used to ask people to say how likely they were to vote, taking those who said they were very likely. The problem was that the proportion of people who said they would definitely vote was invariably higher than the actual turnout recorded at elections. It was necessary to more finely define people’s likelihood to vote.
So instead MORI (indeed, all the pollsters) use a ten point scale, with 10 being the most likely to vote and 1 being definitely not voting. The proportion of people who rate their likelihood of voting at 10/10 has in the past been pretty close the proportion of people who actually vote, so MORI use this as a their turnout filter, taking only those people who say their chances of voting are 10/10 and assuming all other respondents will actually not bother to vote. Since 2003 MORI have done this for all their polls, though they still produce figures that are not adjusted for likelihood to vote to allow people to draw comparisons with pre-2003 data.
The pattern of turnout in the UK is broadly consistent – Conservative voters are far more likely to tell pollsters they will actually vote than Labour voters are. Ipsos-MORI’s “10/10 only” filter is the most demanding used by any pollster and therefore the effect this has is to strongly favour the Conservative party. The reason that MORI’s polls do not normally show larger Conservative leads than other pollsters is that their weighting regime is more favourable to Labour than that of ICM and Populus, as we’ve seen here.
The problem with MORI’s technique is that, while the proportion of 10/10s is roughly the same as actual turnout, not all those 10/10 people actually do vote, and many of those who say 9/10, 8/10 etc actually do. The graph below is from the 2005 British Election Study and shows people’s answers to the likelihood to vote question compared to whether they actually said they had voted when contacted after the election.
As you can see, it’s not far off a straight line up the middle, the more likely people say their are to vote, the more likely they are to report that they did. In fact, we don’t even need to rely on their own reports, as many people will know, the electoral registers at polling stations on which people’s names are marked off as they vote are available for public examination after the election. As well as asking them, the BES also doubled checked each person against the marked electoral register to see if they actually voted…
Postal votes mess things up a bit, but you can see that in reality it is the case that the more likely people say they are to vote, the more likely they are to do so. On this basis Populus weight their voting intention figures according to how likely people say they are to vote. If someone says they are 10/10 likely to vote, they are given a weighting of 1.0 (in other words, they count as a full person). If someone rates their likelihood of voting at 8/10 they are given a weighting of 0.8, at 5/10 a weighting of 0.5 and so on down.
ICM use a similar method. (see update below). ComRes weight all those who rate their chances of voting at 5 and above in the same way as Populus, but exclude people who rate their chances at below five.
While there are subtle differences between ICM, Populus and ComRes’s weighting by turnout, they should be very minor and are unlikely to produce any significant variation. ICM have experimented using the three different methods, and report it makes no significant difference.
Finally we come to YouGov. YouGov do not normally factor in turnout at all. All people giving a voting intention are assumed to be equally likely to vote. In theory, given the pattern of likelihood to vote we see in polls, this should produce polls with a higher level of support for Labour. In practice however it doesn’t, and when YouGov have produced figures that take into account likelihood to vote, such as at the 2004 London mayoral election, they have been less, not more, accurate. The reason for this is unclear, it may be that the social acceptance pressure that makes people claim they will vote when really they will not is not the same on an online survey, or perhaps people who don’t bother to vote also don’t bother to join a survey panel.
UPDATE: Actually ICM seem to have changed their method, and are filtering their data, taking only those who say they are 7/10 or more likely to vote.