Labour's reliance on non-voters
Polls which show Labour up also show the party relying heavily on non-voters and voters of other smaller parties
This week's polls have given very different snapshots of the 2017 general election.
This is not a difference which can be explained by invoking normal sampling variability. The differences stem from particular choices that polling companies have made, and in particular choices regarding the way in which people's likelihood of turning out to vote is modelled.
So who's right?
I think the true Conservative lead probably lies in between these figures, but closer to ICM's figures than YouGov's figures. I've got three reasons for thinking this.
The first reason is a boring one: in the absence of any good reason to prefer one pollster over another, we should assign all their figures equal probability. I think I have some reasons for preferring ICM's figures, but they're not conclusive, and so there's a very mealy-mouthed part of me that assumes the true figure lies somewhere in between.
The second reason is systemic: I think that the polling industry as a whole will over-estimate Labour (and under-estimate the Conservatives) to the same extent that it historically has (see the 2015 Polling Inquiry report, ch. 4).
This means that I think the average Conservative lead in the polls is an under-estimate of the eventual Conservative lead on the day. If the average Conservative lead in the polls is equidistant between YouGov and ICM, I tilt towards ICM.
The third reason, though, is particular to those polls which show Labour within touching distance. Looking at these polls, support for Labour seems to rely to a large extent on those who either did not vote in 2015, or voted for a minor party. Since I am skeptical that many of those who did not vote in 2015 will vote now, I am skeptical of any figure which relies heavily on these people.
Enter the (transition) matrix…
To make this point, I'll be looking at some of the cross-tabs in the most recent YouGov poll.
I'm going to look not at the headline vote intention figures, but on the "raw" vote intention figures, which don't include any adjustments for respondents' likelihood of turning out to vote.
I'm looking at these figures for two reasons. First, they're not too different from the headline figures. The "raw" figures suggests that 31% of people will vote Conservative, and 30% Labour — but some 23% said that either didn’t know how they would vote, or that they didn’t know. If you remove these people and divide each party’s share of the vote so that they add up to 100%, you get Conservatives on 40% and Labour on 39%. That's not too far from the headline figures (42% and 39%). Presumably the Conservatives reach 42% because they are more likely to vote — although table 2 of the report doesn’t suggest that.
Second — and much less impressively — I can't back out the turnout weighting, so I can't say much about the cells in the headline figures if I haven't understood how they've been created.
Here's the table, with respondents' vote intentions by 2015 vote choice.
Only four categories for 2015 vote are given — Conservative, Labour, Lib Dem and UKIP — so we must imagine a fifth, residual category, which includes people who voted for another party, people who were eligible to vote in 2015 but did not, and people who were not eligible to vote in 2015 but are now eligible to vote.
The percentages in these rows give some indication of where the 2015 vote has gone. For the moment, let’s assume that the electorate in 2017 is exactly the same as in 2015. That assumption is false — some people have died since then, and others have come of age — but it’s a useful simplifying assumption.
In that case, the 2017 Conservative vote would come from 74% of those people who voted Conservative in 2015, 9% of those who voted Labour in 2015, and so on.
If you multiply these row percentages by parties’ GB-only share of the electorate in 2015 (which we can do because we’ve assumed the electorate is the same), and if we assume that no-one who voted for any other party now votes Conservative, then you get the following result
0.74 * 25.1 (Cons->Cons)+
0.09 * 20.7 (Lab->Cons)+
0.09 * 5.4 (LDem- >Cons) +
0.40 * 8.6 (UKIP -> Cons)+
0.00 * (100 — (25.1 + 20.7 + 5.4 + 8.6))
If you do the same for Labour, you get
0.07 * 25.1 (Cons->Lab)+
0.72 * 20.7 (Lab->Lab) +
0.25 * 5.4 (LDem -> Lab+
0.16 * 8.6 (UKIP->Lab)+
0.00 * (100 — (25.1 + 20.7 + 5.4 + 8.6))
This is quite far from the published figures, but then we’ve made the unrealistic assumption that everyone who didn’t vote for one of the four largest UK-wide parties (40.2% of the electorate) doesn’t vote at all.
We can therefore repeat the above equations, but solve for two unknowns. For the Conservatives, we set
24.4 (the number from before) + p * 40.2 = 31%
which works out to p = 0.16.
For Labour, we get
19.4 (the number from before) + q * 40.2 = 30%
which works out to q = 0.26.
This means that, compared to the Conservatives, Labour is drawing more than one and a half times more voters from this pool of smaller party voters and 2015 non-voters.
With additional assumptions about voters who voted for other parties (SNP, Plaid, Greens), you can work out how many 2015 non-voters say they would vote Labour (if they voted). For example, suppose all Green voters vote, and Labour takes 50% of those, and suppose further that all SNP and Plaid voters vote, and Labour takes none of those. In that case, Labour gets an additional 1.3% of the electorate, then Labour gets an additional 1.3% of the electorate, and would draw upon 27% of the electorate who didn’t vote. This is a bigger share, because the SNP and Plaid vote (from which, by construction, Labour get zero votes) is bigger than the Green vote.
I think these disparities are worth noting, because the two features of this data (non-voters break heavily Labour, and non-voters vote at consequential rates) are slightly controversial. There's a lot on the habitual character of voting which suggests that non-voters are unlikely to return to being voters. There’s also a fairly large literature which suggests that non-voters are not systematically different from voters in their beliefs. If you don't believe that, here's an earlier YouGov poll on the subject:
Certainly the latter turnout assumption seems an important one, and is behind most of the divergence in the polls. I've already said that I can't back out the turnout weights from the information YouGov publish (although maybe I'm missing a trick here). But it's possible to construct alternative headline figures based on different rules of thumb regarding turnout. On one (stupid) rule of thumb, all of those who voted in 2015 vote, and none of those who didn't vote in 2015 vote now.
In that case, you get a Conservative lead of closer to eight percent than to the three percent reported by YouGov. That alone would more than half the difference between YouGov and ICM.
Eighteen months ago, the answer to the question "Can non-voters win the election for Labour" would have been a resounding no. Some of the polls seem to suggest that non-voters can get Labour close. That, to me, sounds unlikely — but we'll know within a week.