Question wording:

Since our first poll after the General Election was called we have used an updated voting intention question wording. Now that candidates have been announced we are showing respondents their ballot paper with named candidates where possible.

How we deal with turnout:

In our polls we ask people how likely they are to vote. Before the campaign, we included people in our voting intention calculation if they put their likelihood of voting in an election called tomorrow at at least 7 out of 10. It’s hard for people to predict whether they will vote in a theoretical election. But now that the election is less than a month away, we think it makes sense to raise the threshold. We also know that whether someone voted in the last election tends to be a better predictor of whether they will vote than their self-reported likelihood to vote.

So in our new methodology, we include people who were too young to vote last time if they say their likelihood to vote is at least 9 out of 10. If people were old enough to vote we use the information about whether they voted – if they did vote we include them if they put their likelihood as 8 or more out of 10, if they didn’t vote we only include them if they put their likelihood as 10 out of 10.

How we deal with “don’t knows”:

When someone answers “don’t know” to the voting intention question in our polls we ask them a ‘squeeze’ question (how they would vote if they were forced to choose?) and if they give an answer we take this as their intended vote. Previously those who answered “don’t know” to this squeeze question were then excluded. This effectively means we assume that if this group do vote they will do so in proportion with those who have expressed a choice. That means even if someone lives in a rural area, is over 75, and voted Conservative in 2019 – if they respond they don’t know who they’ll vote for, they were treated as the average voter i.e. more likely to vote Labour than Conservative.

But we can make a better guess. We now model the voting intention of these “double don’t knows” based on their demographics and previous voting behaviour. We do this using information about how people with the same demographics and voting behaviours said they would vote, from the same poll. So if someone lives in a rural area, is over 75 and voted Conservative in 2019, the model can use the fact that most over 75s in rural areas who voted Conservative in 2019 and do know who they’ll vote for say they will vote Conservative, to guess that if they do vote it will likely be for the Conservatives.

What difference does it make?

The cumulative effect of these changes, applied to our most recent poll, is to add two points to Conservatives and take two points from Labour, decreasing the Labour lead by four points.

Why change methodology now?

In general, pollsters try to keep the way that they calculate voting intention the same between polls – this way changes in results can indicate changes in public opinion rather than changes to methodology. But as the election draws closer we think it is more valuable to make changes to the methodology which we think provide a clearer picture of how the country will vote on 4th July, than to prioritise consistency.

We also recently released our first seat-level projection based on MRP. For this analysis we updated the voting intention methodology, in particular including information about whether respondents voted last time. This means the voting intention in the MRP is different from our regular voting intention polls, and we want to align these. Going forward we will use the methodology outlined above in both our regular polls and MRP modelling.

But what about tracking through time?

As aggregate analyses rely on consistent methodology we will continue to publish our voting intention as per the old methodology in our data tables.

Is this your last methodology change?

Once voting registration closes, we will exclude respondents who say they aren’t registered to vote.