Using MRP for our voting intention polling

June 04, 2024, 2:23 PM GMT+0

Following the launch of our general election MRP model this week, we will be using the same MRP technique on our regular weekly voting intention polls during the election for Sky and the Times.

Why are we doing this?

In recent months our MRP model and our traditional voting intention figures have tended to show a slightly different picture. Our MRP model gives Labour a marginally lower vote share and the Liberal Democrats a slightly higher share. In the run up to the polling day, we want to put out clear and consistent data, using the method that we think is the most accurate. That is MRP.

What will the changes mean?

This is primarily an "under the bonnet" change. Our regularly weekly polls for Sky and the Times will still be ~2,000 people and still show national vote shares for the country. All the other questions on the poll will be run and weighted in the traditional way. But rather than using rim weighting or raking to weight the voting intention numbers, we will instead put the same data through our MRP model to produce vote shares for the country.

Is this the same as your MRP seat estimates?

No. It uses the same approach and the same model but to do a different thing. For our seat estimates we will put 60,000+ respondents through the model and project shares for 631 individual constituencies, using the demographics of each individual constituency. For our regular polls we will put ~2,000 people through the model, and project to just one geography – Great Britain as a whole.

What impact will this have on our voting intention figures?

In the present political climate, the changes will likely produce a slightly smaller Labour lead, in line with that in our larger MRP model. There are several reasons that using the MRP method currently produces a slightly lower Labour lead:

  1. Constituency question: We drive the MRP off a question asking people how they will vote in their specific constituency, which better picks up tactical voting considerations. At the present time, this tends to increase Liberal Democrat support by around 2 points and decrease Labour support by a similar amount.
  2. Don’t knows: The MRP model effectively reallocates don't knows. Rather than historical approaches that have done this by assuming those people will vote the way they did in the past, it does so by modelling their vote based on those of similar political background and demographics.
  3. Turnout: The MRP approach addresses the question of turnout by basing it upon demographics, rather than self-reported turnout, where people tend to consistently overestimate their own likelihood to turnout, leading to problems of understating the turnout gap between old and young voters.

When we publish our first figures using the MRP method, we will show what the lead would have been under the old approach so that people can correctly assess whether any differences are down to the change in methodology, or genuine changes in public opinion.