NATIONWIDE OPINION polls in Britain have long made clear that the Labour Party is on track to win the overall popular vote by a thumping margin. In Britain’s first-past-the-post electoral system, however, translating votes into seats is no easy task. The leading statistical method to produce such estimates is called multi-level regression and post-stratification (MRP). This year, in partnership with WeThink, a polling firm, The Economist has conducted its first-ever MRP analysis of a British general election. This mega-poll finds that Labour is on track to win 465 of the 632 seats in England, Scotland and Wales, giving it the biggest majority since the second world war. Meanwhile, the ruling Conservative Party, which won 365 seats in 2019, is set to collapse to a mere 76, the fewest in its history.
To produce these figures, WeThink surveyed a representative sample of 18,595 British adults between May 30th and June 21st. The firm asked respondents which party they planned to support, along with where they lived and basic demographic information about them. Using these data, we built a statistical model—the “multi-level regression” of MRP—to predict voting intentions for each of 16m possible unique combinations of voters’ age group, sex, ethnicity, education level, constituency and voting history. This model is similar to our British “build-a-voter” tool that you can explore here. For example, we estimate that a white woman in Bromsgrove aged 50 to 54, whose highest educational qualification is GCSE, who voted to leave the EU in 2016 and Conservative in 2019, has a 45% chance of voting Conservative again this time and a 29% chance of voting for Reform UK.
The next step, known as “post-stratification”, involves estimating how many people with each of these 16m profiles live in each constituency. In Bromsgrove, we reckon that there are around 185 people in the group described above, whereas in Bethnal Green and Stepney there are only five or so. To produce the final results, we simply multiply the expected vote shares for each party in a given demographic group by the number of people in each constituency who belong to that group.
Our MRP paints a remarkably bleak picture for the Conservatives. A large share of the party’s electoral base is abandoning it for Reform UK, a populist right-wing party that has surged in the polls since Nigel Farage, a leader of the successful campaign for Britain to leave the EU in 2016, announced that he would stand for Parliament. The MRP expects Reform UK, which has never won a seat before, to secure 14% of the national vote and three seats. The Tories are also faring poorly in races where the Liberal Democrats are their primary opposition: the MRP expects Britain’s third party to win 52 seats, the highest number since 2010. Even in seats where the Conservatives are not competitive, however, Labour is on track to make gains—particularly in Scotland, where the Scottish National Party is estimated to fall from 48 seats to 29.
Although MRP relies on statistical modelling to estimate seat-level results from a nationwide survey, it is still ultimately the product of a single poll. MRPs did yield broadly accurate results in the 2017 and 2019 campaigns, but they were rare before then, and there is no guarantee that they will repeat such successes this year. In addition to the sources of error intrinsic to all polling—respondents may not be representative of the eventual electorate, and voters can change their minds between the time they are interviewed and the election—MRPs also face unique risks. They can model the relationship between demography and voting intentions incorrectly, or miscalculate how many people in each demographic group will turn out to vote.
Although we are pleased to contribute a high-quality survey to this year’s polling landscape, many other MRPs use similarly rigorous samples and methods and have produced somewhat different results. Moreover, if there is a particular dynamic about the current election that the MRP approach fails to capture, all MRPs are likely to misfire in the same direction. As a result, our best prediction of the final results is an “ensemble” model that combines the “regional-swing” method, whose forecasts we have published every day since March, with an average of all public MRP estimates—including our own. You can see these blended numbers here.
