THERE IS LITTLE doubt about the likely victor of Britain’s general election on July 4th: with a lead of 20 percentage points in national opinion polls, the Labour Party is extremely likely to win. But there is uncertainty about the size of Labour’s majority in Britain’s 650-seat House of Commons. Some polling firms have published seat predictions using a novel technique known as multi-level regression and post-stratification (MRP). What are these polls—and how accurate are they?
Because Britain uses the first-past-the-post system to elect MPs for each constituency, the number of seats a party will have in the Commons does not closely reflect national vote shares. In 2015 the UK Independence Party got over 12% of the vote but won just one seat. In the past, pollsters would predict a party’s seat total using “uniform national swing” (UNS). This means applying the national change in its vote share—from the previous general election to that reported by opinion polls—to every constituency. Assuming that constituencies will swing uniformly is a great simplification. But the technique has been reasonably accurate. In 15 general elections between 1959 and 2019 UNS correctly predicted the largest party in 11 of them. But UNS performs poorly when elections are tight or the electorate—as has been the case in recent years—behaves in unexpected ways.
MRP attempts to produce better predictions by segmenting voters into fine-grained groups. First the pollster builds a model to predict how demography—age, sex, race, education—and other factors such as past voting behaviour affect voting intention. How, for example, is a white British university-educated woman aged 25-34 who previously voted Green likely to vote this time? Then the pollster produces a population estimate for each of these groups for every constituency: this is called a post-stratification frame.
The technique relies on a very large sample. Conventional polls need to survey only around 1,200 people to get an accurate sample of Britain’s roughly 46m registered voters. Pollsters try to make sure their samples are broadly representative of the whole population and apply weights to different types of respondents to adjust for their propensity to answer surveys. But once you start splitting voters into smaller demographic groups, a much larger sample is needed. Some MRP seat predictions—dubbed “megapolls” by the press—have been based on sample sizes exceeding 15,000.An MRP published by YouGov in 2019 had a sample size of 100,000 people.
MRP polls should give a solid estimate of each party’s final seat count. But they are not foolproof. The surveys they rely on might be poor quality, perhaps because they are not sufficiently stirred or the weights used to make them representative are inaccurate. And political scientists must make pretty heroic decisions about who will vote, as people who say in surveys that they will turn out often fail to do so on election day. Voters may behave unpredictably, too, for example by voting tactically to oust a disliked incumbent. Errors can creep in at almost every step.
MRP polls have not taken root in America, in part because state-level polling that allows prognosticators to forecast the presidential electoral-college vote is pretty accurate. The first prominent British MRP forecast was made by YouGov a week before the general election in 2017. It said that the Conservatives would win between 274 and 345 seats, challenging the expectation that the Tories would secure a majority. Sterling fell by 0.5% on the date of the prediction’s release. Sure enough, the Tories got 317 seats. In 2019, two days before the general election, YouGov estimated that the Conservatives would win 339 seats; they got 365. Two other MRP predictions, from FocalData (351) and Electoral Calculus (337), also underestimated the scale of the Tory victory.
After two rounds of reasonable results, MRP’s reputation now precedes it. There have been four MRP polls since the 2024 election was called that predict that Labour will win between 382 and 505 seats. Not all of them can be right. The Economist’s seat prediction—which is not an MRP, but uses a variant of UNS that incorporates information about regional voting behaviour and is tested on 15 previous general elections—expects Labour to win between 305 and 500 seats (its median estimate is 390). The Economist will also publish an MRP in collaboration with WeThink, a pollster, shortly before polling day. Whatever the result, the most accurate soothsayer is sure to claim bragging rights. ■