A Voter's Guide to MRP

This page first posted 4 June 2024

Why are seat predictions for the General Election so varied? Last Friday, Electoral Calculus and Find Out Now published a large-scale MRP poll for the Daily Mail and GB News, the first of the election cycle. It produced some "eye-catching" numbers, with Labour predicted to win a majority of over 300, and the Lib Dems a whisker away from becoming the official opposition, once tactical voting is accounted for.

These results, if they are to materialise, would be unprecedented in British electoral history. Given the nature of these seat totals, it is natural to be slightly sceptical. It is also worth understanding what an MRP is, what it isn't, and what might be leading to rather sizeable differences in MRP seat predictions at this stage in the election cycle.

What is an "MRP" Poll?

MRP (which stands for Multi-level Regression and Post-stratification) is a relatively new polling technique, which has started to compete with conventional polling methods in recent elections. Conventional polling methods often apply a 'uniform national swing' (or UNS) to predict how many seats each party will win in an election.

The UNS approach assumes that there will be the same change in vote share for a given party evenly throughout the country. But this doesn't happen in practice. In 2019, the Conservative vote share went up by 5% in seats which clearly voted Leave in the EU referendum, but went down by 3% in seats which voted to Remain. Geographically, Labour lost 13% in the North East, but only lost 6% in the south of England. These differences would not be captured by UNS. Therefore, uniform national swing models can overestimate the performance of parties in some areas, and underestimate it in others.

House of Commons 2019

Every seat counts: MRP tries to predict who will be elected to each House of Commons seat

The MRP method tries to solve this problem by working out the relationship between peoples' voting intention and their individual demographic characteristics like age, income, educational background and a variety of other data sources, including past voting behaviour. Then, you can count how many of these different types of voters there are in each constituency, to make a prediction of how parties will perform on a seat-by-seat basis.

For example, the MRP in 2019 detected that Leavers are more likely to vote Conservative, and Remainers less likely. This would automatically feed through into an increased Conservative vote in Leave seats, and a lower vote in Remain seats.

How do you know the MRP is right?

MRP has a good track record. It predicted the results of the last two UK elections pretty well. At the 2019 election, Electoral Calculus used MRP to make the most accurate of all the pre-poll predictions of the result.

We have also double-checked our figures using alternative methods, and our results are consistent with previous landslide elections, such as Tony Blair's win in 1997.

We think MRP has a good track record, but no method is foolproof and we definitely won't get every single seat right.

Why do MRPs have differing predicted seat totals for the main parties?

Firstly, not every MRP is the same. The statistical 'building blocks' used to model voter behaviour will vary from pollster to pollster, which will in turn lead to different vote shares and seat totals.

All different pollsters will get different national vote shares. The overall national vote share is still an important driver of how many seats each party will win. In our latest MRP poll, we used the recent national average of opinion polls, which shows a Labour lead of 22%. Other pollsters have different Labour leads, and that will result in different seat predictions.

YouGov, meanwhile, released an MRP yesterday with a Labour majority of 194 seats, which is lower than our own prediction. YouGov see the Labour lead at around 18%, which is a bit lower than the polling consensus, and explains much of the difference between the two seat predictions. Time will reveal which method is more accurate.

Will Labour really win nearly 500 seats on election night?

For all the strengths of the MRP method, it cannot predict future changes in public opinion. MRP polls, like any other, present a snapshot of voting intention at a specific point in time. Given that election day is over a month away, there is every chance that the polls narrow or widen, and the political map changes as a result. For this reason, we'll be updating and refining our seat forecast throughout the campaign. However, at the moment, it looks like a Labour majority exceeding that of 1997 is likely.

Within the polling industry, there is a memory of the 1992 election, where the Conservatives won a surprise majority at odds with public expectations. However, despite the political turbulence of recent years, 2024 is not like 1992. Exactly one month before election day in 1992, Labour and the Conservatives were roughly neck-and-neck in the polls. Today, however, with just over a month until election day, Labour are still 20 points ahead, which is a very large lead in historical terms. Even in the event of a significant, industry-wide overestimation of Labour support, the election still looks like Labour's to lose.