Explainer-UK election: What is the MRP method of modelling opinion polls?

LONDON (Reuters) - As Britain's election campaign enters its final stretch, the work of opinion pollsters is back in the spotlight with several recent projections of a record victory for the opposition Labour Party grabbing the headlines.

Labour's ample 20-point opinion poll lead has hardly budged since Prime Minister Rishi Sunak announced the July 4 election last month, shifting the focus to the question of how big Keir Starmer's win will be rather than whether it will happen.

But Britain's first-past-the-post electoral system means the number of seats each party wins does not closely reflect the national share of vote they receive, so pollsters use so-called MRP modelling in a bid to more accurately estimate the result.


MRP stands for Multilevel Regression and Post-stratification and it is used by pollsters to estimate public opinion at a local level from large national samples. Pollsters describe it as a model that uses polling data, rather than a poll itself.


Pollsters construct a statistical model which summarises how voting intention differs depending on the characteristics of survey respondents and where they live. This will take into account factors such as age, income, educational background and past voting behaviour.

This model is then used to produce estimates of the voting intentions among different types of people living in different areas of the country.

Pollsters combine that with official data on the numbers of people of each type living in each area to generate an estimate of overall voting intention for the constituency.

The exact model used to predict voter behaviour varies from pollster to pollster.


Conventional polling methods often apply a uniform national swing to predict how many seats a party will win.

This assumes there will be the same change in vote share for each party throughout the country, which is rarely the case, meaning it can overestimate the performance of a party in some areas and underestimate it in others.

MRP sample size is also much higher. Typical political polls rely on between 1,000 and 2,000 responses, while MRP modelling uses data from tens of thousands of voters.


MRP is a relatively new technique.

After polling companies miscalled an election in 2015 and underestimated support for Brexit in the 2016 referendum, many looked to use more sophisticated data analysis to come up with seat-by-seat results.

MRP was used by YouGov in 2017 to accurately predict Conservative Prime Minister Theresa May would fall short of an overall majority. YouGov said its model called 93% of seats correctly that year.

The method had some success in 2019, with YouGov's MRP predicting a clear majority for the Conservatives although underestimating the scale of it.


British voters have become more unpredictable, with Brexit scrambling traditional political allegiances.

More voters switched between the two main parties at a 2017 election than in any ballot dating back to 1966, according to research by the British Election Study. The more people change their minds, the harder it is to draw a representative sample.

"MRP only works when there is a strong link between, on the one hand, the characteristics of individuals and areas and, on the other, the opinion being modelled," the British Polling Council says on their website.

Pollsters describe MRP as an estimate of the range of possible results.

Savanta's MRP published on June 19 projected Labour could win a whopping 516 seats in the 650-strong House of Commons, with the Conservatives on 53. But it also noted nearly 200 seats had less than 7.5 percentage points between the parties in first and second place, deeming them as 'too close to call' and meaning the end result could be very different.

YouGov's MRP published on the same day gave Labour 425 seats and the Conservatives 108, but classified 109 constituencies as "tossups" with fewer than five points between the parties in first and second place.

There may also be specific issues at play in individual seats which MRP modelling is unable to capture.

(Reporting by Kylie MacLellan; editing by Mark Heinrich)