A conventional national voting-intention headline does not tell you who is ahead in a particular constituency. MRP tries to answer that smaller-area question by combining survey answers with a statistical model and information about the people living in each area.
An MRP constituency figure is a modelled estimate. It is not simply the percentage chosen by a large representative sample interviewed separately in every seat.
What does MRP stand for?
The name is multilevel regression and poststratification. In broad terms, the regression estimates relationships between voting intention and characteristics of respondents and places. The poststratification stage combines estimated preferences with the composition of the relevant population.
The British Polling Council’s MRP briefing explains these stages. The term describes a family of approaches, rather than a single model used identically by every organisation.
Following a small worked example
Consider an illustrative constituency divided into two population groups. The model estimates Party A support at 40% in the first and 20% in the second. Population information puts 30% of relevant voters in the first group and 70% in the second.
The combined estimate is (0.30 × 40%) + (0.70 × 20%) = 26%. A constituency split 70:30 between the same groups would produce 34%. The estimated relationships can be identical while the constituency figures differ because the populations differ.
Real models use many characteristics and more elaborate relationships. The example illustrates the aggregation step, not the full fitting procedure, treatment of turnout or propagation of uncertainty.
Why does the multilevel element matter?
Some types of respondent are scarce in particular places. A model which tried to estimate every local combination independently would encounter thin or empty cells. Multilevel modelling allows information to be shared through broader patterns while allowing relevant differences between groups or areas.
The resulting estimate depends on which patterns the model represents. Variables, interactions, geographic structure, past-election information and decisions about undecided voters can differ between providers.
More in Common’s published January 2026 methodology gives one example of how these choices are described for an actual release. It should not be assumed to describe other organisations’ models.
What are the main limitations?
Population information can be imperfect or insufficiently current. Survey recruitment can miss people. A model can omit a local influence, and a constituency’s relationship with its demographic profile can change.
A well-known candidate, an independent campaign or a newly important issue may be difficult to represent if the available data do not capture it. The geography of support matters, but geography cannot replace evidence about individual contests.
Larger samples help supply information; they do not make modelling assumptions disappear. Nor does the MRP label tell you that every quoted seat total comes with a properly reported probability distribution.
From shares to winners
Under First Past the Post, the highest local vote share wins. Small errors around a close contest can change the projected winner while barely changing the national vote.
Read the constituency margins, fieldwork, methods and uncertainty alongside any total. Our MRP section collects releases; the guide to comparing seat forecasts explains why two similar-looking models can produce quite different parliamentary outcomes.
Sources and further reading
Reviewed 7 October 2026. Worked examples are illustrative unless an election or poll is identified.