Explainer
What is MRP Polling? Multi-Level Regression and Post-Stratification Explained
MRP — Multi-Level Regression and Post-Stratification — is the most powerful tool in modern UK electoral polling. Unlike standard national polls that produce a single set of headline figures, MRP can estimate how every single UK constituency might vote. It is the method behind YouGov’s now-legendary 2017 prediction, which — alone among major polls — correctly called a hung parliament.
Understanding MRP is essential for interpreting the difference between voting intention polls (which show national numbers) and seat projections (which show what those numbers mean under First Past the Post).
How MRP works: the two steps
Step 1: Multi-Level Regression (the ML part)
A very large survey — typically 10,000 to 50,000 respondents — is conducted. A statistical model is built that estimates the relationship between demographic characteristics and voting behaviour. The model asks: given someone’s age, education, gender, housing tenure, ethnicity, and region, what is the probability that they will vote for each party?
The “multi-level” part means the model allows for variation both within demographics across constituencies (a university graduate in Hartlepool may vote differently from a university graduate in Chelsea) and within constituencies across demographics. This flexibility is what gives MRP its power over simpler uniform swing models.
Step 2: Post-Stratification (the PS part)
The model is applied to the actual demographic profile of every UK constituency. Using census data, analysts know roughly how many people of each age, education level, housing type, and ethnic background live in each of the 650 constituencies. By feeding these demographics into the model, the pollster estimates how each constituency would vote under current polling conditions.
The result is a projected vote share for each party in each constituency — which can then be converted into a projected seat count. This is fundamentally more sophisticated than the older “Uniform National Swing” (UNS) method, which applies the same national movement in every constituency regardless of local context.
Which UK pollsters use MRP?
- YouGov: Pioneer of UK MRP. Published the first high-profile MRP in 2017 with a sample of ~50,000, correctly predicting a hung parliament when all standard polls showed a large Conservative majority. Published multiple MRP models ahead of the 2024 General Election.
- Survation: Published constituency-level MRP analysis ahead of the 2024 General Election with strong accuracy.
- More in Common: Publishes frequent MRP polling, including constituency-level estimates between elections. Particularly useful for tracking how demographic changes affect individual seats.
The House of Commons Library publishes comprehensive constituency-level data that MRP models draw upon for their post-stratification step.
MRP vs. standard voting intention polls
| Feature | Standard VI Poll | MRP Poll |
|---|---|---|
| Sample size | 1,000–2,000 | 10,000–50,000 |
| Geographic detail | National only | All 650 constituencies |
| Frequency | Weekly | Pre-election; occasional between elections |
| Seat forecasts | Requires uniform swing assumption | Direct constituency-level modelling |
| Cost | Low (£5,000–£20,000) | High (£100,000+) |
MRP accuracy: the track record
YouGov’s 2017 MRP was the defining moment for the method. Published two weeks before polling day, it projected 20–50 Conservative seats short of a majority. Every other major poll showed a large Conservative majority. The actual result: a hung parliament, exactly as MRP predicted.
At the 2024 General Election, YouGov’s final MRP projected Labour winning between 385 and 431 seats. The actual result was 412 seats — precisely in the centre of the range. See our 2024 results analysis for a full comparison of how every pollster performed.
MRP limitations
MRP is powerful but not infallible:
- Constituency-level error: While national totals are often accurate, individual constituency-level predictions can have wide error bars, particularly in marginal seats where the model has less historical data
- Assumes stable demographic-vote relationships: If a major shift in voter behaviour is underway (as happened in Scotland in 2015), the model may miss it until it is visible in the large sample
- Cost: Full MRP is expensive, limiting how often it is published between elections
- Not available between elections: Current headline VI polls cannot be directly converted to MRP seat projections without running the full model
For between-election estimates, see our GE 2029 forecast, which synthesises published MRP data with uniform swing models to project likely seat ranges.
Frequently Asked Questions
What does MRP stand for in polling?
MRP stands for Multi-Level Regression and Post-Stratification. It combines a large national survey with census data to estimate voting behaviour at constituency level. The multi-level regression models how demographics predict vote choice; post-stratification applies that model to each constituency’s actual demographic composition.
Which UK pollsters use MRP?
YouGov, Survation, and More in Common are the main UK pollsters that publish MRP models. YouGov published the first high-profile MRP poll ahead of the 2017 general election, correctly predicting a hung parliament when all standard polls showed a large Conservative majority.
How accurate is MRP polling?
MRP has a strong track record. YouGov MRP correctly predicted a hung parliament in 2017 and a large Labour majority in 2024, with the actual result (412 seats) within their projected range of 385–431. Constituency-level predictions can still have significant error bars in marginal seats.
What sample size does MRP require?
Most UK MRP models use between 10,000 and 50,000 respondents. The large sample allows meaningful estimates at constituency level, where a standard 2,000-person poll would have far too few respondents from each of the 650 seats.
How is MRP different from uniform national swing?
Uniform National Swing (UNS) applies the same vote change to every constituency. MRP models each constituency individually based on its demographic profile, allowing it to capture differential swings across the country. This is why MRP is far more accurate than UNS in elections with large regional variation.
Can MRP models be run between elections to track opinion?
Yes, though at higher cost than standard polls. More in Common publishes regular between-election MRP to track constituency-level shifts. Our GE 2029 tracker combines published between-election MRP data with uniform swing adjustments for continuously updated seat projections. The limitation: between-election MRP is less reliable because campaign dynamics have not yet activated and the demographic-vote relationship is less stable outside election periods.