Multilevel regression and post-stratification (MRP) is a way of using large national samples to estimate public opinion at a local level.
Here, a “large” national sample can be anything from 6,000 to 100,000 respondents. Most often MRP is used to estimate opinion in Westminster constituencies.
The first stage of MRP is the construction of a statistical model that summarises how public opinion or voting intention differs depending on the characteristics of (i) individual survey respondents and (ii) the area in which they live.
This model is then used to generate estimates of the balance of opinion or vote intentions among different types of people living in different areas (e.g. vote intention among 55-64 year old men who voted Conservative at the last election and who have a university degree).
Meanwhile, the numbers of people of each type in each area is ascertained from external sources (such as Census and other data).
Finally, taking into account how many of each type of voter there are in each area, an overall estimate of public opinion or vote intention for an area is generated by adding up the estimated opinion or vote intentions for each type of voter in that area.
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. For example, MRP is good at modelling the proportion of people in each area who will vote, say, Labour, because how many people voted Labour in this constituency last time is a very strong predictor of how many people will vote Labour this time.
Conversely, MRP would be a bad predictor of modelling the proportion of people in each area who have a particular hair colour, because it is hard to predict hair colour on the basis of their personal characteristics or the area in which they live.
MRP also only works when the same “opinion object” is being asked about in each area. MRP should not be used to investigate how satisfied different areas are with their MP, because the relationship between “past vote” and “support for the local MP” looks very different depending on the party to which the local MP belongs.
Many of the questions you should ask when given a regular poll have their equivalent in MRP analyses.
Just as in the case of a regular poll, it is important to understand how the sample was obtained.
Much as in an ordinary poll you would ask how the data have been weighted, in a MRP poll you should ask which characteristics were used in the modelling of the survey data.
Instead of asking about the margin of error, you might ask about the level of uncertainty associated with the MRP estimates.
The officers of the British Polling Council would like to thank Professor Chris Hanretty for developing this explainer on MRP. It will feature in the 2023 edition of our “Opinion Polls: Guidance for Journalists” publication, which will be released later in the summer.