What is MRP
MRP stands for Multilevel Regression with Post-stratification. Though this sounds very complicated, it basically means taking into account the likely individual preferences of certain voting groups, then applying this onto sub-regions, such as UK parliamentary constituencies. The first step is to work out how different demographics affect voting behaviour, usually by conducting a huge number of interviews (far higher than for national polls). Then you perform what is called a multilevel regression, finding the best fit for voting behaviour based on your sample. If you don’t have large sample sizes, the whole thing falls apart, as you are progressively more likely to make bad estimates. The next step is the stratification, involving the projection of this information onto each constituency in the UK, in an attempt to work out exactly how each one will vote in the election.
YouGov’s use of this technique correctly predicted the winner in 93% of UK constituencies last time, and it is considered to be a far better technique than simply estimating national vote share, and then assuming uniform swing. No prediction is 100% accurate, but MRP is perhaps the best that has been developed so far for predicting elections.
Why doesn’t every pollster do this?
Time and resources, simply. As stated, huge sample sizes are required to make this technique at all accurate, and most pollsters do not have the time or resources to put into this.
Now, onto YouGov’s MRP, released 27th November, incorporating interviews taken from 20th November to 26th November. This predicts that the Conservatives will win 359 seats, a parliamentary majority of 68, with Labour on 211, the SNP on 43 and the Liberal Democrats on 13.
However, it should be noted that the prediction includes a margin of error, like all sampling from a larger population group. Here, the prediction ranges from 328 seats, only just enough for a majority, to 385, which would be considered a landslide. Furthermore, there are still two weeks remaining in the campaign, and the situation could still change significantly.
How much faith should we put in this result?
In 2017, this model was the most accurate, although it did underestimate the Conservatives final seat total.
However, like any poll, it should be treated with a healthy scepticism, particularly because we are still two weeks out from the polling day. Small changes could completely change the result in a lot of marginal seats, so everything is still to play for. The only totally accurate poll will be the one on election day.