Comparing Outputs from Seat Models
Seven different pollsters have published seat estimates based on MRP (Multi-level regression and post stratification) during this election campaign. We wanted to see how the results compare with each other and with the outputs of our model. MRP combines a large sample of voters with demographic data to forecast each seat. In recent elections it has had some success in forecasting the number of seats each party will win. Our model does not use MRP but instead relies on aggregated polling and past voting intention to forecast the outcome.
More in Common have published two different MRPs, one with the News Agents and a second with Electoral Calculus, we have included both sets of results in our analysis.
The results of our comparison are very interesting, there is a great variation in what the MRPs are telling us. This is because they use different underlying data, different assumptions and slightly different techniques. A good example is the predictions for the Conservative seat count, which varies from 53 for Savanta to 155 for More in Common (our model says 173). There are also sharp disagreements in some seats, like Clacton where Ipsos make it safe for Reform, YouGov likely Reform, Savanta and Survation a toss up and everyone else safe for the Conservatives.
This all suggests that MRP is not a magic bullet to clear the uncertainty surrounding the election result. People should be wary about relying on only one of these models if they are tactically voting, particularly in Scotland.