&Elections: The Methodology
A summary of the methodology of &Elections' live Westminster projection.
Stage One: The Foundational Model
The foundational model of my projection is comprised of a mix of two key elements: the linear elections trends adjustment (LETA) model and the multi-level regression with post-stratification (MRP) aggregation model. Where few MRP polls exist, or where none have been conducted for a while, the foundational model will move toward favouring the LETA model and vice versa.
The LETA model assumes that the swing between the 2nd most recent general election (2019) and &Elections’ all parties standing everywhere estimate (APSEE) of the most recent general election (2024) for all parties in every constituency maintains a 20% constancy to the present. The APSEE involves inferring the results for parties in constituencies where no candidate stood in the previous election by using demographic data from the 2021 England and Wales and 2022 Scotland censuses, coupled with &Elections’ model for how different demographics voted in 2024.
The MRP aggregation model involves weighing MRP polls based on the accuracy of the pollster’s final MRP poll for the 2024 general election; the sample size of the poll; and the recency of the poll. As some pollsters conduct MRP polls on a much more frequent basis than others, to ensure that the methodologies of each pollster are accounted for proportionately, recency weighting does not significantly account for how recently every poll was conducted, but rather weighing based on the recency of each poll in the chronology of polls by any specific organisation. Furthermore, &Elections’ model goes further by weighing each candidate in each constituency based on the accuracy of each pollster in forecasting their vote share in the 2024 cycle, followed by also marginally adjusting the result based on its inaccuracy too.
Stage Two: The Final Model
The final model involves mapping the current voting intention for each region onto the cumulative results for each region under the foundational model, followed by applying constituency-level opinion polls (CLOPs), if applicable.
The current Westminster voting intention (WVI) for each region is determined by a combination of national and regional opinion polls. As not all regions are polled, it is initially assumed that each region has a uniform swing between the last general election and current WVI, before polling for regions is added (which currently only exists for Scotland, Wales and London) with the results for other regions being adjusted to ensure WVI figures are consistent with the cumulative projection for each region. These results are mapped onto the findings of the foundational model for each region to give an accurate projection for every constituency consistent with current WVI. Finally, where CLOPs and by-elections exist, the results of these are weighted and adjusted, and the findings are then applied to the constituencies where they apply to produce the final projection.
Nationwide polling and regional polling are both weighted by recency and sample size. As regional polls are scarcer, polls are more statistically significant for a longer period. Nationwide polling also accounts for accuracy, based on the accuracy of each pollster’s final WVI poll before an election in all elections since 1970. Both national and regional WVI polls have the trends among recent polls mapped onto the projected result to allow the model to be better at registering when significant changes happen in polling and so that it more reflects voting intention now, rather than simply reflecting the weighted polls over a recent period. Furthermore, this is useful for regional polls, as it allows the model to maintain accuracy when polls are more infrequent by following the general trends of the region, although the total significance of regional polls for each region is weighted down when this is the case. CLOPs are weighted slightly differently – whilst sample size is weighted broadly similarly, for the reason that they do tend to provide a much more accurate snapshot of the political landscape than any MRP poll, to ensure that they can maintain their statistical significance over time, each constituency poll is adjusted by the uniform national swing between when they were conducted and the present.