Smaller parties, such as the Liberal Democrats and the Greens, are harder to model. National opinion polls can tell us whether these parties are doing better or worse than they did at the last election. But that doesn't translate into a seat prediction very easily. And uniform national swing (UNS) models are not very reliable. Sometimes the Lib Dems can gain seats in an election where their overall vote share declines, which would not be predicted by UNS.
A better approach is to use MRP polls. These work by matching the demographics of poll respondents with the demographics of each constituency, to get a seat-by-seat prediction of the election result.
An extra complication is that the demographic profile of (say) Liberal Democrat voters can be quite similar to that of Labour voters. If the MRP only looks at demographic factors, it can be less good as distinguishing between Lib Dem and Labour voters. For two voters with similar characteristics, one might vote Labour if they are in a 'Labour area', but the other might vote Lib Dem if they are in a 'Lib Dem area'.
Ignoring this leads to an overly uniform Lib Dem vote share across seats, which will under-predict the number of seats won by the Lib Dems.
In reality, support for the Liberal Democrats and the Greens tends to be 'lumpy': it is quite high in a small number of seats, and quite low in most other seats. It is a known problem that MRP models are not always good at identifying the areas of smaller party support.
A solution to this problem is to use some selected neighbourhood characteristics in the MRP regression. MRP works by finding correlations between the respondents' voting intention and the characteristics of the respondent. Normally, these characteristics are geo-demographic characteristics of the respondent, such as
But it is also possible to add in additional characteristics which are a function of each respondents' local area. These can include government statistical data, such as
Using these neighbourhood characteristics may be more helpful than asking the respondent directly, as there is consistency of definitions and it avoids refusal to answer (which is a marked problem for income questions).
But the real power of neighbourhood characteristics comes from local political variates. These can include
The last of these can be derived from lists of target seats, either from the party itself (eg Labour's list), or by estimates of likely targets.
Using a recent large MRP poll, Electoral Calculus checked the effect of these additional variates on the MRP regression. The long-term general election strength was a very important factor, which was clearly statistically significant. Additionally, local election strength was also an important factor. House prices and campaigning strength were more minor factors, but still made a difference.
In other words, voters are influenced in their political choices by their neighbours and how their neighbours have voted in the past. People who live in an area with historic Lib Dem strength are more likely to vote Lib Dem than similar voters living elsewhere. And people who live in an area with several Green local councillors are more likely to vote Green than similar voters living elsewhere.
In preparation for the 2024 election, Electoral Calculus has upgraded its MRP model to include these additional neighbourhood characteristics. The model algorithm decides whether any particular characteristic is statistically useful, based on the polling evidence and the strength of any correlation between the characteristic and voting intention. If there is no statistical evidence that the characteristic is related to voting intention, then it will not usually be used in the regression. But if a characteristic is statistically meaningful, then it will be used to predict voter behaviour.
We enabled this model for the January 2024 prediction. The difference between the old and the new models for the same polling data is shown in the table below (old boundary seats).
| Party | Old method | New method | Change |
|---|---|---|---|
| CON | 156 | 149 | −7 |
| LAB | 412 | 417 | 5 |
| LIB | 30 | 32 | 2 |
| Reform | 0 | 0 | 0 |
| Green | 1 | 1 | 0 |
| SNP | 29 | 29 | 0 |
| Plaid | 4 | 4 | 0 |
The figures show that the Liberal Democrats gain two more seats (and are the runner-up in about half-a-dozen more seats). Labour gain five seats, but the Conservatives lose 7 seats more.
From January 2024 and onwards, the Electoral Calculus prediction model will be based on a regression which has the option to include neighbourhood characteristics. We hope that this will lead to more accurate predictions for the smaller parties.