New Model Features 2017

This page first posted 23 May 2017

Several new model features have been added to the main model for the 2017 election. This features were developed in response to several novel features of the election. There are four main features:

This article describes the model features that have been developed for these four features, and assesses their likely impact on the final result.

1. Candidates

In a number of seats, some of the non-Conservative parties have announced that they will not field a candidate, in order to increase the chance of beating the Conservatives. But also UKIP is not standing in many seats, which helps the Conservatives.

Electoral Calculus has updated its main prediction model to allow for this new development. Released on 8 May 2017, this new version of the model transfers votes from non-standing parties to the parties which would be expected to benefit from tactical voting. This modelling applies to both the headline prediction as well as the user-defined predictions.

All GB seats have now been updated from the actual candidate lists which were finalised on 11 May 2017. The summary statistics are:

The transition matrix used was a mixture of estimates from polling (particularly by ICM) and reasonable estimates where there was no polling evidence. The matrix used is:

CONLABLIBUKIPGreenNATMINOTH
CON0%34%42%11%0%0%0%13%
LAB36%0%34%11%0%0%0%19%
LIB50%38%0%2%0%0%0%10%
UKIP75%13%7%0%0%0%0%5%
Green5%20%30%0%0%40%0%5%
NAT30%30%30%0%0%0%0%10%
MIN20%30%40%0%0%0%0%10%
OTH0%0%0%0%0%0%0%100%

Each row gives the distribution of defecting voters from that row's party. For instance, the UKIP row shows that 75% of defecting UKIP voters go to the Conservatives, 13% go to Labour, and so on. Rows should sum to 100%. Where several candidates are not standing, those columns are set to zero and rows are re-weighted to add up to 100%.

The average worth of this effect can be calculated in the following way. We take current opinion polls, as at 23 May 2017, and use them as a baseline. Ten additional scenarios are computed using a swing from Conservatives to Labour of 1%, 2%, 3% and so on up to 10%. The impact of the model change at each scenario is computed, and then the model impacts are averaged over the scenarios to get the final net effect. This averaging is performed because the discretization of seats causes the impact to vary from scenario to scenario.

The average impact of the candidate lists in terms of national seats predicted to be won is:

CONLABLIBNAT
+5−40−1

2. EU Referendum vote

Polling evidence from ICM shows that the voter migration of voters depends on whether they voted for leave or remain at EU Referendum in 2016. An average of three ICM polls for the Guardian (14-17 April, 18 April, and 21-24 April) gives the following transition matrices for Remain and Leave voters. The first table is for Remain voters:

REMAINConLabLibUKIPGreenOther
Con89%3%8%0%0%0%
Lab5%79%11%1%2%1%
Lib9%11%77%0%2%1%
Ukip46%17%8%29%0%0%

Each row represents the transition probabilities of someone who voted for that row's party in 2015 (and who voted Remain in 2016), and rows should add up to 100%. We see that Remain voters are fairly 'sticky', with 89% of Conservative Remain voters sticking with the Conservatives. Crucially, they are not defecting to the Liberal Democrats in significant numbers, against Lib Dem expectations. The UKIP row is not very meaningful as there were so few UKIP Remain voters.

The corresponding table for Leave voters is:

LEAVEConLabLibUKIPGreenOther
Con94%2%0%2%1%1%
Lab17%64%4%7%5%2%
Lib28%13%55%2%0%1%
Ukip33%5%3%58%0%1%

Here is a different story. Leave voters from all parties are defecting to the Conservatives. The UKIP to Conservative migration is particularly striking with about a third of UKIP voters changing sides.

We can also see these migrations in graphical form for respectively Remain:

and for Leave:

Given these transition matrices, we can apply them to constituencies. To do this, we need to know how people in each seat voted in both the 2015 election and 2016 referendum, and also how many voted for each combination, such as Con-Remain, Con-Leave, Lab-Remain, Lab-Leave and so on. This information is not directly available, but it can be inferred from election results and census demography in the same way as the seat-wise EU Referendum results and the political extremes.

The seat-wise votes for the election-referendum combinations can be added up to give an estimate of the national picture:

RemainLeaveTotal
Con15%23%38%
Lab18%13%31%
Lib6%2%8%
Ukip1%12%13%
Green4%0%4%
Other4%2%6%
Total48%52%100%

Most Conservative voters also voted for Leave, but most Labour voters chose to Remain.

The average net effect of the EU Referendum transition model is a modest gain for the Conservatives against Labour and the Lib Dems.

CONLABLIBNAT
+5−3−2−1

This model feature was the subject of the Guardian online article "Lib Dems shouldn't count on Remain votes" by Martin Robbins and Martin Baxter (27 April 2017).

3. Regional Swing

YouGov conducted their traditional large-scale survey of 14,395 adults across the country from 24 April to 5 May 2017. The full tables are available.

The advantage of this large-scale survey is that it gives a unique insight into the behaviour of each region of the country. The model already has an estimate, given the national opinion polls, for the party support in each region. The YouGov levels can be compared with the model's estimate, to get the relative difference between them.

The differences between YouGov levels less the model's estimates are shown in this table:

AreaCONLABLIBUKIPGreenNAT
Scotland0.0%0.0%0.0%0.0%0.0%0.0%
North East6.8%-2.9%-3.8%0.0%0.0%0.0%
North West3.4%-1.5%-1.5%-0.6%0.2%0.0%
Yorks/Humber2.0%0.4%-1.5%-0.7%0.0%0.0%
Wales6.5%-0.7%-2.4%-2.6%-0.5%-1.1%
West Midlands1.6%-3.4%0.3%1.5%0.2%0.0%
East Midlands3.1%-2.2%-0.7%-0.6%0.3%0.0%
Anglia-0.9%-1.7%0.5%1.2%-0.2%0.0%
South West-2.8%5.4%-2.4%-0.6%-0.4%0.0%
London-5.1%-1.9%3.8%2.0%0.2%0.0%
South East-3.4%2.0%2.4%-1.1%0.1%0.0%

Large values (outside +/-3%) are highlighted in bold face. The Scotland line is zero because the model already uses Scotland-wide polling directly.

If this poll is correct, we see the Conservatives gaining over Labour in the North, Wales and the Midlands. Conversely, Labour gains over the Conservatives in the South, with the Lib Dems doing well in London. The Conservative moves are more important because there are many marginal seats in the North and Midlands. Conservative majorities are typically so large in the South that the swing to Labour there does not change many seats.

Because this is just one poll, and there is a degree of uncertainty around it, it was given a weighting of 50%. This means that every cell in the matrix was multiplied by 0.50 before it was used.

On this basis, the net effect of the Regional Swing information is another gain to the Conservtives of around three seats:

CONLABLIBNAT
+3−4+10

4. Incumbency and Tactical Voting

Even with all these features, the model will not capture local effects such as particular local issues or the popularity of the incumbent MP. This is particularly important for smaller parties like the Liberal Democrats and the Greens. For example, there is evidence from constituency polling in Brighton Pavilion that many Labour voters are planning to vote tactically to back the Green party leader Caroline Lucas. Similarly, revious elections have shown that Liberal Democrat MPs can do better than national polling would indicate. (Though they are not invulnerable, as 2015 showed too.)

To allow for this, some manual estimates were made in affected Constituencies. This added an incumbency effect for Lib Dem MPs, and the tactical voting effect in Brighton Pavilion. This should make some seats a little more accurate, but does not have a material effect on the election result.

Summary of model changes

Taken together, these model changes represent an attempt to model some of the complexities of the 2017 election. It remains to be seen how effective they are in that objective. Let us also remember that it is not possible to predict every seat correctly, and several dozen seats will not go as predicted (even if the polls are correct). And if the polls are wrong, then the output of any model, however sophisticated, will also be wrong.

Using current polls, then the net effect of all these model features is the following:

CONLABLIBNAT
15−12−2−2

The Conservatives are the net beneficiaries of these changes for different reasons. At the moment, Brexit is a good issue for the Conservatives. Their supporters are staying with them, plus they are gaining Leave voters from the other parties. The collapse of UKIP is particuarly useful to them as well, especially in those many seats where UKIP is not even standing. If the YouGov poll is correct, then the Conservatives also benefit from a swing towards them in areas where they have been traditionally weaker.

Paradoxically, even though the Labour vote share is holding up well compared with 2015 (as at 23 May 2017), the Conservatives still seem on course for a large majority due to the collapse of UKIP, defection of Leave voters, and the technical factors described here.


Return to Psephology home page.