Elections Etc

Polls-based forecast for the 2017 British general election

Advertisements

by Stephen Fisher and Josh Goldenberg.

Our central forecast for the result of today’s general election is as follows.

  GB share of the vote Seats 90% prediction intervals
Conservative 44 349 318-385
Labour 34.5 223 192-252
Liberal Democrat 9 9 3-15
SNP 4 47 39-53
PC 1 3  
UKIP 4 0  
Green 2.5 1  
 

This implies a Conservative vote share lead over Labour of 9.5 points and a majority of 48.  This represents only a modest improvement over the party’s performance 2015 when they achieved a majority of 12 with a 6.6 point lead.

From this central forecast, our estimated probability of a Conservative majority is 87%. Our analysis gives just a 1% chance of the Conservatives winning a 100+ landslide majority.

This forecast is based on the average of the last two published opinion polls by Kantar adjusted by +1 for the Conservatives and -1 for Labour. The reason for this is we are most persuaded by the Kantar methodology, especially regarding full-population weighting of 2015 non-voters and their 2017 turnout model. Other pollsters appear to weight those who did not vote in 2015 down to a lower share of the population than they really constitute. This creates the danger of suppressing turnout differentials. Weighting the vote-intention sample to the pattern of turnout in 2015 might resolve this problem somewhat, but only with respect to some measured characteristics (like age) and it risks going wrong if the pattern of turnout changes (as it did in the EU referendum and President Trump’s election).

We use the average of Kantar’s last two polls because the other pollsters are not showing much consistent change over the last week and because Kantar’s methodology involves up-weighting a relatively volatile group of former non-voters with a relatively small raw sample size.  By averaging two polls we hope that the estimate will be better by reducing sampling variation.

Averaging over more polls and benchmarking them to Kantar’s “house effect” would be another approach, but we do not think that the house effects have been sufficiently stable.

We apply an adjustment to reflect historical polling error because our view is that the problems of the polls in 2015 and other previous elections can only be partly solved by better weighting and adjustments (in effect we are assuming roughly 60% of the historic bias has been solved). Even after the adjustment, our forecast shares of the vote are still well within the range of predicted vote shares for each party in the final polls. Our estimated Conservative lead is towards the high end of the distribution.

The seats based forecast is based on constituency-level regression analysis of kindly provided opinion poll data within England and Wales and uniform change within Scotland from the average of the last four Scottish polls.

The regression analyses suggest that the Conservatives will do better in areas where there was a relatively large UKIP vote in 2015 and Leave vote in 2016. Labour change is relatively flat but higher in poorer areas with more young people. The Liberal Democrats appear to be doing slightly less well in such areas, but better where there are more graduates. UKIP meanwhile appear to be losing most heavily where they started strongest, and in less prosperous areas where there are relatively more pensioners.

While this variation affects the character of the seats most likely to change hands, our central seats forecast is not appreciably different from that which a uniform change change projection (for Scotland separately from England and Wales) would yield.

The uncertainty estimates are generated by simulating election results from a multivariate normal distribution. At each stage random error is added to our central constituency predictions and a new seats tally is generated. The average error for each party is zero. The variance and correlation between the random errors is informed by the historical pattern of errors of polls published within a week of election day for elections since 1992.

We will explain the choice of methodology in further detail in due course. It took quite some time to settle on, and we are by no means sure that we have found an ideal approach!

Acknowledgements

We are grateful to various people for data and helpful conversations that was used to inform this forecast, including (in alpha order) John Curtice, Patrick English, Rob Ford, Chris Hanretty, Jouni Kuha, Eilidh Macfarlane, Jon Mellon and others. They in no way endorse this forecast. We are especially grateful to pollsters who provided individual or constituency level data.

Advertisements