#GE2017 Forecast Methodology

This page covers the methodology for my 2017 General Election forecast. The model predicted the wrong result - eventually calling a small but significant workable majority for the Conservatives. However, of the 'losers', this model performed best of all published and made public in advance of the night itself.

The PME Politics Forecasting Model used a 5-day rolling average of polling data to first produce a uniform swing model, which is reported at the beginning of each article.

However, it is well known that in contemporary British Politics, while uniform swing is able to predict much the General Election story, it misses some of the detailed peculiarities of the constituency based system employed in UK elections.

As such, the PME Politics Forecasting Model used a ‘polls plus’ model as its main forecast.

The ‘polls plus’ model attempted to capture several distinct stories that should be in play during the 2017 General Election. Each are assumptions based on previous election outcomes (including local and regional), polling evidence, and theory. As such, the model is empirically grounded but does rely on a range of assumptions (which will not necessarily hold from one election to the next).

The 'polls plus' model factored in separate information on Welsh and Scottish swings where polls are available, and uses the British Election Study 2015 General Election database (BES 2016) for contextual factors. Alterations to swing in each seat (if applicable) were applied (summatively) in the order that follows.

The first four were assumptions which can be used in any General Election:

  1. Swing will dampen as we move closer to election time (see Elections Etc. 2015). When the election draws nearer, it is noted that party support tends to drift somewhat back towards previous levels. The model thus accounts for this by reducing the overall swing to/away from each party. 
  2. Swing tends to be larger when seats are neither too marginal nor too safe, but somewhere in the middle (see Barkovic-Parsons, Hodgdon, and Maloney 2017). The model thus redistributes uniform swing to be slightly higher in semi-marginal to semi-safe seats, and less in more marginal and more safe seats (all others). 
  3. Incumbent MPs will enjoy a slight advantage in terms of swing compared to new candidates contesting 'open seats' (where the incumbent MP has stepped down and is not seeking re-election) (see Gelman and King 1990; Katz and King 1999). When a seat is recorded as open, swing against the defending party and toward the challengers is increased by 2 points
  4. Polling consistently underestimates Conservative Party performances. In all but one election (1983), the polls overstated Labour support and understated Conservative support. Though after the big polling miss of 2015 an enquiry was launched which made several recommendations for pollsters regarding trying to improve this, but we have no evidence as of yet to suggest that this has been effective. As such, the model slightly adjusts the predicted swing to give a slight boost to the Conservatives from the Labour Party.

The following four were then specific to that particular election, and as such will not (necessarily) be replicable in following contests. They were each my own assumptions, and therefore entirely open to debate and discussion:
  1. Labour Students: the Labour Party swing will be much different in ‘student cities and towns’. It is well known that through Jeremy Corbyn’s leadership, the Labour Party have established a strong network of support among young people, and particularly students (as evidenced by this poll in early May - https://www.theguardian.com/politics/2017/may/04/voter-registration-soars-students-backing-labour-corbyn-general-election?CMP=twt_gu ) . Thus, it is reasonable to assume that in seats with a high density of students (according to Census data), we should expect Labour to do better than their average nationwide (and for the Conservative swing to be not so large). 
  2. Brexit swings: swing to the Conservatives will be higher where ‘Leave’ voting in the referendum was also high, and Liberal Democrat vote share higher in high ‘Remain’ voting seats (according to Chris Hanretty’s (2016) estimates). The Conservatives are very much trying to steal UKIP’s clothes regarding being ‘the party of Brexit’. Theresa May’s “Brexit means Brexit” promise is likely to resonate highly with ‘leave’ voters, and so she should attract more support in areas in which the majority of those who turned out on June 23rd opted to vote leave. On the flip side, the Liberal Democrats have successfully modelled themselves as an ‘anti-Tory Brexit’ party and are attempting to hoover up Remain voter support while Labour dwindle between a rock and a hard place on Brexit policy. That strategy worked in Richmond Park, so we can assume it will have some impact on voting in the General Election. 
  3. Lib Dem fight-back: the Liberal Democrats will do particularly well in seats where they won in 2010 but lost in 2015. The model will assume that the 'Lib Dem fightback' will happen, to an extent, but that their vote share will be moving upwards mostly in constituencies where they won in 2010, but lost in 2015. We often speak about the popularity and high levels of organisation of the Liberal Democrats at local levels, and the model assumes that this impact will result in a stronger swing towards the Lib Dems in areas they represented before 2015 (think Twickenham, Bermondsey and Old Southwark, Colchester, and so on). 
  4. Progressive Alliance (and UKIP Brexiteer withdrawal): though the tactical voting element of this phenomenon won't be directly measured, if any party exits a contest and instructs its voters to tick an alternative box, then the model will account for that. This is happening particularly with the Greens and Labour, and the model also will account for areas where UKIP are standing down in favour of whom they label "life-long Brexiteers". The redistribution of votes from UKIP to other parties reflects recent BES analysis by Chris Hanretty.
The model then calculated the predict result in each constituency using the adjusted uniform swing, and calculated for a final vote share in each seat. Winners were then identified in each contest, and the total number of seat wins summed and displayed for a final result.


References:

Barkovic-Parsons, C., Hodgson, R. & Maloney, J. (2017) “Are Marginals Different? Evidence from British Elections 1950-2010”, Public Choice, doi:10.1007/s11127-017-0438-8
British Election Study (2016) 2015 Constituency Results, Version 2.2., DOI: 10.13140/RG.2.1.1162.1844
Fisher, S. and Jones, J. (2015) Methods, Elections Etc., https://electionsetc.com/methods/
Gelman, A. & King, G., (1990) Estimating incumbency advantage without bias. American Journal of Political Science, pp.1142-1164.
Hanretty, C. (2016) Revised estimates of Leave vote share in Westminster constituencies, available online: https://docs.google.com/spreadsheets/d/1wTK5dV2_YjCMsUYlwg0l48uWWf44sKgG8uFVMv5OWlA/edit?usp=sharing
Katz, J. N., & King, G. (1999). A statistical model for multiparty electoral data. American Political Science Review,  pp.15-32.

No comments:

Post a Comment