Blog/Polling

On a seat-by-seat basis, these final estimates are quite similar to our previous release last week. Across the 5,048 constituency cells (i.e. 631 constituencies * 8 parties) there are less than 6% of occurrences where the vote share estimates show a change of 3 percentage points or more. 11% of our estimates show a change of 2 points or more.

Looking at the changes in more detail, we find that: 

1. The most likely winner has changed in 30 seats since our last forecast. A large majority of these constituencies are highly marginal, with margins of victory as razor-thin as 0.003pp (Lichfield). However, eight constituencies have changed hands with non-trivial vote share leads since our last forecast. These include Bicester and Woodstock; Chesham and Amersham; Chippingham – all now predicted to be Liberal Democrat pickups from the Conservatives; Alloa and Grangemouth, and East Renfrewshire – Labour gains from the SNP; Sittingbourne and Sheppey, and Wyre Forest – Labour pickups off the Conservatives; and the SNP is now forecast to keep hold of Kilmarnock and Loudoun.
2. Reform UK are ahead in 2 seats (Ashfield and Boston and Skegness) instead of the single seat we forecast last week. Clacton is too close to call with 1.6pp between Conservatives and Reform UK. We see Reform UK reaching a 16% national vote share (+0.5pp).
3. The Liberal Democrats are forecast to win an additional 7 seats since our last model, including Witney, Stratford-on-Avon and South Cotswolds. 

Headline Findings

1. Focaldata projects a Labour majority of 238 seats based on a probability forecast. This takes into account the likelihood of parties winning in each individual seat, with total seat counts calculated as a sum of the probabilities across all seats.
2. The upper and lower range estimates for each party are:some text
   1. Labour on 444 (with a lower range of 433 and upper range of 456)
   2. Conservative on 108 (with a lower range of 94 and upper range of 123)
   3. Liberal Democrats on 57 (with a lower range of 51 and upper range of 63)
   4. Scottish National Party on 15 (with a lower range of 12 and upper range of 19)
   5. Plaid Cymru on 2 (with a lower range of 2 and upper range of 3)
   6. Green Party on 1 (with a lower range of 1 and upper range of 2)
   7. Reform UK on 2 (with a lower range of 1 and upper range of 4)
3. The implied national vote shares in Great Britain are Labour on 40.0%, Conservatives on 23.2%, Reform UK on 16.0%, Liberal Democrats on 11.5%, Green on 5.4%, SNP on 2.6%, Plaid Cymru 0.5% and others on 0.9%.
4. We find that 111 seats are ‘marginal’ (+2 since our last call), a term we use throughout to denote a seat in which the difference between the first and second largest parties by vote share is less than 5 percentage points. Marginality pinches the Conservatives the most, with the party defending a total 33 seats with a <0.05 margin (just under a third of its projected total). The number of Conservative marginal wins is down from 52 at our last forecast however, suggesting that our prediction is better representing the ‘floor’ of the Conservative seat share.  
5. Most notably, our model projects 34 seats in which Reform UK achieve over a quarter (25%) of the vote share (+6 since our last forecast), and 2 pickups. 
6. Looking at specific MPs, our model projects losses for senior Conservative MPs including Jonny Mercer (losing Plymouth Moor view to Labour on a 21-point swing), Grant Shapps (losing to Labour on an 18-point swing in Welwyn Hatfield) and David TC Davies (losing Monmouthshire to Labour on a 17-point swing). Both Jeremy Hunt (Godalming and Ash) and Alex Chalk (Cheltenham) are due to lose their seats to the Liberal Democrats. Penny Mordaunt keeps her seat in Portsmouth North by 0.2pp as does James Cleverly in Braintree by 0.05pp. Maidenhead, Theresa May’s former constituency, is currently a LibDem/Tory toss-up with a 0.8pp margin. Rishi Sunak safely keeps his Richmond and Northallerton seat despite a projected 17-point swing against the Conservatives with a 13-point margin of victory.

‍Data for all constituencies is available here. 

How might we be wrong? 

#1 Unwinding adjustment is too aggressive 

We deploy an adjustment – which YouGov coined as “unwinding”. This adjustment seeks to cater for some of the historical problems with MRP relating to regularisation and attenuation. We think it’s best to be really transparent about how material this adjustment is. Our model learns from past election distributions to suppress MRP tendencies to move beyond proportional swing. Our adjustment results in significant change in seat count of around 40 seats for the Conservatives, meaning that without it, we would be towards the lower end of MRP forecasts for the Conservatives. For transparency, our ‘raw’ seat counts (i.e. without the unwinding algorithm) would be Labour on 476, Conservatives on 68, Liberal Democrats on 62, SNP on 17, Reform UK on 4, Green on 2 and Plaid Cymru on 2.

Our political judgement remains that beyond proportional swing is unlikely – but we accept entirely that we don’t know where the “slope” is likely to be between uniform national swing and proportional. The evidence from the 2024 local elections was firmly towards (if not entirely) proportional. Please see a long-form discussion of this here.

However, even with unwinding, looking at the marginality of seats, the Conservatives only hold 65 seats with a margin of over 5 percentage points, and are acutely vulnerable to changes in vote share up to election day. Labour by contrast lead in 398 seats by a margin of over 5 points. This is a staggering difference. To clarify, we are not forecasting c.60 seats for the Conservatives but it is entirely reasonable to assume this could happen if our unwinding adjustment has been too aggressive.

#2 Fieldwork is too late

The other potential source of error is our fieldwork dates. Interviewing people from the 10 June – 1 July means we may miss the latest political movements.

#3 Tactical voting 

We have tactical voting effects in the model, but acknowledge that there is great uncertainty over the Liberal Democrat specific seat gains. 

#4 Change in turnout patterns

Our validated turnout model uses historic voting patterns going back to 2015. Should the turnout patterns of the electorate change in a material and drastic manner - particularly around youth turnout and older voter abstention we could be overestimating the Conservative vote.

#5 Minor party vote distribution patterns 

MRP can struggle with minor parties - and particularly can suppress the vote shares of smaller parties at the top end of the scale. There is a strong possibility that our model may be doing this for Liberal Democrats, Greens and in particular for Reform. Should this be the case we can envisage a scenario where the Liberal Democrats exceed our 57-seat forecast, the Greens pick up seats and Reform ends up with a handful of 4-7 MPs. An extension of this is where there are prominent independents. Our model could underplay significant vote shares for independents running in more diverse constituencies challenging Labour from the left.

FAQs

What is multilevel regression with post-stratification (MRP)?

Multi-level regression with post-stratification (MRP) is a statistical technique for estimating public opinion in small geographic areas or sub-groups using national opinion surveys. It originated in America, and was used by academics to estimate state-level opinion cheaply, given the expense of doing polls throughout the country. 

MRP has two main elements. The first is to use a survey to build a multi-level regression model that predicts opinion (or any quantity of interest) from certain variables, normally demographics. The second is to weight (post-stratify) your results by the relevant population frequency, to get population level (or constituency level) estimates.

At the end of this process the aim is to get more accurate, more granular (thus more actionable) estimates of public opinion than traditional polling. There are, however, significant technical challenges to implementing it effectively. These include large data requirements, dedicated cloud computing resources, and an understanding of Bayesian statistics.

Why are everyone’s forecasts so different?

Thankfully we cover that here in a long-form essay. TL/DR is that every pollster has different vote shares, and the distribution of each party’s vote share - specifically its slope is drastically different between pollsters

Who built the model? 

This new UK model was built by Dr Ben Lobo and Dr Adam Higgins. Our modelers received technical support from Dr Pete Logg who led our successful 2019 UK MRP efforts, and Dr Matt Chennells and James Alster and the wider technical team, and domain support from our Chief Executive Justin Ibbett and Chief Research Officer James Kanagasooriam.

Background to the Focaldata / Prolific MRP

Focaldata was set up to provide high quality and rapid turnaround MRP in 2017. Since then we have conducted thousands of models for hundreds of clients - corporate and political ones. For this election we built up an entirely new MRP model vs our 2019 model - which strongly performed and captured the dynamics of the race and voting shapes efficiently of each party. For this election we have partnered with the firm Prolific to supplement our sample to provide our model with the largest possible sample in a restrained time period given campaign vote share change -  to forecast the general election. 

The component parts of our UK 2024 “MRP” model was built from scratch include:

1. A new post-stratification frame built from the England & Wales 2021 Census and recent ONS data incorporating interlocking cells for [Westminster constituency, education level, ethnicity, housing tenure, 2019 voting intention and age]. Scottish post-stratification frame pertained to 2011 census data, and was adjusted through proportional fitting to match more recent data (The “P” part of MRP). The frame uses the new constituency boundaries that came into force this election
2. We estimate a multi-level regression model where individual respondents are nested within constituencies, which in turn are nested within synthetic regions. We include individual-level characteristics such as age and past vote as random effects. We also include numerous fixed effects such as constituency-level party vote share or constituency population density. 
3. We have deployed an “Unwinding” adjustment to the forecasts which uses prior election vote distributions so that forecasts don’t forecast beyond proportional swing. We have written a detailed blog here on this. This adjustment makes no difference to national vote shares implied by MRP, and correlates to - but does not use - the vote distributions observed at the recent 2024 local elections
4. New Turnout model which uses the British Election Study 2019 random probability post-election survey which has a validated turnout field; in other words a survey where we know that someone has voted, not just said they would vote. We have adapted our turnout model to bring in this new data, but also include turnout patterns from 2015 and 2017 elections respectively 

Other notes about the model

This MRP uses ‘Rallings and Thrasher’ 2019 notional results parliamentary boundaries. 

Technical notes

Focaldata interviewed 36,726 British adults along with our survey partners Prolific from 10 June to 1 July 2024

We provide estimates of each party vote share in each constituency. The estimates show point estimates along with credible intervals i.e. the high and low estimates. To calculate these, ​following estimation of the multilevel model, we draw 500 samples from the posterior, using the poststratification frame as new data. We then calculate the median, low (5%) and high (95%) confidence intervals. The intervals indicate that according to our model, there is a 90% chance of the outcome lying between the low and high estimate - while the point estimate is the median value across all our 500 draws. 

For each constituency, we provide probabilities of a party winning a given seat. These probabilities are the percentage of the 500 draws that the party wins a seat. Our final seat count is calculated by summing the probabilities for each party across all constituencies. This means our seat count takes into account the uncertainty around our estimates for each party in each constituency. 

We include a time parameter in our model so that our estimates take into account any temporal changes in vote choice and ensure that our estimates are weighted to the most recent time period.

About Focaldata

We're on a mission to power the world’s understanding of what people think and do.

We research public opinion for data-driven organisations. With Focaldata, decision makers and researchers get the most rigorous data and analysis at 5x the speed — reducing the time to decision while delivering uncommonly actionable insights. We are a team of engineers, data scientists, product specialists and researchers building outstanding technology and next-generation services. Our services are global. We are non-political and non-ideological. 

https://www.focaldata.com

About Prolific

Prolific is a platform that enables researchers to collect high-quality human-powered data at scale.

Using the Prolific platform researchers can target, contact and manage research participants from Prolific’s diverse, vetted and fairly-treated pool – to deliver world-changing research and the next generation of AI.

www.prolific.com/polling