More in Common's MRP model released today predicts Labour will win 382 seats in the General Election on 4 July - a majority of 114, while the Conservatives are expected to hold just 180 seats.
Luke Tryl, Executive Director of More in Common UK, said:
"While many things could change between now and 4 July, Labour is on course to win a comfortable majority, with the most Labour gains in a single election since 1945, nearly doubling their seat count compared to 2019.
The Conservatives on the other hand are forecast to enter opposition holding only marginally more seats than they did after the 1997 landslide, suggesting a steep path to recovery.
It’s worth noting however, that there are currently 43 seats that the Conservatives are projected to lose by less than 4 points, which means any tightening in the race could see a much smaller Labour majority. There are also 49 seats which the Conservatives are currently projected to hang on to by just 4%, which means it wouldn’t take much movement for the Conservatives to head to a record defeat, potentially far worse than that seen in 1997.”
Key findings
- More in Common's MRP projects Labour will win between 377 and 402 seats
- The Labour Party is on course to take back their heartlands, while the Conservatives could see their number of MPs fall to its lowest level since 2005
- Four cabinet ministers projected to lose their seats
- Labour expected to regain some ground in Scotland, but voter distribution still gives the Scottish National Party an edge
- Conservatives set to be nearly wiped out in Wales, holding onto only one seat
- 117 seats projected to have a majority of less than 4 points
The MRP model is based on voting intention data collected between 9 April and 29 May from 15,089 adults in Great Britain.
Labour on course for healthy majority
Based on this model, Labour would receive 43% of the total vote share, and 59% of the 632 Parliamentary constituencies in Great Britain, while the Conservatives would win 28% of the total seats with 29% of the vote. The Liberal Democrats would make gains to reach 30 seats, while the SNP would suffer losses.
Big Names at Risk
The MRP projects the following Cabinet Ministers are set to lose their seats:
- Justice Secretary, Alex Chalk would lose his Cheltenham seat to the Liberal Democrats.
- Defence Secretary, Grant Shapps would lose Welwyn Hatfield to Labour.
- Leader of the House, Penny Mordaunt would lose her seat of Portsmouth North to Labour.
- Welsh Secretary David Davies would lose Monmouthshire to Labour.
Over a hundred seats currently too close to call
In 117 seats, the second candidate’s vote share is 4 points or less behind the winner. Accounting for uncertainty within the model, these races can be thought of as currently too close to call.
14 SNP-Labour marginals in Scotland
If Labour win all the seats where they are currently within 4% of the SNP, they will win 28 seats in Scotland.
Range of outcomes possible, but a Labour majority in each
In 500 simulations of possible General Election outcomes using the results of the MRP, Labour wins between 377 and 402 seats.
The best scenario for the Conservatives sees them winning 187 seats, and the worst scenario gives them 168 seats. While these are possible values and should not be thought of as predictions, they highlight the limited set of possible outcomes facing the Conservative Party.
FAQ
What is an MRP?
‘Multilevel Regression with Post-stratification’ (MRP) uses data from a voting intention poll to model how people will vote based on their demographics, voting behaviour and information about their constituency. These results are then applied to the demographic and electoral makeup of each constituency to make a constituency-level prediction. The model is 'multilevel' because it uses both individual and constituency-level data.
How is this different from your normal voting intention poll?
The voting intention regularly published by More in Common is a national estimate based on a nationally-representative sample of at least 2,000 people. It indicates roughly how many people in Great Britain intend to vote for one party or another. This is simple to calculate and allows us to track changes through time.
But if you want to project a national seat count, this isn’t as useful. If you applied the headline figure to every seat, Labour would win 632 constituencies in Great Britain and other parties would win zero - but of course this won’t happen. You can instead look at the ‘swing’ between the voting intention today and the vote share at the last General Election. If you assume every constituency will have the same swing you get a better estimate - around 390 seats for Labour and 180 seats for the Conservatives. But swings don’t tend to be uniform - different constituencies are different. MRP is one way of trying to take this into account. The model uses census information about the people who live in every constituency across the country to make seat-level projections. We add these up for a national count.
How does the model account for those who don't know how they will vote?
When we ask people their voting intention some people say they don’t know. We push them to say who they would vote for if they were forced to choose, and we use this response as their expected vote. Some people, when asked to imagine that they were forced to choose, still don’t know who they would vote for. Normally, when estimating vote share on a national level, we exclude these responses - effectively estimating that if this group of ‘double don’t knows’ ultimately votes, they would do so in proportion with those who have expressed a choice. That means even if someone lives in a rural area, is over 75, and voted Conservative in 2019 - if they respond they don’t know who they’ll vote for - the assumption is that they are the average voter i.e. more likely to vote Labour than Conservative.
Using our MRP model, we’re able to make a better guess. When training the model to predict people’s voting intention based on their demographics, voting behaviour and information about their constituency, we excluded the responses of people who didn’t know who they would vote for (after the squeeze) from the training data. When we apply the model to all the voters in the constituency, it effectively means we estimate the votes of people who don’t know, according to how people like them (in terms of demographics and past voting behaviour) but who do know, intend to vote. So if someone lives in a rural area, is over 75 and voted Conservative in 2019, the model uses the fact that most over 75s in rural areas who voted Conservative in 2019 and do know who they’ll vote for say they will vote Conservative, to guess that if they do vote it will likely be for the Conservatives.
How does the model account for who will vote?
When we survey the public about who they will vote for, many of the responses come from people who will not vote in the General Election. Since past election turnout is one of the strongest and most consistent predictors of future turnout, we trained the model on people who voted in the last election. If they were too young to vote at the last election, they are still included if they say they are certain to vote.
Is this a snapshot or a projection?
Some voting intention models attempt to take a snapshot of who the public would vote for if an election was called today. So they don’t factor in things that could affect election results on a future date, such as which parties stand in different seats, who might vote tactically, and who might make up their mind during the campaign. Others take factors like these into account - a projection of the General Election rather than a snapshot of today. This introduces uncertainty and requires making assumptions, but may give a clearer picture of how the public is likely to vote, based on where we are today. That is why in our model we have included historical information - such as who actually voted in the last General Election - as well as current information - such as who people say is their second choice, or whether they’d vote differently if their candidate was unlikely to win.
Is this your final model?
A lot could change between now and 4th July! That’s why we will continue to inform our model with fresh snapshots throughout the campaign and update the projections at least twice between now and election day. We will update the model assumptions if circumstances change during the campaign in a way that suggests we should do.
How do you deal with smaller parties?
MRPs can be less effective at predicting the seat distribution for smaller parties. To address this, we have taken into account additional information about historic vote efficiency, local presence and the relationship between opinion polling performance and realised election results. In our model this has the effect of boosting the Liberal Democrat seat total and reducing the implied level of national support for Reform UK.
When creating any standard voting intention or MRP projection, it’s necessary to make subjective choices about which variables to include and what to adjust. Not making adjustments or including relevant information is one such choice. We have instead tried to include fair assumptions to better represent the true picture.
Why does the model show X party winning in Y constituency?
MRP models are very good at indicating how the parties might perform across different constituencies based on their demographic makeup. But they don’t account for local factors that impact a small number of constituencies, such as a popular incumbent, well known or controversial council policy Therefore it would be a mistake to draw too much from the projected vote share in an individual constituency.
Why is your fieldwork range so large?
To train a model to predict someone’s voting intention requires lots of information about lots of people. We surveyed over 10,000 people in April and used this data to develop our baseline model. Since the General Election date was announced, we have surveyed over 5,000 more people about their current voting intention to take into account the changed circumstances. The data is weighted by recency, so that the newer data influences the model more than the older data.