Don’t believe that Labour landslide poll

Focaldata’s new seat-by-seat projection predicts a Labour majority of 290—but that’s unlikely

June 08, 2023
Image: Nicola Ferrari RF / Alamy Stock Photo
A recent poll predicted Labour winning 470 seats. Image: Nicola Ferrari RF / Alamy Stock Photo

Watch out Rishi Sunak. I’m sure you feel safe in your local constituency, defending a majority of 24,000, under the new boundaries. But according to the latest seat-by-seat projection by Focaldata, Labour has a real chance of kicking you out of Richmond & Northallerton. Your projected majority is barely 1,000—well within the margin of error. 

Still, you are better placed than Tories standing in other supposedly safe seats. Aldridge-Brownhills, Cannock Chase and Basildon & Billericay (each with a majority of around 20,000) would all fall to Labour. Penny Mordaunt’s 16,000 majority in Portsmouth North would be blown away on a 21 per cent swing. Even Liz Truss’s 24,000 majority in South West Norfolk would be too little to save her from defeat. 

I don’t believe it. To see why, let’s look at what Focaldata has done. It has conducted a large national survey of more than 10,000. Its overall figures are in line with other recent polls. Among those who name a party, 45 per cent would vote Labour, 30 per cent Conservative and 9 per cent Liberal Democrat. I do not quarrel with these figures.

The problem is what Focaldata has done with its data. It did not perform the traditional calculation of assuming uniform national swing (UNS). This would have taken the national change in party vote shares since the last general election (Conservatives down 15 percentage points, Labour up 12, Liberal Democrat down 3) and applied them to each constituency. In reality, of course, the swing to Labour would be larger in some seats and smaller in others. But historically, these variations have tended to cancel each other out, so UNS has generally provided a fairly good guide to the overall numbers of seats won by each party.

If we do this calculation with Focaldata’s overall figures, Labour would end up with a majority of 68. There would be 359 Labour MPs, 216 Conservatives, 33 SNP, 22 Lib Dems and 20 others, mainly Northern Ireland. (As with Focaldata’s projections, all the figures cited in this analysis are based on the new boundaries likely to apply at the coming election.) Sunak, Mordaunt and Truss would see their majorities cut, but not wiped out.

Focaldata’s figures are very different. They project a Labour majority of as much as 290. Labour would win 470 seats, the Tories 129, SNP 26, Lib Dems 5, Others 20.

How come? The answer lies with the method Focaldata uses to estimate the figures for each seat. It’s known as MRP (short for multilevel regression and poststratification). I recently wrote a detailed blog dissecting MRP. The method requires some heavy-duty computer calculations, but in my view the method has a fundamental flaw. For reasons that my blog sets out, the effect of MRP is to cause parties that are out of favour to lose most votes in seats where they are strongest, and lose fewest in seats where they are weakest.

Under UNS, we would expect the Conservative vote to fall on average by around 15 points in almost all kinds of seat—safe Tory, marginal, or Labour. (“Almost” because, plainly, if the Tories start locally with, say 10 per cent, their vote share cannot fall by 15 points. However, MRP also hits trouble in such seats. The latest Focaldata projection shows the Conservatives implausibly gaining support in Liverpool and parts of inner London, even as their voters flock to other parties everywhere else.) What really matters is how the rival models perform in seats where the party losing ground starts off with more than 25 per cent—in other words, to any seat that might change hands in an election. 

Let’s apply the rival models to Sunak’s seat. On the new boundaries, he starts off with his 64 per cent share in 2019, against 16 per cent for Labour. If we applied UNS to Focaldata’s national figures for changes in party support, he would expect to win 49 per cent just now (down 15 points), while Labour would climb 12 points to 28 per cent. This would give Sunak a majority of around 11,000—much less than last time, but nowhere near squeaky-bum territory. 

However, Focaldata reckons his share would be down by 25 points to 39 per cent, while Labour would more than double its vote to 37 per cent (up 21). Richmond & Northallerton would be a toss-up, possibly giving us the novel experience of the prime minister enduring a recount on election night.

If Focaldata is right (and recent MRP surveys by other companies have produced similarly massive falls in previously safe Tory seats), then something astonishing has happened to the way Britons vote in elections. The pattern of swing will have changed completely. Election-night swingometers in TV studios won’t work. Most textbooks on the way voters behave will have to be ripped up. 

Could this be the case?

There are bits of evidence that might seem to point that way. Historically, UNS has not always applied to midterm polls and by-elections. It also went awol in two specific groups in 2015, when Labour’s support collapsed in Scotland and the Lib Dems’ support collapsed across Britain. In those cases, out-of-favour parties did indeed lose most votes in their strongest seats, as MRP would predict. But these were exceptions to the wider truth that has applied throughout the modern era. So far, UNS has reasserted itself at every general election in the main two-party battles across Britain between Labour and Conservatives.

The central problem with MRP is that, by its nature it assumes that an out-of-favour party will lose most support in its safest seats. If, as the next election approaches, the pattern of party support reverts to UNS, as it always has in the past, then MRP will be in trouble. A significant swing to Labour will indeed make Keir Starmer prime minister. But without big changes to its algorithms, MRP will exaggerate the number of seats the Tories lose. And the bigger the swing, the wider the gulf between MRP and reality.