Why you should take YouGov's MRP with a pinch of salt

It's just one poll - and its existence could, in of itself, change the result.

NS

Sign Up

Get the New Statesman's Morning Call email.

YouGov has released its detailed model of what it thinks the 2019 election result will be. A technique called multilevel regression and post-stratification (MRP) it allows you to use polls to predict what that means at a constituency level. Here's what it shows: the Conservatives on 359 seats, up from 317, Labour on 211, down from 262, the SNP up seven to 43, and the Liberal Democrats doubling their vote - to get just one extra seat than in 2017, despite that election being their worst electoral performance in their history.

At the last election, YouGov’s MRP called the result broadly correct, showing a hung parliament, though it underestimated the Conservatives (showing them on 302 seats, which would seen them lose office, not 318 as they ultimately got), to the benefit of Labour and the SNP (who it projected getting 269 and 44 seats not 262 and 35), while underestimating Plaid Cymru (who got four not two seats). The exit poll – the detailed survey of voters conducted by the broadcasters – also slightly underestimated the Conservatives, but to a lesser extent.

At the last election, the MRP model was treated with perhaps excessive scepticism: it showed us broadly what we’d expect given the direction of travel in the polls. This time, I suspect that it will be given the opposite treatment, and viewed almost as holy writ. We shouldn’t treat an MRP any differently to how we should any other form of poll. YouGov’s MRP may be right – Survation, who using conventional techniques also got the 2017 election right within the margin of error, may be right. IpsosMori, who got the 2019 European elections may be right. Kantar, who got both wrong, may be right. We can’t say with confidence.

Remember that our electoral system means that the results will be decided in marginal constituencies which are just that, marginal. If I can build a model that accurately tells you from the behaviour of less than a 100 people that Diane Abbott will be re-elected with say, 67 per cent of the vote in Hackney North and Stoke Newington, and she actually gets 65 or 71 per cent, that’s still pretty impressive – I’ve managed to predict the behaviour of 70,000 people in a very small  geographic area using a yet smaller sub-sample. But underestimating the Labour vote by even one per cent would have changed the result in Southampton Itchen, Richmond Park, Stirling, St Ives, Pudsey, Hastings and Rye, Chipping Barnet, Thurrock, Preseli Pembrokeshire, Calder Valley, Norwich North, Broxtowe, Stoke-On-Trent South, Telford and Bolton West, and with it put Jeremy Corbyn in Downing Street.  Overestimating it by even one per cent would likewise have changed the result in eight constituencies and given Theresa May a majority.

In addition, the very act of publishing YouGov’s MRP model may change the result.  We are reasonably confident that telling Conservative Remainers the result is uncertain changes how they behave – that is the view of both Conservatives and Liberal Democrats. Several Tories in the Conservative-Liberal Democrat battleground have told me that what they wanted more than anything was for YouGov’s MRP to show a hung parliament – and with their seat tipping the balance. They credit the personalised letters from David Cameron in 2015 warning of a Labour-SNP government with their victories in 2015, and think that the seemingly distant prospect of a Corbyn government in 2017 is what cost their fellow 2015 intake MPs in Bath, Oxford West, Eastbourne, Kingston and Surbiton and Twickenham.

So all in all, while tonight’s MRP might be right, it might not be: and even if it is right, its very existence might change the results.

Stephen Bush is political editor of the New Statesman. His daily briefing, Morning Call, provides a quick and essential guide to domestic and global politics.