Through a Glass Darkly
Glen Newey
So here it is, the last throes of a campaign that began with Brexit and has ended with terrorism. In between, Theresa May’s foibles have blossomed like a suburban cannabis plantation under the arc-lights of scrutiny. For all the cosseting from her handlers, May has looked ever more frail and flailing as the campaign has worn on, which the terrorism brouhaha has barely concealed. May can scarcely moan about the personalisation of an election that she sought to fight on her own personality. Verbal and other tics obtrude. A pet tag is ‘I have been completely clear that’, which generally prefaces either content-free blather, or a false denial of having fudged or U-turned. Her angularity, faulty judgment and sheer want of imagination make her look like a beneficiary of the Peter principle.
Early in the campaign I wrote that Labour resembled a driverless car, but it's undeniable that Corbyn has grown in stature as it's gone on, to the chagrin of continuity Neo-Labs like Lord Mandelson. His hope was that Corbyn would fall to crushing defeat after a disastrous campaign. That, palpably, has failed to happen, whatever the result. Corbyn, or the movement that bears his name, is not going to evaporate between the close of the polls tonight and tomorrow morning. The Lib Dems' (pause for a minute's silence) impending failure casts doubt on the national appetite for a new Europhile centre party – though if they fail to recapture Labour, Mandy, Blair and the rest may try to start one.
Undaunted by previous forecasting flops, the Telegraph has been using a 'unique forecasting model' developed by UEA's Chris Hanretty, which 'combined with historical data and some very clever maths can give surprisingly accurate results'. It had a try-out in the 2015 election, and got ‘many things … exactly right – including the single seat each for Ukip and the Green Party'. Alas, it blooped spectacularly on the bigger picture, tipping a hung Parliament like everyone else. 'A model can be sophisticated, produce consistent and plausible estimates,’ Hanretty comments, ‘and still be wrong.' A more cautious approach seems favoured this time. Of Ukip, for instance, which had no MPs at the dissolution, Douglas Carswell having jumped ship, 'seat loss' is said to be 'moderately unlikely'.
Hanretty's model farms data from YouGov covering an average of 175 voters per constituency. That may sound quite a lot, but at the 95 per cent confidence level used in calculating the 'margin of error' on polls, a sample of that size leaves a confidence interval of around 7.5 per cent: in other words, a given figure for, say, Labour support in a certain constituency lies within a 15 per cent range. That is far too broad to be of any practical use in predicting outcomes, and not only with marginals. Polls with much larger sample sizes have been historically bad at predicting even the national shares of the vote.
So I peer into my crystal ball and see through a glass darkly. With little historical data and no clever maths, the figures below arise from having the following thoughts and then inventing a number. As judged by the best extant data, namely the returns from 2015, Labour's vote is liable to be less efficient in votes-per-seat terms than the Tories' (in 2015 Labour needed 40,290 votes for each MP, the Tories 34,243). Most polls, YouGov's outlier last week excepted, have Labour several points or more behind. When polls get it wrong, they usually do so by overstating Labour's support. Incumbency effects are usually more pronounced when alarms arise over security, as with the London and Manchester attacks. So I would put the Tories at between 365 and 375, which would give them a majority of 80 to 100. Other parties than Tory or Labour will probably get about 80 between them, mostly SNP. That would leave Labour with between 205 and 215.
If you gamble on these numbers and it goes wrong, then – as Nixon said to Kissinger before they launched the secret bombing of Cambodia – Henry, it's your ass. If you gamble on them and they're right, then it's everyone's ass.
Comments
This is undoubtedly a boon for us youths who are are hoping for a brighter day. I did not want to challenge views more informed than my own, but I think - at least in what remains of the left-leaning British press - a mea culpa is in order: Labour has done well, not purely due to the pure youth vote (and certainly not because they offered a university tuition fee abolition as the Tories seem to be spinning) but because they offered a left-of-field candidate.
We haven't won, but the Tory's plan to acquire, on the cheap, the nation's Brexit triumphalism have fallen short; Corbyn has shown the Tories for what they are in this election (I doubt I need to waste words here) and we now have a better position to push on. Corbyn's pitch has broad appeal, but can we make it broader? He's totemic now but I wouldn't be opposed, despite their previous, to so-called big-beasts returning to the fold.
(To be clear, I don't want to slag off older voters: I know full well not everyone over forty is a Tory!)
You write that "Hanretty’s model farms data from YouGov". This was true of the 2015 model. The 2017 model uses data principally from the British Election Study, with fewer responses per constituency.
You also write that 175 responses per constituency is not enough accurately to estimate constituency results. You cite margin of error calculators in support of this. This misunderstands the method. By trying to model the responses using characteristics of respondents and constituencies, it's possible to extract more information than can be extracted from the responses alone.
As proof of this, don't take my 2017 model, but rather the excellent YouGov model, which correctly predicted a hung parliament.