Where the polls went wrong

What went wrong

So what was it? The last parliament was unusual in many respects. But in terms of substantial shifts of voters, there were three of note: the collapse of the Lib Dems, the surge of UKIP in England and Wales and the surge of the SNP in Scotland. Since the errors were confined to England and Wales, there was clearly no direct impact from the last of these (indirect effects – namely fear of the SNP – are a discussion for another time).

The interaction of these events, with added twists such as the Green surge, created a state of electoral flux, something that I highlighted in my pre-election piece as a significant risk to the pollsters. In the old days, the electorate consisted of two tribes and a single, relatively homogenous group of swing voters, on a single axis, and swings tended to be uniform, unlike 2015 where the pendulum swung 7.7% to the Conservatives in North East Somerset but 10.8% to Labour in Ealing Southall. The patterns of voter flows really mattered.

The flows that the polls showed were extremely clear – the Lib Dems were going disproportionately to Labour and the UKIP vote was coming disproportionately from the Conservatives. John Curtice called the former a “crutch” and the latter a “wound”. Intuitively it made complete sense – one of the Lib Dems’ parent parties was founded by moderates that quit Labour under Michael Foot in the early 1980s, while UKIP’s origins lie among disgruntled Eurosceptic ex-Tories in the early 1990s.

After last year’s council elections, Steve Fisher of Oxford (and also of the broadcasters’ exit poll team and currently the BPC polling inquiry) pointed out on his personal blog that local contests had shown an interesting pattern. If the 2010 Lib Dem vote were going entirely to Labour and the Greens, and UKIP’s votes coming entirely from the BNP and the Conservatives, Labour would have won that round of local elections by 10 points more than they actually did. Reallocating the vote flows to make the numbers add up suggested that ex-Lib Dems were splitting evenly between the Tories and Labour and that UKIP’s votes were coming almost equally from the big two.

Unlike much of Steve’s other work (notably his Elections etc website), this insight never really got the attention it deserved, and those that did comment tended to explain the pattern away by local-versus-national deviations in voting patterns and turnout.

While hats were being eaten

On election night, interpreting the wildly different swings was made a lot easier by Rob Ford and Philip Cowley tweeting the comparable exit poll prediction for each seat swing as it came in, so those of us that were up to our eyeballs in caffeine didn’t have to think too hard about them.

Hampstead and Kilburn, where I had once lived when it was Hampstead and Highgate and Glenda Jackson had a majority of almost 8,000, but was now retiring on a majority of just 42 votes, was a result that got my attention. With the Lib Dems having been on 31.2% in 2010 (in what was then a tight three-way marginal) and likely collapsing into the single digits, it hadn’t been a surprise when Lord Ashcroft’s poll showed that the Labour majority was set to balloon. As it happened, the Lib Dem vote did collapse, by 25.6 points. But the Tories were up 9.6 points and Labour up only 2 points more. That swing of just 1% turned out to be exactly in line with the England and Wales average.

Elsewhere, a new three-way marginal was Thurrock, where UKIP surged 24.4 points to 31.7%. Though it seems this was driven in part by former BNP voters, it was certainly topped up by Labour and Tory defectors. Yet Labour’s vote share declined 4 points – about a point more than the Tories’.

Clearly, conclusions drawn on cherry-picked constituency results can be horribly wrong. So I’ve modelled the voting patterns more generally using what I call relative net flow analysis. The motivation is simple – changes in vote share from seat to seat differ for all kinds of reason, many of which are insignificant from a national perspective. But by filtering through the noise, signals emerge.

What I did was to take all 572 seats in England and Wales excluding Buckingham (the speaker’s seat), order them by the change in Conservative vote share, from the largest gain for the greatest fall, then average them across interval blocks of 10 seats (11 in a couple of cases to avoid odd lots). For each group I also averaged the changes in UKIP and Lib Dem vote shares. Then I repeated the exercise with Labour vote share changes, grouped in order.

(Continued, please navigate using numbered tabs…)





About The Author

Related Posts