Friday, July 11, 2014

Unrounded Survation figures confirm the No lead has fallen to its lowest level so far

As I suspected they would, the newly-released Survation datasets confirm two things - a) the Yes vote has reached an all-time high, even after undecideds are excluded, and b) the No lead has fallen to its lowest ever level, regardless of whether undecideds are excluded or not.  Here are the figures when rounded only to one decimal place -

Should Scotland be an independent country? 

Excluding Don't Knows :

Yes 47.1% (+0.5)
No 52.9% (-0.5)

Not excluding Don't Knows :

Yes 40.9% (+2.1)
No 46.0% (+1.6)

So it's not actually true to suggest, as a commenter did on the previous thread, that Don't Knows are "at best" splitting more or less evenly.  Obviously we have to take account of the margin of error of this poll, but the above figures indicate there's a slightly greater than 50% chance that the undecideds who have recently jumped one way or the other have broken more for Yes than for No.

Unfortunately, I have to put the same health warning on this poll that I put on last month's Survation poll.  The most Yes-friendly age group is 16-24 year olds, and they've been upweighted almost three-fold from 44 real respondents to 125 'virtual' respondents.  That effectively increases the poll's margin of error, and makes it more likely that Survation will produce volatile results (although admittedly that hasn't been the case so far), because any random sampling variation in the small sample of young voters will be magnified in the overall results.  However, the fact that roughly the same findings for 16-24s have been produced twice in a row decreases the likelihood that this is just a freakish occurrence, and it has to be said that it's not as if an 11% lead for Yes among young people has been an unusual finding for other pollsters - even the saintly "we're right and everyone else is wrong" YouGov have been known to produce similar numbers on more than one occasion.

In last month's Survation poll, the raw numbers were very similar to the weighted results, but that isn't the case this time - Yes have been upweighted from 42.7% to 47.1%.  Normally when other pollsters show a disparity like that, there's a very obvious, elephant-sized explanation - Yes-friendly lower income people have needed to be weighted up sharply, and No-friendly older people have needed to be weighted down sharply.  But that isn't the case with Survation - they haven't had to adjust the figures by social class much at all, and although over-65s are as usual breaking heavily for No, that age group have actually been upweighted slightly.  I thought the explanation might lie in the regional weighting, but again, that seems to be helping No - the No-friendly samples in the Highlands and the south have been upweighted, while the Yes-friendly samples in Glasgow and the north-east have been downweighted.  So I'm slightly baffled as to where the overall upweighting for Yes is coming from.  OK, women have had to be significantly downweighted, and there's the aforementioned issue with young respondents, but that wouldn't explain all of the disparity.  I can only assume it must be an accrual of a large number of relatively minor factors (for example Tory and Lib Dem voters from 2011 have had to be downweighted a bit).

One issue that Peter Kellner made a song and dance about in his recent attack on Survation was recall of 2010 vote - although both YouGov and Survation find that far more people recall voting SNP than actually did (almost certainly because they're getting mixed up with how they voted for the Scottish Parliament a year later), Kellner pointed out that Survation were actually showing that more people claim to have voted SNP than Labour, which he regarded as self-evidently ludicrous.  Survation noted in their very robust response that if it was accepted that 2010 vote recall is fundamentally unreliable, it seemed distinctly peculiar to attempt to use that as a measure of a firm's overall accuracy.  As it happens, though, the divergence between Survation and YouGov on 2010 vote recall has now narrowed somewhat - in this poll, 35.8% of respondents recall voting Labour, and 32.0% recall voting SNP.  Presumably if he's being logically consistent, Kellner will now declare that Survation's referendum findings must be a bit more accurate than before?


  1. As i said in the other post I think figures excluding don't knows are pointless, as it conflates "don't know" with "won't vote"

    In fact, the current method for excluding don't knows actually assumes a worse than 50/50 split of don't knows for yes, as it inflates both yes and no by the known split in yes/no intentions.

    I'd just add them on as an upper error bound to both Yes and No :-)

  2. "As i said in the other post I think figures excluding don't knows are pointless, as it conflates "don't know" with "won't vote""

    It doesn't really - it assumes for the sake of convenience that Don't Knows will break in the same way as everyone else. It's certainly standard to exclude Don't Knows in normal election polls - I've always been a bit baffled as to why referendums are treated differently.