Many thanks once again to Ivor Knox for sending me the Panelbase datasets - and the big news is that the weighted results for the whole sample are even tighter than the published results (which are filtered by likelihood to vote).
Should Scotland be an independent country? (Whole sample, Don't Knows excluded)
Yes 49.8% (+3.1)
No 50.2% (-3.1)
If those had been the headline numbers (and there are two pollsters that don't filter by likelihood to vote) they would have been rounded up to Yes 50%, No 50%. In absolute numbers, there were 452 respondents after weighting who said they will vote Yes, and 455 respondents who said they will vote No. Even on the raw unweighted data, which usually favours No more, the position is closer than I can ever remember seeing in a Panelbase poll - Yes 49.2%, No 50.8%.
Incidentally, on the published headline figures that take account of Don't Knows, the Yes vote stands at 46.1%, which is higher than in any previous Panelbase poll - and that includes the poll from last September which is normally disregarded because of an unusual question sequence.
As you'll probably remember, Panelbase made two methodological adjustments a couple of months ago, one of which had a Yes-friendly effect, while the other had a No-friendly effect. The Yes-friendly one was the introduction of weighting by country of birth - the logic for which is inescapable, given that we know online polling panels have a disproportionately high number of English-born people on them. YouGov have since followed suit. But the No-friendly move towards weighting by recalled European Parliament vote is much more controversial, and isn't a procedure that is used by any other firm. It looks absolutely certain that if the latter change hadn't been implemented, there would be a Yes lead in this poll, because the 268 people who recall voting SNP in May have been downweighted to count as just 220.
Turning to the ICM poll, the most frustrating thing about the datasets is that two different sets of fieldwork dates are given - we're first told that the poll was conducted only on Wednesday and Thursday, but then a period of Wednesday to Friday is given. The latter version was the one reported by the Telegraph (which commissioned the poll), so hopefully that's correct - obviously the more up-to-date that a poll showing Yes ahead is, the better.
Of course when I first heard that a poll had shown the Yes campaign on 54%, the first thought that entered my head was that they had probably been flattered by the rounding. But that isn't really the case - on the unrounded numbers, the position is Yes 53.8%, No 46.2%. ICM weight their headline figures to take account of differential turnout, which in this case helps Yes slightly, but even without turnout weighting the numbers are healthy enough - Yes 53.3%, No 46.7%.
If the No campaign are scrabbling around in the datasets for some hope that the headline numbers may have been distorted, they might focus on the slightly implausible Yes lead in the South of Scotland sample, which has been upweighted almost two-fold. But nothing else is leaping out at me - ICM presumably realised that they wouldn't get a satisfactory sample of 16-24 year olds, so in a repeat of what they did in a poll earlier this year, they've opted out of the problem altogether by weighting 16-34 year olds together.
Finally, there's big news about Opinium's methodology, which I'd suggest calls into question the credibility of the No lead they reported last night. Unlike every other pollster, they've weighted their results to bring them into line with how people voted in the 2010 general election. That's in spite of the very clear evidence that there is a particular problem with false recall in respect of 2010, with many voters wrongly thinking they voted SNP because of what they did a year later. To be fair, Opinium have additionally weighted by 2011 recall, which dilutes the problem, but it certainly doesn't eliminate it - in this poll, 258 people who claimed they voted SNP in 2010 have been downweighted to count as just 179 people. There's no real doubt that if Opinium had followed the same recalled vote weighting procedures that most other firms use, the Yes vote in the poll would be even higher than the 47.4% reported.
It looks like neither ICM nor Opinium are weighting by country of birth, which is a deficiency that in all probability is leading to the reported No vote being inflated.