just another note on the polls...

nanos has the liberals at 35%, and the conservatives at 31% on his three-day rolling averages. a mainstreet poll was released today with the conservatives at 38% and the liberals at 29%. they both have the ndp in distant third. these differences are well outside of the margins of error. how is this possible?

if you were naive, you might point to mainstreet having the larger sample size, but that's what margins of error are for. the larger sample size means there's a smaller margin. it's not enough to explain drastically different results that are outside the margins.

the first thing i noticed was that nanos is not undersampling young people. this is a change that some of the polling firms are doing to compensate for not getting the bc election right a few years ago; they've concluded that weighting young people relative to the census is flawed because young people are less likely to vote. the polls that you see with the conservatives above 32-33 are, without exception, overweighting older people on purpose under the assumption that they're more likely to vote. the consensus seems to be to weight 65+ at 40%, 50-65 at 30%, 35-50 at 20% and 18-35 at 10%, +/- a bit. that's obviously not "relative to the census" as claimed, but rather relative to a guess as to how likely they are to vote. it's blatant data manipulation. but, it's hard for me to get on them too hard about it, because i fully realize that it might actually be right.

at the least, this should be understood.

but, i noticed that the mainstreet polling also had the conservatives ahead almost 2:1 with people under 35, which is pretty weird. reweighting would not create consistency. it's not just that, there's something else.

if you look carefully, you see that something else: the mainstreet poll also had undecideds at around 20%, whereas nanos has them around 10%. if you look at the actual numbers, they have the conservatives at 32%, which is more consistent. but, that also takes the liberals down to 24% and the ndp down to 20%. consistency is then possible to create by disproportionately distributing the undecideds to the liberals, and giving the rest to the ndp. multiple polls have suggested that the conservatives are getting almost none of the undecided vote, so that does seem consistent, if somewhat creative.

but, why is mainstreet picking up this huge number of undecideds and nanos picking up huge numbers of liberal voters, and lesser numbers of ndp voters?

it may come down to the question. mainstreet may be deciding to file weak support under undecided, while nanos is deciding to file it as leaning.

it's impossible to say which is more accurate. but, it may suggest that the ndp is still under consideration by a large number of people that are claiming they'll vote liberal.

at least, they can both be right that way. and, these are both reliable firms, so it's hard to come to the conclusion that one or the other is so drastically wrong.

if that analysis is correct, it would mean that the greater sample size is useful at zeroing in on conservative support locally (so, they had them at 36% in ontario, whereas nanos has them at 32, and that is in the nanos margin), but not as useful in determining liberal or ndp support locally, due to the larger number of undecideds - with the caveat that the nanos data seems to suggest that this undecided is currently leaning disproportionately liberal.

or, you could look at it the other way - that it's useful in determining core liberal and ndp support, and useful in measuring the size of the swing that exists between them, which is quite large across the board.

that would have the effect of skewing conservative support upwards, and it would give the ndp a little hope that they're not totally out of it.

hopefully, that undecided comes down a little in their next set of polls. because a larger sample size really ought to be more useful, not less useful. but, factoring in such a large level of undecideds may very well be a more accurate reading, for the moment, too.