“I did think that perhaps the first table in their pdf results was showing the unweighted recalled past vote, which ComRes then use to generate their target weightings. ”

That’s the only way the % would work, isn’t it? e.g. 20 18-24yr old supporting Labour is only 17% if it is of the weighted amount (117) not the unweighted (85).

For example, 52% of the adult population are female. One month the raw sample, thanks to random sample error, might be 50% female, in which case a weighting of 1.04 would be applied to all the women in the sample. The next month the raw sample might be 55% female, in which case the weighting applied to all women would be 0.95. These change according to the make up of the raw sample.

However, I don’t think that’s actually what Richard is asking about! I think he is asking about the variation in the past vote figures that ComRes are weighting to, not the weightings they are using to achieve it. Assuming we are reading the tables correctly, 2005 Lib Dems were weighted to a target of 9% of the sample this time, but to a target of 12% last time. I’m sure there is some explanation – but we don’t know what it is.

But when you actually go and ring up 1,123 names out of the phone book, the fact you chose them at random means they won’t be exactly representative. If your selection process is genuinely random, eventually the differences will average out over many years – which doesn’t help you now. So you count the codgers, Liberals, etc, work out the %s, and compare that to known data, then you weight each group according to the ratio between its size in the sample and its size in the general population.

This brings up some other problems; for small groups (like, say, RESPECT voters), you might only have five respondents even though they make up a significant percentage of the population. If you weight them up enough to match the population profile, though, you have the problem that one or two people’s views are being taken as representative of 5-10% of the nation. So you need to oversample the tiddlers, and then weight them down to their real size.

Therefore, the weightings are never the same, because the size of the weighting needed to correct the sampling error is dependent on the size of the sampling error, which fluctuates randomly around the actual value.

Thanks, but surely such a large change in weighting in less than a fortnight is still inexplicable?

We know this for certain, since if you take 1000 people and ask them how they voted the day after the election, and then ask the same people 6 months later, you’ll find their answers have changed.

The main factors seem to be people forget they voted for minor parties and the Lib Dems. People who didn’t vote, claim they did since it’s the socially responsible thing to do (and those people tend to claim they voted Labour) and people who voted tactically for the Lib Dems but actually supported Labour say they voted Labour, rather than Lib Dem.

In practice this means pollsters need to weight their sample to an estimate of what a representative sample would say they had voted, rather than how people actually voted. The debate is how quickly this changes – ICM and Populus believe it changes very, very slowly – so weight to pretty much the same figures month-in and month-out. MORI believe it has the potential to shift quickly, so reject weighting by past vote.