More on the Angus Reid poll
The full tables for Angus Reid's first UK voting intention poll are now up here, so we can have a dig around. From the top, the sample is drawn from an online panel of paid volunteers, in much the same way as YouGov.
The survey asks voting intention using the question "If a General Election were held tomorrow, which one of the following parties would you be most likely to support in your constituency?". This refers specifically to how people will vote in their own constituencies, in the hope that it will pick up some tactical voting. To be honest, I doubt it will make a massive difference (my theory is that the reason it does so in the PoliticsHome marginal poll is specifically because people are asked the question both ways).
Respondents are prompted by just the main parties, if they say "other" they are given a second question to find which "other" party they would support. I had pondered whether the high level of other support was the result of including smaller parties in the prompt, but that is not the case (I can see no obvious reason why Angus Reid found support for "other" parties so much higher than other pollsters - recently ICM have had them at 10%, Populus 12%, YouGov 12%, ComRes 13%, MORI 11%. AngusReid have them at 17%. I guess time will tell if there's something methodological causing it, or it's just a blip.)
Respondents who say don't know are then given a "squeeze question", in the way Ipsos-MORI and ComRes do, asking them which party they are leaning towards. This had virtually no effect on the final figures, switching them from 41/23/20 to 40/23/20.
Like YouGov, there is no filtering or weighting by likelihood to vote. This is something that should in theory help Labour - filtering by turnout normally increases the reported Conservative vote, since their supporters invariably say they are more likely to vote. However, our only real point of comparison is the only regular UK online pollster, YouGov - they don't filter by likelihood to vote, and it doesn't seem to produce figures that are kinder to Labour. As I've suggested before, perhaps its because the lack of an interviewer effect means people are happier to say they won't vote if they really won't, or perhaps the sort of people who don't vote don't join internet panels.
Finally we get to the question of weighting. AngusReid are weighting by the normal demographics, plus newspaper readership, and by recalled 2005 vote. There is no adjustment for false recall - they weight to the actual shares of the vote in 2005, which they have found to work in the past in Canada (where they do indeed have a good record).
Intuitively this may seem right... but no other UK pollsters do it because of false recall. Evidence from panel studies shows that people do not accurately report their past voting. There are various explanations for this: people are embarrassed to say they didn't bother to vote, so give the party they would have voted for; people who voted tactically give their first preference, rather than the party they put an X next too; people saying the party they wish they'd voted for with hindsight. Whatever the reasons, there is broad concensus amongst British pollsters that false recall means more people claim to have voted Labour than actually did, and invariably fewer people report having voted Lib Dem than actually did. I discuss it at (slightly) greater length here.
The lessons pollsters draw from this are different - ICM, Populus and ComRes adjust their past vote weighting targets to account for false recall. MORI take the view that false recall makes past vote entirely unsuitable for weighting. Putting aside the theory, here's what it means in practice. Below is the recalled 2005 past vote from the pollsters who provide it (YouGov weight by party ID instead so we cannot compare to them). I've also shown it repercentaged to exclude the don't knows and won't votes.
ICM: CON 19%, LAB 23%, LDEM 13%, Others 5%, DNV 32%, Ref & DK 8%. Repercentaged - CON 33%, LAB 38%, LDEM 22%
Populus: CON 19%, LAB 22%, LDEM 12%, Others 5%, DNV 35%, Ref & DK 7% Repercentaged - CON 33%, LAB 38%, LDEM 21%
ComRes: CON 18%, LAB 21%, LDEM 12%, Others 5%, DNV 35%, Ref & DK 10% Repercentaged - CON 32%, LAB 38%, LDEM 21%
Ipsos MORI: CON 20%, LAB 28%, LDEM 10%, Others 6%, DNV 27%, Ref & DK 10% Repercentaged - CON 32%, LAB 43%, LDEM 16%
Angus Reid: CON 25%, LAB 26%, LDEM 17%, Others 5%, DNV 26%, Ref & DK n/a Repercentaged - CON 34%, LAB 36%, LDEM 23%
As you can see ICM, Populus and ComRes have very close figures indeed. MORI do not weight by past vote, and hence their sample has a much higher proportion of people saying they voted Labour in 2005 (though without weighting by these figures, MORI's past vote figures to vary from month to month). Evidently Angus Reid are weighting Labour to a lower level of support than the other pollsters, and weighting the Liberal Democrats and Conservatives to a higher level.
The figures aren't necessarily directly comparable, since we don't know whether there is the same level of false recall in a telephone survey using live interviewers, and an internet survey with no interviewer effect - for example, if some of false recall is a social desirability bias against admitting to not voting, it may be smaller when people are asked online. Prima facie though we should expect their figures to show a lower level of Labour support and a higher level of support for the Liberal Democrats than some of their rivals. Whether that happens in reality, of course, remains to be seen.
