Listed land-line use hasn’t just reduced marginally – which appears to be the basis of your underlying argument. It has plummeted. In 1996 it was about 79% in my electorate, now it is 53%. In 1996 Mangere was close to 60%, now it is about 35%. In the North Shore it was in the high 80′s, now it is about 68%. The same trend is happening over the whole country and it is accelerating,

As Anita says, if there was an inherent sampling error common to all of the polling companies, then a poll of polls merely compounds the error.

That is exactly what I’m asserting – there is an inherent bias happening in the poll methodology. Using landlines is heavily biased against younger people, heavily biased against people on low incomes, and for that matter heavily biased against people I know (who you can’t find in phone books – they’re technophiles). The reduction has been considerably less in areas that vote conservative., considerably more in areas that vote progressive.

If the polling companies weight on demographics then the effect gets compounded, not reduced. Imagine that you’re looking at 25-30 years olds by phone – hard to find. You’re likely to wind up with someone who does things traditionally, like having a listed land-line, and then multiplying it.

Similarly if you get someone in Mangere by phone, what is the bet you get someone who is both relatively affluent AND conservative. They have a listed landline. As far as I can see having a listed landline is indicator that you’re more likely to be affluent, technophobic, and conservative.

Of course the polls close towards reality the closer you get towards an election. You get more of the undecided answering (which is why the missing figures are more critical than who answers). I’ll give you a guess who I think is more likely to answer polls further away from an election. They aren’t younger left voters or older less affluent voters – they’re too busy surviving or doing their own thing to think about politics. It is the social conservatives…

That is why both you and Hooten quote close (3 weeks) to the election polls when you start talking about poll accuracy. Hell they even start closing up between polling companies before the election. Try looking at the polls further away from the election say at 8 weeks. Look at how much variance you get – that is undecided voters. This year has already shown far more voliatility between polls than past election years. I think that the reason is the steadily reducing numbers of listed landlines.

If the polls are wide really close to an election then I’d get more concerned. But at present I see media holding on to the polls as if they are the holy grail. It annoys me because that is not what I’m seeing in the phone polling we do, or the door knocking to fill in the areas that don’t have phones. It is a far more robust technique than pollsters use because we’re targeting the wavers and enrolled non-vote.

It is going to be a hellishly close election when you factor in the coalition politics. Especially since the undecided are higher than I’ve ever seen in the last 18 years. It is all going to come down to turnout.

Which of course is where the polls come back in – effective at stopping people voting. Thats why Hooten likes them, it helps with the spin.

Helen would step aside end of 2009 or early 2010.

The current published polling means something very differen if 95% of the electorate has decided from if 60% of the electorate is decided.

A consolidated combination of a range of poll data (i.e., polls of polls), with a much larger combined sample, will mute any sampling or methodology errors.

Unless they have the same sampling or methodological errors. Given they’re all outbound phone polling we can comfortably assume they have some sources of errors in common.

Currently polls are useful for trends, which is why Helen probably looks at the UNR data. If I was looking at the polling data in the way that you say Helen is, I’d be looking at the effects of specific proposed or announced policies in specific sub-samples. If it indicates a positive (or negative) change in voting behaviour, then that is useful information.

I disagree, LP. That’s what you use focus groups for, which UMR certainly does do for the Labour Party. A poll might include fifteen questions; a few more for a purely political poll (as used by political parties) and to drill down on attack lines, a few less for a major polling company such as Colmar-Brunton or Digipoll, were political questions are included in wider consumer surveying to make up their political poll. Most of the published polling we see is part of a wider consumer survey.

You can see me pointing out in comments the delta change of series of successive polls. A series like -3%, +5%, +1%, 0% in successive polls from the same company targeted in the same way is useful (if they haven’t changed their methodology). It shows the trend espcially when read in conjunction with dates, events, and policy releases.

I disagree LP. A series like 53%, 47%, 53%, 47% in successive polls from the same company with a sample of 500 may not show any bumps at all in support: it probably shows quite a steady level of support. Each of the differences is likely to be within the margin of error. If you are relying on one poll to base a trend, then you are dealing with what is probably a relatively high margin of error in each of them. Polling companies do not call the same people: if you generated a sample of people, and tracked that same sample over the period, then you could probably say that individual policies have had an influence on the difference.

A consolidated combination of a range of poll data (i.e., polls of polls), with a much larger combined sample, will mute any sampling or methodology errors.

However the absolute percentages are rubbish for actually figuring out an outcome is useless. That is what most of the media do – ie saying how many seats each party would get.

I disagree here as well. The 2005 election result was remarkably close to a poll of polls of the last 3 weeks of the election campaign.

Averaging the various polls tends to give better results. However if the base data that each poll is based on has an inherent error (land line access) then they’d all be biased.

You have said that land line access is an inherent error. I don’t know where you get the basis for this. You have said that landline access has decreased marginally over the last three years, but there doesn’t seem to me to be so consequential as to dramatically undermine polling integrity.

Labour and National were last neck and neck in consolidated polling in January 2007. Landline access hasn’t changed dramatically since then. If low landline access creates a bias against the Left in poll results, then that doesn’t explain why in January 2007 the consolidated results were so close.

I suspect also that landline access issues are a bit of a myth, for two reasons. The first one is that landline access is as low among medium-high income apartment-dwellers as it is among low-income tenants. The second reason is that polling companies, when they take their samples, weight their results accordingly. They don’t simply say: “Right, we’re going to conduct two thousand phone calls, and take down the results of anybody who answers” (which the survey linked to this article, ironically, confesses to doing. They ensure their responses match the population in general, including age, ethnicity, income, geographical region, and historical voting preferences. It just isn’t credible to say that poor people miss out on getting surveyed, because polling methodology requires that they do.

As you pointed out, a 2% change between the average of the sampling polls and the real poll was sufficient to change the outcome of the election.

I’m not saying that there is no sampling or methodological bias. There probably is. But if it exists, it is within a 2% range, as we saw at the last election. It isn’t the ten percent range that might suddenly be made up between Labour and National this election. If the polls show a 15 point gap between Labour+Greens and National+Act in the week before the election, then really you’re just living in dreamland if you think that can be explained away by landline access.

People who do canvassing know this because what we see when we’re phone canvassing and especially door knocking varies a *lot* from published polls all of the time. As the elections go by I keep seeing bigger and bigger variances between the polls and the canvassing.

When you’re spending an afternoon doorknocking, you’re only doing so in a single meshblock. It is highly likely within a statistical meshblock, you’re going to see very little variance in voting preferences. Which is why if you’re door-knocking Victoria Avenue in Remuera, you get 80% voting National, or 80% voting Labour in many parts of Mangere. But you would have to be quite deluded to say that the voting pattern in four hundred doors you’ve successively knocked on is representative of the voting pattern generally, because that isn’t sampling at all.

My money is on Cullen. I don’t think he is ready to retire yet (he can point to McCain’s age as a precedent) plus I reckon he could pull together the numbers to fend off Goff.