Opinion polls did reasonably well in the lead up to the ACT election, given the difficulty of translating votes to seats in the Hare-Clark system.
The Canberra Time-Patterson poll got the Labor vote in Ginninderra and the Liberal vote in Molonglo spot on. And was reasonably accurate with the major parties in all three electorates, except the Liberal vote in Ginninderra which it got spot on a week ago at 34 per cent but under-estimated by 8 per cent the day before the election.
The poll was fundamentally sound when you look at the percentage vote across Canberra. Trouble comes when you start breaking the votes into electorates because the sample size is so much smaller and the margin for error exponentially greater.
A simple example shows the point. Say you throw a coin twice in each electorate. That is six throws. To correctly guess and 3-3 heads-tails split across Canberra you have two changes in six or one in three. To correctly guess and 1-1 heads-tails split in each electorate you have a one in two chance in each electorate, which across three electorate is two times two times two which equals one in eight.
The reason polls are accurate is that the heads and tails (Liberal and Labor) even each other out the more you throw. If you slice the sample there is less capacity for even-ing out.
The poll appeared to over-estimate Green support substantially, but got it better in the later poll than the earlier.
The problem came in interpreting the results. Green support in an opinion poll would never translate at the ballot box. In the poll voters are unaware of other options (Motorists, Pangallo, Community Alliance etc). So if they don’t like the majors they say Green.
Come voting time, the other options are before them on the ballot paper.
The other interpretation problem came in translating voting intention to seats. The last seat in each electorate comes down to fairly complicated preference flows. The big Electoral Commission Excel spreadsheets from previous elections show that the Greens do not attract many preferences from either excluded candidates or from overflow of elected candidates. They are a love ‘em (first preference) or hate ‘em (no preferences) lot.
Both weeks of The Canberra Times-Patterson poll gave the Greens four seats in Molonglo. It was not going to happen.
The other poll was done by students at the University of Canberra journalism school as a class exercise. It had a smaller sample of 300 against 1200.
(I have to declare I have an interest here because I teach journalism at UC and had a fair amount to do with the poll.)
The students did fairly well, given they are not professional pollsters and the smallness of the sample. The UC poll got the ACT Labor vote within one per cent but over-estimated the Liberal vote by 3.9 per cent. It, too, had the Greens vote too high for the same reason as Patterson.
But its interpretation into seats was more accurate than Patterson. Patterson had Labor 8; Liberals 5; and Greens 4 – out by several seats. The UC poll had Liberals 7, Labor 7 or 6 and Greens 3 or 4 – spot on.
Both polls, however, under-covered an underlying trend that was largely unknown among the populace in the 18-month lead up to the election – a large swing away from Labor to the Greens.
If that had been drip fed with a poll every few weeks (as it was federally in the lead up to the 2007 election) it is quite possible that Jon Stanhope would have met the same fate as Kim Beazley.