Monday, December 25, 2017

Queensland 2017: Final Results And Polling Accuracy

Queensland: ALP 48 LNP 39 KAP 3 PHON 1 GREEN 1 IND 1
2PP Estimate 51.2 to Labor (+0.1 from 2015)

It's taken a while but I've finally found some time to put up something about the final results of the 2017 Queensland state election.  I try to always put something out on Christmas Day, though last year nasty weather interfered with that plan.

In a nutshell, the 2017 Queensland election was one where a great many dramatic things could have happened, but virtually none of them did, as the following sections explain:

Hardly any seats changed hands

You don't turn 89 seats into 93 without breaking a few eggs, but the level of seat transfer between the parties at this election was remarkably low.  On a notional basis and ignoring retirements and mid-term defections, just nine seats changed hands at this election, most of them marginal anyway.  The Liberal National Party lost Redlands (1.2%), Gaven (2.8%) and Aspley (3.2%) to Labor, and would have lost Maiwar (3.0%) to Labor as well but the Greens snatched it instead.  Labor lost Bundaberg (0.5% and which was a freak win last time anyway) and Burdekin (notionally theirs by 1.4% but LNP-occupied) to the LNP, and might have lost Mirani (3.8) to the LNP had not One Nation helped itself to its only win.  The LNP also dropped Noosa (6.6) to independent Sandy Bolton, and Hinchinbrook (3.4) to KAP's Nick Dametto.  In Hinchinbrook, Dametto (who according to his party had only been campaigning for four weeks) pulled off a duplicate of Andrew Wilkie's Denison 2010 winning method of coming third and getting everyone's preferences.

There was very little 2PP swing

On the night, 2PP estimates were running at about 52-48, but after a close look at the final numbers my estimate is just 51.2% to Labor, a swing of 0.1% from 2015.  Perhaps an exact figure will be derived from the ballot papers, but in the meantime mine is based on the actual numbers of preferences that flowed to Labor and the LNP in cases where both were still in the count, and estimates for the remaining preference flows based on actual flows observed at the election.  For the Greens I've used their Maiwar flow, for Strelow I've assumed her preferences would flow between the ALP and LNP much as they did between the ALP and One Nation, and for Bolton I've assumed 50-50.  Overall a shade over 90% of votes can be exactly accounted for, and those for which estimates are required are mostly One Nation.  I'm most uneasy about the estimate for KAP (51-49 to Labor), because KAP inconveniently failed to come third in any of the ten seats they contested, so there was no opportunity to see how strongly their how-to-vote card was followed between the major parties.

Few seats were won on preferences

As Tim Colebatch has noted, only nine seats were won from behind on preferences, with the LNP losing five primary vote leads to the left plus Hinchinbrook, and Labor losing three to the right.  This is the same number of seats as won from behind under the old OPV system in 2015, but all nine in 2015 were won by Labor.

The change to compulsory preferencing had little measurable impact

Had the same primary votes been cast at an election with optional preferential voting, very few seats would have changed hands. Voter anger at the Newman government in 2015, which drove Labor's many wins from behind on preferences despite OPV that year, was replaced with apathy about both Palaszczuk's replacement and its opponents.  The change to compulsory preferencing was seen as having backfired when the One Nation vote soared in polls, but in the end Labor would have received a slightly greater advantage statewide off Green preferences than the LNP would have off One Nation preferences, even had all the latter been distributed.  To what extent that advantage would have been different under OPV - if any - is speculative.

However, compulsory preferencing had an intangible impact, since it forced major parties to make decisions about preference strategies and policy positioning in order to secure preferences.  The LNP wanted to cause Labor to lose seats to One Nation, and hence had a reason to preference it in most seats (except where its candidates were too much of a liability) but this made it much easier to damage the LNP by linking it to said party, and to talk up a potential coalition involving One Nation.  One Nation only won one seat as a result, so it's likely the LNP made a strategic error by overestimating the likely seat yield of their strategy, in spite of the published modelling that showed One Nation's vote needed to be around 20% before they started winning lots of seats.  Labor seemed wide open to attack on its switch on Adani loans during the campaign as a possible sop for Green preferences, but my impression is that attacks on Labor over this (and indeed on the Premier's obvious conflict of interest) failed to gain the traction they might have gained.

My perception is that far from harming Labor as was widely expected, the switch to compulsory preferencing was in the end close to neutral in terms of the impact on preferencing, and that the strategy of handling it actually played out in Labor's favour.

One Nation - overhyped more than fizzer

As mentioned, One Nation won only a single seat (even worse than the three my modelling expected) but this was only relative to expectations that were never supported by a credible reading of the bulk of the polling.  It was possible One Nation might win a raft of seats but it was seldom likely.  As for the seats my model had them most likely to win, they won one, KAP poached Hinchinbrook, and in Lockyer PHON were let down by a dud showing from the ALP.

One Nation's preferences did flow more strongly to the Coalition than at other recent elections, but the flow was "only" around 65% (perhaps a little stronger but disguised by flows from other parties) and not the 75% some respondent-allocated polls were getting.

Where One Nation really fizzed was in the much-vaunted impact of their preferencing decisions.  As noted by Colebatch, there was not that much difference in preference flows between the seats where they preferenced Labor and those where they preferenced the Coalition.  The one seat these decisions may have swung to Labor was Apsley, where 49.9% of One Nation preferences went to the LNP, 33.7% to Labor and 16.3% to the Greens.  Assuming the state average of about 11% of One Nation voters were following the card (which preferenced Labor), the seat would have been line-ball had those voters flipped.  The other one that has received some attention was Redlands, but I don't think that's a valid example.  In Redlands, more preferences from One Nation went to Labor than the LNP (the only seat where that happened) but that was because the One Nation votes included votes from the past LNP member (pun intended) Peter Dowling, who preferenced One Nation then Labor on his card.  This would also have diluted the flow had One Nation preferenced the LNP.

There was also one seat the LNP won because One Nation preferenced them ahead of Labor, and probably wouldn't have won otherwise - Pumicestone (notionally theirs by a very small margin, but occupied by an expelled ex-Labor incumbent).

Other claims about the One Nation decision costing the LNP seats just ignore the fact that a party can never cause all its supporters to vote the same way.  Meanwhile the LNP didn't seem too bothered that it may have gifted Labor a majority itself, by preferencing Labor ahead of the Greens in South Brisbane and One Nation ahead of the (admittedly Labor-ish) independent Strelow in Rockhampton.

Polling Accuracy: State Polls

I mentioned in my early post-election coverage that Newspoll had done stunningly well on the primary votes and this remains the case.  However if my estimate of the 2PP is factored in, the accuracy picture becomes less clear.  If a final figure is produced that differs from my estimate I will edit this section accordingly.

Below is a table of the final polls by each pollster and my assessments of their accuracy.  The first RMSQ column gives the root mean square error for primary votes, the lower the better. I like using RMSQ because it punishes a big error on one party more than several small errors.  The RMSQ2 column includes the error on the 2PP in the calculation, with a weighting of four.

I've included two ReachTELs because one ("RT late") was released on election day.  I found it very frustrating that someone would commission a poll in time for it to only be released about an hour before the polls closed.  I also think it is not ideal to compare such a poll with one released before polls open on election day, since in theory such a poll might be influenced by on-the-day exit data.  (That said, polls generally include more and more exit data because of increasing pre-polling.)  In future I will not include polls released on election day after polling opens in polling accuracy assessments.

On the primary votes Newspoll is easily the best and Essential impressively beats both ReachTELs and Galaxy.  With my 2PP estimate and weighting factored in, Newspoll still wins but there is little difference in accuracy across the pollsters (excepting the earlier ReachTEL which had One Nation too high and the LNP too low, with the closest-to-the-pin 2PP reducing the damage somewhat).

Seat Polling

I can find only two cases where public pollsters went head-to-head in a seat.  In Thuringowa, ReachTEL's 50-50 was much closer than Galaxy's 54-46 to One Nation (it was actually 54.1% to Labor).  In Whitsunday, both pollsters wrongly had Labor winning, but still their errors were small (Galaxy 51% for Labor, ReachTEL again closer with 50.5, actual 49.3).

The quality of the remaining Galaxy/Newspolls was good, picking correct winners and getting the 2PP margins close in Cairns, Bundaberg, Burdekin (a hard seat to poll), Mansfield and Ipswich West.  They also had correct winners, with moderate misses on the margins, in Logan and Hervey Bay, and the correct winner but exclusion order issues in Rockhampton (where Strelow most likely picked up support after the poll was taken).  The only incorrect winners were in Mundingburra, Gaven and South Brisbane and these were all within margin of error of the samples.  The remaining seat ReachTEL, of Ferny Grove, had the right winner and also was less than a point out on 2PP.

Waiting For Brexot

Pundits have frequently been keen to talk up impending elections as our "Brexit/Trump moment" when the polls will be completely wrong (not that they were in those cases anyway) and our political system will be suddenly overrun by right-wing populism.  However this didn't happen at the 2016 federal election, or the same-sex marriage postal survey, or the WA state election, or the Queensland one, or in Bennelong where Cory Bernardi's attempt to bring a form of Trumping to Australia barely got its deposit back.  Time to get a new line as the repetition of these overseas mantras in the Australian political context is looking more and more like an absurdist script rather than real analysis.


  1. Thanks for your efforts throughout the year Kevin. Have a Merry Christmas and a great new year. It's looking likely that you're going to be busy in the coming year. Tassie going to the polls and I fully expect a federal election before the end of 2018.

  2. Great analysis for a Christmas morning read. Goes to once again demonstrate that scientific polls are the best way to forecast an election. I think Brexit and Trump are examples of low probability events in a two horse race actually happening.

  3. Thanks so much. You're the no.1 psephologist imho

  4. Weatherill getting re-elected in 2014 was unexpected, and he lost the popular vote by a significant margin. Not the same side of politics but I think that's our closest contender to a Trump moment (not sure what the polls looked like).

    It's interesting seeing recurring patterns in politics however. UK 2017 was eerily similar to Australia 2016.