Sunday, February 15, 2015

Queensland: Final Results And Poll Accuracy

The final results of the Queensland election are out.  Three years after being reduced to something that could fit in a Tarago, the ALP has won office in minority by what may just be a single seat.  (Katter's Australian Party spent so long milking press cover over their agonising decision on who to support that in the end government was formed without waiting for their decision.)  In one amazing echo, Peter Wellington gives Labor the numbers exactly as he did for a few months seventeen years ago.  In another, Antony Green estimates the 2PP at 50.9 to Labor, so virtually the same as 2009.  It is much as if 2012 just never happened.

There are two competing baseline 2PPs for the 2012 election, one of 62.8% to LNP and one of 63.1% to LNP, mainly depending on how you treat Gladstone.  On the latter the swing comes out at about 14%, on the former slightly less.  There may be slight revisions to the 2PP estimate for this election too, but in any case the swing was not much short of 14%, if at all.

In fact, the LNP were a trifle lucky to get as close to hanging on in seat terms as they did.  If the numbers are plugged into the ABC calculator (and crossbench defector/retiree seats assigned as they fell), it suggests 48-38-3 to Labor.  My own seat model, taking personal vote effects and probabilities into account, suggests 45-41-3 to Labor for that 2PP.  The LNP managed one more mainly because they had more very narrow seat wins.

The graph below shows the overall pattern of how the swings fell in most of the classic seats compared to the swings required for the seats to change hands.  The closer a seat is to the diagonal line, the closer the result.  The LNP's three closest seat wins against Labor (Mt Ommaney (0.23%), Whitsunday (0.38%), Mansfield (0.55%) were all closer than Labor's closest (Ferny Grove (0.82%).  The LNP also had a very close win against Pauline Hanson in Lockyer, which isn't shown.

The swing by seat was rather variable, with a standard deviation of about four points.  This sort of thing is common when a state swing is large, and it meant that Labor won Bundaberg (18.2% required) while missing Toowoomba North (9.6%). (Toowoomba North was the seat in which ReachTEL received the ultimate flattery of having a fake poll circulated showing Labor winning the seat easily.) In all, Labor missed seven or eight seats below the state 2PP swing that they got, and won two or three seats above it - exact numbers depending on the 2PP baseline and the assumed 2PP for Maryborough.

Sophomore effect (the theory that new MPs pick up a personal vote in their first term) was vindicated again at this election, with the average swing in seats won by the LNP from Labor at the last election being 12.8%, which is just over a point lower than the state average.  However since these seats were nearly half the state total, this means that the swing in the remaining seats would have been nearly a point higher than the state average to compensate.  So double sophomore effect (where an LNP incumbent had displaced a Labor incumbent at the 2012 election) was probably worth about two points, and mainly explains why Labor didn't quite get full value for the swing and won only 44 seats with the same 2PP that won it 51 seats in 2009.

The collapse of the Labor vote in 2012 resulted in a massive 17 seats at the 2012 election being "non-classic" contests (seats where the final two included someone other than Labor and LNP). .  Only seven remained so this time around: seven of the eleven LNP-vs-KAP seats reverted to LNP vs Labor (another two became LNP-vs-PUP and LNP-vs-One Nation.)  Three seats that were LNP-vs-IND in 2012 also reverted.  However, Noosa amusingly remained LNP vs Green.

Poll accuracy

At previous elections I have looked at both the primary vote and the 2PP vote to measure poll accuracy in order to give those who want to use non-standard preference estimate methods a chance to shine on the 2PP front.  At this election everyone used much the same 2PP method and the pollsters who were the closest on 2PP - Essential and Morgan - were actually the most right because their estimates of the primary vote for the major parties were the most wrong.  So while Morgan are not surprisingly taking it that they correctly declared the election "too close to call" while others were trotting out 52-48s, this is more as a result of errors cancelling out than anything else.

The 2PP issue is one I've covered in other pieces: I was expecting that as in other elections the under-estimation of preference flows to Labor would probably be cancelled out by pollsters overestimating the Green vote if not the Labor vote as well.  However in the end there was very little error in the unweighted average of the 2PPs in the five polls, while the preference issue bit about three times as hard as in other recent elections.

Putting all the final polls on the same playing field by combining results here is a comparison of the final polls from five pollsters (table revised 18 Feb 2015):

On the right I have two versions of the root mean square error score, a score that punishes small numbers of large errors more than large numbers of small errors.  The lower the better.  The first version is confined to the primary votes and the second includes the 2PP vote.  Normally I multiply the 2PP component by four but since everyone used the same 2PP method I haven't done that.

ReachTEL had the most accurate primary votes, ahead of Galaxy.  Much of the difference between the two was down to ReachTEL's publication of primaries to one decimal.

It can be also seen that while Morgan wins the version with the 2PP included for the reasons stated above, Morgan's primary-vote performance was not bad either, especially considering it was taken so far out.  Newspoll had the gap between the majors right but had the Others vote way too high and the Green vote too low.  The former has been a hallmark of Newspoll's federal polling lately too.  It was surprising to see seven out of ten polls published since January underestimate the Green vote, given that it was overestimated by every final poll, every Tasmanian state election poll, all 17 polls in the Victorian leadup and also every NZ election poll.

After being more accurate overall than every final poll in Victoria, my state aggregate method had a poor outing in its second mainland state attempt, being beaten by ReachTEL and Galaxy on primaries and everyone bar Newspoll once the 2PP is thrown into the mix.

Seat polls  

Generally the pollsters who attempted seat polling did reasonably well at this election.

ReachTEL had a lot of skin in the Ashgrove game having published numerous polls that all showed Campbell Newman losing, whereas others did not poll the seat so heavily.  Newman did indeed lose and ReachTEL had a very accurate 2PP, albeit because a 3.3 point overestimate of the primary gap between Labor and Newman was cancelled out by preferences. (Something similar applies to Galaxy which was slightly less accurate for this electorate.) Newspoll had the primary vote gap between the majors closer than ReachTEL or Galaxy, but had both major party votes too high and hence had an inaccurate 2PP, and would have had the margin too low under any preferencing assumption.

Newspoll polled the three seats of Cairns, Keppel and Ipswich West finding a 13.6% average swing.  The actual swing was 14.5% so that poll was quite close to the mark.  Galaxy's poll of eight seats with a 10.7% average swing was also quite good - the actual swing was 12.3%, and preferencing changes would have caused the difference (and possibly then some - these mostly weren't high Green-vote seats though).  But for the preferencing issue, it's likely this poll would have projected all eight seats as Labor wins (which they were).

So what has been learnt about polling here?  I'd say firstly that if the idea that robopolling is always inferior had any life left in it after ReachTEL's credible performances in the Victorian, federal and Tasmanian elections, it is certainly an ex-parrot now.  The debate about the quality of different polling methods is a silly one; the real debate should be not about live calling vs robopolls or landlines vs mobiles, but rather about how well different pollsters adjust for the limitations of their chosen methods.

As for Morgan, while their SMS polls had a rocky debut in Victoria, their single Queensland poll taken this year was decent enough to suggest that they are usable data (and indeed their polls of various states do not seem to have shown the same volatility as seen in Victoria).  They are usable data, provided we take the Greens figures with a large grain of salt.

And there's probably been enough said about the whole preference behaviour shift from 2012 except to add this: I thought that the impact of any shift in preferencing behaviour could be muted by optional preferencing, because not all voters direct preferences.  It seems now that optional preferencing can actually increase the volatility of preferencing from election to election.  Why?  Because there are Greens voters (especially) who will never preference the Coalition in a fit and who will always therefore preference Labor under compulsory preferencing.  But under OPV, if these voters are disillusioned with Labor and not especially mad at the Coalition in a given election, they may choose to just vote 1.

The preferencing behaviour issue will be a big challenge for modelling the NSW election, which presents a vaguely similar scenario of an Opposition rebounding from a historical slaughter, though in this case the Coalition government has been much less abrasive and the Labor opposition less impressive in its rebuilding attempt.  After giving NSW no attention to speak of while busy with Queensland and field work, I'll be turning my attentions there very soon.

PS (16 Feb): Adrian Beaumont has advised me that Newspoll did publish a figure of 3% for PUP in a footnote.

PS (17 Feb):  Antony Green's review is also now up.  Antony gives the final 2PP as 51.0 to ALP and notes these estimated 2PPs for the non-classic seats (figures are for Labor): Callide (LNP) 37.2, Dalrymple (KAP) 50.9, Gladstone (ALP) 75.1, Lockyer (LNP) 49.3, Mt Isa (KAP) 47.1, Nicklin (Ind) 47.1, Noosa (LNP) 41.2.

Mt Isa and Dalrymple deserve special comment here.  Mt Isa was won by Labor in 2009 by five points more than the state average and has a mostly ALP history.  Dalrymple was won by Shane Knuth for the LNP against One Nation in 2009 with Labor not even making the final two (admittedly, Knuth and One Nation's Rosa Lee Long were both sitting members for part of the new seat.) Knuth later joined KAP.

It seems rather strange that Mt Isa should return a pro-LNP 2PP and Dalrymple a pro-Labor one and I suspect this reflects a lower campaign effort in Mt Isa, or perhaps a greater appeal of Rob Katter to Labor supporters given that he was not ex-LNP.  Perhaps the 2PP estimate for Mt Isa is correct but I would not be too surprised if it failed to follow the model.

Also, David Barry has a fantastic Queensland election scatterplots page that will tell you (in graphical form) everything you never knew you never knew (yes that was deliberate there) about recent Queensland election patterns!

Table Edits (18 Feb): The original version of this article incorrectly had the Essential December poll as the Essential January poll as a result of taking the figures from a source that had misread Essential's tables.  This has been rectified and the table reposted in the body of the main article.  The correction makes very little difference to my formula's assessment of Essential's accuracy.

(Update: As noted by little_stranger in comments, Antony has increased the 2PP estimate to 51.1 to Labor.)


  1. Hi Kevin - thanks for your great work. Just as a postscript, Antony Green has just updated the 2PP for Queensland to ALP 51.1%.