This post just rounds up a number of issues in modelling the Queensland election. There have been no fresh state polls since my previous post and I still think it's most likely the LNP will retain office with a narrow to moderate (but greatly reduced) majority. I could have added these as updates to the roundup from the weekend or saved them for next time, but thought they'd work quite well as an article by themselves to stop the last one or the next one being too long.
ReachTEL Poll Of Ashgrove
(Scroll down for Newspoll update Jan 17)
The first seat poll of the campaign has been released and there are no prizes for guessing which seat it covers. A ReachTEL with a slightly larger sample size than last week's state Newspoll and Galaxy offers the candidates in ballot order and finds Campbell Newman effectively up 2.4 points in just over a month (after redistributing "undecided" from last time), and Kate Jones effectively down one point, with the other significant change being the disappearance of the Palmer United Party's three points (they're not running). The 2PP is 53:47 to Jones and voters are almost exactly split as to who they think will win the seat.
This apparent improvement for Newman isn't a massive surprise given the apparent statewide improvement in LNP polling. 47% 2PP is a slightly better figure than would be expected at this stage, but not by enough to rule out statistical noise. However, it's also possible that Newman is starting to punch above his weight (based on the expected swing statewide) or that this reflects a shift in state voting intention since last week's polls were taken.
There's quite a split at the moment between popular expectations of a close fight and the orthodox pendulum (or union polling) derived view that if the LNP doesn't increase its state vote share then Newman is all but gone. There's a view I've seen a few times that voters will pick up for Newman close to the election because they will realise the LNP is winning the election, and think about the benefits of having a Premier as local member. After all, if you just want goodies for your electorate then what could be better than saving the Premier by a whisker and putting the seat on tenterhooks again for next time? We don't really know how much this sort of dynamic affects voters since a truly equivalent precedent is difficult to find. It's a reason to be cautious about declaring the Premier's career to be dead by pendulum while the seat is within, say, 5%. All the same, if the desperate efforts to shore up Ashgrove are going to actually work, one might have expected more sign than this of it by now.
There is still a long way to go, since at this stage in 2012 a ReachTEL had Newman trailing 51.5:48.5. He did benefit from Labor's stupid and easily parodied Newman's Web advertisements, and by a few points in statewide vote share drift, so it's unlikely he'll again gain seven points in three weeks.
Seat Betting Changes
In the previous roundup I noted that while polling seemed to imply an LNP seat tally in the upper 40s (taking personal votes into account), the LNP were only favourites on the Sportsbet list in 42 seats and tied in another two. Since then the LNP have become favourites in the two previously tied seats and four others. A second market on most seats (Sportingbet/Centrebet) has also opened and between the two markets the LNP is favourite in 46 seats, Labor in 36, the crossbench in four and three are split between the two markets. On the pendulum Labor is favourite in everything up to 10.2% except Barron River (9.5) and Mundingburra (10.2). The Coalition is favourite in all the classic 2PP seats up from 10.2% except Mirani (11.2), Albert (11.9) and Kallangur (12.4) (these three are split.)
The persistence of Kallangur as a Labor favourite on one exchange (currently 1.50 vs 2.50) while other seats switched to the LNP intrigued me, and there's been an arbitrage on this seat between the two exchanges for a day and a half now. About all I can find as to why this should be the case is that Kallangur was the seat where a Young LNP member was investigated for alleged sign theft. Generally such incidents do not make a big difference to voter intention so I would welcome any hints as to why punters (or bookies) might have such a view of this seat. [Update: This arbitrage finally disappeared on Saturday 17th as a result of which Kallangur is now a tossup on Sportsbet and a very narrow LNP favourite on Sportingbet).
Overall there's been a huge blowout in Labor's overall odds (now out to $7.50 and $9), but we should remember that in SA Labor won at longer odds than that, so it doesn't necessarily mean anything. (As usual, the aim of any betting coverage here is to monitor whether betting odds prove reliable, and not necessarily to claim that they do.)
What About Redcliffe And Stafford?
This section is about Wonk Factor 3/5. Beware!
The seats of Redcliffe and Stafford changed hands in by-elections between 2012 and now. It's customary on pendulums to display seats that have changed hands in this way as belonging to the new party with the new margin. By this convention these are now Labor's safest seats (given that, in the context of the likely swing, all Labor seats are normally considered safe.)
Redcliffe is especially interesting because if you ignore the by-election in between, it's an LNP seat on a margin of 10.1%. That puts it on the kind of margin that would make it extremely dicey if you ignored the by-election, hence the following:
I thought I'd have a quick look at this: what are the historic properties of seats that have changed party from government to opposition? My sample was a quickly gathered one (10 such federal, 3 Queensland and 7 NSW such by-elections) and more data could be added from other states and older by-elections with more effort, but even from a quick sample of 20 cases the patterns are quite compelling.
I'll call a seat that changes party mid-term in this way a disrupted seat. In this sample of disrupted seats, the governments of the day recovered ten (half) of them at the next general election. But that's not useful here, because most of those governments did so in the context of much smaller state swings. Only two out of twenty recorded swings back on the scale needed to recover Redcliffe. The more useful figure is that on average the swing from election to election in disrupted seats (ignoring the by-election) was 5.9 points smaller than the state swing. That estimate of the average is +/- 1.3 points, but the standard deviation for the individual results is a rather high 5.8 points.
It's not correct to see the by-election as having reset the seat's margin to the by-election result either: in this sample governments that lost seats at by-elections on average recovered 2.5 points in those seats at the next election, despite on average suffering overall swings of 3.4 points against them. That means that in this sample, governments that lost seats mid-term outperformed the by-election result plus the overall state/federal swing by ... 5.9 points. So by-election results and previous election results seem about equally (in)accurate as predictors of what will happen when a disrupted seat next goes to the polls.
There's more: it seems that when a seat's ownership is disrupted by a by-election, the swing in the disrupted seat between the two regular elections is an inflated version of the overall state/federal swing.
That's quite a strong relationship given all the different cases and timeframes over which such by-elections can arise.
It seems that rather than thinking of Redcliffe as an LNP +10.1% seat as at the last election, or an ALP +7.1% seat as at the by-election, we could really treat it as a seat where the LNP would have a slight advantage (1.5 to 4 points depending on calculation method) in the absence of any statewide swing. That means that in the current situation it would be an easy retain for the ALP, except that disrupted seats are less predictable. Even so, for a statewide 2PP of 51.8, my modelling would only give the LNP a 10-15% change in Redcliffe, which is along vaguely similar lines to the current betting odds. The current odds should be more tempting if you think the state 2PP is going to be 54 or more, but this also applies to many other seats.
Hobbyhorse Alert: Uniform Swing Pendulums
If you plugged the 52:48 from Galaxy into the ABC Calculator you would have got only 45 LNP seats, and if you plugged the seemingly similar 51.8 from my Queensland aggregate only 42. Yet my seat projection off 51.8% was 47 LNP seats (and I think it would be reasonable to give them another from the crossbench and take it to 48). Why the difference?
Obviously, personal votes are part of it (they're worth about two seats) but I want to pick up on an Antony Green comment about the calculator. Antony writes:
"It is true that uniform swing models assume that swings are randomly and normally distributed around the average seat swing. The more that the random swing assumption is broken, the less reliable the uniform swing model becomes."
The context of this comment is cases in which one side's seats swing more than the other. But uniform swing models can also be wrong even without that and even ignoreing personal votes, and the reason is that the quoted statement is actually false. Uniform swing models do not assume swings are randomly and normally distributed - they assume swings are exactly the same in every seat with no randomness and no distribution. They break the random swing assumption in this way and make themselves less reliable compared to a model that allows for random variation.
This is very well illustrated by the calculator's output for 51.8%:
On this calculator Labor would win 42 seats with a 2PP of a mere 48.2%. However, in doing so they would have been astonishingly lucky, winning four seats compared to none by under half a percent, and twelve compared to four by under two. In reality, very slight random variation would put paid to a number of Labor's projected close wins. The uniform swing model treats a projected win by .003% or 30% as the same, one seat win, but a probability-based model treats the former as more or less a tossup and the latter as a certainty. That's why my model gives the LNP three more seats off 51.8% than the ABC calculator before it even thinks about personal votes.
Labor Don't Recycle
A question from last election was how many of Labor's defeated members would run again. In theory, re-running a former member could help blunt the personal-vote impacts of a change of seat ownership. But there's also a case to be made that a party that re-runs too many of its defeated former members is just running the same old gang who were tossed out at the last election again.
I was quite surprised to find just how few of the ALP candidates are the beaten former member having another go. Of the 33 seats in which Labor might use this tactic, they are only doing so in five: Ashgrove (Kate Jones), Brisbane Central (Grace Grace), Bulimba (Di Farmer), Morayfield (Mark Ryan) and Toowoomba North (Kerry Shine). I have not checked how many Labor candidates are former MLAs from different electorates (Cameron Dick has relocated from Greenslopes to Labor-held Woodridge - and see comments for more), but if Labor actually somehow wins this election they will have at most 11 (by my count) members who served in the previous government (plus two in a previous Federal government), and a large majority will not have served in parliament at all. The Labor team can therefore easily be packaged as fresh or charged with inexperience, depending on which way you want to spin it.
Update Jan 17: Newspoll Giveth And Galaxy Taketh Away (Supposedly)
Overnight we've had a rather encouraging Newspoll for the Premier in Ashgrove, showing him trailing only 51-49 off primaries of 47 each for him and Kate Jones, a mere 5 for the Greens and 1 for others. The poll also shows Ashgrove voters giving Newman the thumbs up as preferred Premier (51-34) and local member (51-42), though the latter +9 netsat pales compared to the +53 and +43 results recorded by Jones when she held the seat. Those stratospheric personal approvals didn't stop Jones losing to Newman last time, and won't necessarily prevent a repeat this time, though voters still (52-44) think Jones would make the better local member. Reading between the lines, it's likely there are some voters, maybe just a few percent, who are at least unconvinced Newman is the better local member but intend to vote for him anyway.
The poll is not much fresher than the ReachTEL, being taken on 13-14 Jan, and has a slightly smaller sample size. The combined meaning we should take from these two polls is that Newman really is competitive in Ashgrove now, though his position remains dicey. A cautionary note is that seat polls are often a bit unreliable, as we saw in the 2013 federal election, so things may not actually be quite that close.
The online Courier Mail has carried a rather breathless preliminary report that a Galaxy of selected electorates points to a hung parliament: "massive swings of up to 12 per cent against the Newman Government across southeast Queensland", "could be wiped out across north Queensland", "wild variations in swings occurring across the state." Now, I don't doubt there will be regional swing variation and we really do need to get a handle on the extent of it, but until I've seen the full figures of this polling I'm treating this with caution - it's a pre-election "closer than you think" line that's older than the hills and has a dismal predictive strike rate. The news that David Crisafulli's seat is close, for instance, is no surprise given that Mundingburra is on 10.2%, and nor are widespread seat losses in Brisbane and the far north news to anyone. I will have further comments when I've seen the full figures but for now ... don't get excited.
Saturday morning: And yep, the cold figures present a completely different and far more orthodox tale to the one the Courier-Mail's article is trying to spin. We have (ALP first, swing in brackets)
Townsville 58:42 (12.8)
Thuringowa 52:48 (8.7)
Pumicestone 48:52 (10.1)
Mundingburra 51:49 (11.2)
Mulgrave 61:39 (9.9)
Greenslopes 59:41 (11.5)
Cairns 53:47 (11.9)
Barron River 50:50 (9.5)
The average swing is 10.7% (more or less exactly the same as my aggregate's current statewide estimate of 11%) with all eight seats falling within 2.1% of that swing. The standard deviation in the swings is below 1.4 points. Assuming these are Galaxy's usual samples of about 550, the expected standard deviation in measured swings for such a sample size assuming the swing was exactly the same in every electorate would be about two points. (The difference between that and 1.4 points might be roughly cancelled out by the tendency of Galaxy polls to have less bouncy results than expected by chance. Maybe.) So far from being evidence of wild variation in swings between the surveyed regions, these results are actually evidence against it. The year is only two and a half weeks old and already we have a frontrunner for its most spurious report about opinion polling.
Only two of the seats polled by Galaxy are really competitive on current swings - Mundingburra and Barron River - and to have the LNP a point behind across the two of them is almost exactly as my model predicts based on recent state polling. Betting markets (or bookies) are presumably banking on a slight lift in the LNP vote to get them just over the line in these two.
[EDIT: It is good to note that the erroneous claim that Annastacia Palaszczuk "is on track to become Premier in a hung parliament, according to an exclusive poll of key electorates" has now disappeared from the online article linked above, replaced with that she merely "could become the state's accidental premier". It may be that there were two versions of the article with one version having been pasted over the other at the same URL.]