Saturday, November 29, 2014

Victorian Election: Final Aggregate And Seat Model

 Live analysis here on a new thread from 6 pm on election night. 

2PP Aggregate using 2010 preferences: 51.7 to Labor
Polls point to highly likely but not certain ALP victory, general picture of modest swing and minimal to low seat turnover
Seat Projection: ALP 47 Coalition 41
(A very small number of these seats may be won by Greens or independents, but there is not enough objective evidence to back such wins.)

Going into the main voting day for the Victorian election, I have to say it's a bit different to elections I've devoted major modelling efforts to on this site (the 2013 federal election and this year's Tasmanian state election) in that there is still some room for doubt about who will win.  Although Labor has led in almost every poll in the last year and a bit, including 15 or 16 (depends how you count 'em) of the 17 polls released in the last six weeks, their lead is not much, and sitting-member effects left over from the last election make the deck a slightly unfriendly one.

The polls suggest that probably Labor will get home with a seat tally in the high 40s, but other still-plausible scenarios (in decreasing order of likelihood) are a fairly comfortable Labor win with around 50 seats (just maybe if the polls have herded a few more), a scraped and very lucky Coalition win, or some kind of tied or otherwise hung parliament.  Anything else (eg a Coalition win where they win the 2PP as well, or a really lopsided Labor win) would be a serious surprise.

The final polls

The previous article (Endless 52:48 Edition) gave results for what are believed to be the final shots from Morgan, Essential and Galaxy, though with Morgan one never knows.  The Morgan, as noted, had a 50:50 2PP and indeed on my own calculations the primaries of that poll would normally imply a 2PP more like 50.5 to the Coalition. 

Coming from a volatile poll series by an under-tested method the Morgan wasn't taken too seriously but it had company with the release of the final Ipsos, which came in as another 52:48 by respondent-allocated preferences, but by batched last-election preferences was 50:50.  (My estimate without the batching: 50.3% to ALP).  This was off primaries of Coalition 42 (highest in any poll in the campaign bar the final Morgan), Labor 35 (lowest by anyone other than Morgan) Greens 15 and Others 8.

The Ipsos leadership scores had Napthine keeping a +9 netsat (49:40), Andrews down a little to -1 (42:43) but Andrews, as in Galaxy, very competitive on the preferred Premier question (trailing only 42:44) especially for such a close poll.   It's tempting to take Andrews' strong Preferred Premier results as a smoke signal that preferences are going to favour Labor much more strongly than in 2010, but Preferred Premier is such a messy indicator of anything that I'd rather not play with that fire.

ReachTEL came out with a 52:48 2PP based on the largest and freshest of the final samples: 2155 all on Thursday night. Coalition 39.7 Labor 38.3 Green 13.5 Other 8.5.  The 2PP was released as 52:48 (down one from their earlier 53:47) but the 2PP differences as implied by the decimalised primaries would have been trivial, with the Coalition-leaning Others losing two points of which one went to Labor and the Greens.

Napthine's Good ratings are up 4.2 and his Poor ratings down 2.2 for a net ReachTEL personal rating of -2.6 while Andrews records larger improvements yet still comes out at net -7.4.  I have been inclined to correct these ReachTEL leader scores because of the influence of the "satisfactory" option, but the need to do so seems to be reducing.  It should be noted they could be a bit on the harsh side.

What's notable is that while Others voters dislike both leaders almost equally (Napthine -24, Andrews -26.1), Greens voters like Andrews slightly (+8.3) but are even more anti-Napthine (-60.8, yikes) than Labor voters (-39.4).  All you need to know right there about how many Greens voters will pay the slightest attention should any of those naughty open Green HTVs see the light of day.

ReachTEL shows that voters generally think Labor will win.  However, Coalition voters think the Coalition will win.  I have always thought these questions need to be worded with an explicit "This question asks who you think will win, not who do you want to win?" and perhaps if the voter picks their own party while it's 55:45 behind, a little extra "Are you off your rocker?" would be helpful.

Newspoll with a good sample size came out with a 2PP of ... do I even need to say?  The primaries were 40-39 for the majors (the fourth to get exactly that) but a rather sobering 12 for the Greens.  This is the equal worst result (with one Essential) for the Greens the whole campaign, and I'll predict (confidence 70%) that it will also be the equal most accurate.

Newspoll's leader netsats were -4 (41-45) for Napthine (his worst by a point), -5 for Andrews and another strong performance for Andrews as Better Premier, trailing by only four (37-41).

There is possibly garbled word that the Newspoll tightened in the last two days of the poll, though it would be difficult to distinguish tightening from bouncing in such cases.  If it's real and I can quantify this I may tweak my poll model slightly on Saturday morning to adjust for it.  But for the time being ...


I currently have a 2PP poll aggregate of 51.7% 2PP to Labor, off primaries of Coalition 41 Labor 38.6 Greens 12.1 Others 8.3.  These are rather similar to William Bowe's current BludgerTrack model (51.8 off 41.2-38.6-12.5-7.6).

It may seem odd that when we've had two 50-50s (including Ipsos by last-election preferences) and three 52-48s in the last five polls that I end up in the high 51s and not at, say, 51.2.  However, firstly I am weighting the Galaxy, Newspoll and ReachTEL slightly more heavily than Ipsos (who are an excellent polling firm worldwide but impossible to benchmark against Australian election results at this stage) and a lot more heavily than the Morgan SMS polls.  To the extent that the Morgan SMS polls could be benchmarked against other polls based on repeated evidence they displayed large house effects to the Greens and against Labor, and also high volatility in tracking for their sample size.

The other thing is that most of the 52s in the campaign have been 52s and a bit when I convert the primaries into last-election 2PPs.  I convert the last three (ReachTEL, Galaxy and Newspoll) to 52.6, 52.7 and 52.3 to ALP.

Readers may have noticed a brief flirtation with the idea of an "expected 2PP" that differs from the aggregate.  I've decided not to do that, because at the federal election I could have done the same thing, and I didn't, and what I did then worked.  My expectation then is that if minor party preferences prove extra-friendly for the ALP then this may well be cancelled out by house effects of polling overall, greater conservative resilience to bad polling or whatever.  So we'll see how this goes.

In comparison with the federal election I've found the subjective choices to be made in aggregating all this Vic data are endless.  When using so much data from little-tried poll methods there is no way to make fully objective choices at all times.

Seat Model

The seat model I published yesterday used an "expected 2PP" concept I've decided not to use as the main model, for reasons discussed above, although I have a nagging intuitive feeling that I should have done so anyway.  So this then is the nowcast version for a 2PP of 51.7:

(Probabilities for seats with third parties in contention are conditional on them not winning.)

This model was first outlined in the late October polling article.  Seats listed on the Coalition side of the table include those that are notionally held by the Coalition based on the 2PP vote from the last election.  Five of these are seats actually occupied by Labor members that have been redistributed (hence the ALP in brackets); if there was no swing from last election in those seats the Coalition would win them.  

The table gives expected probabilities of the side of the table the seat appears on winning on the condition that the 2PP is 51.7.  The commonest path to a 47-41 seat result is that Labor holds all the seats it occupies bar Ripon (which is probably less close than the estimate suggests), and wins Carrum (much as the Liberals insist that they will hold it), Mordialloc and Frankston.  The model gives Labor more chance of losing Bellarine than is generally being expected (and it is probably wrong in this).  It also gives the Coalition more show of saving more than one of the four sandbelt seats (Carrum, Bentleigh, Mordialloc, Frankston), or picking off a Labor seat somewhere, than it gives Labor of picking up seats further up the tree.  So while 47-41 is the median, we can easily imagine what might happen if the Greens win Melbourne and the Coalition holds another seat, or Labor loses Ivanhoe and so on.  A hung parliament at least, even on this 2PP, isn't completely out of the question.

I may later regret not using the "expected 2PP" concept, and moving off 48 seats just as others move onto it (I have a habit of making my next-but-last model better than my final one), but for the record the "expected 2PP" comes out at 52.6% to Labor and 48 seats to 40, which is not so greatly different anyway.  I can make the 2PP in the model 54.6 and Labor still only get 52.

Wonky paragraph alert: I think at this stage it's also good to try an unconditional model in which the underlying 2PP is allowed to vary, so as to come up with probabilities that look more like the real chances of a party winning each seat, rather than probabilities that are pinned to a specific 2PP.  Undecided voters choosing on the day will often break one way or another compared to other voters.  To do this right involves digging through past data and then modelling variables that are under the spell of not one but two normal distributions.  I've run out of time for that sort of maths so as a crude substitute I've asked what would happen if I made the standard deviation for each seat five points instead of three.  I then get this:

I think for the closer seats this mostly gives a better rough idea of real-world priorities, though a few individual cases stand out as inaccurate (neither Mordialloc nor Ripon are probably that close, nor Bellarine for that matter).  For the moderately safe seats the model looks under-confident and I'd expect it to be (eg there is not nearly a 5% chance Denis Napthine will be unseated!) Some people like to get probabilities like this, run them through Monte Carlo simulations and report that Labor wins 95% (or whatever) of them, but such models are usually bogus because seat probabilities are not independent of each other.  Trying to get a really objective polling-based fix on Labor's chances would take a lot of historical state polling work.  I suspect such a model would still give them at least 70%.  

We should be extra-cautious about the idea of the Coalition winning at all because of the big-picture issue of federal-state drag and the messy nature of its term, which are enough to imply a likely moderate defeat despite it being a first-term government.  But I do think that Labor's campaign has been riskier than the typical opposition trying to get into power, with its stance on the East-West link hardly one from the modern playbook of "selective differentiation". 

Anyway, this is my broad expectation: a modest swing with minimal seat turnover.  The Coalition to probably gain Ripon and count themselves lucky to pick up anything else, Labor to win most (maybe all) of the sandbelt seats and not a lot (if anything) else.  My model differs from the bookies only in the seat of Bentleigh, and I won't be putting any money on that one.

As for the Greens I think we should take the evidence of the scale of doorknocking and resources thrown into the seat of Melbourne seriously.  But it's objectively going on out on a limb to say, with only one small, commissioned and unreliable seatpoll, that there is enough evidence the seat will fall against the backdrop of a modest if any increase in the Green vote. 

Additions/Around the traps

As mentioned above: Poll Bludger has 51.8% to Labor and 48-40.  Alizarin Indigo has 48-40 with a 90% chance of a Labor win (presumably for its underlying 2PP).  The Australian has a projection of 46-50 seats that seems to me to be just based off a straight pendulum read of the Newspoll and not anything interesting. See also William Bowe's Guardian piece where he explains how there is a path for victory for the Coalition, though they would have to be very lucky to make it.

It's notable that Newspoll has the Greens vote one point lower in "Melbourne Metro" than in the rest of Victoria; this won't happen.  Ipsos has it seven points higher.  Very likely the truth will be somewhere in between and the metro-country gap will be similar to the 4.5 points last time.


  1. I understand the Greens knocked 25,000 doors in Melbourne and 20,000 in Prahran - so while it's clear that Melbourne is their best shot for a Lower House seat, I quite look forward to seeing how Prahran plays out and what sort of numbers they get for the effort they put in.

    1. Yes these were their two target seats and there is some evidence that this sort of intensive doorknocking and conversation-starting model works (and that a lot of the old stuff largely doesn't anymore but people keep doing it anyway).

      The trouble with Prahran is that the natural Liberal vote is so high that knocking it down to the level where a Green could win on Labor preferences (not as easy as the other way round) would take a lot. Any vote the Greens took off Labor is only good for coming second so they'd need to be taking a lot from the Liberals to have a show.

  2. Looks like the. Greens must have picked up a lot of Lib votes since they have taken the seat in Prahan.

  3. Yep, though not as many as I thought they'd need.

    The Liberal primary vote was down only a bit over 3%, when they had a margin of 4.7% over Labor. So there was likely also some tightening of preferences against them (with some voters who voted Labor-Liberal or Green-Liberal at the last election now putting the Liberals below both Labor and Green). But it looks like virtually all those who switched from voting 1 Liberal put the Greens ahead of the Liberals (whether as a #1 vote or otherwise).

    One thing that was wrong with my coment above was that it actually turned out to be slightly easier for the Greens to win on Labor preferences than the other way around; if anyone had told me the Greens would get 87% off Labor before the election I would have told them they were dreaming. In hindsight it makes some sort of sense (Labor voters more likely to follow card, Labor candidate more appealling to left end of potential ALP support base and not to right, etc) But yes, there must have been a two-party swing of at least 4% and probably higher from the Liberals to the Greens.

    My suspicion is that the primary swing from the Liberals to the Greens is actually slightly higher than the number of votes the Liberals lost, and that what actually happened is that Labor lost primary votes to both the Liberals and the Greens in this electorate.