Wednesday, November 21, 2018

A Token Post About Modelling The 2018 Victorian Lower House

VICTORIA (ALP 45 COALITION 38 IND 3 GREEN 2)
Seat modelling on assumed 2PP of 53.4 to Labor currently gives estimate around ALP 47 Green 3-4 LNP 36 IND 1-2 
On current numbers Labor are very likely to win, but at some risk of doing so in minority
Contest badly lacks sufficient recent polling data so any modelling is unreliable

State opinion polling aint quite what it used to be.  At this stage of the 2014 Victorian state election, there had been twelve statewide voting intention polls by six different pollsters released in the previous two months.  This time it's four by either two or three (depending on how you treat Newspoll/Galaxy) and the most recent one was commissioned.  Over a million voters have voted already (including those whose votes are in the mail) and yet so far this month the only statewide poll we've had is a ReachTEL for the Victorian National Parks Association.  There may well be a flood of polling in the final days, but at this stage, those of us trying to predict what might occur have not a lot to work with.


That statewide commissioned ReachTEL was a puzzle because it had Labor at a massive 56-44 lead, the party's biggest for over two years.  An outlier or a continuing trend from the 54-46 Newspoll in late October?  The flow of respondent preferences in that poll might be suspected of being the cause, but in fact it is about the same as the flow of preferences from the 2014 election (which was almost 70% to Labor).  Those inclined to discount the result might also look at basic sample error, or at the question of how uComms "powered by ReachTEL" polls are different, if at all, from standard ReachTELs.  Another possible cause of error would be the following suite of issues questions, which combine priming, pony-polling and skewing to create an impression that got quite a hostile response from some respondents.  Questions after the voting intention question can influence the voting intention result if the respondent hangs up (thereby discarding their data).  A long time ago the company told me the rate of this was negligible but it would be interesting to know if this is still (always) the case.

So there's the hint in one poll of a lopsided contest.  If you want to ignore that one (and commissioned polls always require extra caution) then the Poll Bludger poll tracker had things at 53.3% to Labor a few weeks ago; my aggregate of the recent public polls is similar at 53.4.  I'm also aware of one unpublished recent result that had Labor comfortably in front, which all seems to be consistent with what the parties are saying.

We have not seen many seat polls this election, and those we have seen have been confined to seats Labor could lose, with the exception of Prahran held by the Greens.  There have been no seat polls of seats Labor hopes to win from the Coalition, nor of seats the Coalition could lose to independents.

Seat model output

The usual way I model state elections is through a conditional probability model.  A conditional probability model takes the seat's existing margin, changes in personal votes and (at a low weighting) seat polling into account and estimates the chance of each major party winning the 2PP vote in each seat assuming an otherwise uniform swing of a certain amount.  This seeks to address two problems with the conventional pendulum model - firstly that it is ignorant of personal vote effects and secondly that in practice seat swings vary between seats, meaning that very close predictions by the pendulum method are almost as likely to be wrong as right.  

Because the aggregated 2PP estimate across polls can be wrong, this method doesn't give the same sort of raw chance of each party winning a seat as bookmaker odds can imply (the bookmaker odds are usually closer).  The question I try to model is: if the statewide 2PP is X, how many seats can each party expect to win?  

Normally that method will provide a pretty accurate prediction, but in this case we don't have many state polls and the ones we have are old, so the estimate of the statewide 2PP is very rubbery.  However, I'll give the example output anyway: this is what I get for an assumed statewide 2PP of 53.4 to Labor:



Seats with a large population increase are shown with a P in brackets, meaning their results should be treated with more caution.  Some seats requiring caution on the 2PP front for other reasons are shown with a note to that effect.  Prahran and Shepparton are excluded.

The model finds that on balance for a 53.4% statewide 2PP, Labor would win the 2PP in about three L-NP held seats, and lose the 2PP in about one of its own, for about 49-50 2PP wins.  They would also probably win the 2PP in Prahran, which might or might not help.

However, winning the 2PP doesn't necessarily mean winning the seat.  Four seats in which Labor is certain to win the 2PP are ALP-Green contests, two of them currently occupied by the Greens.  So in the best case for the Greens (five seats) Labor would only be on, on average, for 45-46 seats of its own, and might not even win a majority off 53.4%.  See below for my attempts to model the Greens seats.

Even if I give Labor a thumping 2PP win (say, 55.5%) this is still only good for an expected tally of 54 Labor two-party seats.  If the 2PP weakens from the recent state polls and the Greens poll OK, it is very easy to end up in hung parliament territory.

The other thing notable about the seat model is that for a small swing, it doesn't identify any seats that individually should change hands, except Morwell which must be treated with a lot of caution anyway.  This has been a general issue with modelling this election - it is hard to identify specific seats anywhere that should fall.  Perhaps like Queensland it will be an election with relatively low seat turnover.

More seats of interest

In my intro to this election I noted the seats of Prahran, Northcote, Morwell and Ovens Valley as seats that were interesting for reasons other than being Labor vs Coalition marginals.  Here are some more:

Pascoe Vale 

This is a ridiculously safe Labor seat (16.8%) by normal standards but independent local councillors Oscar Yildiz and John Kavanagh - especially Yildiz who has been running hard for quite a while - are proving to be more than token nuisance value, to the point that Labor has had the sandbags out in a seat that wouldn't normally be on the radar.  Labor's primary vote in Pascoe Vale in 2014 was below 50% and the seat is nicely set up for independents because the high Green vote and low Liberal vote provide ladders that they can climb over.  However, the Greens have preferenced only Kavanagh ahead of Labor, and Yildiz has come under some fire from the left for not saying who he would back in a minority parliament.  There has been serious interest in Yildiz on the betting markets, though whether it is informed or not who can say.

Yan Yean

Yan Yean (3.7%) was interesting anyway because it is a marginal with rapid population growth, but it has also become interesting because the Liberal candidate Meralyn Klein has been disendorsed over her appearance in a video for the Australian Liberty Alliance.  However, Klein has continued campaigning as an independent.  This is a similar circumstance to that in which Pauline Hanson was first elected to parliament in 1996, but I think parallels are overstated.  Yan Yean is not Ipswich and it is more likely the seat will be a 2PP blowout as with other recent cases of late disendorsement.  Labor has predictably failed in its attempt to have ballot papers for this seat pulped in view of Klein's disendorsement but has been able to get a provisional ruling against the Liberals' state how to vote card that still shows Klein as if she were the candidate.  

Benambra

Benambra (Lib vs ALP 9.7%) is another rural seat under fire from independents, in this case Shire of Indigo mayor Jenny O'Connor and also Cathy McGowan advisor Jacqui Hawkins.  No seat polling has been seen but an intervention from local doctors garnered a fair amount of media interest.  

Alas for the doctors, the logic behind their call is not as easy to follow as that for installing Susanna Sheed in Shepparton last time around.  Sheed was installed to punish an incumbent government for neglecting "safe" rural seats (apparently resulting in Labor funding services in the seat to help her hold it) but in this case the call is to make a Coalition seat marginal because the Labor government is ignoring it.  But voters aren't going to vote Labor to punish Labor for ignoring a seat (even if it might be in their interests to do so) so the only way the voters can really get Labor's attention is to actually elect an independent.

Geelong

Geelong (ALP 6.0) isn't looking like going anywhere on a 2PP basis, but it is being disrupted by flamboyant former mayor Darren Lyons, who was unceremoniously sacked by the Government in 2016 along with the rest of his council.  A seat poll of Geelong found Lyons on just 15% to the Liberals' 28%, but even that could create an impetus for strategic voting for Lyons by Liberal voters, since they don't look like beating Labor but he might.  Even so, I am a bit sceptical.  The problem for Lyons is that anecdotally at least, Green voters see him as right-wing and aren't all that likely to preference him in competition with Labor. But it will be interesting to see how much better his flow from the Greens might be than that the Liberals would get - that is, assuming he gets into second and we find out.

Mildura

Mildura (Nat vs IND 8.0%) is being treated with significant caution by bookmakers off the back of the Shepparton result in 2014 and the general rural indie push.  The independent challenger, Ali Cupper, was the Labor candidate in 2010 before running second as an indie in 2014.  I am not aware of anything in particular that would cause an 8% swing here, but someone seems to think they are.

Population Growth Marginals

The seats of Bass (Lib 4.6%) and South Barwon (Lib 2.9%) are of special interest beyond being marginal Liberal seats because of their rapid population increases.  This makes swing results harder to predict as large numbers of voters in these electorates didn't live there or weren't old enough to vote at the last election and there have been arguments that the regional breakup of new voters in Bass especially could produce an inflated swing to the government.  However, in the case of Bass it is worth bearing in mind that the sitting member has a personal vote bonus, having been elected at the 2014 election.  That's why my seat model above isn't all that interested in the seat, but it is probably wrong not to be.  As noted there has been no public polling of these seats.

Labor vs Green seats

The Labor vs Green seats (Melbourne, Brunswick, Richmond, Northcote) continue to be of interest.  One YouGov-Galaxy seat poll had Labor comfortably holding Richmond (54-46).  This is an interesting result because it is counter-intuitive - the Liberals' withdrawal from Richmond altogether, and their issuing of an open how-to-vote card in the others should, all else being equal, sink Labor in all these seats by itself.  Also a major argument against the Greens in Richmond (that Kathleen Maltzahn is a bad candidate) doesn't have legs because she was also a bad candidate last time, so that is already factored into the swing.  

There are parallels here with the Batman by-election because as with that one, the Greens have had an unhappy campaign.  However this time the fresh problems haven't involved the specific candidates for those seats - they have mainly involved a series of social media misadventures for three uncompetitive candidates and a staffer, and it's not really clear if anyone cares.  Also, in the Batman case Labor was running off the baseline of having had an awful candidate, from which they were bound to improve.  

Notably two of the three betting markets I checked are not buying the Galaxy result and have the Greens favourites in Richmond.  

I have constructed a model based on state primary vote polling and taking into account expected changes in Liberal voter preferencing (which I have assumed to reduce flow from Liberal to Labor by 20 points), personal vote effects, and the disruption effect of the Greens winning Northcote at a by-election, as well as the public seat polling at a low weight.  Currently this model gives the following output:

Melbourne expected margin 7.8 to Greens, 100% Greens win (Labor don't agree!)
Northcote 2.2 to Greens, Labor win 24%
Brunswick 1.2 to Greens, Labor win 35%
Richmond 0.4 to Labor, Labor win 56%

However, this is very sensitive to any evidence of a meltdown in the Greens primary vote, which we might see in any further polls that might be released.  

In the case of Prahran there's a lot to be said for the betting markets' confusion about who will win the seat.  Unless Sam Hibbins has picked up a larger than usual personal vote (as Greens have when they acquire Labor seats) then the statewide movements on primary vote would be enough to be slightly more likely than not to push Labor's Neil Pharoah into second.  (Commissioned seat polls suggest much more likely, but I'll treat that with a lot of caution.) Whoever wins the Labor vs Green battle for second will probably beat the Liberals on 2PP, but they might not (the Liberals have a personal vote loss to contend with, so the Liberals' chance of winning comes out at around 25%, depending on who they are up against.  

All up I currently get the Greens on for about 3-4 seats, but if the campaigning issues bite them and this is shown in the final polls they will be marked down and might only then be projected for one.  

Betting markets

Betting markets are not necessarily predictive but they are worth noting to see if they prove any more useful than anything else doing the rounds, especially in a low-information election.

I have compiled the following assessment from three bookmakers.  If a seat has two or more candidates at $3.00 or less on any site I consider it a "close" seat:

Expected ALP loss to Greens (close): Brunswick

Split market (ALP held vs Greens): Richmond

Split market (LNP held vs ALP): South Barwon

Three cornered with no odds-on favourite: Morwell (IND held), Prahran (Greens held)

Close expected L-NP holds: Ripon, Burwood, Bass, Mildura (vs IND)

Close expected ALP holds: Bentleigh, Carrum, Mordialloc, Frankston, Eltham, Pascoe Vale (vs IND)

Close expected Greens holds vs ALP: Melbourne, Northcote

Close expected IND hold vs L-NP: Shepparton

So the markets have only three seats that they agree are likely to change hands, and in two of those cases they can't decide who to.  The other point to note here is that while more ALP seats have made my "close" seat list, they are not as close in the betting as Ripon, Bass and especially South Barwon.  There are also some Coalition seats not far outside my "close seat" threshhold.

An overall read of the seat betting markets is that they collectively don't expect a lot of change and think the election could be very close.  There are also markets for the difference between Labor and Coalition seat totals, which on average are expecting the Labor total to exceed the Coalition total by only five seats.  Overall the betting markets are very slightly more pessimistic for Labor than my models, and may be factoring in expected closing of the 2PP margin seen in the previous polling.  One obvious issue that might be driving this is the Bourke Street terror attack, and we have not yet seen any public state polling taken since that.  

In most respects the Coalition campaign seems somewhat desperate, with Jeff Kennett being dragged out to argue that this is 1999 all over again (it isn't, because the federal drag effect runs in the opposite direction), and even an attack website being set up with the word "sucks" in its URL.  I wouldn't be at all surprised if Labor actually holds office a lot more easily than my modelling suggests (by polling a higher 2PP, by a collapse in the inner-city Greens vote or by a lucky distribution of results in the marginals) but for the moment there is not the evidence in public polling to support that.  Perhaps there will be by Saturday morning.

I will have live commentary of the Victorian election on Saturday night and aim to have extensive postcount coverage of both houses over the coming weeks.

1 comment:

  1. Interestingly, virtually all of Labor's negative advertising in the past week or two has been focussed on the federal Liberals ("send Big Bad Meanie Morrison a message!!" type of stuff). Not much direct attack on Guy or that state issues at all.

    I wonder what that means.

    ReplyDelete