Monday, November 3, 2014

Victorian State Election: Late October Polls And Seat Model

2PP aggregate of recent Victorian state polling:  ALP leads 52.6-47.4
"Nowcast" seat estimate based on this 2PP: ALP 48 Coalition 40

After a week buried deep in the Hobart City Council count I've finally found time to get stuck into the Victorian state election.  This thread starts with a roundup of recent polling, then goes into the early version of a seat-probability model similar to my one that picked 93% of seats correctly at the 2013 federal election.  The federal election version was whipped up very quickly in the shadows of the post; I have much more time to work on this one so it will probably do much worse.

The new polls

Six polls have been released in the last week or so, by a range of methods.  Two of them have had suspiciously high Green votes.  For the Morgan SMS poll it also had a suspiciously high Green vote in late September, so I'll assume this is systematic.  The Ipsos poll is the first of its kind, and their federal poll published today hasn't shown any skew to the Greens, so I'm assuming for now that their methods are no more prone to green skew than the four more established pollsters.  Anyway in aggregating these six polls at the bottom I've pinged the Greens 0.9 of a point in every poll except Morgan SMS, for which I've applied a very lenient deduction of four.  For more information on this decision see my new blog header.


I've also weighted the polls using similar weightings to my federal 2PP aggregate, with Newspoll upweighted to 1.2 for recency.  This didn't make a lot of difference.  I've recalculated my own 2PP from the converted primaries, but it turns out this is the same as the average of the 2PPs published by the pollsters anyway.  Note here that I am using the last-election preferences for the Ipsos poll (53-47 not the respondent-allocated 56-44), though there has been some debate about whether those primaries would normally just round to 54.  The Essential sample was part of a large data dump for the three most populous states, and follows similar Victorian results for recent months.

Here's the table:


Any previous polling can be considered to have pretty much expired now that we have all this lot. An odd aspect is that five of the six pollsters supply a breakdown for PUP, although PUP are not actually polling anything worth mentioning, and have said (for what that may be worth) that they are only running in the upper house anyway.  I haven't applied any house effects based on past elections - especially note that Morgan SMS does not seem so far to share the strong two-party house effect of Morgan Multi-Mode.

Leadership ratings

The following is a roundup of leader ratings for these polls:

Approval/satisfaction:

Napthine:
Newspoll: +5 (46-41)
Ipsos: +9 (47-38)
ReachTEL: -9 (29.5-38.5), converted to 0 (38.7-38.5) by adding one third of "satisfactory" to "good"

Andrews:
Newspoll: -9 (36-45)
Ipsos: -5 (37-42)
ReachTEL: -16.5 (22.8-39.1), converted to -5 (34.3-39,1) by method above.

Preferred/Better Premier:

Newspoll: Napthine 47-34
Galaxy: Napthine 43-27 (Coalition 81-5, Labor 22-53 and I infer others 10-23 with a whopping 67% indifferent)
Ipsos: Napthine 45-36
Morgan SMS: Napthine 52-48 (I believe this method excludes any undecided option)

It's a familiar story.  Napthine is slightly popular, but not spectacularly so.  Andrews is slightly unpopular, but not spectacularly so.  Napthine would be expected to lead as preferred premier given that his government is not being totally towelled on 2PP; his lead is greater than normal for these sorts of voting intention figures, but again not hugely so.

Are 2010 preferences accurate?
With around 22% of voters apparently currently intending not to vote for a major party, minor party preferences are looking very important in this election.  In 2010 about 80% of Green preferences went to Labor and about 40% of other/independent preferences.  There was also a small leak of Coalition preferences in three-cornered contests, worth about 0.1 of a point of 2PP statewide.

In the 2013 federal election, the preference flow of Greens votes to Labor in Victoria increased from 81% to 86%. This wasn't Victorian-specific but consistent with a similar increase nationwide.  I'd tend to attribute that to the increased perceived difference in environmental policy between the major parties, arising from the Gillard government being supported by a Greens crossbencher.  It's not necessarily clear therefore that this will spill over into this election, but it might.  Especially Victorian Greens voters will be confronting Liberal governments at both state and federal levels - the federal one being seen as having especially poor form by them - while memories of issues with the Brumby government's performance may be fading.   In the event that the share of Greens preferences does shift, it's quite a big deal, since 5% of 13% is .65 of a point two-party preferred.

The preferences of the various "others" are also up for grabs.  Respondent-allocated preferences of the "others" in the Ipsos poll showed a substantial lean to Labor (63:37), but past elections suggest that it tends not to turn out like that on polling day.   The preferences of Others in Victoria excluding independents rose from 40% to Labor in the 2010 federal election to 49% in the 2013 .  This is not only because the mix of Others was different but also because the preference share of most of the Others that were the same as 2010 also increased.  

I have deliberately excluded independents from the calculation because their 2013 share consisted largely of Cathy McGowan votes, many of which were cast by Labor and Greens supporters voting tactically.  Even so there is the suggestion that Others votes might not be as Coalition-friendly as in 2010.  Here, again, there's possibly as much as .8 of a point 2PP hanging on this question.

So while polls are showing Labor at about 52.6 2PP based on 2010 Victorian election preferences, there's the possibility they will actually do up to a point and a half better based on preferences than their primaries suggest.  It would probably not be that much, but it might be something substantial anyway; say, half that.  In the federal election an increase in Others and Greens preferences to Labor largely cancelled out a tendency by most pollsters to have the Coalition's primary vote about a point too low.  That was a situation where the wheels were still falling off Labor's cart even into the final days of the campaign. It's possible that that cancelling factor won't be there that time and that Labor's current position is actually well into the 53s.

(My thanks to Adrian Beaumont and William Bowe for cc:ing me into a discussion on this point.)

Seat model ahoy!
I spent much of last night assembling the first version of my seat model.  This is a preliminary run as I hope to later augment it with seat polling, of which there is currently none that is neutrally-conducted that I know of.  The idea of the initial run is to get a feeling of the sort of performance Labor requires to win.

The methods of the model are as follows.  I start with Antony Green's post-redistribution pendulum.  From this I vary the pendulum results based on assumed personal votes.  A personal vote is gained if a party had a new candidate who won in 2010 (unless of course that candidate is retiring at this election).  It is lost if the sitting member retires.  It is counted doubly in "double sophomore" cases where the incumbent is a new member who won the seat from an incumbent of the other party at the last election.  

The redistribution presents some challenges here.  Typically there is compensation to a party (from the redistribution 2PP base) if part of an opposing-party seat is redistributed into one of its seats, since this means the seat no longer has a personal vote for its opposing-party member.  The compensation runs the other way for a same-party seat.  However the baseline is the state of the seat in 2010, not now.  So if a seat was won by the Liberals from Labor in 2010, and part of it is absorbed into a Liberal seat, then this is a PV gain to the Liberals, since that part of the seat had a Labor incumbent going into 2010, but has no incumbent on the ballot paper now.

Frankston was won by Geoff Shaw as a Liberal from Labor, but Shaw has quit the party and will contest as an independent and apparently heavily lose.  This might seem like a loss of personal vote for the Coalition.  In fact, the electorate goes from having a Labor personal vote factored in to having no personal vote factored in, which is effectively a personal vote gain for the Coalition - all else being equal.  It's quite possible dissatisfaction over Shaw will trash the Coalition's vote in this seat, so I've flagged Frankston with a "caution" concerning the model output.

I've set the personal vote factor at 1.5% for now, but may tweak it later when I've had more time to look at how much PVs are worth in Victoria.  Once I've thrown in all the personal-vote effects, a uniform adjustment is thrown in on each side to get the 2PP back to the preset state figure.  (This turns out to be about 0.32 points off the Coalition and onto Labor.  If there is evidence that parties with many sophomore seats are more likely to get better 2PPs than their polling I am not aware of it.)

The model aims to convert a 2PP into an expected median seat result by summing seat probabilities based on a given 2PP (this gets around problems caused by lopsided pendulums where one party has a lot more close seats than others).  From the set swing I can set a standard deviation of swing per seat, which for the initial run I've set at 3 points.  I should stress that the model does not give predictive probabilities that are independent of 2PP.

The following is the output of the model if there is no 2PP swing at all. The "swing" column gives the raw swing required for the other side to win the seat, the "adj" column gives the adjusted swing and the "prob" column gives the modelled chance of that side holding it.  (A couple of seats have been marked "Caution" - eg Morwell is no certainty for the Nats off any state 2PP because of the coal mine fire.)


With no swing at all (ie 51.6% to the Coalition), the 48-40 notional margin to the Coalition is retained.  The reason for this is that while the Coalition has a lot more close seats based on raw swing required, when the swing is adjusted for personal vote effects, only a few Coalition seats are still tossups, while Labor has five targets below 2%.  This is mainly because the Coalition has "double sophomore" on its side in most of Bentleigh, Forest Hill, Prahran, South Barwon (etc).  In Mordialloc and Carrum the Coalition has double sophomore in about half each electorate, and a single net PV effect in the other half.  

Now this is what the model looks like with the same swing as recent polling, ie 4.2% to Labor and a 52.6% to Labor 2PP:


At this stage, 52.6% to Labor results in a gain of about eight seats, for a rather modest 48-40 victory. If I take the most extreme path on preferencing and give Labor 54% then the expected result becomes around 51-37.  If I give the idea of improved preferencing for Labor half weight then they probably get 49 or 50.  

I can also use the model to investigate the share of vote required for a 50% chance of victory.  On this basis if the Labor 2PP is 50.0% or less then the Coalition is likely to achieve an outright majority.  If it is 50.6% or more then Labor is likely to achieve an outright majority.  If it is between those figures then the possibility of a 44-44 tie may mean that neither side is more likely than not to win outright (I have not tried to model the probabilities involved for these cases.)

Now this finding of a slight skew in the Coalition's favour for a given 2PP result is quite different from what you'll get if you just use the pendulum and assume uniform swing without adjusting for personal votes.  Without personal votes, a 1% uniform swing would take (retake in some cases) five seats and win the election for Labor with a 2PP of just 49.4.  In fact this isn't likely to happen at all.

Seats for other parties
Ah, but what about the ABC Calculator that is showing three seats to the Greens?  Well, actually, it depends which version you look at.  The raw version doesn't (unless you manually override them) but for whatever reason versions based on polling (like the Newspoll) have uniformly had three seats set to go to the Greens.

Some of the poll results recently have been such as to suggest the Greens would win seats, but most haven't.  Most have suggested a slight increase in the Green primary but a far larger increase in Labor's, meaning that there is just no basis in these polls for giving the Greens any seats at all.  It may well be that they focus campaigning tightly in these seats, or that the Liberals don't make such a show of preferencing against them (the killer factor in 2010) but we need to see evidence for these things before making any changes.  The current calculator formatting implies the Greens could be reasonably expected to win three seats on most of the included polls, and that's rubbish.

I'm also waiting on seat polling to see if there is any reason to override the model output in other potential non-classic seats such as Morwell, as well as in any conventional 2PP seat for which seat polling happens along.  The approach will be the same crude para-Bayesian method as in my federal model: a seat poll is a probability distribution which can be used to condition a prior assumption about probabilities in a seat.  For a seat the model was saying is lineball, a seat poll can be strong evidence in favour of one party.  However if the model predicts a party to win a seat easily, and a seat poll says otherwise, then the two are in conflict and the truth is probably somewhere in the middle.  For the "caution" seats Morwell and Frankston any seat polls will be much more heavily weighted.

Win Probability
At this stage this model works as a nowcast based on a set 2PP figure only.  It does not attempt to assess the probability of Labor winning, as this would require me to examine the predictiveness of polling this far out from a state election in great detail.  Every election one sees numerous models dragged out that attempt to imply that the party behind has virtually no chance of victory, but do not consider the possibility that the underlying 2PP will change.  Even the standard Monte Carlo runs to give a win chance for a given 2PP can be deceptive, because seat probabilities are not independent.

However, the solidness of Labor's leads thus far, and the historic pattern of same-party federal government drag  are both consistent with Labor being clear favourites at this stage.  (The crude version of the federal government drag model in that link predicts 39 Coalition seats, while on current federal polling the more complex version predicts 36.) The Coalition are probably at least two points from keeping office and at this stage without any specific thing threatening to change it, that's substantial.  But it's not as if the incumbents are being hopelessly thrashed.  

Another model
Another model based on seat probabilities has been conducted by Alizarin Indigo at Daily Kos: Victorian State Election Modelling - With Venom.  This one, based on polls prior to Newspoll, estimates 37 Labor seats.  I am unsure to what extent it takes sitting member effects into account [edit: more than mine at present actually - see comments - it's coming off a higher 2PP estimate for Labor].

Note also that my model replicates conclusions by Peter Brent (What Vote Will Labor Need In Victoria?).  People who are taking personal votes seriously are all getting about the same thing here. 

I will post links to any others seat models that I see that surface publicly.

Comments welcome
The clerical effort in compiling the personal vote effects across the various redistributed seats was a large one especially after the effort I put into the council elections, so if you spot any apparent errors please let me know.

7 comments:

  1. Your model looks great Kevin!

    I haven't updated since the Newspoll because it actually didn't change my calculated two-party preferred result to include it, obviously I calculate these things a bit differently. If I run Venom with a Labor 2PP of 52.6 I get Labor leading the Coalition (in seats) 48.13-39.87. So less than .08 of a seat different from you!

    My sitting member adjustment is based on calculated strength for previous reelections unless members were election for the first time in 2010, or subsequent by-elections, in which case they receive an adjustment of 3.1% (a bit higher than your estimate, and subject to redistribution effects).

    ReplyDelete
  2. Kevin, while PUP are not polling very well, my guess is that name recognition and Palmer money (if there is much left) might give them a decent chance of one or two seats in the upper house if they can do well on preferences. The Democratic Labour Party's success shows that while it is unlikely that a party could win with a tiny share of the vote, 3-4% and good preference deals can give you a good chance. The same goes for the Sex Party, if they can get off the ground.

    ReplyDelete
    Replies
    1. Yes I have not had any time to look at the Upper House yet so at this stage can't rule out that anyone who can rustle up a few percent (perhaps less) might get seats there.

      Delete
  3. Hi Kevin, I must say I'm a little sceptical about the weight being given to "personal votes" and "sophomore effects", by both you and William, in urban seats in state elections. I agree that in country seats personal vote counts for a fair bit, but I very much doubt that many urban voters know or care who their state MP is. (I'm talking about Victoria here, no doubt in Tasmania they do know.) I really don't believe that in four undistinguished years in government obscure backbenchers like Elizabeth Miller or Donna Bauer have accumulated 1.5% worth of personal vote. My longtime belief has been that if there is a swing on in a metropolitan area, the suburban marginal seats go down like dominoes, with very little regard for the qualities of the sitting member. (Adam Carr)

    ReplyDelete
    Replies
    1. Thanks very much for the comment. I hope to have time to review this empirically based on historical data and see whether the model's weightings for PVs need tweaking in either direction or splitting into rural vs urban.

      When there is a swing against a government that has been in for a while that domino effect would be expected irrespective of personal votes, because any personal qualities of a long-term sitting member in a marginal seat would already have been factored into their vote at the previous election. (Hence Howard losing Bennelong etc).

      Delete
    2. Empirical testing bears out a fair bit of the scepticism there. Looking at data since 1992 (thanks to the Psephos website!) the model's current adjustment of 1.5 points for a retirement turns out to be correct on average. However this comes out to an average of about 4 (!) points per rural seat and 1 point per non-rural seat. Looking at double sophomore cases the model's current adjustment of 3 points turns out to be too generous with an average actual effect of 2.3. There doesn't seem to be a difference between urban and rural seats there.

      I will throw this stuff into my model as it should improve it, and especially fix some seats where it is currently more gung-ho for Labor than anyone else seems to be (despite a generally conservative seat result compared to other models).

      Delete
    3. Adjustments applied. Firms Labor in a bunch of marginals but gives Ripon back to Coalition.

      Delete