Saturday, November 15, 2014

Victorian Poll Roundup And Seat Betting Watch: Sandbelt Edition

2PP aggregate of recent Victorian state polling:  ALP leads 53.2-46.8
"Nowcast" seat estimate based on this 2PP: ALP 49 Coalition 39

It's taken a while for enough polling to build up to justify another Victorian pre-election roundup to follow last fortnight's, but that point has been reached with the belated release of results for the Fairfax-Ipsos poll taken last weekend.  Why it has taken the Age until Friday to get this data out there is anyone's guess, but at least we have it now.

If you followed the media comments about the unreliable mutterings of both parties about internal polling and the like, you'd have seen claims from both sides that things have tightened up through this week.  These claims always need to be treated with caution, since the side in the lead has an incentive to make them to discourage complacency, while the side trailing has an incentive to make them to discourage despair.   They might turn out to be true, but with no data fresher than Monday, we'll need to wait for more polling to be sure.

Based on where things stood early this week, there would have to be a lot of improvement to make a serious difference to the picture of a very likely change of government.  At that stage the gap was, if anything, widening slightly, but the broader picture is a lack of any major change for quite a while. 

This week's public polling

There were only two new public polls this week.  A Morgan SMS poll came out with primaries of 38% Coalition, 36% ALP, 18.5% Green, 7.5% for others.  (There was originally 2% for PUP in there, but Morgan decided that since PUP were not running, one point of their support could be donated to each major party.)  That came out to a 2PP of 53.5 to Labor.

The Fairfax Ipsos poll came out with primaries of 39 each for the majors, 16 to the Greens, 2 for PUP and 6 for Others (Ipsos do not force totals to sum to 100).  There is a published last-election "overall preference flow" 2PP of 53% to Labor, which matches what you get with a batched 2010 flow from all minor parties, Greens included (the batched flow can be calculated exactly, while the exact 2PP splits from specific minor parties can't be calculated from the VEC data).  My own estimate from the published primaries using estimated splits for Greens and Others is 54.0% to Labor, and the respondent-allocated figure was again a whopping 56.

The Ipsos poll includes a regional breakdown into Melbourne and non-Melbourne seats.  The sample size for non-Melbourne would be small with error margins of several points, and we shouldn't read too much into the sample's Green vote of 16 points (clearly the Greens would not actually get the same vote outside Melbourne as within it; in 2010 as calculated by William Bowe their vote was 8.5% outside Melbourne and 13% inside) nor the implied 2PP swing of about nine points to ALP.

The sample for Melbourne, if accurate, is actually underwhelming for Labor since it shows a primary vote swing of three points from Liberal to Green, without Labor getting any of it.  That would still win them enough marginals on Green preferences, but could lose them the electorate of Melbourne to the Greens (since when Green votes surge, they seem to do so most intensely in their strongholds) and might result in a fragile seat margin overall.  This is only the breakdown of one poll though; let's wait for more data on regional breakdowns.

The leadership figures for the Ipsos were of interest in that they showed a rare positive net approval rating for Daniel Andrews - it's only +3 but enough to suggest Coalition attempts to make Andrews the issue are not working.  This was also reflected in the preferred premier score with Napthine holding onto the kind of flimsy "lead" (42-39) normally associated with governments as far behind on 2PP as his was at the time this poll was taken. Napthine himself remains well regarded at a net +9, but that's hardly election-saving material by itself.

In trying to aggregate polling the biggest problem is getting the Greens house effects right.  Morgan at least has them way too high.  If I weight the available polls only by recency and my perception of accuracy, I get primaries of L-NP 38.4 ALP 37.6 Green 16 Other 8.  But the two most recent polls (Morgan SMS and Ipsos) have both had the Green vote running high compared to the more established pollsters so I'm more and more inclined to dock their Green votes until we have some fresher data points from established pollsters to see if the Greens are really booming out of control.

Noting that I'm now splitting what I deduct from the Greens 60:40 to Labor (since Morgan's primaries seem to especially suppress the Labor vote), I have current aggregated primaries of L-NP 39.5 ALP 39.4 Green 13.1 Other 8, for an estimated 2PP of 53.2 to Labor.  The caveats in the previous article about preference flows still apply - it could be a fair bit higher. My way of treating the perceived overstatement of the Green vote  Also the Others figure is too high because many polls are still including PUP, but that has little effect on the 2PP estimate.  If pollsters don't stop including PUP before the day I'll make appropriate changes.

Commissioned seat polls

Various "internal polling" has been referred to in the media without any actual figures being provided for the bulk of it.

Of most interest has been two Greens-commissioned Lonergan robopolls of Green target seats Melbourne (full results) and Richmond (full results).  Before we go further on these I remind readers of the poor performance of this pollster at the 2013 federal election, and the general trend that party-commissioned polling that is publicly released skews to the party that commissions it.  Also, the sample sizes, as usual for seat polls, are relatively small. 

Lonergan publish both raw primaries and "forced preferences".  I take this to mean that they initially ask a voting intention question that allows an Unsure option, and then those who are Unsure are asked to make a choice and excluded if still unsure.  For Melbourne the initial figure is Greens 39 ALP 26 Lib 16 Other 8 Unsure 11, panning out to 40-30-18-11 after forcing, for a published 2PP of 53:47 to the Greens.  For Richmond the initial figures are 33-25-19-9-13, panning out to 39-29-21-11, for a 2PP of 54:46.

Two things are unlikely here.  Firstly in the absence of any declared significant fourth-party candidate, there's no reason to believe Others will get 11 or even 8 in Melbourne, so most likely the Others figure hides soft major-party support.  In Richmond, where the Greens are again running controversial candidate Kathleen Maltzahn (a supporter of the "Nordic model" approach of trying to eliminate prostitution by criminalising paying for sex) the implausible aspect is that 6 of 13 points of Unsure vote would really go to the Greens.  Frequently voters who only reluctantly intend to vote for the Greens don't actually end up doing so.  In Richmond's case the Others figure of 11% is believable as an independent candidate polled substantially in 2010 and is running again.

As evidence that the Greens are favoured to win these seats I don't find these two polls convincing, and there is a history of such polling showing close Green inner-city wins that mostly don't end up happening (or even if they do happen, happen by less than the polled margins).  Throw in the general tendency of polls to slightly overstate the Green vote and it might even be said that these polls provide evidence in favour of narrow ALP wins in these electorates - if (and it's a medium-sized if) the same proportion of Liberal voters follow their how-to-vote cards as last time.

Essential Research polling on behalf of the Trades Hall Council (ie, union polling) found Labor ahead 55-45 in Bellarine (notional LIB +2.5), 54-46 in Frankston (LIB +0.4 but independent-occupied) and 52-48 in Mordialloc (LIB +1.5).  All these seats are closer than when they were last surveyed in the same way in late June.  An average swing of about 5% in these three marginals is pretty much what would be expected but I've not yet seen full results of these polls.

Unsafe Tex?

There's been a fair bit of gnashing and wailing about the news that rock musician Tex Perkins (Beasts of Bourbon, The Cruel Sea etc) has nominated for what seemed at first to be a single-issue indie run in ALP-held marginal Albert Park.  Concerns are held on the left that Perkins might preference the Liberals and cause them to win the seat. An article by Helen Razer on this matter raises various reservations about whether Perkins has any idea what he is doing, and discusses the unChain group, whose unofficial candidate scored 9.1% in the seat last time around.  Unfortunately, the article veers into misleading scare territory with implications that the "impact" of "giving your primary vote" to Perkins could include helping the Coalition win.

We do have preferential voting in Australia and this is not like the Senate - the voter not only can but must decide the order of their own preferences.  They can copy their candidate's how-to-vote card to do so if they want to, but fewer than 40% of Victorian voters copy how to vote cards exactly, and those who do may well be often making their major preference decisions the same way they would have done so anyway.  Also, the HTV card follow rate for non-major party candidates is low in general and Perkins won't be that big an exception.  For every voter who votes for him out of an obsessive focus on the state of the Palais Theatre there will be four or five who do so to give the big three parties the finger or because they like Perkins' music or have at least vaguely heard of him.  The two points or so Perkins might be able to genuinely swing with his HTV card might make the difference (if he preferences either major party at all) but on current polling Labor really should be holding this seat by enough for it not to matter. 

Anyway, the important thing here is that if a voter does vote 1 for Perkins then deliver their preferences to Labor and the Greens ahead of the Coalition, then their vote is not going to help the Coalition in any way.

Model remodelled

In my previous roundup I released my first try at a seat-based model.  I meant to improve this using neutrally-commissioned seat polling, but thus far all released seat-polling has been party- or union-commissioned and hence excluded from the model.

However, what I have done, as threatened (and after prodding from a comment by Adam Carr),  is to empirically test and tweak the personal vote settings.  Using data back to 1992 I find the following average outcomes:

* for the retirement (or non-preselection) of a sitting member in a rural seat, party performs an average just over 4 points worse than uniform swing.  (This is based on very limited data - many rural retirements were confounded by changes in which parties contested the final two-candidate vote).

* for likewise in other seats, the party performs on average 1 point worse than uniform swing.  This validates Carr's comment that there is not a lot going on with state-level personal votes in urban seats.

* a party benefiting from double sophomore effect (its candidate won the seat from opposing party's recontesting incumbent) performs on average 2.3 points better than uniform swing.  This doesn't seem to vary much between rural and urban seats, which may seem odd.  But I suspect the explanation is that rural seats tend to be safe and sitting members tend to build up large PVs over several elections, with defeated rural incumbents either being poor candidates or being in the less safe seats, and hence having shorter shelf-lives.

These figures may be a touch conservative because of redistributions in the past muting PV impacts.  But I've decided for now to change the value of a personal vote effect to 1, except for the following:

* double sophomore = 2.3 instead of 2.
* rural retirement = 4 instead of 1.
* rural loss of sitting member on ballot caused by redistribution rather than retirement = 2.5.

Here's the current output based on a 2PP of 53.2:

 Note again that this is a "nowcast" model - it aims to find seat probabilities assuming the modelled 2PP is correct.  Don't go putting your house on Labor in Cranbourne because this model supposedly said it was more or less a sure thing.  If the 2PP changes, then so do the probabilities. 

On the estimated 2PP, my model is now giving Labor 49 seats to the Coalition's 39.  On current polling as of last weekend, Labor most likely loses nothing in its notional column, wins all its five notionally Liberal seats that it occupies, and wins the four so-called "sandbelt seats" (Carrum, Frankston, Bentleigh, Mordialloc).  There is a lot of talk about these four adjacent marginals that are all on very similar margins and could easily all fall (or hold) together.  Probably Labor doesn't need them all for victory and could get away with splitting them 2-2, but if Labor does take the lot then the Coalition's task of even tying becomes extremely difficult. All were won from Labor in 2010, but the normal double-sophomore effect has been weakened by heavy redistributions in Carrum and Mordialloc and the Shaw problem in Franklin. 

Bentleigh, Frankston and Mordialloc all have the same results history going back to 1992 (all being won by Labor in 2002 then lost in 2010) while Carrum has a slightly more ALP-leaning history, being held by the party in the 92-96 and 99-02 terms. 

If the swing is anything like 4.8 points then it's fairly likely seats will change hands in one direction only, but even with a swing that large, the model's projection for Ripon still seems a tad brave.  This is not only a rural retirement but a 15-year incumbent retiring and getting much of a swing there from the notionally Liberal margin could be difficult.  In the final version of this model I will be applying manual overrides to seats in which I strongly believe the model's output to be incorrect.  At the moment I would override just this one seat, but I am not writing Labor off in it yet.

In terms of the inner-city Labor-Green contests we need more evidence as to whether there is a more serious boom in the Green vote before any meaningful attempt to model them can be made. While it may seem the Greens are tracking above their last election result, their support tends to drop off before election day, and indeed in the 2010 leadup they were travelling better in Newspoll than they are this time around.

Seat Betting

I haven't looked at betting odds for this election yet, but here are a few points of interest.  Once again, the point of this coverage is not to claim that betting odds are necessarily predictive (anyone still clinging to that claim was asleep for the $1.01 favourite losing in South Australia) but to see how predictive they are.

At the moment both the major agencies give the Coalition little chance ($5.20 and $6), though that's still better than the $12 they were giving for Labor in SA.  Both agencies have the Greens not winning any seat ($1.37 and $1.45), and Sportsbet has a market on the eventual Green vote in which 10-12% (below all campaign polls so far) is the favoured range with 12-14% second.  The seat totals markets for Labor have Labor winning an average of 49.4 seats on both markets, with 51-52 seats the shortest priced range on one and 50-54 seats on the other.  (This possibly reflects that if the swing is large for Labor it becomes harder to pick up many more seats per point of swing.)

When it comes to individual seats, Sportingbet has Labor favourites in 48 and Sportsbet in 47.  They differ only in Mordialloc (on which there is currently a nice arbitrage for those into such things!) Beyond that they differ from my model only in that both have the Coalition favourite in Ripon - a seat I've flagged with a caution note anyway.

When it comes to close seats (a close seat being defined by me as one where two parties are inside $3 on at least one exchange), of the 47 seats Labor is favourite in on both exchanges, only Carrum, Bellarine and Melbourne (vs the Greens) qualify.  Of the Coalition's 40 such seats, Morwell, Ringwood, Burwood, South Barwon, Forest Hill and Ripon all appear in the close seat list. Some of these seats are barely rating a mention in media discussion.

At the moment there is very little difference between what these markets think and my interpretation of what the polls imply.  So it's interesting then that a Sportsbet 2PP market thinks we'll be finishing up around 52%, which would probably only be good for 47 Labor seats.

I'll probably post minor updates to this piece through this week as new information comes to hand and lay off a major repost of the model until we have a few more major polls. 

(See also Poll Bludger.  Note that the Bludgertrack nowcast model and mine differ on only one seat, Forest Hill, which both regard on current polling as more or less a tossup.  The betting markets don't agree, most likely because they factor in the assumption that the 2PP will more likely narrow, moving that seat out of reach, rather than widen.)

Late Night Update: How I'm Modelling Seat Polls: We have our first "neutral" seatpolls, both by Galaxy in the Herald-Sun and both apparently more Coalition-friendly than last week's statewide polling would imply. The results reported so far are a 2PP lead of 52:48 for the Liberals in Bentleigh, which will raise a few eyebrows, and a 54:46 lead for Labor in Buninyong.

My model averages results from seat polls with the projections based on the state polling, since both are probability distributions with a roughly similar margin of error.  If there are multiple seat polls for a seat then the seat polls carry a higher weighting.  However, after the federal election in which seat polls generally skewed to the Coalition (with Galaxy's the least prone to doing so), seat polls are not allowed to affect the average 2PP estimate.  Also, if seat polls show a systematic lean against either party in its own marginals that is not explained by personal vote effects, then it will be assumed this is probably incorrect and the seat polls corrected to account for this.

I'm waiting for full details of the Galaxy polls before revising my model and setting a weighting for them.  However, Bentleigh was projected at about 52 for Labor, so for the time being a seat poll with 52 to Liberal brings it to somewhere near square.  In the case of Buninyong the poll has little effect on assumptions that on current support levels Labor is fine there.

Sunday: More On Bentleigh Poll: The Bentleigh poll turns out to have a sample size of 704 and primaries of 48 Liberal, 35 Labor, 11 Green.  It's actually quite difficult to get this down to a 2PP of 52 rather than 53, and that, combined with the large sample size, is actually enough to give the Coalition a win chance marginally under 50% in my model, but only after assuming a house effect of a point to the Coalition based on federal election seat poll results. (The model still gives Labor 49 seats since it gives it close enough to 50% in Forest Hill as well.) This should all be treated with caution until we see more statewide polling - but it is vaguely consistent with the Ipsos poll sample with a lack of swing for Labor on primary vote within Melbourne.  Following the poll the Coalition has been slashed from $3 to $2.50 for the seat at Sportsbet.

Wilderness Society Poll Eltham and Prahran: A Wilderness Society commissioned ReachTEL has reported predictable levels of inner-city support for a Great Forest National Park in Eltham and Prahran. 52% of swinging voters are reported as saying the issue would influence their vote, but this is just the usual suggestivity effect - if you offer voters any issue and then ask them in isolation if it matters to them, an inflated proportion will say that it does.  The bipartisan consensus on this issue (and especially the perception that Labor have been pulled into line by the CFMEU) may well provide oxygen for the Greens in inner-city seats.  What is not clear yet is whether voting intention figures were taken and if so what they were.  Also, as usual with activist-group polls, no stock should be placed in the results until the exact question wording and ordering has been seen.

Friday Update: Galaxy

We finally have a new statewide poll and it's a Galaxy, with a 2PP of 52:48, with the Coalition on 40, Labor on 39 and the Greens on 13. Napthine's lead over Andrews is again higher than in the Ipsos, but has still come down to twelve points (42:30).

The update to my model knocks four-tenths of a point off Labor's aggregated lead, giving the Opposition a 52.8% 2PP, which doesn't quite change any seats, though the model has Bentleigh on a knife-edge and the predicted overall total is only just now rounding to 49 rather than 48.   On weekly update effective from midnight tomorrow all polls bar the most recent Galaxy, Newspoll, Ipsos and Morgan will be deleted, which at present improves Labor's position by 0.1% 2PP.

I am not yet using the same method as used in my federal aggregate of boosting the most recent polls in the last weeks of the election campaign.  The reason is that with relatively few polls this could lead to individual polls swamping the weighting, in this case to this single Galaxy carrying 59% of the figure, which for a sample of 924 seems a little bit much, as reliable as Galaxy are.

And Morgan: The new Morgan SMS poll was more than a little difficult to take seriously, with a Green vote of 19.5% (at least they're consistent).  The 2PP of 55:45 also seemed a shade generous after the Galaxy, but perhaps that will prove a shade stingy.  I've further downweighted the Morgans in my aggregate because of their persistently huge Green votes, but even so this Morgan and the Galaxy are about the only data we have got that hasn't gone off in the fridge yet.  Including the Morgan my aggregate for Victoria is now a flat 53:47 to Labor, which still comes out at 49 seats to 39.  I do think that if the Greens can poll what my aggregate currently indicates (which is around 13%) then it's reasonable to throw them a crumb, so for now as a nowcast, 48-39-1.  But I will reconsider the question of the Green seat with more polling as it may well dive closer to the day.


  1. If opinion polls conducted off such low samples have such a big margin of error ( as the Greens commissioned polls will have ) why do established polling organisations bother to conduct them? Surely they are undermining their own integrity and reputation by allowing it if it simply comes down to how much the client is willing to pay? Or is it down to the client wanting a result as quickly as possible, which limits the opportunity to get a more respectable sample size?

    1. It's usually down to the client wanting to strike a balance between cost and accuracy, and the market realises that these limitations come from the client rather than the pollster. Errors in small-sample commissioned polls don't necessarily reflect on whether the pollster will do a good job when asked to do an accurate large sample poll.

      Sometimes with flawed polling it is that way because the client actually wants a misleading answer so they can promote that answer. Especially the Greens love to publish friendly polling for themselves because they think it energises the troops. It doesn't do a pollster's image among potential activist sponsors any harm if it looks like they're willing to do what it takes to get the answer the group wants to publish.

  2. Could the Others figure in Melbourne be Sex Party support? They managed 7% ať the by-election, and they have received a good deal of media coverage.

    1. Any party that calls itself the Sex party isn't serious about going anywhere, despite some reasonable liberal policies, which they seem to have snitched from the Greens. I think people are just having a laugh when they vote for them, to be honest.

    2. Sex Party did get a high vote at the by-election but that by-election had no Liberal candidate (admittedly it did have an enormous pack of indies). Also they were extremely prominent in the lead-up to that by-election, especially after declining to preference the Greens because of anti-sex-industry elements (ie Maltzahn) in the party. They would be a significant part of the Others mix and should get at least 4% but I think they'll be doing very well if they repeat their vote from the by-election.

    3. Presumably, the Sex Party are trading off the media attention generated by their provocative name and advertising. If they were called the Libertarian Party or the Freedom Party, even if they had the same policies, they would get nowhere near the support that they get now.

  3. So it's plausible that clients could ask for 3 mini polls to be conducted around the same time on the same issue and only publicly release the best one, considering the results would likely vary based on the larger margin of error involved? No obligation for the polling company to reveal such a tactic has taken place, in any way? Quite unethical really. Another thing, what do you think of the two recent Roy Morgan polls on state voting intention off really low samples that produced almost identical results. Does that allignment mean the results could be treated as reasonably accurate and on par with poll outcomes ( despite their idiosyncrasies ) from firms EMRS and Reachtel?

    1. Yes, there is no obligation on pollsters or parties to reveal the existence of unreleased polling and nothing to stop parties and unions from cherrypicking which results to release. But this isn't the only way it happens; pollsters hired by parties may subtly and even unintentionally do things that skew either what data they get or the way they analyse it to tell the parties what they think they want to hear, or give them what they want to release. I don't get the impression the Greens are sitting on vast masses of unreleased robopolling the way the big parties and some unions are.

      I assume in the last part you're referring to Morgan's Tasmanian SMS state polls with samples around 300. The answer is: not yet. Although the two polls got very similar results to each other they were way out of whack with the recent EMRS and ReachTEL polls and showing a suspiciously big swing from the state election. One possibility is that voting intention has drastically changed since the last polls by anyone else, but another is that Morgan's SMS polling in Tasmania has a really big left-wing house effect. Not certain of this but I believe Morgan SMS polls sample from a panel (a la Essential) rather than random-sampling the entire population. Morgan's Tasmanian federal samples have often appeared even more ALP-skewed than their results overall. With the release of the next EMRS it will be possible to have a go at benchmarking the Morgan results and then making some kind of use of their future polls of this kind, but I am thinking that to do so I will have to correct them more heavily than the others. Even so, with such small sample sizes any given poll will need to be treated with great caution.

  4. But PUP are running candidates in every upper house electorate?

    1. Yep, they're not running in the lower house, only in the upper house where they think they might get lucky with the group ticket bingo game. So they're not relevant to lower house polling. Will be interesting to see how their upper house vote compares to the 2% or so they've been getting in the lower house polls that include them.

  5. Fair go're showing your Green bias here. So if we take party names as a guide to their policies then the Liberal Party have liberal policies and Labor has policies that support all workers. And the Greens, who have just done a preference deal with Palmer's mining party and if you look closely at the preferences in Vic you'll see they also did a deal with the Shooters....well how Green is that? Family First of course represent gay and lesbian families.The Sex Party's policies are civil libertarian in nature and some of them like legalising recreational marijuana are policies that the Greens have not had since News Ltd attacked them years ago over drugs.