Wednesday, March 23, 2016

Poll Roundup And Seat Betting Watch: And They're Off! (Probably)

2PP Aggregate: 51.3 to Coalition (-0.2 since last week)
Coalition would win election "held now" with much reduced majority

In this issue:
Voting intention and aggregation
Leaderships: Turnbull Goes Negative
Metapoll Or Metaparasite?
New England
Polls Fail In Brisbane
Other Polling (Includes worst poll of the week)
Betting Watch

After one of those days constitution junkies love, Prime Minister Malcolm Turnbull may have just started a 103-day countdown to a July 2 double-dissolution.  We won't know for sure until we see whether the recalled Senate goes to water over the Australian Building and Construction Commission (ABCC) legislation, but the early noises from the crossbench are defiant.  Should the crossbench indeed refuse to pass the bills, we're looking at at least eight weeks in which the election date will be absolutely clear, but perhaps a few more.  We can also expect a 53-day formal campaign. Long campaigns have been in disrepute since Bob Hawke bored the nation into liking Andrew Peacock in 1984, but it was not always thus.  Menzies called a long campaign in 1958 and the experience didn't scare him off doing it again three years later.  (See Malcolm Farnsworth's tabulation of formal campaign lengths).

The government enters this new phase of the leadup in a reasonable but not stellar polling position.  A rapid plunge in its standing in February shows signs of levelling out short of actually gifting Labor the lead.  However, there's a fair bit of variation in the individual poll results.  This fortnight we've seen 2PP scores for the Coalition of 53 from Ipsos, 52 from ReachTEL, 51 from Newspoll, two 50s from Essential and 49.5 from Morgan.  The 50.5-49.5 lead to Labor in Morgan was the first time Labor has led the 2PP in any of the 55 polls taken since the removal of the previous PM Tony Abbott.  Morgan has tended to lean to the Coalition since Turnbull took over, to a recently reducing degree, but the evidence of this result has caused me to now treat Morgan as neutral on average. Perhaps it will soon return to its old ways and start again skewing to Labor.

With the primaries and apparent house effects included I counted the Ipsos at 52.8, ReachTEL 52.4, Newspoll at 50.9, Essential at 50.8 then 50.7 (one point temporary house effect adjustment) and Morgan at 49.3.   Most of the primaries were friendlier to Labor than the published 2PPs suggest. For now we're at 51.3 to Coalition and the rot seems to have stopped for a few weeks.  Here's the smoothed tracking graph:

(Regular viewers of my terrible graphs may notice that I now smooth out the summer holiday periods, although the aggregate is often switched off during them.)

Some other aggregates: Phantom Trend 51.9, Mark the Ballot 51.9, Bludger Track 51.3.  Another aggregate to be noted is Luke Mansillo's, which is regularly seen on Twitter and has now made a blog appearance (at 52.29).

Historically, a government polling at this level three and a half months from an election will usually win.  Historic polling at this stage explains a rather high 45% of variation in final election 2PPs and implies a 77% chance of the Coalition getting at least 49% of the 2PP vote (which should be enough to win).  My more advanced - but not necessarily better - model which includes the historic Labor Fail Factor (the ALP's tendency to underperform compared to polling) gives the Coalition an 84% chance of getting at least 49%.  It is possible the target for an even chance of winning is actually below 49% (I just have to find time to analyse the final demographics for the NSW redistribution) but the biggest potential confounder of any projection is Tony Abbott.  Whether the ex-PM is just obsessed with self-validation or actually determined to ruin his rival at any cost, if we get a repeat of the "backgrounding" leaks against Julia Gillard in 2010 then bets based on general historical trends are off.

Leaderships: Turnbull Goes Negative

Well, if you believe Newspoll he does, anyway.  Newspoll had Turnbull at a net -5 (38:43) netsat this week, his first negative netsat since becoming Prime Minister.  He has now lost 43 netsat points in four months, which is the equal second-worst for a PM in that timeframe (Keating lost 49 in 1993 and Howard 43 in 1996 - but the latter was artificial as discussed last time).  The patterns of Turnbull's Newspoll netsat over time since becoming leader are quite similar to when he was Opposition Leader - in fact, this time around he's taken two weeks less to go below zero:

Turnbull still holds a massive lead over Bill Shorten as Newspoll "better Prime Minister" (52:21), but the don't-know score (27) is remarkably high.  It was last that high for Howard vs Crean in 2003 (a rather similar political dynamic - if we can call it something that sounds that exciting - to now) and last higher in 2000.  Shorten meanwhile is continuing to improve and is now at a personal net -24 (28:52), a 14-point recovery since December and his least worst score since Turnbull came to office.

That's just Newspoll, however.  Over at Ipsos, Turnbull continues to enjoy a cruisy +23 (55-32), even if it is down 15 points on last month.  Ipsos has Shorten on -19 (33-52), up six, and has a 61:22 lead to Turnbull, a narrowing of the same amount.  ReachTEL's forced choice (which doesn't favour incumbents so much) has Turnbull leading Shorten 60:40, a rather dramatic narrowing of 15 points in six weeks.  ReachTEL leadership figures are coming.

Newspoll found Turnbull preferred to manage the economy 54:20 (cf 58:22 last month - see the history of this indicator here) and tax reform 45:25.

Turnbull's attribute scores have mostly gone down according to Essential, but he continues to beat Bill Shorten on everything except "arrogant" and "out of touch", including a massive 30 point lead on "a capable leader" and a 23 point lead on "intelligent".  An Essential smorgasbord poll of possible leaders of the Liberal Party shows Turnbull down three points, but the only beneficiary is "someone else" (up six).

Metapoll Or Metaparasite?

There was a newish poll-shaped object reported in the Guardian; it's called Metapoll.  Anyone reading the Guardian's report can see that what is going on here is quite odd.  Metapoll aggregates the findings of other polls, then adds its own online polling and thus far publishes only the combined outcome (51.8 to Coalition by last-election preferences or 52.4 respondent-allocated).  What is the weighting of its own data to the other results?  How does it age the other results?  How does it weight the other polls?  Which other polls if any does it apply house effects to and how?  All this is at this stage a mystery.

For a polling aggregation service like mine, there is a real problem with including a product that is partly a new poll and partly somebody else's aggregate, as this would make my aggregate partly a meta-aggregate and not entirely a test of my own decisions.  Also to the extent that the Metapoll figures reflected their own data, if their data were good I would be underweighting them and if their data were rubbish I'd be best off not including them at all.  If Metapoll starts publishing its own component separately then I can examine its behaviour and probably include it, but for the time being all I can really do with this "poll" is report it as one of a number of aggregates, noting that it is including data of unknown quantity that nobody else has access to.  However either the existing aggregates in the psephosphere (led by BludgerTrack, which nailed the 2PP and the Coalition seat tally exactly in 2013) are unknown to Metapoll or else it believes in slurring them while boasting about itself:

Most of us doing aggregation in Australia are doing it on a shoestring, funded by donations if at all, as a hobby or public research enterprise.  Generally we share information and views with each other, and my own aggregate's methods and house effect decisions are all public.

Along comes a company that's trying to make money by taking the data from other pollsters then repackaging them, while at the same time so far refusing to make its own fresh polling data available to those who might want to study it or do the same thing with it.  Metapoll (co-founded by ex-Greens staffer and recent Guardian writer Osman Faruqi) has been talking about possibly releasing its own data component and not just the aggregate.  This is a decision that must be taken and implemented right away. As things stand, Metapoll's business model is in my view not only opaque but parasitic.

This aside, it will be interesting to see how Metapoll's attempt to incorporate "betting odds and market behaviour" in an aggregate goes and how this 538-style big-data approach goes at the election.

New England

The only new seat polling I've seen lately has been the seat of New England where independent ex-incumbent Tony Windsor is trying to retake his own seat from the Nationals leader Barnaby Joyce.  A ReachTEL commissioned by an unknown source and one commissioned by The Australia Institute (PDF) showed Joyce leading on primaries 41.6-33.9 and 44.8-38.4 (leaning undecideds redistributed by me) respectively.  A Newspoll then showed Joyce leading 46-44.  These three results would require Windsor to poll 66%, 69% and 60% of preferences respectively, which is very achievable given that they come mostly from Labor and the Greens. Cathy McGowan got a flow of 79% against Sophie Mirabella in similar circumstances (though Mirabella was more hated than Joyce and McGowan's image less controversial than Windsor's).

I am a little bit surprised by these polls.  Joyce still has remnants of his former scruffy buffoon image, but Windsor escaped the verdict of voters on his support of Julia Gillard in the 2010-3 hung parliament by retiring.  Perhaps that role is more saleable now that voters have seen how his alternative dancing partner, Tony Abbott, would have governed, but it still seems difficult to believe that recapturing one's own seat against a party leader is this easy, the unpopular Shenhua coal mine or not.  (Newspoll adds four points to Windsor's 52:48 if the mine goes ahead.)  It also seems that the Windsor campaign is fast becoming a bandwagon for left-wing activists hellbent on unseating Joyce, which won't do Windsor any good at all, especially not if it tempts him to steer his own campaign too far to the left.  But the confirmation of the commissioned ReachTEL results (and then some) by Newspoll do at least flag this as a seat that should be watched.

Polls Fail In Brisbane

A moderate failure of polling has been recorded in the Brisbane lord mayoral race with Galaxy (53:47) and ReachTEL (52:48) close to each other but nowhere near Graham Quirk's (LNP) 59.3% (and rising).  Queensland local politics shouldn't be too hard to poll since voting is compulsory, so results like this are worth keeping an eye on.  Superior Conservative performance in local government elections given what the polls were saying was a large part of how Matt Singh predicted the UK Polling Fail.  Thus far, however, there has been little effort in Australia to determine what (if anything) local government elections can tell us about the health of our national party brands.  Not too much along those lines should be read into the LNP's continuing dominance of Brisbane council seats at this stage, since their seats were on double-digit margins from 2012.  (Note: ReachTEL advises there was a very high undecided level - for them - of 15% for this poll, suggesting low voter engagement despite the "compulsory" nature of voting.)

Other Polling

Some quick notes on other polling:

* Ipsos found the Coalition preferred to Labor as economic managers 43:25 (compared to 41:32 last April when Tony Abbott was PM).

* Ipsos found mild opposition to limiting the tax concessions available for superannuation (35:40) and negative gearing (34:42), the latter result somewhat against the run of previous polling.

* Newspoll finds voters believe the Coalition will win the election (55:25).  Coalition supporters overwhelmingly think they will win while Labor supporters are divided.

* Essential finds 35:17 support for re-establishing the ABCC, even with a question that might be considered slightly loaded against such an outcome.

* Essential finds 34:22 support for the government's current intention to force a double dissolution over the ABCC issue.  (ReachTEL has a similar result).  The ABC (with one C) has made much of the high "meh" factor response to both these questions, but this is something often seen on Essential (and not just Essential) issue polls in general.

* Essential finds that voters want plebiscites on the issues of euthanasia and abortion as well as same-sex marriage - the reason probably being that voters think politicians are too conservative on these issues, though a party breakdown would shed more light on that.  (Essential also showed its highest support level for same-sex marriage so far.)

* Views of what Tony Abbott should do haven't changed since December.

* ReachTEL finds the economy way out in front (by 22 points!) as the issue most important to voters from a chosen list.

* Metapoll (this being its own polling, not an aggregate!) reported strong support for the endangered Safe Schools program (which depending on which side of that debate you're on is either an LGBTI anti-bullying tool or an LGBTI propaganda machine). 55% of respondents supported implementing it from primary school.

The worst poll I saw this week was this, though it's no fault (that I can see) of the pollster. Former Greens candidate and GetUp! activist Simon Sheikh has founded a superannuation firm which has commissioned a robopoll on whether people agree that climate change and renewable energy will influence their vote.  They also asked whether people agreed that Great Barrier Reef policy would influence their vote.  Polling of this kind is worse than useless; it has been shown again and again that voters will say that just about any issue offered in isolation will influence their vote.  Throw in the non-publication of the text of the polling questions and the very likely bias of the commissioning source and all I can do is give this poll on our oceans the Old Boot and award it something that lives in them.

Wirrah Award For Fishy Polling (image source)

Betting Watch

I'm not sure I'll be doing this for every week if the campaign is really going to run for fifteen weeks, but it seems like a good time to kick off some seat betting tracking. The purpose of all this is not to advocate betting, nor to promote the idea that betting markets are good predictors, but to monitor and test their performance - see my mixed review for 2013. At this stage seat markets have little money in them and mostly reflect the modelling of the bookmakers, but that will change.  Mark the Ballot has been tracking the headlines so I'm not going to do that here for now, except to note they're about the same as my historic projections at the moment.

William Hill has an exact seat market in bands of five, that I take as implying about 80 Coalition seats assuming five crossbenchers.  Sportsbet has a market on the winning party's majority (over 75 seats, apparently) which seems to have 92 Coalition seats as the most likely tally, but with a distribution that tails off much faster for a higher tally than a lower one.  Sportsbet's 2PP market appears to be pointing towards a median expected 52.4% 2PP for the Coalition, which would not lose it many seats at all.

I track individual seat betting by classing a seat as "close" if there is more than one party at $3 or less. On this basis the following are the current Sportsbet projections from a starting base of 91 Coalition and 54 Labor (counting the three notionally-Labor Coalition seats as Coalition-held):

Loss (Coalition to Labor): Barton*
Close Loss (Coalition to Labor): Dobell*, Petrie, Capricornia, Lyons, Solomon, Hindmarsh
Tie (Coalition-held): Paterson*
Close Loss (Labor to Coalition): McEwen
Loss (PUP to Coalition): Fairfax

Coalition Close Holds: Braddon, Burt, Eden-Monaro, Cowan, New England (vs Ind)
Labor Close Holds: Parramatta, Moreton

* = notionally Labor

Summing these seats gives a net projection of 85 Coalition, 60 Labor, 4 crossbench and one tie.  The Coalition has more projected close holds, but also more projected close losses. Seats on smokey-watch include Warringah ($8 NXT), Sturt ($4 NXT), Grayndler ($3.24 for Greens to stop Albo taking nonsense about Senate reform) and Kennedy ($3.50 LNP vs Katter).

Of the above, Bludger Track projections have consistently suggested Tasmania is holding up against the national swing, and that makes me very sceptical of Labor's favouritism in Lyons.  However, we'll need to see more Tasmanian polling as the campaign continues.


  1. Comment by Liam Whelan that I stuffed up the approval of (too clumsy to do this by mobile phone):

    What about Mayo? I'd definitely put it on smokey watch.

    1. Mayo is one of a large number of SA seats where either NXT or "Any Other" is in the $5-$7 range at the moment.

  2. I would be extremely skeptical about Hutchinson losing Lyons he has been a very good local member and entrenched himself well. Plus Dick Adams is not on the ballot this time so his personal vote which was huge will be up for grabs.

    I hail from Bridgenorth on the Bass Lyons border and feel Bass is far more vulnerable to Labour than Lyons is. I believe you had a post here a while ago with some seat polls showing a swing to LIB in Lyons and Bass being 50/50. Up here this is very much still the case.

    1. Deepest apologies for only just publishing the above comment - there was a technical issue which meant I didn't get the notification for it and have only just found it!

  3. Kevin - have updated the aggregation: 51.9%

  4. I don't know what to make of MetaPoll. I don't think they're morally obliged to release their own polling (though it'd be particularly ungenerous of them not to) -- they're paying for it out of their own pocket, and if they want to make an end-product set of numbers that no-one else can use, then so be it.

    But it's going to be hard to build up trust in their processes the way that Nate Silver has over the years (since they seem to be modelling themselves in part on 538, I think the comparison is useful). Silver rose to prominence in the political sphere by pseudonymously blogging the 2008 primaries. With dozens of state-based elections to model and predict over a primary season, it's easy to gain a reputation as a high-quality polling/demographics analyst. Silver's since continued to produce good forecasts over several elections, and so even though his team don't give full details of how they generate their forecasts, I'm nevertheless happy to consume them regularly.

    MetaPoll will have one shot at gaining a reputation as being better than the other Australian aggregators, probably on 2 July. But how would we know if they're actually better? You can aggregate to N=30000 and there'll still be a half-percentage-point margin of error. All aggregates should miss by a bit sometimes.

    (Unless they do their own gigantic polls, and get good seat-level forecasts as well as the headline primary/TPP numbers.)

    Still, I look at their website and it's stuff I'd eagerly read -- I don't know if there's much value-add beyond your roundups or the Poll Bludger's summaries, but since they can do their own polls they should have more material. I'll be happy with it if they can turn a profit being a local 538-ish clone.

  5. they stole your graph!

  6. Removed quickly after I told them they were using it without permission. Thanks very much for letting me know.

  7. Kevin, here is a comment I made at pollbludger about metapoll... no one responded, and I would like to know what you think...

    Regarding the “poll of polls” at the guardian

    It seems like their data is wrong. As examples, I think the Feb and March Reachtel points are plotted in the wrong places, and should be 54 and 52 for the LNP, not 46 ad 48.

    I just cant see how their graph and the one marktheballot does are so different in recent weeks…

  8. I agree, those two data points are wrong. Perhaps that is the cause of their graph dipping to 49 in Feb-Mar when nobody else did; I saw that and was puzzled by it. Sometimes data points look different to the reported result on these graphs because the modeller has applied a house effect but that is clearly not the case this time.