Tuesday, August 13, 2013

Poll Roundup and Seat Betting Watch: Scruff Of The Neck Edition (August 13)

2PP Aggregate (Tuesday 13 August): 51.5 TO COALITION (+0.7 since last week)
Individual Seat Betting: Labor favourites in 63 seats (-2, Forde to Labor; Banks, Reid, Brisbane to Coalition)
Seat Total Market: Labor 65 seats (-2) (This figure may be slightly skewed by longshot bias; see below.)

This is week seven in a regular weekly series in the leadup to the federal election.  Week six was here and through it you can click back to the previous weeks.  Or just click the "betting" label at the bottom.  As stated before, the aim of this exercise is not to claim that seat betting markets have predictive value, but to test whether they do, and to see which of the markets and the aggregated polls see the ultimate outcome of the election first.

Things aren't looking too good for the Government after a few weeks of apparent competitiveness.  The polls say they're losing at the moment, the punters strongly think they'll lose, the voters think they'll lose and the modellers think they'll lose.   Still left expecting a Labor victory are really just the true believers, the Abbott-haters (and not even all of those), the Sawford formula devotees and Peter Brent.  There's a general perception that the Coalition now has this election by the scruff of the neck, that the first month of the Rudd comeback polling was partly bounce and that the government won't be going up from here and may indeed be going down further. Is it really quite that hopeless for Labor?

The first leaders' debate has been the subject of widely varying views on the verdict, perhaps reflecting the sharp difference in debating styles of the two leaders.  There's nothing near consensus on who "won" it in commentary or in attempts to measure the result.  There is, perhaps, a slight lean towards Rudd doing so on average, but Abbott avoided major damage, which was really all he had to do.

I should mention in passing, however, that the "worm" methods employed by Seven and Ten on the night, and the ABC's Twitter tug of war, all involved some form of opt-in process, which makes their results stack-prone and meaningless.  Only the Nine worm even attempted to be a valid poll, using 100 supposedly undecided voters.  But even its readings were a bit  hard to credit, with Rudd in positive territory for the entire debate (even including some very risky comments about Sydney and its airports).  Very experienced pollster Mark Textor opined on Twitter this week that vermology is a dated audience assessment method that is flawed and only good for "infotainment".  (I can't find the exact tweet, so I'm mostly paraphrasing.)

This week's polls

This week continued the bad run for Labor from last week.  Nielsen went first late last week with a 52-48.  This was followed by a 51-49 from Galaxy, a 53-47 from ReachTEL, a third consecutive 52-48 from Newspoll, and a 51.5-48.5 from Morgan by last-election preferences (50-50 respondent allocated).  Essential was late for the second time in recent weeks but finally released a 51-49.   I've been on-again-off-again about applying a pro-Coalition house effect for Essential, but it remaining at this level tends to confirm that it was continuing a recent history of sluggish and partial response to major shifts.  As the Rudd bounce has faded, the others have come back to it, so I've removed the +1 point.  Nice chart of the trend over here.

(By the way if you want to know more about all these pollsters generally, I have recently released A Field Guide To Australian Opinion Pollsters.)

The ReachTEL was the first poll taken wholly since the return of Kevin Rudd to give the Coalition 53, though this is not as significant as other polls doing so given the slight lean of ReachTEL national polling so far.  The Newspoll must also have been very close to going to 53-47.  Even with the most extreme possible assumptions about the rounding of the primary figures, the Coalition would have been on at least 52.3, more likely very near to 52.5.    All these polls have pushed my aggregate into the mid 51s, and while some of the others doing the rounds are about the same, there are also some (including Pottinger's and Simon Jackman's) that are in the mid 52s. Jackman's does so because of house effects that have nearly every pollster as Labor-leaning, which surprised me slightly, but may be a result of calibration against "the last federal election" (rather than against a combination of multiple elections and other pollsters as per Bludgertrack, and in simpler form, here.)  At the last election there was a swing from late polling to the final result, but it's debatable whether the methods of the pollsters or those of Julia Gillard should be blamed for this. 

The upshot is that at the moment, whatever aggregate you use, is that the Coalition is clearly winning and would win an election held "right now" fairly comfortably. Bludgertrack (off a base of 51.9%) projects that current polling would produce 82 seats to the Coalition and 65 to ALP. 

The Nielsen poll was notable for returning Tony Abbott's best personal ratings since July 2011, with a net approval score of -7.  (Nielsen personal ratings are milder than Newspoll's) Newspoll also showed an improvement in Abbott's standing with an eight point netsat lift to -14, his best since March, after four polls becalmed in the -20s. Essential did not show a significant change in Abbott's ratings.  Even more notably, a Nielsen question on trust showed Abbott actually leading Rudd 47-40, a remarkable result given Abbott's long-entrenched reputation for shiftiness.  On the other hand, 21% of Coalition supporters think not being Prime Minister material is Abbott's greatest weakness of a list of options offered by Galaxy.

Also according to Galaxy, 30% of voters thought Rudd had improved his style since he was PM last time, while 51% didn't, but this is a typical Galaxy murky question.  It's not clear from the response how many of the 51% think he's got worse (it might be very few) or how many think he was fine to begin with.

Attribute polling on issues outside Labor's historic strengths has been poor for the Government this week.   On economic management, Labor trails 36-48 (Galaxy), 36-58 (Nielsen) and 33-49 (Newspoll).  On asylum seekers, it's 39-47 by Nielsen.  Their Newspoll rating is up one to 27, but the Coalition jumps to a large lead on 42 (with more voters now supporting the Coalition's solution although it hasn't changed and nor really has the debate advanced since the question was last asked.)  On interest rates it's 33-44 (Newspoll again.)  Clear deficits, across the board, on a range of issues on which just weeks ago Labor was much more competitive.

There is one other poll I thought worth mentioning.   The ANU Poll is not a traditional poll but a measurement of political attitudes on things such as compulsory voting, the health of democracy and so on.   A notable finding in this edition is that over 30% of voters think there should be more independents in parliament.  However, it seems that most of that 30% must reckon somebody else should vote them in.

Historical Comparisons

Even a 2PP of around 48.5 (or even 48 or 47.5 if you think it is that bad) is hardly disastrous during an election campaign.  At least four governments have won despite trailing this heavily or slightly worse in campaign polling: Fraser (1980), Keating (1993) Howard (1998), Howard (2004).  In 1980 and 1998 the governments in question had won big at the previous election, which makes it much easier to cushion the impact of 2PP swings through flexibility in deciding which areas to try to save, and in the case of 1998 through sophomore effects for new members elected at the previous poll. So really only 1993 and 2004 are relevant precedents for a win from the current polling position; 1998 at least should really be considered as a precedent for a loss.  (NB added 27 August: 2004 is not a valid comparison wither because of Newspoll's use of respondent-allocated preferences.  Howard did not, in fact, trail heavily in the campaign.)

A simple model based purely on the relationship between polling at this stage of the campaign and the final result will still show the Government with a realistic chance (say 25%) of polling above 50% two-party preferred on the day, because of the large variation in past swings.  Based on historic records, what we are seeing now could conceivably be a few points different from the final outcome, in either direction.  However, the big "but" for such a model remains the possibility that the Rudd bounce is still deflating and that Labor's 2PP, irrespective of campaign events, has further to fall.  If that is the case, Labor's chances become very low indeed.

I interpret much of the shift in polling this week and last as being mainly caused by the Rudd bounce subsiding naturally, rather than mainly by the specific issues that many people like to think drive polling.  There is a widespread assumption that the Coalition now has "momentum" and that further falls in coming weeks for Labor are more or less inevitable.  Judgement should be reserved on that one until we see next week's data.  Momentum in polling is a very elusive quantity. 

Markets Ditch Labor

Betting markets this week have taken a very dim view of Labor's prospects.  This is not surprising given that polling this week was worse than last week and the election is closer, so the proverbial hill for Labor to climb is twice as steep as it was last week.  The markets have also either not agreed that Rudd "won" the first debate, or else (and more likely) assumed that since there was nothing in it to seriously damage Abbott, even a moderate loss in it improved his overall chances.  Headline rates for Labor have moved into the $5-$6 range. 

The Centrebet correct-result market now gives an implied average of 65 Labor seats (down 2) assuming Labor wins Melbourne.  However, it implies a 27.4% chance of an outright Labor win. In a market summing to 146%, I suggest that longshot bias is a factor here (most of the Labor-win seat ranges are well over $10) and that the headline much more accurately captures what the market thinks of Labor's chances (ie <20%).    (EDIT: I've removed an incorrect comment about the seat handicap market that was here.)

On to this week's individual seats graph.  This week I am using Sportingbet, Sportsbet and Luxbet.  There is almost complete consensus between them as to seat favourites, except that Luxbet has Melbourne as a tossup:

The colour-coding again:

Medium blue: A seat in which the Coalition is favoured in all betting markets.
Pale blue: Coalition favoured in some markets, level in others.
Grey: (none this week) All markets tied or both parties ahead in some markets and behind in others.
Orange: (none this week) Labor favoured in some markets, level in others.
White: Labor favoured in all markets.

Bold shows a seat that has changed in colour in the last week. There are five of these.  Banks, Reid and Brisbane have all returned to the Coalition pile, Longman is no longer tied on any market, and only Forde has crossed to Labor.

I also thought it would be useful to run a graphic showing the trend of seat favourites over time in those seats for which there have been changes since I started this series:

Here I can use a clearer set of colours, so:

Red: ALP favourite in all markets
Orange: ALP favourite in some markets, tied in some
Grey (yes I know those two aren't quite the same shade!): All markets tied or different parties favourite in different markets
Light blue: Coalition favourite in some markets, tied in some
Dark blue: Coalition favourite in all markets

Every week the Coalition has been unanimously favoured to hold 70 seats it won at the last election and gain Lyne and New England from independents and Bass, Braddon, Lindsay, Greenway, Corangamite, Deakin, Dobell and LaTrobe from Labor.  Every week Labor has been unanimously favoured to hold 54 of the seats it won last time and to win Melbourne from the Greens.  Every week Katter and Wilkie have been unanimously favoured to retain.

The battleground of seat betting favourite status has consisted of 13 seats and only eight of those have seen an actual change in outright favourite.  Of those, three shifted to Labor after the first week (when seat betting markets were not very informed) and of the five to have moved to Labor since, three have now moved back.   It is, all up, a very static picture.

The close seat tally this week comes to eighteen seats (+3) in which Labor is favourite compared to fourteen (+4) for the Coalition.  However, the Coalition's basket grew by three seats this week while Labor's shrank by one, so Labor's perceived risk factor in close seats is slightly higher.  The close seat list for Labor includes a lot of the usual suspects but there was a surprise appearance from McEwen (Vic, 9.1%), while Blair (Qld, 4.2%) also made its first appearance for a while.

It's also worth keeping an eye out for smokies who the market thinks have some vague sort of show of upsetting the major parties.  PUP leader Clive Palmer is currently around $7 in the seat of Fairfax, and the one that has quite a buzz about it is rural indie Cathy McGowan's attempt to unseat Sophie Mirabella from the seat of Indi.  It will be interesting to see some polling on this one, but it is a difficult race to poll for since there may be a snowball of primary vote support for the independent if voters for other parties decide she has a real chance of winning the race.  So a useful poll would ask respondents for a 2CP of Mirabella vs McGowan. McGowan is currently at around $6.50-$7. 

Another market that amused me was Sportingbet's "Tasmania Special: Labor Wipeout" (No $1.24 Yes $3.80).

In all, this is the second week in a row in which polling has moved towards the seat market picture.  If the current polling accurately foreshadows the end result, then the seat betting markets from the second week on will scrub up very well as predictors.


Update (14 August) - 57-43 From Outer Space: A small Morgan phone sample has been released (sample size 549) with primaries of 52 to Coalition, 31 to Labor, 9 to Green and 8 to Ind/Other.  The supplied 2PP is 57-43 but it is not stated whether it is respondent-allocated or last-election.  However I haven't been able to obtain a last-election result of 57-43 with any assumption on preferences, so I assume it's the former.    I have treated it as 57 for aggregate purposes. 
Historically, Morgan phone samples tend to have little if any skew and to be quite accurate, but this one has a small sample size and a maximum margin of error of around 4%.  Even with that MOE considered, it's probably rogue (and I don't use the term lightly.)  The extreme 2PP compared to other polls, unless confirmed by the next round of polling, invalidates the leadership findings which have Abbott with a better netsat than Rudd (-6 vs -9, the first poll to produce such a result) and Rudd with a very slim lead as preferred Prime Minister (46-43).  If a poll has an unrepresentative 2PP it will usually produce skewed scores on everything else.
Gary Morgan argues that there is a big change in the wind and attributes the massive shift to "the debate and the analysis that followed, including discussion around the use of notes by Prime Minister Kevin Rudd".  It also states that Abbott received a "free kick" from News Corp, ignoring that Abbott has been receiving such on a regular basis through the campaign.  Morgan closes with a prediction that "all telephone polls conducted over the next few days" will show "a jump in support for the Liberal-National Party!" 
Perhaps they will, but if they do it will not be anything to do with silly arguments about the Prime Minister using notes in the debate and whether or not it was against the rules, arguments which affect nobody's standard of living and which not that many people follow or care about, and nor will it be (across the board) to 57%.  This continues a long-running tradition of Morgan poll commentary in which large moves in Morgan polls are ascribed to events unlikely to have produced massive jumps. 
The implication seems to be that because the poll now shows that Abbott "won" the debate (24-16 compared with 24-23 after it - hardly a remarkable move anyway, especially given so much of the sample didn't watch it) that this explains the 2PP shift.  Far, far, far more likely it is the other way round: an unrepresentative sample that has a very high Coalition 2PP will also think Abbott won the debate.
But, as I have pointed out before, even data from a rogue poll (if conducted by appropriate means) is still data as useful as any other, so into the aggregate it goes, downweighted for small sample size.  The net result is a 0.3 point shift to the Coalition to 51.8.

Newspoll of Robertson and Dobell (16 Aug):  Amazingly, there's actually seat polling that isn't by robopoll!  Newspoll have polled the adjacent seats of Robertson and Dobell (apparently combined) with a 2PP of 54-46 to Coalition.  Rudd has a netsat of -15 and Abbott +12; Abbott is preferred PM by 47-41.  At this stage I have not seen any breakdown of the two seats in the sample, or the sample size for each.

These are both seats the Coalition has generally been expected to win, but I did think that Dobell was remarkably obstinate in sitting at longish odds while NSW seats on lower margins became close or crossed, and suspected the Thomson saga was being overvalued in it all.

Update: the sample size for Robertson/Dobell  is 505 respondents, which does leave a great amount of room for doubt about whether Labor is actually cactus in both of these seats, even assuming its position doesn't improve.   After all Labor's result in the best of the two (whichever that is) would be at worst 46 off a sample of c. 252, which translates to a minimum 10% chance of being ahead, and hence a minimum 19% chance that Labor is currently ahead in at least one of the two seats.  If the results are assumed to be less even than 46 apiece, that chance goes up.  Still, it is much better not to be polling such results than to be polling them. 

Meanwhile odds for Labor overall continue to blow out as can be seen here, despite the relative narrowness of national polling and the historic record that does not flag this as a lost position.  Either the markets are justifiably confident that Labor should not now recover or else what we are seeing is a fully-fledged run in which the prices are affected by belief that the prices themselves are predictive.

Lindsay Lonergan (16 Aug): A Lonergan robopoll for the Guardian with a massive sample size reckons Labor is utterly busted in the seat it is so historically obsessed with (60-40, though this was apparently respondent-allocated and slightly closer by last-election preferences).  Lonergan is a robo-pollster similar to ReachTEL or JWS.  An interesting aspect of this poll is that Lonergan asked respondents how they had voted last election in an attempt to back-validate their sample as accurate.  Poll Bludger notes that this doesn't necessarily work as respondents will often over-report voting for the winner;' on that basis it is possible (my extrapolation not PB's) that the polled lead is exaggerated. Various reasons are given in the article for the possible bad result; one that is not flagged is Rudd's extremely risky comments about Sydney in the leaders' debate.

Seatpoll Flurry (17 Aug): More seat polls have emerged.  ReachTEL has:

Bennelong Liberals ahead 65-35
McMahon Liberal 53-47
Kingsford-Smith Liberal 52-48
Blaxland Labor 52-48

Given ReachTEL's apparent Coalition lean and given margin of error issues, the McMahon and Kingsford-Smith deficits are no disasters, but again it is better not to poll such results than to poll them.

JWS Research has:

Forde LNP 60-40
Brisbane LNP 54.1-45.9
Corangamite Liberal 53.3-46.7
Macquarie Liberal 55.1-44.9
Aston Liberal 63.4-36.6
Greenway Labor 51-49
Lindsay Liberal 60.7-39.3
Banks Liberal 52.8-47.2

Lonergan has a similar result for Labor in Forde (primaries of 56-34).  The impression is that the Beattie parachute-in idea has simply not been well received by the locals, who will respond by keeping a sitting MHR whose dumping was almost a done deal according to even Coalition internal polling before that.  Forde is no longer in the Labor seat-betting column.

JWS tables are here.  In all the JWS surveyed seats bar Greenway, the local candidate has a better netsat than Tony Abbott (the Greenway candidate being one of the great campaign duds, Jaymes Diaz).  Aside from Beattie in Forde and Darren Cheeseman in Corangamite, the Labor candidates generally have much better netsats than Kevin Rudd.  The ALP candidates average -4.6 to Rudd's -17.4 and Rudd is especially on the nose in Lindsay.  The Coalition candidates average +16.3 to Abbott's -0.9. These candidate/leader netsat comparisons are nice work by JWS. 

Newspoll has the Nats winning Lyne and New England 59-41 and 66-34 to no surprise.  In New England Rudd's netsat is -25 and Abbott's is -1.  In Lyne it's -28 and -2

It is common to find that pre-election seat polls are not very predictive, and in the past small sample size has been an issue.  But there is more going on here and it is the same mysterious thing that has been happening all year: that public seat polls are almost unanimously dire for Labor, even with large sample sizes, and keep telling a story that is much worse than national 2PP polling.  That the vast majority of seat polls are conducted by robopolling only adds to this mystery. 

There is an increasing feeling in the last few days that the wheels are falling off Labor's campaign - ad hoc announcements, confusion, unclear strategy, poor attack ads, flat recycling of themes, failure to cut through on key issues and so on.  Plan A was a presidential campaign with Rudd as its centre and neutralisation of weak points.  But Rudd is actually barely more popular than Abbott, and the voters have returned to thinking the "neutralised" points are still weak points, so what is plan B?  There doesn't really seem to be one yet, except for saving the furniture and hoping that the Coalition will collapse.  

Maybe national polls this weekend will now show much worse results than last week, and the latest batch of dire seat-polls will all make complete sense, in which case the election will be over as a contest.  If this doesn't happen, it's time to start seriously wondering where the hell the ALP is picking up all the votes elsewhere that it is shedding in droves in almost every marginal or semi-marginal seat poll.

Sportingbet have now unveiled an Exact Result market with 85-86 Coalition seats the shortest-priced offerings apart from Hung Parliament (which is not broken down into seat bands).

61: In August last year I ran a simple historical-polling model which had precisely two terms, the worst rolling average position of the government in office, and the party in power.  It predicted 61 Labor seats (+/- 9).  It was my first attempt to project the result of the election, however facetiously.  I laughed at it a little at the time, without actually dismissing it, but at the moment it is laughing back loudly.  Until further notice, and with the usual low probability of being exactly or very close to right that goes with these things I'm predicting 61 Labor seats.  :)

(A final and more serious prediction, probably wrong, will be published here the day before the election.) 


  1. "Also the seat handicap market thinks Coalition -10.5 will defeat Labor +10.5, implying that that market expects a seat difference exceeding 21"

    Are you sure that this is how to interpret the handicap markets? (I think you've got it wrong, but now I'm second-guessing myself!) My understanding is that it's Coalition-10.5 v Labor, or Coalition v Labor+10.5. In other words, "Will the Coalition win at least 10.5 seats more than Labor, or won't they?" The 0.5 is added to the line to ensure that there can't be a tie.

    It looks like Sportsbet are moving their line to keep the odds even, which is a useful market to have: currently they're giving the Coalition a 14.5 seat lead.

    1. Yep, I had that wrong (not sure how since it's obvious how it works in other sports); I've removed that bit. Ta.

  2. The question of how you calibrate the pollster bias is pretty interesting, of course. Jackman's model calibrates off the last election, as does mine.

    As you have pointed out, you can use a series of past elections to do this. This is fine provided that you think the methodology hasn't changed over that period. How far back are you willing to go to make that assumption? Consider the massive uptake of smartphones (and mobiles more generally), such that a lot of people no longer have a landline.

    I'd probably be happy to use the 2010 and 2007 elections as part of the calibration, but I'd be pretty cautious about going further back than that.

    Another way of looking at this is to plot the post-hoc bias of eg Newpoll against time, and to see whether this quantity is random (as you'd expect). Small sample size will be a hindrance in telling you anything, particularly if they keep changing their methodology (as you'd expect them to do).

    In a world where pollsters constantly change their methodology, using the last election only may be the safest option, although not necessary the most unbiased one.

    1. There's not much of a pattern when looking at the predictive errors of what I take to be the final federal Newspolls (based on their website listings) over time: -2.2, -3.1, +1.9, -0.1, -2, +2, -2.7, +0.7, -0.1. (Error is in 2PP prediction of Coalition vote share).

      The Bludgertrack method is stated here: http://blogs.crikey.com.au/pollbludger/bludgertrack-2013-methodology/ BT uses multiple elections for Newspoll and Galaxy (including state elections, which might be contentious) and then does the "bias" adjustment for the rest on their behaviour relative to the rest of the aggregate. I'm a bit more comfortable with this than either using just the last election or assuming zero-sum across all pollsters, so at a particular point I swiped William's primary loadings for some of the pollsters and translated them to 2PP as I don't have historic data sets for all pollsters. At the moment, BT doesn't have its current loadings up.

  3. Betting market result is interesting. If you assume that Labor will win 63 seats, and solve the cube rule (s/[1-s]=[v/1-v]^3), it suggests a Labor TPP vote around 47.3 per cent.

    If Labor won 65 seats, the cube rule prediction would be a TPP of 47.8 per cent for Labor.

    Link: http://marktheballot.blogspot.com.au/2012/11/cube-law.html

  4. "Either the markets are justifiably confident that Labor should not now recover or else what we are seeing is a fully-fledged run in which the prices are affected by belief that the prices themselves are predictive."

    Heh, nails it.

    Well, maybe not in this particular case... I see very little evidence the ALP is in the game at present--my personal view is that they'll only form government again on the back of a black swan--but I am certain people place way too much stock in predictive accuracy of political betting markets.