Poll Roundup And Seat Betting Watch: The Polls Are Getting Closer (To Each Other)

2PP Aggregate 51.8 to Labor (2016 preferences) (-0.1 since last week)
51.4 to Labor (with One Nation preference adjustment)
Current seat projection assuming polls are accurate c. 80 Labor 65 Coalition 6 crossbench (+/- lots!)
Polls appear to be "herded" which can increase risk of error

Time for another instalment of the week in polling and seat betting, delayed slightly by an interstate trip.  As of last week the United Australia Party were going through a bit of a purple patch, polling 5% in Newspoll and above their 2013 result in a bunch of seat polls.  This hasn't lasted; all their poll results this week have been in the 3-4% range, and in three WA YouGov-Galaxy seat polls they did worse than their 2013 results (in one case much worse; the other two only very slightly.)

This week's action in the minor party primaries came from the Greens who polled 14% in Ipsos, 11 in Morgan, 12 in Essential and 9 in Newspoll.  Ipsos (especially) and Morgan have form for exaggerating the Green vote and Essential's reading of the Green vote lately has been quite volatile, but even so the party doesn't seem to be in too bad shape, with the issues mix at this election helping it (that is an understatement.) That said, the Greens have finally struck the candidate problems that have hitherto affected everybody else.



Ipsos Preference Mystery

The 2PP is the major focus of my efforts and this week we added 51-49s from Newspoll and Morgan and 52-48s from Ipsos and Essential.  Issues with Newspoll preferences were dealt with last week.  This week's Newspoll primaries were entirely consistent with 51-49 by their current method (as best it has been documented) and came out to about 51.8% by last-election preferences.  This week the preference mystery was actually with Ipsos.

Ipsos uses, or in the past has used, a form of preference allocation often referred to as batched preference flow.  Batched preference flow works by treating all parties other than the Coalition and Labor as a single preference source and assuming that the preference flow will be the same whatever the changes in the makeup of minor parties.  In 2016, after adjusting for three-cornered contests, the batched flow was 36.4% to Coalition.  [Note: It's been long believed in the psephological community, and was originally stated, that Ipsos inherited this preference flow method from Nielsen, but this is actually incorrect.]

Off the current Ipsos primaries (36 Coalition, 33 Labor, 14 Green, 5 One Nation, 3 UAP, 9 Others), the batched 2PP should be about 52.85 to Labor, so was this yet another case of the Coalition getting lucky with the rounding (ie it was actually just under 52.5 and got rounded down to 52?)  According to David Crowe's report "Ipsos treats votes for One Nation and the United Australia Party in the same way as votes for other minor parties at the 2016 election, which means 53 per cent of the preferences are allocated to the Coalition and 47 per cent to Labor."  Except (i) this isn't how Ipsos has done its preferences in the past (ii) you can only get a 53-47 breakdown for "other minor parties" from the last election if you remove Nick Xenophon Team (now Centre Alliance, and still running in a few seats) from the Others category.

What is going on here?  All major pollsters should prominently explain what they are doing with preferences on their websites.

The Newspoll set one notable record.  Labor's 54th 2PP win in a row was the most consecutive wins by the same party within a parliamentary term in Newspoll's history, beating the 53 set by Labor after the 2007 election.  Those 53 had followed a pre-election streak of 27 wins by Labor, making the longest streak covering more than one term Labor's run of 80.  If Labor wins the next two polls and the election, they will have another 25 to go to beat that.

By last-election preferences I aggregated Newspoll at 51.8, Ipsos 52.5 (ignoring any possible preference change since there was no information explaining it clearly), Essential 51.9 and Morgan 51.1.  The net impact of this week on the aggregate was an insignificant further narrowing, of one tenth of a point.  Here's the smoothed tracking graph.

My seat-based model currently suggests that for a 2PP around the mid-high 51s, the Coalition should lose about ten seats to Labor and Labor should lose about one or two back.  However at the moment its read of which seats are most likely to fall is close to being straight off the pendulum - it gives a >50% chance of defeat for all Coalition seats on less than 2.2%, and no others, though there are good reasons to doubt this in a fair number of seats.  It also doesn't tip any specific Labor seat to fall.  I've experimented with including a One Nation preference shift factor to make the Coalition do better in seats with high One Nation votes and worse elsewhere, and this affects chances in particular marginal seats such as Capricornia and Flynn a lot (bringing both to pretty much lineball), but it doesn't have much impact on the total.

The main ways Labor does much better than the model are:

* state-based findings of strong results across Queensland holding up (they haven't in the past)
* locally high swings in target inner Melbourne and/or WA seats but no swing in other safe seats generally (unlikely)
* the current 2PP in national polling being wrong.

Off the current 2PP, even if I adjust it aggressively for One Nation and UAP preferences and drag it down to the low 51s, it would be very difficult for the Coalition to remain in government.  Their best hope would be that there were large-ish swings in Melbourne that took only Dunkley and Chisholm but fell short in the other at-risk seats there, and that these were balanced by weak swings or no swing elsewhere, and by swings back and forth in different parts of Queensland.  This would all be incredibly lucky and really the Coalition's best hope is simply that the polls are wrong.

Herding, Herding, Herding ...

The last few weeks have seen the re-appearance of an issue from the 2016 election polling: the appearance of a string of very herded-looking 2PPs.  The following are the 2PPs for Labor in the last eleven polls (three each of Newspoll, Essential and Morgan, one Galaxy and one Ipsos):

52, 52, 51, 52, 51, 51, 51, 52, 51, 51, 52

Mark the Ballot reckons this is suss.  While I add that the Newspoll sample size has lately been more like 2000 than the 1000 used in Mark's calculations, and while Newspoll's often "under-dispersed" nature should also be considered, it's still a remarkable string of numbers.  Using just the last twelve polls (MtB also included the slightly prior Morgan 52.5) I ran a Fakepoll simulation with 250,000 tests and found that if I assume the polls are all random, normally distributed samples then all the values fell between 51 and 52 only 61 times, ie about once in 4,000.  This looks like pretty strong evidence that someone is playing follow-the-leader, though self-herding (anti-volatility measures, often suspected of being employed by Newspoll) may also be contributing.

Herding happens when some pollsters (usually the worse ones) make subjective decisions that, consciously or otherwise, guide polls towards all getting similar results - at present, preferencing is a candidate for this.  I was suspicious of the 2016 polls because of herding in the final polls but in that case the final polls were accurate.  Final polls were also accurate in the Victorian 2014 campaign in which a flock of 52-48 results proved to be bang on the money.  But herded-looking polls have come a cropper in some seat poll cases (notably the Longman by-election) and one day it might happen at national level too.  If it does, it's possibly notable that the last elections in every state except Queensland were less close than final polls said they would be.

Leaderships

This week saw Bill Shorten poll another rather bad Newspoll net rating of -18, which he also polled in February.  Scott Morrison returned to an 11-point lead as Better Prime Minister, which is effectively a 5-point deficit after accounting for the skewed history of preferred leader scores (deduct 16 points from the PM's lead.)  Morrison's ratings continued doing nothing, leaving him on a net rating of -1.

There was some contrast at Ipsos which had only a 5-point lead for Morrison, but that's to be expected given that the Ipsos poll numbers were friendlier for Labor on the whole. 

If this election sees a relatively close Labor win, it could be a good opportunity for the skewed Better Prime Minister metric to fail as a predictor to an extent that it has never failed before.  In three consecutive elections (1993, 1996, 1998) the candidate who was Better PM according to Newspoll lost, but in two of those cases the leader was Better PM by a single point (Keating was Better PM against Howard by five, which didn't prevent a landslide.)  No PM holding a double-digit lead this close to the election has previously lost - because Better PM is correlated with voting intention, and the last three changes of government were decisive.  If, on the other hand, the Coalition somehow wins, much will be said about the hidden predictive power of Better Prime Minister scores, but it will all be nonsense.  If the Coalition wins, it will most likely be because pollsters' reading of basic voting intention was all wrong.

Seat Polling

Seat polling has been relatively scarce at this election.  I'd like to report that this was because media had realised that seat polling was a relative waste of money compared to more national and state-based polling, but rather it seems to have been part of the general downturn in public polling over the last few years.  This downturn is sometimes attributed to the Trump-Clinton result, but I think another big factor is that media are spoonfed internal and campaign poll results by parties and activist groups, and get free stories out of those, so why bother paying?

A notable aspect of seat polling in 2019 is that we have not yet seen one single media-commissioned seat poll that has shown an incumbent trailing.  The seat polls that have shown incumbents trailing were all commissioned by lobby or campaigning groups or parties.  Three Galaxy seat polls in WA continued the trend with Christian Porter ahead 51-49 in Pearce, Anne Aly ahead 51-49 in Cowan and a little more surprisingly Steve Irons ahead 51-49 in Swan.

The Adelaide Advertiser also had a Galaxy with Rebekha Sharkie ahead 57-43 in Mayo (those crying out for seat polling in non-classic seats would be much more interested in any of the others), while a GetUp! commissioned Lonergan poll of Warringah had Tony Abbott trailing Zali Steggall 44-56.  Unfortunately GetUp! has yet again failed to publish the full primary results of its poll.  Various reports have had the primary for Tony Abbott at either 38% or 39%, but as usual lacked detail about the treatment of "undecideds"; suffice to say that whatever it is it wouldn't be enough with Steggall very close behind.  If, that is, the poll is anywhere near accurate.  Every commissioned poll in this seat recently of which any reports have emerged has had Abbott being flogged by about the same amount - and that includes reports of the Liberal Party's internals.

Seat Betting

Seat betting isn't especially accurate, but it's interesting to keep an eye on it given the ongoing debates about the accuracy of betting.  I did another run around the seat betting sites last night and found quite a lot of disagreement in particular seats.  I found arbitrages in at least six - an arbitrage is a situation where the odds between two sites are so different that in theory a punter could bet with both and lock in a profit, assuming that one of the two leading contenders won.  (In practice, many arbitrages are so slim that they're really not worth the bother.)

As of last night this was the lineup (see here for methods):

Expected ALP classic-seat gains (not close): Gilmore, Chisholm, Reid, Dunkley

Expected ALP classic-seat gains (close): Petrie, Dickson, Forde, Hasluck, Bonner, La Trobe, Swan, Flynn, Corangamite, Robertson

Split market in Coalition classic seats: Pearce, Capricornia, Stirling, Deakin, Boothby

Expected Coalition classic-seat holds (close): Casey, Canning, Hinkler,  Flinders,  Menzies, Grey, Brisbane,  Banks, Page, Dawson (tied in some markets), Leichhardt (tied in some markets)

Expected Labor loss to Coalition (close): Herbert
Split market Labor held vs Coalition: Lindsay, Braddon
Expected Labor hold vs Coalition (close): Bass (tied in one market), Solomon

Expected Coalition loss to IND (close): Cowper, Farrer (tied in one market)
Expected Coalition gain from IND (not close): Wentworth
Expected Coalition gain from IND (close): Indi
Expected Coalition hold vs Ind (close): Warringah (tied in one market)
Expected Coalition hold against Shooters (close): Calare
Expected Labor hold vs Green (close): Macnamara, Wills

And here's the colour-coded seat tracker:

Key to colours (more may be added):

Red - Labor favourite in all markets
Orange - Labor favourite in some markets, tied in others
Dark blue - Coalition favourite in all markets
Light blue - Coalition favourite in some markets, tied in others
Grey - all markets tied or different favourites in different markets
Purple - IND favourite in all markets
Pink - IND favourite in some markets, tied in others

The seat betting markets are still either not quite buying that it is as close as the national polls say (because there was a strong prior that it wouldn't be), or else are expecting that Labor will get a favourable seat distribution.  Seat total markets have a similar view, expecting Labor to land around 85 seats and the Coalition in the low 60s.

The markets are also expecting the crossbench to stay around its current size, but with turnover (the Coalition regaining Indi and Wentworth while losing Farrer, Cowper and possibly Warringah).  Of the six non-classic seats that appear to be most seriously in play (these five plus Macnamara) it's not easy to get a read on what's going on in any of them, and even if we had neutral seat polling it wouldn't be useful unless it showed a large lead for one side.  Without in most cases having a strong view on the outcomes, my own comments (on request from a reader!) -

* Cowper - there is objective evidence that this is at least very close to falling based on the larger scale of the Oakeshott campaign this time and the retirement of Luke Hartsuyker.  Whether it is more than that, I'm not sure.

* Farrer - there is a good case that this should fall based on anger in the west as seen in the NSW state election and Kevin Mack's profile as Mayor of the electorate's largest population base.  The existing margin is meaningless because it is a different type of contest.

* Indi - I tend to read the votes here in 2013 and 2016 as relying as much on dislike of Sophie Mirabella as desire to have an independent MP for the seat.  That plus it being a vacancy suggests it could return to the fold.  That said some reports I've seen re internal polling are rubbery here, taking a high Nationals primary as a good sign when in fact Nats preferences are pretty leaky in three-cornered contests.

* Macnamara - very messy.  With the double whammy of factors that ailed the party in the state election and a bad campaign for their candidate, it's going to be tough for the Liberals to even hold station despite the retirement of Michael Danby, which would normally have given them a big boost.  However any votes leaving the Liberals have the potential to muck this up for the Greens in their fight against Labor to make the final two.  My suspicion is such votes are more likely to go to Labor.

* Warringah - every report of a commissioned poll seen from any pollster or source has Tony Abbott losing heavily, which I doubt will happen but should count for something even despite the poor record of seat polls.  Confident expectations that he'll hang on are overstating the predictive value of the previous margins (when this is an entirely different level of contest) and the relevance of Abbott's standing (which was baked into the previous results and still saw a less-than-12% margin against a circus of opponents in 2016.)

* Wentworth - the argument here is that the by-election backlash over the Turnbull booting has subsided and voters will go back to the Liberal Party.  Probably the best precedent is Wills 1993 when Phil Cleary won the seat again after being disqualified the first time, but with a huge 2CP swing back to Labor.  There's not a lot of information on this sort of contest though, so it's hard to be as confident as the markets have been at some stages.

Only nine more days to go!