Wednesday, May 11, 2022

Poll Roundup: Ghosts of 1996 and 2007 Edition

2PP Aggregate: 54.1 to Labor (+0.9 since last week)
(Weighted for time only, no house effects or quality weightings)
If the normal range of polling to result relationships applies, Labor is very likely to win

I've been very time-challenged this election (expending much of last night refuting an American know-all and sundry reply-guys about our preferencing system on Twitter probably didn't help) but finally, maybe, I can get something out in time for an article about the state of national polling. 

As an overall summary, Labor suffered alarmingly rapid losses in support during the first week and a half of the campaign as documented in the previous edition, but the Coalition's position did not improve over the next two weeks, and now in the latest round of polling there has been a bit of a blowout.  Labor's lead is so large now that the Coalition is extremely unlikely to win by normal means.  What hope remains for the current government is that the polls are even more wrong than last time (despite everything some of them have done to repair it) or that they are about or nearly as wrong as last time and the government makes some improvements over the final week and a half then gets lucky with the seat distribution and scrambles back, perhaps with a sub-50 2PP.  It's not all over yet, but on the other hand the prospect of a decisive Labor win is much stronger now than it looked even three weeks ago.

This election has long been compared to 1996 in the sense that a government had had an upset win that was one too many and faced a troubled term where all the opposition had to do was get its ducks into line and abandon its previous errors.  In the 1993-6 term the Keating government from time to time polled well, but only when the Coalition was having leadership crises under Hewson (who lost all authority with the 1993 result) and Downer (who wasn't up to it).  In the 2019-22 term, the government only led significantly during the early pandemic rally-round-the-flag polling phase, and since early 2021 it's just been down and down and down, until a now modest degree of recovery this year. 

If Labor does go on to win decisively I will think of it as having a lot in common with 2007.  John Howard couldn't understand why he was losing in 2007 because the economy was in strong shape, and likewise Liberals now must be bewildered that they have stared down fears of massive unemployment, taken official unemployment to remarkably low levels, saved the economy from the early pandemic, and yet the voters appear thankless.  My view on 2007 is that while the Global Financial Crisis had not yet happened, voters had a sense that a very different economy was coming and didn't trust the Coalition to take sufficient action in response.  It feels the same now and the government's history of often not being seen to do anything when confronted with a challenge (Jobkeeper/Jobseeker excepted) is working against it.  The other problem is that the good economic figures are the proverbial feral abacus in the face of lack of wage growth. 
For sure there are differences - in 2007 a still respected Prime Minister faced a rampantly popular Opposition Leader while in 2022 a tarnished PM (this week a rather bad netsat of -14 in Newspoll) faces a challenger with underwhelming ratings (this week a net -6).  If anyone thinks the latter is an obstacle, it isn't; even Tony Abbott managed to win 90 seats in 2013.

This week's polls

This week's Newspoll came in at 54-46 off primary votes of Coalition 35 Labor 39 Greens 11 One Nation 5 UAP 4 others 6.  The One Nation vote has recently risen in Newspoll simply because One Nation are contesting every seat except Higgins and Kennedy.  Newspoll has attracted some attention for its recent run of very stable 2PPs (54-53-53-53-54) with some echoes of concerns in the previous term about Newspoll being "underdispersed" (ie less variable than a random sample should be).  However there are important respects in which Newspoll's 2PPs are not a random sample.  The first is the use of targeted respondent selection, which may in effect ensure that much of the sample is actually sub-samples of groups that are less variable in their voting intentions than a pool of 50-50 voters.  The second is the use of last-election preferences, which reduce the variability in 2PP results among those who are not voting for major parties.  Anyone wanting to model dispersal or otherwise in Newspoll would be better off using the primary votes.

This week's Ipsos came in at 57-43.  The primary votes were a dog's breakfast because they were reported in raw form (Coalition 29 Labor 35 Green 12 ON 4 UAP 3 others 9 undecided 7) and while that would normally go with a Coalition primary of about 31.2 (Labor 37.6) it was separately published that the Coalition primary was 32.  I got 56.1 by my own estimate of last-election preferences.  The headline 57-43 was higher than any lead for Labor in the previous term and higher than any lead this term from anyone other than Morgan.  This didn't stop Phil Coorey understating things considerably by claiming that the poll merely showed a "sizeable gap" and that Labor "could win government in its own right if the numbers hold."  (It would in fact be the most lopsided result since 1943.)  This wasn't the only case of this from Coorey lately as two weeks ago he argued that a 50-50 subsample in Queensland "buttressed" "the Coalition's belief that it will hold all or most of its seats in Queensland" and "is nonetheless consistent with views from within the major parties [..] that Queensland could remain static".   In fact 50-50 in Queensland would be a disastrous 8.4% swing that if uniform would cost it around eight seats and all else being equal put Labor into majority from Queensland changes alone.

This week's Morgan was 54.5 to Labor by last-election preferences (56 by respondent preferences).  Amazingly, Morgan has stopped using respondent preferences as its headline figure, in what might be uncharitably taken as an attempt to get with the herd.  However it still doesn't seem to have changed how it handles the Independent vote and so it gets an apparently overstated 8.5% for independents, with 34% Coalition, 35.5% Labor, 13% Green, 4% ON, 1% UAP and 4% others.

Polls out last week but not so far this week included Essential at the equivalent of 52-48 (raw primaries Coalition 36 Labor 35 Green 10 ON 4 UAP 3 Ind/other a suspiciously low 5 undecided 6).  Note that Essential has a rather high estimate of the combined major party vote once undecideds are redistributed, with both in the high 30s.  There was also Resolve finally releasing a 2PP at 54-46 off primaries of Coalition 33 Labor 34 Greens 15 ON 5 UAP 5 IND 4 others 4.  The overly high Independent vote in Resolve has gone away as they have adjusted to Independents not being on the ballot everywhere, but the Green vote is hard to take seriously.

Overall I get an aggregated estimate of 54.1-45.9, as a simple average of what the polls are saying, ignoring possible house effects and the quality of different polls, and weighting only for time (albeit conservatively).  If repeated at the election this would be a seat result of around 87-58-6 to Labor, ignoring crossbench gains (most likely from Coalition) which are beyond the model's capacity to project.  To illustrate the government's difficulty, on average a single Newspoll released two weeks out has been within 1.8% of the actual result, with no error greater than 3% thus far (though there was a bigger outlier one week out in 1987).  In terms of aggregates, a repeat of the 2019 error (the largest for decades) from here would only put the Coalition to the low 49s, which is probably a loss but might still be winnable or at least enough to keep Labor to a minority.  But if we are talking about downside polling errors for Labor, we should also consider the upside, which would be a massacre.  
However, historically the incumbent government would still be expected to do better than the current polling.  My very rough median projection from polling since the mid-80s if the normal historic relationships hold is 52.2-47.8 to Labor, which would be a likely Labor majority with about 80 seats to 65 for the Coalition, minus any losses to the crossbench.  Note that Armarium Interreta have reported a far more bruising poll-based projection of 53.5-46.5 with an overall projection of 52.7 to Labor (hello 2007) and a median seat result of 82-60-8.

Historic polling-based models (some with a dusting of other ingredients) now give the Coalition not much chance of re-election at all.  Armarium Interreta has 4% for a Coalition majority and 11% for a hung parliament, Australian Election Forecasts has a 5.5% chance of a clear Coalition win and 13.7% of a negotiable balance of power and Buckleys & None has 22.3% for a hung parliament and 8.1% of a Coalition majority. A win chance range of around 10-20% for the Coalition, including minorities, seems a fair read to me on the assumption that there is not an escalation of the polling errors seen in 2019.  And while it would not be surprising if pollsters were still skewing to the left somewhat despite their best efforts (just as overcorrection a la UK 2017 would not be surprising either) it would be very surprising if they managed to be more wrong immediately after a major polling failure that some of them had done their best to address.  
It is pleasing to see a reduction in the volume of hung parliament hype in the last few weeks, with one of the remaining hosts of this infection being Chris Uhlmann, who on 27 April said the Coalition could not win without a four in front of its primary vote and had not been so low since 1946.  This was wrong on multiple fronts: the Coalition last had a 3 in front of its primary in 1998 (and won) and actually did have a 4 in front in 1946 (he's confusing it with 1943),  Uhlmann also said "all the public polls show neither major party has electoral support to hit [a majority]." when in fact nearly every released poll this year has had Labor in a position that would be majority-winning on election day.  

I mention in passing modelling I have done on the composition of the Others vote (excluding One Nation, UAP and independent). The mix of Others parties has changed greatly but in terms of the expected preference flows I find that the increased field of Liberal Democrats should roughly offset the disappearance of the Christian Democrats and some minor right parties like FACN, meaning that we don't need to adjust our sets on possible preference flows from these parties.  The flows from United Australia and One Nation will be more interesting.  The UAP flow might weaken given that the UAP has recommended against the Coalition in some seats and given its relatively evenhanded campaigning to this stage.  The One Nation flow might in theory strengthen as One Nation is recommending preferences to the Coalition almost everywhere (from memory in 2019 it ran open tickets) but also it is running in more seats where it has no presence, so it might pick up more random voters. 

What's the deal with this MRP thing?

Initial results of a YouGov MRP model similar to models that have been run in the UK have been released.  MRP stands for "multilevel regression with poststratification" and it is better to refer to the individual seat results as a poll-based model of the seat, not a seat poll.  This is an untried method in public polling for Australian federal elections (there have been claims it was used with large-scale union polling in Queensland during the 2015 leadup). The individual seat sample sizes at an average of 125 voters have prompted comments about that being obviously too small, but the model does not consist of isolated tiny samples and usual ideas about sample size do not apply as strongly.  Rather its gamble is that if you have a range of results from seats with similar demographics, you can in effect use these samples collectively to project the results and tone down the bounciness of the individual samples.  It's an imperfect method because while demographics correlates to seat results, it doesn't explain nearly everything about them, and whether it can predict third-party races like the teal indie seats will be especially interesting to see.  If it works, what it should be able to do is get a handle on types of seats where the swings are above and below the national average.  

The Herald-Sun's coverage included figures for 19 seats.  This included 11 with 2PP results, and an average swing in those of 3.7% (consistent with a Labor 2PP in the low 52s - but the seat mix chosen could be very arbitrary).  However the average was blown out by a projected swing of 6.1% to the Coalition in Cowper (which seems a little hard to credit though this year does not have any Oakeshott factor and the new incumbent picks up a personal vote).  Of the crossbench seats, it had Indi, Warringah, Mayo and Clark all retained (though the 61-39 in Clark is an obvious big error based on the stated primaries and I wonder if it was a typo for 69-31).  It had teal indies short of the final two in Boothby, North Sydney and Curtin, losing in Mackellar and Wentworth (a surprisingly heavy 56-44) and only winning in Goldstein and Kooyong (both of those narrowly).  We shouldn't read too much into these individual results; the test will be how the model goes in projecting the overall number of teal indie wins and Green gains (none of the latter at present). I may comment further on the full MRP model, though I am finding at the moment that there are not enough weeks in the average day for me to do everything I'd like to do with this election.

The MRP model has way lower results for teal indies than much of the released polling for those seats, but I believe that the released polling has been extremely cherrypicked, and understand the overall thrust of internal campaign polling is nowhere near as positive for the teal indies as the tip of the iceberg we've been shown above the water.  In this light it was also interesting that after an earlier uComms internal poll showing Monique Ryan with a ridiculous lead, this week Redbridge released partial data finding that she was only now building enough primary vote for a (just) winnable position.  Nonetheless the fact that the Liberals have agreed to high-profile debates between their incumbents and independents in these seats strongly suggests that what is on their own radar is causing them very serious concerns, about Kooyong and Wentworth in particular.  There was one report of Liberal internal polling said to show Tim Wilson in Goldstein on a hopeless 37% and Frydenberg on a dicey 42%.  Methods details were insufficient but the former was strangely not wildly different to an Australia Institute poll with Wilson only on a very difficult to credit 34.5%.  Some of the dire Liberal internals may be put out for expectation management or to discourage complacency about voting for independents.

(Update: as I was writing news came through that the model has an average outcome of 80-63-8 to Labor, gaining Chisholm, Higgins, Brisbane, Robertson, Swan, Pearce, Boothby, Reid and Bass with Bennelong, Lindsay, Ryan, Corangamite (ALP), Longman and Sturt all at 50-50.  Plus Coalition losses to IND in Goldstein and Kooyong.)

(Update 2: A number of odd results have been noticed including large 2PP swings to Labor in Parkes, Mallee and Hunter, very low IND votes in Fowler and Bradfield etc.  The model is bound to produce silly results in some specific seats, the question being how it goes at its core tasks. It has also been noted that the Greens are only 1% behind Labor in Ryan and Brisbane and might conceivably get into 2nd.  Anyway, you can see the full figures here.)

Murdoch tabloid exit polls

"Exit polls" by the Murdoch tabloids have been reported in various seats with sample sizes usually around 100-200.  Examples include Corangamite (Labor ahead), Dawson (close), Capricornia (LNP ahead), Flynn (close), Kooyong (Liberal ahead), Flinders (Liberal ahead), Deakin (Labor ahead), Chisholm (Liberal ahead), McEwen (Labor ahead), Ballarat (close) and Wide Bay (LNP ahead).  Small sample size aside for given seats, these are convenience samples conducted by reporters without weighting or other professional processing and the danger is that prepoll votes in particular booths could be stratified by time of day or day of prepolling (there might, for instance, have been a rush to vote fast among voters keen to kick the government out - or it might be that early prepollers are more pro-Coalition than the late ones, as is the case with postals).  

 One report in the Courier-Mail had a combined 2464 voters from 19 mostly Queensland divisions with totals of LNP 48 Labor 31 Green 6 One Nation 5.  The seats canvassed were on average seats the Coalition won 59.4-40.6 in 2019, and the Coalition was on average a few points better in prepolling too, so an implied 2PP of around 58-42 is actually consistent with a substantial swing (say 4%) to Labor.  But we don't know how evenly the votes were taken.  

Some seat polls

As mentioned above there are a lot of internal Climate 200 or campaign seat polls that are being released (usually incompletely) in teal indie seats, and I am wary of those.  We have also seen a reported (by Ten) leak of NSW Liberal internal polling said to show the party losing Reid and Robertson and behind 45-55 in Bennelong.  Bennelong is vacant and is a likely epicentre for any Chinese Australian backlash against the Morrison government so the seat falling is far from implausible if the swing is on, but there may be some intention behind the apparent "leaking" of these numbers.

Of slightly more interest to me was a uComms of Boothby for the SA Forest Products Association (55-45 to Labor, consistent with an earlier TAI poll at 57-43) because forest industry polling tends not to be susceptible to selective release.  Nonetheless after uComms' 30-point howler in Stuart at the state election I am not weighting their robopolls that heavily either. Particular polls might turn out to be correct but I think that anything that still uses age/sex/location as its only weights after 2019 is confused about whether it is a poll or a dating site.  

Overall the seat polling at this election so far has been dominated by companies that use simplistic or in cases weird methods, has been mostly commissioned by biased sources, and has been variable in terms of transparency.  I have not included any of it in my seat model.


I haven't tracked seat betting as much at this election as I would have liked, but here's the current situation by my usual methods of assessing it:

Projected Labor gains from Coalition (not close): Reid, Boothby, Swan

Projected Labor gains from Coalition (close): Longman, Bass, Braddon, Chisholm, Pearce

Mixed market: Robertson

Projected Coalition holds vs Labor (close): Bennelong, Lindsay, LaTrobe, Robertson, Brisbane, Flynn, Leichhardt, Sturt, Deakin, Hasluck, Tangney, Casey, Higgins

Projected Coalition holds vs Labor/IND (close): North Sydney

Projected IND gains from Coalition (close): Wentworth, Goldstein

Tied market IND vs Coalition: Kooyong

Mixed market IND vs Coalition: Hughes

Projected Coalition holds vs IND (close): Mackellar, Nicholls, Curtin

Projected Labor holds vs Coalition (close): Lingiari, Lyons, Gilmore

Projected Labor holds vs IND (close): Fowler

By favourites, the projected total is Labor 77.5 Coalition 64.5 others 9.  On a close-seat adjusted basis it is Labor 80.2 Coalition 61 others 9.8.  Seat betting is not necessarily predictive and in fact has a poor track record, but I like watching it to keep testing that.

Righto, off to watch the Seven debate then.  It could not be as half as bad as Nine's, could it?


  1. I like 1993/6. fightback was a great document but they could not sell it. Howard then went for a very small target policy and The libs got more popular whenever he made a stumble.

  2. I find the betting markets fairly interesting. I'm finding that the overall win market is quite liquid now on Betfair. There was an obscene volume traded on Betfair for the US presidential election, particularly on the overall win market, and from my recollection it was far more predictive than the poll/poll aggregators in 2020. For this election, there is reasonable volume on the win market, with around $5 million having been matched. You can back the Coalition at $3.35 (implied odds of around 30%). There really isn't sufficient volume on seat betting to be too instructive I'd say. There are two seats with significantly more matched than others:
    Kooyong has around $64k matched with Frydenburg currently a slight favourite, but it isn't overly liquid right now.
    Warringah has had $34k matched, Steggal currently a $1.10 favourite but not a lot of liquidity currently.