Rolling National Poll Roundup: 2022 Final Days

2PP Aggregate: 53.5 to Labor (-0.6 since last week) (not a prediction)
Cross-poll average of polls released in last week: 53.0 to Labor 
(Weighted for time only, no house effects or quality weightings)
If the normal range of polling to result relationships applies, Labor remains very likely to win, probably outright

Historically, Labor has underperformed when it has leads in final week polling

This article covers national polling since the update from mid last week and House of Reps federal polls will continue to be added until a final roundup which I intend to post overnight on Friday following the final Newspoll.  The aggregates and averages in the heading will be updated as new national polls come out. (Ipsos is expected, I am not sure if there will be another Morgan, while the Resolve was described as "final" by Latika Bourke.)  I aim to issue a seat polling and seat betting roundup tomorrow, and Senate polling will soon be updated on my Senate prospects page. 

In the last two days three polls have pointed to some narrowing in Labor's lead, but all this has done so far is put Labor back in about the same position as it was in after the gaffe-ridden opening week and a half, reversing minor gains since.  All four polls out this week have shown smaller leads for Labor than their previous offerings, but only in the case of Resolve is the change substantial, and Resolve is hard to take completely seriously on account of its forced-choice methods and overly high Green vote.  

Voting Intention

* Newspoll 54-46 to Labor.  Coalition 35 Labor 38 Greens 11 ON 6 UAP 3 others 7.  The Coalition is again outside the Newspoll recovery window as it has been for much of the year so far - no-one has polled anything worse than 47 2PP in a single Newspoll this close to an election and won (and no-one has polled worse than 48 this close and won the 2PP), though the old Newspoll in its very first season did throw a rogue from which the Coalition recovered about 4% with one week to go.  This Newspoll isn't quite the same beast as past penultimate-weekend Newspolls; it was released on Friday night.  Leader results saw both leaders on a mediocre, nondescript net -11 and Morrison leading by a meaningless point on the skewed Better PM indicator. 

* Essential 51-49 to Labor (48-46 2PP +) raw primaries Coalition 36 Labor 35 Green 9 One Nation 4 UAP 3 ind/others 6. Reallocated primaries Coalition 38.7 ALP 37.6 Grn 9.7 ON 4.3 UAP 3.2 Ind/other 6.5  Essential has tended to have much closer readings than other pollsters, possibly as a result of weighting by party identification. I still think Ind/other at 6.5% is on the low side; either that or One Nation is.  

* Morgan 53-47 to Labor by their attempt at last-election preferences (I get 53.9).  Coalition 34 Labor 34 Greens 13 ON 4 UAP 1 IND 9 Others 5.  Morgan's IND vote is obviously too high as the number of independents is way below the number of seats and they would need to average 14% each - a handful will get double that or more but most are unknowns who won't get their deposits back.  I hate to complain when a pollster stops doing something that doesn't work, but Morgan's decision to adopt last-election preferences does seem a little bit herdy, given that their explanation for it was illogical. "We believe that as the election draws closer and early voting has now begun – starting yesterday – it is more accurate to estimate a two-party preferred result based on the voting pattern of the most recent Federal Election in 2019."  Surely if respondent preferences were ever going to work at all they would be most likely to do so in the leadup to the current election when people are more likely to be thinking about it?  Their respondent-allocated 2PP was 56.5 but this implausible figure which would require a massive 70% of preferences to flow to Labor didn't rate a mention in their report.  

* Resolve 52-48 to Labor by last election preferences (51 respondent allocated).  Coalition 34 Labor 31 Green 14 One Nation 6 UAP 4 IND 6 others 4.  The Labor primary is the lowest by any pollster this term except for another Resolve poll that also had 31 in September 2021 (a time when Resolve appeared to be skewing to the Coalition). The combined major party primary is the lowest in any poll this term (there was also a 65 in the previous term by the short-lived YouGov-Fifty Acres series).  I don't think anyone should believe it's really this low, and successive Greens readings of 15 then 14 make me think this pollster's doing something, probably multiple somethings, wrong.  

Morgan, Resolve and Essential have frightened a lot of red horses but the movements in Morgan and Essential were only slight.  Morgan's 1.5% 2PP drop appears to be down to a dubious last-election estimate (Labor lost 1.5 points in this poll but none went to the Coalition, so my 2PP estimate is down only 0.8).  Essential's 1% drop is off a 1% shift from Greens to others, which makes only 0.4% difference on last-election preferences (though rounding could also be a factor).  For completeness, I also get this Newspoll about half a point less friendly on expected last-election preferences, but that might not even be real because of rounding.  The only serious movement is Resolve which I have down from 54.1 to 51.8, and Resolve is a volatile poll.  Furthermore this Resolve sample was beefed up with phone polling - a decision fairly transparently telegraphed when Resolve started, but I've seen no details on how the two parts of the sample differed in results from each other (if at all).

As a result, while there's a slight overall narrowing signal in the past week, it's not enormous (and my aggregate is a simple one without house effects that may well blow out again if Ipsos comes in on the high side for Labor for a fourth time.)  If there's a difference, I'd say that whereas a week ago a repeat of the 2019 polling failure would still be a likely ALP win, today it would be a lot more of a tossup.  However, internationally it's rare to get two failures of the same size in the same direction in a row.  (And no, US President 2016/20 wasn't one of them - the national US polls in 2016 were pretty good, with an average error that was equivalent to just under 1% 2PP in our system.  It was largely the polling in a number of specific states that caused the wrong Electoral College winner to be projected.)

Further Adventures In The ALP Fail Factor

If we had a Coalition opposition that was wobbling around between 54 and 53 2PP all interest would have long gone out of the election and there would be general agreement that Labor was clearly smashed.  Indeed that was much the case by the final week in 2013 when the Coalition wasn't even doing that well.  Even if the Coalition's lead narrowed, as it did in 1996, people would say, so what? (The 1996 narrowing at the end proved to be totally illusory.)  But, 1993 excepted, Labor has a bit of a history of falling over compared to its lead-in polling, so should we be especially cautious about Labor's current (and still solid) average lead because it is a Labor government?

I've been aware of what I call the "Labor fail factor" for some time - that Labor tends to underperform its leadup 2PP polling, albeit not on average greatly, but Mark the Ballot has shone a new perspective on it with an excellent graph this week.  He finds that for late campaign polls, the polling average tends historically to be accurate if the Coalition is doing well, but the further ahead Labor is, the more of their lead proves illusory.  His predictive model incorporates this, and still finds a Labor majority likely.

I thought at this stage, it would be useful to look just at the average of polls released around the final week and a half (including late in the penultimate week) - but also, for historic estimate purposes, I decided to remove Morgan.  Morgan had very large misses when it wrongly predicted Labor wins in 1996, 2001 and 2004, and was also at the sinners' table (along with all the rest) in 2019.  For the elections for which I have readings from multiple non-Morgan polls, the same effect is seen, but the size of it is much reduced (more like 1% instead of 2% at the lower end):

The 2PPs are for the Coalition, the dotted line is the line of best fit and the black line is the line of parity.  Overall, an average of polled 2PPs for the last week and a bit without Morgan was a wonderful election predictor until 2019 (the outlier) came along.   

If I include 1987 and 1990, for which the only non-Morgan poll I have is Newspoll, the relationship is a little bit more like MTB's - but the widening is driven entirely by the 1987 rogue.

At this point of time the average of the current non-Morgans is only 52.3 by their released 2PPs (which I also used for old polls for the graphs above), but that is likely to increase given the recent form of Ipsos and Newspoll that still each have a reading to come.  Still, if this level of underperformance is still a thing, Labor would be doing well if they could beat 52-48.  (Which is, incidentally, about the 2PP of the YouGov MRP model, which had a very large sample size and very advanced weighting, though it would now be showing some age if voting intention has changed in the meantime.)

Explaining why this underperformance exists would be a long story, especially when it certainly doesn't exist at state level.  Soft voters? Long-running representative issues?  Something about our federal issues and campaigns?  

Should we assume it's a thing this year?  Maybe, maybe not.  We're coming off a polling failure that only some polls show meaningful signs of having addressed in some fashion - but on the other hand, overcompensation following a polling failure is possible.  

As I write, all the forecast models I've been seeing (Armarium Interreta, Mark the Ballot, Australian Election Forecasts and Buckleys & None) find Labor still more likely than not to win a majority with median estimates in the range 77 to 81 seats, but all find that it is still realistic if unlikely for the Coalition to win (most likely not outright). The YouGov MRP model had a very similar finding despite a clearly lower 2PP than surrounding polling.  The models produce similar overall results, with Buckleys & None the most cautious at the moment (unlike in 2019 when there was a lot more variation in modelling estimates!) but their modelling decisions are quite different.  The betting markets have a different view of this tonight with a Labor majority being seen at prices like $2.15 and $2.35, headline odds giving Labor only a 62-64% chance of winning, and seat totals odds (albeit with the customary over-rake) pointing to Labor falling just short of a majority.  (I have not checked the seat betting yet.)  This is somewhat contradicted by an also over-raked Sportsbet 2PP market that has the Labor 2PP on for a median of 52.5 (almost certainly a majority), though it also has the Coalition 2PP on for around 48.5 (hmmm, I don't really think that it works like that).   Are these markets onto something with this almost here-we-go-again mood, or are they suffering from recency bias and jumping at noise in some relatively flimsy polls (as happened with a rather funny run following a previous Resolve)?  Who is right and who is wrong?  It is only a few more days before we find out!

I may have more to add to this article later.  I made several George RR Martin jokes on Twitter about whether or not I would meet a self-imposed midnight deadline, so I will put it up now and see if anything else occurs to me to add.  For one of our leaders, winter is coming, and it could come a little bit early this year.

Thursday: Ipsos: Ipsos came in at 53-47 to Labor (about 55-45 respondent-allocated), which sounds solid but given its previous three have been very strongly pro-Labor, it will probably narrow some aggregates and projections.  (It did not narrow mine, because mine is very simplified.)  So while it may reassure Labor supporters, Coalition supporters could see it as a signal of hope, at least of denying Labor a majority.  Coalition 35 Labor 36 Greens 13 One Nation 5 UAP 3 others 8.  My estimate of last-election preferences off these primaries is 53.6, but the difference could plausibly be down to Ipsos calculating 2PPs by state rather than federally.  That said the 2PP being 51-44 prior to allocating undecided (which appear on primaries to have broken 2-2-1 Labor-Coalition-Green post-rounding, though these changes might be very different before rounding) is further evidence for the 53 being rounded down.  (In my view pollsters should publish their last-election preference formulae to assist verification.)  

Interestingly Ipsos has had Anthony Albanese as preferred PM in all four polls so far, albeit narrowly (3 points) in the most recent, 42-39.  Ipsos also differs from Newspoll in finding Morrison (net -17) a lot less popular than Albanese (net -4).

Bizarre Morgan PS: Today I was looking at this Morgan table which purports to show their 2PPs by last-election preferences (used for the last two polls) and respondent preferences (applied previously).  As noted above the difference for this poll was 3.5% vs 1.5% on the previous poll.  However, I noticed that the difference is also 1.5 for 33 consecutive polls between August 2020 and the early May 2022 poll.  Given the volatility of sampling respondent preferences, the chance of this difference occurring by chance is zero point a great big bunch of zeros, and for instance on my own estimates off their primaries there was a case of no difference in the last six.  It appears that a uniform adjustment has been applied rather than recalculating - given a spreadsheet with a clear record of national primary votes I could apply a conversion formula within minutes.