Poll Roundup: Outliers At Thirty Paces

2PP Aggregate (Last-Election Preferences) 50.4 to Coalition (-0.2 in last four weeks)

With One Nation Adjustment (recommended) 50.9 to Coalition
If polls are accurate, either side could win election "held now", probably well into minority

It was just about time for another one of these articles anyway but we've had some extra fun in the last few days with something we've not had for a while, big outliers!  Firstly a 55-45 to Coalition by respondent preferences from Resolve (52-48 last election) and secondly a 51-49 to Labor by respondent preferences from Morgan (53-47 last election). Resolve was the worst headline 2PP of the term for Labor by far, while Morgan was Labor's first headline 2PP lead from anyone since late November, and their first lead from anyone who wasn't Morgan since early October.   Morgan of course put it down to the interest rates cut.  Who to believe?  My aggregate says neither. The net impact of these two plus Freshwater was that Labor improved its standing in my estimate by 0.001%.  

Resolve had Labor on 25 Coalition 39 Greens 13 One Nation 9 Independent 9 others 4.  Resolve has typically had the Labor vote lower than other pollsters lately and this reading is the lowest I'm aware of Labor ever recording from anyone in a federal poll.  The primary vote gap of 14% is the largest of the term from anyone.  Resolve offers a generic Independent option everywhere between campaigns which tends to inflate the independent vote compared to what they'd actually get at an election, until we know who is actually on the ballot papers.  This probably affects their estimate of the ALP primary.  Resolve's One Nation estimate of 9% on the same basis may seem very large, but don't adjust your set, this is One Nation's third 9% in recent weeks, discussed further below.  

Through post-election 2022 and into 2023 Resolve consistently had the Labor lead far larger than other polls (the size of the effect was about 4% on a 2PP basis, though Resolve weren't issuing 2PPs).  This more or less abruptly stopped without explanation at the start of 2024.  Since then in effective 2PP terms Resolve has been middle of the road, but the two polls before this one were pretty bad for Labor.  There is a possibility that this isn't just an outlier but that Resolve's behaviour relative to other polls has changed again - which could be natural, not necessarily a methods change - and my aggregate didn't give the Coalition full credit for this poll for that reason.  Also, Resolve has lately had a string of very low primaries for Labor in between-campaign polls (federal and in three states) but not in their final Queensland poll, which was very accurate.  

Morgan, which also uses a generic-ballot style prone to pump up Independents, had Labor on 31.5 Coalition 36.5 Greens 13.5 One Nation 5 Independent 10 and others 3.5.  While silly people on twitter like to claim that Morgan is the most accurate poll and that the others are skewed to the Coalition, in fact Morgan's headline respondent preferences lately have been very bad for Labor.  Prior to this poll the Coalition had averaged 51.65% 2PP by respondent preferences in the previous ten Morgans, winning nine and drawing one.  That is worse than the worst single headline 2PPs from any of Newspoll, YouGov or Redbridge!  Even with this one included, the average for the last eleven is still 51.4 to Coalition.  So while this one was a hark back to the early part of the term when Morgan would attract criticism for bouncing around like a frog in a sock, it isn't by itself any evidence that Morgan has suddenly started skewing to Labor.  Wait for the next few to see if there's anything on that front.

It might seem bizarre that two polls released in the same week can differ from each other by 6%, unless one of them has a house effect.  But actually this is unusual only in that we haven't been seeing anything like it for a long time it til now - polls have been remarkably tightly packed, I would say too tightly packed, for several months.  Assuming that the reality is more or less between them, the chance that a randomly selected pair of polls with no 2PP house effect would differ by this much is about 1 in 1600.  But there have been 169 different pairs of polls released in the same week since the start of January 2024, so the chance of this happening randomly at least once should be about one in ten.

Other recent polls

Other recent national regular voting intention polls have been:

* This week's Freshwater, with Labor on 31, Coalition 41, Greens 13 and the overlumped others 15, and a 2PP of 52-48 to Coalition by an unstated method.  This is the second Freshwater in a row where the 2PP has been a point worse for Labor than last-election preferences would expect, so I suspect that's not what they're using.   It may be an adjusted version similar to Newspoll's, or they may have changed to respondent without me knowing of it.  By my last-election conversions this one came out at 50.9 to Coalition.  That is pretty benign by Freshwater standards, but I also find that Freshwater is lately less ahead of the flock where the Coalition's standing is concerned.  

* Last week's Newspoll, with Labor on 31, Coalition 38, Greens 12, One Nation 7 and others 12 and a 2PP of 51-49 to Coalition by preference flows based on previous elections.  Alongside this Newspoll, the Australian stated that Newspoll was now adjusting its One Nation preference flow in light of the flows from the Queensland election, meaning its 2PP numbers will from now on be weaker for Labor than 2022 preferences.  By 2022 preferences I get 50.2 to Coalition.

* Last week's Essential, with raw primaries of Labor 30 Coalition 35 Greens 12 One Nation 9 UAP (what is that?) 1 and Ind/other 9 with 4% undecided.  Their "2PP+" was 48-48 (equivalent to 50-50) by mainly respondent preferences; my last-election estimate was 50.9 to ALP.

* Last week's Morgan (ALP 28 L-NP 39.5 Grn 12.5 ON 5.5 IND 10 others 4.5, 2PP 51.5 to Coalition by respondent prefs, 51-49 by last-election)

Overall my aggregate has been up and down as good and bad polls arrive for either side but sees little movement in the last few weeks, so comments in the previous edition re where we stand overall still stand.  There's been no further net deterioration for the government, but if one follows pollster-released 2PPs one will see the slide continuing into the mid to low 48s.  This is partly because Newspoll at least has adjusted its preference flows; I believe some others may have also done so without being public about it.

YouGov MRP

A YouGov MRP model finding the Coalition knocking on the door of majority government off a 51.1% 2PP attracted a high level of interest.  The primary votes in this model were Labor 29.1 Coalition 37.4 Green 12.7 One Nation 9.1 IND 8.9 other 2.8.  

An MRP model takes small samples from each seat and uses these small samples and the assumption that similar seats will behave in similar ways to produce a meshed estimate of voting intention for each seat.  One of these models was done by YouGov with different staff in 2022; this model performed outstandingly in classic Labor vs Coalition seats but not too well in non-classic contests (perhaps partly because of late swing in those contests).  YouGov maintain this version is much better at handling teal seats than the 2022 version.

The voting readout in this poll functioned by offering the respondent the choice of Labor, Coalition, Greens, One Nation, Independent or some other party.  The respondent had to rank these six options in order.  Because Indepedendent is on the ballot everywhere when not all seats will actually have independents of any note, this method is likely to pump up the indie vote, probably at the expense of the generic others if not perhaps the majors as well.  The question lately is whether this method may be also overstating One Nation whose string of 9s from generic-ballot polls is quite remarkable.  (On the other hand, One Nation would be benefiting from the United Australia Party not being a significant presence while Clive Palmer is busy transferring his operations to Trumpet of Patriots.)

Very commendably, YouGov have published new preference flows that they are using in their MRP model, which are based on a blend between respondent preferences and flows at past elections (it would be interesting to know how many past elections).  They are assigning preferences at Green 79% to Labor, One Nation 31%, Independent 59%, other 49%.  Typically this produces an estimate that's about 0.8% lower than a last-election figure.  What is significant in their MRP seat projections is that they apply the same estimates everywhere.  This matters in a seat like Eden-Monaro, where the Green vote in 2022 was 9% with 92.2% flowing to Labor.  Even if that 92.2% drops to, say, 85%, Labor is still doing half a point better out of that than a uniform 79% estimate would suggest.  Of course there are seats where Labor's flow from the Greens is below average too, but in those the Green vote is often low, or the seat is often not competitive.

Broadly, YouGov's model implies that in an election "held now" Labor would suffer from a poor distribution of 2PP swings, wasting good 2PP results in teal and safe Coalition seats and getting crunched in marginals and near-marginals, especially in the outer suburbs and NSW regions.  It is a somewhat similar picture to the Redbridge MRP from November 2024.  YouGov's median estimate is 73 Coalition seats and 66 Labor when my seat model for the same gets 67-68.  

YouGov projects that Labor would lose Werriwa, Eden-Monaro, Robertson, Hunter, Shortland, Paterson, Gilmore, Macquarie and Bennelong in NSW,  Aston, McEwen and Chisholm in Victoria, Boothby in South Australia, Lyons in Tasmania, Solomon in NT, and Bullwinkel and Tangney in WA, but also that Labor would recover Ryan, Griffith and Brisbane from the Greens and Fowler from Dai Le.   As a once-off independent seat (rather than a systematic breakout like the teal seats) Fowler probably doesn't play well with the MRP method so I would not place too much weight on that one.  On YouGov's numbers it is fairly likely that the Greens would actually hold Ryan by jumping Labor on minor party preferences, and also hold Griffith by virtue of Labor falling behind both the Greens and Coalition on One Nation and "independent" preferences.  So it could actually be even worse for Labor than their model.  

I applied a couple of tweaks to YouGov's output, which were to install a personal vote effect (something MRP models tend to heavily dilute) and also to make the Greens preference flow variable based on variations in the 2022 preference flow.  When I did this Labor stopped losing Macquarie but started losing Whitlam.  Also a lot of seats became very very close, such that by applying a 0.5% national swing from their model to all seats I could put the Coalition into a four-seat majority government off 51.6% 2PP, or into not having the most seats and staying in opposition off 50.6% 2PP.  Given the prospect that Labor might form government from a few seats behind, I think any favouritism the Coalition might claim in the YouGov projections is very very fragile.

There has been a lot of detailed work on the individual seat projections of the YouGov poll from observers - I mention for instance the Tally Room post here, a YouTube deep dive by the Anglo Elections Insights channel (perhaps I should know who this is but I don't) and also a number of tweets by the Twitter account enigmatically named Wannon Survivor (@NotTurnbull).  Particularly as a result of the high One Nation vote there are some properties of this MRP that are unusual in ways beyond just 'that seat looks odd' and it will be very interesting to see the closer-to-election models when the candidates who are running in each seat are known.  We'll have a richness of projection data we have never seen before; I'm feeling my usual variable swing model is very quaint and primitive by comparison, but will the fancier models be right or wrong?  

Something else I should mention that has created some myths is the sample size and YouGov's description of it.  The poll refers to "40,689 interviews". An interview is simply a case of a respondent doing an online poll via the YouGov panel, often mixed in with general market research polling that they do.  But it also refers to an effective sample size of around 8,732.  A part of the reason for the latter being much lower would be scaling, but many respondents were interviewed multiple times.  Nonetheless, 8732 is around the sort of effective sample size often considered necessary for this sort of thing to work.  

Redbridge Tracking Poll

Redbridge has started a new tracking poll series that mimics marginal seat tracking polls done by the major parties.  A party wants to know how it is going in the marginals overall so will often do daily tracking of samples from a list of perceived marginal seats, some of which will be just outside the AEC's 6% threshhold for that status.

The important thing here is that the 2PP from the tracking poll should not be compared to the 2PP in national polls, because the poll is not a representative selection of the results from 2022.  Rather it's a selection of where the action is expected to be this time.  On average, these seats have a notional 2PP of 50.92 to Labor meaning that all else equal they will run 1.21% above the national 2PP.  So if the 2PP in them is 52-48 to Coalition (as per the first release) that's spot on the general run of polling.  A uniform swing of that amount would cost Labor seven of these marginals plus also Lingiari, which wasn't included because it's pretty much unpollable.

Leaderships and fitness to govern

The most recent Newspoll saw little change with Anthony Albanese down one point to a new net low of -21, Peter Dutton up a point to net -10 and Albanese's inadequate lead on Better Prime Minister (which favours incumbents) up two to net +5.  Freshwater despite a poor 2PP for Labor had Albanese up seven to net -11, Dutton down four to net -8 and the Better PM gap up from zero to 2 points in Albanese's favour.  Perhaps given the extreme nature of its voting intention result, the best that could be said for Resolve for Labor was that none of it's 4-point Better PM lead for Dutton, -22 net rating for Albanese or +5 net rating for Dutton were records; Dutton has persistently had good numbers from Resolve compared to other polls.  Essential's surprise zero netsat for Albanese in January was replaced with a net -6 to Dutton's net -4.  We're back where we were then: all polls have Albanese rating worse than Dutton (in some cases much worse), none have Dutton rating as badly as pretty as many would have had tickets on when he took over as Opposition Leader, and Albanese is leading inadequately in the beauty contest score.

The Newspoll voting intentions were not that toxic for Labor.  What was far more damning was a question finding that only 34% of voters thought Labor deserved to be re-elected, compared to 53% thinking "It's time to give someone else a go".  It is hard to think of any previous government that would have polled so badly on such a question so close to an election and survived; I think not even Keating 1993, though alas the question wasn't asked back then.  However, just maybe the fact that the question mentions "a federal election for the parliament in Canberra" could have made the responses harsher than if the question didn't mention such a location.   But at least with Newspoll we can look up the exact question wording and see that such questions were asked and consider the wordings.    

Unfortunately the transparency behaviour of the leading pollsters going into this election is in general just not good enough. Three are not Australian Polling Council members (none of these being remotely transparent enough).  The AFR and Nine show contempt for their audiences in not requiring their house pollsters to either be APC members or to at least adopt a similar level of disclosure practice.   The others that are members, with the exception of Pyxis/Newspoll who are excellent, are in general often tardy and/or skimpy in their updating of APC requirements, or are posting APC updates somewhere where not even I can quickly find them.