Saturday, May 13, 2023

New South Wales 2023: Final Results, Poll Accuracy And 2PP Pendulum

It's been a long time coming because of other work but here finally is my wrapup piece for the 2023 New South Wales lower house election.  The twelve-year old Coalition led by its fourth Premier Dominic Perrottet was sent packing by the Labor opposition under its fifth leader Chris Minns, but optional preferencing and some luck in the close seats cushioned the blow and the Coalition managed to retain 36 seats.  Labor after looking almost sure to get a majority on counting night (more of that later) ended up with just 45 but no trouble at all forming government given that the crossbench held three Greens and nine independents.  The result snapped a 15-election streak at state and federal level since the last case of a non-majority outcome.

Vote Share, 2PP and Preference Change

The election saw a further slight fall in the major party primary votes.  The primary votes were 36.97% Labor, 35.37% Coalition, 9.70% Greens, 8.76% independents, with the rest led by 2.23% Sustainable Australia, 1.8% One Nation, 1.56% Shooters Fishers and Farmers and 1.28% Legalise Cannabis.  The 2PP was 54.27% to Labor, a swing of 6.29%.  

There was a slight increase in preferencing at this election with 35.2% of non-major party (or minor Coalition partner in three-cornered seats) preferences reaching Labor on a 2PP basis (+2.2%), 16.1% reaching Coalition (+2.1%) and 48.7% (-4.3%) exhausting.  However these figures include many Independent and Green preferences that were not actually distributed in many seats.  The Greens' share of all non-major votes dropped slightly from 38% to 35%.  



Seat Swings

Many interesting things happened with seat swings at this election.  Firstly, they were remarkably variable given that the overall state swing wasn't all that large.  I am used to elections where the standard deviation of seat swings is about 3-4%, but in this election it was 6.2% (5.7% excluding the outlier Kiama).  

This is how the swings in given seats compared to Labor's pre-election margins.



Seats marked in grey were won by Independents and seats marked in Green were retained by the Greens.

Labor gained eight seats (the red dots left of the y-axis) that were notionally Coalition based on the 2019 election: Camden, East Hills, Monaro, Parramatta, Penrith, Riverstone, South Coast and Bega (which it already occupied as a result of the 2022 by-election).  It also became the 2PP winner in Kiama, which delivered a ridiculous 31.76% 2PP swing because the Liberals ran dead and voters for ex-Liberal independent Gareth Ward did not preference them.  The swing was wasted for Labor because Ward retained the seat.

The increased standard deviation should have helped Labor (in my seat model it raises their expected score from 46.5 seats to 47.5 seats).  Yet Labor won only 45 seats - why was this so?  

Firstly Labor were unlucky, winning none of the six closest seats (all won by Coalition by under 1.5%, with Ryde won by 54 votes.)  This shows on the graph in the number of blue dots that are just to the left of the diagonal line.  However there was also a degree of swing management which shows up more in the median swings for various margins than the mean.  In seats notionally held by the Coalition by under 10%, the median swing to Labor was just 3.9%.  Labor did in fact match the pendulum swing (failing to win five seats that were on margins below the actual swing while winning five including Bega that were above it) but given the number of Coalition retirements in this seat range and the disruption effect of the Bega by-election, Labor should, all else being equal, have done better than that.  Aside from the lowish median swing in the Coalition's greatest-risk range, one further reason Labor didn't do better was that nearly 0.3% of its 2PP swing was wasted on the seat of Kiama, and also because of various oddities of optional preferencing the average seat swing was slightly lower than the state swing.

With a swing this large the Labor side of the graph wouldn't normally be interesting (since they were never going to lose anything), but look at this:


There was very little swing in ultra-safe Labor 2PP seats while in ultra-marginal Labor 2PP seats the average swing approached 10% (Kogarah in the top left is an outlier but not a massive one; the trouble spots below the x-axis are Cabramatta and Liverpool).  This repeats a pattern seen in Victoria which rendered the swing to Coalition there ineffective.  In this case it has also compressed the Labor side of the pendulum in a way which might make life difficult for the Coalition in future - as if the global issue of voters not drifting right with age the way they used to doesn't make things hard enough.  In NSW the good news for the Coalition is that there is an entrenched electoral system that gives them some help, but it's always possible Labor in government will find some ways to tinker around the edges of OPV to reduce the damage.  (It could place restrictions on just-vote-1 signs that are not in party colours for example, except that it has in some cases used them itself.)

This election was another to give a massive tick to personal vote effect theory.  Seats where the Coalition incumbent retired returned swings of 1.1% above average to Labor, while the three with a new Labor MP elected at the previous election were 4.8% above average.  Those where the Coalition MP from 2019 had been replaced in a mid-term by-election were 4.4% above average (led by the monster swing in Bega which Labor won mid-term).  In seats with Coalition sophomores, Labor's performance was 4.2% worse than average and Labor's retirements included their two shockers in Liverpool and Cabramatta.  The above average swing in Strathfield (despite a mid-term ALP by-election) was the one counter-point but there were not many.  

Winning from behind, non-classics, OPV, etc

22 of the 93 seats finished as non-classics, up slightly from 19 in 2019.  The Greens defeated the Coalition in one seat and Labor in two.  Independents defeated the Coalition in six seats and Labor in three (one of them Kiama).  The Coalition defeated independents in six seats, and each of the following happened once: Labor defeated Independent, Labor defeated Green, Labor defeated One Nation (this was Cessnock where the disendorsed National finished fourth) and Liberal defeated National.  Of the ten non-classic seats won by a major party in 2019, four reverted to classic status, five remained non-classic but were retained by the same major party (in three cases with the opponent switching from Green to teal), and only one (Wollondilly) was gained by an independent.

The growing minor party vote results in a fair amount of misinformed hostility to two-party preferred polling and 2PP-based projections.  However this was another election that was won and lost on classic 2PP seat shifts where what went on with the crossbench was a sideshow that didn't even determine whether or not Labor won a majority.  Aside from recontesting crossbenchers getting re-elected, the only things that happened were the Independent gains in Wakehurst and Wollondilly which simply rubbed a couple of extra seats into the Coalition's wounds.

The election has been portrayed as a flop for the teal movement a la Victoria but it actually wasn't that bad.  Both Michael Regan and Judith Hannan are not out-and-self-identifying teals and are the sort of local independents who have often won in NSW anyway.  However they do have connections to the movement and/or Climate 200.  Teals were also competitive in defeat despite being hampered by OPV in various other seats.  

Six seats (up from two!) were won by a trailing candidate on preferences, these being the Labor wins in East Hills, Monaro, Penrith and South Coast and the Independent gains in Wakehurst and Wollondilly.  Of these only Wollondilly happened from a significant among behind, the others being typically within 1% on primaries.  Antony Green argues here that optional preferencing cost Labor a majority by depriving it of Ryde and Terrigal and that the Coalition would have also lost Pittwater under compulsory preferences.  This sort of impact (a few seats) is quite common in my modelling of apparent impacts of OPV.

Vanishing Majority

On election night it appeared extremely likely that Labor would get a majority, a result that was pretty much called by the ABC which at one stage had Labor on 51 +/- 3 seats.  (I never called that Labor had a majority but I did say it was overwhelmingly likely).  However the apparent Labor majority disappeared as counting went on in a complete failure of post-election projection.  Given that the 2PP (54.3) ended up not that much short of where it looked to be on election night (around 55), what went wrong?

The first factor that caused projections to fail was that the swing in on the day voting was not repeated in prepolls.  In seats for which 2PP swings are readily available the on the day swing averaged 6.8% and the prepoll swing averaged 4.8%.  This could have been caused simply by changes in the composition of the booth and prepoll voters, but late swing in both 2019 and 2023 is also a likely culprit.  In 2019 Labor had a very bad last week (including a disastrous debate performance) and underperformed compared to polling while in 2023 Labor's result was at the high end of the range of polls (and more or less matched the last poll in field) and a late accusation of ambulance queue-jumping may not have assisted Premier Perrottet.

However, this pattern was especially pronounced in the seats that were close at the end of election night.  Of the seven 2PP seats in most doubt (Drummoyne, Holsworthy, Goulburn, Miranda, Oatley. Terrigal and Ryde), all except Drummoyne and Oatley saw day booth voting vs prepoll differences of at least 4.0%, with Holsworthy at 5.9%, Goulburn 5.2% and Ryde 5.1%.  Considering the swing for Gareth Ward in Kiama as a candidate, there was a difference of 3.5% in Kiama.

As a result of this, while Labor had a sufficient swing in booth voting to win Holsworthy, Goulburn, Oatley, Terrigal, Ryde and Kiama if that swing held up across all remaining vote types, in all these seats except Ryde it was behind on combined booth and prepoll votes.  In some cases there were special factors behind this, most notably Holsworthy where the behaviour of prepolls in 2019 was aberrant and the 2023 prepoll locations were all different (so the projected Labor lead was always flimsy anyway). In Goulburn, the Liberals benefited from the fact that swings in the actual Goulburn city booths were to them and this then flowed on to the large prepoll.

In some of these seats there was a bigger swing to Labor on postals than in on the day voting, meaning that where Labor was already ahead or not far behind they might have been able to minimise the damage on postals and win on absents and provisionals by enough to maintain or recover the lead.  Except the primary problem here was that there were so many more postals than in 2019.  In the case of Ryde, the postal swing was 13.1% compared to the booth swing of 10.9% and the prepoll swing of 5.8%.  But the increase in the size of the postal vote was such that even despite that massive swing, the Liberals' 54.8:45.2 edge on postals was still good enough for a 688 vote win of the postal count, compared to a 678 vote win of a far smaller postal count in 2019.  

[A small section of text has been removed here as it is under revision - KB]

Polling Performance

If the 2023 NSW election was a bad night for projections, it was a rather good night for statewide polling.  Polls generally forecast that Labor would win the 2PP comfortably though most had Labor not getting quite enough for a majority to be likely.  As usual there were various levels of nonsensical spin and hype from media sources about the election being very close when Labor had a decisive lead and was always on track to form a government of some kind. 

As in the federal election I use four different indicators of final poll accuracy:

RMSQ: root mean square error on primary votes.  Punishes outliers
RMSQ2: the same but with the error on the 2PP given a 50% weighting, because I think 2PP is an important part of an Australian pollster's job
AVE: average error on primary votes
AVE2: average error weighted 50% for 2PP and 50% for average error on primary votes (this is my leading indicator)


Note that for Resolve no 2PP was explicitly published online before the election to my knowledge so I have derived one from accompanying reporting of a "4.5-point swing".  Resolve figures to one decimal place were published but I have not seen evidence that they were published before the election, so at this stage they are not included.  (If I see such evidence the table above will be revised, but it will not change much).  

There was a lot of to-ing and fro-ing with the 2PP and the primary vote shares in the postcount of this election, which led to various pollsters prematurely declaring they had been closest to some pin or other.  (Really, it's best to just wait for the final results.)  In the end Newspoll was easily the closest on the 2PP, getting it right to within rounding error, its fourth election in a row within 1%.  Freshwater was slightly closer than Newspoll on average on primary votes, and overall Newspoll was the most accurate considering both the 2PP and primary votes.  All four final polls were, however, at least reasonably good.  

Resolve was the only pollster to release a specific breakdown for independents (8%) and this was accurate enough (actual 8.76%).  Morgan significantly underestimated the combined major party vote.  

There were various adventures on the way to these results, including Morgan in its second-last poll having the One Nation vote as high as 8.5% on the basis of including One Nation as an option everywhere (I suspect an overestimate even on that basis).  The second-last Resolve had Labor with a primary vote lead of 6% which would have resulted in a thumping victory, consistent with a trend since the federal election of Resolve polls producing far greater primary vote leads for Labor than other polling, yet this did not happen in their final poll.  (A similar difference in tracking was seen in Victoria, and both federal and Queensland Resolve non-final-polls are running strongly for Labor lately compared to other polling).  It is notable that the final Resolve includes a small CATI (live polling) component while the lead-in Resolve polls do not, but it is not obvious that either this or anything else explains the difference.  More on this in my Budget federal roundup maybe.

Ah, but didn't Newspoll wrongly predict a Labor majority?  Not as such.  The poll itself did not release any seat projections.  The Australian's article however reported "NSW Labor leader Chris Minns is poised to lead the ALP into government [..] according to Newspoll data that shows the party is on the cusp of claiming a majority. [..] Mr Minns set to claim up to 10 seats, enough to reach the 47 seats necessary to govern without the support of the crossbench."  Unfortunately the Australian did not include the words "which would be" before "enough", which would have made it look a little less like a prediction.  Aside from that, that Labor was in maybe/maybe not territory if it polled 54.5 was a fair read of the poll (in fact polling that would have seen it one seat short.)  The Australian were correct in reporting that the issue as shown by their poll was majority or not, and the election was by this stage clearly not close in terms of who looked most like forming government.

Seat Polls

Seat polling was relatively rare at this election.  Pride of place must go to Freshwater for pretty much nailing the results in Riverstone (54-46 to Labor, actual 53.7) and Pittwater (52-48 to Liberal vs IND, actual 50.66).  That said the AFR's reporting of basic details of these polls was again abysmal.  Redbridge might have joined Freshwater but was let down by implausible 2PP estimates.  They released 2PPs for Labor of 49% for Penrith and 54% for Parramatta and I posted that my estimates off their primaries were 51 and 57 respectively.  Labor duly won both seats with 51.6% and 58.6%.

Community Engagement claimed to have Independent Helen Conway as close as 49.3-50.7 in North Shore but underestimated the Liberal primary by over 7% leading to a 5% 2CP error.  They had the wrong winner in Pittwater but were only 1.4% off on the 2CP in that case.

Sky in a courageous attempt to be even more opaque than the AFR released unnamed industry group polling of various seats (later determined to be by an energy industry source that had it in for Matt Kean).  Amusingly this polling did in fact nail the 2PP in Kean's seat of Hornsby but it wrongly had Kean's primary down 16% (actually 11.6%) and One Nation at 15% (actually 7.9%).  The polling was also all over the shop in five other seats with two wrong winners and a 9.7% 2CP miss in North Shore, but it did provide useful data on Kiama.  Overall these nameless seat polls had an average 2CP error of 4.6% and underestimated the conservative candidates by 2.2%.

An IPA poll by Insighfully of major vs non-major votes in Bathurst, Dubbo, Orange and Upper Hunter was accurate for Orange but underestimated the Nationals massively by 18% in Bathurst and 25% in Dubbo and Labor by 7% in Dubbo and 13% in Upper Hunter.  Such are the dangers of not waiting til the candidates are declared (which happened a week later.)

2PP Post-Election Pendulum

Why do I do it?  I keep putting out 2PP pendulums because I am fed up with supposed pendulums that mix classic and non-classic seats and end up as just lists of seats that each side might lose to somebody, not seats that will fall based on 2PP swing.  And as we saw at this election, 2PP swing is alive and well as a predictor for now despite all the nonsense to the contrary. 

This 2PP pendulum shows classic seats by their 2PP margin, with any 2CP margins in brackets.  Non-classic seats are shown by 2CP margin and opponent, with the 2PP margin in brackets.  Seats that were close against independents might be at risk from independents again, but the example of Dubbo at this election shows how that sort of thing is unreliable.




Seats marked * are seats where a first-term MP is expected to pick up a personal vote bonus if they recontest.  Seats marked ** are "sophomore surge" seats where the first-term MP defeated the other side's incumbent and may therefore get an extra bonus.

What we see here is a pendulum where it's very difficult for the Coalition to get a lot back, because Labor only has six marginal seats and all of them are protected by personal vote effects.  Becoming the largest party again requires a uniform swing of 3.7% all else being equal, but probably in practice a bit more, so something around a 50-50 2PP or slightly better might do it (which seems fair enough).  But getting a majority back without recovering anything from the indies requires a uniform 8.87% swing back (54.6% 2PP) and even that might not be enough because even there there are personal vote advantages for Labor.  

The hazard for the Coalition in this pendulum is that if there isn't too much of a swing back to them, their abundance of marginals could mean that the luck they enjoyed in close seats at this election runs out.  It's even possible Labor could have a swing against them of 1-2% next time but win a majority.

All this said, we don't know for sure that these will be the margins next election as there's potential for population estimate errors to lead to an early redistribution.

All in all, NSW 2023 was a fascinating state election and there is plenty more that could be talked about.  Just as well perhaps, because it could be a while before the next one!

4 comments:

  1. If a 54-46 2PP doesn't produce a majority, then the concept is in some difficulty. Not yet useless, but not the reliable guide it used to be.

    ReplyDelete
    Replies
    1. Sure in NSW's case the days when winning 51-49 was highly likely to be a majority are gone because of increased crossbench seats. But a side that wins 54-46 still easily wins the election, and that Labor might well exceed 54 and still not get a majority was well known before the election. 2PP swing is still doing very well indeed as a predictor of seat swing.

      Delete
  2. This is a complication of opv..Labor tends to need preferences more than the coalition and opv chokes off green preferences to the alp and alp and Green preferences to the teal like candidates. This may possibly have changed the result in Kiama too

    ReplyDelete
  3. I only evaluate what people say before election day; that piece was published during the postcount. After some results are known someone can say their poll predicted a given seat outcome but unless they clearly said that before the election it wasn't on the record. More of an issue in cases where the claimed result actually did occur, rather than this one where it looked like having occurred and then didn't.

    ReplyDelete

The comment system is unreliable. If you cannot submit comments you can email me a comment (via email link in profile) - email must be entitled: Comment for publication, followed by the name of the article you wish to comment on. Comments are accepted in full or not at all. Comments will be published under the name the email is sent from unless an alias is clearly requested and stated. If you submit a comment which is not accepted within a few days you can also email me and I will check if it has been received.