Tuesday, November 19, 2024

Queensland 2024: Final Results And Poll Performance

Queensland: LNP 52 ALP 36 KAP 3 GREEN 1 IND 1

2PP Estimate 53.76% to LNP (+6.89% from 2020)

The 2024 Queensland election is over.  At one stage it looked like it could be a bloodbath, and it was far from close despite some close-looking returns from the day booths, but it still ended up being only a routinely medium-heavy defeat for a decade old federally dragged government.  Labor was criticised for running so far to the left in this election but they did the right thing by ensuring they would hold a mass of seats in Brisbane in an election they were never going to win anyway.  The same strategies that were effective in not merely stopping the Greens taking more Labor seats but in recovering a seat from them (very narrowly) were also effective in holding off the Liberals in most of Brisbane, cutting losses to three seats on the city's eastern fringe.  For the LNP, mission accomplished, with a workable majority but without the hazards of an overly large backbench.  For the ALP, relief.  It could have been a great deal worse.

Antony Green has published a very useful post about this election with an accompanying report.  In this he derives a 2PP estimate of 53.76% to the LNP and explains his working in excellent detail.  I am not in a position to construct a clearly better estimate but I did have a go at estimating the 2PP by a different method for the eight non-classic seats, by applying Antony's preference flows at the 3CP stage rather than to primary votes.  The advantage of this is it is no longer necessary to make assumptions regarding those votes (especially from independents) that did reach one of the major parties during the preference distribution.  A disadvantage is that 3CP-stage flows to the majors from third parties are in general slightly weaker than 2PP flows from those parties' primaries, in cases where both are known, so my method can create distortion.  In this case that is likely to cancel out as there are three seats with a left-breaking non-classic candidate (two with very strong Green flows to Labor) and five with a right-breaking non-classic candidate.  Anyway I did all this and I also got 53.76%.  

Based off the 2020 post-election pendulum, for my swing estimate of 6.89%, Labor should have lost 18 seats to the LNP (this does not include Ipswich West which had been lost in a by-election).  In fact they lost 17.  Below the waterline they saved Mansfield, Pine Rivers, Cairns, Aspley and amazingly Bundaberg (Labor's best 2PP swing in a seat they won in 2020 by nine votes!); above it they lost Rockhampton, Capalaba, Maryborough and Mulgrave.  In my conditional probability model, this slight overperformance was predicted even though there were theories out there that Labor would do worse than the pendulum prediction.  The main reason it was expected was that Labor went into the election with nine seats on margins of 5-6.9% but only three on 7-9%.  Because swings are variable, that means far more seats where (all else equal) Labor had a just under 50% chance of holding than seats where it had a just under 50% chance of losing.  And indeed Labor held four of those nine - but also lost a couple on around 5% above the swing.  

The election again supported the usual theories of personal votes.  In eight seats where a new Labor MP was completing their first term the average swing to LNP was 1.9 points below the state swing, although five of those eight MPs lost.  (In the two "sophomore surge" seats where the new Labor MP had replaced the LNP incumbent in 2020, Labor overperformed by 4.1 and 8.4 points; the latter the retention of Bundaberg by Tom Smith who is obviously doing something right.) In five seats where a Labor MP retired at the election the average swing to LNP was 4.9 points above the state swing, with all of these seats except Sandgate falling.  If including all eight seats where the Labor MPs elected in 2020 did not contest for whatever reason, the gap rises to 5.4, but that includes the unusual cases of Ipswich West (disrupted by the LNP winning it mid-term) and Inala (where the 2020 MP was Premier Palaszczuk).  Slightly offsetting this in the three seats where LNP incumbents retired, the swing to LNP was 0.6 points higher than the state average, but overall the evidence for personal vote effects was very strong.  

The Greens not only underperformed their polling but got swings in the wrong places and as a result missed their targets, apart from retaining Maiwar.  While the LNP's preference decision made it much harder for Amy MacMahon to hold onto South Brisbane if the LNP did remain third (which it did) the primary vote swing of 3.2% against the party in the seat compared poorly with an average of 4.25% for eight previous Greens who have defended single-member seats around the country after their first full term.  Of those eight cases only one previously had a negative swing (Ellen Sandell -2.6% Melbourne (state) 2018) and the other two below +3% featured left competition from Victorian Socialists and Keep Sydney Open.  South Brisbane is an agonising loss for the party because had they been able to keep just 53 more Greens voters from switching to Labor the seat would have been retained.  (It may seem odd that I am talking about swings to Labor when Labor lost primary vote share in the seat too, but I have no doubt Labor both lost votes to the LNP and took votes from the Greens.)  This is the first case in Australian history where a seat won by the Greens as such at the previous general election has been lost at the next.  

Katter's Australian Party looked a good chance to expand its presence beyond the federal Kennedy footprint on election night but its prospects of victory in both Mirani and Mulgrave were dashed by feeble performances on absent prepolls and election day absents. 

Antony has compiled a post-election pendulum in his piece.  Things will change with a redistribution but as usual when there is a change of government on a large swing the LNP will have personal vote benefits in most of its marginal seats, making what is currently an otherwise "fair" pendulum more challenging for Labor.   I add here some 3CP swing figures (which assume a swing from one party to the other with no change in the third):

* South Brisbane: Labor would fall to third, returning the seat to the Greens, on a 3CP swing from Labor to LNP of 0.15%.  

* Mulgrave: A 1.0% 3CP swing from Labor to KAP would have seen KAP make the top two and presumably win the seat.

* Maiwar: A 4.4% 3CP swing from Greens to Labor would result in Labor taking the seat.  

In Labor vs Greens seats the Greens now need 3CP swings of 4.6% in Greenslopes, 5.0% in McConnel and 5.9% in Cooper.  

OPV ... Watch What You Wish For!

A lot was said about optional preferential voting on the campaign trail.  The LNP might be relieved that once they pass OPV, in future elections they will be spared making how to vote card decisions between Labor and the Greens.  (Oh the poor petals, One Nation put out excellent double-sided HTVs so their voters could choose their own major-party adventure in 2020 and didn't whinge about it.)   Apart from that, this election gives every reason to suspect that OPV could bite them.  The LNP won three seats from behind on preferences (Cook where they trailed by 16 votes and Rockhampton and Maryborough where the gaps were more substantial).  

What tends to happen under OPV in NSW is that Greens voters fill out their ballot at far higher rates than voters for other parties, including One Nation.  In the context of Queensland, Labor only got 52.8% of all preferences to the LNP's 47.2% under CPV this year.  If you switch to OPV and the exhaust rate is higher among right-wing preference sources, that could mean that Labor gets a much higher share of all preferences to compensate for the lower volume of them.  It makes it hard for Labor to win inner city seats from way behind on Greens preferences, but they didn't actually do that at this election anyway.  

In seats where the LNP's victory was relatively narrow and relied on 2024's extra-strong preference flow from One Nation, there is a fair chance that the LNP would have lost under OPV; that includes the three they won from second and also Pumicestone.  Labor overtook the LNP on preferences in only Aspley and Pine Rivers, and probably would have done so in Pine Rivers under OPV anyway.  The Greens came from second in Maiwar on Labor preferences and would have done so easily under OPV.  Labor passed the Greens in South Brisbane on LNP preferences after barely avoiding exclusion; under OPV with a just-vote-1 how to vote card Labor would have avoided exclusion more easily but the result between Labor and the Greens could have gone either way.  It is highly likely the LNP would have done worse at this election had preferencing been optional.  

Anyway I have said my piece about the misguided and misleading push for OPV here and here and if the LNP insists on resurrecting it for entirely bad and sooky reasons then far be it from me to interfere further with such foolishness.  Especially when they don't yet have an Electoral Matters process for me to make these points to. If the LNP must restore OPV having rashly promised to do so, I hope they will consider restoring it in a form that encourages voters to at least consider giving preferences to make their votes more powerful, and that bans "just vote 1" material that resembles ECQ or AEC signage.  

Poll Accuracy

This was a mostly great election for polling!  There was a reasonable variety of it (unlike 2020) with five pollsters producing polls during or just before the campaign.  There was no evidence of herding and the spread of values was good.  Not everyone was accurate but the average across the polls was very accurate indeed, with the five "final polls" giving an average 2PP of either 53.5 (or arguably 53.4, see below) compared to the actual 53.76.  Final polls all overestimated the Greens slightly (by an average of 1.7 points) and mostly underestimated combined Others, and were very good on the majors.  

The only unfortunate aspect for poll-watchers was that the last three polls all happened to underestimate Labor's 2PP (in two cases slightly) producing a picture of late narrowing that was probably exaggerated.  Thus my time-weighted average of the polls had the LNP as low as 52.4, but not weighting by time produced a better estimate.   

There are a couple of quirks to mention with the final polls.  Firstly while I have recently usually limited the polls assessed to those taken during the campaign, Queensland campaigns are very short and I don't think I should omit Freshwater which was released on the eve of the formal campaign.  If anyone thinks it shouldn't be included, they are free to ignore what I say about it.  It might seem harsh to include a poll that was four weeks old by election day in an election where voting intention is thought to have changed late in the piece, but if polls and/or media clients want a clear reputation for accuracy then it is best secured by polling close to election day unless the pollster believes that polls are less accurate then.  That way we can avoid the situation where a poll taken well out from the election looks good if it is accurate, but if it isn't accurate there is the excuse that it was taken too far out.  I have not included the Redbridge and Wolf + Smith polls from August; those are too far back.  

Secondly there was a question about what was the 2PP in the YouGov poll.  The Courier-Mail reported it as 55-45 to LNP in both its print and online articles.  However, on the online article if one hovers over the 2PP graph one can see that the graph shows the LNP as actually at 54.5.  Moreover, on my measurement of both the graph in the online article and the different graph in the print version, the 2PP was indeed graphed at 54.5.  The YouGov website does not include any statement on the matter.  I did later find a tweet by Amir Daftari of YouGov when the poll was released that said "2PP rounds up to 55. Final results likely between 54 and 55."  Assuming that the 2PP rounded to 54.5 to half a point and 55.0 to a full point, the 2PP in YouGov's poll was somewhere between 54.5 and 54.75.

This sort of thing could happen in various ways, and I haven't asked - the most likely to me is the pollster supplied a 2PP rounded to half a point (as Newspoll does in its final polls) but the Courier-Mail wanted to round it to a whole number.  Given that YouGov was using previous-election preferences 54.5 seems more likely than 55 (indeed it could easily have been 54.0).  

Below I have used 55.0 in the table which has been most widely accepted as being the "published" 2PP for this poll but I also give the accuracy values for 54.5 in the text; I am included to believe that YouGov did intend 54.5 but got overruled by the innumerate Courier-Mail.  The difference doesn't affect my ranking order but it does affect bragging rights for closest to the pin on 2PP since 54.5 is very slightly closer than Resolve's 53.  However we have the information that the YouGov 2PP was probably between 54.5 and 54.75 which suggests that it is more than 50% likely that Resolve's unrounded figure was closer than YouGov's anyway.  The reader can form their own judgement on that matter (Resolve has also published a figure to one decimal place but I am unaware of it being published anywhere before 8 am on polling day).  

I am laying down the following standard for my articles for any future cases where the same poll has multiple co-equal 2PPs on the loose:

When a pollster, client or pollster/client combination publishes two or more different 2PPs resulting from the same method, it is not clear which is intended as the 2PP (ie not caused by obvious misprints) and the pollster does not clarify publicly via an official web presence which is the correct 2PP before 8 am election day without me having to ask them, I will use whichever value gives the worst accuracy scores and I will not publish any accuracy data for the other value. 

(Resolve published both a respondent preferences figure of 53 and a last-election figure of 52 but the respondent preferences figure was mentioned first in the primary report, alone in some secondary reports and was clearly the leading figure, so that is not the sort of thing I have in mind.)

Here's my table of final polls for this election. Values within 1% are shown in blue and closest to the pin in bold.  Misses by 3% or more are shown in red, irrespective of sample size (there was only one of these, and barely, on uComms' estimate of the Greens vote.) 


As usual, I use four different accuracy indicators but the one I am most interested in is AVE2, which is weighted 50% by average 2PP difference from the result and 50% by average difference on primary figures, noting the importance of the 2PP in election prediction and in reporting of the contest.  AVE is the average difference on primary figures and RMSQ and RMSQ2 are root mean square error figures for primary votes and primary/2PP combined respectively.  (RMSQ punishes one large difference on one party more than several small ones on different parties).  

For YouGov, if 54.5 is used as the 2PP then RMSQ2 drops to 1.41 and AVE2 drops to 1.19, so the order of the polls is the same.

Nobody absolutely nailed it but Resolve Strategic had the most accurate final poll of this election overall by my method, followed fairly closely by Newspoll, although Newspoll had slightly the more accurate primary numbers.  It came down to the 2PP and in this case Resolve's respondent preferences sample caught the preference shift accurately and outperformed previous-election preferences as used by Newspoll, with the two in effect assigning 53.5% and 58% of preferences to Labor in the published figures after rounding respectively (it was actually 52.8%).  Newspoll couldn't keep hitting the 2PP (or Voice equivalent) within one point forever (as it had remarkably done five times in a row) but it wasn't far off making it six.  

Did voting intention change?

YouGov and Freshwater, which went out of field earliest, overestimated the LNP's result at this election, while Resolve and Newspoll had the picture on primary votes basically correct and uComms which came last had it closer than it was.  Given this, did voting intention tighten?

It most likely did, since Newspoll (the only poll to release two polls entirely from September onwards) had a 2.5% shift from mid-September to the end.  That could have been random by itself, but there is more.  There is supporting evidence from the swing in on the day voting being about four points lower than the swing in votes cast before election day (whether prepolls or postals), and in the government performing much better and the LNP worse in leadership polls late in the piece whoever did them.  However there are other possible causes of the differences in swing.  One is that 2020 was a "COVID election" and there's a general pattern that left-wing voters were more concerned about catching COVID than righties and hence became unusually likely to vote by mail that year.  On that basis left voters were more likely to switch from voting pre-election to voting on the day.  

Another is simply that on the day voters are becoming a smaller and smaller proportion of all voters, and thus becoming less and less representative.  The ABC broadcast had several comments early in the night about the numbers being closer than the Newspoll - these comments were irresponsible and wrong, and the ABC should have learned from their wrong majority government call in NSW that you cannot reliably project the final swing primarily off the day booth swing.  2019-level failures excepted, the time to discuss polling accuracy is when all the votes are counted.  

If we average the earlier polls before and around the campaign, the first Newspoll plus YouGov plus Freshwater had an average 2PP of 55.2 or 55.3 while the later Resolve/Newspoll/uComms set averaged 52.2.  But about a point of this three-point difference came from preference flows.  All the first three happened to have relatively low preference flows to Labor in spite of at least two using flows from previous elections, but two of the last three had similar flows to 2020 in spite of one of those two (uComms) being respondent-preferences based.  That puts the polled shift from mid-late September to election day around 2 points, and given the possible house effects in some of these polls I suspect any real shift wasn't even that much, but that there was something.

I do think something meriting some thought is how polls ask about voting intention when they are polling late in the campaign.  Both uComms and Newspoll, according to their disclosure statements, asked a single question as to who respondents would vote for in a state election held today.  Is it possible a respondent who had voted early and then developed buyer's remorse (if it can happen that quickly!) might give an answer different to how they actually voted?  

Exit Poll Debacle

This article would not be complete without a major "told you so" to the Courier-Mail.  Serious pollsters have more or less given up on exit polling because capturing a representative sample of all voters in these days of a fortnight of prepolls is impractical.  However newspapers, especially the Murdoch tabloids, continue conducting and publishing small sample "exit polls" presumably conducted by their staff.  In this case, the Courier-Mail published two, one on the day and one during prepoll.  

The problem with these "exit polls" is that the Courier-Mail does not bother to ensure the data they collect are interpreted credibly but instead puts out silly spin that misrepresents what the numbers mean.  I am not sure if they do this because they lack the ability to understand the data when there are plenty of people out there who do understand it, or because they don't want to print evidence-based interpretations that would get in the way of hype and spin in which everything is either a  demolition or a nailbiter.  In the case of their on the day poll, the Courier-Mail published an unqualified statement that "The Queensland election result will come down to the wire" under the heading "[..] exit poll shockwaves: Labor surges, election too close to call".  The article grossly misinformed readers (if not flat-out lied) by saying the results were from "a cross-section of electorates today" when in fact every electorate polled was a Labor seat, on an average 2020 margin of 7.6% 2PP.  

I reinterpreted these numbers as pointing to a day vote 2PP that was about 5% worse for Labor than the 2020 overall 2PP for those seats, if correct.  I pointed out that on average that meant Labor would be facing a large swing and would lose decisively.  The exit poll was actually not too bad a sample of the day vote.  It had Labor trailing in these seats 33.6-33.9, and Labor actually won the day vote in the selected seats 36.1-34.7 (but did very badly on preference flows in several).  But as pointed out in my live coverage, if Labor were on for a day vote that was about 5 points worse on 2PP than the 2020 election then they were in for something worse than that as an overall total, even though they would hold some of those seats. In fact Labor did 6.5 points worse on day booth 2PP in these seats than in the 2020 election overall 2PP, and ended up saving three of the ten.  

The Courier-Mail thereby missed an opportunity to provide their readers with an excellent broad reading of how election night would later turn out, that could have been provided had they got anyone who understood election stats to comment.  I always find this disappointing because I want people to be informed about what is going on in elections, and that includes people who read the Courier-Mail.  

Learning absolutely nothing from this, the Advertiser in South Australia did it again a few weeks later!  It released a similar poll for the Black by-election where it found the result could "come down to a knife edge" off a sample of early voters that had Labor leading on primaries 47-37.  Anyone with a clue about Australian elections knows that if Labor is leading by 10% on primaries in a 2PP contest then that's a lopsided result, especially in early voting.  As it happens Labor is currently winning declaration votes (which included early voting) 50.4-34.7 (so the sample wasn't that far off given it was only 100 voters) and is winning the by-election in excess of 60-40 2PP.  The Advertiser therefore missed an opportunity to tell its readers the story its own data were telling - that Black was on the verge of delivering not a close one but something far more historically unusual and interesting.

The earlier Courier-Mail poll made different mistakes.  It polled a partly different list of ten Labor seats in early prepoll and on this basis announced that Labor was headed for a "devestating defeat" and the loss of "several senior ministers" including Meghan Scanlon and Mick de Brenni, both of whom retained (Scanlon narrowly).  The error here was treating prepoll as predictive of results overall when within-division prepoll often skews a little bit conservative compared to the overall voter pool, and when voting behaviour isn't necessarily even through the whole prepoll period.  (Early prepolls are likely to be more conservative than later prepolls.)  The poll found a 30-42 primary vote deficit for Labor but it was actually 34.7-40.9 in the seats canvassed in early voting alone, so it didn't even accurately measure early votes.  Both Scanlon and de Brenni actually did lose the prepoll 2PP but did so far more narrowly than the "poll" projected, and then both won their seats anyway.

Even these nonprofessional samples are of some use but the Murdoch tabloids' inability to interpret their own exit polls correctly raises the question of why bother.  If you're going to take the trouble to collect the data why not spend a few minutes getting someone who understands it to help you get the story right.  

So ends the 2024 Queensland election!  I always find Queensland elections fascinating because of the blend of electoral landscapes that make up the state.  This one has delivered a very middle of the road result in terms of what we already knew about federal drag and old state governments.  In four years' time we'll be back to see whether David Crisafulli's government can mount a better first term defence than Campbell Newman's did.  

Monday, November 18, 2024

EMRS: Is Labor Finally Making Some Progress? / Hobart Poll Controversy

EMRS Lib 35 (-1) ALP 31 (+4) Greens 14 (=) JLN 6 (-2) IND 11 (-1) others 3 (=)
IND likely overstated, others likely understated
Liberals would be the largest party but Labor would make seat gains 
Possible seat estimate in election "held now" off this poll LIB 14 ALP 12 Green 5 IND 4

The final EMRS Tasmanian voting intention poll of the year is out and it provides some evidence that the Labor Opposition might be taking baby steps on the road back to government at last.  Labor is up four points, albeit from a poor base.  The Liberal government is at its lowest vote since it got down to 34% in December 2017 (a reading that I doubt was accurate given their rapid recovery months later) and the major party gap is also as low as it has been since then.  This said, Labor still hasn't been above 32% since the last "pre-COVID" poll back in March 2020 and if the ALP is going to make a serious push for government, at some point in the term it will need to break out of the very low 30s.  This is a movement in the right direction for once; let's see if it continues.  The poll comes following a quarter dominated by the Spirits of Tasmania fiasco that led to the forced resignation of Deputy Premier Michael Ferguson from Cabinet.  I suppose that yet again, as with bad polls following the 2023 defections, the Government might say that in it could have been worse.  

Thursday, November 14, 2024

Poll Roundup: Labor Loses The Lead!

 

All things must pass.

2PP Aggregate (Last Election Preferences) 50.1-49.9 to Coalition (+0.4, first Coalition lead of term)
With One Nation adjustment 50.6-49.4 to Coalition
Election "held now" could leave Labor with about 70 seats

--

It finally happened!  After endless months of Labor's 2PP lead being painfully whittled towards zero, last weekend's Newspoll finally put the Coalition into the lead on my last-election 2PP aggregate.  It is true that I said they were eight weeks away from breaking Kevin Rudd's record for the longest aggregated 2PP lead since there was nothing much to aggregate, and they actually lasted nine.  So there is that, but it's not much consolation, and a lot of the other aggregates flipped two or three months back anyway.

Historically, it's no big deal, and perhaps not even a medium one.  Almost every government falls behind in polling at some stage in every term, except the first Hawke government which went to an election not long after the half-time siren.  Governments, albeit the other side's are better at it, have frequently recovered from being well behind and tend to poll badly in the last year before elections.  Labor has the benefit of a friendly pendulum from last time and the Coalition needs to win a lot of seats in some very different places to get another sniff at government in 2025.  

Tuesday, November 12, 2024

Rebecca White and Anne Urquhart To Run For House Of Reps

I have four (!) articles I'm aiming to write for this site in the next week or so but the first cab off the rank should be the one where my local knowledge is most relevant, that being today's news that Rebecca White and Anne Urquhart will be running for Labor in the federal House of Reps seats in Lyons and Braddon respectively, resigning their State and Senate seats to do so.  The candidate to take on Bridget Archer in Bass, Tasmania's other competitive House of Reps seat, has still not been announced.

Rumours about White and Urquhart running have been around for some time and frequently canvassed on the Fontcast podcast and at times in mainstream media, the White one steadily gathering pace despite having been denied by the candidate in a debate for the March state election.  The Urquhart rumour, together with one that Shane Broad might quit state parliament to run for the federal Braddon seat, has been on and off but it is now clearly on, with the Prime Minister's social media announcing both Urquhart and White as candidates during his visit to the state today.

Sunday, November 10, 2024

The Pressure Valve: Does The Defeat Of Same-Party State Governments Help Federal Governments?

It's 1992.  The unpopular Cain/Kirner Victorian Labor government has been sent packing.  In comes Jeff Kennett and some voters are soon alarmed by his New Right agenda.  Cue massive protests.  The Keating federal Labor government has been struggling in the polls but it springs to life soon after Kennett's win (though that was far from the only cause).  In the 1993 election Labor gets a 4.34% swing in Victoria and gains four seats.  Across Bass Strait, where a short-lived Labor government had been removed in early 1992, there's an even bigger swing that yields another three.  The three Tasmanian losses are the first signs on counting night that something has gone terribly wrong with John Hewson's unloseable election, and these seven seats picked up by Labor in these two Liberal states combined are the backbone of Keating's against-the-odds win.  

Victoria 1992 is the paradigm case for a theory that one might call the "pressure valve" theory of state elections, that there is a drag effect of state elections upon federal elections and that federal governments benefit if the voters let off steam by throwing out an unpopular state government of the same party instead of taking their anger with it out on the feds.  Better still if the new state government has started to frighten the horses.  I have talked a lot about "federal drag" on here, which refers to the fact that state governments do much worse at elections, all else being equal, when the same party is in power federally.  Age and federal drag are the two biggest killers of state governments and it is for this reason that the Miles Government was always likely to lose by about as much as it did.  But does it work the other way?