Monday, July 25, 2022

Jacquie Petrusma Resignation And Recount

Updates Aug 15: Today's the day - nominations close at noon; I expect we'll know the result today (parliament resumes tomorrow).  Updates will be posted in this section.

Update: Candidates contesting the recount are Enders and Young (Lib), Brumby (ALP), Cordover (Green) and Flannery (ungrouped).  The latter three have no chance whatsoever.  Having only two Liberals contesting should mean the count is much faster with a majority on first preferences for either Enders or Young likely (unless it's very close).  Even if it is very close it will not then take long to distribute the other candidates.

Update: That was quick, and a slightly surprising result too: Dean Young wins.  Young defeated Enders 51.1% to 46.5% with negligible numbers for the others.  That is a bullet dodged for the government which would not have been wanting Enders on board.  


Before I could even whip out a guide for the upcoming Pembroke by-election prompted by the resignation of Labor's Jo Siejka we have yet another resignation from the Tasmanian Parliament, this time from the lower house.  It has been announced that Jacquie Petrusma is resigning from her Franklin seat for personal and family reasons. I don't have to cover off on the usual speculation about ministries as Felix Ellis is the big winner, landing a collection of portfolios reported as including police, fire, emergency management, resources, skills, training and workforce growth.  That leaves who is contesting, the outcome of the recount and whether it will muck around with the sitting of parliament.

Petrusma is the fourth minister to leave the ministry this term (following Sarah Courtney, Peter Gutwein and Jane Howlett MLC) and the fourth Liberal MP to leave the parliament (following Courtney, Gutwein and technically Adam Brooks, who resigned his seat as soon as he was elected to it.)  Four is the most government MPs to resign from the House of Assembly in one term since the 1972-6 term, which saw six Labor resignations.  

Petrusma first came to Tasmanian electoral notice in the 2004 Tasmanian Senate race where, as a Family First candidate, she at times threatened to defeat Christine Milne under the now discredited Group Ticket Voting system.  She later joined the Liberal Party and defeated Tony Mulder in a within-party race for one Franklin seat in 2010, and was re-elected in 2014, 2018 and 2021.  In 2021 she conveniently topped the poll with over a quota in her own right.  

Hare-Clark recounts consist only of the votes that elected the departing member, so how many primary votes a candidate got in the recount and how close they came to being elected has no direct bearing on the result.  Except in highly unusual circumstances (and this is not one of them) these recounts always elect a member of the same party as the vacating member.  The contenders are therefore whoever contests out of former Huon Valley Mayor Bec Enders, newsagent and 2019 federal candidate Dean Young and Clarence City Councillor and podiatrist James Walker.  Walker was a late replacement in the 2021 Franklin campaign after original candidate Dean Ewington was found to have criticised the government's COVID management. 

At this stage Young has confirmed he intends to contest.  Comment re the others' decisions to contest or not contest will be added when known

Because Petrusma was elected on primary votes, the recount consists solely of her primary votes, and there is information from her surplus about where many of them will go.  What is known of the distribution is as follows:

38.12% 1 Petrusma 2 Street
22.89% 1 Petrusma 2 Enders
18.73% 1 Petrusma 2 Young
13.74% 1 Petrusma 2 Walker
6.52% 1 Petrusma 2 for various non-Liberal candidates

The votes that are 1 Petrusma 2 Street will be thrown to the next available candidate, in most cases being a Liberal.  All the non-Liberal candidates contesting the recount will poll trivial primary vote tallies and be quickly eliminated.  If all the Liberals contest then the one in third place (most likely Walker) will be excluded and their preferences distributed.

Assuming the final two are Enders and Young, Enders starts with a 4.16% lead over Young with 58.38% to throw, meaning he needs something like a 54-46 split on the remaining votes after accounting for exhaust.  There isn't any particular reason to think he would get this so I expect that if Enders does run then she probably wins (but absent of very detailed scrutineering, we'll need to see the recount to be sure as surprises can happen).  Likewise, if Enders doesn't run Young has a lead of nearly 5% and would seem likely to win.  

Recounts can play havoc with the resumption of Parliament and I am awaiting news on what will happen here.  Premier Rockliff has been reported as saying he didn't think the resumption would be affected, but I can't see how this is so (unless Petrusma delays her resignation til late August).  In 2019 the Electoral Act was amended to increase the time for candidate consents to be returned from ten to fourteen days (on account of slow postal services).  This means that even if the notice is published on Tuesday 26 July, the recount could not be held before Tuesday 9 August, the scheduled day for resumption.  The recount takes all day.  The government has so far been reluctant to resume Parliament in similar circumstances lest the non-government MPs use numbers on the floor to embarrass it.   I suspect we will see a delay.  

Bec Enders might not be the most popular MP in the Huon if elected, and would be a target for opposition and crossbench MPs.  Enders was elected Mayor of the Huon Valley Council in 2018 and was initially a very popular Mayor as she set about cleaning up what had been one of the state's more shambolic councils.  However, from late 2021 the council became involved in a new scandal after it was revealed the director of the consultancy firm involved in the recruitment process for a new General Manager was in fact the partner of the person eventually appointed.  The council itself was cleared of any wrongdoing.  Enders resigned as Mayor and from the council in March 2022, at which time she and three other councillors were facing a code of conduct complaint over the issue.  (The complaint against Enders specifically was never ruled on because of her resignation.)

More comments will be added as information comes to hand.  Depending on the timing of the recount, I may or may not be able to post comments on it at the time it is counted.  

(Note: re Pembroke, I expect that to be in September to November, maybe earlier in the piece rather than later.)

Update: Recount set for August 15, so that's one week of scheduled parliamentary sittings under a cloud.   I should be available to cover it.

Update (2 August): As expected parliament will be prorogued for a week (without affecting the number of sitting days). 

Update (3 August): It's been reported today that Bec Enders and Dean Young are contesting the recount while James Walker is not.

Sunday, July 24, 2022

Federal Election 2022: Pollster Performance Review

In 2019 the Australian polling industry had a disaster after decades of reliability - every final poll had the wrong two-party winner, the final polls were bizarrely clustered around the same wrong result, many final polls were individually wrong by more than their claimed margins and the lack of transparency in the industry was such that it was difficult to understand just why it had happened.

Fortunately 2022 has not been a repeat.  The fallout from 2019 saw a great increase in polling transparency, especially via the formation of the Australian Polling Council (though unfortunately not all pollsters have been on board with that) and also more diversity in polling approaches.  No one poll has ended up nailing the remarkable results of this year's election, but collectively, federal polling has bounced back and done well.   This is especially so on two-party-preferred results, where a simple average of the 2PP figures released in the final polls is pretty much a bullseye. The primary vote results were a little less impressive.

Here I discuss polling in several categories.  Overall YouGov (which does Newspoll) made the most useful contributions to forecasting the result, Redbridge's performance in publicly known niche polling during the campaign was very good, and Resolve Strategic's final poll was a useful counterpoint to Newspoll.  The other major polls were so-so on the whole, and many minor pollsters were wildly inaccurate.

Final Polls

Pollsters are judged a lot by final polls, but final polls are only a single poll (so there's a degree of luck involved as to whether it's a good one), and final polls are taken at the time when there is the most data around for any pollster who might be tempted to herd their results in some way to base such herding on.  However, the final poll stage is the only stage where, at least in theory, there's an objective reality to measure the poll against.  A poll taken a few days out from election day should have a good handle on how people will vote, even more so now that so many people have already voted by that point.  Any attempt to determine the accuracy of polls taken well before election day requires much more debatable assumptions.

It's also important to note that often which poll comes out as the most successful depends on how you choose to measure it.  A few comments going into the following table:

1. In previous years I have used root mean square error (RMSQ) as my measure of the accuracy of final polls.  However, I have found that absolute error is almost universally used in other sources' assessments, possibly because the maths is easier.  The difference is that RMSQ punishes a small number of larger errors more than a larger number of small errors.  In an absolute error estimate, being out on four parties' votes by 1% each is the same as getting two dead right and getting two wrong by 2%, but RMSQ says that the former is better.  For this year's tables I give both figures; I think there are arguments for either.  

2. 2PP remains the main way in which Australian elections can be and should be best predicted overall (for now), and while 2PP is partly a modelling exercise as well as a polling exercise, it is a service that most pollsters provide and that those that do not provide should (as it is important for interpretation).  I consider 2PP estimates to be a very important part of Australian polling accuracy, so as well as giving one set of figures for primary vote error, I also give (and prefer) another in which primary vote error and 2PP error are weighted equally.  

3. At this election Essential did not publish net figures with undecided voters removed.  This means that rather than providing an estimate of how voters would vote at the time the poll was taken, the poll was providing an estimate of how 93% of voters would vote while providing no estimate for the remaining 7%.  To compare the accuracy of this poll with other polls it's necessary to convert it by redistributing the undecided, which I've done in the standard way.

4. I've included the YouGov MRP figures in the table for comparison, although it was not a final poll, but a difficulty here is that it published seat by seat projections but not totals.  The primary figures given were not published and were obtained after the election but are virtually identical (three differences of 0.1%) to figures calculated by Ethan of Armarium Interreta from the published primaries before the election.  Had YouGov published totals it may well have rounded them to whole numbers.  In any case as it didn't publish any totals I have avoided directly including it in the main list of final polls; likewise for Resolve I have included only the rounded totals that were published and not totals to one decimal place that were published post-election.  

5. I've also included some irregular polling attempts by ANUpoll, KORE and the Australia Institute Dynata "exit poll", showing how badly these compared with the regular pollsters.  2PPs were not published for any of these but all would have seen massive 2PP wins for Labor, which didn't occur.  Normally I don't include exit polls unless they are released within an hour or so of polls closing, but in this case I've made an exception, as what on earth is the use of drawing conclusions about how people voted (as The Australia Institute did) from an exit poll sample that was obviously garbage.  (There was also an ANUpoll post-exit poll that was even less representative than the ANUpoll in the table, which didn't stop all manner of conclusions being drawn from it).  I wouldn't have included KORE at all given the age of the poll except that they said that "the vote is largely settled".

6. Some pollsters published multiple 2PP estimates for their final poll.  For Resolve I have used their respondent preferences estimate (51) not their last-election estimate (52) as the FAQ published by the client made it very clear that the former was to be regarded as the primary method and there was no indication otherwise.  Ipsos published two different last-election estimates (53 and 54) and a respondent estimate (49-40 = 55-45) while making it clear 53 was to be preferred.  For Morgan, they clearly switched to using a last-election estimate (53) rather than a respondent preferences estimate (56.5) just prior to the election. This appeared to be a case of method-herding to avoid a big error as their stated rationale for doing so ("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.") made the opposite of any sense at all.  Surely if respondent preferencing was ever going to work and last-election preferences fail, that would happen closer to the election not further out from it.  

7. Some assessments of polling error single out errors that were outside the pollster's claimed margin of error.  This, comparatively, rewards polls with large claimed margins of error (eg those with smaller sample sizes or heavier weighting), but I think what is most important is how accurate a poll was, not how accurate it was compared to how accurate a (often primitive) margin of error model claimed it to be.  Therefore I haven't done this.

Here's the table, sorted by a 50-50 weighting of average absolute primary vote error and absolute 2PP error (AVE2).  Figures within 1% are shown in blue, figures 3% out or more are shown in red, and readings that were closest to the mark for the five final polls are shown in bold.  The "pref" figure is the implied share of preferences going to Labor after accounting for contests with two Coalition candidates. The four error figures are shown at the right (in all cases the lower the score the better).  As noted above I consider the AVE2 and RMSQ2 figures the best indicators since they include the 2PP. Ind/Others figures shown in italics are for polls that included ON and/or UAP under this heading.

While Newspoll has narrowly topped the table on the indicator used for sorting, Resolve has outperformed it (in two cases slightly) on all the other three (albeit two of which I regard as less important), so a case can be made for either as the best final poll as published prior to my cutoff time of 8am on polling day.  

The case for Newspoll is that it was closer on published 2PP (noting that Resolve also published a last-election estimate of 52), and only very slightly less accurate on absolute error on primary vote.  The case for Resolve is that its primary vote estimates were slightly more accurate overall and in particular it avoided Newspoll's significant 3.4% error on the ALP primary vote (which is picked up by the RMSQ columns).  Resolve also got the gap between the major party primaries right (an indicator far more relevant to overseas elections under first past the post, but with some predictive value in Australia since major party primary shifts are the main cause of 2PP changes).  Both these polls clearly outperformed the other final polls, but neither was as accurate as the best polls in 2013 or 2016, when each would have been about midfield on the table.  However, this was a more complicated election to poll, and polling has got harder since 2013 at least.

Resolve's post-election-published primaries to one decimal place would have made it clearly the most accurate of the final polls if published before election day.  That said, perhaps if Newspoll's primaries to one decimal place were also published, this wouldn't still be true.  

Morgan would have been a contender had it not had a stubbornly low UAP vote of 1% and continually overestimated the vote for Independents.  I believe the latter came from having Independent on the readout everywhere.  When Resolve started offering voters only the options available in their division, its previous problem with overestimating the IND vote went away.

Essential's result was not great.  Because of the way Essential handled the undecided vote, one might say that the undecided voters all voted for Greens and teals so the poll was closer than it looked, but it still had the Labor primary vote too high even if no undecided voters voted Labor.  (Plus undecided voters don't usually vote Green all that much).  Essential had systematic issues with underestimating the vote for Independents and Others.

All the final polls were outperformed by the YouGov MRP on both indicators that include the 2PP.  This applies irrespective of what set of figures are used for the MRP, including those rounded to whole numbers.

Overall, the five final polls collectively more or less nailed the 2PP as a result of two moderate errors cancelling out: they overestimated the primaries for either Labor (everyone except Resolve) or the Greens (Resolve) but they also - irrespective of the method used for the headline 2PP - underestimated the 2PP preference flow to Labor.  

A notable feature of the 2022 polling was the age of the final polls.  Based on field dates the final poll data had an average age of 6.4 days as of election day, compared to 3.2 days in 2016 and 4.8 days in 2019.  The trajectory of polling as election day approached is consistent with there having been some late swing against Labor.  The pattern of the swing to Labor being higher in votes cast before the day is also consistent with late swing, but it can also be explained at least partly by more voters using postal voting and hence the postal voter pool becoming less conservative.  Late swing might therefore be a genuine factor in these not-so-final final polls overestimating Labor's primary, but I doubt it's the full story, and it needs more investigation.

Winners: YouGov/Newspoll and Resolve

2PP Tracking

As noted, final polls are not everything.  It is also useful to try to get a handle on whether polls accurately portrayed the state of voting intention in the leadup to the election or whether a poll that had either skewed to one side or bounced around like a frog in a sock got lucky, or did something different, with its final reading.  And in this case, some polls did do something different:  Morgan switched its 2PP method, Resolve started publishing 2PPs and Resolve also beefed up its final poll with a phone component (having published in advance it would do so).  Also, some polls altered their methods as candidate lists became available.  Who most likely had an accurate handle on what was going on in the several weeks leading up to the election?

This is a challenging question to answer because it is very hard to answer it without subjecting a poll to some sort of judgement by a jury of its peers and/or itself.  If one poll consistently told a story that other polls did not agree with, and that story ended up being the result, then was that poll right all along with the others herding to it at the end, or was that poll wrong?  Also, a poll moving around is a good thing if (as in 2013) voting intention is clearly moving around during the campaign.  

I suspect you'll find more sophisticated analysis of the tracking question elsewhere (looking at where the dots come out on Mark The Ballot's aggregation graphs is often interesting.)  For my own attempt, I looked at released 2PPs in polls since the start of March and took as a starting point that the average of the final week 2PPs more or less nailed it.  I found provisional house effects for the individual pollsters in the leadup based on the assumption that on average the polls had been right when taken, but split Morgan's 2PP polls into those that used respondent preferences as a headline and those that used last-election preferences. I then "corrected" each individual poll for these apparent house effects.  I found that the scarcity of polls in March meant that the first usable week of data for tracking comparisons was that through to 2 April.  

For each poll I compared the released 2PP (or derived for those polls with no released 2PP) to the average of the "corrected" polls by other pollsters released in that week.  On this basis I estimated an average lean for each poll relative to what other polls found to be going on at the time (as adjusted for their house effects) and I also found the standard deviation (the lower the better) for the errors implied by this method.

The following were the results:

This is a rubbery method but its results are very consistent with my impression of the polls during the election leadup.  Morgan and Ipsos produced at times implausibly strong readings for Labor and Morgan was also too volatile.  Essential generally seemed asleep even when other polls suggested the Coalition was having an especially bad time of it (this is probably because of its use of party ID as a weighting).  This leaves Newspoll and Resolve, neither of which had any skew to speak of but Newspoll was generally steady while Resolve tended to swing from disastrous to fairly benign results for the Coalition.  Only one Newspoll in early April was flagged by this method as any sort of outlier, but this may well have been because the Morgan and Resolve polls in the same week were laying it on a bit thick for Labor (even in the former's case by Morgan's standards).  When I expand the comparison to include the previous week's polls, Newspoll's SD drops further to 0.7 while Resolve's increases to 1.7.  All up I consider that Newspoll had the best 2PP tracking.

I have not tried to analyse primary vote tracking but I think that various pollsters had persistent skews there, eg Newspoll overestimating Labor, Resolve overestimating first Independents then the Greens, Morgan overestimating independents, and Essential overestimating majors and underestimating independents/others.  I should add that while Essential has scrubbed up fairly poorly from an accuracy perspective, by doing something different it did provide a cautionary note against landslide scenarios seen in many other polls.  It is better to do something different even if it's not a very successful experiment than to herd.  

Winner: YouGov/Newspoll

Individual Seat Polling

Seat polling has been under the pump in Australia for a long time.  In 2013, seat polls skewed severely to the Coalition.  In 2016 they skewed to the Coalition to a lesser degree and were also under-dispersed (less variable compared to the previous election than they should have been).  In 2019 they skewed to Labor, though YouGov had a high strike rate in seats where it picked a winner.  They were also poor at some high profile by-elections, such as Longman and Wentworth in 2018. 

At this election there was not a great amount of seat polling seen, especially with YouGov largely switching to its MRP model instead, covered below (the MRP was a nationwide poll, but the results for specific seats are a poll-based model influenced by other seats, and shouldn't be treated as polls of that seat).  What was seen was often commissioned polling, reported in insufficient detail.  Multiple seat polls were seen during the campaign proper by three pollsters: Redbridge (mostly of teal seats for Climate 200), uComms, and Utting Research.  There was also a series of 2PP results released via media for something calling itself the Industry Association, but it was never to my knowledge determined who the pollster was.

For many of the seat polls released there was no released 2PP or 2CP figure so I have had to estimate my own off the (sometimes incomplete) primary votes.  In some cases the poll itself contained enough data for a respondent preferences estimate, in others I have used estimated preference flows.  

In the following table I give the number and percentage of polls for which each pollster or source had the right party winning, and an "ease" figure to indicate whether the seats should have been easy to get right given the eventual margin.  The "ease" figure is the theoretical average strike rate given a poll with an average error of 4%, but it ignores sources of error such as having the wrong party in the 2CP (which happened in one poll in the sample.)  The skew figure is the average extent to which the poll skewed to the non-Coalition side (minus equals skew to Coalition) and the error figure is the average raw error.  Where multiple results were recorded by the same pollster, only the last is included.

Congratulations are here due to Redbridge because while their polls taken just prior to the campaign did not scrub up so well (including pointing to Labor's potential demise in Greenway, which Labor won 61.5-38.5) the five results that surfaced during the campaign (Wentworth, North Sydney, Goldstein, Parramatta and incomplete details for Kooyong) all had the right winner and were also outstanding on the implied 2PP margins.  The Industry Association results of unknown polling source also had a perfect score but were more erratic on average (though still pretty good by seat poll standards).  uComms was erratic and skewed to the left, while the final batch of Utting polls showed a shift back to Coalition that proved to be entirely an illusion.  (An earlier batch had been more accurate.)  That said Utting Research did accurately predict Kate Chaney's narrow win in Curtin.

Once offs during the campaign include Compass Polling of North Sydney (wrong winner and way off except for nailing that TNL's Victor Kline would receive 0.8%), Community Engagement of North Sydney (an early poll and hard to tell who would have won based on it), an escaped YouGov of Pearce (right winner, well off on the margin), an early-campaign Laidlaw poll of Fowler (wrong winner but at least showing Dai Le had potential).  There was also a brace of incomplete 2CPs given to Peter van Onselen that proved to be internal polls from long before they were taken; all four had the right winner but these overestimated the Coalition by 2.4% on average.  And there were all the usual rumours about vague results from internal polling that I have not covered here.  

Winner: Redbridge

Pollster Forecast Models

The totals for the YouGov MRP have been discussed above.  The MRP (which predicted 80 seats to Labor, 63 to the Coalition and 8 for the crossbench) was especially interesting as an attempt to improve on the problems of seat polling by using small samples in each seat but then scaling each sample to the results of seats that were comparable, hence giving each seat a larger effective sample size in an attempt to smooth out the bumps.

In the seats won by major parties, the MRP was very successful, predicting 131 of 135 seats correctly (the only incorrect predictions being Lindsay, Bass, Tangney and Hasluck, two on each side).  In seats which the MRP predicted to finish as two-party contests, the average 2PP skew was 0.3% to Coalition and the average raw error was 3.3%.  There was, however, a significant 2PP skew in WA where the MRP underestimated the Labor charge by 4.7%. 

It has, over time, been very difficult for any model to reliably do better than getting about 10 marginal seats wrong so (allowing for a few marginals becoming non-classics) to only get four wrong is an exceptional result.  On the other hand the MRP did less well with non-classic seats, missing the three Green gains in Queensland (although projecting the LNP would lose Brisbane) and missing four of the six teal gains and Dai Le's win in Fowler.  The latter was down to seat-specific factors hence not predictable by the MRP's methods.  In the case of Brisbane it is not clear if the error lay with the poll or the interpretation of the numbers in terms of the Greens getting from third into second on minor party preferences. 

Overall the MRP's projection of 80 seats to Labor also had benefits because although pretty much every poll suggested Labor was either on track to a majority or capable of winning a majority, commentators at this election were exceptionally prone to make stuff up about what the polling meant if true.  To have a major pollster put out an actual forecast of the results, not just a set of numbers that mass media commentators would then misinterpret or simply ignore while churning out a preconceived narrative, had its uses.  

The only other loosely poll-based projection forecast model by a pollster I'm aware of was by KORE, which did have a predictive success in that its "Effective Vote" table was very close to the actual seat tally, at 81-54-16 (Labor-Coalition-Other) compared to the result of 77-58-16.  The major party seat results being anywhere near right here was, however, a result of two very large errors cancelling out.  The actual "effective vote" (a tally of 3CP votes recorded by each party and others) was 45.6 for each major party and 9.2 for others, nothing like KORE's 54-35.7-10.3 to Labor which would have resulted in an enormous seatslide.  But also the relationship between "effective vote" and seat share is nowhere near proportional - a small advantage could convert into a big seat margin.  In this case Labor got a big seat margin with no effective vote advantage, largely because the Coalition racked up useless 3CP votes in 13 mostly close seats that it lost while Labor only made the 3CP while losing in three, two of them lopsided.

Winner: YouGov (essentially unopposed but the MRP was good)

Senate Polling

Senate polling has a history of being even worse than seat polling.  There were not many attempts this year and those that there were were mostly showing their age by election day, so I consider everything from the start of March.

The Australia Institute issued Dynata polling for every State except Tasmania.  The four-sample poll underestimated the Coalition in every State by an average of 3.5% and overestimated Labor in every state by 6.0%, equivalent to a nearly 5% 2PP error and hence wildly inaccurate. It was fairly accurate on the Greens vote (0.6% under), overestimated One Nation in every state except Queensland, but by less than in 2019 (1.2% over), overestimated UAP slightly (0.6% over), underestimated Liberal Democrats in every state (1.4% over), overestimated Nick Xenophon in SA (3% over) while getting Rex Patrick right, and underestimated parties not listed in every state and by an average of 3.6%.  This wasn't an age of data thing as the final sample had even greater errors on Labor and unlisted parties.

Results were seen of two Redbridge polls for the ACT Senate.  The last of these correctly predicted David Pocock's victory, getting the votes for the Liberals, Pocock and the Greens almost exactly right but having Labor 6.3 points too low, Kim Rubenstein 1.6 too high and UAP 3.8 too high (overestimating the UAP vote is a common feature of Redbridge polls).  On these numbers Pocock would have won less comfortably as preferences that flowed to him strongly in the election would have split between him and Labor, but he still would have won.  A Community Engagement poll from late March had Pocock much too low, as did Redbridge at the same time, but it's likely Pocock built up steam as the campaign went on, so these polls may have been accurate when taken.

A Lonergan Senate poll of SA, of unknown commissioning source and dates, was incompletely reported by the Advertiser on 18 May.  The results as reported (with Labor on 34 and Liberals on 23, and it being unclear what was done with undecided) proved way off as it was the Liberals who ended up with 34 and Labor on 32.3.  Finally a uComms of Braddon, Tasmania in March overestimated Labor by 8%, and underestimated the Liberals by 3.6% and parties not listed in the readout by about 5%.

Winner:  It's a low, low bar for this one but Redbridge again

I intend in a separate article some time later to review the major polling story of this cycle, being the formation of the Australian Polling Council and the increased disclosure efforts by many pollsters, and to review how the attempt to improve transparency in polling panned out and where it can go from here.   

Friday, July 22, 2022

2022 House Of Reps Figures Finalised

Yesterday the 2022 House of Representatives figures were added to the archive of election results, making lots of the usual preference flow goodies available. Although all the preference throws had been completed and uploaded in rough form some time ago, the final figures importantly include the two-party preference flows by party and two-candidate preference flows by party per seat.  As well as this piece I will also be putting out a full analysis of polling accuracy, I expect within the next few days.

Some of the ground that I normally cover in this article was already covered in Two Party Swing Decided This Election (Plus Pendulum).  That article showed that Labor won the election on normal two-party swing in classic Labor vs Coalition seat contests, with changes in the seat share for the major parties pretty much exactly matching historic patterns, and that the groundbreaking defeats for the Coalition at the hands of six new teal independents and two Greens were nonetheless a sideshow in terms of explaining how the election was won.  

The article also noted:

* that Labor would have won the election under any system (contrary to the nonsense of the "3 in 10 voters" Sky right disinfo crowd who are wrongly claiming the Coalition would have won under first past the post - a different system would have seen different voter behaviour)

* that the view that Labor's primary vote was greatly damaged by strategic voting for teal independents is incorrect

* that Labor's win was assisted by gaining higher swings in Coalition seats on any margin and in marginal Labor seats, while in very safe Labor seats the swing was weaker to zero

Preference Shifting and Card Impacts

The official 2PP is 52.13% to Labor and 47.87% to the Coalition, a 3.66% swing to Labor. (There was an unlikely 2PP flow in a booth in North Sydney which if changed would have made the 2PP 52.14, but no changes have been made there; I do not know if the 2PP flow was rechecked.)

The 2019 election saw an unusual if modest preference shift in the Coalition's favour - the largest to the Coalition since the 1950s - but Labor still received more preferences than the Coalition in that year.  This election largely reversed the 2019 shift, with Labor's 2PP coming in 0.99 points higher than would have been expected based on the primary votes and the 2019 flows by party from Greens, UAP, One Nation, independent and others.  2022 therefore follows 1990, 2013 and 2019 as uncommon examples of such a preference shift in the last 40 years.  It will be interesting to see if 2019 was an anomaly or if 2025 sees a shift back to the Coalition again.  

The change in flow has happened across the board rather than a sharp shift from any one source:

* Greens 85.66% to Labor (+3.45% and a record high)

* One Nation 35.70% to Labor (+0.92)

* United Australia 38.14% to Labor (+3.28)

* Independent 63.77% to Labor (+4.37)

* all others 45.33% to Labor (+0.63)

The 2PP loss on three-cornered contests from the Coalition side was 0.03%.  

So for polls that break out all of these, the formula for Labor's 2PP by 2022 election preferences will be:

2PP = Labor + .8566*Green +.3570*ON +.3814*UAP +.6377*IND +.4533*Others + 0.03

The following are flows to Labor for some combined categories pollsters may employ:

* Others including IND 54.73

* Others including UAP 42.11

* Others including IND and UAP 50.01

* Others including ON and UAP 39.86

* Others including IND, ON and UAP (all non-Greens) 46.37

* Green+IND+ON+UAP+Others (all non-majors) 61.54

After splitting to the Coalition over Labor in every classic seat in 2019, One Nation preferences did the same again, except in Gorton where Brendan O'Connor got a 53.71% share.  The strongest split from One Nation to Coalition was 81.91% to Darren Chester (Gippsland).   One Nation voters also preferred Rebekah Sharkie to the Coalition in Mayo, and preferred independents to the Coalition in Calare, Groom, Wentworth, Wannon and Indi.  

I have records of five seats in which preferences were distributed between the major parties and One Nation recommended a preference to Labor on its how to vote card (Bass, Cook, Franklin, Lyons, Sturt).  In these the average flow of One Nation preferences to Labor was 7.8% higher than average, suggesting that the cards had not much impact.  

United Australia voters preferred Labor to the Coalition in ten classic seats (up from four) with Spence (59.65 to Labor) heading the list; the rest were Greenway, Forrest, Blaxland, Scullin, Canning, Macarthur, Swan, Ballarat and Jagajaga.  The strongest UAP to Coalition flow was 81.52% again to that popular MP, Darren Chester in Gippsland.  UAP voters also preferred Rebekha Sharkie in Mayo to the Coalition, and preferred independents to the Coalition in North Sydney, Cowper, Indi and Groom.

I have records of four seats outside WA that finished as 2PPs where UAP recommended a preference to Labor (Banks, Cook, Dickson, Maranoa).  In these the average flow of UAP preferences to Labor was a mere 4.9% stronger than overall. For Western Australia I saw but did not confirm reports that the UAP pursued an anti-incumbent strategy on their cards.  The preference flows suggest they indeed did this with a 65.0% flow to Coalition in Labor seats but a 53.7% flow to Coalition in Coalition seats (a modest difference of 11.3%).  Again, the evidence is that minor party how to vote cards don't make a lot of difference.  

The weakest 2PP flow involving the Greens was 71.95% to Labor in Parkes.  There were weaker 2CP splits off the Greens in three non-classic seats - Kennedy 64.21 to Katter vs LNP, Clark (where the Greens issued an open card) 66.52 to Wilkie vs Labor and Fowler 64.47 to Labor vs Dai Le.

Flows from the Coalition to Labor over Greens in the Labor vs Greens 2CP seats ranged from 58.65 (Canberra) to 73.33 (Wills) with an average of 65.6%.

In the semi-optional-preferencing Senate, there were far greater shifts in the 2PP preferencing behaviour of UAP voters (shift from Coalition to exhaust) in particular, and overall Labor won the Senate 2PP 52.93-47.07, 0.8% higher than its win in the Reps despite optional preferencing systems having a reputation for not helping the trailing party.  I will probably explore this in more detail but any idea that the Coalition would have done much better under optional preferential voting (which JSCEM recommended but the previous government didn't pursue) is delusional. It's possible they would have lost more heavily in seat terms as UAP voters especially would have exhausted their preferences rather than reluctantly preferencing the Coalition (a la Queensland 2015, where OPV saw a huge preferencing shift in Labor's favour against a disliked LNP government.) 

Non-Classic Seats

There were 27 non-classic seats at this election (up from 15).  These are seats where the final pairing wasn't Coalition vs Labor:

ALP vs Green (6): Grayndler, Cooper, Wills, Canberra*, Melbourne, Sydney*

Coalition vs Green (3): Brisbane*, Griffith*, Ryan*

Labor vs IND (2): Clark, Fowler*

Coalition vs IND (14): Indi, Warringah, Wentworth, Kooyong, North Sydney*, Mackellar*, Curtin*, Goldstein*, Calare*, Groom*, Nicholls*, Bradfield*, Cowper, Wannon*

Coalition vs Centre Alliance (1): Mayo

Coalition vs KAP (1): Kennedy

Those marked * were not non-classics last time.  Three seats moved from non-classic to classic status (Maranoa, Farrer, New England).  Melbourne and Kooyong both shifted from one non-classic status to another while Cowper had a different independent in the 2CP to 2019.

Labor 2PP winners failed to finish in the top two in Brisbane, Griffith, Ryan and Mayo.  In the first three of these Labor were also probably the Condorcet winners (the candidate, if there is one, who would win head-to-head against any other candidate), based on the Coalition to Labor flows recorded elsewhere.  I may discuss this in more detail sometime too.

As explained way back in 2013 I like to explore Labor-vs-Coalition 2PP splits for those voters who preferred the "non-classic" candidate to the majors, as this reveals what sort of voters a potential crossbencher might be beholden to.  Sometimes this can be done exactly for the voters who put the non-classic candidate first, rather than just for those who put them above the majors.  I understand from Antony Green that this year the AEC was in a position to extract exact splits for every seat, which would save me a lot of work, but I've not yet seen those published.  Here's a table showing 2PP preference flows from the non-classic contender to Labor in the non-classic seats:

The "To ALP" column shows the percentage of the 3CP voters for the non-classic contender that put Labor ahead of the Coalition.  In some cases the figure for the primary votes for the non-classic contender is also available, and this is shown in brackets. The %primary figure shows how much of the non-classic contender's 3CP vote is their primary vote (the higher this is, the more accurate the "To ALP" figure is likely to be as an estimate of the 2PP split of their primary votes.)

Overall this table shows that those who voted for independents, or preferenced them at 3CP level, generally strongly preferred Labor to the Coalition.  The exceptions were Groom (where most of Suzie Holt's 3CP votes were primary votes for other candidates), Fowler (where Dai Le's campaign attracted mostly voters who preferred the Liberals in this instance, though possibly in disgust with Labor's preselection or by following a Dai Le how to vote card) and Nicholls.  In most cases, teal independent voters or preferencers preferred Labor to the Coalition about 70-30, which isn't surprising.  Rebekha Sharkie's voters were more Coalition-friendly than last time, but this probably reflects her taking more votes from a struggling Liberal candidate.

Strongest preference flows

The following are the strongest preference flows I could find evidence of, whether those preferences were distributed or not. Estimated flows shown in bold:

93.8 Tim Hollo (Green) to Alicia Payne (Labor), Canberra
93.2 Adam Bandt (Green) to Keir Paterson (Labor), Melbourne (NB Bandt was elected)
93.1 Sarah Jefford (Green) to Peter Khalil (Labor), Wills
92.9 Liz Chase (Green) to Kate Thwaites (Labor), Jagajaga
92.2 Vivian Harris (Green) to Kirsty McBain (Labor), Eden-Monaro
92.2 Tony Hickey (Green) to Susan Templeman (Labor), Macquarie
91.6 Cate Sinclair (Green) to Lisa Chesters (Labor), Bendigo
91.3 Chetan Sahai (Green) to Tanya Plibersek (Labor), Sydney
91.2 Charlotte McCabe (Green) to Sharon Claydon (Labor), Newcastle
90.8 Jade Darko (Green) to Julie Collins (Labor), Franklin
90.7 Rachael Jacobs (Green)  to Anthony Albanese (Labor), Grayndler
90.6 Sam Wainwright (Socialist Alliance) to Josh Wilson (Labor), Fremantle

The highest flow to a losing candidate was 90.2% Greens to Labor in Deakin.  The flows in bold are likely to be slight underestimates, so it is not clear (pending further data, which may emerge) whether the Greens to Labor flows in Wills and Melbourne might have been stronger than Canberra.

Winning (or not) from behind and on minor party preferences

Sixteen seats (up four from 2016) were won by a candidate who didn't lead on primaries.  Labor won Gilmore, Lyons, Bennelong, Higgins, Robertson, Tangney and Boothby, overtaking the Liberals on Greens preferences.  The Greens overtook the LNP on Labor preferences in Ryan and Brisbane, in the latter coming from third on primaries (by 11 votes) in the first case of a candidate winning from third since Andrew Wilkie, Denison 2010.  Independent Dai Le passed Labor on Liberal preferences in Fowler, and new teal independents passed the Liberals on left preferences in Kooyong, Curtin, North Sydney, Goldstein, Wentworth and Mackellar.

In Labor's closest win, Gilmore, Labor needed an above-average 86.4% of Greens preferences to win but got 88.0% and won anyway.  Lyons (Labor needed 79.1%, got 87.1%), Lingiari (needed 68%, got 76.6%) and Bennelong (needed 74.8%, got 83.4%) were all close enough that it's likely that Labor owed these four seats specifically to the Greens' decision to recommend preferences to Labor on how to vote cards.  

In the Coalition's closest win, Deakin, Labor needed 91.5% so a very strong 90.2% was not quite enough.  Other seats where an impossible 100% Greens to Labor flow would have won Labor the seat had it occurred were Sturt (needed 90.6, got 87.9), Moore (needed 87.6, got 82.9), Menzies (needed 91.4, got 86.6), and Casey (needed 95.4, got 84.0).  In all these cases Labor needed figures that were above the national average and that were unrealistic in the seats in question.  On the ALP side of that ledger, independents Nicolette Boele (Bradfield) and Caz Heise (Cowper) could in theory have won with a 100% flow of Labor preferences, but their target percentages of Labor preferences (94.1% and 98.9% respectively) weren't realistic.  Preferences never flow 100% and it is pointless to reproach any party or its voters for this.  

There were also three seats (an unusually low number) where Labor led on primaries but would have lost with an even split of Greens voter preferences.  These were Lingiari, McEwen and Richmond, making a total of ten seats (down from fifteen) where Labor would not have won had Greens preferences split evenly.  The Greens also depended on Labor preferences favouring them in two of their four wins (Brisbane and Ryan), as well as on minor party preferences favouring them at the 3CP stage to get them into second in Brisbane.

Post 2019 there was much fuss for the next three years about United Australia preferences supposedly delivering the Coalition victory (which was false as there was only one seat, Bass, where an even split of 2019 UAP voter preferences would have made a difference).  In this case, because the Coalition has lost, nobody seems to care how many seats the UAP did or didn't save it.  In fact, there were two seats where had a specific right-wing party's preferences split 50-50, the Liberals would have lost: Deakin (any one of LDP, UAP or One Nation) and Menzies (Liberal Democrats only).  There were also other seats where some combination of multiple right-wing party votes (generally UAP plus One Nation) splitting 50-50 would have caused the Coalition to lose - Sturt, Moore, Casey.

As for the independents, unsurprisingly all the new indies needed some help from one side of politics or other and were always going to get it, but the extent of their reliance on preferences varied from seat to seat:

* As well as needing Liberal preferences to win from behind, Dai Le (Fowler) would also have lost had both UAP and One Nation voter preferences split evenly.

* Kate Chaney (Curtin) won because both Labor and Green voter preferences favoured her over the Liberals and would have lost had either not done so.

* Zoe Daniel (Goldstein) would not have won had Labor voters split their preferences evenly, and would also not have won had both Green voters and either DHJP or Sustainable Australia voters done so.

* Kylea Tink (North Sydney) also needed Labor preferences to win, and would have also lost if Greens voters and (voters for TNL, IMOP or both UAP and Sustainable Australia) not preferred her to the Liberals.

* Sophie Scamps (Mackellar) would still have won had only voters who voted 1 Labor split their preferences evenly, but would have lost had both Labor and either Greens or TNL voters done so.  (Scamps came from second on the final Labor exclusion but Labor was carrying preferences from other candidates.)

* Monique Ryan (Kooyong) and Allegra Spender (Wentworth) would still have won if either Labor or Greens voters had split their preferences evenly, but not both.  This is the same as for Rebekha Sharkie in 2019.

The crossbenchers elected in 2019 all won so easily that an even split of preferences from all the parties whose voters favoured them would not have stopped them winning.

Preference flows and tactical voting

Tactical voting arguments were attempted at this election by supporters of teal independents, and also to a much lower degree by a small number of Labor supporters attempting to defend Queensland seats from the Greens.  The tactical voting argument for putting a teal independent ahead of Labor was that Labor voter preferences would favour the teal independent more strongly than teal independent voter preferences would favour Labor, and therefore the teal independent (if second) might win in a case where Labor lost.  

The 3CP flow from urban teal independents to Labor was 70.5%, compared to 78.6% the other way.  (Kooyong, uniquely, had a stronger 3CP flow from the winning independent to Labor than vice versa.) However perhaps the narrowness of this gap speaks to some degree to the success of the tactical voting argument in converting intending Labor voters to teal voters.  Also, the argument only really applies to a hypothetical situation where Labor and the teal have about the same 3CP vote.  In a case like Kooyong, if teal voters who preferenced Labor shifted to Labor, this would reduce the remaining teal to Labor flow and increase the Labor to teal flow, so the strategic voting argument would then say that it is better to vote teal anyway.  

The correct test of whether it is better to vote teal if voting tactically is therefore not the preference flows between the parties but the actual 2PP.  On this the teal argument was vindicated because in all fifteen cases where Labor was excluded, Labor's 2PP was lower than the teal or teal-ish candidate's 2CP.  Furthermore, in nine seats teal-ish candidates won the 2CP and the seat but Labor lost the 2PP.  

In the three Brisbane area seats won by the Greens, the Greens to Labor 3CP flow was in the low 86s while the Labor to Greens 2CP flows were 81 in Griffith, 81.6 in Ryan, 83.1 in Brisbane.  In Ryan, however, the Greens were so far ahead of Labor that their 2CP exceeded the Labor 2PP.  In the other two, Labor's 2PP was above the Greens 2CP, so had the LNP been closer to retaining these seats, there is a very narrow range (0.67% in Brisbane, 0.61% in Griffith) in which Labor would have won but not the Greens.  This is the opposite to the general pattern in previous cases federally and in other compulsory preferencing states.  Overall, Labor and Greens preference flows to each other are more or less interchangeable but there will be rare cases where one would win but the other wouldn't (the only practical case being the Greens' win in Prahran 2014.)

Other sections may be added to this piece if I notice anything worth adding, or if there are any interesting requests that are practical to add here.

Wednesday, July 13, 2022

The Spurious Linking Of "One Vote, One Value" With Territory Senator Numbers

After each election comes a new season in which the Joint Standing Committee on Electoral Matters receives submissions and considers proposals for changes to electoral law.  This JSCEM season has special significance because as well as a change of government in the lower house, there has been a serious shift to the left in the Senate.  Any ALP legislation that is supported by the Greens and ACT Senator David Pocock will have the numbers to pass.

There have been several media articles commenting about this, though it is not always clear to what extent the articles are reporting on what Labor wants, and to what extent they are reporting on what other actors would like Labor to do.  A common theme in these articles (here's the latest) is that a proposal for more ACT and NT Senators appears in the context of a discussion of "one vote, one value" (a principle to which Labor's policy platform included a general commitment without any specifics.)  The linkage of the issue to "one vote, one value" is spurious.  From a pure one vote, one value perspective, the proposal looks like an attempt to rig the Senate to favour the left.