Tuesday, August 3, 2021

The 2022 Pendulum Only Slightly Favours The Coalition

 With redistributions in Victoria and Western Australia complete, the AEC has gazetted new boundaries.  Antony Green has released an estimated pendulum for the next election, now looking most likely to be held in 2022 rather than 2021.  There were some slight differences between Antony's initial estimates and those of William Bowe and Ben Raue, and I expect many of these would still apply to the final version.  Furthermore the AEC will release its own estimates later.  But pending the AEC estimates I thought I would use Antony's estimates as a starting point for a look at the 2022 pendulum and how much it helps or harms each side based on what we know so far about candidates.  This article is fairly mathsy and has been rated 3/5 on the Wonk Factor scale.  

There are three questions I am most interested in here:

1. All else being equal, what national two-party preferred (2PP) vote does the Coalition need for a better than even chance of a majority?

2. All else being equal, what 2PP vote does each side need for a better than even chance of winning more seats than the other?

3. All else being equal, what 2PP vote does Labor need for a better than even chance of a majority?


For these purposes I assume - pending strong polling evidence otherwise - that all seats won by crossbenchers in 2019 are won by the same crossbenchers again, that no new crossbenchers win, and that Hughes reverts to being a 2PP seat with Craig Kelly defeated.  The Coalition does have prospects of recovering seats from crossbenchers if things go well: Indi was very close last time and Helen Haines' popularity as successor to Cathy McGowan has not been tested (though Haines has had a term to pick up a personal vote), Warringah was lost to an indie because of Tony Abbott's extreme personal unpopularity in the seat, and Mayo is technically marginal.  If the Coalition does recover one or more seats from the crossbench then its 2PP requirement for the above targets will reduce slightly.

Question 2 is only a rough proxy for forming government in a hung parliament.  If all the 2019 crossbenchers are returned, then 75 Coalition seats would certainly result in government.  It is not entirely sure that 73 or 74 seats would be enough, but 72 (with 73 to Labor) would seem unlikely to be sufficient.  On the one hand, four of the 2019 crossbenchers are serving in Coalition-2PP seats, but it's also the case that the voters who elected Zali Steggall and Rebekha Sharkie after preferences were overwhelmingly ALP supporters on a 2PP basis.  There is some history of independents in Coalition-2PP seats at various levels deciding to back Labor governments anyway, and quite often getting away with it too.

Uniform swing paints an odd picture

Antony's 2022 pendulum is a good example of why simply assuming a uniform swing in each seat and then projecting seat totals from that is risky.  If we apply uniform swing, then the Coalition needs 51.2% for a majority, 48.5% to win more seats than Labor and 48.3% to prevent Labor winning a majority, so Labor wins a majority with 51.8%.  

But this ignores that the Coalition has some advantages if there isn't much swing either way.  Firstly, the Coalition has fewer very close seats than Labor.  Labor's average margin in its ten closest seats is 1.34% while the Coalition's is 2.72% - a remarkable difference.  The swing is never the same in every seat so if the average swing is about zero but there is random variation between seats, Labor is going to have a much harder time holding onto its close seats.  I get the standard deviation in 2PP swings across the last six federal elections at around 3.3%.  If there is zero swing, then just based on this difference between how close each side's marginals are, Labor would be expected to drop around five of its seats to the Coalition, while the Coalition would only drop around two back the other way.  

However, if there is a swing to Labor of around 2.5-3%, the Coalition has a bunch of seats on just above 3%, and Labor becomes more likely to outperform the pendulum.  That means Labor's target for winning more seats than the Coalition shouldn't be as high as 51.6, and its target for a majority might not be as high as 51.8.

Personal votes may also be relevant.  The Coalition has sophomore effect (a boost from the personal vote of a first-term member) on its side in seven seats that it holds by 5% or less, compared to just four such seats for Labor.  In three of these Coalition seats, its new MP defeated Labor's incumbent (which all else equal brings a further bonus) compared to two for Labor.  On the other hand, the Coalition currently has known retirements coming in Boothby (1.6%) and Casey (4.6%) while Labor's closest seat with a retirement is Lingiari (5.5%).  Eden-Monaro (0.8%) might be treated as a semi-retirement, because at the last election Labor had Mike Kelly, but in this case Kristy McBain has only held the seat for half a term.

Something else worth considering is the inflated margin in Lyons (ALP, 5.2%).  Lyons was one of four seats (the others being Isaacs, Wills and Scullin) where the Coalition candidate was disendorsed or "resigned" after their name was printed on the ballot paper, in this case resulting in the Liberal Party endorsing the Nationals candidate.  The disendorsed Liberal candidate finished second anyway, but Labor got a swing to it in Lyons while losing Bass and Braddon (which often move together with Lyons, although Lyons is more Labor-friendly) with hefty swings.  

Conditional Probability Model

My standard seat model for federal elections is a conditional probability model which aims to work out how many seats each side should win for a given 2PP (crossbench seats excluded).  It includes allowances for variable swing, personal vote effects, and the opportunity to add extras on a seat by seat basis (such as an adjustment for an opposing candidate disendorsement at the last election, or based on the dubious arts of seat polling).  State by state adjustments can also be added based on polling evidence, though there is a long history of state-specific federal polling failing in Queensland especially.  The probabilities for given seats will sometimes be off, but on the whole if the assumed 2PP is correct, the model does well (including when back-applied to 2019, with Bass the only seat that much surprises it.)

This is some sample output for a zero swing result based on the following adjustments:

* Standard deviation 3.3%
* Sophomore effect, sophomore surge effect and retirements all 1% (I may derive more accurate historic estimates later - eg compare William Bowe's estimates here)  
* Mid-term MP changes in by-elections treated as half a retirement.
* Opposing candidate disendorsed at last election -5% correction for inflated result (this too may be refined later, there are not many data points for this one!)
* No state or poll adjustments yet
* No allowances for redistribution except that Hawke is modelled as a Labor retirement because the entire division had various Labor sitting members in 2019.

(Click for larger clearer version):



Some seats where the numbers given should be treated with special caution are marked ( C ).  The columns are the current 2PP, the projected 2PP based on the assumptions given, and the estimated chance that that side wins the 2PP.  Note also that there's an overall adjustment to make the numbers add up after including personal vote effects, and as a result any Coalition seat with no personal vote effects has a projected slight swing against.

For an unchanged 2PP of 51.53%, no specific seat is projected to fall in the model, but a number are tossups and Labor's estimated probabilities of holding seats are lower, largely because their seats are on smaller margins.  As a result, if there is no 2PP swing, the model predicts the Coalition to make a net gain of two seats.

The model's estimates for the targets I started the article with are:

* More than 50% chance of majority Coalition government: 50.7% for Coalition (0.8% swing)

* Coalition wins more seats than Labor: 49.7% for Coalition (1.7% swing)
* Labor wins more seats than Coalition: 49.6% for Coalition (1.8% swing)

* More than 50% chance of majority Labor government: 48.7% for Coalition (2.7% swing)

Again, the pendulum becomes more favourable for the Coalition if it recovers one or more seats from the crossbench.

And here are some seat tally estimates for whole number Coalition 2PPs of the sort we see in polling for the current parliament, assuming the 2019 crossbenchers all retain and no others join them.

54%: 87-58-6
53%: 83-62-6
52%: 80-65-6
51%: 77-68-6
50%: 74-71-6 (hung parliament, probable Liberal government)
49%: 70-75-6 (Labor minority government)
48%: 67-78-6 (Labor majority from here down)
47%: 63-82-6
46%: 59-86-6

If the election is lopsided then a given 2PP should give almost the same seat return to the winner irrespective of which side gets that 2PP, though the Coalition does about one seat better.  If the election is close the Coalition does about two seats better for a given 2PP than Labor does.  On the whole it's very close on paper to a balanced pendulum at the moment, and more so than I expected before doing this model.  This may change as more retirements are announced.

At the level of Labor's recent polling (average around the low 52s) my model suggests Labor would almost certainly win, and would probably have a modest majority (around 80 seats).  That's better than I initially thought, and the reason turns out to be that at that level of swing the Coalition has a large number of seats where the model gives it an only slightly better than even chance.  However, because governments tend to recover from bad polling, the fact that polls say Labor would very probably win an election "held now", and do so with a good chance of a majority, means very little predictively.  Any but the most extreme polling evidence should be treated with caution after an election where polls were on average wrong by around 3%, especially since only one of the four active federal polls appears to have done anything useful about it, and another is untested.  (A repeat of the same error at the moment would mean the Coalition are still ahead.)

At this stage there isn't any poll-based evidence that the government is probably going to lose, and to even say that the polls now were showing Labor with an even-ish chance at this time I would want to see runs of 55-45ish leads for Labor. That doesn't mean those throwing money at Labor with the bookies are necessarily wrong, but they are mugs if they are doing it on account of the polling alone.   It will be interesting to see whether the current extended lockdown in New South Wales especially, and the connection being drawn with the government's slow vaccine rollout, results in much bigger leads in coming weeks and months or not.  

General cautions

While I've assessed the current landscape as only slightly favouring the Coalition for a given 2PP result (based on Antony's estimates and I suspect any others to be released), there are things that could unsettle that in either direction:

1.  One side could outperform the other in marginal seats.  This actually happened in 2019, and would have made a difference had the Coalition's overall 2PP been weaker (for instance had the Coalition's 2019 2PP been 2.5% lower in every seat it would have won the 2PP in 74 non-crossbench seats compared to the model's estimate of 67).  

2. State effects.  If one side is performing better than expected in a state that's unusually rich in close seats then it may outperform the model.  Big swings in Queensland and WA polling are a familiar trait of state by state federal polling but the usual outcome is that Labor underperforms in Queensland and nothing much happens in WA.  This time the low base in Queensland and the nature of pandemic politics in WA suggest a greater chance that these swings are real.  However, WA is not unusually rich in marginal seats and Queensland at present is only slightly over-represented.

3. Issues mix.  This looks like a very unusual election in terms of issues and where they might play out.  The 2019 election saw a lot of policy daylight between the parties and this generated big 2PP swings in some areas, resulting in 2019 having a larger variation between seat swings than normal (the largest in the last six except for 2010).  In the current environment we've seen the Liberal Party ditch previous anti-debt approaches and take a big-spending approach to keeping the economy afloat, much as Labor did during the Global Financial Crisis only more so.  This kind of ripping up of ideological lines can be hard for oppositions to counter and the early signs are that Labor's approach will be to move towards what it thinks is the centre itself by reducing tax and climate policy differences and, if still relevant, running against the government's COVID management record.  Whether this is the selective differentiation that worked for Kevin Rudd in 2007 or the small target approach that failed for Kim Beazley in 2001 (and was on the way to failing before Tampa/S11) remains to be seen. 

Both sides are annoying some usually sympathetic forces with their tactics.  A trickle of anti-lockdown types whose Sky News loyalties have trumped their Liberal/LNP memberships appear to be heading for the Liberal Democrats, and there has been plenty of criticism of the ALP on taxation from the left and centre.  The latter might play out in some interesting ways in inner city seats, including ALP vs Green seats where Labor's program might be less well received than in 2019, but the whole matter of Green seat prospects is probably a matter for a different article.  For the time being, I don't expect the Greens to win a swag of seats at a Reps election anytime soon, but one of these days they'll win something other than Melbourne.  

 Based on how the 2022 campaign develops, we might see big swings back to Labor in seat types where their failures in 2019 were more than just a reversal of the unusual inner city/outer city dynamic from Malcolm Turnbull's 2016 campaign.  It's probably too early to speculate about which seat types Labor has the best prospects in, other than that there is some risk of a swing back to Labor in Queensland being largely wasted in seats where the Coalition built a very large margin out of nowhere much last time.

More re retirements

The news of the impending retirement of the Speaker, Tony Smith, resulted in a small level of social media speculation that his retirement indicated he  believed the Coalition would lose.  This stuff following retirement announcements was ubiquitous in 2019 but in fact the number of retirements in 2019 was unremarkable and a Coalition victory was consistent with that.   So far there have been six government retirements announced from the two Houses combined, and there will probably be more, but the number has to at least double before history suggests there is anything unusual to see.  (I count Andrew Laming as a retirement in spite of him technically seeking preselection after announcing he wouldn't contest, while Kevin Andrews does not count as a retirement as he was seriously seeking to recontest but was deselected.)

As usual, I may add more comments to this article if I think of anything else relevant. It's difficult to avoid small errors in modelling like this so don't be too surprised if the table above changes slightly.

---------------------------------------------------------------------------

UPDATE 12/8: I have made one change to this model, reducing "sophomore surge" to half a point in line with declining evidence at recent federal elections.  I have also corrected a spreadsheet bug.  The article has been edited accordingly - results are slightly more favourable for the Coalition, moving Labor's 2PP targets for a seat plurality and seat majority by 0.1% apiece.  Many of the seat tally estimates for given 2PPs changed but a lot of these were a matter of a small change plus rounding.  

14 comments:

  1. Great analysis, thanks Kevin. Would the suspected drop in the female vote alter the coalitions chances much in some seats more than others do you think?

    ReplyDelete
    Replies
    1. The variation in numbers of enrolled female electors by seat isn't large (there are none over 53%) so it would have to come down to a division that was competitive having an unusual proportion of women who were particularly influenced by the Coalition's recent gender issues.

      That said I'm not that convinced there has been a widening of the gender gap. One pollster (Essential) has been finding a big gender gap in its sampling but others generally haven't been finding anything unusual compared to recent elections.

      Delete
  2. Another potential factor in some seats is the dramatic increase work from home allowing a significant movement of working age professionals to some regional areas. This presumably more effects individual seat margins, rather than overall polling, presumably mainly in sea-change and tree-change electorates.

    It is little tested. The WA election was such a large landslide it was presumably harder to extract evidence of any significant underlying swings. It may have been a factor in the Queensland election, although that was only a few months in, the flow has continued and the intrastate migration trend appears to be different in Queensland and WA. Tasmania also appears to have been relatively less effected, although I may not be correctly adjusting for different population sizes and thus the effects of that at the state election were presumably minimal.


    The ABS has released the internal migration statistics for the March Quarter today:

    https://www.abs.gov.au/statistics/people/population/regional-internal-migration-estimates-provisional/latest-release

    Presumably the trend will also continue, particularly involving outflow from Melbourne and Sydney, with housing supply constraints and new house building time staggering the movement of people to regional areas.

    ReplyDelete
  3. Liberal candidate in Isaacs was also disendorsed.
    https://www.abc.net.au/news/2019-05-01/federal-election-liberal-candidate-dumped-anti-islamic-comments/11061480?nw=0

    ReplyDelete
    Replies
    1. Yes, I mentioned that in the article but didn't place as much emphasis on it as Lyons because I have the feeling Labor took their foot off the pedal in that one (whereas in Lyons they were quite concerned about a potential Hanson situation). Nonetheless my model includes an adjustment for Isaacs and this is why its win probability is much lower than surrounding seats.

      Delete
  4. "only one of the four active federal polls appears to have done anything useful about it, and another is untested."

    Which two polls have made these changes?

    ReplyDelete
    Replies
    1. There are four significant pollsters that have recently produced national polling: Newspoll (administered by YouGov), Essential, Roy Morgan and Resolve Political Monitor. Of these:

      * YouGov/Newspoll has made direct methods changes targeted at apparent causes of the 2019 failure.

      * Resolve Political Monitor is untested as it was not polling at the last election.

      * Essential and Roy Morgan do not seem to have done anything that is usefully directed at the possible causes of the polling failure. (Essential's documented changes are in my view irrelevant, while Morgan has switched from face-to-face to phone/online polling but did so in response to the pandemic rather than the polling failure.)

      Delete
  5. Have you tested the hypothesis that the more marginal a seat the lower the percentage swing?

    ReplyDelete
    Replies
    1. Not across multiple elections, but for 2019 the pattern was if anything very slightly the opposite. Especially Capricornia and Dawson saw double-digit 2PP swings. I'm not sure if there's any reason to expect that more marginal seats would have lower swings - very marginal seats tend to get attention from both sides, but very safe seats often get attention from neither.

      Delete
    2. Ethan of Armarium Interreta here.

      If I plot 2pp margin before the election versus size 2pp swing for elections 1993-2019, there is practically no correlation (R^2 = 0.0002).

      Averaging the size of the swing for marginal seats (defined by the AEC as being seats held on 56% or less) produces no difference with non-marginals. Both round to an average 3.8% swing, with the difference being somewhere in the 2nd or 3rd decimal place.

      I think the problem here is that there are elastic and non-elastic marginals. Some seats are almost always very close no matter which side wins (e.g. Canning, Eden-Monaro). Others are very elastic - they jump all over the place (e.g. Banks, Braddon). This may be down to various factors, e.g. which areas the seat is comprised of (off the top of my head Eden-Monaro is mostly combined from rusted-on Labor areas and rusted-on Coalition areas), how swingy the area's voters are, how diverse an electorate is (I suspect an electorate which combines wealthy areas and less-wealthy areas would swing less, because gains from a policy supported by the wealthy would be cancelled out by losses among the less-wealthy and vice versa) etc.

      This also applies to safe seats too - there are safe seats which barely budge from election cycle to election cycle (e.g. Bradfield, Canberra) and there are some which jump around a lot (e.g. Blaxland, Barker). The latter group may not ever swing hard enough for the other side to win (Labor has never lost Blaxland, and Barker has always been Coalition-held), but they are as elastic, if not more so, as any marginal.

      Delete
  6. Assuming that all the crossbenchers retain their seats, how many seats must Labor win to have a 50% chance of forming government? Obviously this depends on the inclinations of the crossbenchers; if all six members of the crossbench came from One Nation, it'd be unlikely that Labor would be able to form government even with 75 seats (one short of a majority).

    With 74 seats, it's unlikely either Bandt or Wilkie would support a Coalition government under either Morrison or Dutton. I also think that Katter, Steggall or Sharkie would support a Labor government in any form. However, there is a chance that Helen Haines might do a Tony Windsor and support a minority Labor government of 73 seats.

    So the minimum would probably be 73-72-6. Logical given that it is also the lowest number of seats whereby Labor would have more seats than the Coalition. Your thoughts?

    ReplyDelete
    Replies
    1. I discuss this a bit near the top of the article. I'm not really sure if the answer is 73 or 72 - with 73 Labor would be quite confident at least. Previously Labor governed with support from independents from Coalition 2PP seats so I wouldn't rule out Labor governing with 72. If Steggall thought she was going to lose at the next election anyway she might throw in with Labor for more action on climate change for example. On the other hand, it is very hard to see Labor governing with 71 and needing to constantly herd every crossbencher except Katter and then pass votes by casting vote, and with the Coalition declaring the government illegitimate and running a hard line on pairing.

      Delete
  7. Kevin,

    Re your reduction of the sophomore surge adjustment: it may be better to simply assume sophomore surge is roughly the same size as regular sophomore. Once you adjust for state 2pp swing (instead of using deviation from national 2pp swing), sophomore surge is very slightly larger than regular sophomore:

    https://twitter.com/ArmariumI/status/1425732666771480580

    ReplyDelete