Wednesday, September 29, 2021

Tasmania Doesn't Swing With The Nation Like It Used To

Advance Summary

1. Tasmania is routinely omitted from national state-by-state federal polling breakdowns because of small sample size.

2. In theory Tasmanian seats could be projected by just assuming Tasmania would get the national swing, but in practice this isn't very accurate.

3. In elections in recent decades, Tasmanian two-party preferred federal swings have been less similar to the national swing than was the case before the 1970s.

4. One cause of this is that while national swings at federal elections have tended to be much lower in recent decades, 2PP swings in Tasmania are not much lower.

5. Another cause is the tendency of groups of Tasmanian seats to swing together.

6. As a result of this, the national swing that has actually occurred has only correctly predicted the Tasmanian seat result once at the past ten elections.

7. While the swing in current national polling (if accurate) implies the Coalition would not win any Tasmanian seats in an election "held now", a more accurate read based on the history of projected shutouts is that the Coalition would probably retain one Tasmanian seat at the moment.

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

This week was Newspoll quarterly aggregate week, which saw a fresh round of media "analyses" of how many seats the Coalition was standing to lose.  It's all a nonsense game at this stage -  projected landslides in Western Australia have turned into nothing much on election day and projected braces of Queensland Labor gains have turned to nothing or losses often enough before. The 2PP on election day is historically more likely to be much closer to 50-50 than whatever the heck it is today.  As noted in my recent polling roundup, three governments polling worse than this at about this time ended up winning the 2PP and gaining seat share.  The most inspiration I drew from this whole exercise came, oddly, from people on Twitter who again queried the omission of Tasmania from the quarterly figures, and that eventually led to this article, which is quite numbery in places and rated 3/5 on the Wonk Factor scale.    
Of course, Tasmania is left out because of insufficient sample size, but that's no reason to ignore it as a place where seats might change hands.  Three of Tasmania's five seats are in play:

* Bass (Lib 0.4%), which has booted its incumbent MP at eight of its last nine opportunities and enters the election as probably the Coalition's most marginal seat.

* Braddon (Lib 3.1%), which is not far behind Bass in electoral bloodlust and which is also among the Coalition's most marginal seats.

* Lyons (ALP 5.2%), which might not seem too risky for Labor on 5.2%, but which is on an artificially inflated margin after the Liberals had to disendorse their 2019 candidate.  

An obvious approach to "nowcasting" seats off the Newspoll aggregate would be to just assume Tasmania follows the national swing (in which case Labor would recapture Bass and Braddon, adding two seats to the mere "six to ten" (eleven by uniform swing) that Simon Benson had them losing.)  But how well does this work, and what is Tasmania's history of following or not following national swing trends?

First I used David Barry's 2PP figures to find 2PP swings nationally and for Tasmania, though these are rubbery for the earlier elections where seats on the mainland were sometimes filled unopposed and seats in Tasmania were sometimes contested only by one major party.  But the rubbery nature of the early 2PPs generally doesn't bite in what I looked at.  Here's the graph:



Well that looks pretty predictive, though the outlier in the bottom left (1931) is doing a fair bit of the lifting here (the r-squared drops to 0.55 without it).  But when I split the sample into early and later halves, something happens:


Even with the 1931 outlier removed to stop it from swamping the regression, this is a very tight relationship.  Most of these elections are more or less bang on the trendline (albeit with a slightly muted average swing for Tasmania), with only three seriously out.  But the second half is a different story:


Tasmania still usually swings in the same direction as the rest of the country, but the relationship has become a lot looser, with a lot more cases where it swings in the other direction.  Only a few of the dots are especially close to the trendline, and the one that lands on it is 2019, which does so only by accident given the candidate malfunction in Lyons.

What has happened here?  One clue comes from the average 2PP swings per election.  For the elections from 1922 to 1969 the average national 2PP swing was 4.47% (3.67% excluding 1931) and the average Tasmanian swing was 4.57% (3.57%) - very similar figures.  But for 1972 onwards the average federal 2PP swing has fallen to 2.84% while the average Tasmanian swing is still quite high at 3.96%

One possible reason for this (see comment by Ethan) is the "two Tasmanias" pattern in which at many elections Bass, Braddon and Lyons do one thing and Clark and Franklin do another.  

How does this play out in terms of seat projections for a given expected 2PP swing?  As a way of getting a handle on this I thought I'd try the following: how many Tasmanian seats would have been predicted correctly by assuming the Tasmanian swing was the same as the national swing?  (For these purposes I'm interested in the 2PP winner only, and if either major party failed to make the 2CP at an election then I assume the other party will win the seat at the next one).  Here I made much use of figures from Psephos; any errors are mine but are unlikely to affect the overall picture.  



The "proj ALP" and "proj Coal." columns give the number of Tasmanian Reps seats each side would have been expected to win on a 2PP basis based on the national swing to Labor, and the "ALP" and "Coal." columns give the number actually won.  (Since 2010 one seat, Clark has been held by an independent but Labor has continued to win the 2PP easily.)  The "error" column gives the error in projecting the result for Tasmania off the national swing.  

What is interesting here is that for elections through to 1990, the national swing predicted the Tasmanian seat result (though not always every specific seat) perfectly about half the time, and was only more than one seat out three times (1934, 1943, 1983) in 28 attempts.  

From 1993 onwards, the national swing has predicted the Tasmanian seat result perfectly just once in ten elections (2007) and has been more than one seat out half the time (1993, 2004, 2010, 2013, 2019).  

One might still expect that the national swing would easily beat a dummy prediction like "ignore the national swing and just assume all the incumbents win".  But for elections from 1977 onwards, this dummy prediction has actually beaten the national swing as a predictor for Tasmania (19 2PP errors to 21).  (Note that the start point of 1977 is cherrypicked for effect as the dummy prediction did very badly in 1975.)

Have Tasmanian seats become more marginal over time?  Excluding elections in the 1920s (in which there were some 2PP walkovers), not much - the average margin per seat contest was 6.7% from 1931-1969 and 6.0% from 1972 onwards.  This difference is almost entirely down to the 1931 blowout and the medians of the two groups are the same.  

Rather, the main reason national swing has become a poorer predictor of seat results in Tasmania is that Tasmanian 2PPs have diverged more from the national 2PP at recent elections.  The mean difference from 1972 onwards has increased from 2.3% to 3.1% and the median is up from 1.7% to 2.4%.

Cases where national swing projection failed

Is there a reason for this, or is it just a cluster of elections with unusual characteristics?  I'm not sure.  But it's worth a look at some of the elections where Tasmania has diverged from national swing on both a 2PP basis and a seat basis:

1934 (Labor overperformance 7.4% 2PP and 2.5 seats - one seat is tied based on national swing on Psephos estimates): Based on the national swing (5.0%) Denison should have been line-ball and the other UAP seats safe, but not only was the swing in Tasmania (12.4%) way higher than the national swing, but it was also uneven enough to take down Franklin (UAP 13.0%) and Bass (UAP 14.5%).  I am not sure why there was a monster swing to Labor in Tasmania in this year.  

1983 (Labor underperformance 7.1% 2PP and 3 seats): This was the Franklin Dam election.  Labor under Bob Hawke promised to use Commonwealth powers to stop the Tasmanian government's proposed Gordon-below-Franklin Dam.  This resulted in a swing against Labor in Tasmania at an election it won, and no seat gains.  (Labor had got within 80 votes of taking Wilmot in 1980 but had to wait til 1993 to claim it.)

1993 (Labor overperformance 5.2% and 3 seats): This was the Keating vs Hewson GST election.  Tasmanians are vulnerable to cost of living issues so the campaign against the GST was especially successful here with Bass, Franklin (which was vacant and had seen a messy Liberal preselection) and Lyons falling on above national average swings in an early sign on election night that the Coalition's attempt to seize office was belly up.

2010 (Labor overperformance 7.0% and 2 seats): Nobody really knows why the Liberal Party was so woeful in Tasmania at this election.  Was it a hopeless broadband policy? Eric Abetz at the top of the Liberal Senate ticket?  Tasmanians liked Julia Gillard?  The Liberals not making an effort?  Whatever the reason Labor had a good swing to it in Tasmania despite the national swing against.

2013 (Labor underperformance 5.8% and 3 seats): This was a forestry election and Labor was smashed in the northern three seats over the state Labor/Green government's forestry "peace deal".  Another Labor/Green government in Canberra made an easy target and three seats that would have been safe based on the national swing were lost.  A correction from whatever drove the 2010 anomaly was probably also a factor here.  

The following are some cases where the Tasmanian swing wasn't that unusual but the seat results defied the national pendulum prediction anyway:

1943: (Labor underperformance 2.2% and 2 seats): The national 2PP swing of 9.3% would on paper have won the UAP two seats but the smaller Tasmanian swing was enough to account for the UAP retaining Darwin (UAP 9.2%).  In Wilmot (UAP 5.0%) the UAP's margin from 1940 had been deflated by the seat being occupied by Labor as a result of a by-election and hence the 1943 swing to Labor was reduced.

2004: (Labor underperformance 1.85% and 2 seats): The national 2PP swing of 1.79% would on paper not have dislodged any of Tasmania's 5 Labor MPs but this was another forestry election and disagreement with Labor's forests policy contributed to stronger swings against Labor in the northern three seats and the fall of Bass and Braddon.  (In Bass the link with forestry areas in the swings was less pronounced and there were other issues in the outer suburbs of Launceston that were probably more decisive.)

2019: (Labor underperformance 0.23% and 2 seats): A uniform national swing would not have changed anything in Tasmania but Braddon would have been very close.  This was another election with a north/south divide, driven partly by the nature of the election but also by parochial campaigning against Hobart-centric Labor election commitments.

Nature of seat projection errors

Overall, there are 19 elections for which the national swing projects no change in Tasmanian seats from the last election.  Nine of those actually saw no change, six saw change in the direction of the national 2PP swing and two saw change against it (in one case there was no national swing to speak of).  Four in the former category are recent (1993, 2004, 2013, 2019) so when national polling swings project no changes in Tasmania, this should be treated with caution since Tasmania may have a larger or a more effective swing than the rest of the country.

There are two elections for which the national swing leads to a fence-sit projection in one seat. One saw no change, one saw a three-seat change (1934).

There are 17 elections for which the national swing projects a change in Tasmanian seats.  In four cases (three of them early plus 2007) the change occurred as predicted, in five cases there was no change, in five cases there was more change than expected, in three cases the change was less than expected (one of those by two seats).  On average, when a change was predicted the prediction was wrong by an average of one seat, rising to 1.2 seats in the second half of the sample.  Therefore projections that seats will change hands in Tasmania off the national swing - even when based on supremely accurate polling - have a high error margin. 

There was at least, however, no case where a change was predicted in one direction and occurred in the opposite.  So that's something.

There are 11 cases for which national swing implies a 5-0 2PP shutout to one side.  Five of the shutouts occurred (1925, 1974, 1977, 1984 and 2007).  Six didn't (1943, 1969, 1987, 2004, 2013, 2019).  On average when national swing implied Tasmania should have been a 5-0 shutout, the actual result was 4-1.

Is Tasmania a weird unit lately?

We shouldn't expect national swings to perfectly predict state gains and losses (and whether they do or not is beside the point for states that have state-specific federal polling, if that state-specific polling is better than assuming national swing.)  However, Tasmania's run of errors in this regard lately seems unusual and does suggest that the state is behaving more erratically than other states.  At the last ten elections, national swing has on average predicted the Tasmanian Reps seat tally wrongly by 1.8 seats.  At a quick check I estimate that the average error for the other states (excluding territories) is 1.45 seats per state per election.  That's although the average mainland state has over five times Tasmania's seat count.  

How should Tasmania be "nowcast"?

For the purposes of deciding how many seats would be expected to fall in Tasmania in an election "held now" and assuming national polling is accurate, the above suggests the national swing figure should be treated with caution.  The maximum possible government loss is two seats but given how often seat gains projected off national swings have been overstated or not happened at all, it may be one seat is a better read.  

Morgan

Finally, Morgan is the only pollster currently providing Tasmanian federal breakdowns, albeit in the form of respondent-allocated 2PPs only.  These small Morgan samples are volatile (probably based on fewer than 100 voters per poll), and individual Tasmanian samples from Morgan should be ignored and never reported by media.  However, aggregating a run of, say, five of them in a row may be vaguely useful, though the sample size is still very small.  In 2013, Morgan samples both had Labor too high on Tasmanian 2PP and had the difference between Labor's national and Tasmanian 2PPs too high.  Probably the error on the Tasmanian 2PP after adjustement for national swing after the last such poll was about 4%.  In 2016, the last five Morgan samples released (before Morgan stopped releasing detailed national polls a month out) had an average Labor 2PP in Tasmania of 54.9 (2.5 points below the election figure), but with a range from 45.5 to 63.  In 2019, Morgan did not release state breakdowns.  

For the last five Morgan polls in 2021, the average national 2PP was 53.7 to Labor, a figure exaggerated by around a point by the use of respondent preferencing.  The average Tasmanian 2PP has been 56.4 to Labor, an implied swing of only 0.4%, and that potentially inflated by respondent preferencing.   Given the past errors in Morgan's estimates of the Tasmanian 2PP, this could be consistent with an election held now returning any Tasmanian result from Labor perhaps even losing Lyons to Labor regaining both Bass and Braddon.  (Figures updated for Morgan released 30 Sep).

4 comments:

  1. G'day Kevin.
    The table that projects Lab and Lib seats actually has proj Coal twice. Unless I'm missing something it's just a tad confusing. Otherwise an excellent post as per usual.
    Barry

    ReplyDelete
  2. This might be relevant:

    https://twitter.com/EthanOfHouseK/status/1425338566335340544

    Basically, the uncertainty on a uniform swing projection for Tassie is about 3x higher than it is for NSW. I've reproduced the standard deviation from uniform swing for each state (measure of uncertainty, higher = bigger deviations from uniform swing) below:

    NSW 1.2
    VIC 2.1
    QLD 2.3
    WA 2.7
    SA 2.3
    TAS 3.4
    ACT 1.7
    NT 2.1

    You can see Tassie stands out in terms of how much it defies a uniform swing; with its stdev being 1.25x higher than the next state's (WA, funnily enough).

    This might also be of interest:

    https://twitter.com/EthanOfHouseK/status/1425346204980813826

    It's basically the deviation from uniform swing for each Tassie HoR division. Couple of things stand out to me:

    1. Oftentimes, when one electorate defies uniform swing in a certain way (e.g. swinging more to Labor), at least one other electorate tags along. In a majority of elections you see two groups of Tassie seats - the Bass/Braddon/Lyons group and the Franklin/Clark group. Usually deviation from uniform swing is correlated within each group - e.g. if Bass swings harder to the Coalition then Braddon will also swing harder to the Coalition, and both groups will usually deviate from uniform swing together.

    2. The non-Hobart electorates - Bass, Braddon, Lyons and Franklin - have a much higher deviation from uniform swing than an average electoral division. Again, reproducing the standard deviation from uniform swing for each:

    Bass: 5%
    Braddon: 4.9%
    Lyons: 4.7%
    Franklin: 5.2%
    Clark: 3.2%

    For reference, the average deviation from uniform swing for all electorates is 3.3%. That means the uncertainty on a national uniform swing projection for the four non-Hobart electorates is 42%-57% higher than it is for an average electoral division.

    ReplyDelete
    Replies
    1. Thanks for that. Yep, this is all consistent with my experience; the "two Tasmanias" thing of Bass/Braddon/Lyons (BaBLyon for short!) often doing one thing and Franklin/Clark doing another, though factors in specific seats will often mess up the pattern.

      Delete