SA State Election March 15

South Australia goes to the polls on March 15, the same day as Tasmania. Tasmanian readers should not be cross with SA for this; after all, they were certainly there first, and no-one has advanced any really good reason why Tasmania had to follow suit.  An unfortunate impact of this is that election analysis resources around the country will be strained, and I especially will be too busy with the Tasmanian poll to devote much attention to the South Australian situation.  But I do think SA is interesting too, and here are my comments on the SA election generally and especially one quirk of the SA electoral system.

South Australia loosely parallels Tasmania in that both see ancient Labor governments looking to the presence of a new Liberal government in Canberra as their primary hope to escape a drubbing.  Both are perceived as having won one election too many (in SA's case despite losing the 2PP vote by an embarrassing 48.4%:51.6% two-party margin).  But the problems in Tasmania run much deeper, as the age of the government is even greater and it has spent the last four years in a widely disliked coalition with the Greens.  The SA government has at least enjoyed an outright majority, with 26 of the 47 seats to 18 for the Liberals and three independents.



There was some hope therefore that the South Australian government might buck the trend and be able to turn the federal Abbott factor to its advantage.  However a recent run of especially lopsided polls (55-45 from ReachTEL and Galaxy and 54-46 from Newspoll) together with the poor resilience record of struggling state governments suggest this isn't happening and that the election is practically over as a contest.  Another historical death rattle that is evident in this polling is the closeness of the preferred/better premier polling, with Newspoll showing Weatherill a pitiable point ahead and Galaxy showing Marshall favoured.  Because preferred leader scores heavily favour incumbents (except when polled by ReachTEL) winning premiers in any state except Tasmania have always had large leads on these questions close to elections.  Weatherill does not, even though he is not all that unpopular.

The polling blowout into decisive-loss territory closely followed (and in my view was probably caused by) an extraordinary interview in which Premier Jay Weatherill threatened to resign before the election if former Senator Don Farrell, famed for his role in the dumping of Kevin Rudd in favour of Julia Gillard in 2010, was preselected by his party.

Why should this have been a damaging episode, one may well ask.  After all, if the factional machinations of the "faceless men" are a bad thing, then surely having a Premier willing to put them to the sword is good and could be popular?  The problem is that if Weatherill had absolute authority over the party he would not have needed to threaten to blow it up by quitting on the eve of an election in order to get his way concerning Farrell.  The episode showed either that Weatherill did not have control over his party and would probably soon be dumped if the government was re-elected, or that he was prone to pointless melodrama.  Little wonder the radio host who received this confession asked Weatherill if he was serious.

The Use And Abuses Of Uniform Swing

The main argument against the view that this election is done and dusted has come via the use of the pendulum and the assumption of uniform swing.  The Liberals need to take six seats for outright victory, and even assuming they take Mount Gambier from independent Don Pegler, that still leaves five.  With the fifth seat on the pendulum, Mitchell, coming in at a 2.4% swing, it may seem that the Liberals need a 2PP vote of around 54% for an even chance and could even be struggling to form government with more than that if really unlucky.  Indeed, Antony Green tweeted today that the SA Election Calculator produces a hung parliament.

As people will hold out false hopes that Labor can salvage a miracle win despite being thrashed on the 2PP, I am going to be blunt about this: the calculator's output is misleading.  It is misleading because it assumes an exactly uniform swing, but exactly uniform swings never happen.

Swings can be unusually non-uniform for various reasons.  These can include sophomore effects favouring the side that gained seats at the previous election, superior marginals campaigning by one side or another or an unusual regional distribution of the swing.  But even assuming that none of these apply there is a more important problem, one that applies in all cases.  It is that rather than being exactly the same everywhere, swings vary from seat to seat as a result of various small-scale factors.  Seat swings tend to be normally distributed around the mean swing, with an average difference from it of a few points.

If the state swing is 2.4% to the Liberals, then even assuming none of the above issues apply, some seats will swing by 1% and some by 4%, some will swing by several points and some will even swing to Labor.  In the 2002 state election, the last South Australian state election to see a relatively modest swing, the swing to Labor in classic ALP-Liberal two-candidate-preferred contests was 0.30 points with a standard deviation of 2.87; sixteen such seats actually swing to the Liberals.  And the bigger the swing, the bigger the standard deviation - in 2010 the mean swing in classic seats was 9.28 points to the Liberals with an SD of 3.75.

Usually this doesn't matter because the gains and losses resulting from the vagaries of candidate selection and specific seat issues more or less cancel out, making the pendulum a reasonable model.  But if we look at the calculator's implied 2PP result for the 54-46 Newspoll, it is glaringly asymmetric:

(Click for larger version).  On this model, the ALP is shown as retaining three seats by margins of less than 1% and 10 by margins of less than 5%.  The Liberals in contrast are shown as gaining four seats by margins between 1.8% and 2.3% and holding everything else by over 5%.

Now, instead of predicting seats on a win-loss basis, we should take account of the likelihood of a degree of variation in the swings between seats and convert each one to a probability.  If the uniform swing model predicts victory by 0.03% then that really means that, all else being equal, Labor's chance of victory in that seat is barely over 50%.  On the other hand, a 1.8% margin is a more likely but by no means certain win. 

This is how the 44 projected "classic" seats come out if we convert them to probabilities based on a 2.4% swing with a 2.8% standard deviation - a relatively conservative assumption:

 (Click for larger version.)

This model suggests that because so many close seats are notionally Labor's, given a 2.4% swing, they would actually lose about three of the 22 seats assigned to them by the calculator.  Meanwhile, all else being equal, the Liberals would be expected to win 21 of their assigned 22.  So just based on taking into account the known fact that seat swings vary, the hung parliament projected by the ABC calculator for a 54:46 result becomes a slightly more probable than not Liberal majority win.  It becomes a more likely Liberal win if they're thrown at least one of the Independent-occupied seats as well. 

Of course, the probabilities in the above seats could be wildly inaccurate in specific cases (even assuming a 54:46 2PP, and it could well be larger.)  For instance, the loss of the seat of Adelaide in the last election is widely seen as having been abnormal, and even with a small 2PP swing against Labor, many would think Labor's chances of recovering it would be a lot better than 1%.  Also, Labor has a sophomore advantage in the seat of Mitchell.  But when you start fiddling the expected swing in one seat, that does imply a greater swing over the rest to make up for it.

A 54:46 2PP would, all else being equal, be good for a thumping win with around, say, 28 seats, so it is intriguing that even the fixed-up uniform swing model puts the Liberals only barely and probably over the line.  And this brings up the long and intriguing history of South Australian election seat tallies not aligning with the 2PP for a wide range of often unrelated reasons, and the history of attempts to fix it.

The Playmander Problem And Its Successors

Sir Thomas Playford IV was Liberal and Country League Premier of South Australia from 1938 to 1965 and in this time "won" eight state elections.  At least two of those (1953 and 1962) were only won as a result of a system of pro-rural malapportionment (nicknamed the Playmander, although malapportionment and gerrymandering are different) that grossly favoured the LCL. 

A toned-down version of the Playmander saw the LCL under Steele Hall win another undeserved victory in 1968 despite scoring a two-party preferred vote of just 46.8%.  One-vote-one-value legislation passed in 1975 might have seemed to have fixed the problem, with rural malapportionment and hence the Playmander removed for good, but John Bannon then won the 1989 state election for Labor with 48.1% of the 2PP vote. 

At the time Bannon was himself accused of presiding over a "Labor gerrymander" in which enrolment sizes varied wildly between electorates as a result of inadequate redistribution processes.  Looking at the 1989 results I find not that much evidence that this problem (while serious) was the main cause of the Bannon win - the electorates won by the Liberals had only marginally higher turnouts than those won by Labor on average.

While the Playmander itself is decades dead, it seemed its legacy persisted in the form of a heightened concern about results in which parties won "undeservedly".  The 1989 election
resulted in a 1991 referendum that entrenched a review of boundaries after each election, and more contentiously introduced a "fairness clause" that requires redistributions to attempt to cause whichever party (or coalition) receives over 50% of the two-party vote to win the election. 

The fairness clause itself then totally failed to achieve its aim in 2010 when Labor won a majority with 48.4%.  The swing of 8.4% would have typically wiped out around seven Labor seats, at least costing the party its majority, but in fact Labor lost only two of the seven seats below the swing line, and one above it, and also won a seat from an independent, for a net loss of just two.  An important factor here was that Labor had won the 2006 election with an increased seat margin, meaning that in six of the seven target seats, it had the benefit of sophomore effect in 2010.

The Electoral Commission decided the variation between swings in 2010 was about the strength of campaigns and that the boundaries themselves were not unfair and therefore refused to conduct the major redistribution that may have been expected (see Antony Green on this here.)  However, while the Rann government certainly did sandbag its marginals remarkably well in 2010, sophomore effect (which contributed to this) is not a campaigning issue but rather an innate property of name recognition gained in a previous term.  It is not something Oppositions can be expected to effectively counter.

This is one reason why I find the whole "fairness clause" (which is embedded by referendum and can only be removed by referendum) perplexing and even self-defeating.  If a party has gained several seats at a previous election, then all else being equal, sophomore effect will result in that party winning the next election even with slightly less than 50% of the 2PP vote.  However, it is difficult for electoral authorities to make assumptions about something like that in advance.  Furthermore, it can work the other way: a government that has suffered large seat losses (but remained in power) will probably lose the next election with a 50% 2PP if the "fairness clause" requirement is strictly applied.

Another problem is that if a state naturally has an urban area that leans mildly to one party and a rural area that leans strongly to another, then drawing electorate boundaries so as to set up a scenario in which 50% of the 2PP vote wins 50% of the seats becomes almost impossible, without violating the principles of commonality of interest that are normally used in electorate design and producing something that really does look like a salamander.

Finally, the whole argument that a party that has won the election while losing the 2PP has necessarily won undeservedly (federal example Howard in 1998) is in my view a furphy.  Parties play to win in seat terms under the rules that exist.   A party may deliberately decide to sacrifice a large amount of support in a safe seat to gain a small amount in a marginal.  But this does not mean the same party would have lost had the criterion for victory been winning the 2PP.  It just would have made different tactical calls and pushed different lots of pork to different places.

Really, if the overarching concern is that the two-party preferred vote determine the formation of government, the answer is not continually trying to adapt the single-seat system to a purpose to which it is unsuited.  Rather the answer is to find a new electoral system, which requires a new referendum.

It seems from the most recent redistribution that the Electoral Districts Boundaries Commission finds itself saddled with a rule that is an ambiguous and inflexible overreaction to a history of problems in the 2PP/seat tally conversion area, but that is nonetheless entrenched in the constitution.  It therefore takes the same approach to the rule as has often been taken by American courts to disapproved provisions of the USA constitution: interpret it out of existence. 

While the Liberals claimed that the lack of a significant redistribution was unfair and claimed they would need 54% 2PP to win, I've shown above that the victory hurdle, on average, isn't quite that high.  Indeed, if an overall swing of zero from the 2010 result is assumed, the abundance of Labor seats on margins below 1% means that random variation in swing size between seats alone would be expected to cost Labor 4-5 seats, meaning that any swing at all should at least cost Labor its majority.

Furthermore the Advertiser-Galaxy poll,  albeit with a relatively small sample size, provides indications that the swing to the Coalition could be greater in the urban seats where Labor is at most risk.  Both sides played down the prospect of an 11-seat Liberal gain, but they would do that even if it was true, so that doesn't really prove a lot.  As evidence against that trend, a new poll of the seat of Colton isn't much more than a sniff (since the seat should be pretty close anyway) but it is interesting that Labor can poll a 50-50 2PP in a seat where Steven Marshall is preferred premier.  The report on the Colton poll claims the 3.6% swing in Colton is "significantly less than a general 5.3 per cent swing to the Liberals in broader metropolitan Adelaide" but in fact the difference is not even close to statistically significant.  Incumbent Paul Caica has a outstanding +36 net rating.

I am expecting a Liberal win in this election, though there is a fairly high chance of  the margin being underwhelming in seat terms for the 2PP result.  I have not had time to look in any detail at the Upper House (a Senate-like exercise involving group voting tickets ... shudder!) but those interested may find the comments of The Tally Room and Casey Briggs of interest. 

This piece is intended as a broad overview of themes relevant to an election I haven't been following at a fine-scale level.  Comments from those on the ground and with knowledge of polling especially in specific seats very welcome.

Addendum (March 3):  Very experienced electoral researcher and former AEC staffer (now retired) Michael Maley has sent me some further comments on the problems with the ideals underlying the SA "fairness clause".  One of these that was in my mind a few months ago but that I forgot to add into the article is "that any process of boundary drawing to produce a particular outcome is a blunt instrument rather than the type of scalpel which the fairness clause assumes." 

Maley writes:

"The typical gerrymanderer seeks to pack, crack and stack in order to win, but doesn’t care so much if the winning majority is 10% or 15%.  The fairness clause on the face of it requires a party to lose with as close as possible to 100% certainty if it polls 49.9% of the vote, but to win with similar certainty with 50.1% of the vote.  But at best, if an even split of the vote produces a 50/50 chance of victory for either party, a slight shift of the vote will only shift the overall probability of victory slightly.  The whole idea of basing a fairness clause on a deterministic rather than probabilistic model of the seats-votes relationship is therefore fundamentally misconceived."

When I asked for permission to publish the above, Michael also sent a further comment:

"[..] there’s one more point, slightly technical but illuminating, which you might also want to include.  The overall result in a single-member constituency system can be very well modelled using the generalised binomial (sometimes called the Poisson binomial) probability distribution.  This is like a binomial distribution, except that the  probability of “success” (which in this case means a win in a seat by whichever of the two parties we choose to analyse) varies from draw to draw.  (See http://en.wikipedia.org/wiki/Poisson_binomial_distribution for details.)

The interesting thing about this distribution is that its variance is maximised when the probability of each “success” is ½; in electoral terms, when every seat is marginal.  The insight into the fairness clause which flows from this is that the seats-votes relationship can only be even predictable if this variance is reduced by creating a significant number of safe or fairly safe seats.  But when you have a legislature with as few seats as South Australia, that means that the absolute number of marginal seats has to be quite small, and once that is the case, the operation of the seats-votes relationship around the 50/50 split of the total two-party preferred vote is highly susceptible to disruption via targetted “sandbagging” of a handful of seats.  This is yet another reason why the fairness clause is a truly nutty idea in a place like South Australia."

The following links written by Jenni Newton-Farrelly (SA Parliamentary Research Library) are very useful for those interested in detailed discussion of the fairness clause and the issues it raises. The third goes into the historic background behind the formation of the "fairness clause" in far more detail than my article above:

http://www.law.unimelb.edu.au/files/dmfile/WP_9_Farrelly4.pdf 

http://www.law.unimelb.edu.au/files/dmfile/WP_6_Farrelly2.pdf 

http://www.law.unimelb.edu.au/files/dmfile/WP_4_Farrelly2.pdf

Update March 5: A Galaxy poll of the seat of Adelaide shows this won't be the easy recapture that Labor was dreaming of - the Liberals are up 54:46, off a primary of 49%.  Incumbent Rachel Sanderson has an excellent netsat of +25.

More polls March 13:  In an encouraging sign for Labor's chances of an unrepresentative win or at least a hung parliament, Galaxy is reporting slim 51:49 2PP leads to Labor in Newland and Mitchell.  Full figures are not yet available.  If the 2PP was 54:46 statewide then these two seats would be expected to be right on the wire and to most likely split 1-1, but these two polls imply the chance of holding both to be more like a half than a quarter (even without considering the Hanna factor), and just slightly improve Labor's prognosis if the 2PP is around that level.  Alternatively, in the absence of other state polling, these results might be taken as weak indicators of possible state improvement.  Either way they make the Liberals' position a little bit trickier than before, though I still think they will get there (if maybe only just).

Newspoll March 14: Figures from the first 1000 respondents show the Liberals leading only 53-47 and that this is still very much a live contest with a significant chance the Liberals won't get a majority.  

Apologies for the font inconsistencies in this article.   

Election Day March 15: I will be too focused on Tasmania to say anything much more until tomorrow but Newspoll has come in with a 2PP of only 52.3 to the Liberals, while ReachTEL's final is a crushing 55:45.  The good news for the Liberals in the Newspoll is that as with the earlier Galaxy the swing appears to be higher in Adelaide, giving the party good chances of cleaning out many Labor's marginals with a 2PP that otherwise would probably not be enough.  Of course the ReachTEL result would be an easy (if perhaps underwhelming in seat terms) win.  As I mentioned above, the idea that 54:46 would deliver a hung parliament was far-fetched, but some modelling I have seen by statistician Paul Eckermann that extends the basic principles of my analysis above, has confirmed that the deck is stacked in Labor's favour if the swing is not regionally skewed.  Ignoring the issue of the distribution of the swing, the Newspoll 2PP would make for a roughly even contest.  But the increased swing in the city areas suggests even the Newspoll result implies a probable Liberal majority.

My expectation is that even if the 2PP is modest, the Liberals will punch above their pendulum weighting, cancelling out what happened last time, and win the election.  But it is not the foregone conclusion that the betting markets have had it as and if the 2PP is more like Newspoll's than ReachTEL's it may yet be an interesting night.  You'll just have to enjoy it without me.  :)

March 16: Oh dear, it seems to have happened: a $1.01 favourite has failed to win a majority thanks to a combination of a modest 2PP vote (as per the extremely accurate Newspoll) and good sandbagging; barring late-counting surprises SA could be heading for 23-22-2 in Labor's favour, and a hung parliament in which Labor may still continue in office.  The 2PP was about 52.5% which was well within the zone in which anything could happen.  And it has.  

Further coverage of this surprise can be found on my post-count article.