Saturday, March 10, 2018

SA Election: Some General Modelling Comments

Note to media of all kinds: this long weekend (10-12 March) I am not available for in-person interviews.  My phone will be switched off most of the long weekend - you may be able to get me on Saturday morning or Monday night, or if you leave a message with an after-hours number I may be able to return your call at night on Saturday or Sunday.

Note to posters: Comment clearing may be slow and replies slower for the next few days.


The two-week gap between this year's Tasmanian and South Australian elections means I will at least be able to do live comments on South Australia.  However, the Tasmanian campaign hasn't done wonders for my ability to devote energy to the SA contest.  I may be able to do another piece with more detailed modelling on South Australia next week, but I'm not sure I will have time for this yet.  This piece just makes a range of general comments that I think are important to trying to model the outcome.

The Backdrop

Historically, the two most important factors in understanding state election results have been the age of the government and whether it is the same party as the power in Canberra.  (See What Kills State Governments: Age Or Canberra?) Young governments tend to be re-elected and governments that are the opposite party to the party in power in Canberra tend to be re-elected.  Here we have a government that is of the opposite party to the party in power in Canberra, and at a time when the federal Liberal government is unpopular, but the state government is also very old.  Labor have been in power in SA since 2002 and are going for five in a row.  The last two state governments to face an election at about this age (NSW in 2011 and Tasmania in 2014) were smashed.  Not since Queensland 1986 has a state government older than this been re-elected. 

After being thumpingly re-elected in 2006, Labor "should have" lost in 2010, but held on by an extraordinary feat of sandbagging that allowed them to confine their losses to just two seats despite an 8.4% swing against them.  The body charged with administering the state's infamous "fairness clause" concluded that Labor's win despite losing the 2PP 48.4% to 51.6% was down to campaign factors rather than an unfair distribution of votes, so didn't do anything about it.  Then in 2014 Labor lost the 2PP even more heavily (47:53) but clung on in minority.  The Liberals might have been expected to get more swing back in those seats that had been sandbagged in 2010 but were on the whole unable to do so. 

Later in 2014, in a humiliating result for the Liberals, Labor won the Fisher by-election by nine votes and gained a majority of one.  However this was then lost in 2017 when Frances Bedford quit Labor after being overlooked for preselection.  Labor still had a comfortable hold on government, and former Liberal leader Martin Hamilton-Smith had even defected to the crossbench and become a Minister during the term.  

This is the last election at which the "fairness clause" is expected to operate, as it was abolished in a classic piece of last-sitting-day skulduggery at the end of 2017.  The redistribution carried out under the fairness clause is designed to eliminate any skew in converting the two-party preferred vote into seats, so that whoever wins the 2PP should in theory win the election.  However with the possibility of the two-party system being dismantled at this election, it is high time this well-motivated but practically troublesome clause was done away with, before it might have had to face a test it could not even have made sense of.    

Labor needs a 3.1% uniform swing to win the 2PP in a majority of seats, but it should be noted that Antony Green's pendulum (which I am using as a base for the reasons he states) shows a cluster of seats around that margin.  This favours Labor as, if just random variation is taken into account, they should have a 50% chance of winning the 2PP in 24 seats on just a 2.6% swing (49.6% 2PP).  Furthermore, independents offer complications - for instance, Geoff Brock in Frome has backed Labor twice although coming from a Liberal-leaning electorate.

With Labor seemingly looking down the barrel in that the Weatherill government needs to improve its 2PP vote despite having been in government for a very long time, another lifeline has magically appeared.  That is that if SA-BEST take enough seats from the Liberals, the 2PP might not be very relevant, and Labor might lose the 2PP again but still form a minority government with SA-BEST or other crossbench support.

The SA-BEST Situation

Elections where third parties try to break into single-seat lower houses are very difficult to model.  We've had a couple of these in recent years - firstly the Nick Xenophon Team attempt at the 2016 federal election, and secondly One Nation in the 2017 Queensland election.  Both these attempts in the end yielded just a single seat (my forecasts were two and three respectively, but NXT had a few narrow losses and one of One Nation's target seats was snapped up by KAP.)

At the time Nick Xenophon announced he was quitting Canberra to go back to state politics, his intention was to hold the balance of power.  Polls in late 2017 suggested his rebaged party, Nick Xenophon's SA-BEST, was in a ludicrously strong position and might win government outright.  Following this there was a gradual upcreep in the number of seats SA-BEST is contesting, finishing at 36 of 47. (The exceptions: Adelaide, Black, Bragg, Flinders, Florey, Frome, Kaurna, Light, Stuart, Torrens, West Torrens.)  Oddly there is actually no real difference in the party's underlying support (based on NXT results in 2016) between the seats the party is contesting and the eleven it isn't.

Newspoll has been declining to publish two-party preferred results on the grounds of the high SA-BEST vote rendering them meaningless.  With a general perception that traditional pendulum methods are useless at this election, those trying to model it have often turned to Senate data instead.  Check the Tally Room thread which refers to Alex Jago's data derived from 2016 Senate results.  There's also an online model by Jack Larkin, based off statewide vote share, though it seems to give some rather exciting results for SA-BEST.

Senate data are useful but have some limitations.  Firstly, it is somewhat difficult to extrapolate from an optional preferential election to a compulsory preferential one, especially when the optional preferencing involves a very large number of parties.  Secondly, there is a discrepancy between Senate data and state election data.  When I derived notional two-party preferred votes from the 3PP figures posted by Jago I found that these 2PPs were over-compressed compared to the 2PPs from the last election.  That is, the 2PPs one might expect from Senate votes were less variable than those that had actually happened at the last SA state election, by almost a third.

One reason for this is that NXT votes have come disproportionately from strong Liberal areas (indeed if one compares past state 2PPs with Senate NXT shares the correlation is even stronger.) One might expect that they would then flow more strongly back to the Liberals as preferences in those strong Liberal areas, but actually, they don't.  At least as concerns Reps data (though I haven't checked Senate) there doesn't seem to be a link between how Liberal-leaning a seat is and how strongly NXT voters in it preferenced Labor.  So it seems that at the 2016 election, NXT acted as a gateway drug; some voters who had voted 1 Liberal before switched to voting 1 NXT 2 Labor.  But there's likely to be more to it than just that.  I think that maybe some safe seats at state level are parts of less safe seats federally and hence are more hotly contested across the electorate at federal polls.

I've constructed the basics of a model to try to convert state vote shares into an estimate of about how many SA-BEST seats might be won.  In the case of SA-BEST I've used the state seat 2PP results in the past to modify the Senate 3PP results, and then assumed that preferences will flow roughly as per the federal House of Reps election (where there were effectively open how-to-vote preferences between the big three).

This model still suggests SA-BEST could do quite well off its Newspoll vote shares, giving it around nine seats on average.  (Hartley isn't one of them but a Xenophon factor could be installed.)  However, the run of seat polling (for what it's worth) has had SA-BEST trailing or tied in three of these seats, leading to general scepticism about the party winning more than a few.  Back in January it had them winning one seat very marginally outside my win list, but it's possible the SA-BEST vote has come down a little since then.

In the coming week I hope to find time to add in components dealing with preferences based on how-to-vote cards for a more accurate model.  However the track record of such models tending to slightly overestimate third-party success means that I will probably end up agreeing (one way or the other) that SA-BEST are only on for a few to several seats and not more.

The Newspoll Problem

The most recent Newspoll created a slew of misleading reporting that the SA-BEST vote had crashed spectacularly, because it had fallen from 32% to 21%.  Such reporting is potentially a problem if it creates the impression that the party is running a disastrous campaign, and that impression then causes voters not to vote for it.  In fact, the party was polling 27% but Newspoll had rightly adjusted its vote downwards to 21% to adjust for the seats it wasn't contesting (whether by just multiplying by 36/47 isn't clear, though that would be reasonable in the circumstances - some other reports suggest they dropped SA-B from the readout where not running).  What adjustments were made to the other parties to compensate wasn't made clear.  (See Tom Richardson's report based on interviews with Galaxy.)

That's all I have time to add about SA for now, hopefully I can add some more detail during the week - another state poll would be nice!


  1. I'm curious here, the betting markets seem to have the ALP favoured, do you think this is fair?

    From where I'm sitting there are just too many variables to know for sure, from the SA Best preference flows, to the age of the ALP government, to the seats they need to win (and retain) to win the election seem to be pointing towards if not a minor Liberal edge, at the very least 50:50, down the line.

    If you think it's an apt description of the status of South Australia, could you give a bit of an explanation? Thank you!

    1. I think the markets are factoring an apparently high chance of a hung parliament and the apparent likelihood of that resulting in a Labor minority government with SAB support.

  2. Hmmmmmm... Intriguing...

    Two particularly intriguing things. First, Newspoll's hamfisted "explanation" of the 21% or 27% for SAB. If the average punter is asked "who do you intend to vote for?" I wonder how many think "Hmmm I'd like to vote for the Xenos but they don't have a candidate in my seat so I'd better not say that"? Without knowing that you can't begin to make a correction!

    Secondly, Alex Jago's numbers show that there is surprisingly litle variation from seat to seat in the Xeno vote. Whereas Lib and Lab support both vary from 17 or 19 to 51 or 52, presumably depending on the traditional factors such as wealth, self-identification by social class, employment status, or youthful desire for change versus senile fear of it, the Xenos are in the 20s in nearly all seats, with a few low 30s, and a 34, 35, 37 and 38. What makes a Xeno voter? I suppose disillusion with both major parties and finding their combativeness tiring, but is there anything else? And what explains their almost uniform distribution, with peaks in Finnis, Kavel and Heysen? PhD students, get your grant applications ready!

    My prediction - Lab to get more seats than the Libs, and Xeno to get anything between none and 30! Alright, trying to be a bit more definite, anything between say 4 and 18. Nothing will surprise me.

  3. usually any Government is lucky to make it past 3 terms......the problem with sa is that alp, sa best and the libs are each on 25% to 33% on a state wide basis....this probably means all will win seats......the change in boundaries means labor needs a swing of 2 to 3% to win in theory
    also it appears that sa best is losing votes..... what will they do with Balance of power if they have it? also what if Mr X is not the band decides without the bank leader?

    1. In theory, one would assume that it'd be a little bit easier for the ALP to get a swing of 2-3% in South Australia, than anywhere else - simply due to the enormous 2PP win the Liberals got last time, 55-45, and still lost. I hear your point though.

    2. Oh, my bad. Nonetheless, that's a huge 2PP to have without taking government.

  4. I live in Hartley and have literally been polled 15-20 times over the past few months, sometimes twice on the same night via both home and mobile phone. I also know people who are deliberately giving bogus answers to polls as they are sick of being asked so often. So not sure how much that affects the validity of poll results. In any case, interesting that Tarzia's odds have moved from $10 to $1.85 over the past month despite no poll results being released publicly. So either some people have strong gut feels or they are confident in the results on polls that haven't been made public.

  5. There has now been a poll released for Hartley and other seats. This puts Tarzia ahead of Xenophon 51-49 but is probably not worth much. A bit more surprising, but again maybe worthless, is that the SA Best candidate for the safe Labor seat of Taylor, Sonja Taylor, is also only trailing 51-49. Probably only means that there is an extra seat to add to what on earth is happening list of seats.
    Anecdotal evidence from Heysen is that the Liberal candidate has been adding a large amount of posters to already large amount so might think he is in trouble.

  6. I suspect using 2016 Senate data to model SABest votes/preferences this election is not very useful for a number of related reasons: 1) 2016 Senate was essentially a vote for Nick Xenophon himself who has broad appeal across the state - in this election people will be voting for a loose coalition of very diverse candidates. 2) These candidates come from both sides of politics and hence will be targeting different demographics and face different issues in each seat 3) The Briggs factor which inflated the Mayo 2016 NXT vote is no longer relevant, hence I suspect the SABest vote in Mayo seats has been overestimated for this election 4) The amount of resources SABest are putting into different seats appears to be very variable and I suspect many booths in seats they aren't targeting hard won't have any SABest presence on Saturday 5) Local issues will be a much bigger factor in many seats. eg the SABest candidate in Schubert is hardly known at all in that seat (he is from Kangaroo Island) so will presumably poll quite low, while I suspect the good SABest performance in the Taylor poll was in part a protest vote against Vlahos/Oakden since that was her seat. Hence that poll result may not be much help in predicting other northern suburbs seats.

  7. On federal figures would labor win Dunstan and Adelaide?

    1. I haven't got federal Reps figures for those seats but if you apply the difference between the 2014 2PP and the federal 2PP to them then yes.