Recent Newspolls Do Not Prove That A Hung Parliament Is Likely

Today's Australian reports claims by federal Greens Leader Adam Bandt that recent Newspolls suggest Australia is headed for another minority government situation similar to that which occurred (for the first time in many decades) in 2010.  It's a typical case of using polls for a purpose for which they are neither intended nor fit.  

The most relevant quotes:

"The Greens leader used ­research from the parliamentary library to argue that Labor’s only chance of forming government, based on the past six months of Newspoll, is to gain the support of the Greens and independents on supply and confidence.

The research, based on a uniform swing in each electorate, predicts Labor would have reached 75 seats on its two best polls this year – one seat short of a majority government.

“The maths just says we are heading towards a power-sharing parliament; (there) is a swing against the government,” Mr Bandt told The Weekend Australian. “It shows Morrison being pushed out of majority government but not enough for Labor to win in its own right.”

This article is not about the research, which I haven't seen.  It is about Bandt's interpretation of the stated result (ie that Labor would not have won a majority based on uniform swing in any of this year's Newspolls).  

The Newspolls conducted so far this year have shown the Coalition on between 48% and 50% two-party preferred.  It is true that if there was a uniform swing, Labor would not win government outright with 52% 2PP, nor the Coalition with 50% 2PP.  However, swings are never uniform, and are sometimes significantly non-uniform.  So it's quite possible either party would win outright at their end of that 2PP range.

The bigger problem with the argument is that it ignores that polls, at best, are snapshots, and should not be treated as predictions until very close to the election.  The state of polling between around 19 and 25 months into a term (the yardstick implied above) is not a very good predictor of the election outcome because ... Things Happen!  In 1991-2 Labor was still going through the painful removal of Bob Hawke and the fallout from that.  In 1994-5 the Coalition was churning through dud leaders.  In 2000 we hadn't even got to "mean, tricky and out of touch", let alone Tampa and 9/11.  In 2003 Simon Crean was still Labor leader, in 2006 Kim Beazley was.  In 2009 Kevin Rudd, his own position and re-election seemingly assured, was pummelling Malcolm Turnbull until Turnbull was replaced at the end of the cycle.  In 2012 Labor under Julia Gillard was polling awfully and Rudd had yet to be returned.  In 2015 the sample was mostly Tony Abbott with a bit of Turnbull at the end, and in 2018 the sample was all Turnbull, who was soon to be removed.   

Another factor in the final case specifically was polling error: at the end of the 2016-9 term polls were again in the same 47-49% range as they had been in the later mid-term range, but the final polls were wrong by 3%.  Even if the recent Newspoll range of 48-50% was the final range going into the election, it would be unwise to be overconfident about the 2PP actually landing in that range.  Despite the best efforts of YouGov the 2019 failure might repeat in one direction or other, which would result in a Labor landslide at the low end of the range and a much increased Coalition majority at the high end.  

This two-point range of relevant Newspoll 2PP results in the past six months has been narrower than normal because (i) Newspoll polls less often than it used to, and (ii) Newspoll has been less volatile for various reasons (including increased sample size) since YouGov started running it.  In the last three terms, the spread of Newspoll 2PPs over the same six-month period has been five points (because of the Turnbull ascension), two points and two points, for an average of three.  In nine terms under the old Newspoll, the average 2PP spread for the comparable six-month period was nearly seven points.  Even so, the eventual 2PP still fell outside the entire Newspoll 2PP spread for that period in 1993, 2001and 2010. 

Therefore to usefully comment on how predictive Newspolls over this sort of time period really are, it's better to look at the six-month average rather than the spread of values.  The following graph shows the relationship between the average of Newspoll 2PPs this far into the term and the next election result:

If a predictive relationship even exists here, it's weak.  An r-squared of 0.07 isn't remotely near statistically significant evidence off only 12 data points, so it's doubtful that there's any relationship at all.  Here are the data:

The average change from the state of Newspolls at this time to the final result is 2.3%.  Based on the long-term average change, if it was a real relationship it would imply a 50% chance that the Coalition 2PP at the next election will be somewhere between 46.8 and 51.4.  That range would include some majority governments for either side even assuming the swing is exactly the same in every seat, but there would also be a 50% chance that the 2PP will be outside that range, an automatic majority government for one side or other assuming uniform swing.   

2.3% might sound OK but the average state of play of Newspolls at this stage of the game is actually a worse predictor than assuming the final 2PP will always be 50-50, which for this sample has an average error of only 1.6%.  The polling average at this stage has predicted the wrong 2PP winner six times out of twelve.  

There's more.  Governments have on average improved by 0.6 points, but this can be divided into Coalition governments (5 out of 6 improved, average gain 1.4 points) and Labor (3 out of 6 improved, average loss 0.2 points).  Six of each is of course a trivially small sample size, but this is consistent with the pattern in other historic comparisons of 2PP polling with results: the Coalition has more often overperformed.  

On the other hand, there has been a relationship between the level of leadership chaos and how good or bad late mid-term Newspolls are as a predictor.  Assigning 4 points for a change of Prime Minister after the poll period, 3 points for a change during the poll period, and 2 and 1 points for Opposition Leaders respectively to form a leadership chaos index, it has quite a strong relationship with how bad a predictor the polling at this stage was.

So we might expect that if both parties keep their current leaders then the outcome is more likely to fall near to the current polls but (i) this is getting rather overfitted off just a dozen data points (ii) we don't even know if they will.  

Swings aren't uniform

Even if voting intention doesn't change, swings are never truly uniform and are often meaningfully non-uniform.   Assuming Craig Kelly's seat is reclaimed by the Coalition, all crossbenchers hold their seats, no new crossbenchers win, and that there are no further changes in the redistribution process, then the Coalition would need about 51.3% 2PP for a majority given uniform swing (Stirling is abolished and on some estimates Chisholm becomes more marginal than Bass), while Labor would need 52.4%. (Regarding new crossbenchers, I assume that Wentworth reverts to safer Coalition status after being disrupted by a by-election party change).  However, some relevant considerations here are:

* Labor's marginals are on average closer than the Coalition's.  This is subject to the redistribution still ongoing in Victoria, but the average notional margin in Labor's ten closest 2PP marginals looks like being around 1.4% while for the Coalition it's about 2.6%.  Because swings are never actually uniform between seats and at best resemble a random distribution, if there was actually more or less no swing this factor alone would be expected by chance to roughly cancel out the Coalition's loss of Stirling and Labor's gain of Hawke in the redistribution.  

* Personal vote effects from the 2019 result very slightly favour the Coalition.  In its ten closest 2PP marginals, the Coalition has a double personal vote effect (its incumbent defeated a Labor incumbent) in four seats, a single personal vote effect (a new Coalition MP at the end of their first term) in two, and a vacancy in one (Boothby).  Labor has a double personal vote effect on its side in two of its ten seats, a single effect in two (Gilmore was a government vacancy) and possible complications in Eden-Monaro where its current incumbent has served a partial term and replaces a very popular previous incumbent. 

* We also sometimes see elections where the swing is non-uniform in a way not predicted by random variation.  2019 was a striking example - the Coalition got strong swings to it in its own marginals on average, while swings against it were wasted in safe seats.  

Finally, there has been a historic pattern that oppositions generally don't win without first getting large polling leads.  That's not to say it cannot happen - the pattern could have an exception, or the large leads could be still to come (as they were at this point in the 2004-7 cycle).  

None of this is to say that a hung parliament is impossible or hugely unlikely.  The increasing size of the crossbench is increasing the chance of it, with one actual case and two very near misses in the last four elections.  But we're dealing with two rather rubbery relationships at the moment: (i) between Newspoll 2PPs this far out and final actual 2PPs (ii) between the pendulum conversion of a 2PP to a seat count and the actual conversion.  There isn't any basis on the current evidence for concluding at this stage that a hung parliament is more likely than not.  There may not be such a basis even on election eve. There wasn't even in 2010 (the one time when it actually happened) as the polls on the whole implied a reduced Labor majority as the more likely result.  

Poll spinning 

Poll spinning by parties to try to drive up supporter morale, get media attention or even just provoke other parties is common behaviour and the Greens are especially prone to it.  Probably this sort of exercise is better than simply stating flatly that one's party will win a majority and citing no evidence whatsoever that this is even remotely likely (as Anthony Albanese has done in response) but I really don't see a need for parties to engage in this sort of thing at all.   It's more credible to just say that one is hoping for a certain outcome and has good reasons to believe it's a real possibility, but even that carries some risks and it may be better to just let good results do the talking.  

While the Greens habitually get the balance of power in the ACT, at state and federal level the last 13 consecutive elections at which they have won seats have all seen majority governments elected. This is although at least half of those saw lead-up speculation about a possible or supposedly probable hung parliament where the Greens might hold the balance of power.   In Tasmania at least, and I think also perhaps in Victoria in 2018, this "hung parliament club" stuff (my term for any narrative that maintains that minority parliaments are both clearly preferable and perpetually likely) has seemed to frighten the horses.  Voters open to voting for either major party may respond to the prospect of a Greens-supported Labor minority government by backing whichever major party seems most likely to win in its own right.  That's even though most possible minority scenarios would provide Labor with easy ways to work around any Green demands, since there are very likely to be more non-Greens on the crossbench than Greens.  In any situation in which Labor needs the votes of nearly all the current crossbench including the Greens to form government, it probably won't be forming government anyway.