Tuesday, September 10, 2019

Did A Late Switch-Off From Shorten Cause Labor To Lose?

(Note for Tas readers and anyone else interested: Scott Bacon recount thread is here)

Nearly four months after the election, Labor and its supporters are still having great trouble working out what happened.  Ahead in the (faulty) polls for an entire term, well ahead in them for much of it, Labor managed to lose to a government that had seemingly imploded nine months earlier.  There are basically three possible explanations.  The first is that Labor should have won the election, but that at least some central parts of its policy platform were wrong.  The second is that Labor should have won the election and that its policies were sound, but it was let down largely by tactical issues.  The third, about which little has been said, is that Labor could not have won anyway.  (The idea here is that voters no longer care about governance scandals or internal party turmoil so long as they like the PM and the basic way that he is leading.)

A version of the second theory - and by the way, I don't subscribe to any version of the second theory - says that Labor's policy mix was OK but Labor was undone by spurious "death tax" scare campaigns and a massive advertising spend by Clive Palmer against Bill Shorten.  (Those arguing this tend to oversimplify things as if the United Australia Party did little in the campaign but attack Shorten.)  Adherents of this theory seem to have taken succour from findings of a recent ANU study that has been reported as finding that Labor lost because the Coalition made net primary vote gains in a volatile environment during the campaign, and also that a big part of Labor's failure to do likewise was voters switching from Labor to other parties because they became more negative towards Bill Shorten.

Bill Shorten was not a popular Opposition Leader and may well have been electoral lead in Labor's saddlebags all along, although it's not that easy to test that.  But how strong is the evidence that voters deserted Labor in droves during the campaign because their view of Bill Shorten became more negative, whether because of attacks on Bill Shorten or otherwise?  Well actually, looking closely at this study, it's much weaker than the media reports are making it sound.  That's not to say Shorten himself wasn't a major factor in Labor's defeat, just that the evidence that he only became a problem at the end does not stand up.

Methods

The survey, among other things, resampled 1692 voters who had given a voting intention (which in a few percent of cases was uncommitted) at some stage during April 2019.  In June 2019 it asked these voters whether they had voted and if so who for.

Some general limitations of such studies should be mentioned.  Firstly, studies involving voters saying how they voted can be unreliable.  There's a vast literature on this going back decades (example summary here).  While recall errors can be linked to memory, especially for voters whose votes jump about a lot, they can also be linked to the voter's present opinions or to voters over-reporting that they voted for the winner.

Secondly, the sample is, after all, another poll at a time when polls have been in trouble.  While its sample size is decent, it is still smaller than some of the major polls that failed.  Its final results (scaled) for voters who did vote were Coalition 42.2 (actual 41.4), Labor 35.4 (33.3), Greens 13.7 (10.4), Others 8.7 (14.8).  This places it way outside the theoretical margin of error for both Greens and Others, and while that's not necessarily a major problem, over-representation of Greens voters is always a worry in terms of potential over-sampling of politically engaged or educated voters generally.

I also want to discuss closely where the Bill Shorten finding comes from.  Voters were asked (Fig 2 p 14) to choose "all that apply from a list of eight pre-specified options, and an ‘Other (please specify)’ category" (my bolding) as reasons why their vote changed.  The following were the pre-specified options:

* Your view on the Prime Minister Scott Morrison changed
* Your view on the Opposition Leader Bill Shorten changed
* Your view on your local candidates changed
* Pressure or advice from your family or friends
* A policy announcement or announcements by the Government
* A policy announcement or announcements by the Opposition
* Your individual circumstances changed
* You moved electorates

My concern about this wording is that limiting the policy aspect to "a policy announcement or announcements" is quite narrow.  It might well focus the respondent's thinking on announcements made during the campaign rather than encouraging them to consider policy more broadly.  For instance if a voter's reason to switch was that they became convinced a policy Labor had announced some time ago was bad, would they necessarily pick this answer?  Or what about if during the campaign they became convinced that Labor could not manage a particular issue area?  I suspect the question design results in an underestimate of the impact of policy on vote-switching.  Given that it's an all-that-apply question, that might not necessarily affect how often voters cited views of the leaders as a factor, but it could well affect the prominence of issues as a factor in vote-switching, including in the subsequent media reporting.

(Peter Brent has reminded me of another point I meant to include in this article but forgot: voters' claimed reasons about why their vote changed could well have been coloured by the aftermath of the election - a point especially applicable to Shorten since he is being widely held responsible for the defeat. After all it's much easier to say your view on Shorten changed than that you were sucked in by a death tax scare campaign.)

Effect sizes

Even ignoring all the above potential issues, the major problem with the media and social media claims about changed views of Shorten blowing the election is that the effect sizes do not bear it out.

32.1% of the April sample intended to vote Labor at that time.  Of these 20.8% said they didn't actually vote Labor, so that's 6.7% who switched away from Labor.  The breakdown of switchers away from Labor was 39.6% Coalition, 27.4% Greens, 7.6% other and 25.4% didn't vote.  At the election 82% of Greens voters preferenced Labor anyway, and if a voter shifted from Labor to Greens then they are probably even more likely to have preferenced Labor, so those switching to Greens would have had very little impact on the 2PP outcome.  44% of Others voters preferenced Labor, but possibly those who switched from Labor during the campaign would have been more likely to do so than those who never intended to vote Labor at all?  Also, a Labor voter not voting only has half the 2PP impact of one defecting to the Coalition.  All up, in this particular sample Labor's total 2PP loss caused by vote-switching was probably not more than 3.9 points.  That's still a lot, but in this particular sample there is a lot of volatility, and all parties are losing points to someone.

We also have a percentage of previous-Labor-intenders citing Shorten as one of their reasons for shifting: 27.7%.  We don't have crosstabs to show whether these were more likely (compared to other switchers from Labor) to switch to the Coalition, the Greens or not voting.  If they were typical, then Labor's net 2PP loss to which perceptions of Shorten contributed drops to about 1.1 points.  This is already not enough to swing the election as only two Coalition seats (Bass and Chisholm) were won by less than this.  But furthermore, it's a multiple-answer question, so some of these voters may have switched for other reasons even without the Shorten factor.

Furthermore it was not only Labor voters who reported switching because their view of Shorten changed.  This was also reported by 14.8% of intending Coalition voters.  Coalition voters who switched (16.5% of an initial 36.9%, so 6.1% of the initial total) reported switching to: Labor 37.2% Greens 16.8%, others 21.2%, didn't vote 24.8%.  But - and this is one of the most crucial issues with this whole thing - while a Labor supporter citing Shorten as a reason to switch from a Labor primary vote could well have still preferenced Labor, it would make no sense for a Coalition supporter citing a changed view of Shorten as a reason for switching their vote to still preference the Coalition.  It wouldn't make that much sense for them to not vote either, although this is still vaguely possible especially if the respondent is giving dishonest reasons for not voting.

Most likely, nearly all the 0.9% 2PP worth of former Coalition voters who cited a changed view of Shorten as a reason for switching from the Coalition, would have either voted Labor or preferenced Labor.  At this point the likely maximum negative impact of Shorten-caused switching on Labor's 2PP vote is dropping still further, to perhaps a few tenths of a point.  And, in a survey of this size, a few tenths of a point is statistical noise.  It's something like five voters.  And that's before we get into any 2PP gains Labor might have made from voters for the Greens or Others who were initially intending not to even preference Labor but were impressed by Shorten during the campaign.  (That's not as silly as it sounds - for instance, Shorten did try to show much more character during the debates than his previous wooden performances.)

This asymmetry applies to Scott Morrison too.  The survey suggests changing views of Scott Morrison were cited in very similar proportions by Labor defectors and Coalition defectors.  Labor defectors were very slightly more likely to cite changed views of Morrison, and as noted above there were also slightly more Labor defectors.  But the important point again is that a Coalition defector who went cold on Scott Morrison may well have not voted, or if defecting to Greens or Others may well have preferenced the Coalition anyway.  But a Labor defector who warmed to Morrison has no reason not to vote, and no reason to defect to Greens or Others while still preferencing Labor.  On this basis, it's unlikely that changed perceptions of Shorten were a major 2PP drain for Labor - it's much more likely that changing perceptions of Morrison were a handy 2PP asset for the Coalition.

Fitting the pattern

There are also good reasons why we should regard the theory that Labor was sunk by voters turning away from Shorten (a theory that I should make clear is not promoted by the study itself, but rather in interpretations of it) as unlikely on the face of it.  Firstly, Shorten was not new; he was a known commodity who had been leader of the party for almost six years including a previous election.  Secondly Opposition Leader personal ratings have historically shown a very weak relationship if any to voter intention (and that applies through the times when Newspoll was accurate, so let's not dismiss that too lightly on account of the 2019 polling failure.)  Elections are usually referendums on governments and Prime Ministers.  The 1993 election was a clear exception and in that case a relationship did develop between the Opposition Leader's personal rating and voting intentions.  This one is not so clear.  Polls underestimated support for the Coalition, but did capture around 1.5 points of movement towards it during April.  However, through this time Bill Shorten's ratings didn't change and basically hadn't changed since October 2018.

Vote shifting and poll failure

It's not relevant to the Shorten narrative, but I should also mention the study in relation to the polling failure.  The study has been used to support a narrative that the polls failed because of a high degree of voter volatility that may have contributed to the ancient enemy of pollsters, late swing.  But what the sample actually shows is not a big deal in 2PP terms:


In the figures above I have redistributed Don't know (April) and Didn't vote (May) and found an expected 2PP based on the actual election preferences.  The actual results are also shown for comparison.

The Coalition made greater primary vote gains than Labor in the sample.  However, Labor also gained - it is just that the raw table disguises this because there are more "Didn't vote" respondents in May than "Don't know"s in April.  Both majors supposedly gained at the expense of Others.  In 2PP terms Others (including UAP and One Nation) were favourable to the Coalition, so this only washes through to about a one-point 2PP shift over the sample period - which is about what the mainstream pollsters got.  The only new thing here then is the idea that the Coalition picked up votes from Others.  But because the results are so weird (with the very low Others vote in the May sample meaning that the ANU's "after" result is even less accurate on an average-error basis than it's "before") that looks a lot more likely to be a peculiarity of this specific study or its sample.  This is backed up by pollsters having been in general wrong not about the Others primary vote, but about the Coalition and Labor primary votes (except for Ipsos who were wrong about the Coalition and the Greens).

The study also cannot demonstrate late swing (eg in the last few days of the campaign after polls had left the field) because its baseline is in April.  And late swing is a theory that is up against the evidence anyway.

I hope this article shows that media narratives about elections based on single studies need to be treated with a great degree of caution.  I considered doing this article as part of a Poll Roundup but was long enough by itself and stood better alone.  A new Poll Roundup may or may not follow in coming days.

See also

Excellent article also drawing sceptical conclusions about the late-swing narrative by Murray Goot here.