Thursday, September 23, 2021

Poll Roundup: Why Is Resolve Diverging?

I last wrote a federal poll roundup in early August.  At that time I noted that the implied 2PP polling numbers of Newspoll, Morgan, Essential and Resolve this year had been more or less identical.  No new Essential voting intention polling has been seen since mid-July though another dump may appear within days.  

This is what we've seen since the last roundup:

* Newspoll (late August) 54-46 to Labor (Coalition 36 Labor 40 Green 10 PHON 3 others 11)

* Newspoll (mid-Sept) 53-47 to Labor (Coalition 37 Labor 38 Green 10 PHON 3 others 12)

* Morgan (mid-August) 54-46 to Labor (Coalition 37.5 Labor 37.5 Green 12.5 PHON 3.5 others 9)

* Morgan (late August) 54.5-45.5 to Labor (Coalition 37.5 Labor 38.5 Green 11.5 PHON 3 others 9.5)

* Morgan (mid-Sept) 52.5-47.5 to Labor (Coalition 38.5 Labor 35 Green 13 PHON 3 others 10.5)

* Resolve (mid-August) Coalition 40 Labor 32 Green 12 PHON 2 IND 10 others 3.  Last election 2PP would be very roughly 50-50 (my formula gives 50.3 to Labor) 

* Resolve (mid-Sept) Coalition 39 Labor 31 Green 10 PHON 4 IND 9 others 7.  Last election 2PP would be very roughly 51-49 to Coalition (my formula gives 51.1)

Now, it's true that Morgan's use of respondent preferences is inflating its leads for Labor (in the last three polls by an average of 0.9 points compared to last election preferences).  And it could be that the independent voters being (over-)reported by Resolve are actually much more left-leaning than the ones who split about 60-40 to Labor on preferences at the last election.  But still, one of these polls is very clearly not like the others.  Resolve has shown two consecutive Coalition primary leads of 8% while polls by other pollsters have had an average lead of minus 0.5.  Two Resolve polls in a row have come out as massive, beyond margin-of-error outliers. 

Unusual properties of Resolve

I have discussed Resolve's apparent overestimation of the Independent vote before.  Part of this would be caused by offering Independent as an option everywhere, but because the Independent vote is heavily concentrated (as noted last time) it would mostly be caused by other things.  Unless the 6% or so indicating "independent" despite not having voted that way in 2019 are all former Labor voters, however, there seems to be more than just the issue with independents going on here, because Labor is running 6.3 points lower in Resolve than the other polls.

Attention has focused on some of the other unusual properties of Resolve.  Firstly, unlike other current pollsters, Resolve does not provide an uncommitted option on voting intention.  The voter must choose an option to proceed in the survey and receive the small credit points reward that awaits them at the end of it.  This may result in undecided voters choosing "independent" by default.  The claim is that this represents the actual voting process, but it doesn't entirely, because 2-2.5% of voters will deliberately vote informal.  

Secondly it appears that Resolve asks voters to fill in a simulated ballot with full preferences:  

"A core set of political questions is asked each month, including Federal and State voting intention (of registered voters only) using ranked preferences [..]" (source)

"“While we do collect people’s full preferences in a numbered ballot paper, it is their first choice that is the most important and reliable indicator of who truly likes you and wants you to govern,” says Reed." (source)

It would be helpful to know how many options are provided on each ballot (I suspect more than actually run in most divisions), whether the ballot is rotated as opposed to, eg, the Coalition always being on top, and if there is a single ballot order what is the rate of donkey voting.  Filling out a full preferential ballot is the sort of task a busy survey recipient answering questions for meagre rewards is likely to satisfice.  It also doesn't replicate the voting experience for a fair proportion of major party voters, approaching half of whom copy their party's how-to-vote card.  

A further unresolved (or should that be un-Resolved) methods issue is how Resolve handles three-cornered contests, or Liberal voting in seats contested only by Nationals.  Its results are being presented with the Coalition primary simply showing as "LNP".

I'm still not convinced that any of these methods issues are enough to entirely explain the recent differences between Newspoll/Morgan and Resolve.  It will be interesting to see whether the next Resolve poll continues the divergence. If anyone is polled by Resolve (it is an online poll which uses a variety of panels, and it is not clear yet whether participants would know Resolve was the pollster) it would be useful to see screenshots of how it operates.

On the question of which to believe, the problem with Resolve's finding of a crash in the Labor primary vote in the last two months is that there is no obvious reason why that should occur to the extent suggested.  The heat has been largely on the government over the NSW and Victoria lockdowns and the contribution of the vaccine rollout to them.  It seems a strange time for an Opposition primary vote to crash.  

Others surge

A modest rise in the polled vote for "others" (in this context meaning candidates other than Coalition, Labor, Greens and One Nation) has been noticeable in Newspoll and Resolve, but not Morgan, in polls taken in the last six weeks.  Across the three pollsters this category has grown by about 1.7%.  

In my view this increase is probably a result of bulk advertising by the United Australia Party, which comments from Resolve and Essential suggest is now polling a few percent, probably at the expense of the major parties.  (There is a suggestion this could be from Labor especially, though the logic behind that isn't obvious.)  This party's vote tends to die down to 1% or less between elections and flare up again close to polling day or when it is campaigning heavily.  Left-wing posters on Twitter have attributed the rise to such forces as "Voices of ..." independents, but I don't believe that there is any more buzz around independents yet than there was by the end of the previous cycle.  Media continue to run puff pieces about independent prospects, but winning new crossbench seats is hard, and mainly occurs when seats are vacant or have very tarnished incumbents.  


The current Newspoll showed Scott Morrison on his lowest net rating since March 2020 (-4, 46-50, incidentally about the same as US President Joe Biden, but Biden is newer to his position).  However Anthony Albanese also hit negative double figures for the first time (-11, 37-48).  Morrison also had his smallest lead on the skewed Better Prime Minister indicator since March 2020 (47-35).  That is still a much larger lead for a leader whose party trails 47-53 than normal.  This sort of discrepancy was even higher in the previous poll, in which Morrison's 16-point lead was the largest except for John Howard leading Kim Beazley by 21, 22 and 24 points while trailing 46-54 on two-party preferred in late 2005.

Resolve and Essential, which both seem to push undecided respondents less than Newspoll does, both have larger leads for Morrison on Better/Preferred PM (45-26 and 47-26).  Essential tends to have milder ratings for both leaders compared to Newspoll, while Resolve tends to have worse ratings for Albanese specifically - perhaps because of whatever is also causing it to register Labor with a very low primary vote.

While Morrison's positive net satisfaction Newspoll streak ended recently, another streak continues - he has had a higher net satisfaction rating than Albanese for 24 polls in a row.  I have added a section on comparative netsats to the Federal Newspoll records page but it is worth noting that the longest winning streak on this indicator was held by Kim Beazley against John Howard (51 polls), without Beazley ever winning an election.

The overall picture - does it mean anything?

While Labor still lead comfortably in two of the three polls, the lead has dropped in all of them, by an average of about 1.5 points in the current cycle.  A straight aggregate of the recent polls would have Labor's lead down to 51.5-48.5, but adding in assumptions relating to house effects, sample sizes and the questionable reliability of Morgan (historic inaccuracy) and Resolve (untested) could bump that up into the 52s.  In any case it's not a very big lead at the moment, and for now the blowout that had been taking shape has stalled.  We have not yet seen the Opposition lead move into the high 50s, from which incumbent governments have only sometimes recovered.  

Polling at this stage just isn't very predictive.  Another excellent Armarium Interrata post covers a lot of the ground here: historically the 2PP at this point (assuming a March to May 2022 election) is on average wrong by about 3 points.  Governments tend to go through bad patches between about a month and a year prior to elections.  Whoever is behind at this stage tends to gain (even when the lead is small) and at this stage the polled 2PP is historically a less accurate predictor than just assuming the result will be 50-50.  

Is this just a result of old or dodgy polls fouling the mix?  Not at all.  For instance let's assume the net election is about seven months away (late April) though it may be six or eight is more likely, and a snap election in November seems highly unlikely but isn't completely off the table.  Considering the average of the two most recent Newspoll 2PPs (derived from previous-election preferences where necessary) seven months prior to elections in the Newspoll era, incumbent governments have been behind, 48-52 or worse, in every election cycle except 1987 (tie), 1990 (tie), 2010 (ahead) and 2016 (ahead).  Yet five of the eight governments that trailed were re-elected, and four of those (1993, 2001, 2004, 2019) won the 2PP by more than both of the governments that were well ahead (2010 and 2016.) 

In this time the government on average gained 1.5 points by election day, but in five of the twelve cycles the government gained more than 4 points.  (In the 2019 case, mostly through polling error.)  The side trailing on average gained 3.3 points and only went backwards (slightly) once.  The average shift to the final 2PP was 2.9%.  Assuming the final 2PP would be 50-50 had an average error of only 1.6%.    There's a slight average tendency for governments to do better if they're ahead at this stage and worse if they're behind, but it isn't remotely statistically significant and the slope line is weak, because of the narrowing effect.  (It was mockingly called the Narrowing in the 2007 leadup, but it's actually clearly a thing.) Finally, you may as well have flipped a coin as assumed that the party ahead on 2PP at this point was going to win the election - half the time they did, and half the time they didn't.  

At the moment I think the polling only shows Labor to be competitive.  Being ahead by a modest margin does not mean an especially large chance of winning.  It is still better than not being ahead though, because oppositions that are level or behind this close to an election or closer almost always lose.  Even there there are exceptions - the Chifley government in 1949 was ahead until July (at which point a coal mining strike appears to have damaged it). Also while the McMahon government in 1972 is usually portrayed as long-doomed it was actually ahead on a 2PP basis between July and September, losing the lead only in the last two months.  

No comments:

Post a Comment