Poll Roundup: What Is Going On With Newspoll Preferences?

2PP Aggregate (Last-election preferences): 52.4 to Labor (-0.4 since last week,  -1.2 points in five weeks)
With One Nation adjustment, 51.8 to Labor 
Closest position since May last year
Labor would very probably win election held "right now", but would probably have a small to moderate majority

[Updated on 26 April at bottom of post]

------------------------

Normally I issue Poll Roundups every second Newspoll. Since the last one there's been the slight distraction of the Coalition's 30th consecutive 2PP Newspoll loss. This week's Newspoll was number 31 in a row, just two shy of equalling the longest losing streak held by Labor under Julia Gillard, but it was more significant for the discussion it has sparked about (i) Newspoll's preferencing methods (ii) the prospect of a Coalition recovery.  So firstly, a detailed discussion of the preferencing issue, and then a slightly shorter discussion of where things now stand in federal polling.

Newspoll as a brand name linked to a standard set of questions has a 33-year history in Australia.  However, in 2015 the company previously doing Newspoll for the Australian was wound up.  The Australian retained control over the brand, but its operation was contracted out to Galaxy Research (which in turn was recently acquired by YouGov).  While the questions didn't change when Galaxy took over, the methods did, with a mixture of robopolling and online polling replacing the old Newspoll's landline-only live phone polling, and an increase in average sample size.



Over time it's been apparent that it's not only the methods that have changed, but also some properties of the polling.  In the leadup to previous federal elections, Galaxy's polls (which were very accurate) had been noted for their uncanny lack of volatility.  Once Malcolm Turnbull's honeymoon bounce as PM was out of the way, it soon became clear the new Newspoll shared this characteristic.  As former Nielsen pollster John Stirton mentioned, Newspoll has been so steady (with an average poll-to-poll 2PP change since the election of just 0.7%) that it has been hard to avoid speculating Galaxy does something to reduce the risk of rogue results to below what would be expected randomly.

But now, there's more, because it appears to be the case that Galaxy has taken Newspoll away from the purely last-election-based preferencing method that the brand had used continuously from 2005.  This follows similar behaviour by the pollster in the leadup to the 2017 Queensland election, in which Galaxy-run statewide polls used averages of preference flows from multiple previous elections to try to predict how preferences might flow.  One consideration here was the unusually pro-Labor behaviour of Others preferences in Queensland in 2015, and another was that Queensland had switched from optional to compulsory preferencing.  The changes helped, but they were generally not very clearly documented.

Here's a graph of the 33 federal Newspolls this term so far in sequence showing the difference between the published Coalition 2PP (which is rounded to the nearest whole number) and the preferences that I get if applying the 2016 election preferences to the published primaries to two decimal places:

The first two Newspolls in this sequence were taken before the 2016 election preferences were available, and when the third was taken they were only just available.  After those three, the average difference was a probably trivial 0.11 in favour of the Coalition until Newspoll number 26 (early December 2017).  From number 26 onwards, the average difference has been 0.77 points.

On a poll-by-poll basis, even differences of up to 1.2 points between the published 2PP and the expected 2PP off the published primaries may not mean anything. This was discussed in microscopic detail in this Wonk Central piece.  Say the published primaries in a poll would normally give 52.1 to Labor, but the published Labor and Green primaries had actually been rounded up by nearly half a point, and the published Coalition and Others primaries had been rounded down by the same amount.  That could mean the 2PP from the unpublished, pre-rounded, primaries was really 51.45, and on that basis the published Labor 2PP would be 51.

What we have in the case of the current Newspoll, is published primaries that would normally give a 52.4% 2PP for Labor, but a published 2PP of 51.  Rounding can't account for a gap quite this large by itself, but at a big stretch some combination of rounding and state-by-state preferencing impacts might do it.  Even so, it would be an extremely rare event (and that's assuming any pollsters still use state-by-state preferencing, and I cannot say for sure that any do).

What's happened since December is a sudden series of mostly large differences in the Coalition's favour, commencing immediately after the Queensland election at which One Nation preferences flowed strongly to the Coalition.  (Also, commencing the week before it was announced that YouGov had acquired Galaxy, for anyone who might think there could be anything in that.)  As William Bowe notes, what we've seen is what we would expect to see if Galaxy had suddenly changed the One Nation preference allocation from 50.5% to Coalition (the 2016 value) to something like 60%.  That also exactly matches my average difference estimate above.  I have seen reports that Dennis Shanahan said Newspoll had actually switched to respondent preferences, but at this stage I doubt it has.  Indeed I've had what I am taking as a degree of informal confirmation that there has just been a change in the allocation of One Nation preferences.

Could this run of bullish published 2PPs for the Coalition off the primaries as published have happened by chance?  Seems hugely unlikely, but I had a quick go at simulating it.  I ran a Monte Carlo of 2150 randomised Fakepolls with the same average primaries as the current Newspoll and a sample size of 1621 fake voters each, and only 8% of the Fakepolls had differences over 0.7 in the Coalition's favour in this regard.  Yet this has now happened in seven in the last eight real-world Newspolls, which would come out as about a one in 10,000,000 chance at that rate.  Throw in weighting, special modifications, scaling, state preferencing and whatever else you like and you could knock a couple of zeros of that, maybe, but it's still going to be a very unlikely event.

(Note: Initially this exercise had a sample size of 1797 per poll but that seems to be somewhat on the large side given recent Newspoll sample sizes.  So I redid it with a more representative sample size, which made no difference to speak of.)

And by the way, the average poll-to-poll published-2PP shift in my Fakepolls is 1.2 points, nearly twice that in Newspoll, even though nothing whatsoever happened in underlying voting intention in the 110 fake years I ran them over.  This further confirms that the current Newspoll is under-dispersed, as Essential was in the 2010-3 term.

Why might such a change have happened?

The behaviour of One Nation preferences at the 2016 election in the House of Representatives was unusual.  The party contested very few Lower House seats, and in one of those (Longman), One Nation decided to blow up the career of the LNP's Wyatt Roy, who they labelled "a little leftie".  (Roy got only 43.5% of One Nation preferences as a result, needing 52% to win.)

Only 50.5% of One Nation preferences flowed to the Coalition nationwide, but this is lower than the flow at past federal elections (generally in the mid to high 50s) and much lower than at the WA and Queensland state elections (60-40 in WA, 65-35 in Queensland).  With One Nation regularly polling around 7%, and with disgruntled Tony Abbott supporters still attracted to the party, no-one seriously believes that One Nation preferences would flow anything like as weakly as 50-50 in an election held "right now".  Election day when it comes about could be a different story.

So it can well be argued that last-election preferences are currently a bit favourable to Labor and that a more accurate predictive model should adjust for this at least for the time being.  This should be argued with some caution, because the track record of last-election preferences is incredible. And the example of Queensland 1998 (with a much stronger One Nation preference flow to the conservatives than seen at federal elections around that time) should show that state elections aren't always good models for federal preference flows.

But changing the One Nation preference allocation is at least a reasonable and perhaps even a good methods decision.  It is quite silly to say such a change would have been made to advantage the Coalition, and if anything a pollster that continued using only last-election figures without any acknowledgement of the One Nation issue at all might be accused of stacking the deck against the Government.  (Certainly, using exact last-election preferences for so long was no help in breaking the "30 losses" streak).

So what's the problem then?

The problem here is the lack of public detail - rather typical of an Australian polling industry that is remarkably and increasingly opaque.  The Australian continues to publish Newspoll with a statement that preferences are based on preference flow at the 2016 election.  It appears that in the case of One Nation, this may not be technically true and may not have been so for  months.  Especially in the context of a sudden return to competitive polling via the Coalition's first 2PP above 48% since 31 polls ago, conspiracy theorists on Twitter have alleged that a change has been made at the behest of the Murdoch empire to break the Turnbull losing streak.  Some Twitter partisans have also alleged that the change was made for this poll specifically, when it seems to have started eight polls ago (the cause of the one-point change from the last poll being that the Greens are down a point on primary vote, plus rounding.)

Normally, I find it depressing to be of the same species as those who entertain Newspoll conspiracy theories.  But in this case, while the conspiracy theories are as clueless as ever, I feel that someone out of The Australian, YouGov or Galaxy, or some combination of the three, may have brought this situation on themselves.  While announcing the change (whatever exactly it was) at the time publicly would have caused some people to claim the change was being made to try to help PM Turnbull break his losing streak before it got to 30, it would have been a better look than making a publicly undocumented change which would sooner or later be inferred from the evidence anyway.

The state of play

The minor controversy about Newspoll preferences (and the lack of detail concerning them) comes at a mildly interesting time in federal polling - the Coalition seems to be recovering significantly.  However it is not unusual for struggling governments to get a run of more benign polling at some stage only to regress to the mean or worse thereafter - eg Labor in late 2012.  So it remains to be seen if there is any substance in this shift or if it is just another modest resurgence like the one early last year.

In aggregating Newspoll, I'm assuming for the time being that all they've done is altered the flow of One Nation preferences, and on that basis I've come up with a rather complex method for finding a last-election 2PP for Newspoll.  So for the current poll, the primaries would normally imply, on average, 52.4 by last-election preferences.  But even after changing the flow of One Nation preferences, they would still convert to something that rounds to 52.  This implies that the primaries are actually a little stronger for the Coalition than the rounded version, and that Newspoll were probably looking at something around 51.3 which they then rounded to 51.  (Of course, if they've actually switched to respondent preferences, all of this is irrelevant).  So on that basis I've fed the Newspoll into my aggregate at 52.1 to ALP.  In terms of primary votes, the main reason that this poll rounded to 51 instead of 52 by whatever methods Newspoll now uses is that the Greens are down a point.

The only other new poll so far this week is Essential, which in its usual fashion contradicts the vibe and goes in at 53.1.  The polls from two weeks ago all went in somewhere in the 52s, so it's a few weeks since the last really lopsided poll (the ReachTEL).  My aggregate is also throwing the Coalition a little morsel on account of the demise of YouGov (meaning that I think the remaining pollsters on average lean to Labor very slightly) so my current aggregate is 52.4 (51.8 with One Nation adjustment).  Here's the smoothed tracking graph:

(A reminder that my aggregate continues to run on 2016 preferences, but it's the One Nation adjusted figure that I'll use for seat projections should One Nation still be a serious force by the time of the campaign.)

The overall trend is more important than the one-point shift in the latest Newspoll. The Coalition is in its best position for fifty weeks - things were last this mediocre for them just before the 2017 Budget.  The 2017 Budget was supposed to free the government from bad polls thanks to the powers of the mighty "budget bounce", which the Coalition placed great hopes in despite the evidence that budget bounces are very rare events.  In fact the Budget marked the end of a brief rally and things got gradually worse for the rest of the year.

It's amazing what difference a single competitive Newspoll makes.  Only last fortnight the Prime Minister was dissing Newspoll and saying it couldn't be trusted after failures in recent elections (one of the examples being Tasmania, where it actually wasn't in the field at all!)  Now suddenly "the political contest in Australia is very close".  There is even talk that the government might scoot off to an election a year early to avoid the worst of the redistribution* (and clashes with state elections, Easter, school holidays, etc) should its Budget succeed in producing a bounce.  (Has nobody heard that you can't bounce an election?  Or what voters think of being dragged to the polls a year early for no remotely fathomable reason?) We are also starting to see waffle about the mythical concept of polling "momentum", and journalists have even invented the concept of a "pre-budget bounce" (spare me).

[* This probably wouldn't make much difference as the replacement mini-redistribution would still give Labor at least an extra seat in the ACT]

On net satisfaction, PM Turnbull is up eight points to -17 (36-53) and Bill Shorten is up five to -20 (34-54).  Neither has cracked 40% satisfaction since the election, and the 33-poll streak of sub-40 results for both leaders at the same time is the second longest ever.  (Gillard and Abbott had 44 in a row like this.)  Turnbull's lead as better PM (38-35) is now rather meagre given the Coalition's improved 2PP, contrary to his past habit of polling large leads while the 2PP was bad.

Other Polling

Issues polling this week had a strong focus on immigration issues.  Essential found a marked increase in the level of concern about immigration compared to in 2016 (bearing in mind that Essential tends to show higher levels of concern compared to some other polls, as noted by the Scanlon Foundation (PDF)).  Overall, Essential found 64% thinking there had been too much immigration in the last 10 years and only 5% too little.  In Newspoll, 56% said Australia's current target level was too high and 10% said it was too low.

Essential also found the curious result that Greens supporters were a little less fussed about population growth than the major parties. Although the sample of Greens voters would be small and the differences hence not significant, it's not that long ago that Greens supporters tended to be the first to argue strongly against population increase - especially in Australia which they argued to be a fragile continent - on environmental grounds.  What has happened there within the green movement is that the old zero-growth or negative-growth approach has largely come to be seen as synonymous with low immigration and hence with anti-refugee attitudes.  Supporting refugees is these days a strong marker of difference among Greens politicians and supporters so they're nowhere near as likely to complain about population growth as they used to be.

Losing Streaks

Lastly, I recently collaborated with RMIT ABC Fact Check who put out a little graphic showing the longest streaks of Newspoll 2PP losses by leader, including estimates for polls for which no 2PP was published.  With all the focus on consecutive losses, I also thought the non-consecutive loss rates would be of interest, so here they are by PM, updated for this week's poll:

Hawke* 30/85 (36%)
Keating* 77/109 (71%)
Howard* 129/294 (44%)
Rudd 9/66 (14%)
Gillard 52/67 (78%)
Abbott 35/39 (90%)
Turnbull 35/51 (69%)

(* = includes estimated 2PPs)

It should be noted that Howard is slightly disadvantaged by Newspoll's use of respondent preferences around the 2004 election.  If these are replaced with estimates, his loss rate drops to about 123/294 (42%).  For the purposes of all this stuff and to keep attribution simple for the Fact Check folks I recalculated my own last-election 2PPs for all the Newspolls from before 2005.  These are generally similar to Peter Brent's dataset that I have used a lot in the past, but here and there there's a point or two of difference.

Update 26 April: 

We finally have some comment from Galaxy, which has tweeted:

"A quick comment about the issue of minor party preference flows. Our longstanding policy has been to look at previous election preferences and use those to calculate 2PP figures. This has not changed."

"The 2PP figure is determined based on the mix of the minor parties and their preference behaviour at recent elections. This is entirely the prerogative of YouGov Galaxy and not any other organisation. Our track record at federal and state elections speaks for itself."

The reference to "recent elections" (plural) confirms as suspected that Galaxy's preferences in this instance are not based solely on the 2016 federal election, as has been falsely claimed by The Australian for what appears to be at least the last five months.  The tweets also confirm that the speculation about Newspoll being switched to respondent preferences was (fortunately) wrong.

The tweets do raise the question of whether the preferences used in Newspoll since 2015 might have included any other adjustments that created departures from last-election preferences that were too small to be easily detected.  I also suspect the same approach could be taken regarding Australian Conservatives (a new party) or whatever NXT are calling themselves this week (60-40 to Labor at 2016 federal, but only 50-50 as SA-BEST in the SA state election.)

Update 29 April:

This issue was featured in a promo for ABC's Insiders (and I assume in the program itself).  Unfortunately the promo supported the incorrect claim that the change has just happened for the most recent Newspoll, thus leading to another mass spawning of nonsense about the change.  (Galaxy Research retweeted my tweet pointing out that the change probably happened eight polls ago).  The promo also referred to my 52.4% typical last-election preferences calculation for the primaries, but not to my estimate that the underlying last-election 2PP was probably about 52.1. (A very technical point about extra information that modifies a prior assumption.)

Update 30 April:

And finally, confirmation.

See also:

Pollbludger article on historic One Nation preferences, which among other things notes that the donkey vote in Paterson (NSW) was in fact a bigger influence on the One Nation 2016 preference split than the Longman situation.