Wednesday, March 22, 2017

Poll Roundup: Attack Of The Poll-Shaped Objects

2PP Aggregate: 52.8 to Labor (-0.8 since last week)
Labor would comfortably win election "held right now"

This week has been a confusing week for many poll-watchers, and an amusing week for those of us who watch the antics of the poll-watching partisans.  Newspoll was eagerly anticipated following its 55:45 to Labor three weeks ago, and widely expected to come out with more of the same if not then some, but pulled up at only 52:48.  Oh well, the line went, maybe it was never 55:45 to begin with, after all Essential had only got out to 53s and the odd 54.  But then Essential jumped to 55:45, so two polls not usually noted for volatility had delivered it not just in spades but also in opposite directions.  At the moment we don't have a third opinion, since the others are very inactive lately.

In trying to decide between these competing figures, it is worth bearing in mind that Newspoll is now administered by Galaxy.  Galaxy was the best pollster of the 2013 election, and the Galaxy/Newspoll stable triumphed again in 2016.  Essential was poor in 2013, and while its final poll in 2016 was very good, its tracking performance suggests it was probably lucky or herding.  So my aggregate comes down more on the side of Newspoll, crediting Labor with 52.8% 2PP.

Anyway, that is a recovery of sorts, but we have seen false dawns before.  Here's the smoothed tracking graph:


For anyone with an interest in my input values, this Newspoll went in at 45.1% for the Coalition after considering the primaries, and the last three Essentials have been 46.9 (published 47), 46.7 (47) and 45.2 (45).



The three-point rise in Newspoll was interpreted by the Australian as caused by Prime Minister Turnbull's Snowy River 2.0 hydro power announcement, but it could have been caused partly or entirely by many other things or none.  It might even be random bouncing. In particular, if a policy change announced by the PM is supposedly the major cause of a three-point rise on voting intention, one would have expected a much bigger rise in the PM's own ratings.  This wasn't the case.

Leaderships

This week's Newspoll had Malcolm Turnbull's net satisfaction up three points to -27 (30-57) with Bill Shorten down two to -28 (29-57).  The last time we saw both leaders polling this badly at once was just before Tony Abbott was rolled.  Turnbull has jumped out to a 43-29 lead as "better Prime Minister", his second-highest lead since the election, but that's not much to write home about.

Roy Morgan Research had a leadership poll release that made me think "Morgan poll, hmmm, oh, is that still a thing?" Morgan found Turnbull to have dropped two points since October to -24 (30-54) and Shorten to have dropped ten to -28, the same as Newspoll (28-56).  Morgan has Turnbull up 49-32 as "better Prime Minister".  Finally, last week's Essential also found both leaders in the slop with Turnbull at -17 (33-50), his worst to date, Shorten at -19 (30-49), his worst since a year ago, and Turnbull up 38-26 in the least-ugly contest.

Morgan also had a preferred leader poll that suggests some of the modest-to-begin-with support for making Tony Abbott Liberal leader again has gone over to Peter Dutton, but otherwise didn't report anything new.

I won't cover the other reputable polling results in much detail this week because in the last few weeks they have been outnumbered by polls that were at least badly reported and in cases unsound.   What follows is a selected parade of poll-shaped objects: public releases based on "polls" that have been poorly reported, are badly designed or that fail appropriate disclosure standards.  (The term is by analogy with piano-shaped object, something that appears to be a piano but sounds horrible or is impossible to play - whether for reason of shoddy construction, mistreatment or decay.)

I should make it clear that in many cases below there is nothing evident wrong with the poll, and the issue is the way the media have trumpeted it.  In other cases the polls themselves have problems.  I've left enough information for readers to hopefully work out which is which.

PSO 1: Peter Dutton Losing His Seat? 

Wirrah Award For Fishy Poll Reporting (image source)

I'll start with this one because I think it's one of the worst examples of bad poll reporting I've seen in a while, but it also underlines the importance of clear communication between pollsters, sponsors and the media and immediate publication of full results.  The Australia Institute has done a ReachTEL of Peter Dutton's seat of Dickson.  TAI-commissioned polls frequently purport to show that voters in marginal and even safe Coalition seats actually support ideas espoused by the political left, especially in the field of renewable energy.  These polls are common and I don't have anything new to say about them for now, but what was interesting about this one is the reports on voting intention polling concerning Peter Dutton.

Polling surrounding Dutton is of interest at the moment because there is, for some reason, speculation that he might be the next Liberal leader; indeed he remains the $1.87 favourite to do so with one bookmaker.  That said he has been barely a blip on the radar in smorgasbord polling since the hype began, scoring a massive 2% (the same as Christopher Pyne and Scott Morrison) in Essential's effort before breaking away from these formidable rivals to lead 5% to their 4% each in this week's Morgan.  Malcolm Fraser went from 5% to leader in three months, so there is hope for #Pynementum yet...

Even if he never becomes leader, Dutton is an influential Coalition head-kicker (and heel-dragger) and so the fate of his marginal (1.6%) seat is of much interest.  And this certainly got some attention when the figures first came out on the Guardian website and on Twitter.

The article's original subheading had "Poll says immigration minister would lose election if held today", and then the article went on to say the following, among the other things it still does say:

"The poll also shows a large One Nation primary vote in the seat (16.8%) and Dutton on track to lose his seat by five percentage points if an election were held today.

[..]

The poll found Labor leading the LNP in Dickson on 55.7% to 44.3% in two-party-preferred terms.

[..]

The ReachTel poll suggests Dutton would lose his seat if an election were held today but, as respondents were asked to allocate their own preference between Labor and the LNP, it may understate the flow of minor party preferences to the LNP.

[..]

Some 3.7% said they intended to vote for another candidate and 5.5% were undecided. Exactly half (50%) of undecided voters said they leaned to the LNP."

The problem was that by the primary votes published by the Guardian, and ignoring the information about undecided voters for want of a full breakdown, the poll would come out at 50.2% to Dutton by nationwide last-election preferences.  And, while the use of respondent-allocated preferences can skew polls by a few points, six points seems a bit much.

Questions were asked, and soon all the references to Dutton losing vanished from the article, as did the line about half the undecided voters leaning to the LNP (which was strange since that bit was actually true.)  A note was added to this effect: "This story was amended on 21 March to remove an incorrect reference to two-party-preferred share in the poll."  - a note which went nowhere near describing everything that had been removed.

This raised so many further questions.  Was there ever a 2PP?  If there was, was the 2PP incorrect as supplied by the pollster to TAI, as supplied by TAI to the Guardian, or had the Guardian misinterpreted it on their own accord?  If there was a real 2PP what was it, and why was it not being reported instead?

It has taken much of the day for the facts to become clearer.  The respondent-allocated 2PP was 52% to Peter Dutton, which would be an encouraging result for him if seat polling was more reliable than it is.  This is pretty close to the 51.0% obtained by 2016 preferences using the full primary figures.  The source of the confusion was that some recent ReachTEL results have had a question 1b that asks the respondent who they would preference out of Labor and the Coalition, but this question is only asked of those who did not already choose Labor or the Coalition.  So 55.7:44.3 is not the two-party preferred for the poll - it is the two-party preferred among the non-major party voters.

ReachTEL have stated on Twitter that "The original document did have a 2pp but it sounds as though that wasn’t communicated."  The version that appears on the Australia Institute website has no 2PP and no clear indication that question 1b isn't the actual 2PP.  I do not know what explains this discrepancy, but what is clear at the moment is that the Guardian article made much of the supposed fact that Peter Dutton was trailing badly but there has been no rush from the Guardian to fully correct the article and ensure readers are informed of the real story.

Amusing as it is, poll watchers shouldn't have to be chasing up this nonsense to hold poll reporting to account.  It should be standard practice that once a poll is referred to in a media publication, the full details of the poll are immediately published, either on that publication's website, or as a link from the article.  The howlers that are being caught are the tip of the iceberg.  Among all the reported results of polls that have been seen only by a journalist and whoever showed them the poll, there are bound to be many errors that cannot be checked.

(And by the way, no, I don't support subsidising the Adani coalmine either.  I'd probably be up for paying it to go away.)

PSO2: Public Support For Amending 18C?

On the other side of the coin, we have the Australian's doubtless specially-commissioned Newspoll on proposed changes to Section 18C of the Racial Discrimination Act, an issue that said paper is notoriously focused on.  Here we did get the question wording upfront, and also interesting party breakdowns (showing that the issue is a lot less partisan than culture-warriors would think) but the problem is the preamble is skewed:

"The Racial Discrimination Act makes it an offence to 'insult' or 'offend' someone on the basis of their race. Supporters of the act say it protects people from racial abuse but opponents say it goes too far and limits free speech.  If the words 'insult' and 'offend' are replaced by the word 'harass' in the act, it will set a higher benchmark for complaints. On balance, would you be in favour or opposed to this wording change to the Racial Discrimination Act?"

This preamble is misleading firstly because it creates an impression that insulting or offending someone on the basis of their race is always illegal.  The preamble fails to mention the Section 18D defences which provide protection for insulting and offensive statements under some circumstances. The preamble also fails to explore the distinction between an offence and an unlawful act.

Another sneaky word in this preamble is "but".  The form of words A-says-X-but-B-says-Y makes it sound a lot more like A is wrong than if the "but" is not there.  Then the next sentence is simply flying a flag for a positive quality of the proposed change and blatantly priming the respondent.

Given the level of skewing in this preamble, the overall result (47-39 in favour of the change) is about as useful as an Andrew Bolt attempt to determine somebody's ethnic identity.  Next ...

PSO3: Daniel Andrews Non-Preferred Premier?

If Daniel Andrews' Victorian Labor government is in serious polling pain then it utterly deserves it.  The recent entitlements scandal that saw both the Speaker and the Deputy Speaker quit and the latter go to the crossbench is exactly the sort of thing that makes people disgusted with politicians in general, especially when it seems that pollies of all parties simply rotate whose turn it is to engage in compulsive entitlement rorting.  After Geoff Shaw in the previous parliament this is the last thing public trust in politics needs, and manna from heaven for populist parties crammed with people who given the chance would probably be even worse.

That said, the Herald-Sun's coverage of a recent commissioned ReachTEL supposed to show Labor trailing 46:54 (Labor Party faces Victorian election 2018 wipeout - may be paywalled) was especially concerning.  It published what was clearly somebody's internal/"private" polling of some sort without saying whose it was, and without even making it all that clear that this poll was something different to the normal run of media polling.  All we got by way of disclosure was "The Herald Sun has obtained a ReachTEL poll, the first commissioned since an expenses scandal claimed the scalps of the Speaker and deputy speaker, [..]"

A possible hint within the poll results comes from the commissioned question about the Firefighters Union boss, suggesting that this could be someone who either has an ax to grind over the issues that led to the departure of Jane Garrett, or someone who just thinks the issue is useful for attacking Andrews on.  That doesn't narrow the field all that much.

Anyway, at least one commentator (I don't now have the link) seized on the finding that Andrews was trailing as better premier and noted how unusual this was for an incumbent, but an important point was overlooked here.  The way in which the question was commissioned, giving options of Andrews, Guy, Other and Undecided, is completely irregular by including an unspecified "Other" option, and therefore there is no benchmark to assess it against.  Even a party breakdown would be useful here to say more. My suspicion would be that when a diehard Greens voter hears Other they will often think "That's us! They've given us a question to be special on!", and hence that this method is not kind to Labor.

PSO4: Federal Libs Losing In The West?

Ah yes again we go, the wailing sirens of looming electoral doom as a sensational new poll makes a startling finding about a swing that is caused by a single thing, the GST carveup, and is not in any way contaminated by the state election held a few days earlier.  Dubious timing aside, there's another issue with this one: respondent preferencing.  Based on the converted primaries, the 2PP by last-election preferences comes out to 50.2 to Coalition using national preferences, or 49.9 to Liberal using WA-specific preferences (which includes treating Nationals as separate) but the national figure for One Nation.  Yet the published 2PP was 53:47 to Labor.

While there have been some recent state elections with massive primary vote swings where last-election preferences were all at sea (Queensland and NSW) those were conducted using optional preferential voting.  In WA last-election preferences (using the last federal election or an assumption for One Nation) seem to have worked better than respondent preferences overall despite a substantial swing in Green preferences to Labor, and the track record of last-election preferences at federal elections is very good.  It's useful to have respondent preferences to keep an eye on shifts and on any evidence of potential for last-election preferences to be badly wrong, but it baffles me that anyone would use them as a headline without even publishing a last-election result.  In this case the choice of preferencing method makes about two seats' difference to the projected outcome. Let's see how this goes when the new state government has some time to settle in.

PSO 5: Christian Porter To Lose Seat Over Marriage Equality?

Hang on, wasn't he losing his seat over the GST carve-up as well?

My Twitter profile has a rainbow flag and describes me as "Totally pro-SSM".  This probably looks like virtue signalling to someone out there but it is actually more like a beware-of-the-dog sign.  It's a warning to moral reactionaries that I may rant about marriage equality from time to time and if they expect me to only tweet about psephology, then I expect to flame them to a crisp.

The reality is that I would do anything for equal love, but I won't do that, where that is let people on either side get away with sloppy poll-reporting practices.  Polling data belongs out of the closet, fully out and parading proudly, not just half-out to a few special friends. The findings of a recent campaign-commissioned poll are quite emphatic, especially as there have sometimes been issues with obtaining results fully comparable with other polling methods by robopolling in the past.  I don't see any reason to doubt the primary result that voters in the listed conservative seats support marriage equality.  But when I see a barrage of results to different questions reported together I'd at least like to see what all the questions were and what order they were asked in, just in case the earlier questions might have potential to influence the later ones.

And when it comes to that staple of issues polling, the "would you be more likely to vote for blah if blah takes this position on issue X?" type question, it should be well understood that this type of question has no predictive value whatsoever.  All that style of question shows is that voters will say that almost anything is important and might change their vote in isolation.  It says nothing about the relative salience of different issues when they come to deciding votes.  Today I told a pollster that adult literacy and numeracy rates were a "moderate problem" in Tasmania, but if a charity caller rang up asking me to donate $1 to a literacy and numeracy campaign I'd tell them I wasn't interested.

The lack of salience of marriage equality as a vote-shifter has been frustrating, but it is what it is. Fear not; the Coalition has a vast capacity for self-harm when it comes to gay rights issues, and has shown in the past that it's perfectly capable of destroying itself on such matters without needing any help at the ballot box.

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

There are so many more poll-shaped objects out there.  Every few days on average I will come across a new one.  I've even considered setting up a monthly spreadsheet of them all, categorised with which forms of polling or reporting failure each one displays, but I'm not sure I could keep up with the workload!

No comments:

Post a Comment

The comment system is unreliable. If you cannot submit comments you can email me a comment (via email link in profile) - email must be entitled: Comment for publication, followed by the name of the article you wish to comment on. Comments are accepted in full or not at all. If you submit a comment which is not accepted within a few days you can also email me and I will check if it has been received.