Monday, October 2, 2017

Divergence In The Queensland Polls Is Caused By Preferences

Newspoll 53-47 to Labor, ReachTEL 52-48 to LNP - what gives?

A Queensland state election is coming soon, possibly very soon.  Electoral history tells us little of what to expect here. First-term state governments are usually returned, as are state governments that are of the opposite party to the party in power federally.  However, the former might not apply when the government was elected almost accidentally, and the latter is most at risk of falling over when a government has been a messy minority regime.  But if we turn to polling for the answers, whether the Palaszczuk government is cruising or crashing depends on which pollster you ask.

Through 2017 there has been a major divergence between the media-commissioned polls of ReachTEL and those by the Galaxy stable (sometimes branded as Galaxy, sometimes as Newspoll).  Three media ReachTEL polls have all shown the government trailing in the two-party-preferred contest (47% in February, 49% in June and 48% just now.)  Three Galaxys and a Newspoll have all shown the government ahead (51% February, 52% May, 51% August, 53% July-September quarterly).  On average, that's a 3.75-point two-party difference between the two stables, way too large over so many polls to be explained by chance or fluctuations from month to month in support.  Either the truth is somewhere in the middle and an election now would be extremely close, or somebody is right and someone's wrong.

We saw divergence between these stables in the leadup to the Western Australian state election, where the Galaxy stable consistently showed healthy leads for Labor while ReachTEL showed leads of varying sizes (and in one case no lead at all).  ReachTEL's final 2PP margin was the same as Newspoll's (54-46), and in fact Labor (55.5%) beat both of them and beat every poll bar one taken in the whole of the term.  Score one to Galaxy/Newspoll for being more predictive further out.  But Queensland is ReachTEL's home ground, and they had the most accurate primary votes (by a whisker) last time.  So why are these big differences arising?

Of course, the 2PP is not the be-all-and-end-all in this election.  The very large One Nation vote offers the prospect that that party could win a bunch of seats, making a hung parliament likely if the 2PP falls inside 51% one way or the other.  As previously looked at, One Nation may not get much bang for their buck if their vote lands in the mid to high teens, but if they can crack 20% they could get several seats, and if they can get well into the 20s they could do even better than in 1998.  Their vote did appear to be slipping back to the mid-teens, but the current ReachTEL (based on a sample of 4000!) doesn't think so.

Primary and Preferences Breakdown

Reporting on ReachTEL polls is proving to be increasingly difficult because the full details often aren't published.  The use of headline figures that include so-called "undecided" voters creates a false impression that the primary votes of both major parties are a few points lower than they are.  The pollster has a breakdown of who these "undecided" voters lean to, but it is not always reported, making it difficult to establish whether oddities in the 2PP result come from the preferences of the voters or from the breakdown of these voters between parties they are leaning to.

Preferencing is a further complication.  The 2015 Queensland state election was conducted by optional preferential voting, but that system has been replaced with compulsory preferential voting.  Also One Nation are a very significant factor in current polling, which would make "last election" preferences especially meaningless.  ReachTEL are using respondent preferences (normally not a very accurate method in compulsory preferential elections) while the methods used by Galaxy/Newspoll have been quite difficult to track down.  In the case of the September Newspoll, the following preferences were assumed: One Nation 55% to Coalition, Greens 80% to Labor, Others 50-50.

In the table below I give, for each poll:

* the primary vote results (in the ReachTEL cases, exactly if I've got them, or extrapolated by distributing the undecided if not)

* the 2PP results as published by the pollster (2PP (PUB) column)

* the 2PP results based on the primaries found by applying Newspoll's assumed preferences (above) to the published figures.

As well as the three media ReachTELs, I've also included two commissioned ReachTELs taken in May.  These were both for left-wing sources (Together union and Australia Institute) and were slightly better for Labor than the media ReachTELs.  I give averages both for the three media ReachTELs (Media RT) and for all five ReachTELs including the commissioned ones (All RT), and also for the Newspoll-Galaxy polls.  In the line DIFF1 I show average differences between the Newspoll-Galaxy polls and the media ReachTELs.  In the line DIFF 2 I show average differences between the Newspoll-Galaxy polls and the full set of ReachTELs.

So, what is the story here?

The differences betwen the Newspoll-Galaxy polls and the ReachTELs are generally less pronounced if we include the two commissioned ReachTELs in the ReachTEL data set (something I wouldn't necessarily recommend doing, since the unions may well have had worse ReachTELs for Labor and hidden them under the nest of a Night Parrot or something.)  But nonetheless some differences are consistent whichever way these numbers are sliced:

* ReachTEL finds stronger primary results for the minors (especially One Nation) and weaker results for the majors (but particularly Labor).

* If both pollsters used the same 2PP method, then their 2PPs would be pretty similar.  The media ReachTELs would on average be 0.9 points weaker for Labor than the Newspoll-Galaxys (which isn't even convincing evidence of difference), while if the full ReachTEL sample is used there is basically no difference.

* The difference is mainly caused by ReachTEL getting something very different in its respondent preferencing to the preferencing assumptions used by Galaxy.

Comparing the media ReachTELs with the Newspoll-Galaxy results, 2.8 points of the 3.7 point difference is accounted for by different preferences.  Comparing the full ReachTEL set, 2.6 points of the 2.8 point difference is accounted the same way.

So the story of who is actually "winning" the current Queensland polling comes down, largely, to what is going on with preferences.  If the preferences break roughly as per the last federal election (but One Nation's break along similar lines to their long-term federal average) then on current primaries Labor are well ahead and could even win a small majority off a primary of, say, 36.  If, on the other hand, One Nation preferences have suddenly swung wildly to the LNP, then the LNP might either be winning outright or at least in a strong position to form some kind of minority government.

On the primary votes, it looks a lot like One Nation have taken a lot more of their votes from the LNP than from Labor.  If votes taken by One Nation from the majors just return to them as preferences (an unreliable prospect), then the election could be decided by the "others" voters who voted for PUP and so on last time but have now switched to One Nation.  If the strange preference flows from parties like PUP and KAP last time were just about their voters disliking Campbell Newman, then on current primaries Labor could struggle.

Pollsters and psephologists were burned last Queensland election when preferences did something extraordinary, and we'll be very careful this time round in case the same thing should happen again.
However I remain sceptical of some of these more dramatic preferencing-shifts (both in ReachTEL and in the new federal YouGov online polling).  Even before considering the possible impact of how-to-vote cards and their surrounding politics, I would like to see claims of radical preferencing shifts blessed by a method that at least randomises the order of the parties in the readout (nobody does this).

A bit of a hint might come from the preferred premier scores, although they are a very messy indicator.  Newspoll preferred leader scores come with a baseline advantage to the incumbent, but they are also connected to the 2PP and, to a lesser degree, individual leader ratings.  Normally, an incumbent leader with a 53-47 2PP lead would have a large lead as preferred Premier, but Annastacia Palaszczuk's lead as preferred Premier over Tim Nicholls in the recent Newspoll was only 10 points (43-33).  ReachTEL, which doesn't skew to incumbents much, gave her no lead at all.  That's although Palaszczuk is not herself unpopular (having only recently slipped into mildly negative netsat ratings for the first time) while Nicholls isn't popular at all (his most recent Newspoll netsat was -16).  This all follows the last election where Campbell Newman went to the poll with a Better Premier lead of just one point, and lost, despite the poll giving his party a 52-48 2PP "lead" based on incorrect preference assumptions.

Just maybe we can use preferred leader scores to spot cases where an election might be closer than a headline 2PP suggests.  I should warn that I haven't tested this extensively, so please do not try this at home without so doing!

(PS: For completeness in advance of new polling, noting here the Oct 10-12 Newspoll which had almost identical primaries to the previous one, but a 52-48 2PP because of different preference assumptions.  Newspoll now assumes 60% of One Nation preferences to LNP).


  1. What methods do they use for their polling Kevin? I have been rung on the landline by ReachTEL quite a few times over the last few years, but by none of the others except that odd mob from Perth. I think I've had one poll on my mobile - forget who that was. If ReachTEL are selectively hammering the landlines you might expect a bit of conservative and ON bias - despite my best endeavours!

    1. ReachTEL is 100% robocall with a mix of landline and mobile (more landline than mobile). Newspoll/Galaxy is a mix of online panel polling and phone polling. I think the phone polling is all robopolling now, but not completely sure about that, and also have seen varying reports about the extent to which the phone polling may include mobile as well as landlines. In any case, the online component captures people who have no landline.

      Nobody much still does all-landline polling anymore.