Friday, July 10, 2020

Divergent Polling In The Northern Territory

Advance Summary

1.  There have been two massively different recent polls of the Northern Territory - a MediaReach internal poll for the Territory Alliance covering all bar four seats and a uComms poll for an environmental group covering Greater Darwin.

2. Although the uComms poll is much smaller, uComms has the better track record nationally so far (from limited evidence) while MediaReach is a mysterious pollster with a weaker public track record. Furthermore internal party polls that are released tend to favour their sponsor.

3. The MediaReach poll implies a probable hung parliament with Labor losing its majority and the Territory Alliance supplanting the Country Liberal Party as the largest conservative party, and possibly even forming government.

4. The uComms poll, however, implies that at least in Greater Darwin the Territory Alliance is unlikely to win many seats, and is consistent with Labor remaining in majority unless there are larger swings against it outside its survey area.  

5. Neither of these polls are necessarily reliable.  Media should not rely strongly on either in setting narratives for the upcoming election.

6. The history of "federal drag" - parties tending to perform best when in opposition federally - appears to apply to the Northern Territory very much as it does to state elections.


It's only six weeks until the 2020 Northern Territory election!  Some observers are bemused that psephologists pay any attention to NT elections, given that they have such low enrolment and low turnout.  The electorates seem much more like council wards and MPs get up to stuff that in the rest of the country is usually pruned out at the student politics stage. However aside from the frequent entertainment value of Northern Territory politics, there are two points of special interest in this year's election.  The first is that the NT will be the first state or territory to test whether the COVID-19 pandemic is a boon to state and territory leaders at the ballot box as well as in approval ratings.  The second is the third-party disruption attempt by the Territory Alliance, led by former Chief Minister Terry Mills, who may or may not have been Opposition Leader for between one and six days earlier this year.  The Gunner Labor Government is widely expected to come down to some degree from its 2016 thrashing of the Giles CLP rabble, but how much?  Could it lose its majority?  Will the CLP rebuild from its current holding of a feeble two seats out of 25, or will the Territory Alliance strengthen its position as the largest non-Labor party in the Parliament?

While Labor won the last election with a whopping 57.5% 2PP, its position is inflated by huge margins in a few of its safest seats.  Even ignoring the Territory Alliance, a uniform two-party swing of 5.2% (following a redistribution) would cost Labor six of the eighteen seats it won in 2016 and, all else being equal, its majority.  It has also suffered a net loss of two MPs to the crossbenches, including Fong Lim MP Jeff Collins to the Territory Alliance, so in theory a loss of five seats (a uniform swing of 3.0%) could lead to a loss of majority if Collins retains his seat.  As against that, personal vote effects tend to be strong in Territory elections because the electorates are so small, so many of Labor's marginals might be safer than they look on paper. (Labor's other defector Scott McConnell is targeting Adam Giles' old seat of Braitling, now a Labor marginal, so McConnell's seat of Stuart is free for Labor to recover it without having to compete with him.)

Polling is rare (and highly unreliable) in the NT, but at the moment we have two recent NT polls, and they say very different things.  In late June a MediaReach poll of 21 seats was produced which purported to show Labor in a majority-losing position, with the Territory Alliance polling ahead of the CLP in many seats and likely to win at least seven seats, perhaps at the most a small majority.  This poll was described by the NT News as "independent polling" but one had to scratch around different articles to see that it was in fact commissioned by the Territory Alliance themselves.  The NT News did trumpet MediaReach's record at the previous election and in federal contests, but in fact while the pollster got those results right, it has had significant errors on the 2PP in doing so (6.5 points and 4.7 points on the two 2016 polls; I haven't found their 2019 results).  Such errors didn't matter in forecasting the 2016 demolition, but for this one they may well.

While polling in the NT is known to be very difficult and uComms may not do as well there as elsewhere, there is also the example of the Braddon by-election, in which the Liberal Party used MediaReach to claim they had successfully driven the vote of independent Craig Garland down from 9% to below 5% by attacking him and taken the lead.  In fact the attacks backfired; Garland polled 10.6% and his preferences decided the seat in Labor's favour.  On the other hand, the poll seemed to perform well in the leadup to the 2018 Tasmanian election.  

This is not the first highly adverse MediaReach poll for the Gunner Government in its 2016-2020 term.  A previous poll last September, commissioned by somewhat controversial conservative businessman Owen Pike (and not by Sky News as has been misreported in some places) found the CLP leading Labor 39-29, with Michael Gunner on a net satisfaction rating of -47 (23-70, worse than any state Premier rating in Newspoll history, for example) and with preferred Premier figures of 29% for Gunner, 21% for then CLP leader Gary Higgins and 37% for Mills.  

Overall the current MediaReach poll was an enormous exercise involving around 200 respondents per seat, which given that only about 4000 people per seat actually vote would have involved reaching a very high proportion of reachable respondents, and probably a lot of scaling.  A new smaller poll of Greater Darwin is now available, from uComms, commissioned by a local environmental group (see Poll Bludger report here and the poll report here.)

Despite the uComms poll's smaller size and more restricted coverage, there are good reasons to give at least some weight to it.  Firstly while it is commissioned by a lobby group, the lobby group's primary interest is in getting attention for the poll's issue findings. The voting intentions are polled to enable party support breakdowns and released as a hook to make the media interested in the issue findings.  (As for those issues findings, the respondents become more and more primed as the series of questions goes on, so I suggest especially ignoring the ones towards the end.)

The results of lobby-group-commissioned uComms seat polls have been good by the low standards of seat polling at recent elections.  In contrast, MediaReach is little tested on the public record at actual elections, and party-commissioned polling that is released by parties tends to favour those parties compared to the actual results.  (We saw this again in Eden-Monaro recently with the Nationals getting at least 5% less than their internal polling said.  See also this US overview.)  There are two possible reasons for this: firstly if parties do not like their internal polls they won't release them, and secondly pollsters may find it difficult to avoid the subconscious temptation to make subjective decisions that give encouraging results for the sponsor.  MediaReach is, overall, a very mysterious pollster and detailed original reports of their polling are seldom seen despite its frequent use by mainly conservative clients.  

Overall, the uComms poll finds the vote in Greater Darwin (after redistributing undecided) to be 39.3% Labor, 31.0% CLP, 13.7% Territory Alliance, 7.2% Greens, 8.8% others.  William Bowe gives that as an 8.6% swing against Labor and a 2.6% swing to the CLP.  There is no two-party-preferred figure published, but from the primaries I get the 2PP at 54.4% to Labor, only a 3.2% swing from 2016.  The reason the swing is not larger is that the primary vote swing from Labor is going mainly to the Greens, who were included across the sample despite contesting only three seats in 2016.  At this stage they have announced candidates for five seats in Greater Darwin and one outside.  If they do not contest all seats, or maybe even if they do, then they will get a lower vote than polled.  

In comparison, after redistributing the large "Don't Know" component in the Media Reach poll on a seat by seat basis, I get its results in Greater Darwin as 36.2% Labor (3.1 points lower than uComms), 23.0% CLP (8 points lower), 26.8% Territory Alliance (13.1 points higher!), and 14% for others including Greens.  There is no point in estimating a 2PP for these figures as many seats would see the Territory Alliance place first or second.  So there is an enormous difference between the Territory Alliance's own poll and the uComms poll which is not for a party.

The 200-vote samples in the MediaReach are extremely rubbery.  In their study of seat polling at the 2016 election, Simon Jackman and Luke Mansillo found that seat polls behaved as if they had one sixth of their actual sample size.  If this is applied to a 200-vote sample, it goes from having a theoretical maximum margin of error of 7% to having one of 17%.  It might not really be that bad, but the point that we shouldn't be projecting individual seats off them with any confidence at all (unless someone is polling, say, 70%) remains.  However, all else being equal, a party's seat result would be something like what you would get from adding up the winners in those samples, provided that the polling overall was accurate.  

Below I use a seat breakdown model - which is not any kind of prediction of anything - to showcase how massively different the uComms and MediaReach results are.  The NT News got a lovely detailed free polling story out of the Territory Alliance poll with an abundance of statistical goodies for its readers (including me, an indirect subscriber!) but more scepticism about assuming that such an internal poll actually "proves" anything would have been healthy.  

TA Poll Realigned To Match uComms Poll

To do this, what I've done is, as best I can, taken the MediaReach seat by seat results and adjusted them to match the difference with the uComms.  The purpose is to give example seat projections if we assume that the MediaReach poll is wrong by the same amount pretty much everywhere while the uComms poll is accurate.  That may or may not be the case, so I've also done a hybrid version that assumes the two polls are equally wrong.  I do not claim that that is necessarily more likely to be the case either.  

Several caveats:

* The uComms poll does not say which seats exactly were considered to be part of Greater Darwin.  I've included 16 that I believe at least partly enter the area, but if I'm wrong about one or two of them it shouldn't make a massive difference.  Corrections welcome.  

* I've followed the TA poll's assignments of the four seats it didn't poll, except Barkly.  The CLP did not make the final two in Barkly in 2016 with Labor beating an independent 58-42.  I saw no reason to believe the CLP would be in play here if the Territory-wide swings were not so strong as in the MediaReach poll.   However, since I published this article, a reader has pointed to the retirement of the 12-year Labor incumbent and the CLP candidate being a local mayor as reasons to consider the seat in play despite Labor having held it since 1990.  (Antony Green estimates Labor's 2PP margin at a formidable 15.9%).

Also, in 2012 there were massive swings to the CLP among Indigenous voters in the remote and unpollable seats of Arnham, Arafura and Stuart (now renamed Gwoja) resulting in the CLP unexpectedly winning these seats.  There's no data to say that this couldn't happen again, in which case all the models would be optimistic for Labor.

* I've made no adjustments to the MediaReach results in either Fannie Bay (where the TA polled vote is negligible) or Mulka (where the TA isn't running).  I've made proportionally greater adjustments in other Greater Darwin seats as a result.

* Excepting Mulka I've adjusted the seats outside Greater Darwin by the same amount as those inside.

* I've used yellow to highlight some results and interpretations that should be treated with special caution:

- former speaker Kezia Purick's seat of Goyder was an easy retain for her in the TA poll, but she has since been dumped as Speaker after multiple adverse ICAC findings, so it may or may not still be safe.

- the TA poll interprets Johnston as a Labor or TA win but Labor would win easily, as its numbers are better than in the by-election which it also won.  I don't believe how-to-vote card decisions would make enough difference to change that.  

- the TA poll interprets Karama as a Labor or TA win but this is the seat nearly retained by de-selected Delia Lawrie as an independent and there is enough "other" vote for an independent to potentially win it if they could corral enough of it (that said, I can't yet find evidence of one running.)

- the TA poll calls Port Darwin "open" but the TA wouldn't get out of third on those numbers.  The ludicrously high 27% don't know might be taken as a reason for calling it "open" but that could apply to many other seats.  It's handy for the TA that the table gives the TA a shot here without giving Labor a shot in several others, because it leaves the TA as the only party in the hunt for a majority, which is the message the TA is keen to convey as it tries to break through.  (I don't know whether the assessments of who would win were prepared by the TA or the pollster.)

* I'd also be surprised if Labor was really in the mix in Nelson as the adjusted and hybrid models suggest, as it was one of only three seats to return a CLP 2PP win last time (and a hefty one - 57%)

* In trying to project the results I've assumed the following re preference flows:

- preferences could flow either weakly or strongly between the two conservative parties, but Labor won't get more than an even split in any case.

- Labor preferences will flow to TA in preference to CLP, perhaps strongly so.

All that said, here's the table (Click for full size, clearer version):

* Namatjira is notionally CLP following a redistribution.  See Antony Green re the redistribution here.

In the adjusted model that assumes the uComms poll is correct, all the Territory Alliance's surefire wins disappear and indeed it only remains competitive in two of the three seats it currently holds.  Labor now wins seats it was borderline in before, and the CLP now has six likely wins instead of two.  

The hybrid model gives the Territory Alliance two of its three current seats and has it competitive in a few others.  What is especially notable about this model is that while it only gives Labor nine clear wins, Labor is competitive in a lot of seats.  

Here are the comparative seat count ranges for the three models:

The TA poll and the adjusted model derived from the uComms poll are completely different political universes.  Based on the swings in the uComms poll, Labor still governs in majority (though it's well within the margin of error of the seat samples that they could fall just short and need to rely on one or two crossbenchers).  The CLP rebuilds while the Territory Alliance holds either some of its seats or else nothing.   There is certainly room for scepticism about this version of events too since at the Johnston by-election the CLP performed woefully while the Territory Alliance did quite well.  I would also be very surprised if Mills at least did not retain.

In the hybrid version, the extremes of the Labor range should probably be ignored.  There's an extremely wide range of Labor seat outcomes depending on preference flows and exclusion orders, but it's unlikely they would either win nearly all or lose nearly all of the several ambiguous seats.  The key point is that only Labor can win a majority in this version, though whether they do or don't is unclear.  At the lower end of Labor's range, the TA holds the balance of power.  It's possible in this version that the TA wins more seats than the CLP does, but it's not that likely, and if this does happen Labor has probably been re-elected.

None of the current polls have provided any fresh leadership ratings for Gunner.  I don't regard the MediaReach from last September as a completely reliable measure of his rating anyway, because it was not a "public poll" commissioned by media or conducted by a pollster of its own accord - that said, anecdotally his government was severely on the nose at the time with the NT suffering an economic downturn. However it is likely his approval has been lifted by the NT's success so far in containing COVID-19.

In short, the polls are so divergent, and are all affected by issues (whether it is sample size, limited coverage or non-neutral sponsorship) that what they, taken together, say about what will happen on August 22 is very little.

Federal Drag And NT Elections

Federal drag is one of the most important factors for understanding Australian state elections (What Kills State Governments: Age Or Canberra?).  All else being equal, voters prefer to elect a different party to run the state government as is in power federally, perhaps because they feel that a same-party state government will be too compliant.  However, the longer a state government has been in power, the more vulnerable it is to being thrown out anyway.

I have never looked at NT results in this light but it is clear that they display the same pattern.  The CLP won overwhelmingly in 1974 while Gough Whitlam was in power federally and was able to remain in power for 27 years running what was often basically a one-party regime.  However it lost many seats, including its leader (!), in 1977 with the Coalition in power federally, lost another seat in 1980, and then rebuilt in 1983 opposite the Hawke Labor federal government.  Against the run of the pattern it did gain a seat in 1997 during the Howard Government's first term, but it was then defeated in 2001 with Howard in power, and lost more seats in 2005 ditto.  The Howard government in its middle to later years seems to have been especially toxic to state and territory conservative campaigns.  The CLP then rebuilt in 2008 (assisted by the departure of Labor Chief Minister Clare Martin) and won in 2012, both under Labor federally.  In 2016 with the Coalition in power federally the CLP was smashed (though that would surely have happened to some degree anyway).

Overall, the CLP has averaged 66% of seats at elections with Labor in power federally, but only 43% with the Coalition in charge in Canberra.  For this reason a defeat of the Gunner Government by the opposition CLP would be a very unusual outcome.  We know less about the resilience of governments to third-party insurgencies, though in recent years several of the latter in a range of state elections have been hyped and fallen flat.  Is Mills' "comeback kid" story enough to mess up that pattern too?  We'll be finding out soon enough!

I expect to have a live thread here on the night of August 22, with updates through the post-count.  My 2016 count thread was here.


  1. Has the release of HTV cards and final candidate lists changed any of your analysis? CLP and TA are generally directing preferences to each other over Labor (but TA not in every seat). The Greens are surprisingly high on HTVs too, with CLP directing Green>Labor the way Liberals used to before Bandt got elected on those preferences.

    1. Good question. The impact of preferencing decisions may be muted by the NT's ban on canvassing within 100 metres of a polling station (that said the exclusion zone is apparently only 10 metres for prepolling, which is odd). All the same the decision of the CLP to preference the TA is significant and, in my models above, should be taken as giving the TA the edge in all the Labor/TAP contests above, eg Port Darwin in the adjusted model and Braitling, Fong Lim, Katherine and Sanderson in the hybrid model.

      The TAP are not running in Mulka, Arafura, Gwoja or Barkly and are preferencing Labor in Braitling, Port Darwin and Drysdale.

      I would therefore change the following assessments in the poll-based models (which could of course be rubbish anyway as one or both polls could be just completely wrong; they are just a tool for interpreting these polls):


      Araleun: TAP
      Blain: CLP/TAP
      Nelson: CLP/IND
      Port Darwin: CLP/Labor, but probably Labor


      Braitling: TAP
      Brennan: CLP
      Drysdale: Labor
      Fong Lim: TAP
      Karama: Labor (no IND running)
      Nelson: CLP/Ind
      Port Darwin: Labor
      Sanderson: TAP

      In the hybrid model Labor lands on 12 seats with the two opposition parties quite closely split (probably the TAP is better placed.) Something like 12 Labor, 5-6 TAP, 4-6 CLP, 2-3 IND is the output.