Tuesday, October 6, 2020

The Queensland Poll Drought Is Finally Over

Note added 16 Oct: The Newspoll just out is virtually identical to the YouGov poll below so I have nothing to add to this article on its behalf, save that it is not just one poll saying this now, but two by the same company, and with much less time remaining.  It should be noted though that the Greens primary is down 1% on the YouGov poll, which makes the 52-48 2PP off those primaries rather than 51-49 seem more tenuous, especially with the information that the UAP is polling something in the Others tray.  

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

YouGov 2PP 52-48 to Labor

Projected seat result if this poll is accurate: ALP win, but majority status touch and go (approx ALP 48-9 LNP 37 KAP 3 Green 2-3 ON 1 Ind 1 - almost no net change from 2017)

Over three months since the last, and less than four weeks til polls close, we finally have another public voting intention poll for the 2020 Queensland election.  Finally, for a week at least, I will be able to scan social media in search of polling data in some peace without the Violent Femmes singing "WHY CAN'T I GET JUST ONE POLL?" or the Angels responding "AM I EVER GONNA SEE A POLL AGAIN?" in my head.  The release of a rather large statewide YouGov sample that has Labor leading 52-48 finally gives me something to apply a statewide poll model to so I can talk about modelling and this election.  Though in comparison to 2017, there's really nowhere near as much to say.  


And, well, it's still just one poll, and with four weeks to go, so even if it's accurate now, things can still change.  YouGov's previous combined robopoll/online method (abandoned late last year) had an average final poll 2PP error of 1.8 points (still good by world standards).  The new one may well be better, but this will be its first serious test.  I do not assert that the polling is necessarily perfect or predict on its account that Labor will win, but I am here to say what sort of seat result it should convert to if correct.

The basics

Although the Courier-Mail touts this as "One of the largest weighted opinion polls ever in Queensland", the word "public" should have found its way in there somewhere, as there have been several larger union robopolls.  Nonetheless, a sample size of around 2000, using YouGov's new methods that are designed to cut down on the need to scale up hard-to-reach demographics, is not small fry.  

The poll has Labor leading statewide 52-48 off primaries of 37 each for the major parties, 12% for the Greens, 9% to One Nation, 2% to KAP and just 3% for Others (I always doubt Others will be quite that low!)  The suggestion is that primary votes are leaving One Nation and going to the major parties but that there may be also some small movement from the LNP to Labor.  This becomes clearer in the regional breakdowns.

The regional breakdowns have Labor leading 57-43 (-1 compared to 2017) in Greater Brisbane, the LNP leading 53-47 in regional Queensland (+1) but the LNP's lead in the Gold and Sunshine Coasts shrinking by three points to 54-46.  In the latter, the primary vote motion appears at first glance to be from One Nation (down 5) to Labor (up 6) and several people have queried this.  However this would actually represent One Nation losing votes to both the LNP and Labor, but the LNP (up three) would also lose some of its gains to Labor, thereby swelling Labor's gain compared to 2017.  

The 2PP estimate assumes preferences will flow in roughly the same way as they did in 2017.  There's some potential for this not to be the case - for instance Katters Australian Party voters might swing to preferencing the LNP, or the United Australia Party might bob up for a percent or so statewide which might flow that way.  However it's also possible that Greens preferences, which have a history of flowing somewhat weakly in Queensland state elections, might flow more strongly to Labor.  

The poll shows Annastacia Palaszczuk with a +30 net rating, to Deb Frecklington's -3, and Palaszczuk leads 48-22 on the skewed Better Premier metric, which favours incumbents.  As noted in my coverage of the recent leadership Newspoll, leaders this popular, historically, do not lose.  However, COVID-19 is a potential exception because huge personal ratings do not reliably translate into huge voting intention leads during it, and the latter are what drives election victories.  So we should not see Palaszczuk's personal ratings as a signal strong enough to disregard the closeness of this poll.  

Seat modelling

My usual approach to converting a notional statewide 2PP (from a poll, a poll aggregate, or anywhere else) to an expected average seat tally is a conditional probability model.  It applies basic corrections to seats for house effects based on the retirement of sitting members or the gaining of new personal votes, and then assumes that all seat swings vary within a standard deviation (usually of 3%).  This is a basic version of the more deluxe models used from time to time at Poll Bludger, and usually produces very similar results.  It avoids some weaknesses of the uniform swing/pendulum model, which can overestimate seat losses by a government that gained seats at the previous election, and can make errors if a large number of seats held by one party have margins just on one side or the other of the expected swing.  A basic finding of my model here is that if the election is relatively close, Labor should have a slight advantage in converting 2PP votes into seats, as a result of seats that they won from the LNP at the previous election and recent LNP retirements.  However, this is only worth about half a seat.  

For this model's purpose I exclude all seats won by crossbenchers (as such at the time) in 2017 (all of which I expect them to retain, though with considerable doubt about Mirani).  And I am only interested in the 2PP ALP vs LNP margin, or my best approximation of it.  

This is a much simpler model than the 2017 models I used to try to get a handle on the One Nation vote, which was a very difficult task.  If the One Nation vote falls as much as suggested, then any One Nation seat wins, or seats where they trigger an exclusion order issue that determines the outcome, become unlikely and very difficult to predict using a model.  The high success rate of crossbench MPs in holding their seats generally might mean they hold up better in Mirani than elsewhere, but so few One Nation MPs have ever survived a full term within the party that we have very little to go on for them.  Note also that this model does not seek to project the chances (if any) of independent challengers - one seat of particular interest is Oodgeroo, but there is no data basis for projecting it.  As concerns Katters Australian Party, I am simply assuming that they hold their existing seats, but Poll Bludger poster Bird of paradox raises the possibility that in the fascinating seat of Thuringowa, if One Nation's vote collapses, the Katter candidate (Julianne Wood, apparently a high profile anti-crime crusader) could win that seat.  

This is what the close end of the model looks like assuming a 2PP of 52-48 and a uniformly distributed swing around the state.  Raw is the seat's raw 2PP margin, Model is its model margin under these assumptions, and Prob is the probability of it being retained given the assumptions that the model makes.


The model projects Labor losing about two of its own seats (1.8 on average) and the LNP losing about four of its (4.2), though it doesn't tag any seat as particularly likely to go, and some of the individual seat odds (especially around Cairns and Townsville) are overconfident because of the dubious assumption that the swing varies only randomly from place to place. Overall it projects Labor to win the 2PP in 50-51 of the 87 seats in the model, based on this poll.  However, assuming the LNP stick to their pledge to preference Labor last everywhere, and assuming what the poll says about the Green vote in Brisbane is true, the Greens should win South Brisbane and may win McConnel, so Labor might end up with about what it has now (48-49, though variation in individual seats could put it a few better or worse than that).

Now here's a version that takes into account the regional swings predicted in the poll's local sampling.  This version assumes a 3% swing to Labor in the Gold and Sunshine Coast seats and a 1% swing to LNP everywhere else.  While this picture is not surprising given general speculation about likely areas of swing, it sometimes happens that everyone expects a swing to break in a certain way, the polls say it will do so, and then it doesn't.  

In this case Labor on average loses the 2PP in 3.5 Labor seats, and the LNP in 5.1 LNP seats. Labor holds only one of the seats that are treated differently, and the LNP holds a stack of them that become probable Labor pickups or tossups.  But the projected swing pattern shores up the LNP's regional and Brisbane seats while making Labor weaker in almost all its seats (Mundingburra and Townsville probably fall and a few others become very shaky), so the pattern is actually slightly better for the LNP's seat tally overall than the basic version.  Labor comes out with 49-50 2PP wins, and is touch and go for majority government if the Greens take their two target seats in this version.

A note re McConnel: The 2PP model predicts that Labor will win the 2PP but doesn't project the contest with the Greens.  LNP preferences are unlikely to be relevant in this seat.  The swing in the Brisbane section of the YouGov poll on primary votes is not quite enough to put the Greens into second if it is uniform, as it would only cut an existing 5.9% 3PP margin to about 1.9%.  However surges in the Green vote (if they occur) are likely to concentrate in target seats.  

Whatever, while Labor could easily lose (or increase) their majority on this poll, it would take some very creative slicing and dicing to fashion it into an LNP victory.  Unless this poll is wrong, the LNP have work to do if they are to win office.  Indeed, a few more like this would damage their claim to be able to win majority government.  A perception that only one side was able to win majority government could be damaging at this time.

Something I do not consider at this stage is One Nation preference shifting.  If One Nation maintain a policy of preferencing against incumbents, except the ones they like, then there could be some seats such as Aspley and Redlands where the preference flow for Labor is more difficult than last time.  At this stage One Nation have not announced candidates for either, and it is unknown if they are running.

Betting

It will be interesting to watch whether betting markets scramble to repair their webs in the wake of this poll.  Until this poll, markets have been slightly favouring the LNP, and have echoed a general assumption that the LNP is ahead with around 51-52% of the 2PP vote.  It is unclear what basis for that assumption, if any, exists beyond public polls having shown this in June and July.  As of a few days ago one market had Labor gaining nothing and was projecting Labor to lose Apsley, Barron River, Mundingburra, Mansfield, Townsville and South Brisbane with ties in Redlands, Cairns, Keppel (with One Nation) and McConnel (with Greens).  Relatively little has changed as I write.  Another smaller market, however, disagrees about several of these seats.  

We may see other useable polls soon, so I will comment about what seat breakdowns they would translate to if accurate at the time; if any emerge in the next few days this article will be updated.  

Also see my preview for the Guardian

Update: Economic Question

While I have not seen the exact wording, a question on management of the economy and jobs found Palaszczuk (possibly plus her party) leading Frecklington (ditto) 40-27.  Economy is a natural Coalition strength area so if this is anywhere near accurate it's a bad sign for the LNP and they have very little time to improve perceptions in that area.  Some inauspicious Coalition comparisons - Campbell Newman led 52-36 on such a question and lost in 2015, while in Victoria Matthew Guy trailed only 42-43 before an election at which his party was thrashed senseless.  Possibly mention of jobs has diluted the conservative advantage on this question but however this was worded it's a wakeup alarm for the LNP.

18 comments:

  1. Cook is possibly another potential KAP gain, with the fall of the ONP vote (if it falls bellow the KAP 4PP and the preferences flow KAP-ward), although the ALP primary is higher.

    ReplyDelete
    Replies
    1. Yes, especially if the ALP primary drops.

      Delete
    2. I wouldn`t guarantee the ALP primary falling though, given that this is Lui`s first term, so she could get a small sophomore surge, potentially offsetting downward pressure from wider trands. This might be particularly significant given there may have been an element of protest vote against the ALP at the last election because of the reasons the previous member was booted from the ALP during the term.

      Delete
  2. From memory the ALP vs LNP margin in Thuringowa (from the on the night check count) was significantly closer than the ALP vs PHON margin.

    ReplyDelete
    Replies
    1. Yes, I've estimated it at 2.3% based on average ON to LNP flows but it may have been even closer than that. Normally the ALP vs PHON margin would be closer but Thuringowa was one of the seats where the LNP preferenced Labor over PHON because of an especially contentious PHON candidate.

      Delete
  3. One of the theories going around is a theory I'm dubbing the left/right lens theory. That is, that ALP voters who switched to PHON last time are more likely to go to LNP rather than back to ALP as PHON declines, and are more likely to preference LNP over ALP, as they end up in the right wing sphere of influence on social media. What would just a slightly higher PHON -> LNP flow do? If this theory pans out and the scale of the swing to the LNP maps to the 2017 PHON vote (bigger PHON vote = bigger proportion of the overall swing), would that affect the outcome in your model?

    Similarly would a higher flow of Greens -> ALP preferences compared to last time change the outcome of any seats?

    ReplyDelete
    Replies
    1. Assuming a higher PHON to LNP flow but changing nothing else shifts the 2PP and would be obviously bad news for Labor in marginals with high ON votes (Townsville, Mundingburra, Barron River, Thuringowa, Redlands, to a lesser degree Cairns) on the assumption that ON finishes third behind the majors in those seats.

      Assuming however that the 2PP stays the same but that the 2PP swing varies in relation to the 2017 ON vote then that helps the LNP in all the above listed seats but would also help Labor in its marginals where One Nation didn't run or did poorly. There are just as many of those but on average they are safer on paper for Labor (especially after adding personal vote adjustments).

      The LNP would probably do well in its own marginals out of a shift of that kind because most of its marginals, the exceptions often being Brisbane seats ON didn't contest, had high One Nation votes. Overall such a pattern would help the LNP significantly if it did occur, but it's not especially consistent with the YouGov regional primary vote breakdowns (which could of course be wrong).

      There was no seat where the final margin was below 10% of the Green vote. An interesting one is Clayfield where 76% of Greens votes went to Labor but 89% would have won the seat. It's unrealistic to actually get 89% but if Labor can increase their share then that brings the primary vote swing needed down.

      Delete
    2. Cheers. I've heard reports the Greens ran dead in Clayfield last time and didn't staff booths. This time both Labor and Greens seem to be targeting it (though Greens are on their 2nd candidate and didn't have one for all of September). Andrew Bartlett's Green preferences split 84.6% to ALP last year. I think Clayfield is in the picture.

      Delete
  4. Hi Kevin,
    I'm just curious as to how you got those 2pp figures for non-classic seats, as to the best of my knowledge ECQ doesn't release (or for that matter, calculate) a Labor-LNP 2pp estimate for each seat.
    Additionally, how applicable are the last-election preferencing assumptions used here (especially wrt to the Others' preference, which as far as I can tell broke about 50-50 last state election), considering that UAP is planning to run? (assuming they will likely put LNP above Labor on their how-to-votes)
    Finally, will you be releasing some of the outputs of this model closer to the state election?

    ReplyDelete
  5. Oh, also, I'd like to note that from the data I have, in QLD state elections, the stdev from statewide uniform swing for electorate 2pp seems closer to 4% than 3%.

    ReplyDelete
    Replies
    1. What I do is find the point where a major party was excluded in the preference distribution for the seat, and at that point apply an estimated preference flow for any non-major parties ahead of them. This estimated flow is usually based on the statewide flow of such preferences as were distributed from the non-major parties ahead of them (usually One Nation) - I didn't include such wrinkles as which seats One Nation preferenced Labor or the LNP in.

      Others preferences may well flow more strongly to LNP if UAP is taking a significant share of them, but at the time of the YouGov poll I don't believe that was the case. I expect that if YouGov at least do find UAP polling something significant then they will break out UAP and apply a separate flow to them, maybe based on the federal election. I am also interested in the effect of One Nation contesting virtually every seat, especially if One Nation issues a more or less blanket preference recommendation.

      Whether I release further outputs will depend on whether there are other polls that show different enough results to make it worth doing.

      Re swing standard deviation, Queensland had OPV for a long time, excepting 2017 which was a shift from OPV to CPV, and also had two recent elections with monster swings. I think all these things could drive up the variation in swings compared to CPV-to-CPV shifts in elections with modest swings, though it's also possible Queensland's regional variation or the smallish size of Queensland seats are factors. Also, a small part of the swing variation is caused by personal vote effects which are included separately in the model. In the second version of the model above, changing the SD from 3 to 4 causes each major party to lose about one extra seat to the other.

      Delete
    2. That...sounds like a lot of work for something the ECQ could just calculate for non-classic seats using the ballots it receives. That's some impressive work to calculate preference flows that way.

      I ran a multiple linear regression on polling booths (kindly collated by Ben Raue of the Tally Room) which had the Labor and LNP candidate as their final two, and these were the preference flows I got:
      Greens 0.7895
      One Nation 0.3015
      KAP 0.48
      Others 0.5291

      One Nation's probably has to be adjusted for the fact that they made the final two in more Labor-leaning electorates than LNP-leaning electorates (12 Labor wins vs 8 LNP wins in 2017, if I'm not mistaken).

      On the effect of One Nation contesting virtually every seat, based on my demographic regression of One Nation vote, if One Nation had contested the same electorates in 2017 as they are doing now, they would have probably gotten 2 - 3% more of the statewide vote (so they'd reach about 16%)
      As for the impact of the increase in electorates ON is contesting on preference flows, given that the additional seats they're contesting are approximately an even mix for each side (14 Labor + 1 Green, 15 LNP), assuming they continue their anti-incumbent preference recommendation I don't think the 2pp preference flow would shift much.

      What interests me is this YouGov poll. As is the case with a lot of Australian polling, they don't give much information as to their methodology. In particular, how they got that 9% figure for One Nation is somewhat problematic; in federal polling it's common to adjust for the fact that One Nation rarely runs in most electorates, and at the time this poll came out it looked as if One Nation would be running fewer, not more candidates than last time. Without knowing whether they weighted their One Nation result to account for the number of electorates One Nation was apparently running it at the time, we can't tell whether a statewide One Nation vote would be 9% or more like 12%. The former would be a strong swing against One Nation, likely endangering Mirani, while under the latter Andrews would have decent odds of survival. Furthermore, I'm not sure how they decided which voters are in Brisbane vs the coasts - would voters in seats like Redlands/Oodgeroo/Redcliffe/Pumicestone be considered "Brisbane" voters or "Sunshine/Gold Coast" voters?

      Delete
    3. Fortunately I had already done most of the work in the course of estimating the statewide 2PP. However that estimate was based on preference flows when each party was excluded, which are impure because of mixing in of preferences from multiple sources and hence produce slightly weaker flows than what actually occurs. It is disappointing that some state commissions do not bother throwing the ballots to obtain 2PPs for all seats.

      It is disappointing that we don't have data about how YouGov made certain decisions for this poll. At one stage YouGov was publishing results of the Courier-Mail polls separately but this doesn't seem to be happening now. Regarding Mirani, I have noticed that once crossbenchers actually win seats their success is often not much influenced by their party's overall result; crossbenchers may receive a larger incumbency benefit than major party MPs. So I'm not assuming that if the One Nation vote in general goes down that that seat will necessarily fall.

      Delete
    4. On the minor party incumbency advantage, I used a simple model to estimate the minor's sophomore surge:

      Sophomore surge = (First election as incumbent 2CP) - ((Last election 2CP) + (Minor party swing in seats it contest) - (Opponent swing in seats it contest))

      With the above, I get an average sophomore surge of 11.5%, although a lot of the data in that calculation was from the 1998 One Nation surge and hence I'm unsure how well those calculations would apply to a generic minor MLA. If I exclude those elections, I get a figure closer to 4% (mostly based off KAP wins).

      Tying this back into the One Nation vote - if their statewide vote was 9%, that would imply a swing of 6 - 7% (from my estimate of their statewide vote had they contested more electorates at the last election), which could be enough to cancel out their sophomore surge and endanger Mirani (especially considering a large chunk of the swing against them seems to be coming in the regions).

      On the other hand, if their statewide vote was 12% (adjusting for the seats that they had a declared candidate as of 1/Oct, the final date of the poll), that would imply a swing of about 3 - 5%, probably not sufficient to overcome a typical minor party's sophomore surge. This one decision by YouGov changes a lot of seat modelling, and the fact that they are unclear about it makes it highly disappointing.

      Kevin, maybe the next time you're tapped to write something for news media, you should consider dropping in a line noting the lack of transparency in Australian polling compared to its international counterparts (if they'll let you, of course). After the (relatively large) bungling of the 2018 Vic state and 2019 federal election polling, it seems like our pollsters haven't changed much at all.

      Delete
    5. Sorry, I didn't get a notification for the above comment (happens sometimes) so I have only just found it and cleared it.

      Delete
  6. Hi Kevin,
    Any comments on the new Morgan poll out? While it has similarities to the YouGov/Newspoll polling that we've seen recently, it does have One Nation higher than the Greens unlike YouGov/Newspoll since June.

    ReplyDelete
    Replies
    1. It is hard to take that Morgan poll very seriously because the regional distributions (eg Greens higher in regional Qld than Brisbane) are obviously nonsense. Morgan SMS polling doesn't have a good track record.

      Delete