Saturday, August 24, 2013

ReachTEL Says Tas Labor Still Losing Three

REACHTEL: Bass Liberal 58.4:41.6, Braddon Liberal 56.6:43.4, Lyons Liberal 55.8:44.2, Franklin Labor 50.6:49.4, Denison Wilkie 45.5, Lib 24, ALP 18.7

Approx State 2PP 52:48 to Coalition 

(State analysis will be published on Sunday and some updates on age/gender stuff on Monday).

A new Mercury/ReachTEL poll has been released.   To summarise, the poll results, if reasonably accurate, would result in Labor losing three of its four House of Representatives seats and probably barely saving the fourth.  They would also result in Labor losing its third Senate seat, probably to the Liberals. (See Prospects for the Tasmanian Senate Race)

There are two things that can overturn this picture.  The first is a major turnaround in the national situation to a position in which Labor wins the election - which looks highly unlikely.  The second is that ReachTEL federal polling in one or more of the seats is consistently faulty - a view for which there is no convincing public evidence, but we must remember that this company has not been tested at a federal election in Tasmania before.


The third some might suggest - the margin of error of small seat samples: well, it's time to forget about that one.  The ReachTEL results from Tasmania this year have been so consistent that if the polling method is reasonably accurate, and the national 2PP does not improve sharply for Labor, Labor will almost certainly lose all three of Bass, Braddon and Lyons.   I will cover the key question of whether ReachTEL is accurate later in this article.

Poll Questions

I was actually polled in this poll and so I can reveal what was on the menu (can post verbatim questions on request).  Questions were:

1. Federal voting intention.  Candidates were offered in a set and presumably non-rotated order with a button to press for each one and an undecided option.  Each candidate was named with their party (if any) stated.  Some minor Denison candidates were amalgamated as others.

2. Preferred prime minister.  Forced choice Rudd vs Abbott

3. Issue importance.  Choice of jobs, housing, health, the environment, education, the economy or same-sex marriage.

4. State voting intention.  Choice of ALP, Liberal, Green, Ind/Other or undecided.

5. Rate the performance of the state government.  Very good/good/average/poor/very poor

After this just two demographic questions were asked: age (in broad groupings) and gender.  Other demographic indicators could presumably be extrapolated from the distribution of phone numbers against census data.

Federal Findings 

I have converted the primary figures into percentages with undecided removed, and tabulated them below:


PUP = Palmer United, RUA = Rise Up Australia, KAP = Katters Australian Party, FF = Family First, AC = Australian Christians, SEC = Secular Party, OTHER = lumped Denison others (Democratic Labour Party, Sex Party, Rise Up Australia, Stable Population Party).  I have shown the undecided percentage as (und) at the bottom, although I have redistributed them.

The sample size in each electorate is from 541 to 588.  The notional maximum margin of error is from 4 to 4.2 points.  However, in practice, the margin of error is slightly larger because of scaling and the redistribution of the undecided in the above table (undecided results listed at the bottom).  For small percentages the margin of error decreases; for instance for a result of 10% the notional MOE is 2.5 points.

Because of the large number of readings taken it is likely that there will be some inaccurate results in given seats.  The Braddon Green vote looks obviously too low, and the Franklin Green vote a bit too high in the context of the Green decline elsewhere.  The results for the various micro-parties have very little significance generally as most would represent just a handful of voters in each seat.  PUP seem to be a bit more than just a blip on the radar, but not polling strongly enough to be likely to overcome their rather ordinary hand in the game of Senate preference dealing. 

It is not possible to exactly project an ALP/LIB two-party-preferred score for Denison because it is not clear which parties Andrew Wilkie's votes have come from.  I attempted this by two methods.  One involved use of Wilkie's own preferences from the last election, although these had to be estimated because the exact measure of them included preferences from the Greens. (Although Wilkie won, an exact Labor/Coalition breakdown of the votes he had at the stage where there were three candidates remaining can be determined from AEC data).  The second involved extrapolating from the Preferred Prime Minister score.  These methods gave Denison estimates of 57% and 59.2% to Labor respectively.  On this basis, a state average of 52.3:47.7 to the Coalition, an implied 12.9% state swing, is obtained.

Federal two-party-preferred polling has shown that ReachTEL tends to record higher results for the Coalition than other pollsters, but not by much.  At the time of ReachTEL's most recent polls it differed from the average of other polls by about a point.  Current aggregations suggest its results are lately not that different to Newspoll.  Even if the Tasmanian poll is adjusted by a point or so, a swing of 12% is still enormous.

Discussion of previous Tasmanian federal polling by ReachTEL can be seen in the following articles:

Federal Labor Getting Smashed In Bass
Federal Labor Getting Smashed All Over Tasmania
Tas Federal BaByLon still falling?

There is a very steady pattern in all this polling.  Labor was uncompetitive in Bass, Braddon and Lyons under Gillard earlier this year.  The return of Kevin Rudd produced two-party swings of 7 points back to Labor in Bass, 5.4 points in Braddon and 4.6 points in Lyons, based on the Examiner poll of 28 July.  These figures are, on average, more or less exactly equal to the national lift in Labor's fortunes.

Since that poll, however, Labor has dropped back again by 4.4 points in Bass, 1.4 points in the Lyons sample, and gained 0.2 points in Braddon.  The differences between these particular swings could well be meaningless and down to sample error on either end of the scale, but the average drift back of 1.9 points again matches the flow of the national tide.

Another possibility is that Geoff Lyons in Bass has copped some blowback for a bad gaffe in which he incorrectly downplayed the military service record of his opponent, only to follow that up by having his own awards history incorrectly described in his own promotional material. However, I thought the Labor vote in the previous ReachTEL was a bit high in Bass compared to Lyons and Braddon anyway.

Franklin was not polled in July but the small difference between the Gillard-era poll and this one (1.6 points) is also consistent with the Rudd bounce deflating.

Preferred Prime Minister results in the various seats are not of great interest but show Rudd's PPM score running an average of four points ahead of the seat 2PP (slightly more than the normal incumbent advantage in ReachTEL polls.)  

Outcomes

If an election was proverbially "held now", and assuming these polls are reasonably accurate (re which see below) the following would be the results:

* In Denison, Andrew Wilkie would be returned, either comfortably or perhaps fairly narrowly depending on the order of exclusions.  The only thing that makes it any kind of contest is that the ALP have, in a very strange decision that contradicted their attack on Wilkie the previous week, preferenced the Liberal Party ahead of him on their How To Vote card (see The Anti-Wilkie Denison Billboard Stoush.)  This opens up the possibility that Labor, coming third after preferences and preferencing the Liberals, might elect unheralded Liberal Tanya Denison.

On the poll figures, Labor is in third place on primaries, and it would be touch and go as to whether or not they would overhaul the Liberals on preferences.  (The Greens have preferenced Labor on their how-to-vote card, but Green voters aren't as prone to HTV-card sheepery as the majors, so it would come down to what Green voters thought of Wilkie at the moment.)  However, on the poll figures even if Labor are excluded and it is assumed 100% of Labor voters follow the card, the Liberals' Tanya Denison still does not catch Wilkie's primary tally, and the preferences of Greens (including via Labor) and others elect Wilkie. 

* In Franklin, a surprisingly close race is suggested.  If a one-point correction is (debatably) applied for ReachTEL's apparent national lean, it becomes 51.6% to Labor, which would give Labor a more than 70% chance of holding the seat.  What is surprising is that the Liberals, even after a cringe-inducing if probably not especially damaging gaffe of their own (hey there, rockstar) would still be in this one at all.  Apparently not yet a foregone conclusion.

* In Bass, Braddon and Lyons then if the poll results are reasonably accurate, Labor needs a large swing back because it will not save these seats just on margin of error issues.  The most doubt about these must attach to Lyons on the grounds of the massive swing required to win it, the modest name-recognition for the Liberal candidate Eric Hutchinson in previous polling, and it being the seat with marginally the closest polling.  Lyons is also likely to be the most difficult of the three seats to poll accurately because of its dispersed and demographically diverse population.  The point about margin of error here is that while Lyons is only just beyond the implied MOE for the seat sample size, comparison with the July poll decreases the probability that this is a chance of random sample error.   The results fit the pattern of an underlying picture that has turned away from Labor and is now just flowing up and down with the federal scene, with the opposing porkbarrelling of the two parties (and there has been a lot of it!) having little net impact. 

So if this poll is reasonably accurate, with only a modest house effect, then Labor is very probably just losing all these seats.  Yes, it may defy normal expectations that a seat would be lost on a 12.3% margin, but I have mentioned before that Franklin was lost on a larger margin in the 1975 wipeout. This therefore brings up the important question:

Are Local Robopolls Just Nonsense?

Seat robopolls are the subject of heated debate in Australian poll analysis at the moment.  They are generally producing results that seem much harsher than the national or even state two-party preferred polling picture.  Supposedly Labor is trailing by only a modest margin nationwide, but being thrashed with large swings in many specific seats, far more than would be normally expected, even accounting for cherry-picking.  The extreme swing suggested in Lyons especially is an example of the problem.

But this is a complex and messy debate right now, especially after the most extreme robopollster thus far, Lonergan, had its striking finding that Kevin Rudd was trailing in Griffith replicated exactly by Newspoll (to my embarrassment; for more on all that see the comments in updates here.)

Several of the seat-poll results that have been obtained by relatively new robo-poll companies have now been obtained by established companies too, either using robo-polling themselves or by conventional phone-poll means.  At this point, of the three newish robo-pollsters active, ReachTEL is the one not showing an average skew to the Coalition in seat polls, as compared to established polling.  It seems this is about something more significant than just suspicious or new polling methods, and perhaps for once the repeatedly-discredited scenario of the national polls being inaccurate while the local polls - on average, with the usual errors caused by small sample size -  reveal a more accurate picture might have something in it.  Something like this happened in the last US Presidential election.

Looking at the past elections ReachTEL has polled for (and a previous piece had examples from Queensland's state election) I can see there have been cases where ReachTEL polls with small sample sizes have had large errors, but not in a consistent direction.  I can't see anything to indicate that ReachTEL has a large and systematic house effect compared to other pollsters generally.  Because they've never been tested at a federal election, we don't know for sure that their method works well in Tasmania until the night.  But in terms of convincing arguments for why the ReachTEL Tasmanian federal picture must be wrong, I can't find any.

The only other pollster that has been releasing Tasmanian data, Roy Morgan Research, continues to tell a very different story.  The average Roy Morgan multi-mode (combined online, SMS and face-to-face) result by last-election preferences for the last four weeks has been 57.1:42.9 in Labor's favour from a sample of 609 voters, a difference of nine points from this ReachTEL poll, based on respondent-allocated preferences.  The Morgan series with respondent-allocated preferences has been skewing to Labor by at least two points, and even on a result of 55:45, the uneven distribution of swing around the state would probably still see Bass and Braddon fall, while Lyons would probably be retained.

However, while ReachTEL's methods have at least been tested against state election results elsewhere, Morgan's new multi-mode method has not been tested against any election at all.  My concern is that Morgan's methods are more likely to be prone to local glitches in the recruitment of an online panel.  Even if we assume the Tasmanian picture is midway between the two pollsters as adjusted for apparent national house effects (say, 52:48 to ALP) that is still a situation in which Labor would lose two seats and could lose three.  It would be useful to see Tasmanian federal polling by Newspoll too, and I suspect we will see this before election day.

It will be interesting to see what noises, if any, are made by parties concerning internal polling in response to this particular poll.  Labor has been quite insistent that its internal polling does not match the very bleak picture presented by ReachTEL, but has not to this point released any actual evidence to the contrary.

Updates (25 Aug): Recently I was surprised when it was put to me in an interview that Brett Whiteley in Braddon will have a problem getting votes outside of Burnie.  An example of the argument may be found in the Advocate here.  Sean Ford demonstrates that Whiteley performed poorly in the last state election east of Burnie, without dominating the Liberal count in Burnie to the extent that he was dominated by other Liberal candidates elsewhere.  The problem with extrapolating state results to federal results in this way is that state results only tell you about competition for votes within parties.  In this case, while Whitely polled 40.6% of the Liberal vote at the Burnie booth selected by Ford to only 7.4% at the Nixon Street booth, the Liberal lineup contained only two candidates living in Burnie but five living east of it.  On that basis it was to be expected that a Burnie candidate would be less competitive for Liberal ticket votes east of Burnie than the leading non-Burnie candidate would be within Burnie.  The likelihood of a voter east of Burnie knowing at least one of the non-Burnie candidates would be so much higher.  Comparisons involving the distribution of the Whiteley vote should also take into account that in total he polled only half the vote of Jeremy Rockliff and substantially less than Adam Brooks.  (Admittedly, it might be argued this is because he was unpopular east of Burnie - it's easy to be circular here.)

If this sort of state-election analysis illustrated dangers for Whiteley then it would also have illustrated dangers for Leonie Hiscutt in Montgomery at the Legislative Council elections.  It turned out that Hiscutt did not poll badly in Burnie.  Most likely therefore candidate impressions will not have a great impact on the vote, and if it is closer than the polls suggest, that will be so for other reasons (like the house effect of the poll, sample error, or subsequent tightening in voting intention).  Whiteley does have a hardline reputation, but again that's also true of Hiscutt, and it didn't do the latter any damage.

An uncredited ABC report of the poll is unfortunately of poor quality.  It states leads by primary vote rather than two-party preferred, and also states "The ReachTEL poll also found that the Palmer United Party would outpoll the Greens in the north-west seat of Braddon."  Polls do not find that an event involving small percentage differences "would" occur; rather, they indicate a probability; in this case around 65% without adjusting the margin of error for scaling, or perhaps 60% with scaling taken into account.  But it's one of those findings that, while amusing, requires a lot of caution because of the tiny sample sizes of PUP and Green supporters involved.  Kevin Morgan is a chance to outpoll the Greens in the seat because of his profile as a recent LegCo candidate and the Greens' perennially low standing in Braddon.  But it's the kind of thing better called after it happens as polls won't reliably predict it.

Update (28 Aug): Full findings are now up on ReachTEL's website:

Bass
Braddon
Denison
Franklin
Lyons

The differences in issue ratings across all the electorates are quite informative.  Braddon, Lyons and Bass voters are much more likely to be concerned primarily about jobs than voters in Denison especially.  Environmental issues are barely a blip on the radar in Bass and Braddon, and across all the electorates, concern about them is least week among the middle age ranges, rather than the 18-34 bracket as some might expect.  Female voters care more about health services than male voters in every seat.  Same-sex marriage, which I saw amusingly described alongside asylum-seekers as a "baby seal" issue of the election (from a post on Pollbludger), is much more strongly supported by young voters than others.  However, as the recent Nielsen also showed, it's an issue that most people support but few give a high importance rating to.

If you add the percentage who rated Jobs first and the percentage who rated Economy first, you get something pretty close to (usually a few points below) the Coalition's 2PP in each electorate. 

Those seeking hope that the margins in the northern seats might at least be much closer than the ReachTEL surveys indicate should note that ReachTEL did do a small amount of federal polling in Queensland in 2010.  Their polls picked the winner in eight of nine seats surveyed, but those I have found (covering seven seats) tended to overstate the Coalition 2PP by an average of nearly five points.  Those polls had much smaller sample sizes and were taken at a very early stage of the company's polling history, so they are not necessarily a reliable predictor of current performance.  I also think it's worth mentioning that seat polling in the leadup to 2007 by conventional phone polling correctly indicated that Labor would recapture Bass and Braddon, but overstated the margin by which they eventually did. 

2 comments:

  1. After a few days fiddling, I've got my spreadsheet up and going. I've set it up so that each time it runs, it throws a variance on to some nominal expectations I've put in for each of the tickets and simulates the count.

    Originally I went with a 40-40-10-10 (others combined), and adopted some of the figures from above (44-33-12-11) when I saw them here tonight.

    Variance is +- 1/5 of the primary vote at the top end, +- 1/1 the primary vote at the little end, with a centrally weighted distribution.

    Generalising

    - The only 'other' that ever (without artificially depressing the FF vote) gets up is Family First.

    - At 44-33-12-11, Labor gets a third seat about 1 shot in 10.
    - Libs miss their third seat about 1 time in 10
    - Greens get one up about 80% of the time
    - Family first gets one up about 40% of the time.

    About 20% of the time, there is a 3-2-0-1 result. i.e if the Greens miss out, the extra goes to FF, not to Labor.

    Obviously the other vote is historically high in these, but I suspect that this will be the case this time around.



    ReplyDelete
    Replies
    1. The above comments relate to Tasmania Senate projections - readers can also see http://kevinbonham.blogspot.com.au/2013/06/prospects-for-tasmanian-senate-race.html for more of this discussion and my latest modelling of the Tas Senate picture.

      Delete