Friday, September 6, 2013

Saturday (Final!) Federal Election Projection Model and Seat Betting Roundup

Note: I will be at themercury.com.au on election night.  A pointer to the coverage will be here.

PROJECTION 10 AM SATURDAY(FINAL)
COALITION WINS ELECTION (probability >99%)
SEAT TOTAL PROJECTION 95 COALITION 52 LABOR 3 REST
(This is an approximate projection only.)

Finally the time is reached for pollsters to release their final polls, modellers to release final projections and politicians to make final attempts to woo the "undecided vote" (which is usually, in real terms, a lot smaller than a lot of people think.)

Late And Final Polls

Essential was first of the regular pollsters to release what seems to be its final result, a 52:48 based on 1035 interviews from Sep 1-4.  If the result for Labor is benign or even an improvement compared to current polling by other sources, Essential will look very good, but if election day is a rout then there will be questions asked about going too early, using a smallish sample and the pollster's trending behaviour relative to others. Galaxy has released what seems to be a pretty safe 53:47, again showing statistical properties more akin to an aggregate than a poll; most likely if they're wrong at all it won't be by much (I am unsure if that's their final offer.)  ReachTEL is providing useful daily tracking and finals from Newspoll, Nielsen and others will arrive later today.  Morgan has released a multi-mode with respondent preferences at 53.5:46.5 (54:46 by last election); they will have a final poll tonight. (Update: this turned out to be just extra data for 54.5: 46.5)



Polls added through the day were Nielsen 54:46 (with Rudd with a meaningless single-point lead as preferred Prime Minister) and Newspoll 54:46 (with Abbott PPM by two points, 45:43). Nielsen had its usual intolerably mild leadership ratings with a mere -9 for Rudd and -5 for Abbott.  Newspoll probably won't be as friendly to Rudd.

There was also a strange mobile-phone only Lonergan showing 50.8:49.2 that ran in the Guardian, with a sample size of 862.  This would be an interesting experiment if it had been run regularly, if it had a decent sample size, and if it wasn't run by a company that has been criticised for its unorthodox scaling methods, but as it is its overwhelmingly likely failure will tell us very little.  Seriously, Guardian, what were you thinking with the small sample size? Could your staff only find $147 on a whip around to pay for this poll or something? 

Great importance is placed on final polls, but the whole business of judging a pollster by a final poll is a bit silly anyway, because really we would like a pollster to provide accurate readings of public sentiment at all times, and not just in the days before the election. Both a pollster's lean to one party or another, and its overall accuracy, are best judged from a series of poll results.  (To illustrate the difference, a pollster that is steadily one point too friendly for one side is fairly accurate and very useful, but has a small house effect.  A pollster that bounces from being wrong in one direction by three points to the other has no house effect, but its individual readings are practically useless.) Final polls may be more accurate than others because a pollster may notice their assumptions being out of whack and modify them to get a result more like the others.  But final polls have great bragging rights importance because they are the benchmark that the general public best understands.

With all these included my aggregate stands at 53.5 to the Coalition. (It was 53.4 when I did the initial run of the modelling below.)

The naive historic data model I've mentioned in previous weeks (based on Newspoll) will still give Labor a c. 3% chance of winning and a mid-47s 2PP, assuming no significant change in the remaining final polls; even if there should be a slew of polls showing 54, that won't change greatly. It's interesting though that Newspoll's biggest errors from final poll to election came at its first two elections, 1987 and 1990 (about three points at each - the error in 2004 was lower once preferencing assumptions are accounted for.)  Although those results were the Coalition's best relative to its final polling, removing them from the model would actually make the projection worse for Labor by tightening the relationship between polling and results -  a 47ish 2PP and a winning chance of 1.5%.

This week's data have not, so far, strongly confirmed the downwards pattern of the last five weeks, but nor have they disproven it either, with a particularly nasty Morgan today for the government. The 2PP figure I am giving in my aggregate is open to the following counter-arguments:

* That there might be late recovery by the government.
* That there might be a further late swing to the Coalition as undecided voters make up their minds.
* That there may still be drift to the Coalition resulting from the Rudd bounce still washing out from the system.
* That the modelling of the aggregate off combined state and federal election results may produce a 2PP that is too friendly for Labor because the most recent two federal elections specifically showed this feature in final polling (though in the case of 2007 most of the blame goes to Nielsen for releasing their final poll too early)
* That Palmer United preferences may be more friendly to Labor than the modelling of Others preferences by pollsters suggests, because their votes seem to be coming fairly evenly from the parties and some polls have their supporters saying they will preference Labor.
* That Palmer United preferences may be less friendly to Labor than the modelling of Others preferences by pollsters suggests, because their how-to-vote cards preference the Coalition.

None of these arguments are in my view individually compelling, though the second seems much more realistic than the first.  I am therefore using my aggregate as it stands (and as adjusted by polling received through the day and overnight) as a baseline for my model.  If it is significantly wrong I will be less surprised if it is too generous to Labor, than not generous enough.

Result Projection

I of course project that the Coalition will win the election. I predict they will win it comfortably.  They may win it about as heavily as 1996 or perhaps even more so.

Seat Model

I am attempting here to produce a seat model that forecasts seat results off a combination of the 2PP, characteristics of the seats, expected state swings, and known seat poll data, and that does not use betting odds in any form. I haven't done this on the national level to anything like such a detailed degree before, so it will be interesting to see how it goes; it is experimental and may yet be totally wrong.  Previously my final predictions at national level going back to 1993, generally made on the basis of subjective assessment of national polls and past trends, have always picked the right winner, but with errors in the seat total for the winning party ranging from two seats (2001) to twelve (1996).   I'm especially concerned that in the past my subjective assessments have had a pro-ALP house effect.  In the last seven elections I've overestimated the ALP seat tally six times. 

People who have watched US politics would be dazzled by the performance of Nate Silver's model that correctly predicted every state in the US presidential election, having correctly predicted all bar one the election before.  It is not possible to reliably replicate that success here.  Our seats are much smaller, polling information at local level is sparse and unreliable, and we have sharp objects such as preferences to consider (although at least we don't need to work out who will vote).  Anything within half a dozen seats or so is a good outcome.

Normally for a c. 3.7% swing, a loss by the pendulum of only 13 seats would be expected, leaving Labor with 60, or 61 if Melbourne is recovered.   The Mackerras pendulum has a formidable predictive record, matching the number of seats gained (though not necessarily exact seats) to the 2PP swing within a few seats in 1996, 2004, 2007 and 2010.  But it experienced failures in 1998 (when Labor was short-changed for the swing it received, probably in large part because of sophomore effects from the 1996 landslide) and 2001 (when the Coalition's swing back should have netted it about eight seats but was only good for a net gain of two, perhaps for the same reason.)

Thirteen seats is also the projected loss based on my charts here which show the transfer of seats by swing to or from the government (for a swing of 4.2%, ie 54:46, it rises to 17 seats by the pendulum or 15 by the chart.)

Yet all year we have had seat polling that suggests, whether on an aggregated basis or by electorate, that the government is doing much worse in its own marginals than in the aggregated 2PP.  This is the big problem of this election, and is one of the contributors to a position in which aggregated polls point to seat losses in the low teens while seat totals markets believe Labor will shed around 18 seats (varying slightly by market) and individual seat markets have an even grimmer view (re which see below).

There are actually two reasons to suspect that the seat gain for the Coalition might outperform the pendulum this time, even ignoring seat polls.  The first is that although a swing of the expected size would usually still see the government winning a few seats from the Opposition, there is very little evidence that any specific, or indeed any, such seats will fall.  Neither current state swings nor seat polls have implied it.  And it makes sense that it should be like this because the seats on low margins for the Coalition are mostly protected by sophomore effect, and because Labor barely targeted them until Kevin Rudd returned (and then wasted energy trying to knock off Aston and the like).  The second is that Tasmania is a bit of a special case.  It makes little difference to the national 2PP for its seat numbers because its seats have lower populations (due to constitutional restrictions) and it seems to be facing a localised monster swing. It's possible for the Coalition to get a double-digit swing in Tasmania that when compensated for around the rest of the country amounts to about 0.2% per seat and barely saves anything against the pendulum.

Now perhaps this seat polling is saying that the national 2PP is a point or so out.  Or perhaps it's saying that for once the old saw about one party being clubbed in its own marginals more than the pendulum suggests is actually true.  Or perhaps the seat polls (dominated by robocalls with a small input from Newspoll, which has been pretty bullish for the Coalition lately) are systematically wrong by a large amount even in a way that national polls by the same firms are not.  I just don't believe you should pick and choose and throw away polling data because of a preconception that any of these things must be true or false.

The following are basic details of my methods for this model:

1. My own current 2PP aggregate is used as a base.

2. A state breakdown is produced by scaling the Bludgertrack state weightings here to my aggregate.

3. I additionally modify the Queensland breakdown from the Bludgertrack weightings because I have a different assumption about PUP preferences, and for Tasmania I insert whatever my own estimate is at the time.

4. For each Labor-held seat a notional position is calculated.  This is the swing required to win the seat modified by the current state swing, and with further modifiers added for sophomore effects and retirement of sitting members.

5. If seat polling exists (JWS, Newspoll, Galaxy and ReachTEL only - Lonergan excluded) it is scaled to the current 2PP by modifying it by the approximate 2PP change since the poll was taken.  Multiple seat polls for the same seat are then added.

6. Aggregated multi-seat Newspolls are not included in step 5 with the exception of the combined Newspoll of Dobell and Robertson, for which I assume the 2PP in Dobell to have been two points worse than Robertson and weight the poll at half value per seat.

7. For seats where seat polls exist, a merged estimate is formed using the seat poll and the notional poll.  The scaling is such that a modified single seat poll typically needs to show a swing to Coalition that is on average about 1.7 times the notional Labor margin before the seats corrected 2PP goes below 50 for Labor.

8. If there are multiple seat polls for a seat, the seat polling is weighted more heavily (in the case of the most polled seat, Griffith, the notional position carries hardly any weight.)

9. A seat with a notional 2PP below 49 is assigned to the Coalition.  A seat with a notional 2PP between 49 and 50 is assigned to the Coalition if it has also featured in an aggregated Newspoll that implies it would be easily lost assuming uniform swing across the sampled seats.

10. Other seats with notional Labor 2PPs between 49 and 51 are retained for my individual judgement or lack thereof, taking into account all polling I've seen relevant to that seat.  If I don't have a strong view about a seat in this list, I leave it where it is.

11. No Coalition seats are modelled.  With the exception of Indi there is no Coalition seat for which evidence of likely defeat exists.

The following is the current output for this model:

MODEL OUTPUT AT 10 AM SATURDAY.  BASED OFF AGGREGATE OF 53.5

Group One: Consensus Losses (13)

These are seats where both the notional position and the seat polling point to Labor losing, or one points to a loss and the other is tied or does not exist:

Braddon, Corangamite, Bass, Deakin, Lindsay, Banks, Lingiari, Greenway, Robertson, Dobell, LaTrobe, Kingsford-Smith, Reid

(Cautions: Lingiari - no seat poll.  Greenway - model projects as relatively close and Coalition has bad candidate)

Group Two: Losses By Seatpoll (1)

The notional position implies a weak hold but the seat polling overrides it:

Lyons

(Caution: Lyons - seat polls exclusively by untested company.)

Group Three:  Losses By Seatpoll With Newspoll Aggregate Poll Used (2)
The notional position combined with the seat poll position (if any) weakly indicates a loss and the seat has also featured in a Newspoll aggregate that also weakly implies loss (because Newspoll aggregates don't imply much clearly about given seats.)

 Parramatta, Lilley (2)

(Caution: Lilley - seat poll exclusively by JWS robocall which performed indifferently at last election)

Group Four: Retained for individual judgement or lack thereof (10)

Barton 
Hindmarsh
Franklin 
Page
Eden-Monaro
Moreton
Capricornia 
McMahon
Werriwa
Chisholm

In the current version Barton, Hindmarsh, Franklin project negatively for Labor and the remaining seven project to Labor's favour.  I don't trust the projection on Franklin, while Page, E-M, Moreton and Capricornia all featured in unpleasant aggregate polling.  I'm inclined on this basis to shift Page, E-M and Moreton to Coalition thus making the above a 5-5 split.

The model's projections for individual very close seats will very often be in error and I am more interested in how close the model's seat totals are than how many errors it makes with specific seats.  Indeed I would expect it to get at least ten individual seats wrong.


All up the model currently projects 21 ALP losses. It disagrees with the seat markets in projecting Barton as a loss and in not projecting Brand, Petrie, Blair, Werriwa or McEwen as losses (though all these are well within its margins of error).   Compared to the seat markets it has the limitation that it knows nothing about internal polling (even if I know, I don't include it) and nothing about candidates on the ground.  Betters in individual seats may have superior knowledge of these things and this site should not be used as betting advice on specific seats.

I am expecting, on average, one crossbencher besides Wilkie and Katter, though it's possible there will be none or even three.  Two Coalition seats (Indi and Fairfax) are at significant risk from potential new crossbenchers, and there is also the question of Bandt in Melbourne (I think his chances are better than general wisdom reckons.)  The crossbencher situation is tricky in all cases - at the start of Friday I thought more likely it would be McGowan who got up, now I think more likely Bandt, while the possibility of PUP winning seats is a known unknown for sheer lack of seat polls in Palmer's area.  Anyway, on balance I've docked the Coalition one.

Current projection is therefore: 95-52-3.  This comes with an implied margin of error of several seats, so perhaps Labor will save all of the shaky ones above and emerge in the low 60s, or perhaps we will still see a full-scale rout taking out nearly all the above and a few others and knocking Labor down even to the 40s.  Figures in the Morgan poll just out, with voters who have already voted breaking 60:40 (compared to 53:47-ish last election) are rather ominous for Labor, unless there is some dishonest self-reporting going on.  That said the sample was fairly small.

This model will have at least one published revision tonight, and probably another tomorrow during the day. An election-watcher's guide will probably be the last piece I publish apart from revisions of various pieces.

Around The Traps

This section will list final (or nearly so) predictions by other modellers as I become aware of them.

Poll Bludger: Currently has 90-57-3 off an identical aggregated 2PP to mine.

Mark the Ballot: Aggregation has 88-60-2 off a 2PP of 53.1% but that includes zero-sum assumption, and author explicitly disclaims voters must make up their own minds on that. Author's prediction: "Yesterday I thought it might be in the high 50s for the number of seats won by Labor. Today, with this latest suite of polls, I suspect the low 50s for Labor is more likely. But the high 40s cannot be ruled out."

Pottinger Model Final Prediction: The Pottinger final model by Julian King is out here.  Using quite different baseline assumptions to mine and, by now, almost completely different methods (since historical comparisons have been negated in my current assumptions), it gets an almost identical seat projection of 97-51-2  The Pottinger model comes from a much lower 2PP baseline than mine (presently 1.3 1.8! points different), mainly because of house-effect conditioning from the last election exclusively, but it also assumes swing by the pendulum for the final result.  Mine effectively hedges its bets on the swing matching the pendulum, accepting seat swing as leading evidence that either the swing is non-uniform in a bad way for Labor, or else that my underlying 2PP is too generous.  There is the added possibility that Julian's base 2PP is correct and the swing is non-uniform in a bad way, in which case tomorrow will be the biggest pasting since 1975-77.   (Note: Pottinger model slight final update here.)

AFR Election Explorer: These guys seem to have been doing some similar historical comparison stuff to what I have through the campaign.  They're giving 89-58-3.

electionlab: Seat based model 98-49-3; more coming.  Final model 99-48-3.

Simon Jackman:  50 Labor seats based on seat betting; or perhaps low 50s if Labor lucky.  While Simon notes that the massive swings in 2007 confounded "bookies and punters alike", I recall that most modellers and psephs actually overshot for Labor (this includes me; I changed from a call wrong by one to a call wrong by four a few days out.)  One exception was Adam Carr who tipped just 80 Labor seats, a call which really stood out at the time - and held up much better than most. Jackman's final 2PP is 53.8.

Adam Carr: Tipped 20 losses on Pollbludger last night.

mumble: No exact prediction I can see but shares my model's willingness to buck the pendulum slightly in the direction of low 50s.

theredandtheblue: Yale Stephens is a Liberal blogger whose dislike for Rudd is even more palpable than my intermittent Abbott-bashing.  I've found his comments about polling quite astute.  His prediction as of last night is exactly the same as my final model's (though last night mine was one seat better for the Coalition).

ABC 7:30 Report: No idea what their model was or if there was one but they had about 56 for Labor.

Mackerras: Malcolm Mackerras (famous for bold calls that are frequently wrong, so I'm nervous about being near his seat tally) predicts 94-54-2 with Rudd losing his own seat. I add that Rudd is not completely safe on my model either but I would expect that if he falls Labor will do much worse than 54. 

The Senate:

I have not had time to attempt a nationwide Senate model.  I think it fairly likely that the Coalition will have a path for passing things that will not involve the Greens, but that a situation where the Greens scrape a share of the balance of power is still very possible.  For detailed Senate projections based on modelling the Truth Seeker blog is well worth a look.

Seat Betting Roundup:

One more time!

As the election night seat betting situation is often the subject of erroneous claims years later here is where we stand tonight.

The "Correct Election Result" market projects to 54.6 Labor seats on the assumption that there are two crossbenchers, but the market probably now expects three, so let's say 54.

The exact seat market still reckons 91-94 Coalition seats (average 92.5)

The seat betting market is quite staggering:


The colour-coding again:

Medium blue: A seat in which the Coalition is favoured in all betting markets.
Pale blue:  Coalition favoured in some markets, level in others.
Grey: All markets tied or both parties ahead in some markets and behind in others.
Orange: (none this week) Labor favoured in some markets, level in others.
White: (not too much of this one left either) Labor favoured in all markets.

Bold shows a seat that has changed in colour since Tuesday.  McEwen, Blair and Werriwa have gone from white to blue. Page has gone from white to light blue and Capricornia from white to grey.  Hindmarsh has gone from orange to light blue, and Petrie from orange to dark blue.  Lilley and Kingsford-Smith have gone from light to dark blue.  But wait, there is hope at the end of the tunnel for Labor because someone has realised no-one really has a clue what is going on in the Northern Territory, and made the seat a tossup.

Note that Luxbet are not included in this roundup as their market is shut.

Here is the trend tracker, which shows how seats not on the seat betting radar since Gillard was PM (or in some cases even at all) have appeared out of nowhere as the wheels fall off in the last days of the campaign:


As for the close seat analysis, the pattern of this fall is that today's close Labor seat is tomorrow's dogmeat, but for now the following are close seats still in the Labor fold:

Barton, McMahon, Franklin, Bendigo, Melbourne (over Green), Newcastle, Rankin, Chisholm.  Total eight.

For the Coalition the following, all of which recently crossed from Labor, are considered close: Blair, Lilley, Petrie, Kingsford-Smith, Page, Werriwa, Hindmarsh, Lyons, McEwen, Brand and Lingiari.  Indi (over indie) is also considered close and was briefly tied.  Total twelve.

Seat markets collectively believe this will be an absolute rout, with every Labor seat on 5.2% likely to fall bar a solitary tie, for the sort of result that comes with a 2PP of about 56:44.  Most likely punters are figuring on a continuing surge to the Coalition and also on the pendulum bouncing badly for Labor.

How low can they go?



5 comments:

  1. Very interesting as always KB!

    Things are certainly a bit crazy atm, with a really significant discord between seat polling, national polling and the betting markets. If those are to be believed it's plausible that the ALP will end up with less than 50 seats.

    I don't know what to think!

    ReplyDelete
  2. I suspect Bandt will be safe in Melbourne by virtue of increasing his 2010 primary vote, owing to having been a good local member and prominent nationally, and that the ALP/LNP voters are probably sophisticated enough to buck the How to Vote recommendations from their nominal party of allegiance. I’m wondering if Wilkie has a chance in Denison, and whether you’d blogged about his chances somewhere in the blog archive; winning from third place as he did in 2010 is a rare occurrence at the best of times and I’d doubt that it would be so easy to repeat it.

    Overall, I agree pessimistically with your conclusions, especially since in 2010 NSW held up extremely well for Labor despite the 2PP figures, and I see no sign of that holding true tomorrow. But I’m stubbornly hanging out for that 2% chance you mention. ;)

    ReplyDelete
    Replies
    1. ReachTEL polling has shown Wilkie's primary surging to about double what it was in 2010 and taking votes from everyone. The tracking has been quite steady and if anything moving further in his favour recently. Internal union polls are said to show similar. Unless the polling is totally wrong he should win comfortably. The biggest danger on current polling is if his vote is much lower than polled, and the Labor vote is a bit higher, and the Liberal vote is much higher - then it is just possible for Labor to gift the seat to the Liberals. However the Libs are not exactly trying and it seems a rather artificial scenario to me. I'm not totally certain Wilkie will hold but it does seem very likely.

      Delete
  3. One of the things interesting about a 60:40 split in prepolling is that given around 10% pre polling, the results on the night could well shift another two points towards the coalition in the real counts later on as pre-polls are added in.

    So really, anything more than 49% on the night is likely to trip over to a win given an even spread with those pre-poll results. Not that an even spread can realistically be assumed, either in proportion or in number.

    ReplyDelete
    Replies
    1. This is going to be a big thing to watch for - last election the 2PP moved half a point to the Coalition from the night. Could be more this time. I suspect Morgan's result over-eggs it a bit because its sample size of already-voteds is <1000 and that the true Coalition early-vote 2PP will not be quite as strong as 60:40, but it may not be too far from it.

      Delete