Saturday, December 15, 2018

Victoria 2018: Final Lower House Results, Poll Performance and 2PP Pendulum

There is no single method of calculating 2PP for this election.  The following are examples of possible figures:

57.57% to Labor (Uniform swing applied to Richmond - probably fairest method)
57.84% to Labor (Richmond treated as 100% to Labor)
57.30% to Labor (Richmond excluded)

With two-party results for all Lower House seats now available it's time to wrap up my Victorian election coverage for 2018, on a high because at least that's the one house where I can talk about the results without constantly losing my temper at the system.  The article again includes a 2PP pendulum.  While this will be of less use for the future than the 2014 one was, given that there is a major redistribution coming, I think it is still useful for looking at the results, and especially at whether the Coalition was lucky not to lose even more seats than it did.

Final vote share results

The final primaries for the Lower House were Labor 42.86%, Coalition 35.19, Greens 10.71, Others 11.24.  As @sorceror43 notes, excluding Richmond, "Labor won 67.4% of all minor party prefs, slightly down from 69.8% in 2014. That's despite Greens share of all minors dropping from 57.8% in 2014 to 48.4%."  I have not attempted to measure shifts in the preferences of Greens and Others specifically yet, but suspect there was very little change.

There is unfortunately no unique way to calculate the statewide 2PP for this election because of the Liberals' ill-fated decision to not contest the district of Richmond.  Richmond is a very pro-Labor seat on a 2PP basis so removing Richmond from the results has the effect of downplaying the swing across other districts.  While official sources will probably need to give a 2PP percentage based on an actual vote tally, I think that applying the swing from all electorates other than Richmond to Richmond on a notional basis creates the fairest estimate of 57.57% to Labor, a swing after rounding of 5.58 points.  The average swing per seat was slightly lower at 5.43%, as a result of population distortions.  The standard deviation of seat swings on a 2PP basis was 3.84%.

As well as the estimates available by omitting Richmond or treating Richmond as 100% Labor, Antony Green has tweeted a further statewide estimate of 57.7% obtained by applying derived Greens flows from other seats to Richmond.  However this implies an 85.9% (+12.8%) estimated 2PP in Richmond, which is out of step for other comparable seats.  Another rough way to get a handle on the possible swing in Richmond (had the Liberals run a candidate is to look at the Upper House votes).  In the remaining Northern Metropolitan seats the average 2PP swing was 3.2%.  In Upper House primary vote swings, Richmond was 1.3 points worse for the Liberals and 2.6 points better for Labor than the Northcote average swing.  It was also 6 points worse for the Greens, much of which can be explained by switching to other left-wing parties, primarily Victorian Socialists.  On this basis I feel comfortable that the 2PP swing in Richmond would have been around the state swing had the Liberals actually contested.  Perhaps slightly higher if we assume that ALP voters in Richmond were disproportionately likely to switch to Victorian Socialists.

For the benefit of those who care about exact totals, I have the total of 2PP votes excluding Richmond as 1988434 - 1481976, with 44394 formal votes in Richmond. There is a 100-vote discrepancy between the VEC's figures for Carrum on the seat page and in their 2PP total and a 1-vote discrepancy in Albert Park; I have also made some corrections to my own figures because of a clerical error in my figures for Tarneit. Remaining discrepancies with the primary vote statewide totals (a few hundred votes) would be caused by rechecking issues in seats not thrown to a final result.

(As a word of caution about VEC 2PPs for seats that don't go to a full preference throw, a data entry error affecting the margin in Eltham by about 1540 votes very narrowly missed out on making it into the final 2PP.)

Seat results, swings and personal votes

The graph below shows the relationship between the 2014 margin and the 2PP swing in each seat.  The winners of each seat are colour-coded (blue for Liberals, dark green for Nationals, red for Labor, grey for indies and you tell me for Greens).  The gap between the two thick black lines (marked "ALP Gains") is the area in which the 2PP status of seats changed hands.  Uniform swing has been applied in the case of Richmond.  Click for larger clearer version.

Labor obtained a 2PP swing in all bar six seats.  It won the 2PP vote in eleven seats where the 2PP vote was won by the Coalition in 2014.  However these eleven included Prahran (retained by the Greens) and Morwell (retained by Russell Northe as an independent after previously holding it as a National).  As well as losing nine seats to Labor and Morwell to Northe, the Coalition also lost Mildura to independent Ali Cupper.

As a general rule the Coalition vote held up best in rural seats, with no 2PP swing to speak of in most Nationals seats or in Liberal-held Benambra and Ripon.  The swing was modest in Ovens Valley (where the Nationals incumbent had had some unfriendly publicity) but Mildura was a massive 2PP outlier, not only falling to Cupper but also being Labor's largest 2PP swing of the election.

The Coalition saved four seats that would have fallen had the swing been uniform (Ripon, Eildon, Caulfield and Forest Hill) but also lost three on margins above the state swing (Box Hill, Nepean and Hawthorn).  They saved eight seats and lost four by 2PP margins below 2%, and saved twelve and lost six by less than 3%, so it could have easily been even worse.  The 59-29 2PP split of seats is exactly what my seat model (which took into account personal votes, population changes and the variability of seat swings) would have expected for the 2PP recorded.  While Labor fell seven seats short of the 2002 "Brackslide" off roughly the same 2PP, four of those were caused by the seats being occupied by crossbenchers.

In total, non-classic two-candidate results occurred in the six seats won by crossbenchers (Brunswick, Prahran, Melbourne, Morwell, Mildura and Shepparton), two seats won by Labor against the Greens (Richmond, Northcote), two seats won by Labor against independents (Pascoe Vale and Geelong) and one seat won by the Liberals against an independent (Benambra).  The seat of Preston had a 2CP preference throw between Labor and the Greens but it was not determined exhaustively whether the Greens were second in it (though based on the results it is likely they would have been).  In Footscray and Williamstown the Greens may have finished second had preferences been fully thrown, but they weren't.

The non-classic case that presented the biggest problems was Werribee, where independent Joe Garra certainly finished second, but because Tim Pallas had over 50% with two opponents remaining, preferences were never fully thrown.  This creates the weird situation that there is no official victory margin between the candidates finishing first and second.  The VEC had the option to realign the seat as they did for Pascoe Vale and Geelong, but chose not to do so.  Whether this is because they were not convinced Garra was going to be second, or because they saw the result as a foregone conclusion, remains unclear.  In the event that Garra runs again, we now do not know what 2CP swing he needs to unseat the incumbent, although the precedent of Preston suggests the VEC would conduct the quick 2CP count between Labor and Garra in this instance.  I don't think it's ideal that whether or not a challenger gets to know what they lost by depends on an opaque decision on whether or not the count is realigned and, while budget constraints doubtless are a factor, I think this should be looked at for the future.  [UPDATE April 4 2019: The VEC has now conducted a 2CP count for Werribee, with Pallas getting 58.78%.]

As for Labor-Green seats, Labor lost ground and the seat in Brunswick (where its sitting member retired) and, compared to the last election, and also Northcote (which had been disrupted by the Greens winning a by-election, so recapturing the seat by any margin was a good result for Labor). In Prahran, Labor and Green primaries both improved by a similar amount and the Greens kept the seat.  That leaves Melbourne, where there was a small swing to Labor, and Richmond, where there was a larger one.  There was some theory - the logic behind which I never grasped - that Liberal voters in Richmond would be more likely to preference the Greens if the party didn't run at all rather than if it ran on an open ticket.  The Liberals may have been keen to try to get Kathleen Maltzahn into parliament because of the impact of her sex work views on the Greens party room.  But it failed.  The increase in the informal vote in Richmond was 2.22 points compared to 1.31 in Brunswick, 1.21 in Melbourne and 0.08 in Northcote.  It appears that some Liberal voters, rather than voting Green as their party seemed to want them to do, just didn't bother voting at all.

One might expect the huge size of the swing would swamp personal vote effects.  In fact this was yet another election that confirms in spades that personal votes are a theory and a fact.  Compared to the modelled overall state swing of 5.58%:

* in the three seats where Labor had double sophomore effect on their side (defeated sitting MP at previous election), the swing was 11.06%
* in nine seats where Labor had a single sophomore boost (first term MP elected at previous election) the swing was 6.48%
* in in ten seats where Labor incumbents retired the average swing was 4.55%
* in seven seats where the Coalition had a single sophomore boost the average swing was 3.45%
* in five seats where the Coalition MP retired or switched to run as an independent the average swing was 7.00%.

We shouldn't take the first result too seriously as these seats were all part of the hitherto crucial "sandbelt", so regional factors could well be at play (which is a nice way to say things without trying to remember how many "l"s there are in "porkbarrel"). However, that can also be part of why new sitting MPs tend to do well.  And overall this is very strong evidence even ignoring the double-sophomore cases.  In noticing that the Coalition held up well in rural seats, we should bear in mind that they had assistance from personal vote effects in some of them.

Polling Accuracy

The lack of a single 2PP figure makes it a little difficult to compare 2PP polling with the actual vote.  While I have little doubt that ReachTEL would have treated the Coalition as running in all seats,  and suspect Morgan would have done the same, I have no idea what Galaxy would have done about the situation.  I am using 57.57% (Richmond treated as uniform swing) for now, but correspondence will be accepted if any pollster used a different measure and sends me detailed evidence of it. In any case it will make little difference.

The table below lists the state polls taken in November, with ReachTEL "undecided" reallocated and the two halves of Morgan's sample (for which they provided no overall result) merged.  The table also shows the expected preference flow to Labor based on the published 2PPs.  The accuracy scores are in two forms - average primary vote error (error (P)) and the average of the 2PP error and the primary vote error (error (2)).  The polls are ranked by error (2).  Closest to the pin before the poll, and any estimate within 1% before the poll, are highlighted in blue.

What to make of this?

1. Everyone had the Labor primary vote too low.

2. The polls mostly had the preference flow to Labor too low, whether as a result of respondent preferencing (ReachTEL and Morgan) or pollster assumptions or unlucky rounding (YouGov-Galaxy).  Pollsters would have been better off using last-election preferences with separate estimates for Greens and Others.

3. Compared with the 2014 results, the statewide polling wasn't very accurate.  But it did paint a picture (also consistent in YouGov's polls in September and October) that Labor was comfortably ahead and likely to win outright.  Nonetheless the average miss on the 2PP (3.5 points and all in the same direction) was pretty high.  The 2PP error on the final Newspoll wasn't unprecedented, but was the worst I've seen from the Newspoll brand at a state election for several years.

4. It seems only fair to include two ReachTELs (one activist-group commissioned, one media-commissioned) given that YouGov-Galaxy gets two bites of the cherry via the Galaxy and Newspoll brands.  But even we include only the Age ReachTEL and not the more accurate VNPA one (widely considered an unreliable outlier at the time) ReachTEL has outperformed YouGov-Galaxy in this instance.  This would have been more striking had ReachTEL used respondent preferences for the Age poll.  On the other hand, YouGov-Galaxy had superior tracking, showing solid margins in three polls in September and October while the sole ReachTEL in that period did not.

Why did polls miss the size of Labor's victory (the scale of which was hardly unprecedented at state level)?  One possibility is late swing.  There was ample evidence that votes cast at booths on the day swung more than votes cast in pre-polling, causing a bunch of Coalition seats that were projecting for 53-54% to Labor based on booth votes to come back to 50-50 or so by the end of all counting.  However, more evidence is needed on whether this was because voters switched their vote in the final days, or because more conservative voters became more likely to pre-poll upon the requirement to have a reason to do so being removed.  I have heard that internal polling in the sandbelt seats showed a surge to Labor late in the campaign, but many of the seats where large swings happened were never polled by anybody because nobody on either side expected them to fall.

I'm usually sceptical of late swing and a reason for scepticism about it at this election is the shortage of obvious reasons for it.  The big-picture battlegrounds were laid out well in advance and nothing really happened in the last weeks of the campaign, which became dominated by media coverage of candidate malfunctions and other things of little interest to voters.  There might however have been a bit of a Tasmania-style bandwagon effect in which voters, seeing that there was a risk of the Greens winning the balance of power and considering that the Liberals weren't much chop anyway, decided to lock in majority government.

If one wants to ignore possible campaign causes for the systematic error, then the familiar suspects of hedging and herding (playing safe) are difficult to ignore here, though there's no positive evidence of either.  As with the 2014 Victorian poll the final 2PPs were very close together, but in that case they were right and in this case way short of the mark.  However, with only a few pollsters in the field it's impossible to say that this is robust evidence of herding.

In advance of the election, I did note some degree of hunch that the 2PP result for Labor might be stronger than pollsters were predicting for various reasons, but I could not see any basis in the data for predicting a blowout on the scale of what occurred.  The pressure on pollsters to avoid large errors could be a factor here - if they predict a solid win and the result is a huge win, they still look pretty good to the public in general.  But if they predict a huge win and the margin is close, that's a different story.

Something I possibly should have looked at more closely in advance is the historic non-polling equations from my article What Kills State Governments: Age Or Canberra, which if applied to this election would have predicted the Andrews government to win 52 +/- 8 seats.  However I need to update those sometime for elections since that piece was written - Queensland 2015 in particular could affect them greatly.

Seat Polling Flops Again

The uselessness of seat polling is a common theme on this site and this election provided still more evidence of it.  From August on, I am aware of six Galaxy polls of seats that were expected to be 2PP seats, of which five had the right winner but the average 2PP error was 8.3 points with all the errors in the same direction; this even includes a 52-48 in Mordialloc a few weeks out (end result 62.9% to Labor).  When it comes to Labor vs Green seats there was one gem amid the garbage with Galaxy pretty much nailing Richmond (54-46 compared to actual 55.5) but two commissioned ReachTELs were both hopeless in Prahran.  One of them had Labor 5.8% ahead of the Greens on primary votes (though it did nail the 2PP) while the other had the gap at 12.8%; it was actually 0.9% and the Greens won.

This repeats themes of failure from previous elections.  ReachTELs have never had much handle on inner city contests in areas with high Green votes, and the Galaxy pattern here tends to echo what happened at the 2016 federal election, with a picture of low swing and close 2PP results that then didn't eventuate.

Betting Doesn't Beat Polling

This election is a nice example of one where if betting added predictive value and could detect where polls were wrong, it could have shown it, but it didn't.  Betting overall was very consistent with the pattern in poll-based modelling, which was that Labor seemed likely to get a modest 2PP swing but have little to show for it in seat terms.  (The results are consistent with this being a fair read based off the 2PPs pollsters were getting, because Labor only gained one seat by a margin exceeding 4%). The markets on election day had consensus on a net Labor gain of two (excluding Melton), the same as my poll-based prediction, but the actual net gain was nine (even despite one of the expected gains, Ripon, not happening!)  The seat total market was offering $4 on its highest range of Labor-minus-Coalition seat tallies, yet Labor went more than a dozen seats above that mark.  Either the betting reflected bookies' modelling that was no better or worse than anyone else's who was reading the polls, or else any input from the markets didn't help.

2PP Pendulum (With A Few Extras)

Lastly, here's a 2PP pendulum with a few non-classic bells and whistles to fill up all the empty space for seats the Coalition didn't win.  It will become a lot less relevant after the redistribution, as a result of which the Coalition will probably drop a 2PP seat or two to Labor somewhere.  I'm not a big fan of 2CP pendulums in which margins against all comers are treated as the same, because the purpose of a pendulum shouldn't be just to show which seats are close and should be to show which seats would be picked up by a uniform swing from one party to another on a 2PP basis.  This does get messed around by non-classic contests but 2CP pendulums are easily misread.

For the seat of Werribee after looking at comparable seats (there aren't many) I have assumed a 78% preference flow from Liberal to Joe Garra.  It may be he would have done better than that but I would want to see evidence of it.  Andrew Wilkie in Denison 2010 achieved an 83% flow but he was running against a low-profile candidate rather than a state treasurer.

Click for larger clearer versions:

[UPDATE: The correct margin for Werribee is 8.78% vs IND]

That's all I've got planned for Victoria for this year!  There is a Paypal button in the sidebar for anyone who wants to throw money (please only give if you can afford it).  I will soon be turning my attention to the murkier contest that is unfolding in the state north of the border ...


  1. Thanks Kevin - quite interesting. As to Richmond, surely the most plausible guess would be, not the average swing in all other seats, but the average swing in other seats already held by Labor with margins in the 20s and 30s. Looking at your graph I suspect that would be in the 4-5% range rather than 5.63%. Not that it would make much difference to the overall swing estimate or to the fact that it was an utter disaster for the Liberals.

  2. Oh and let me add that the graph of swing vs previous margin is particularly useful. We always hear talk of how many seats would change hands on the basis of a uniform swing, and indeed the number changing hands may _often_ end up close to the uniform-swing-estimate but that's a matter of swings (in a different sense!) and roundabouts. But of course swings are never uniform and a graph like yours reminds us of that fact. Something similar should be included in every after-the-event analysis.

    1. I also found that graph useful, and found myself wondering what the equivalent ALP vs GRN graph would look like.

  3. look what an extra 2.5% would have done for labor