1. This article seeks to model what two-party preferred vote Labor would on average need to win the 2016 federal election, based on known information about personal vote advantages, retirements and optionally state-level federal polling.
2. A simple reading of the Mackerras pendulum projects Labor as requiring about 50.3% 2PP to form government and 50.55% to govern in majority.
3. However, the pendulum is prone to over-predict how many seats a party will gain for a given swing if that party lost many seats at the previous election.
4. The reason for this is that the new personal votes of sitting MPs in marginal seats make them more difficult to defeat.
5. With this considered, my current model estimates that Labor requires a 50.9% two-party preferred for an even chance of forming government at all, or 51.4% for a 50-50 chance of a majority.
6. These numbers are estimates only and it is possible Labor could form government with a slightly lower figure or fail to form government with a slightly higher one.
7. These numbers also don't take into account the possibility of a substantial number of new non-Green crossbenchers in the Lower House, as it is not possible to model the chances of that happening yet.
It's been a really long time coming (waiting for the current redistribution to be finished and fully documented) but I am finally in a position to roughly model the 2PP score Labor needs for victory at the 2016 federal election, and also to post seat total projections based on what my 2PP aggregate (on the sidebar) says at any given time. I warn that this article is very numbery and thoroughly dry, and comes out at about 4/5 on the Wonk Factor scale.
There is always a lot of focus on the two-party preferred lead in polling, but a quick look at election results shows that it doesn't tell the full story. At least four times (1961, 1969, 1990, 1998, and some sources add 1940) - including twice since the introduction of "one vote, one value" - a federal Opposition has narrowly won the national two-party preferred vote but lost the election. The reverse has never happened (though there was a near miss in 2010 when the Opposition won the 2PP in 76 seats despite narrowly losing the 2PP, but failed to secure government because Independents in two Coalition-2PP seats backed Labor.)
The Mackerras pendulum is one method of assessing what 2PP score is really needed to win - simply line up the list of seats, count off as many as are needed for the Opposition to win, and take the swing of the last domino to fall. However, it doesn't always work. In 1998, the pendulum had government falling with a 2PP swing of 3.9%. The Labor Opposition recorded a 2PP swing of 4.6%, which "should" have given it a 10-12 seat majority, but actually lost by 13 seats. At the time this was assumed to be evidence of a brilliant "marginal seats campaign" by the Howard government, but in fact there were other reasons. One was the erratic preferencing behaviour of the One Nation party, which tended to give Labor lots of useless preferences in safe Coalition seats. Another, however, was sophomore effect.
As commonly discussed here, sophomore effect occurs when a new sitting member faces their first election. Over their first term they have built up name recognition and profile, which helps to get a few more votes than when they were less well known. If they defeated an incumbent from the previous party, even better - at the last election that incumbent had a personal vote, but at this election the new member has one, so the impact is (in theory) doubled.
People often try to disprove the existence of sophomore effect by pointing to some seat or even election (like 1993) where no sign of it was seen, but the evidence of it tending to happen across a large number of state and federal elections is so strong that they may as well deny the existence of gravity. Of course some new sitting members are so terrible that they actually cost their party votes, but the theory is only that a benefit from having a new sitting member will apply on average. On the other hand, estimates of the size of personal votes vary. This chart posted by Peter Brent sets personal votes as 2 points and double sophomore effect as 4, but I find that on average since 1987 MPs with double sophomore effect on their side have only done about 1.5 points better than the national swing. Sometimes the effect has been blunted by redistributions, so I am comfortable treating double sophomore effect as worth about 2 points.
(Why is this so when sitting members appear to have personal votes worth over 2 points based on comparisons with Senate votes? One important reason is that it's common for the defeated MP to recontest against the MP who defeated them. It could also be that some marginals have a tendency to bounce backwards and forwards, and that such volatility could mask personal vote effects.)
The other reason pendulums don't always work is that even once seat swings are adjusted for personal vote effects and state variation in swings, they are still not going to be uniform. They display variation which looks more or less random until you know a lot about individual seats (and sometimes even then). So, for instance, if you have a pendulum that projects, for a certain swing, that Party X will gain ten seats (winning all by at least 1%), while Party Y will retain ten of its seats by between 0 and 1% each, then most likely Party X will actually win some of the seats it is projected to just miss, and gain something like thirteen seats rather than ten.
What Counts As "Winning"
The obvious target for winning the election is 76 seats, so that a party governs in majority. However in projecting what Labor needs to win the election we should also consider what they might need to govern in minority.
Here I have assumed that Clive Palmer will lose Fairfax (if he contests it) and that all the remaining Lower House crossbenchers will be returned. I have doubts about Bob Katter in Kennedy given that he almost lost last time, but this is not very relevant given he would most likely back the Coalition anyway. The remaining three crossbenchers are all left-leaning. It's difficult to see Labor forming government if the seat breakdown is 74-72-4 to the Coalition, but 73-73-4 seems quite plausible. Also, assuming that any elected Greens support Labor, any wildcard crossbench wins are more likely to come at the Coalition's expense (eg New England or NXT in South Australia). So I'm taking 73 ALP seats as a probable winning target (any Green wins from Labor not affecting the picture).
On that basis, the ALP currently holds 54 seats (one has been abolished), though a further three Coalition seats are notionally Labor's. This gives Labor a target of 19 pickups for a probable minority government or 22 for an outright win, probably including the three notionals in each case. On the pendulum, the sixteenth notionally Coalition seat, Gilmore, falls on a 3.8% swing (50.3% 2PP) and the nineteenth, Bass, goes on 4.05% (50.55%).
The question is whether this is actually accurate given the problems with pendulums I've outlined above. As far as I can tell, Labor's challenge is more difficult than that.
The form of model I use aims to find an adjusted margin for each seat and a probability of winning that seat for a given national two-party preferred vote, assuming seat swings vary randomly. The adjusted margins take into account personal votes and the impact of redistributions on them, and can also be adjusted to include differences in polled swing by state (I consider Bludgertrack the best available measure of these) and seatpolls. At the last election, seat polling was very skewed to the Coalition and so I am going to treat it with much more caution in this year's model; however this far out there is really no point including the very few public non-commissioned seatpolls at all.
I make the following assumptions, all of which are contestable but minor changes to which should make little difference:
* personal vote = 1 point, thus double sophomore = 2 points. I have found in state elections that the loss of a rural sitting member has a much greater impact, and William Bowe has found evidence of some difference federally, so I may consider rural loadings later. However the difference between state and federal rural electorates stands to reason because state federal electorates are smaller and also state elections are further apart.
* if more than 5% of voters in an electorate were redistributed into it, then personal votes are proportionally lost from those areas. (When part of a seat held by the opposing party is redistributed into a seat, this means a loss of personal vote for the opposing party in the moved area, which is a small bonus for the holder of the seat receiving those voters. This is, however, more than offset by the moved area being likely to contain a lot of voters for the opposing party).
* residual variation in seat results (beyond what personal votes explain) has a standard deviation of 2.5%, normally distributed.
* whatever advantage one side has from personal votes is redistributed equally across all seats. (For instance if the assumed Coalition 2PP is 53 but the Coalition is assigned a 1-point advantage based on personal votes in 10% of seats, then I deduct 0.1 from the assumed 2PP for every seat, so that the total 2PP will still equal 53.)
* seat swings below state level are assumed independent of each other. This is almost certainly false (swings can have a sub-state local, or urban-vs- rural, component) but for the purposes of a model aiming to predict an average seat total it shouldn't matter much.
For the time being the model has not retired either Bronwyn Bishop from Mackellar or Nickolas Varvaris from Barton, though the former is in well-deserved danger of losing preselection and the latter was reported as not renominating for preselection in time.
These are example outputs for the current aggregated swing of 2.2% to Labor (51.3% to Coalition 2PP). For each seat "Swing" is the swing required for the other side to win the seat according to the revised pendulum as posted by Antony Green, "Adj" is the adjusted required swing, and "Prob" is the model's probability that the party holding the seat wins the 2PP against the other major party in that seat. I treat Fairfax as LNP-vs-ALP, and the remaining four crossbench seats are excluded from the model. Firstly the output for the closest seats without including state polling:
(There are more "safe" Coalition seats all of which the model treats as 100% or 99.9% chances.)
Now if I apply the current Bludgertrack state swings:
The main difference is Western Australia. BludgerTrack has the Coalition polling badly compared to 2013 in WA, presumably because of drag from the unpopular Barnett state government. The Coalition has several seats in the 5-10% band that the state-swing model considers very much in play. Mainly for this reason the expected Labor seat tally under the first version (62.3 seats) is slightly lower than if state swings are included (63.1).
Please do not go putting your hard-earned on Labor in Barton at $1.05 (for instance) on account of these models! Not only is the model output conditional on the swing at the top, but even when seat swings are normally distributed, there will always be reasons why particular seats swing more than others; the model is blind to these until seat polls are included. Overestimates of winning chances in one seat for this reason will often be cancelled out by underestimates in others. There's also always the possibility that the swing isn't normally distributed, or the swing happens to fall much more in one side's seats than the other, so the model could be overconfident about some of the seats it calls done deals.
Overall though, the model implies that all else being equal, Labor's target score is substantially higher than the pendulum suggests. In the simple version of the model, Labor is only expected to deprive the Coalition of a majority at all with 50.4% 2PP, "wins" (73 seats) with 50.9% and wins a majority with 51.4%. For the model with state data included, these figures are 50.3, 50.8 and 51.3. As state swing data can change a lot before an election even with minor changes in the overall 2PP, I'll be treating the 50.9% figure as the target for now, but monitoring how it changes as more retirements or deselections are announced. While this does place Labor at a disadvantage, the target figure is not quite as high as my estimates before the redistribution was all done.
These are only estimates and it's possible that Labor could, for instance, win with a 2PP of 50% or lose with 51.8, just by getting very lucky or unlucky with the swings in certain seats. What I find though is consistent with what should be expected based on the 2013 result. Mainly, the Coalition has a personal vote advantage in all bar four of the seats on 4% or below, making these in-play seats significantly harder on average for Labor to win. When this advantage is subtracted across all the other seats to cancel it out, the changes occur mainly in safe seats.
A final note: this model does not directly assess the probability of the ALP winning the election. Every election we see models that claim to use this kind of probability working (or one derived from betting odds) to show that one side or other has virtually no chance of winning the election. Such conclusions are generally wrong because seat probabilities aren't independent - if a party wins one seat it is not expected to win, then that increases the chance that it has outperformed its expected 2PP result, which improves its chance of winning in other seats as well.
Corrections of any apparent errors (eg typos, or if I have obviously missed that someone is retiring) are welcome. (NB Perth sitting member retiring has been fixed.)
Update 21 June:
Polls (both aggregate and seat) are suggesting the Nick Xenophon Team (NXT) is on track to win some Lower House seats and that these are more likely to come at the expense of the Coalition. It is unclear which side if any NXT would support in a hung parliament of, say, 72-72-6. I have remodelled this situation and as a result reduced Labor's win target to 50.8%. This is based on a model version that includes state data.