With a bag of polling data bursting at the seams, it’s that time of year again where we not only update our Pollytrend metrics, but run some election simulation goodness to see how the polling data for the last three months would have played out if it was reflected in any hypothetical election.
We have the usual Newspoll and Nielsen polls for the simulations, plus a Galaxy for Qld, a couple of unpublished national polls and unpublished NSW and Victorian polls – giving us an aggregated sample size just shy of 15,000 over the quarterly period from July 1st through to last Tuesday. (made up of around 11,900 from the published and 4000 odd from the others – where there’s nothing significantly different in those unpublished jobbies that we aren’t seeing in Newspoll, Nielsen and Galaxy anyway, it just tightens up our uncertainty a bit)
First up, the trend measures.
Last month’s Pollytrend update saw a small movement back to Labor – small, but the largest movement towards them that they had experienced since the election last year. Suffice to say, that movement back to the ALP troughed out and is now in slight reverse.
The current point estimates on the trend have the Coalition sitting on 57.3% to Labor’s 42.7% on the two party preferred.
On the primary vote trends (conveniently available on the sidebar to your right) we have Labor bouncing around in palliative care on 28.3, the Coalition hanging around 48.7 and the Greens moving up a point or so over the last month to come in at 12.3. This gives us a whopping 9.7% primary vote swing against the government, washing out into a 7.5% two party preferred swing towards the Coalition (off the back of a 5.1% primary vote swing towards the Opposition). The Greens meanwhile are slightly ahead of their 2010 election results by 0.6%.
Looking at the longer term primary vote trends, we get:
Nothing particularly exciting there (especially if you happen to be the ALP!), except to note that the big ALP crunch on the primary vote came back in the May to July period and has been pretty much faffing about ever since. That primary vote level is roughly what we saw Labor achieve at the last NSW State election and, barring Campbell Newman getting caught riding a Citycycle naked through Queen Street Mall in the middle of a crack bender, it’s similar to what we’ll also see Labor achieve at the upcoming Qld state election
Next up – the election simulations. Monte Carlo based, state swing focused simulations using our quasi-event dependence framework etc etc.. you’ve heard the spiel before. This time we’re updating the system to incorporate the new electoral redistributions as outlined by Antony Green over at his place.
If you’re an ALP member – probably best to look away about now. If you thought those trends were bad, you ain’t seen nothing yet.
If you’re a Coalition member, well then – break out the popcorn!
The raw state swing data from the aggregated polling looks like this:
The whole country has moved blue since the election, with SA and NSW leading the charge against Labor. WA (the home of Stephen Smith – being spruiked as a possible Gillard replacement by the very people that caused this mess in the first place) comes in showing the smallest swing against the government.
Starting with the basic distribution of the results of a hypothetical federal election held during the last 3 months (usual caveats of the election results matching the polling data etc – the simulations stabilised out around the 20,000 iteration mark for those in to such nerdery), looking at the number of seats won by the ALP, we have (click to expand)
It’s showing us a result clustered around the 41 to 44 seats won mark in a Parliament of 150 – between 27% and 29% of the seats available. That would give the Coalition around 103 to 107 seats, depending on how they fared against the Indies.
Changing the format of the charts, so it tells us the implied probability of the ALP winning *at least* any given number of seats, starting with the full distribution, we get:
The ALP is rock solid up to winning at least 38 seats, before the distribution starts falling away.
Zooming in to the key results to have a closer look:
The most likely result is the ALP winning at least 42 seats, with an implied probability of 61%. The 43rd seat mark was sitting just under the 50% threshold at 49.5% implied probability. The implied probabilities then start to fall off a cliff, with only a 38% likelihood of getting that 44th seat.
Now here’s where the ouch factor comes in. Comparing the September simulation with the June and March simulations, we can see how the likely number of seats that the ALP would have won have dramatically reduced as the months have rolled on. We’ll run that chart above (giving us the implied probabilities of the ALP winning at least X number of seats) overlayed with the same charts from the last two quarterly simulations (click to expand)
To use an example to explain it – in the March quarter, there was a 60% implied probability of the ALP winning at least 68 seats. By June, that 60% implied probability level had reduced to 52 seats. Now in September, it’s sitting on 42 seats – a full 26 seat contraction over the year from an already election- losing position.
At the critical 50% threshold level, it went from 69 seats in March, to 53 seats in June, down to 42 seats now (as the 43rd seat is just below the 50% implied probability mark) – a 27 seat contraction at the critical 50% threshold.
So what seats would be lost?
I’m going to do something I rarely do and post the individual seat baseline implied probabilities from the simulation. The reason I rarely do this is because it gets taken out of context too often. So to hopefully reduce the likelihood of that happening, it’s best to give a long explanation of what it means.
The simulation results are based on the statistical sledgehammer of monte carlo simulations applied to state swings (with a quasi-event dependency framework so that individual seats aren’t treated as independent events, but are semi dependent on clusters of seats within each state that exhibit historical correlation from the mean state swing, as well as an overarching state swing quasi-event dependence). We need a quasi-event dependence with simulations like this to reflect observable reality and to prevent our results from becoming what is a technical term known as “horseshit”
The reason we use the swing of federal polling in each state as the main driver for the simulations, is simply because the best predictor of the result of any given seat at any given Federal election, is the swing at the state level of that federal election that such an “any given seat” happens to reside in.
Seats swing together more at the state level than at the national level – where a perfect example is the 2010 Federal election. Victoria, South Australia and Tasmania had a swing towards the ALP, while the rest of the States had a swing towards the Coalition.
It’s this state based nature of the simulation, as well as its accommodation of uncertainty (both the standard deviations of the swing within each state and the uncertainty of the polling ) and the quasi-event dependency framework (that allows seats to move together on the basis of both history *and* any localised, regional polling results within a state) that makes it at least as accurate as a standard pendulum approach, usually more accurate (sometimes much more) – but never in the last 20 years of elections I’ve retrospectively applied it to for testing, has it been less accurate than a standard pendulum approach.
The ordinary election pendulum has the property where a swing of 5% doesn’t mean that every seat under a margin of 5% will change. What it means is that the average swing would be 5% and that for every seat under a 5% margin that doesn’t change hands, some seat above the margin will swing by more to make up the weight.
The simulations exhibit this same property, but at a state level.
Using an example – let’s say the South Australian seat of Adelaide held by Kate Ellis. Adelaide sits on a margin of 6.1% and South Australia is facing a swing of 8.8% against the Labor Party, meaning that a uniform swing would see this seat fall to the Coalition and be held by them on a margin of 1.3% . If Kate Ellis managed to hold the seat, then some other seat or groups of seats would need to swing more to make up the weight to give us the average swing.
This seat currently has an implied probability of ALP victory of 36.7% based on this quarter’s polling results. That doesn’t mean that the actual, real life probability of Kate Ellis winning her seat is exactly 36.7%. What the 36.7% implied probability represents is the generic starting probability based on the state swing, the historical variance of the seat of Adelaide in terms of the state swing and geographical factors (in this case, being a capital city seat) with all else being equal.
Of course, not all things are equal.
Most seats have varying numbers of local issues running at various strengths, which will impact the swing in a given seat. Occasionally that impact is large, but often it’s either not very large, or different local issues end up cancelling each other out in net terms as candidates stake out their local issue grounds and it becomes a game of swings and roundabouts, pardon the pun.
Sometimes, opposition candidates end up being complete duds that act to reduce the swing in a particular seat.
Sometimes, some local members are strong local members.
Going back to our example of Adelaide, Kate Ellis is a strong local member, so that would increase her actual probability of holding the seat beyond the generic probability. By how much? Maybe 10 to 15% – looking back at previous elections, strong local members seem to have around a 10-15% boost in implied probabilities of holding their seats.
However, in any given election, for any given seat, the state swing – particularly a big state swing – will generally take out those before it in a manner highly correlated to their generic, seat based, implied probability.
The implied probability of winning is effectively a measure of the *distance* between a given seat’s margin and the swing we are seeing, controlling for what we know (geography, historical variance and historical behaviour) and what we think we know (the actual state swing itself). What it doesn’t take into consideration is local and seat dependent issues at the individual seat level.
Anyone using the results below inappropriately, without taking all that rant into consideration, will now be sent to the naughty corner 😛
So, this is what the implied probability of the ALP winning the seats they currently hold looks like if an election were held during the last three months. I’ve only listed ALP/Coalition head to head seats for the mainland states, and no Coalition held seats are listed because the ALP would only pick any of them up on these polling results through some black swan event.
All up – 14 kinds of gruesome for Labor.
BTW – dealing with a spam problem, so some of the comments from new posters might be held up from appearing for a small amount of time.