With the new Newspoll quarterly aggregates out, we can combine these State based breakdowns with the breakdowns from the 2 Nieslen polls taken over the same period and run some election simulations to find out what would most likely have happened were an election held last quarter and had results the same as the polling.
You can the see the mechanics of how the monte carlo simulation works over here.
The only difference is that for this sim I used 20,000 iterations rather than 10,000 and have adjusted the simulation to take into account the uncertainty involved caused by the polling itself – via a widening of the standard errors attached to the swing in each state, based on the respective State sample sizes of the aggregated polling.
After running our 20K elections, the results come in like this:
That first chart shows the distribution of the simulation results in terms of the number of ALP seats gained, where the most frequent result was a 17 seat gain to the ALP.
You might also notice that the shape of that distribution is close but not quite perfectly normal (or isn’t perfectly like a bell curve) – that’s a consequence of where the seats sit on the pendulum in each State relative to the swing suggested by the polling in each State.
The last two charts are by far the most interesting – they give the implied probabilities of the number of seats gained, should an election have been held and the polling results replicated (where the sim takes account of the margin of error of the polling)
Those last two charts are basically identical, where only the bottom axis has been changed – where one chart represents the results of the simulation as the number of seats gained by the ALP, while the other represents the results as the number of ALP seats there would be in a new hypothetical Parliament.
To read the charts, simply choose a seat number on the bottom of those two charts, trace up until it intersects the red line, then trace across to the left to read the associated implied probability. This tells us the probability that the ALP would have gained *at least* that number of seats were an election held over the last 3 months where the results replicated the polling.
For instance, there is just over a 60% implied probability that were an election held and the results replicated the polling, the ALP would have gained at least 16 additional seats.
Similarly, using the other chart, if an election were held during the last 3 months etc , there was just over a 50% implied probability that the ALP would have 100 seats in the new Parliament.
22 Comments
lol
The Coalition are clearly going to have to think hard about pre-selections. They desperately need some new fresh talent, but the loss of 17 seats is going to give them a lot to choose from in the party room. They are going to have to be very pro-active about moving on the deadwood in safe seats.
I realise you don’t get from your simulation exactly which seats are likely to be lost by the Coalition (any more than you do from just looking at the last election margins), but do you know if any of the high profile opposition front-bench are in danger of losing their seats based on the recent polls? Maybe that’s more of an Anthony Green question….
Actually that might be an interesting flashy app, a (probabilistic) seat by seat predictor based on the swing required and the poll results (with uncertainties). Might make a good post title e.g. “Utegate doubles Turnbulls odds of losing his own seat….”
Should read posts more carefully, obviously second sentence should read “…goint to give them a lot less to choose from…”
So Poss, given the polls as they now stand – the probability of the Libs gaining seats at the next election is zero?
it would seem so – from this position they need to win 17 just to stay in the same position they were after the 2007 election, to turn that around and even come close to winning would be incredible.
ru
Move along the horizontal axis and subtract from 1.00
Rua,
According to the last 3 months worth of polls, the probability of the Coalition gaining seats in net terms is exactly 0. Not a single election simulation out 20,000 had them gaining any seats as a whole.
All I’d like to know is how likely is it, presuming Malcolm Turnbull keeps not being able to get out of his own way from now until the election, that he could lose his own seat, and become the 2nd leader of the Liberal Party in a row to be turfed out by the electorate?
Do you think the Liberal Party might finally get the message then?
Ah.. well that was fun I guess
Of course Possum correctly observes that the data is backwards looking. And as much fun as this is, its hard to read much into this regarding the next election. Other than Labor is doing fine and probably will win some seats.
Possum, I have a couple of questions for you.
First, is it possible for you to put up a PollyTrend for all the data since the last election? I find the current PollyTrend graphs a bit claustrophobic.
Second, is there a statistic you can apply that essentially measures the predictive ability of the PollyTrend into the near future (up to a month) – based on where the trend line is (or was) headed, versus the actual data 2 or 3 months after the fact. So much polling data looks to me (and I think you’ve observed this too) like a sawtooth – full of straight line trends with points of inflexion. And you’d expect the trend line to be less predictive around the points of inflexion.
btw.. finally figured out how to change my handle here.. this one feels more comfy. yay
Wot?
That’s most erudite of you to say Ms Twain
Just glad to see you can spell ‘erudite”, Poss. Have no idea what the hell you are on about otherwise.
Uh oh, Cud Chewer is no longer Harmless!
That’s not a bad idea on the post-election Pollytrend – I’ll get onto that over the weekend.
I can forecast out the regression lines for Pollytrend, but the vote estimates on the polling wobble around according to the events of the day – so forecasting out short term movements will just be nothing but pure luck if they’re ever found to be right in hindsight.
What Pollytrend does is just show us where current poll numbers sit in the context of the last 12 months worth of polling trends, and can be ramped up to try and determine whether a couple of poll results have changed whatever trend might have been in recent action – it’s not really built for forecasting unfortunately.
I had to look it up. I’m sure I had a rr in there originally.
This is a post for the nerdy end of town – but the link in the article to how the mechanics work was written to help non-stats folks to get their head around it.
Basically, we can run 20,000 simulated elections based on the state by state results of the last 3 months worth of federal polling from Newspoll and Nielsen. With any election, we hear about a uniform swing of X% – but that uniform swing also has a standard deviation of Y attached to it (or the variation around that uniform swing).With these simulations, we take into account the historical variation of each state at federal elections, as well as the extra uncertainty caused by the margins of error from the polling.
When we tally up the outcomes of those 20K simulated elections, we can work out the implied probabilities of an election result delivering any number of seats based on the polling – which are the charts above.
I know, Possum, I’m just having you on. Surely as a fellow bananabender you realise this? Perhaps not. As Madonna once said, Queenslanders don’t get irony. Or something like that.
Anyway, I’ve been pondering a post you did a while back, and my brain has been having issues with this pondering situation ever since. This should come as no surprise, considering how slowly my brain works, being a Queenslander and all.
The post I’m referring to is this one: http://blogs.crikey.com.au/pollytics/2009/06/05/when-analysts-are-no-longer-enough/ where you make the point that statisticians have to be able to speak to the public.
One bit really caught my eye: you were banging on about how “public demand continues to grow for what was only recently very specialised fields” and then said something about how you now have to be a storyteller.
Apart from the horrifying vision of the general public storming the streets demanding more statistical analysis – What do we want? Polynomial regression! When do we want it? In the second percentile! – surely you are going to have to admit that what you do is very valuable but very niche, and that try as you might, the non-nerdy end of town is still not gunna get it. So it might be an idea not to dismiss these people as mentally deficient, which is wot you do, eh.
Apropos a post in the other place.
I see that with 0.0125% certainty, there is a chance for labor to get the magical 110.
at least truffles can sleep easy on that one then.
It’s sometimes hard to tell with these things – although I should have known with you
Monte carlo sims are a bit information dense – anyone going “wot” is an entirely fair call. It also hooks into something else you said:
Of course – it always was that way, and always will be. But that said, while what goes on here is ostensibly aimed at the pointy end of electoral politics that is by nature stats heavy, not everyone has maths running through their veins. I try (to varying degrees of success!) to be aware of that, and try to bring the non-maths folk along for the ride in terms of whatever point I’m trying to get across.
Hence the copious use of charts and data visualisation tools generally (wank term of the week!)
Sometimes I’m just not very good at, sometimes it seems to work (the inbox seems to be the ultimate judge of that) – but I am sensitive to the fact that posts that are information dense often generate questions that I need to answer, usually because I didnt do a good enough job at explaining them to begin with – hence my treating your piss taking as an honest WTF. Questions are encouraged as a way for people to get their heads around some of this if it’s new to them – the very first “non-stating the obvious point” in the moderation policy of this place says “Dont be a smart arse to people asking honest questions”.
I apologise to all and sundry if anyone thinks I’m treating them as mentally deficient (the funny thing is it’s actually the opposite) – I certainly don’t mean to come across that way. I realise that what goes on here is complicated at times, and that you could have an IQ of 180, but if you don’t have a maths background, some of this stuff could be a hard slog.
However the readership here isnt ordinary – it’s simply not the general public. To even stumble across this place you would have to have a particularly pointed interest in politics – certainly to the point where the average reader is far more intellectually curious than, say, the average reader of The Oz – let alone The Daily Terror or things like The Punch.
Because of this I tend to treat everyone as if they know what I’m talking about because on any given thing, they not only probably do, but half of them know more about it that I do!
I swear you’re Canadian Gus
Ms Twain, just to add something.
I suppose what I’m getting that is that while I try to write for the widest possible audience, the potential width of that audience outside of election campaigns is limited to those that give a really big shit about politics – and since such folks are mostly highly intelligent, I try to treat them that way.
I can see how sometimes that might come across as paternal, or even treating people like they’re mentally deficient by trying to hold their hands to walk them through some stats ditty – but I do that not because I think they’re dopes (after all, they’re on a blog that uses the term ‘polynomial regression’ as often as some blogs use the term TomKat…ugh) but to explain to as wider audience as possible the complexities of some of the stuff I do.
Poss/Twain – it’s all good. It seems generally to work…
Poss, perhaps my main point was lost amid my snarkiness. The point is that while you understand what you are talking about and communicate it quite well to a knowledgeable audience, that audience is minute. A journalist on a daily newspaper is attempting to communicate the events of the day to a very wide and varied readership, a readership many thousands-fold larger than yours and almost certainly one that is not interested in the details.
It is perhaps best to stick to what you know, and cut out the politicking that infects some of your writing. You quite rightly criticise News Ltd journalists for allowing ideology to overcome their news sense – although this is generally an issue with subeditors replying to editorial directives, not the journos themselves – so perhaps you should take some of your own advice. You are a blogger and therefore not held to the standards one would expect of a journalist, but you are doing yourself no favours in criticising bias in one sphere, and indulging your own prejudices in another.
Take a look at Nate Silver and what he does on FiveThirtyEight. Certainly his politics are there for all to see, and I note that you make no bones about yours, but his ideology does not infect his writing. William Bowe is good at this as well. You have the ability to be a very good analyst and you do have a talent for writing, punctuation issues aside. Just leave off the bullshit and your value will rise exponentially. Surely a statistician can see the value in this?
I weally weally feel your pain, mark twain
Irrelevance,thy name is limited news