We are expecting the latest quarterly Newspoll demographic and geographical breakdown to be released any day now, so I thought it might be worth having a look at the history of these polls vs actual election results. If we take the final aggregated poll released by Newspoll for the 1998, 2001, 2004 and 2007 election periods (where the final polls of these type were released at the back end of the last week of the campaigns) and compare the primary vote results estimated for each State against the actual primary vote results for those States, we get an interesting little chart.
But first off, the raw numbers for both the ALP and the Coalition (pretty self-explanatory)
If we now run a scatterplot of these results as well as a quick linear regression for both the ALP and Coalition results across these 4 elections, we get:
The ALP primary vote results are consistently 1.5 to 2 points higher on average in the polling than they turn out to be at the election, across all values of primary votes, and where the exageration happens on average across all states.
The Coalition’s primary vote on the other hand is a bit more complicated, being on average, slightly overcooked in the polls when their actual vote is below 42 and slightly undercooked in the polls when the Coalition primary vote is above 42.
The dotted diagonal line shows where the points would be if the polling was perfectly accurate.
That’s all interesting in and of itself and worthy of some fat chewing, but I was wondering – since we’ll be using these Newspoll quarterly results (combined with Nielsen data) for running the Election Simulations, would it be worthwhile running two sets of simulations where one set is simply the polling data as it is and the second set would be adjusted slightly to account for these polling vs. election result patterns?
9 Comments
Interesting post Poss. Keep them coming (but don’t forget about Nerdy Sunday!).
Yes, I think it would be worth running the two sets of numbers.
I assume as you are going to combine the Nielsen state quarterly figures as well in the simulation, you would need to determine what the adjustment would be including past Nielsens, as the above relates only to Newspoll. Do you have access to those?
Then of course, you could run the results against the pendulum by state, and look at how this translates to potential seat changes.
Boy, Possum, you save me a helluva lot of work!
Ari, I’ve actually got the state breakdowns for each monthly Nielsen poll so I was thinking of combining the quarterly Newspolls with the same 3 months worth of aggregated Nielsen data to arrive at a polling estimate for each State. That would reduce the uncertainty a bit on each figure giving us some tighter estimates on the simulation (because the simulations will also adjust for margins of error)
With the pendulums, I’m not quite sure how to incorporate them yet. I’m hesitant to do what we did last election because too many people that should know better misinterpret how pendulums actually work. They see some seat inside the swing and assume it must be on track to fall when that’s not necessarily the case.
Anyway – glad I could be of assistance!
I understand what you are planning to do, by including Nielsen’s monthly results going forward and incorporating those results with Newspoll’s three monthly results.
That will be very good.
But regarding the adjusted data, if I understand correctly, you also have the Nielsen monthly data by state breakdown going back to 1998, to include in determining the adjustments to the actual election results. So that all the data, going forward and going back includes Newspoll and Nielsen. Is this right?
As for the pendulum issue, I know what you’re saying, but still it might worth calculating, with the requisite caveats.
Perhaps, as each simulation is run, you could email the results to the sitting members whose seats are under threat. Just as a courtesy!!
Great site and keep up the good work.
Do you interpret the 1.5 – 2 point difference as an ALP-effect or an Opposition-effect?
Is there an incumbancy effect?
Yes. The poll closest to the result (except for Qld) was 1998 where incumbency would have been less of a factor. Later their were all those fear and wedge factors which maybe stopped people from following through at the end of the day. Wonder how the figures would look for the Hawker/Keating era?
Hawke