Because I have way too many things to do right now I decided in my usual fashion to do one more that isn't any of them! Introducing my 2025-2028 federal 2PP polling aggregate, which at this very early stage sits at 56.2 to Labor, with an overall pattern of basically no 2PP movement since it had enough data to wake up on 29 June. The above is a 7-day smoothed aggregate though it has been as high as 57.2 on individual daily readings, and as low (a 0.8 point outlier lasting one day only!) as 55.1. By the end of the term who knows if 2PP will even still exist the way the Australian right are going after this year's drubbing, but for the meantime, here we are. Differences will be detected with aggregates that use pollster-released 2PPs (these tend to have Labor losing support, though I don't believe that is really happening yet) and also my estimate is currently running about a point below Bludger Track but with a similarly flat trajectory.
The aim of the aggregate is to present a frequently updated figure for what the current polls should be taken as saying collectively about the state of the two-party preferred contest. This is never a prediction or a statement that the polls are right, it is just putting a number on where they're at.
This aggregate works quite differently from previous aggregates that had a simple 5-3-2-1 week of release formula, and does so mainly because of the increasing frequency of polls with long in field dates or late releases. The mathematics are kept simple enough that I should be able to understand if something is going wrong, but are no longer readily hand calculable to make my treatment of data less chunky and arbitrary. The working of this year's aggregate is below:
* Last election aggregate: The poll is based on my last-election 2PP estimates based off primary votes released by pollsters, using the formulae in this post (unless we get better data on Bradfield). 2PP figures released by pollsters themselves do not affect the aggregate. In the 2022-5 term there was vast speculation about preference flows to the Coalition being stronger than in 2022, but this amounted in the end to virtually nothing, the latest crushing victory for last-election preferences over other approaches (though admittedly the latter had come around to the idea preferences wouldn't shift by the end).
* Update frequency: The aggregate in theory updates daily, but it often doesn't change if no new polls have been released for a while. I will only usually check or log values in real time when a new poll comes out or at the end of the polling week (Friday midnight), and even then not always if too busy with other things.
* Age weighting: A poll is added from the day of its release. It is given a weighted data age related score which is based on a 2:1 weighting of the youngest and oldest days in field (these will be estimated if not known). Outside of campaign periods this is set at a maximum weighting of 5 for a poll where the weighted age score is 7 days or less, and thereafter decays at the rate of x0.618 per week (spot the golden ratio!) calculated daily. Polls discovered long after release will probably not be back-included.
* Anti-swamping: Only the two heaviest age weighted polls by any pollster are included at any one time and the rest are dropped. If a specific poll is swamping the sample by releasing weekly it will be downweighted. The most recent polls by pollsters that release irregularly are not dropped from the sample (unlike my old aggregate) but will soon be downweighted by data age to the point of having virtually no influence.
* Accuracy weighting: A poll is weighted by an accuracy statistic calculated from its table rankings in all included elections from the 2022 federal election onwards. This will be updated after fresh elections that are included, however an election must be polled by at least four different pollsters to affect the rankings. This works on a scale where a poll that consistently comes last is weighted at 0.50, a poll that is consistently midfield is weighted at 1.00 and a poll that consistently tops the table is weighted at 1.50. (As I start this aggregate Newspoll has the highest weighting at 1.35 followed by Redbridge at 1.11). A new pollster gets a weighting of 0.8 and a pollster that has less than five elections is capped at 1.1 for one election, 1.2 for two (etc). Party/lobby group commissioned polls are excluded, as are polls I consider to be junk polls and polls banned for ludicrous incompetence. There is no sample size weighting except that any poll with a sample below 900 is downweighted by 50%.
* House effects: Polls may be assigned house effects based on their average difference from the aggregate during the term, or based on the average of the last six differences in the case of a strong appearance of a temporary divergence. This will only be done where the divergence is at least half a point.
* Inclusion: Polls must include at minimum primary votes for both majors and the Greens. Polls that do not also include a One Nation breakout are downweighted by 50%, and are also taken out behind the shed and flogged. Polls included at present are: Newspoll, YouGov, Redbridge, Resolve, Roy Morgan, DemosAU, Essential, Freshwater, Spectre, Ipsos, though not all have yet released results in this term.
* Automatic reset: The aggregate resets if there is a change in Prime Minister.
Other aggregates
Bludger Track - BT aggregates primary votes then determines a 2PP from them.
Wikipedia - based on pollster-released 2PPs
Mark the Ballot - also based on pollster-released 2PPs.
Updates
Substantial updates to methods will be logged here if there are any.
Heyo!
ReplyDeleteWas curious about your weighting for PHON given their relative surge recently?
PHON preferences are allocated .2550 to Labor / .7450 to Coalition. I just take whatever the pollster has published as a PHON primary and convert it according to that. I probably won't be doing a primary-vote aggregate, so I suggest people use Bludger Track and Mark the Ballot for that.
Delete