Sunday, December 3, 2023

2PP Federal Polling Aggregate Relaunched

Introduction (December 2023)

In recent weeks I've relaunched the 2PP aggregate on the sidebar that was a feature here between mid-2013 and the 2019 federal election.  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.  A couple of things encouraged me to do this and the first one was a desire to have an up-to-date figure readily available to media now that things are actually happening (my 52.9 estimate from this article was being quoted after my estimate had fallen below 52.)  

The second is that following the Newspoll 50-50 last week, I am starting to see some usual suspects on social media pretend that the media interest in the Newspoll is voided by Morgan then coming out with 52.5-47.5.   The reality is that Morgan is a pretty bouncy poll from week to week, it had the Coalition just ahead on its 2PP the week before and Newspoll has a stronger track record.  So the commentators are right to take the Newspoll as the strongest yet signal of a return to competitive polling, but it doesn't mean that Labor is doomed or even that Labor would lose an election "held now".   (Even based off the Newspoll, Labor would probably lose just five or so seats to the Coalition and one or two to the Greens.)  John Howard's government trailed for much of the second half of its first term but was still re-elected despite backflipping on a previous promise to "never, ever" introduce a GST.  The first Whitlam government was also re-elected after a term of mostly weak polling.  

Someone is bound to ask what is the point of doing a 2PP aggregate at all, but I would again reply that despite the recent crossbench gains, the 2PP contest remains the main thing that decides federal elections, including 2022.  Perhaps in the future the major party votes will continue to decline and we will eventually reach a point where 2PP is close to useless, but we're not even nearly there yet.  

I've put up some earlier aggregation graphs at various points of this term but this is the first time I have gone through the polling data to this point and done something similar to my previous aggregates.  

The following are the basic rules of this aggregate.

1. Inclusion: Polls are included if they are released unilaterally or commissioned by established media.  Polls commissioned by political parties and lobby and industry groups are ignored.  To be included polls must release primary votes for at least Labor, the Coalition, the Greens and enough information to determine the total for others.  Polls considered by me to be junk polls are excluded.  Thus far the following have done polls that have been included: Newspoll (Pyxis and previously YouGov Newspolls treated as the same poll), YouGov (non-Newspoll), Roy Morgan, Resolve Strategic, Essential, Freshwater Strategy, Redbridge. Others may be added.  

2. Anti-Swamping: Only the two most recent independent samples by any pollster are included.  

3. Preference Methods: This is a last-election preferences aggregate.  I estimate a two-party preferred vote off every poll based on how preferences flowed in 2022, because previous election preferences are historically more reliable than respondent preferences.  If strong evidence emerges based on electoral events or party-specific preference polling that last-election preferences are unreliable, an adjusted figure may also be issued.  (A possible example of this is that One Nation preference flows to Coalition could increase because One Nation voters may like Peter Dutton.)  Where a pollster is known to use previous-election preferences (eg Newspoll) then I'll constrain my estimate to round to the same thing as theirs if I regard their estimates as reliable, but if I have reason to doubt a particular pollster's estimates I'll calculate my own.  

4. Recency Weighting: A polling week commences on Saturday (this day has been chosen based on recent release days for various polls).  Unless decided otherwise (eg if a poll is released too slowly or has an unusually long in-field period), polls released in a given week are weighted 5, those from the previous week 3, those from the week before that 2 and those from the week before that 1.  

During the final three weeks before the next election, polls meeting certain age criteria will be weighted 8 or 10; I am still deciding what the rules for that for the next election will be.  

5. Track Record Weighting: Polls are weighted for track record at elections and other ballots (including the Voice referendum and the Marriage Law Postal Survey).  On the principle that it is better to have something of this kind and not nothing, the weightings are based on average relative position in final poll "league tables" published on this site, in cases where at least four different pollsters released final polls at an election.  The principle of the weightings is that a poll that is always the most accurate is weighted at 1.5 while a poll that is always the least accurate (excluding junk polls) is weighted 0.5.  Polls with fewer than five elections are capped at a maximum weighting of 1.1 for one election, 1.2 for two etc.  Events since 2020 are weighted double and subjective calls are made on which current pollster to count past results to in cases of major staff movement and ownership changes etc.  (eg the current YouGov series does not include YouGov polls from the YouGov Newspoll era, they are counted as Newspoll.)

(By the way I considered using some measure of average error in final polls instead, but the issue with that is that some elections seem harder to poll correctly than others.)

6. Apparent House Effect Adjustment: Poll 2PPs are adjusted if their converted 2PPs display strong differences to the consensus of other polls, and to polls that have a historic record of not displaying substantial 2PP skew at elections (eg Newspoll).  These adjustments are muted when the number of polls by the pollster is less than five.  A house effect may be applied either temporarily or continually depending on the evidence.  The striking divergence of Resolve from other polls since the election has been noted various times on this site.  

7. Activation: The aggregate is not active when the number of polls in it falls below three; during this time readings are interpolated later based on the start and end point of the break period.  

8. Restart event: If the Prime Minister changes then the aggregate is halted with all old data disregarded going forwards and the aggregate restarts when there are three new polls.  This also happens at each election.

9. Retrospectiveness: During this term there have been a large number of cases (mostly Morgan) where data have become available long after a poll was taken (in some cases this has involved previously completely unreleased polls being published, in other cases just more precise breakdowns).  As at its start time in December 2023 the aggregate has included all this data retrospectively, but going forward any new retro-data won't necessarily be added. Likewise while the accuracy weightings for the 2022-3 period include electoral events that have happened since some of the polls, going forward accuracy weighting changes following further events (if any qualify) will not be back-applied.

Morgan 2022 break 

The aggregate is rather light on for polls and Morgan-heavy for the period up to the start of December 2022, from which point Essential resumed and Newspoll became more common.  While Labor maintained generally honeymoon-level polling from the election until the start of September 2023, the mostly back-released Morgan numbers suggest it was initially not that strong a honeymoon.  However, there is a sudden snap in Morgan in the second week of December 2022, at which point the Coalition loses 4% to minor parties and doesn't get it back.  

It looks like Morgan must have changed something in their methods around this time, and their Coalition primaries prior to that (as high at times as 40%) look much too high.  But there is also weak evidence from Resolve and Essential (but not Newspoll) that supports a surge to the government around this time, for no obvious reason although there is a historic pattern that governments tend to do well at the end of the year.  I found compelling evidence to apply a 1.2% correction in Labor's favour for all Morgans up to the first week of December 2022, but not clear enough evidence to apply any correction after that (unless I want to turn it into a Newspoll-centred aggregate, which I don't).  

Other aggregates

There are a few other aggregates around that have been running for a while, generally more advanced statistically than mine (which is a "simple math" version designed for me to be able to calculate manually as required), but with different assumptions.  

The Bludger Track aggregate includes Morgan in its data tabs, but I believe doesn't actually aggregate Morgan, which probably explains why its take for late 2022 is pretty much stable at 56 to Labor without the summer 2022-3 peak that I and Mark The Ballot both get.  

Mark the Ballot uses the 2PPs provided by the pollsters themselves, and currently has a closer race than I have as a result.  There have been some odd numbers out of pollsters using respondent preferences, including one Morgan that had the Coalition in front on 2PP off primaries that were better for Labor than the 2022 election.

Wikipedia uses pollster-provided 2PPs where available and estimates where not, but I don't believe it assumes any house effects.  

I will add notes to this article on any significant further methods decisions, but unlike in previous terms I am not going to log every minor change.   

Methods updates

Resolve and other house effects (April 21 2024): This year's first three Resolves have shown significantly less divergence from other polls than Resolves between the 2022 election and the end of 2023.  The system I am using to deal with this, which I am also using to correct for other polls that have displayed a house effect, is to use the average divergence from the six most recent polls by that pollster.  I've been using this system since February.  

Frequency weighting (June 11 2024): At present Morgan releases every week, Essential every second week, Newspoll and YouGov about every third, Resolve and Freshwater about monthly and Redbridge sporadically.  This tends to give Morgan and Essential more influence on the aggregate than the average poll.  I've therefore included a corrective weighting that reduces the influence of houses that release very often while still recognising that a series that releases very often still provides more data than one that doesn't do this.   For each poll I find an average recency weighting for that series at the end of the week based on its release schedule, setting this at a minimum of the average for a release every four weeks.  I then divide that by the figure for a poll that releases every three weeks, and take the square root of that figure.  No correction is applied for Newspoll and YouGov which usually release every three weeks with some variation.  The corrections are 0.68 for Morgan, 0.78 for Essential and 1.17 for Resolve, Freshwater and Redbridge.  These corrections will be dropped when an election is called.  At present I have not back-run this weighting.  

Morgan 2PPs (July 17 2024): A note that Morgan in its last three polls has released last-election preferences.  So far my last-election estimates agree with theirs so I am using the same released-2PP constraint as for Newspoll.  Also, I back-ran the frequency weighting above for January specifically to reduce the impact of the mid-January Morgan outlier on the term tracking graph (rerunning it elsewhere will have little impact).  


  1. Thanks for providing so much information on your methods. One question - what statistical method do you use to calculate the aggregate 2PP? Just interested whether it is a local regression, weighted average or some other technique? Thanks in advance, Mark.

    1. Weighted average with weighting for recency (section 4) and estimated poll quality (section 5). Designed to be mathematically simple so that I can calculate it quickly whenever needed.


The comment system is unreliable. If you cannot submit comments you can email me a comment (via email link in profile) - email must be entitled: Comment for publication, followed by the name of the article you wish to comment on. Comments are accepted in full or not at all. Comments will be published under the name the email is sent from unless an alias is clearly requested and stated. If you submit a comment which is not accepted within a few days you can also email me and I will check if it has been received.