Tuesday, October 8, 2013

2PP aggregate restarted: transitional arrangements

Transitional phase closed (Dec 10): This article is no longer current as new 2PP estimate methods have been released.


There has been a small trickle of federal polling under the new Abbott government, all of it so far from Essential (three polls that did not include any pre-election data) and Morgan Multi-Mode (two polls).

I'm restarting my 2PP aggregate in the sidebar, but these early readings will be extremely rough,  because of the paucity of polling and the lack of precise 2013 preferencing data.  Essential have been using 2010 preference data while Morgan's calculations stated to be based on 2013 election data look extremely dubious to me given their primary results.

The methods for the aggregate I started early this year can be found here.  The aggregate was experimental and underwent a lot of mostly minor changes through the year; the aim was to create a simple aggregate method, capable of quick hand calculation where needed, that would reasonably accurately capture the national 2PP based on a range of polls at any given time.  As the aggregate's final projection of 53.5% 2PP to Coalition proved very close to the final outcome I will be retaining the basic method.  However some small changes will be made.  Most of these will not be determined until I have seen the final preference flows by party from the 2013 election, and conducted a detailed review of all polls.

The methods outlined in the link above are being retained for the moment, with the following transitional exceptions.  Some of these are likely to become permanent exceptions.

1. Up to two most recent Morgan Multi-Modes are included at any time.  My provisional view is that the performance of this poll during the campaign was good enough for it to be treated as an established poll.

2. A new Essential (including two weeks of data) is currently weighted at four, whereas another new poll is weighted at five. Only the most recent Essential is included.  My provisional view is that the performance of this poll during the campaign was poor. 

3. Pending full review of 2013 campaign data, no house effect is applied to any of the regular pollsters (Morgan, Newspoll, Essential, Galaxy, ReachTEL, Nielsen). 

4. Pending full review of 2013 campaign data, the global house effect (previously +0.15 to ALP) is not applied.

5. Until the arrival of 2013 preference flows by party, and given that (for once) last-election preference flows were a tad unreliable at the 2013 election, the results of a multiple regression are temporarily used to estimate the Coalition 2PP from the primaries published by pollsters.

For those pollsters that include PUP the regression in use is

L-NP 2PP = L-NP Primary +.092*Green+.569*PUP+.449*Ind/Other+1.1%

For those pollsters lumping PUP with Ind/Other the regression in use is:

L-NP 2PP = L-NP Primary +.092*Green+.495*Ind/Other+1.1%

(The regressions I actually conducted based on seat figures gave the added constant at 1.0% but I have tweaked them to compensate for slight errors possibly caused by three-cornered contests.)

Note that these formulae do not mean the Coalition really only gained 9.2% of Green preferences.  A portion of the Green, PUP and Ind/Other preferences is likely to be swallowed in the constant term of the regression.  It is also possible that, for instance, the multipliers for Green and PUP are polarised by a tendency of Ind/Others candidates to preference the Coalition in seats where PUP performed well, and preference Labor in seats where the Greens polled well.  Therefore it should not be expected that the multipliers too closely resemble the real preference flows. 

The opening bid of 50.5 to Coalition is extremely rubbery in the circumstances and is strongly skewed to Morgan because Essential is in the doghouse as a result of my provisional view of its campaign polling.  However when processed using the above regression, all the five polls available so far show either equality or weak leads for the Coalition (between 50 and 51.5).   Morgan and Essential at least are returning very lukewarm primary vote figures for the new government (indeed probably the weakest polling for a brand new government relative to its election result in the history of modern polling).  It will be interesting to see whether or not other polls confirm this once the ALP has elected its new leader.

I will be conducting a full review of the performance of all polls once the final two-party preferred result is known.  It's a shame it takes so long to arrive. 

I'm retaining data for all polls included in the aggregate.  There might be graphing, eventually!

A new full methods post will appear once the dust of party preferences, the final 2PP and the performance review of polling has settled.

A Closer Look At Morgan

(some corrections made to this section)

Andrew Iliadis on Twitter noticed something I had not - not just one but two of the Morgans in the last couple of weeks of the 2013 campaign dropped the face-to-face component.  (I originally thought it was all of them, but Adrian Beaumont advises that Aug 31-Sep 1 was business as normal - full documentation of it is missing from the Morgan website, apparently for technical reasons.)

A quick summary of the performance of Morgan Multi-Mode 2PPs (both their figures by last-election preferences and those by respondent preferences) as measured against my aggregate and going from the return of Kevin Rudd as Prime Minister is below:

Figures in two rightmost columns are average lean to ALP.
During the period following the Rudd return, for a month or so Morgan delivered figures that were Labor-skewed even when measured by preferences from the last election, and wildly so when measured by respondent-allocated preferences.  This peaked with the July 5-7 poll which was 52.5 to ALP by 2010-election preferences and 54.5 to ALP respondent-allocated (my aggregate at the time was 49.8 for Labor). 

In August however, Morgan's figures using 2010 election preferences perfectly tracked the aggregate while their respondent-allocated figures had a one-point Labor skew. The last-election-preferences figures also tracked the aggregate well while Gillard was Prime Minister.

It would seem that what happened in the first few weeks of the Rudd return with Morgan Multi-Mode was an aberration: that, for whatever reason, its methods registered enthusiasm for the return of Rudd much more strongly than any other pollster.  In August and the first weekend of September, Morgan's 2PP results by last-election preferences were very normal (and very consistently so: +0.3, 0.0, 0.0, -0.2, +0.6 compared to my aggregate.)

A general issue with aggregating polls from the 2013 campaign is that the established polls tended in the end to (on average) get the 2PP more or less dead right because two minor sources of error cancelled out.  They slightly underestimated the difference in primaries between the two sides, but the flow of preferences to the Coalition was weaker than in 2010.   If the 2010 flow of preferences had repeated then the Coalition would have got about 54.5% 2PP*; as it is, it looks likely to settle 0.8-0.9 points below that once the non-classic divisions are added.

(* This estimate uses the 2010 preference flows here, including adjustment for three-cornered contests.)

I'll be looking at this closer later when deciding what size of global correction for the average leans of all the pollsters that I aggregate will be required.  It's important not to repeat the error made by some of placing too much weight on a single election in doing this, but whatever I decide is not going to make a huge difference anyway.

ReachTEL added: At the current stage of 2PP data-deprivation I'll throw in anything I can get!  The new ReachTEL commissioned by the Fairfax press canvassed voting intention in Victoria and NSW along with ALP leadership views.  In both NSW and Victoria the poll shows a primary vote swing to the ALP and against the Coalition compared to the election (+2.6/-0.5 for NSW and +0.8/-0.7 for Vic).  I model the NSW result as 52.7:47.3 to Coalition off the differences in primaries for the various groupings and the known NSW 2PP as adjusted for the sole non-classic division of New England.  For Victoria there are lots of non-classic divisions, which makes things messy, and I model the 2PP as 49.3:50.7 (ie Labor slightly in front) using the multiple regression model above.  The Victorian 2PP estimate isn't critically important since it is easy to estimate the Victorian swing off the primary swings.  All up I get an estimated swing against the Coalition since the election of 1.5 points in NSW and 1.1 points in Victoria.

As this ReachTEL covers 57% of the electorate I've applied that swing to the likely national 2PP and thrown it in at 50% of normal weighting.

Another Morgan Method Shift (Nov 7): The latest Morgan "multi-mode" uses only face-to-face and SMS polling without online polling.  It also records a result (50-50 by respondent-allocated preferences, and I believe similar by 2013 election preferences) that seems suspiciously good for the ALP.  It's possible that this poll has a much greater proportion of face-to-face sampling than usual.  I've decided to treat this form of Morgan "multi-mode" as essentially a new poll and leave it out of the aggregate pending sufficient tracking data on its performance; if there are three polls with this combination of methods I will back-include it.  The difference is that without this poll the aggregate gives a reading of 53.0 to Coalition; with it 51.8.  Note also that Bludgertrack has reopened with an opening offer of 53.6 to Coalition.  I am still waiting for the AEC's final data uploads so I can calibrate the performance of polls in the last election campaign accurately.  (I could do a rough job now but I'd rather just do it once!)

Nielsen Bomb (Nov 24): The new Nielsen created a strange situation in which the aggregate was better for Labor with Morgan excluded than with Morgan included, although Morgan is only suspended because of suspicion that it is favouring Labor.  While this situation persists (which may be for just a few days) I am using whichever of the two is strongest for the Coalition, so at the moment the last two Morgans are included.  Figures for Morgan will be back-revised from next week if the SMS/F2F method is used again.

Revised Treatment for Morgan (Dec 2): A third Morgan F2F/SMS multi-mode is out.  It was released as 51.5:48.5 to ALP (both respondent allocated and last election); I make the last-election preferences 51.1:48.9 to Labor.

The three Morgan F2F/SMS multi-modes are on average 1.6 points below the Coalition average for the aggregate. I am applying an assumed house effect correction for this poll series of 1.5 points in the Coalition's favour for now pending further results from this and other poll series.  I may reduce this correction based on subsequent results from the same pollster.

No comments:

Post a Comment