Mayhem in the Marginals 3: Mayhem Everywhere

Advance Summary
1. A new multi-seat poll by JWS Research again provides seat loss projections for Labor that are above what is expected based on national polling and the assumption of uniform swing.  The poll projects losses of 32-35 seats, while national polling points to a loss of around 26.

2. State-of-seat issues may have some influence on this difference, but it is unclear whether the main finding of about six expected extra seat losses is sound, or whether it may result from the polling method having a slight house effect.

3. The pollster's claim that "marginality analysis" adds another three expected seat losses for Labor is unsound as it is based on false assumptions that seats within a given bracket would swing uniformly.  

4. In particular, the firm projects the loss of five Labor seats within the 6-9% bracket by very small margins.  In practice, it is more likely that some of those seats would fall and some not.

5. The poll's findings that voters do not want Kevin Rudd to challenge for the Labor leadership do not demonstrate whether voters want Kevin Rudd to become Labor leader, because the question design is flawed and too restrictive.

6. The poll's core finding that seat losses will be worse than modelled by a "uniform swing" (national pendulum), or by assumptions of uniform swing at state level, is consistent with results from other polling sources.

7. However all of the polling sources producing this finding are open to a similar range of reservations.


=====================================================================
This article continues a series analysing the ECG Solutions/Financial Review series of aggregated marginal JWS Research robopolls.  ECG Solutions is a corporate strategic advice unit linked to Peter Costello and other Liberals - for more info on it see the first article in the series. That first piece, Mayhem In The Marginals, was written when the world was young and Labor was in its last weeks of competitive polling.  I criticised the first poll for its difficult-to-explain discrepancies with the national 2PP at the time, the poll's non-neutral commissioning (which was not immediately obvious), and the confusing presentation of often unsound projections which overstated what the poll was actually saying.  The poll was also criticised by Peter Brent for having too small a sample of young voters.  The second article, Marginals Mayhem 2: Labor Losing The Lot, concerned a slightly less dubious exercise in late March, which nonetheless didn't really tell us much we didn't already know.  The poll analysis also wrongly assigned seats in batches on the assumption that the batched seats would all swing by the same amount.

Now, there's another JWS monster robopoll out - see the report here.  Claimed findings are that in its 0-12 point seats, Labor is facing a 7.6% swing.  This is said to lead to a loss of 32 seats assuming uniform swing, 35 seats based on findings that the swing is higher in the higher range, and 32 seats on a state basis.  It's hardly a sensational claim anymore with Labor's internal polling supposed to show even worse, albeit off the back of state swings that are not confirmed in public breakdowns.

Before discussing the JWS in detail, a brief look again at the current national polling context.  Labor's federal polling position is about as bad as it has been all year, with no net recovery since the late March failed spill.  Current 2PP aggregate estimates include 55.7% for Coalition at Mark the Ballot, 56.5 at Bludgertrack (as of mid last week), and 56.1 on my own rough aggregator.  So there's a swing on of probably around six points or maybe even a bit more.  On a uniform swing this would claim 26 ALP seats leaving Labor with 46, or 47 if Melbourne is recovered. The Coalition is at very short odds in cab 27 off the rank, Bass.

(Incidentally, the relatively worthless Preferred Prime Minister indicator in the latest Newspoll showed Tony Abbott with his largest lead over Julia Gillard so far, and Gillard with the lowest PPM of an incumbent PM since John Howard in July 1998.  A bit over two months after returning that rating, Howard was re-elected, while the biggest PPM lead for an Opposition Leader in Newspoll history belongs to Alexander Downer.  You may see, again, why I don't take PPM very seriously.)

Worse than the uniform swing?

The JWS polls have consistently showed a swing in the seats selected that exceeds the aggregated national swing at the time.  This can be explained in two ways - firstly that the marginal seats surveyed are actually swinging by more than the national average (which is more plausible now in the context of a large swing than it was initially in the context of a small one) and secondly that the pollster has a slight house effect in favour of the Coalition. 

One way to look at the first aspect is to consider what we know about state differences.  In this case, I've taken the current Bludgertrack state readings and simply multiplied them by the state distribution of the seats polled.  On that basis, a swing of 6.9% would be expected in the polled seats on account of state breakdown alone (for a model predicting a national swing of 6.6%).  So the state breakdown of the polled seats may contribute a small part of the difference, and the difference between the average polling and the national picture is probably only about a point. 

(Something else I've noticed is that the Tasmanian seats are oversampled and the South Australian seats undersampled compared to the rest of the sample, but this makes little difference as these are the two states with the largest swings.)

Reserving judgement on whether there is a house effect at play and assuming that this really is the average swing in Labor's 47 notionally riskiest seats, do the report's projections stack up?  The important thing to consider here again is that swings are never uniform; rather, they are distributed by seat around the national swing.  Mark the Ballot gives a standard deviation of about 3.2 points for this distribution in past elections.

Looking at the results listed when Labor seats are distributed by margin:

* JWS projects the loss of all 10 seats in the 0-3% range off a swing of 5.9%.  The most likely seat loss number based on their results is actually nine.
* JWS projects the loss of all 15 seats in the 3-6% range off a swing of 7.5%.  The most likely seat loss number is actually 14.
* JWS projects the loss of 10 of 12 seats in the 6-9% range off a swing of 8.0%.  The most likely seat loss number is actually about seven.  (Labor has five seats within 0.5 points below the uniform swing result)
* JWS projects the loss of none of ten seats in the 9-12% range off a swing of 8.6%.  The most likely seat loss number is actually about three.

Thus, the claimed "marginality analysis" actually overstates the seat losses within the 0-12% group by about two, leaving a net likely loss in this group of 33, not 35.   However, on the basis of even a 6% swing nationwide, it would be expected that Labor would lose at least one seat on a margin above 12% (there are nine targets between 12 and 13% alone).

Other Vaguely Notable Stuff In This Poll

The poll report unhelpfully aggregates Greens with Ind/Other voters in the leadership aggregate questions.  However looking at these it's notable that even with about half of this category being Ind/Others voters, Tony Abbott has a shocking netsat rating (-57, and that's even with a neutral option) among such voters.  It's likely therefore that the Ind/Others voters in these seats are more left-leaning than in the nation overall and include substantial numbers of "Rudd refugees" (ex-Labor voters intending to only preference the party and vote 1 for someone else).  This is further confirmed by the poll's overall 2PP compared to the primary vote.  JWS use respondent-allocated preferences and these are about a point more friendly to Labor's position in this case than calculation by last election results nationwide (I have not checked it for the individual seats.)

It is also interesting that undecided (on primary intention) voters in this poll think the government has performed poorly but have a better opinion of Gillard (-17) than Abbott (-38).

The poll's question on the Labor leadership finds a new way to ask the question incorrectly.  The question is "Do you think Kevin Rudd should challenge Julia Gillard for the Labor leadership before the next election?"  The response is 33-54.  But just as questions on whether Gillard should step aside for Rudd ignored the possibility that some voters preferred Rudd but believed he should prove his support in a challenge, so this question design ignores the possibility that some voters prefer Rudd but believe he should become leader through Gillard resigning (and otherwise not challenge.)  Questions of this sort are producing results out of whack with polling of preferred Labor leader (even when broken down by party support) because of faulty question design.

Consistency with other findings

The idea that Labor might fare worse in the seat tally than the pendulum, or a state based model like BludgerTrack, predicts, is one that seems to be supported by all the polling that has looked at it, whether it is the JWS Research polling, the individual ReachTEL polling of selected seats (although those seats are far from a random selection) and the noises from Labor sources about their internal polling.

However, internal polling that is selectively released always needs to be treated with caution, and the published polling evidence on these issues is derived entirely from robopolls, the house effects of which (if any) are difficult to establish.  Because robopolls are cheaper, we are now seeing that virtually all seat polling is being conducted by that method.  It would be nice to see a few seat samples conducted by "conventional means".  (Morgan is sitting on a lot of aggregated data for particular electorates, but at $1950 per electorate, no thanks!)

I've also updated my Tasmanian coverage to reflect these results and consider the ReachTEL results further - see Federal Labor Getting Smashed All Over Tasmania (update at bottom).