The reality of Rudd’s last days

Hidden among the events of Thursday’s spill was Newspoll’s quarterly release of their polling demographic breakdowns between April and June in The Oz. In what can only be described as very convenient timing, this allows us to pool these Newspoll results with the Nieslen results over the same period (plus one  unpublished phone poll I have from a company that was testing their sampling frame), to not only give us geographical and demographic cross-tabs of the last 3 months of the Rudd government – but also the demographic baseline that the Julia Gillard Prime Ministership starts from.

With these numbers, we can run them through our updated election simulation engine to see how an election would have played out on the ground were it held between April and June – providing results that can only be described as very, very interesting. You may even be surprised.

This is quite a long post, so it may be worth going and making yourself a cuppa first.

On the demographics, over the period between April and June, the Newspoll sample was 5766, the combined Nielsen sample was 4200 and our unpublished poll was 1500 – giving us a pooled sample of 11466 respondents. From that pooled sample, we can drill down and look at how the primary votes were behaving by geography, gender and age cohort, as well as how the two party preferred was operating by geography – all with fairly significant sub-sample sizes, giving us a goodly amount of certainty on the results.

First up, the primary votes. As we can see, there were pretty massive swings underway against the ALP across the demographic and geographical spectrum, yet swings which weren’t transposing into consistent primary vote gains for the Coalition. The numbers sort of speak for themselves.

The Greens were the Party feasting on the spoils of ALP primary decline, particularly among the young and particularly in the capital cities – although not just directly through ex-ALP voters moving across into their column. As we saw with Nielsen preference dynamics and Essential voter firmness behaviour, there appeared to be significant compositional shifts at play, with the Greens also picking up ex-Coalition voters as the Coalition was itself picking up ex-ALP voters. If we look at what was happening at the state level of data, with the primary vote and the primary vote swing since the last election – we find that there was some quite complicated dynamics occurring between the states.

In Victoria, we had the interesting situation where the Greens vote had  increased by more since the last election than the ALP vote had fallen over the same period – leading to the rather odd outcome of the ALP two party preferred vote actually increasing (as we shall see in a minute). Yet the other states had the ALP primary vote falling by such large amounts that it could not be completely offset in two party preferred terms by the strong growth in the Greens primary. If we look at the ALP two party preferred results, we can see how the ALP primary decline and the Greens primary vote growth differentials came out in the wash.

This is where it gets particularly interesting. As we can see, there were large swings against the government occurring in NSW particularly, and also capital cities generally. This gets us onto what the two party preferred polling actually meant at the end of the Rudd era.

If we look at our Pollytrend vote measures for all the polling done over the last 12 months while Rudd was PM, we can that there was a recent turnaround in the decline of the ALP two party preferred.

Collectively, the polls had generally been trending upwards for the ALP over the last few cycles. Our Pollytrend measure is not a slave to any one poll or any one pollster – it looks through the polling noise and potential sampling error to find underlying trend movements in the true state of public opinion, by comparing how each pollster’s poll results change compared to their recent polling results.

Bob McMullan had argued in the party room and on Sky Agenda that the polls and their reality were being misrepresented in the media, and that the ALP was in a much stronger position than the reporting was suggesting *in terms of the way election dynamics historically play out.* A federal government has never lost an election if its two party preferred vote at the national level had a 5 in front of it – such is the power of incumbency when it comes to an election campaign. With the latest Newspoll being 52, McMullan argued, and quite accurately in that historical context, that the ALP was substantially in front of the Coalition and improving in the electoral race.

Others, one Newspaper in particular, have argued that because of the low ALP primary vote, only the ALP primary vote mattered and the two party preferred had become essentially meaningless.

That latter argument fails to grasp the complexity of what Bob McMullen was saying. Under compulsory preferential voting of the type we have in Australia, to win any seat, a party needs to get 50%+1 vote of the two party preferred vote share. That we know to be true.

With the low ALP primary but high Greens vote, the ALP was becoming heavily reliant on Greens preferences to win an election. As a consequence of this dynamic, one of two things *would* have occurred.

Either: as we witnessed with our voting firmness analysis, the Greens vote share was being inflated by soft voters and when an election was called, that Greens vote would decline and move back to the major parties, some of it moving back to the ALP and increasing the size of their primary vote (more likely than not, more of it moving back to the ALP than to the Coalition if history is any guide)

Or: If the Greens vote did not decline then those preferences would have to be distributed.

Regardless of what ultimately may have happened at the election later this year, far from the two party preferred vote being irrelevant, it was as important as it always has been – being the only measure of public opinion we have on the actual behaviour that ultimately decides elections. It just became more complicated as a consequence of an increased amount of uncertainty being attached to it – uncertainty that Greens preferences may not flow as strongly back to the ALP as they did at the 2007 election in the case of Newspoll results, and uncertainty that the stated preference of Greens voters with Nielsen polls may not be consistent with the revealed preferences of Greens voters come election day.

But just because something gets more complicated is simply no reason to ignore it. You just have to work your way through it with more sophisticated methods. This gets us on to how the ALP would have performed were an election held over the last three months and the election results matched the polling.

If we use an orthodox approach to this and plug just the latest Newspoll quarterly results into Antony’s spiffy electoral pendulum, it suggests that were an election held over the last three months, the ALP would have won around 81 seats to the Coalition’s 66 seats – give or take a couple either way as a result of the NSW Newspoll swing also containing the swing in the ACT.

A superficially good result for the ALP. However, what this result doesn’t take into account is how a large number of ALP seats sit right on the border of these swings. What we would actually see in an election is more variation on the pendulum – with some of those ALP seats on the margins of the swing falling to the Coalition and some seats the Coalition would be expected to take with those swings, not falling to the ALP.

If we move to the larger pooled sample of our demographic breakdowns, we can run those results through our election simulation engine that uses monte carlo analysis to accommodate for sampling error of the polls and the historical variation we see in the swings at the state level. This gives us a more robust, probabilistic view of how polling results transfer into likely seat numbers – as it accounts for seats with margins on the edges of the nominal polling swings, as well as the uncertainty of the polling swings themselves. Our simulation stabilised at 20,000 iterations (or 20,000 simulated elections) – giving us a nice smooth distribution of the results (click to expand).

As we can see, once we account for the expected variation of the swings and polling uncertainty, the ALP was indeed in very real danger of losing an election were one held over the last three months and the result was compatible with the complete polling aggregates.

The most likely result was the ALP winning 76 seats, for a government majority of a single seat. If we turn those results into a cumulative probability curve, it becomes more stark.

This chart tells us the implied probability of the ALP winning at least X number of seats. Simply pick a number from the bottom axis, trace it vertically until it intersects the red line, then trace it horizontally to the left axis to see the implied probability of the ALP winning at least that number of seats in an election held between April and June.

There was only a 65% implied probability that the ALP would have been able to form a majority in its own right.

If we rerun the simulation, but this time use Nielsen preference data as the two party preferred generator (where respondents get to allocate their own two party preferred preferences rather than have them allocated on the basis of the 2007 election flows), the most likely result would have been for the ALP to win only 69 seats, with there being only a 10% implied probability of the ALP winning government in its own right and a 30% implied probability of winning enough seats to even be able to attempt to form government with the support of all three independents that will be comfortably re-elected.

This makes the ALP “internal polling” that was being bandied around before the spill as quite important, as the purported numbers involved are entirely consistent with the polling we’ve been seeing over the last three months once we adjust for respondent based preference allocations.

It also makes them only a “moderately worse case scenario” under the assumption of preference flows repeating their 2007 election pattern – meaning that even if the internal party polling results were being slightly over-hyped to MPs and the media to encourage the dumping of Rudd, they contained not so much nuggets of truth about the most likely electoral situation, but large ingots of truth.

From a brutal hard-politics perspective, this adds considerable weight to the public opinion argument for replacing Rudd. However, it also needs to be kept in mind that the polling was improving, so the results would have been a little better than these simulations.

Not greatly better, but around half a dozen seats better in each instance.

This begs the question of whether the recent turn in the polling forced the hand of those in the Labor party wanting to remove Rudd – putting them in the position where if they did not strike last week, any further improvement in the polls, any further improvement in the polling trends, would effectively close any window of opportunity to replace Rudd with Gillard until well into the next term of government. Especially as any further improvement in the two party preferred would make many people start to question the authenticity and age of internal polling suggesting a wipeout.

Finally, and unfortunately in some respects (but entirely predictable in others), this gets us on to the tiring business of dealing with our old mate Dennis Shanahan and yet another one of his outbursts in The Australian. He states:

McMullan’s argument was that since Labor’s support crashed in April, the past four Newspoll surveys had shown Labor’s two-party preferred figures as 49, 50, 51 and 52 per cent and hence on the rise and on the way to victory. This was embraced by optimistic Labor supporters and academic or amateur pollsters who argued the primary vote and preferred prime minister were immaterial.

Read the whole article if you can be bothered, it’s the usual paint by numbers stuff.

Unlike Dennis, we analysed respondent allocated preference flows to attempt to understand and get on top of what was actually playing out with the two party preferred vote at the preference level. Unlike Dennis we actually analysed voter firmness to get a better grip on the dynamics of the large third party vote we have recently witnessed. Then again, unlike that sanctimonious windbag from The Australian, we actually do analysis here rather than word salad swimming in acidic dressing, where the substance of the meal rarely lives up to the hype of the menu.

Some of us that he’s presumably complaining about are professionals at what we do by any definition of the word – we are university qualified, we have professional experience, we have professional accreditation and operate as professionals in the fields of quantitative research –  while Dennis is, and always will be, just a journalist.

And without blowing one’s own vuvuzela, that can be seen in the content of this very article – unlike Dennis, rather than hide from uncertainty because it’s too difficult, taking the easy route of substituting proper analysis with noise, some of us have the professional skills to be able to embrace that uncertainty, adapt to the complexity it adds and come out the end with robust results from robust analysis using robust methodologies that can give rise to robust discussion. Results that actually contribute something positive to the sum total of human knowledge.

There are certainly amateurs and professionals when it comes to quantitative analysis – and then there’s those that lack the self-awareness to realise just which category it is that they actually fall into.

UPDATE:

In the next post, I’ve added another round of simulation results comparing the April-June period to Rudd’s final Newspoll.