Thursday, January 17, 2013
Morgan And The Myth Of Excessive Bouncing /The 1993 Analogy
Note added later: This article refers to a Morgan polling method that is no longer in use.
1. Face-to-face polls conducted by Roy Morgan Research are no more volatile than Newspoll.
2. In the case of both Morgan Face-to-Face polls and Newspoll, the level of bouncing from poll to poll is unsurprising, unsuspicious, and mostly explained by random variation.
3. Morgan Face-to-Face polls do however favour Labor, and this is especially true of their figures which allocate preferences according to results of the last election. Morgan phone polls do not seem to favour one side or the other very much.
4. The current polling position and history of the Gillard Government is similar to that of both Keating governments. Both became competitive about a year from the election, but continued to trail modestly for several months.
5. The first Keating government took the lead suddenly four and a half months from the election, while the second Keating government never managed to close down the Opposition's modest lead.
6. Recent articles by Peter Brent (Mumble) have repeatedly argued that a Gillard vs Abbott contest results in a certain or near-certain win for Abbott.
7. But an August 2012 article by the same writer sets criteria according to which the Keating polling situation was competitive throughout 1992, and then unfavourably compares Gillard's position then.
8. If the same criteria are applied, Labor under Gillard is competitive now.
Federal polling returned this week, and the picture is not too different to last year so far. Essential (which leans to the Coalition) returned another 54:46, Newspoll a 51:49, and the Labor-leaning Morgan Face-to-face series returned a 52:48 (as measured by respondents' claimed preferences) or 51:49 (as measured by the results of the last election). (All these in the Coalition's favour).
The Newspoll was amusing because a 54:46 from that company at the end of last year spawned an incredible amount of nonsense, some of which I disposed of in Is the Silly Season Real? and The Silly Season 2: End of Year Poll Myths . It's especially amusing to note that the single poll that, according to Peter van Onsolen's article, was supposed to make it difficult for the government to obtain momentum and doom it to a repeat of the fate of Howard in 2007, was followed by a 3-point change in Labor's direction. The Newspoll series since the early September shift now runs (figures are for Coalition) 50-54-50-51-51-54-51. I think that you can spot the odd ones out!
The Morgan was of most interest because the previous Morgan Face-to-Face had provided an unbelievably strong result for Labor (52.5-47.5 ahead, 53.5-46.5 using last-election preferences). It was no surprise at all to see a large correction. As usual Morgan waxed lyrical about the abandonment of the surplus promise, the single-parent Centrelink changes and the Macklin gaffe as "obvious" causes of the apparent loss in support, and as usual the main cause of the change was probably just that the previous poll in the series was a crock. After all if the government was taking a multiple-point hit over these matters, there is no way it would go up 3 points in Newspoll at the same time.
Those who have read some of my previous articles, especially the two above, will know that I've pointed out that a lot of the variation from federal poll to poll is meaningless bouncing, a point also strongly made in Possum's Trends, The Horserace and Random Numbers. Yet it is common to hear claims that a particular poll is all over the place and therefore unreliable. This criticism has been levelled a lot at Newspoll, and today Morgan Face-to-Face copped a lot of it for its 4.5-point swing. And Morgan has had the misfortune to release a couple of very serious pro-Labor outliers (even once they are adjusted for house effect) in the last three months.
But looking at the poll-to-poll changes, it doesn't seem that either pollster is unreasonably bouncy. In 50 Newspolls since the last election, the mean poll-to-poll change in the two-party preferred vote has been 2.02 points. Considering just the Morgan Face-to-face series, and ignoring their less frequent phone polls, the mean poll-to-poll change across 75 polls of various kinds has been 1.95 points (respondent-allocated preferences) or 1.73 points (preferences allocated by the last election result, the more reliable measure.) (For the Morgan phone series, the average changes are 2.69 points and 2.58 points respectively, but there's a very good reason for that - wider spacing.)
Some might think these levels of bouncing are still suspicious. But as a quick example of why they are not, I constructed my own random series. For each random poll, I used a random number generator (Excel's in this case) to poll 1000 imaginary respondents in a hypothetical Coalition-leaning nation, so that the chance of each respondent preferring the Coalition was 54%. Then I rounded the results of each poll to give a Newspoll-like whole number figure and repeated the exercise 100 times.
My series had phases of apparent "volatility", like 54-52-56-51-54. It had periods of "stability", like 54-54-54-53-54-54-54-53. It had a clear "rogue" (54-53-53-60) though the rogue was by chance followed by another outlier and a string of above-average Coalition results, 58-56-56-56-55. The average poll-to-poll change was 1.74 points. And, if the simulated poll gained points in one direction, on average the next poll shed 35.5% of the points that had been gained (not much less than the 40.6% I found for actual Newspoll results since 1986).
I could model all this much more precisely by factoring in the exact sample sizes of the two polls, and obtaining theoretical values for poll-to-poll changes for a random sample rather than doing an experiment, but it would be a tedious process given Morgan's erratic sample sizes, and it isn't necessary. Possum has already shown that Newspoll variation from his trend figure fits a normal random distribution, and the general pattern is that neither Newspoll nor Morgan are much more volatile from poll to poll than would be expected in a random sample with static properties. We know that the voting population isn't static - there are systematic movements over time and there can also be short-term impacts that wash out of the system - so the pattern of poll-to-poll changes in both Newspoll and Morgan face-to-face is as we should expect.
Furthermore, despite the constant rubbishing of Morgan face-to-face as being all over the shop, it simply isn't. It's at least no worse in this regard than Newspoll, with the slightly better results no great surprise given that Morgan polls more frequently and hence real changes in sentiment between polls will on average be smaller. It's also no great surprise that Morgan's last-election preferences are less volatile than their respondent-allocated ones, since allowing respondents to allocate their own preferences introduces an extra source of possible variation in the results. (There is strong evidence, covered by Pollbludger from time to time, that the way in which Morgan respondents have allocated preferences has not been constant over time.)
The only real problems with Morgan face-to-face polling as a data source are:
(i) It has a house effect in favour of Labor, as compared both to other polling and the company's phone polling.
(ii) The company confuses the presentation by including respondent-allocated preferences although these are known to be less reliable.
Some figures relating to the house-effect problem and average Morgan results (it's a rather confusing picture) are as follows. Morgan phone polling is generally considered to have no skew either way to speak of, and has an excellent record in final polls (effectively on a par with Newspoll). In this term, Morgan phone polling has produced average results of 53.9% 2PP for Coalition (respondent allocated) or 54.1% (by last election).
Morgan face polling has produced average results of 53.4 (respondent allocated) or 51.4 (last election). But if it sounds like we can just deduct one from the other, it doesn't necessarily work that way, because Morgan phone polls have not been evenly distributed through the term.
Another way to try to compare the two is to look at differences between adjacent phone and face-to-face polls. In this case, when a phone poll has followed a face-to-face poll, the Coalition vote on the phone poll has been 0.9 points higher (respondent allocated) or 3.1 points higher (last election). When a face-to-face poll has followed a phone poll, the Coalition vote on the phone poll has been 0.6 points lower (respondent-allocated) or 2.9 points lower (last election). The sample size for this exercise is only 19-20 changes in each case, so these averages are not all that reliable. But they are consistent with the picture of the raw averages: that face-to-face with respondent-allocated preferences slightly favours Labor, while face-to-face with preferences allocated by the last election favours Labor by over 2 points.
Another useful comparison is Newspoll, currently thought to favour Labor by about half a point. The Newspoll mean for the term has been 53.8 to Coalition. As Newspoll and Morgan Face-to-Face are both very commonly conducted, we can compare averages directly, although looking at changes over the term of the parliament is worth doing too. On the latter, I'll defer to the Mark the Ballot aggregator, which currently shows Morgan Face-to-Face with last-election-allocated preferences as leaning Labor by 2.3 points.
On this basis, the current Morgan is really a 53:47, and with an Essential that's really a 53:47 and a Morgan that's probably really 51.5:48.5, we're still where we spent quite a lot of time near the end of last year, with the Coalition somewhere in the 52s. MtB has it at 52.3:47.7.
Two paths from the past
The Gillard government, after a long period of uncompetitive polling, began polling competitively (still behind, but well inside of 54:46) in mid-September, 11-12 months from the generally expected election date. It has now stayed in this new situation, with minor ups and downs, for four months.
Two past polling trajectories show us that governments have both won and lost from this sort of position. The 1990-3 Labor Government trailed for most of its first two years, and was polling uncompetitively until a year prior to the eventual election. It then was in a "competitively trailing" position (rolling averages between 51.2 and 53.6) for the next eight months, and then a sudden snap of 8 points in a single fortnight amid rising concern about Fightback (and, probably, Jeff Kennett) put Labor back in the lead (which they later lost and then regained during the campaign.) As I've demonstrated with the example of Menzies' win in 1954, Keating's win may have been pretty startling at the time, but it is by no means unique.
The 1993-6 Labor Government had the advantage of Opposition turmoil through the ends of the Hewson and Downer leaderships, but was uncompetitive in March-April 1995. As the Howard honeymoon washed out of the system, the government again sat in the 51s and 52s (rolling average of Mumble's derived 2PPs) but could never close down the last point or two to reach an even position. (And if you think long runs with no change to speak of like the one I gave from my random "poll" don't happen, check out this one, from Sep-Dec 1995: 52-53-52-52-53-52-52-52!)
I think these two are much more useful comparisons than 2001 (an election year unusually strongly affected by unexpected events), 1998 and 2004 (governments trailed in these cycles, but not uncompetitively for any length of time) and 2007 (the government was never competitive once Rudd was installed as Labor leader). The key differences between now and both Keating governments include that this government is younger, but also that it is encumbered with its minority status and perceived links to the Greens. The economy is much better (at the moment) but on the other hand the "broken promise" baggage at the moment looks worse than in 1993.
Again, as throughout Keating's time, we have a Labor government that is not well regarded and that the electorate would rather like to be rid of, but I totally agree with Mumble that the election in these cases (and all cases) comes down to these basic questions:
"(1) am I ready to throw the government out? And if the answer to that is ‘yes’, then (2) is the alternative palatable, acceptable? Or is it too scary?"
Mumble's response to question 2 is:
"Question two is difficult. It elicits sullen responses. Must we? But the strength of that first ‘yes’ outweighs misgivings about an Abbott prime ministership."
Mumble gives 2004 as an instance of voters saying "yes" to 1 and "no" to 2. I'm not sure "yes" to 1 for 2004 is all that clear. Aside from the Latham honeymoon period, Labor did not have big leads in that year. Moreover, Howard had good personal ratings throughout it. And yes, Latham's danger factor did not become apparent until people balked in the shadow of the ballot-box, but the Coalition may well have won even without it.
The clearer case of "yes" to 1 and "no" to 2 is obviously 1993. And plenty of pundits, including the psephs, were assuming in around July 1992 that the strength of the "yes" to throwing out Keating (at that time, polling personal netsats such as -44) and his economically moribund government outweighed misgivings about imposing a GST. After all, Fightback had been Liberal policy for many months, and they were still leading comfortably, even if it had closed in a bit, right? Anyone, at that point, who said that a few months later the Coalition would suddenly drop from 53s to the mid-40s, and be forced into a massive, and arguably fatally confusing, revision of Fightback would have been dismissed as not merely a true believer but a mad one. The voters will suddenly ignore everything wrong with the terrible, poorly-communicating, left-skewed incumbent government and start worrying about the Opposition's policy shortcomings? Surely not!
It is true that in 1993 the main problem generating the "no" was policy, and that the blunders made by the Coalition in that case are unlikely to be repeated anytime soon. Likewise, in 2004 the leadership problems that generated the "no" (and the policy problems arising from the leadership problems) became most apparent (at least outside Tasmania) at a very late stage. If Abbott's leadership is the Coalition's only serious problem in 2013, it may be already as factored-in to the Coalition's standing as it is going to get, in which case the Coalition may just sit on its 52-ish average all the way to the ballot-box, much as Howard did in 1996. It remains to be seen if it is really all so easy for them when they have to put policy meat on the table, and which side is really marching calmly towards the cliff.
Embracing the 1992 Model, Then Running Away From It
I would also suggest that those following the current Mumble line check out this article also by him from August 2012, just before Julia Gillard abruptly became less unpopular and her government more competitive on 2PP. In seeking to explore the causes of the 1993 Keating win, Mumble refers to:
"the election on 3 October of a Coalition government under Jeff Kennett and the new premier’s decision to ignore urgings to wait until after the federal election before administering harsh medicine.
Megalogenis’s theory goes that Australians got a glimpse of what a federal Coalition government might look like and it unsettled them."
The parallel with Queensland today is discussed. (I would also throw in that the Victorian state Liberal government is not travelling well.)
"If Newspoll had published 2pps there would have been a greater appreciation throughout 1992 that Keating was at least competitive." [my emphasis]
If Keating was "at least competitive" "throughout 1992" (even I wouldn't go that far!) then why is Gillard not "at least competitive" now?
Indeed, suppose you're an incumbent PM and you're offered the following run of 2PPs for the period from twelve to eight months out from an election:
(i) 50-46-50-49-49-46-month off-49 (mean 48.4)
(ii) 47-51-49-47-46-45-47-49-48 (mean 47.7)
Which one do you take? Assuming a September 2013 election, choice (i) is Gillard now, choice (ii) is Keating 1992.
"See how close the polls were over that whole period by today’s standards. With hindsight the 1993 election result shouldn’t have been so surprising."
But if Gillard beats Abbott from a similarly "close" polling position, that will be not just surprising, but impossible?
"George suggests the Kennett-Hewson dynamic might be replicated this year, with Queensland premier Campbell Newman in Kennett’s role. That part doesn’t quite work for me. For one thing Gillard is in a much, much deeper hole than Keating was"
Again, in 2PP terms, and assuming a Sep 2013 election, that wasn't true comparing like for like (in terms of expected time to the election) at the time Mumble wrote his article, and it certainly isn't true now.
"Most of all Keating was a superior politician to any of today’s bunch.
By March 1993 he was a fully incumbent prime minister, with authority."
And a disapproval rating of 60%.
A lot of what is written about the mystery of 1993 - the election which almost no models predict - is annotation-by-result. Keating must have been brilliant, because he won the unwinnable election. But that had as much to do with the unnecessary strategic bungling of his opponents as with his exploitation of it, and had they played their cards correctly that "superior politician" would be just an unpopular, contentious bookend to the glorious Hawke era.
Few were calling him a great politician at the time.
Also see: The 2013 election is looming: Is it really unloseable? by Stephen Koukoulas for the economically literate version of the argument that the election is at this stage competitive whatever some people may say otherwise. (A minor correction: Hewson was not consistently popular, and generally rated indifferently to badly from June 1992 onwards. He was nonetheless much less unpopular than Abbott)