Saturday, June 1, 2013

How Common Is A Five-Point Swing In Under Four Months?

Advance Summary

1. The Gillard government now probably needs a swing back in public opinion of around 5% in three and a half months to win the federal election. [Update:  After Monday's polls, it's even more than that.]

2. A recent article by Simon Jackman draws attention to swings of this magnitude, from one election to the next, being rare events that mainly (not exclusively!) accompany changes of power.

3. It is not correct to infer that swings of this size in a much shorter period are even less probable, because in fact 5-point swings within 16 weeks are commoner than 5-point swings between elections.

4. However, 5-point swings in shorter time periods typically result from temporary factors - leadership and event bounces, honeymoon effects, policy mistakes that are corrected, bad patches etc.  

5. Fast swings that have occurred in the leadup to elections have usually washed out of the system partly or entirely by election day.

6. The past history of quick 5-point swings in polling provides no sound basis for belief that even an extreme event such as the S11 attacks or leadership change would save the Government now.  (This does not mean it is impossible, just that there is no precedent.)



=======================================================================

Not much is happening in federal polling at the moment.  Since the polling horrors of the late March leadership-spill-that-wasn't, there has been a gradual shift back to the Gillard Government, but a recovery of a point or so is not much to write home about when it takes two months to happen and you only have three and a half months to gain another five points.  I say five rather than the 4.5 implied by my current aggregator because I believe that if the 2PP is 50:50 then the Government will probably lose (as it would have done last time with a 2PP of 50.2 had not two independents elected in seats that preferred the Coalition to Labor decided to support a Labor government).

(Update 3/6: after today's Newspoll shocker my aggregator has gone to 55.4, so perhaps we should be asking how common is a six-point swing in three and a half months!)

The election looks over as a contest unless something very unusual happens, and even if most of the "something very unusuals" that might be plausibly nominated occur.  We are seeing some pretty bold statements from not just psephs and pundits, but also statistical modellers, to the effect that it is more or less done and dusted, from Simon Jackman's article on a Guardian blog called The Swing to this Bayesian model by Julian King (Pottinger).

Jackman declares:

 "I expect some narrowing to be more likely than the Coalition improving on 55-45 TPP. But bottom line: Labor just can’t get to an election winning position from here."

The Pottinger model - the first detailed attempt I've seen to model the final result based on assumptions about changes from present polling and not just the state of present polling - gives the Coalition a 93.6% chance of winning outright.  It does this even though it projects the ALP's final vote to increase to 47.2% 2PP, on the basis of Bayesian arguments about using the range of past results to condition expectations of where the vote is likely to end up.  I am actually sceptical of the Bayesian assumptions in this case, since the experience of both NSW 2011 and Queensland 2012 shows that a party polling near or outside the boundaries of the range of prior results cannot be trusted to recover.  Yes, state and federal are different, but it could even be that the Pottinger model gives a generous assessment of Labor's chances on this basis. 

But I was intrigued by this statement from Simon Jackman:

"So, assume we’re currently at 55-45 TPP in favour of the Coalition. The Coalition would have to suffer a five-point TPP swing at this point (or worse) to be at risk of losing. Swings of that magnitude between elections are rare in Australian political history — let alone over 16 weeks — and are reserved for “change” elections."

What intrigued me here is the question of whether a five-point swing between elections is actually a more common event than a five-point swing over 16 weeks.  I suspected that it would be the other way around.


How common is a five-point swing between elections? 


Two-party preferred results for elections are available or have been estimated back to at least 1937 (officially for elections from 1949) though in the early half of the 20th century the frequency of seats being decided unopposed creates a lot of room for error in the calculations.  A version of the two-party system has existed since the 1910 election won by Labor, although it was severely disrupted by Billy Hughes' defection to the Nationalists, and although the "Labor" side has been severely fragmented at points during splits.  All up there have been 40 elections fought under essentially two-party lines, so there are 39 swings that can be calculated or (however roughly) estimated.  Of these, eight should be considered as 2PP swings exceeding 5% - 1917 (a mongrel example as noted), 1931, 1943, 1949, 1969, 1975, 1996, 2007.  Seven of these were essentially "change" elections in that they returned a different government to the one elected at the previous election, though in the cases of 1917 and 1943 the change of government had occurred mid-term.  One of these (1969) did not see a change of government of any sort, though the swing in that case was amplified by Labor's terrible performance under Calwell in the 1966 election.


How common is a five-point swing over 16 weeks or less?


I have used my Newspoll rolling average data set, such as it is, to investigate this question.  As I've previously noted, it's a rather slow-moving animal since it includes the three previous Newspolls as well as the current one, although the current one is weighted double.  Nonetheless, until such time as I get around to obtaining a full back-set of all polls in order to do aggregations into the distant past, it's best to be cautious when using a run from a single polling company and hence relatively few very fresh data points in each case.  What I have looked for, then, is cases since 1986 in which my crusty old Newspoll rolling average (based in part on Peter Brent's conversion of old 2PPs) indicates a five-point polling swing that covers a period of 16 weeks or less.

I've also included a "lag exclusion" rule.  Because the rolling average of a reading from months earlier includes data from outside the 16-week period, I've only considered a five-point swing as demonstrated if the current rolling average differs by five or more points from both a rolling average in the previous 16 weeks, and a specific poll reading in the previous 16 weeks.  This increases the chance that these swings are actually real over that time and not just baggage from before it.  I've omitted cases where the 16-week span includes an election, and I would have omitted any cases that were caused by single obviously rogue poll readings, but it happened that there weren't any.  There might be a few false positives in the list below, but if anything it's more likely that a few borderline positives have been missed out. 

Anyway I found twelve apparent cases of a five-point (or more) swing over 16 weeks or less.  Many of these were multi-week cases in which the 2PP more or less consistently differed from one or more rolled-average results in the prior 16 weeks by five points or more, over a number of consecutive weeks.  The twelve cases are in the table below (click for larger version).


Table legend: Poll Date = poll month at the end of the largest swing in each sequence Start = Govt 2PP at start of swing, End = Govt 2PP at end of swing, Weeks = weeks between high and low points creating the gap, Dur = approx number of weeks for which gap exceeded five points in this sequence.

Note that while the number of 16s in the Weeks column may create an impression that some of these just scraped in, that's not necessarily true. In most of these cases there were slightly smaller (but still > 5-point) swings over much less than 16 weeks in the same sequence.  The rolling average actually showed cases of 5-point swings over six and even four weeks, but the rolling method stops it from picking up any faster ones unless the swing is ridiculously high.  Sometimes, as in Oct-Nov 1992, August 1993, and September 2001, it seems that public opinion just "snaps", by several points, in a matter of days.

The period from 1986 to present includes only the two elections (1996 and 2007) that saw five-point plus swings from the previous election.  But far from it being rarer to see a five-point swing over 16 weeks than it is to see a five-point swing between two elections, it is something that happens much more often.  This is partly because opinion polls are common but elections are rare.  But that is not the only reason.

All up, around 7.5% of all Newspolls since 1986, excluding those taken within four months after an election, have been taken at a time at which a five-point swing in public opinion had apparently happened in the previous sixteen weeks.  It's simply not a very rare event, even if a minority of these readings would have been cases where random bouncing made the swing look larger than it was.  Most of the cases above look real enough in that there are obvious explanations in terms of what was going on at the time.  (I've left late 97 blank as an exercise for the reader since I'm not too sure what might have been driving that one.)

So Does That Mean It Might Happen Now?

I can sense Labor supporter hopes rising through the previous paragraphs but there is really not a lot to get excited about on account of these results.  Firstly, of the twelve swings mentioned, eight of them favoured the Opposition at the time, and only four the Government.  Over half involved either leadership instability or a post-election honeymoon effect washing out of the system.  When the swings to Governments are considered, two involved leadership issues, one involved major events (and the 2001 bounce was much more S11 than Tampa) and only one involved policy.  And the one that involved policy involved Fightback, which was the seminal policy blunder lesson on how to lose elections from a winning position, a lesson the Coalition is still very careful to learn from. 

The swings in question generally didn't happen in the shadows of the winning post.  Only the two Fightback-related swings (which came close to cancelling each other out, only for the Opposition to lose to a much smaller late swing) and the Tampa/S11 bounce happened very close to an election.  The S11 bounce was mostly gone by polling day.  There is a view that if something remarkable like S11 happened again it might create a similar swing and win the election for Gillard, but I think that view is wrong.  The net impact of S11, once the bounce had come and gone, was not that great, beyond depriving Labor of an opportunity to really compete for the victory.  I've mentioned before Peter Brent's insight that bounces (whether caused by events or leadership changes) are not real.  They don't last til the election, and most likely the closer the election the shorter their longevity.  You can't orchestrate a bounce event a month from the election and win the election on the back of it.  Not even a Dothraki wedding for Julia Gillard would do it.

Who Stole My Volatility?

It's also a commonly clutched straw that polling in recent years has been extremely volatile, given the roller-coaster ride from late 2006 that started fairly even, then went to huge Labor leads, then to equality, then to huge Coalition leads, the Gillard Recovery, the Spill Cliff and so on.  But despite some extreme values in this time frame, it seems from the above that this volatility exists only over a fairly long time scale and that the last six years have not been a time in which public opinion has changed very rapidly.  Rather what we are seeing is sustained shifts where there are prolonged moves in one direction, but not especially steep transformations.  We've seen sustained 2PP changes of just over a point per month at times in this period, and one case through early-mid 2011 of a 5.2 point swing over eighteen weeks (close but no cigar!) but the real days of polling volatility in the last 26 years were those days of old when knights were bold, the Keating era. 

And going back before then, the Keating days had nothing on 1975, in which Morgan data shows Whitlam's government leading on primaries 49-44 in March, behind 33-60 in July, ahead 47-43 three days before getting the sack in November, and behind 41-55 four weeks later, before finally losing 43-53.  On a two-party basis that's two swings of about 15% that each took about four months, followed by another one of several points in four weeks.  Admittedly, these are based on individual poll results, and a richer set of polling data that allowed for more meaningful rolling averages would trim a few points from some or all of these.

Live By The Sword ...

I've basically rejected the idea that a swing of five points in sixteen weeks is too steep to climb (by showing that such swings are actually commoner in this time frame than between elections) but I strongly believe that this particular mountain is too steep.  The reason for this is that history giveth and history taketh away: exactly the same factor that makes precedents for a five-point climb fairly common also makes those precedents mostly irrelevant.

The twelve cases I've found mostly involve polling being artificially inflated for one side or other, either at the start or at the end. And this applies to all the examples in which governments have gained such a swing so quickly.  The first Keating swing was from an artificially depressed position because Bob Hawke had passed his use-by date and had to be removed.  The other three such swings obtained by governments were the products of an Opposition policy blunder that was mitigated, a dud Opposition leader who was replaced, and a major world event that soon largely washed out of the polling.  In the two past cases in which Governments were bounced well above 50% by a 5+ point quick swing, and went on to win the next election not long after, those Governments did so with 2PPs 3.2 and 4.2 points below the inflated end points of those swings.  Even if the Government finds something (and other than a new leader, it is hard to imagine what) and goes up seven points then down three from here, it would still very probably lose.

So the issue isn't that a 5-point recovery in sixteen weeks is too steep a mountain to climb, since it's a mountain that has been scaled and then some in polling in the past, not too uncommonly.  The real problem is that where federal governments are concerned, elections aren't steep mountaintop finishes. 

(And the same, I think, applies at state level too.  Massive late swings at state level tend to be at the expense of governments and to cause them to unexpectedly lose.)

Update (3 June)  Tonight's Newspoll was a shocker for the Government with a reading of 58-42.  However Morgan (55.5-44.5 by last-election preferences) and Essential (55-45) didn't come to the party so I reckon this Newspoll, like the last 58:42 in March, is overcooked.  All the same, the Government is again moving in the wrong direction in today's polls considered together.

So much so that we can now say the Government's position is no better than Howard's Coalition at the same stage in 2007.  Indeed on my Newspoll rolling averages it is a whole point worse, but that is likely to be skewed by the current reading.

No-one should be surprised the Government has polled badly this week after the electoral funding debacle - which is not to say that is certainly the cause, but there was certainly enough in it to stop any recovery attempt in its tracks. 

No comments:

Post a Comment

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.