Thursday, February 7, 2013

Rogues, Recoveries and Irrelevant Portents

Admin note for Tas readers especially: this month I am running Not-A-Poll: Best Tasmanian Premier of the last 30 years.  Please vote (on right) if you have an opinion!  Readers may be surprised to see the Premier of one of Tasmania's least popular governments ever (ie the current one) challenging for the lead, as well as the big lead for the Labor premiers over the Liberal ones, but there are actually quite a few reasons why this might occur.  When the not-a-poll is finished I'll post some comments about the exercise and what it means. Of course, the results are not representative of the general population - and that's even assuming nobody stacks the thing!

-------------------------------------------------------------------------------------------------
Advance Summary

1. Following a very turbulent week in federal politics, this week's polls on average show a small move to the Coalition.

2. The Newspoll showing a result of 56:44 to the Coalition is unrepresentative and probably exceeds the real figure by at least 2 points.

3. The Newspoll reading is probably not, however, a "rogue poll".

4. Over-calling of supposed "rogue polls" is a very common problem in the online poll-watching community.  This article provides many cautionary notes about use of the term "rogue" to describe a poll result.

5. The Government's polling position is now worse than that of the Keating Government in 1992-3 was at the equivalent or any later stage.

6. There are possibly still as many as six cases of governments recovering from worse polling positions than Labor's current position (with the time to go until the election factored in), however in three of those cases the data are very limited and at least one of the others is probably an invalid comparison.

7. Claimed evidence that either the declaration of the election date well in advance, or the chosen election date itself, are bad portents for Labor is not valid.

8. This article concludes with some discussion of contradictory voter attitudes to the economy, which may be posing a major problem for Labor at the moment.  



--------------------------------------------------------------------------------------------------

A big batch of federal polls came in this week, handily following a very turbulent week for the Gillard government.  Here's a table of the four polls with Coalition 2PPs adjusted for house effect:


A few comments about the above are needed.  Firstly, Essential poll results are two-week rolling averages, but while the 2PP did not change this week (Essential is a slow-moving animal: see Essential: Not Bouncy Enough?) the ALP primary vote did; it was down by one point.  The Essential 2PP was probably therefore very close to rolling over to 55:45 and the probability that this week's sample was a 55 is high. Against that, the latest Mark the Ballot benchmarking shows Essential's lean to the Coalition at somewhat over a point.  So there is some case for treating the adjusted 53 for Essential as higher, but if so that might be more like 53.5 than 54.

Secondly I have chosen not to adjust the Newspoll for its recent tendency to favour Labor by about half a point, on the arguably circular grounds that the current Newspoll constitutes evidence against it still doing so.  Thirdly, for Morgan Face to Face I have used the last-election preferences, because they are easier to benchmark consistently.

Whatever is done with all these calls on given polls, the overall effect is the same.  All these polls are on the upside, for the Coalition, of the previous rolling average of about 52.5.  However most of them are only slightly worse, with Newspoll the big exception.  Whatever is done, Newspoll is stranded high and dry in comparison to the others.  Its most generous treatment by the various aggregators came from Bludgertrack, which returned an overall figure of 54.7, but that is largely because Bludgertrack is weighted in Newspoll's favour because of its strong performance at past elections.  (Pollytrend shows about 54, and Mark the Ballot 53.1).   As previous data shouldn't be completely thrown away even when you have four new polls, I'm inclined to see the current real 2PP situation as about 53.5 to Coalition.  Whether the gains last remains to be seen.

The Newspoll has received some flak because of a stated exclusion of natural disaster areas from polling.   I wouldn't have thought the disaster areas would have been especially pro-Labor overall and hence I doubt excluding them would have given the Coalition any advantage, let alone a significant one.

The over-the-top result from Newspoll resulted in many people calling it a "rogue".  Which raises the obvious question:

What is a rogue poll?

A rogue poll is a poll that is outside the margin of error for that polling sample. The margin of error (MOE) of each sample is a function of (i) the actual support level for the given party or question being polled (ii) the size of the sample.  If the poll has no house effects, then 95% of poll readings should fall within the margin of error of the actual support level.  If the poll has house effects, then after those house effects are accounted for, the same should be true.

Newspoll gives its maximum margin of error as 3 points.  The margin is highest for readings that are close to 50%, and lower for readings that are low.  For instance, a Green vote of 9% off a sample size of 1150 has a margin of error of 1.65%, although this is a slight underestimate because of rounding.

Greater knowledge of rogue polls has led to people being more familiar with the possibility that a given single poll is simply complete nonsense.  But unfortunately, it has also led to people over-applying the concept.  Here are a few things to bear in mind.

1. A poll isn't a rogue just because it changed more from the previous poll by the same pollster than that poll's MOE.  For instance if a party's actual support level is 52, and one poll is 50 and the next is 54 with an MOE of three points, the change is four points but both polls are within the MOE of 52.

2. A poll isn't a rogue just because it is out of whack with other polling.  Sometimes it is clear that one poll is not following the trend of others and is a bit inaccurate, but that doesn't by itself mean that poll is outside the margin of error.  It needs to be out by a sufficient amount.

3. A rogue is still relevant data.  Suppose I want to test whether a coin is loaded and I do four runs of an experiment in which it is tossed 100 times.  The coin comes down heads 48, 44, 45 and 62 times in the four experiments.  The 62 is a rogue sample, but that doesn't mean I should just throw it out and take the average of the others.  Indeed, in this case, it may well be that the coin is fair and that the rogue 62 just happens to cancel out a below-average run in the previous three cases.

4.  When dealing with a run of results by a single pollster, a rogue can only be detected after a few more polls.  The reason for this is that a poll that appears rogue compared to past polls may be a harbinger of significant change.  For instance in late 1992, the following sequence of Coalition 2PPs appeared: 54-52-50-53-53-54-46.  The 46 was not rogue at all; it signified a backlash against the Coalition's policies and was followed by 48-45-44-46.

5. Even when a poll looks rogue by comparison to other polls in the same week, it is still necessary to see whether or not its result is confirmed in following weeks.  It is possible, but rather unlikely, that one pollster correctly picks the new trend but subtle variations in research methods mean that the others do not. 

6. Rogues are not always detectable.  A rogue may not look rogue if, by chance, a number of surrounding results lean in the same direction.

7. A fortnightly pollster will only produce an average of just over one rogue a year.  The chance of a pollster that issues 24 polls per year throwing three rogues in a year is only 6%, and for four it drops to about 1%.  If you're calling a given poll "rogue" many times a year, you're getting it wrong.

8. An apparent rogue is not a rogue if there are good reasons for it.  Sometimes issues or leadership events do affect polls strongly and briefly.  That said, it's common for large changes to generate erroneous commentary declaring them to have been caused by events that interest hardly anyone but avid politics-watchers. 

Clearcut rogues are pretty rare.  A famous Newspoll rogue occurred in August 1995 when Labor polled figures equivalent to 51:49 in the middle of a string of results that were all 52s-54s for the Coalition.  The most recent really clearcut-looking Newspoll rogue was October-November 2009 (Labor polled only 52 amid a sea of mid-high 50s) though there is also a roguish look about the 56:44 to Labor in late March 2010 (surrounded by much closer results), and the 51:49 to Labor in late March 2011 (surrounded by Coalition-leaning results before and clearly pro-Coalition results after.)  The 57:43 to Coalition in November 2011 and the 59:41 to Coalition in April 2012 were both suspiciously high, but the latter was the only 2012 poll that raised a blip on a basic "rogue detector test" I ran (average of the four surrounding polls).  Where the current poll one scrubs up through the rear-view mirror remains to be seen, but based on the results of other pollsters, in my view it is not a rogue, but not far short of being one.  

A similar problem to over-frequent calls of "rogue" is a tendency by many poll-watchers to assume there is never anything happening (or at least nothing bad for their side).  Just because we know that a lot of poll-to-poll movement is random noise, does not mean we should write off all poll movement as necessarily just noise, or seek to show how a bunch of polls from different weeks are all within MOE of some "trend" and hence the trend must be stable.  Indeed, if you find yourself making such an argument on the basis of multiple polls with differences close to the supposed MOE, it's time to reconsider it.  Even polls close to rogue status aren't that common.

Even A Point Is A Bad Hit

Suppose we accept that the 56:44 from Newspoll, uncorroborated by even the Coalition-leaning pollsters, just can't be taken all that seriously and that the real state of play after last week is in the mid-high 53s.  Well, that's still bad news for Labor if we stack up their current position against historical recoveries by oppositions that were trailing.  Anyone who thinks Labor is going to be able to poll figures like this all the way to election week and then suddenly win with a massive last-minute swing is very probably kidding themselves; it has never happened on that sort of scale before, and Labor does not have the advantages that past governments with majorities had when it comes to being able to win the election while losing the 2PP.  As a general rule, the larger the opposition lead for a given time from the election, the more likely it is that the government will be defeated.  As time goes by, the size of lead that can be considered to strongly predict a Coalition win will shrink.

Even in the present situation, at least three governments have still won from positions that were as bad as this or worse closer to the election.  The 1998 Howard government had a rolling average Newspoll 2PP no better than the current three months from the election and yet won, and the 2004 Howard government was likewise just less than six months out.  The 2001 Howard government was in an even worse pickle, down about 55:45 with six months to go.  But it's arguable whether 1998 counts (given that Howard was able to get away with losing the 2PP).  What's perhaps notable is that this week's results drop Labor below even the Keating 1993 recovery path.  Keating was in about this position nine months out rather than seven; in the last seven months prior to the 1993 election the Coalition's rolling average was never higher than the mid-52s.

Aside from the Howard cases mentioned above, governments that may have recovered from worse polling closer to the election to win include Fraser's in 1980 and Whitlam's in the 1974 early election. However, in each case there is just a single old Morgan poll result suggesting this, and not enough surrounding data to say how reliable those results were or weren't. It's also not clear exactly when Menzies in the leadup to 1954 became competitive, as there was no polling between nine and six months out (in which time the gap closed from 55:45ish to 52:48ish) in that case.

How to play the precedent game

The Australian in particular has been prone to run articles that pick through the entrails of poll data and come up with a conclusion that such-and-such precedent means the Coalition wins.   These may look prescient and objections to them (including mine) may look pedantic if the Coalition does win comfortably, but in fact should this happen, they will be right for the wrong reasons.  In The Silly Season 2: End-of-year poll myths I debunked an example of this sort of pointless psephastrology by Peter van Onsolen. 

Now here's a similar effort by Christian Kerr.  The main argument this time is that of five previous governments that have called elections early, four (1954, 1958, 1961, 1984) have ended with a swing against the sitting government, while only one came up trumps (1966).  A second argument is that Labor has had four previous September elections (1914, 1934, 1940, 1946), and while he only spells it out in the case of 1946, I assume the point is that the incumbent government went backwards in all of them (and in 1914 even lost).

But looking at the eight elections that form his case closely, in most cases the swing against the government can be explained without referring to the length of the campaign or the month of the election:

* In all of 1958, 1934 and 1946, the incumbent government was coming off a big win at the previous election.  A swing against a government that wins a big win is no great surprise.  (This is also true of 1984, but in that case the government's polling was so good in the leadup to the election that perhaps it wasn't avoidable.)

* In 1914 the cause of the defeat was very likely not the election being in September but rather the election being amid the intensification of World War I.

* In 1954, the victory was in fact a great result for the Menzies-led Liberals given the appalling polling they had endured through much of their term. 

There is also a data issue with the 1954 election.  Kerr writes "The Liberal vote fell by 2.31 per cent. The Country Party's dropped too, by 1.2 per cent."  However the primary votes are a misleading indicator because the Coalition won six seats unopposed in 1954 and these are not included in its primary vote (compared to one in 1951).  With those taken into account the swing against the Coalition in 1954 was only 1.4 points.  And in the case of 1958, while the Coalition vote did go backwards, the main beneficiary was the DLP, and the 2PP swing was negligible (the Coalition in fact gained seats.)

That the majority of the remaining elections considered by Kerr saw swings against governments is no great surprise, because that is what usually happens (perhaps a more relevant portent!)  Swings to incumbent governments have happened only nine times, two of them related to the mid-term formation of the government in question.

To show how easy it is to play the "portent" game however you like, here's how to write portent punditbabble that points to a Labor victory!

--------------------------------------------------------------------------------------------------------------
"In the 13-15 February 1998 Newspoll, the Coalition's 2PP under Howard went backwards by 3 points.  In the 9-11 Feb 2001 Newspoll, the Coalition's 2PP under Howard went backwards by 5 points.  In the 6-8 Feb 2004 Newspoll, the Coalition's 2PP under Howard went backwards by 3 points.  In the 2-4 Feb 2007 Newspoll, the Coalition's 2PP under Howard only went backwards by 1 point.  The Howard government won the first three elections but lost the fourth.  

Clearly a bad poll early in February at around the time parliament resumes helps sharpen a government's awareness of its own electoral failings and warn it against complacency before the battle ahead!  On this basis, the Gillard government, having gone backwards by 5 points in the equivalent poll, is well placed to win the election!"
-------------------------------------------------------------------------------------------------------------------

Now that would be quite a silly argument; after all, the current Newspoll reading is actually worse than Howard's even after all of those hits (including the one the year he lost, in which it was obvious he was struggling prior to that.) It's probably largely coincidence that Howard had bad polls around this time all those times he came back and won. 

For a "portent" to mean anything, it is not enough to know it has worked in the past.  There has to be a valid logic behind it. Is there a reason why September elections should be bad (and no, I don't buy football season or Yom Kippur as convincing arguments)? If not, then a pattern of them being so for the government, even if real, would most likely be just coincidence.

Attribute Hell

In my view the worst aspect of the Newspoll was not the easily-refuted 56:44 but rather the attribute polling that came with it.  There is a useful historical review of the question about handling of the economy by Mumble here.  As pointed out in that review, it's just unlucky for Labor that this particular attribute poll has been taken as part of what just happens to be a bad sample.  That would influence the results of the attribute polling - but not by very much.

The general pattern of the polling is that Labor holds slender leads in all the areas regarded as its strengths (though in the case of climate change, this is deceptive, with the anti-Coalition preferences being split between Labor and the Greens) while in traditional Coalition strength areas, the Coalition's strength is massive.  And the 50-28 figure on economic management is one that has aroused a lot of attention.  It's even a shade worse than the 47-28 in October 2011, when Labor's polling was consistently terrible. 

It's strange because, objectively, economic management is probably among the least of the government's worries, while the Opposition Leader is often considered to have not much idea in the area (including by some on his own side).  Economics isn't always Labor's strong suit, but Australia is seen as having come out of the GFC comparatively well, and the extent to which credit should go to the previous government must by now be fading to a fairly large degree.  An article by Gordon Graham shows how recent Essential polling on the budget surplus question reveals remarkable contradictions in voter attitudes, especially when placed alongside the Newspoll results.  Voters believe the budget should be in surplus, but they tend (mildly) to approve of it not being in surplus this year, and they (very strongly) don't believe the Coalition would deliver a budget surplus in its first year either.  This, plus the Newspoll response on economic management, adds up to a very strange combination of attitudes.

Graham argues that voters have a kind of emotional identification with the surplus "because of the emotions it represents to people in relation to how they see Australia."  Voters are more concerned with the treatment of the ideal of the surplus than with whether it is actually delivered in concrete circumstances.  The Coalition uses the surplus issue to push the message that the government is untrustworthy on the economy, although it is not the government's management of the economy that has been demonstrated to be bad.  Rather, the government can't be trusted to make the right combination of noises that people want to hear.

Of course, the Coalition's lead on economic management might be seen to be about something other than the surplus - the carbon tax, the MRRT, etc.  But asking questions about these does not tend to lead to any clear sense of disapproval of any specific policy.  It would be interesting to ask the 50% who believe that the Coalition would be better economic managers, why they actually think that.

Last September I observed that "a bizarre government exists in the times that seem to suit it least" and tied this in to the themes raised in Scott Steel's much linked (perhaps even canonical!) Great Unhinging Revisited.  There's just so much that is odd about the Gillard minority government, in the contexts of a simplicity-craving time in politics, that many voters may be unable to fully accept the idea that for all its backflips, ructions, scandals, broken promises and errors, it still runs the economy at least as well in current circumstances as an Abbott-led Coalition.  There's a certain level on which they buy it, and a certain level on which they apparently don't.  If this is so, I really don't know what the way around this is, and I'm not sure that the government does either.   If Labor remains unable to sell itself to save itself (which in my view is a big part of its problem), its hopes of re-election will depend on the Coalition finding a way, beyond just Abbott, to lose.

1 comment:

  1. Kevin,

    I have recommended you as a must read by anyone interested in elections here.

    Keep it up.

    You are virtually always on my around the traps each friday

    ReplyDelete