Sunday, December 30, 2012

Uneven Swing to Liberals in Tasmanian State Election Polling

Advance summary:

1. Figures from three 2012 polls of Tasmanian state voting intention, show that swings between different electorates since the 2010 election are clearly not uniform.

2. Especially, Labor is polling very much better than the state swing suggests in Franklin, but generally worse in the northern seats and Lyons.

3. For a Liberal vote of around 50, the uneven swing pattern makes little difference to the seat total, with the party projected to win 13 seats (the barest possible majority).

4. For a Liberal vote well above 50, as in the November 2012 EMRS poll, the uneven swing pattern means that the party does not necessarily win more than 13 seats. 

5. However the uneven swing pattern also makes it possible (but not at all likely at this stage) for the party to win as many as 17 seats.


In the previous state article Tasmanian State Election Patterns Since 1989 I looked at the history of voting patterns in the five Tasmanian state electorates in the period since the state's present three-party system emerged.  I found that swings by electorate have a non-linear history, at least for the Liberals and Greens, such that the vote for each party tends to be more volatile in areas where they are performing well, and more consistent where (and when) their vote is low.  For Labor there is no clear pattern.

At the end of that article I promised to "have a go at seeing whether the limited and mostly dated electorate sampling available thus far provides any useful insight into electorate swings." and wrote that "Given the huge amount of bouncing between 200-vote samples, and the problems with data from some pollsters, I'm not overjoyed about the prospects there, but if there are strong regional differences they may pop up."

In fact, this was a much more useful exercise than I expected.

I am aware of three polls in 2012 for which electorate-by-electorate breakdowns have been publicly released.  They are the EMRS polls of February and May and the Liberal-commissioned ReachTel poll of October.  I think it is completely acceptable to use the latter for an exercise of this kind, for two reasons.  Firstly while internal polls that have been publicly released should not generally be used as good evidence of voting intention (because parties may cherry-pick which polls to release or deliberately time the conduct of a poll for the most suitable result), there is no reason to believe this would affect the distribution of vote by electorate compared to the sample as a whole.  Secondly the more recent EMRS poll, plus another unreleased poll of state voting intention that I have seen, show that there is nothing suspicious or unrepresentative about the Liberal-commissioned ReachTel in terms of overall vote distribution.  If anything, it is kinder to the Greens than the more recent sources.

I've taken the following approach to modelling possible electorate swings based on these three polls:

1. Assume that only the three parties will be competitive at the next election.  This assumption may well prove false (and indeed I for one hope it is false) but at this stage there is no known evidence of a specific fourth force that is threatening to win seats based on polling, and the only fourth force who has been a threat in the last three state elections (Andrew Wilkie) is currently in federal parliament and may well be re-elected there.

2. For each poll, find the difference between each electorate's results for each party and the state average.  Thus, in a given poll, if the state average for party X is 32, and the party polled 37 in electorate Y, then party X scores +5 for electorate Y.  This should largely address the problem of EMRS overestimating the Green vote, because an electorate sample with an overrepresented Green vote is being compared with a state total with the same issue.  Some errors may remain because the extent to which the Green vote is overestimated will not be constant around the state.

3. Average the results across the three polls.  Although one data source (ReachTEL) is much more recent than the others, we are still presumably a long way from the next election, and weighting the sample would have the effect of reducing the effective sample size (presently at about 600 voters per electorate).

4. Adjust the sum of swings for each electorate to zero, and also adjust the sum of swings for each party to zero.  (This eliminates some issues caused by varying levels of measured support for "other" and also some pooling discrepancies - in the case of ReachTEL these were discussed here, including comments 7 and 11.)

The result is an indication of how each party appears to be travelling relative to the state swing across these three polls.  (The state swing varies from poll to poll and the November EMRS poll shows a state swing that is more extreme than any of them):

So, for instance, the Greens appear to be travelling 2.2 points better than the state swing in Braddon.  If there was a 6-point swing against the Greens statewide (resulting in a state vote of 15.6%), these figures would predict a 3.8 point swing against the Greens in Braddon, leaving them with, on average, 10%.

The margins of error for each modelled electorate result are about 4 points for the Liberals, just over 3 points for Labor and just under 3 points for the Greens. 

A number of results here deserve comment.  Firstly in Franklin there is very strong evidence of the Labor vote from last election holding up better than in the state overall (with the swing to the Liberals in that seat being much lower than in other electorates.)  In the 2010 state election Labor underperformed in Franklin by 6.4 points compared to the state average, but in the three electorate breakdowns included in this sample, Labor has polled seven and 4.9 points above their state average in two, and level with it in the third. (Labor also polled six points above their state average in Franklin in the Nov 2011 poll, and even with their state average in the two polls before that.)

The 2010 Franklin Labor result can be attributed to a weak team containing two incumbents with low profiles (Ross Butler and Daniel Hulme, who were elected on recounts following the resignations of Paul Lennon and Paula Wriedt), while both Liberal and Green 2010 teams were headed by their state leaders, both of whom went into the election with strong net satisfaction scores.  In comparison, the Labor team in Franklin now looks much stronger than it did, with Lara Giddings elevated to Premier and David O'Byrne the most prominent of Labor's new faces and seen as a likely future leader.  Although there are no post-2010 polling figures that directly assess Nick McKim's popularity, I would expect it is now much lower than before the last election.    There may also be a dose of "premier pride" in the observed figures (similar to how federal voting intention is sometimes boosted in a state by the Prime Minister being from that state), but I also think that personnel is not the only factor, and that there is a north-south split in the current strength of feeling against the Labor-Green coalition government. 

The picture in Denison is very much clouded by the Wilkie factor and the persistently high rates of preference for Independent or Other candidates in polling (extremely high in the ReachTEL poll especially).  There is no clear sign that there is anything else interesting going on there but there do seem to be signs that the Green vote is holding up reasonably well there compared to the state swing, especially given the potential for the Ind/Other vote to damage the Green vote disproportionately.  The results for Braddon and Franklin for the Greens are fairly consistent with expectations based on the previous article.

Applying these figures to recent polling

The significance of this sort of finding about variation in swings between electorates can be seen if we apply this finding to the most recent EMRS poll.  The recent EMRS poll was great for the Liberals and exceptionally bad for the Greens (see full analysis at EMRS: Libs on course for massive win). Because of the scarcity of polling in Tasmania it is not possible to know whether the swings are really so large or whether they were amplified a little by random sampling noise (my suspicion, especially in the case of the Greens' supposed seven-point crash in three months, is the latter).  Federally, several dozen polls with sample sizes similar to EMRS's are conducted every year and an outlier or rogue is soon shown up, but when there are only polls every few months it is difficult to pick the duds.

But assuming that this EMRS poll (as modified by me to take into account the series' known foibles) actually represents either the way things are, or the way they will be by the time the election is held, here's what the election might look like with a uniform swing of the (massive) magnitude suggested by that poll (again, assuming no significant 4th-party candidates):

Translating polled outcomes to Hare-Clark seats is notoriously difficult, because Hare-Clark is about candidate totals and not necessarily party totals, and strange pitfalls like the dreaded Ginninderra Effect sometimes lie in wait for those who may believe they've got it in the bag.  Nonetheless all other things being equal the above results would very likely result in a 15-7-3 parliament.  They also wouldn't really happen quite like that, because the idea that the Green vote would crash from 13.8% to 5.6% in Braddon, even with a general swing of that size against the party, isn't credible. Indeed the Greens have always polled at least 8% in Braddon since 1989, even when their state vote was only about 10%. Other aspects of the uniform swing model, like a Liberal vote of 59% to 22% for Labor in Franklin, are pretty far-fetched too.

Here's what the same poll result looks like with the same assumptions except for applying the 2012 electorate swing findings:

The projected Green vote in Braddon on this model is probably still too low, but at least it's vaguely plausible if the swing was really bad.  The impact of varying the electorate swings is that the Liberals' third seats in Denison and Franklin drop into the danger zone (3.04 quotas vs 1.94 quotas is not a safe three if your leader polls most of the votes and has a massive leak-prone surplus).  However, they also come into contention for a fourth seat in each of Bass and Braddon.  Thus on these assumptions, the most recent EMRS poll delivers anything from a scraped one-seat victory (13-9-3) to a one-party state (17-6-2).  More likely it would be something in or near the middle.

(Note that 17 seats would, in the unlikely case that it happened, set a new record for the highest percentage of seats won by a party under Hare-Clark in Tasmania.  The current record is two-thirds (20/30), won by Labor in 1941 with 62.6% of the vote.  As far-fetched as these sort of scenarios may seem, I think it's worth keeping an eye on such possibilities.  Sometimes, as Queensland showed, these sort of things actually happen.)

Applying these figures to a less extreme state breakdown

Assuming that the November EMRS was a bit of an outlier, or that things will close up before the election, here's a different version of the same exercise.  To get the base figures in this case I've averaged my interpretation scores for all the year's EMRS polls, to get a base result (before excluding "others") of 52-29-16.25 instead of the 56-28-13 I gave for the last EMRS poll.  It's not a massive difference, but here it is with the assumption of uniform swings:

 (The Liberals might here, if very lucky, take a seat from someone in Lyons, but it wouldn't be counted on.  They gain seats from Labor in Bass and Franklin and a seat from the Greens in Braddon.)

And again with the assumption of different seat swings derived from the 2012 electorate breakdowns:

In this case, the Liberals no longer gain a seat in Franklin but they now gain a seat from one of the other parties in Lyons.

Uniform swing: an unsound model

This article and its predecessor have now shown three reasons why uniform swing should not be assumed in any attempt to project the 2014 state election:

1. Historical evidence shows that at least two parties achieve (or suffer!) greater swings when/where their vote is higher and smaller swings where it is lower.

2. When a uniform swing model is applied to a large statewide swing, it produces some silly results.

3. Electorate breakdowns available for 2012 show in some cases, and suggest in others, that at present, both regional and candidate factors are creating non-uniform swings.

Factor 3, especially, is so strong that even models based on historical evidence alone are not likely to be sufficient.  Effective modelling of the 2014 election will require careful use of electorate breakdowns from multiple polls (hopefully) issued closer to the date.

An important common factor in all these simulations is that it is very difficult to turn realistic numbers based on recent polling into a scenario in which the Liberal Party doesn't win outright.  As amusing as a parliament with exactly twelve Liberals and Kim Booth in it might be, the Liberals' polling would have to come down a fair way before we had to take the prospect seriously.

No comments:

Post a Comment