2021 Tasmanian State Election Polling Drought

One of the regular services I've provided on this site and, before that, on Tasmanian Times, is polling-based forecasting of Tasmanian elections.  At all elections since 2006 there have been multiple public polls that have allowed me to do this, and some of these polls have had very good sample sizes indeed. However this year (link to my 2021 guide main page) we have had no public polling since February.  By public polling, I mean a poll either released by a polling company off its own bat, or commissioned by a media outlet that is always going to publish the results.  We have had a single commissioned poll with voting intentions data released (uComms commissioned by the left-wing Australia Institute), and some rumours about party polling.

I would like to be able to present a forecast and say that it should be as reliable as in previous years, but with such a low level of polling data I can't do that.  There might be more voting intentions polling to come, but I don't have any specific reason to expect any.  This especially follows the news that EMRS, who last polled in February, will not bring forwards their May omnibus to include a poll for this election.  

Before I get into some fine detail about the maths of such a lack of polls, I want to cover some general points: how did we get here and why does it matter?  The tail end of this article is more technical and gets up to around 3/5 on the Wonk Factor scale.  

How we got here

Australian state polling experienced a golden age in the mid-2010s (especially with the Victoria 2014 and NSW 2015 elections, which were remarkably richly polled) but has been in a steep decline ever since.  This decline was happening before the 2019 federal polling failure, but it certainly didn't help.  The major driver of the slump has been cost-cutting by the media.   

In Tasmania's case, Newspoll polled at every state election from 1989 to 2014, and sometimes in between.  In 2018, Newspoll had switched to a mixed robopoll/online method and for whatever reason (unsure if financial or logistic) was not commissioned by The Australian to conduct a Tasmanian state election poll.  I have not heard anything regarding whether there will be a Newspoll for this year.  

EMRS has polled or near every state election since 2002.  However it normally polls four times a year.  Sometimes one of its polling periods happens to land just before the election, and sometimes even when this isn't the case it has done an extra poll that has appeared in The Examiner.  However, I believe The Examiner stopped spending money on EMRS polls many years ago.  In 2014, EMRS polled in February but not in March (the month of the election).  In 2021 this practice appears to be being repeated.  

ReachTEL polled at the 2014 and 2018 elections with very large robopoll samples that were outstanding in the latter case and good enough in the former.  ReachTEL has since been sold and is doing very little polling in Australia.  It has been supplanted by uComms, which uses the same technology but is a different pollster.  However, many media sources refuse to hire uComms because of their union links.  

Roy Morgan conducted regular phone polling of Tasmania until 2006.  Since then it briefly issued SMS state polling in 2014 and 2016, but not in the leadup to an election.

Some other pollsters have now and then emerged, for instance TasPoll, which polled in the 2006 election campaign.  Various other pollsters are also used in party polling, for which results are only selectively released when it suits parties to release them.  

Why it matters

Especially following the 2019 federal failure, there's been a lot of whinging about the existence of polls - much of it coming from people who placed too much stock in narrow polling leads that were never guarantees of Labor victory.  But there are many reasons why we should have more polls than we have seen this campaign.

The first reason is that voters deserve the information.  In this election, one party (Liberals) is claiming that only it can win a majority, but that it will be a close thing.  The leader of this party has said he will resign as leader unless he does win a majority.  Suppose a voter is genuinely tossing up between voting Liberal to keep Peter Gutwein as Premier and voting independent because they want more diversity in the parliament.  If the election is genuinely unclear, they might do the former, but if there's strong enough evidence that the Liberals will probably either easily get a majority or miss by a lot, they might be more inclined to vote independent.  Voters considering becoming active on the campaign trail in some way also deserve to know what is going on.  Decisions by busy people about whether to get involved might hinge on how a party seems to be actually travelling.  

Expectation management is also important for some voters.  While there is always the possibility that polls might be a somewhat wrong, they tend not to be way wrong, whereas unguided expectations can very easily get caught up in social media bubbles and go completely off the rails.  

A second reason is that polls help us understand what caused the result.  Suppose on Saturday the Liberals poll 44% and lose their majority by 1000 votes in Clark.  With no polling between the EMRS poll in February and the TAI poll in the final week, we will never know where the wheels fell off.  Was it the early election call?  Was it mistakes or shortcomings exposed in the final weeks of the campaign?  Or had the government's COVID bounce already worn off before the campaign started?  Or, on the Labor side, suppose the party polls 25% and wins only seven seats.  Was the result baked in because of COVID, or might the party have been more strongly placed before the campaign started, only to unravel during the campaign?  Having some kind of tracking helps us get some idea of what decisions were mistakes and what worked.  It helps us know what parties should do for the future.  Of course, parties have internal polls for this purpose too, but what we find out about them is always filtered by people with axes to grind, and almost never published in enough detail anyway.  (And quite a lot of party polls are rubbish.)

A third reason is that without polls you get pundits.  You get punditry to a degree anyway, but it's easy for narratives of who is winning or losing to run astray and dominate the media without anything to ground them in reality.  At any given time social media is awash with potent arguments about why something that either major party is doing will massively cost it support.  Tasmanian Labor started the campaign consumed by faction-fights, but did the voters actually care?  They've been criticised for their pokies policy backflip - was it actually a winner?  The Liberals have been criticised, perhaps above all else, for under-delivering on past health promises yet responding with more promises, but is it cutting through and actually hurting them?   The Greens have had a positive and relaxed campaign but haven't had much attention amid the major party fights and are at risk from indies - will they go up or down?   None of this is to say that picking out the impact of specific issues is easy even with rich polling, but it's easier to get some hints than it is with none.  Also, without good polls you get bad "polls".  Opt-ins treated as if they were scientific polling and so on.  

Poll-based forecasting has worked

I've published election forecasts based on aggregated polling from the two or three pollsters doing public polling each election for the last four state elections.  Below are the projected and actual seat totals based on these aggregates:

The success of these forecasts hasn't been because the polling has been super-accurate (it hasn't).  Rather, it's been because in Hare-Clark most seats are not close.  There are only five electorates, and most elections have some contests where you can move 5% between any party and any other party and the seat result is still the same.  

So if you have a few polls with large sample sizes and electorate by electorate breakdowns, it's possible to get pretty close.  In this case only one poll is less than two months old, that is a commissioned poll from a pollster untested at a Tasmanian state election, and there has been no electorate by electorate voting intention polling released at all.

Old rope is less reliable (unless it's green rope ...)

Generally as polling gets closer to election day it should, and worldwide does, become more accurate.  This effect can sometimes be especially important in Tasmania because of bandwagon effects such as those seen in 2006 and 2018, both of them fuelled by controversial third-party campaigns.  That's why it's risky to place too much stock in EMRS polls from months ago.  

I thought I'd have a look at how accurate EMRS polls have been in terms of raw errors by days (from last day in field) to polling day.  This is what I got for polls from 2002-2018, but I am missing the numbers for the final 2002 poll, though I know it was reasonably accurate.  Some of the others have had to be retrieved from media reporting in NewsBank and elsewhere so may not be exactly the correct poll numbers.  

Liberals

The horizontal axis is days before polling day and the vertical axis is the absolute difference between the EMRS poll result (undecided redistributed) and the actual outcome.  

Liberal Party results have tended to become more reliable as election day approaches especially in the last 50 days where the poll has usually been within 2%.  (Note: the relationship isn't statistically significant off so few data points; it's just what should be expected.)  However there was one truly massive outlier a few months out in late 2017, when the polled Liberal vote was a whole 16% below what the party actually got on election day, so it's possible that polls a couple of months out could be also out by double digits.

Labor

There isn't any overall evidence of EMRS estimates for Labor getting closer as the election does.  There are a few outliers of about the same size at different time scales (these all come from the 2006 campaign, at which EMRS was way out on the Labor vote less than two months out - that's not to say the poll was wrong, more likely voting intention changed.)  However, the three cases where EMRS released polls in the final week saw it much closer than average twice, and around the average error once.

Greens

Oh dear, EMRS poll estimates for the Greens have actually historically got worse as polling day gets closer!  Nearly always overestimates too.  (The overestimate was less in 2018 and their new methods may well have fixed this.  We just won't know if they don't poll close to election day.)

Another Aggregate Attempt - This Is Not A Forecast Or A Prediction

After noticing that it was possible to have a rough go at benchmarking the uComms polls using EMRS polls at a similar time, I decided to use this to have another go at an aggregation-based model the very poor collection of available polls and semi-public polls we have available at this election.  I mainly did this for a baseline to use if another poll becomes available.

For this one I estimated the house effect of EMRS from past election results and then benchmarked uComms house effects off that.  I found that every last-in-series EMRS poll I had (2002 excluded since I don't have it) had underestimated the Liberals (but never by very much) and overestimated the Greens.  However, maybe their new series has fixed that, maybe it hasn't.  And maybe the differences between recent uComms and recent EMRS polls indicate that the uComms polls are skewing to the left, but also, maybe they don't.  So I halved the weighting on all the house effect assumptions.  I also raised the weighting on the recent uComms to 30%, because if you can benchmark a poll (even off a sample size of two others) you're a bit closer to knowing something about it.

This is what came out:

Whether Labor wins two in Bass depends on whether its top two candidates are even enough to get both over either the Greens or the fourth Liberal.  Clark is a mess but the Liberals probably win two, even if I increase the Independent loading.   The Liberals probably don't get four in Bass but might.  Probably 13-9-2-1 would be the most likely result here.  There is some possibility of the indies both losing if the split between them is too even and the preference flow too weak.  

And this is the same thing with all the house effect assumptions removed:

This looks more like 13-8-3-1.  Even with the house effects removed, weighting the recent uComms at 30% doesn't go close to knocking the Liberals off a majority, though the removal of house effects does stop the Liberals winning four in Bass, and makes it far more likely the Greens will win there unless the split in Labor is very even.  And to knock the Liberals off a majority I'd need to do something like more or less discard the old polls so the uComms is most of the sample, and even then I'd need to assume the uComms had fairly weak house effects.  (I've arbitrarily assumed the Others vote in Clark will be double the state average and in Braddon on the state average.  Perhaps the first assumption will turn out to be conservative.)  

The above isn't a prediction of what will happen, because the data quality is not good enough.  It's just saying that it's currently hard for me to find mathematical evidence that there should be change in any particular seat without placing a very high weighting on a single uComms poll.  That said, in Western Australia it was generally assumed the first Newspoll with 68:32 to Labor was too lopsided and the election would probably be closer.  In fact Labor won by even more than that.  So I won't assume that any error in the uComms poll must have gone against the government.