A long and prominent association with News Limited has given the poll a disproportionate amount of influence
It has been one of those weeks that reminded us of the outsized influence of Newspoll on Australian politics and on commentary on Australian politics.
Let's begin by recognising that polling is tough. Survey research is an ever-competitive business, with web-based upstarts intruding on the higher-cost, "live interviewer", random-digit dial mode used by Newspoll. As this week testifies, political polling is subject to detailed scrutiny and commentary. Election polling also brings with it a "day of reckoning" of sorts, when the hitherto, unknown target you've been trying to hit is revealed on election night.
Newspoll had a good performance with its final poll ahead of the 2010 election. Newspoll fielded a survey of 2,507 voters on 17-19 August, 2010; the election was on Saturday, 21 August. Newspoll's estimates were close to the actual outcomes: Labor 36.2% (actual 38.0%), Coalition 43.4% (actual 43.6%), Greens 13.9% (actual 11.8%), Labor two-party preferred 50.2% (actual 50.1%).
Newspoll's 1.8-point underestimate of the Labor primary vote in 2010 was largely offset by a 2.1-point overestimate of the Greens' House of Representatives vote – both misses are very close to or outside a conventional 95% credible interval or "margin of error" for a poll with a sample size this large – resulting in a two-party preferred estimate that was virtually dead-on.
Newspoll also has a good track record in earlier federal elections. In 2007 I found Newspoll to slightly overestimate Labor's primary vote, by 1.2 percentage points, not enough to trip conventional levels of statistical significance. But 2004 was not such a good outing for Newspoll, with its final pre-election poll (October 6-7, 2004) underestimating the Coalition's two-party preferred result by a substantial margin: 50% versus the actual result of 52.7%, well outside a conventional 95% confidence interval (n = 2,500, so +/- 2.0pp).
Much of Newspoll's influence stems from its longevity, frequency and prominence, via its association with News Limited (now rebranded as News Corp Australia) and its newspaper the Australian. Any other poll with a comparable track record simply wouldn't have as much impact on Australian political discourse without the visibility and support that Newspoll enjoys via the Oz.
So with that as background, what about this week's reaction to Newspoll?
As is often the case in tense political times, there was a good deal of anticipation and speculation ahead of the release of Newspoll, with journalists writing and tweeting that a "good Newspoll" (for Labor) would see Rudd under increased pressure to bring the election on.
One journalist tweeted that they simply "couldn't pick it" (ie, would it be a better or worse Newspoll for Labor?), which is not much different than trying to guess the outcome of a coin flip, when you think about (1) the likely magnitude of opinion change in the two weeks since the last Newspoll and (2) the magnitude of the sampling error associated with Newpoll's samples of size (1,100 or so); more on that in a second, but such is commentary in an age of "140 characters or less" (the length of a tweet, for the Twitter virgins).
Within hours of the Newspoll results being released – Labor -1 on primaries, Coalition +3, Greens +1, Labor TPP 48% (-2) – the betting markets reacted. Labor had been at $3.30 at Centrebet on Sunday, 21 July and into Monday 22 July, into $3.10 on Monday night, but out to $3.40 late Monday night once the Newspoll figures were out. By 10am Tuesday morning, Labor had eased further, $3.55 to the Coalition's $1.28, or an implied probability of Labor winning of 26.5%.
A slightly bigger swing occurred at Sportsbet, with Labor easing from $3.00 on Sunday to $3.25 on Monday afternoon, but then to $3.75 by Monday morning, against the Coalition's $1.25. In years of closely watching the Australian betting markets, I can confidently attest that no other poll moves the betting markets more than Newspoll, further evidence of its disproportionate influence.
Comments on some widely read blogs said that the downturn in Labor's Newspoll result would mean that speculation about an 31 August election would end, as would murmurings about Turnbull replacing Abbott. Is Rudd's honeymoon over? Has Labor hit its high water mark?
Lost in the clamour were a few voices trying to stress that the movement suggested by the poll was inside the "margin of error". It is worth stressing that these references to the margin of error did include Newspoll's Martin O'Shaughnessy and the Australian's Dennis Shanahan.
So what is the appropriate conclusion to draw? This week's Newspoll is just one of several nationally representative polls fielded recently. I've developed a statistical model for pooling survey results, taking into account (1) estimates of voting intentions based on earlier polling, (2) the track record of each pollster, when known, (3) the different sample sizes used by each pollster. I've deployed this model in Australian and American elections since 2000; see the references, below, for more details.
A graphical representation of the model's output given the recent stream of national polls appears below. Labor's TPP vote share jumped about 5.5 percentage points in reaction to Rudd's return to the leadership, followed by about another 1.5 percentage point improvement to 50% in the week that followed. The model output does not suggest any further improvement since an apparent Labor TPP peak of 50.1% (+/- 1.3pp) on 6 July, and, if anything, a decline to 49.6% (+/- 1.3pp) given the very latest national polling.
If we omit this last 48-52 Newspoll from the modelling, the estimate of Labor's current TPP level goes up slightly, to 49.8% (+/- 1.4pp), indistinguishable from the 49.6% result obtained by including this last Newspoll. In short, the Newspoll sample size (1,141) and the 48-52 result is largely overwhelmed by the force of the other available data.
Only about one thing is unambiguous in the data: Labor got a six- to seven-point boost in TPP from switching back to Rudd, to be polling at levels indistinguishable from its 2010 election result.
There is insufficient evidence to confidently conclude that Labor's current TPP levels are falling, although the data are more consistent towards a fall than a continued rise in Labor's TPP. Including the last Newspoll, I find that there is a 74% chance that Labor is currently polling below the apparent 50.1% TPP peak on 6 July (the "Labor's falling" hypothesis is three times more likely than the "Labor's rising" hypothesis). Excluding the last Newspoll, the probability that Labor's TPP has fallen below a July 6 peak falls to 60%.
The bottom line remains that the national polling points to a close contest, perhaps with the Coalition narrowly ahead. And when it's this close, we're going to need a lot more polling to authoritatively state who is ahead, or which way opinion might be trending. Indeed, when it is this close, the more valuable set of polls would be from marginal seats.
My analysis highlights the relatively small impact of Newspoll – or any single poll – on a hard-headed statistical analysis of the data.
Newspoll is a poll, and a good poll at that. But from the dispassionate perspective of a statistical model, it's just another 1,100 interviews, bumping against the other 3,000 to 4,000 interviews spread out through the other polls – not that you would have picked up on much of that from media commentary and Twitter traffic earlier in the week.