The consistent Labor lead before the election led most observers to expect a Shorten government. Now the pollsters may have to adjust their techniques.
2019’s federal election appears to be the most spectacular polling failure in Australian politics since regular polling began.
Polls in general suggested Labor would win roughly 51.5% of the two-party preferred vote – that is, the vote after minor parties’ preferences were distributed, under Australia’s preferential voting system. That was also the result in the final survey by Australia’s best-known political polling organisation, Newspoll.
Before election day, 17 straight polls predicted the result would go to Labor 51-49 or 52-48. A final Newspoll gave Labor 51.5%, suggesting a late swing to it. Even exit polling had it ahead.
The 2019 polls didn’t even show the thin signs of change that appeared late in 1993’s election, considered by some to be ‘unlosable’ for the Coalition under John Hewson. (In 1993, polling for The Age and The Sydney Morning Herald had Keating and Labor just in front on the Friday before the election. The papers reported the result without obviously believing it; it turned out to be correct.)
Two-party-preferred vote as shown by Newspoll, June 2017 to May 2019
Reverse the polling numbers
The miss of roughly 3 percentage points is a far more dire error than anything in the US polls. Those polls in some cases underestimated Trump’s chances of winning. They correctly predicted Clinton’s popular vote win, but did not clearly signal how that would play out in key Midwestern states.
In Australia, seat-by-seat polls in key Queensland and Victorian seats sometimes got the party winner right in a particular seat, but were often wildly off on the margins: Queensland’s Herbert and Forde were both 50-50 in Galaxy polls during the final week, but the Coalition has won them by at least 58-42.
“This is a national total polling failure more similar to Brexit or to recent UK national elections,” wrote independent election analyst Kevin Bonham deep on Saturday night. As he noted, betting markets failed, too; they had been so strong for Labor that SportsBet actually paid out on a Labor win before polling day.
— Sportsbet.com.au (@sportsbetcomau) May 18, 2019
Some bitterly disappointed pro-Labor tweeters declared on Saturday night that the polls are essentially invented, or based just on calls to the shrinking population that still has a landline.
In fact, the pollsters now use other tools than just landlines. And the parties’ own internal polls, while pointing a little less clearly to Labor, did not flag a Coalition win. Almost no-one within the Liberal Party seems to have expected such a result.
Find the polling problem
So what actually went wrong with the polls? Experts offer several possibilities:
- The ABC’s Antony Green argued on Saturday night that Australian polling has indeed been changed by the decline of landlines, so that pollsters have had more trouble finding a representative sample of voters.
- Monash University statistician Adrian Beaumont yesterday suggested a more specific sampling problem: not all pollsters ask for the educational backgrounds of their survey subjects, and it’s not clear that any adjust their results for this. Beaumont wrote that such adjustment might improve the outcomes.
- Beaumont and Bonham yesterday both pointed to a third factor – signs of obvious ‘herding’ in that recent run of 17 poll results that had Labor on 51% or 52%. The herding theory holds that nervous pollsters tweak their results to more closely resemble those of their rivals. An anonymous econometrician styling himself Mark the Ballot offered pre-election calculations suggesting the 2019 poll results were under-dispersed with less variations than unadjusted polls would show.
US pollsters have been warned by experts that they’ve relied too heavily on robocalls, internet surveys and incomplete voter information as they have tried to cut costs. Australian pollsters may have to absorb these lessons, too.