Research · When the polls missed

The miracle that was a sampling problem: Australia's 2019 polling failure

Abstract. All five final polls had Labor winning the two-party-preferred vote 51.4 to 48.6; the Coalition won 51.5 to 48.5. The industry inquiry blamed samples that held too many engaged, highly educated voters and were never weighted to correct it.

Polling Forecast · 2 November 2026 · 6 min read · Australia, 2019 federal election, polling error, final polls, inquiry, pollster methods

Summary

Every national poll published in the 2019 campaign had Labor ahead on the two-party-preferred vote. The five final polls averaged Labor 51.4, Coalition 48.6. On 18 May the Coalition won 51.5 to 48.5 and kept government with 77 seats. The error on the lead was 5.9 points, and the industry's own inquiry found the cause in the samples: they held too many politically engaged, university-educated voters, and nobody weighted that away.

What the polls said

The final polls from the five national series, as set out in the AMSRS/AMSRO inquiry's Table 3:

Pollster	Fieldwork end	Mode	Coalition	Labor	Greens	One Nation	2PP (Coalition–Labor)
Essential	14 May	Online panel	38.5	36.2	9.1	6.6	48.5–51.5
Ipsos	15 May	Phone (landline and mobile)	39	33	13	4	49–51
Newspoll	16 May	Robopoll plus online	38	37	9	3	48.5–51.5
Roy Morgan	12 May	Face to face	38.5	35.5	10.0	4.0	48–52
YouGov Galaxy	15 May	Online panel	39	37	9	3	49–51
Average			38.6	35.7	10.0	4.1	48.6–51.4

Five pollsters, four different modes, different days in the field and samples from 1,004 to 3,038, and their 2PP figures sat within one point of each other. The inquiry noted that the 16 estimates of the Coalition's 2PP over the last four weeks of the campaign all fell between 48 and 49.

What happened

The Australian Electoral Commission's figures: Coalition 41.44% of first preferences, Labor 33.34%, Greens 10.40%, United Australia Party 3.43%, One Nation 3.08%. On the national 2PP the Coalition took 51.53% and Labor 48.47%. The Coalition won 77 of 151 seats, Labor 68, and six went to the Greens, Katter's Australian Party, Centre Alliance and three independents.

Against the average, the polls had the Coalition 2.8 points too low and Labor 2.4 too high. The Greens (0.4 too low) and One Nation (1.0 too high) were close. The whole miss sat in the two big parties: the polls had the Coalition 2.9 points ahead on first preferences, it won by 8.1, and Labor's 2.8-point 2PP lead turned into a 3.1-point deficit.

Why it happened: what the pollsters found

On the Monday after the election the Association of Market and Social Research Organisations (AMSRO) set up an inquiry, joined by the Statistical Society of Australia. The panel was chaired by Darren Pennay of the Social Research Centre and included Murray Goot and Jill Sheppard, with Patrick Sturgis (who led the British inquiry into 2015) among its international advisers. A discussion paper on disclosure came out in May 2020 and the final report, dated October 2020, was published in November.

Its conclusions:

It was "a 'polling failure' not just a 'polling miss'": the polls erred significantly, all in the same direction and by a similar amount, and the source of the error lay in the polls themselves.
It ruled out a late swing "except possibly to a very minor extent", along with shy conservatives, question wording, respondents misleading pollsters, early voting and ballot order. Preference allocation added a little error to some 2PP estimates but was not a major cause.
The most likely cause was that "the samples were unrepresentative and inadequately adjusted", skewed towards more politically engaged and better educated voters. In the post-election CSES survey, 54% of tertiary-educated voters backed Labor or the Greens against 33% for the Coalition. Among those with no post-school qualification it was 40% to 48%. In Australian samples 31–44% held a bachelor's degree, against 20% of adults in the 2016 Census.
It could not rule out herding, because none of the pollsters handed over raw data. It put the chance of five final polls being this similar and this wrong at about 1 in 1.5 million. Lonergan Research admitted it had held back one Queensland poll that looked out of step.

The inquiry reweighted the CSES survey to show the effect. Adding education to sex, age and location cut the Coalition's error from 2.1 points to 0.6. If the polls had moved by the same amount, the Coalition would have been on about 40.5% and the first-preference gap 4.7 points rather than 2.8. Adding self-assessed health took the error on the gap close to zero.

YouGov, which ran both Newspoll and YouGov Galaxy, offered its own account. Campbell White said robopolls were "answered largely by older people or those who are very interested in politics". In a submission to Labor's review YouGov blamed "a reliance on 2016 vote recall as a weighting factor". The inquiry's methods table shows YouGov Galaxy already weighted by education as well as past vote, and it still missed the 2PP by 2.5 points.

What our data shows

Our record of final polls (see the Australian polling error paper) puts 2019 in context:

It was the whole industry. The Coalition's first-preference error is 98% shared: every final poll had it between 2.4 and 3.4 points too low. For Labor the shared part is 72%. Ipsos came closest on first preferences (Labor 33, almost exact), but its Greens figure of 13 was the highest of the five and its 2PP of 49–51 was still 2.5 points out.
The minor parties were fine. The average missed the Greens, One Nation and others by between 0.1 and 1.0 points. This was a miss on the Labor-Coalition split, which fits a sample with too many Labor-leaning, highly engaged voters better than a turnout or preference story.
It was not unusually large, only unusually decisive. The average miss per main party on first preferences was 1.87 points, against 2.74 in 2010 and 1.97 in 2025. What set 2019 apart was that every poll put the wrong side ahead on the measure that decides government.
The lean did not last. Labor was overstated at eight of the ten elections in our record. Then in 2025 all ten final polls had Labor's 2PP too low, by 2.2 to 4.2 points: a 2PP miss of 2.9, the same size as 2019 in the other direction.

What the pollsters changed

Newspoll. YouGov announced on 19 September 2019 that it would drop the robocall-plus-online mix it had used since 2015 and run Newspoll entirely online, using "the synthetic sampling procedures we use globally". The new series started in late November 2019, weighted by age interlocked with education, with household income mentioned too and quotas set by electorate type. In August 2023 Newspoll moved to Pyxis Polling and Insights, formed by YouGov's former head of polling.
YouGov took the same overseas methods into its other Australian polling, moved its seat polls to live phone, and ran an MRP model in 2022.
Essential published more detail and from July 2020 began reporting a "2PP+" that keeps undecided voters in the figure.
Roy Morgan stopped face-to-face interviewing in early 2020 and moved to an online and phone mix.
Ipsos lost its Nine newspapers contract. Resolve Strategic took over Nine's series in April 2021.
Disclosure. The inquiry made ten recommendations: a code of conduct on the British Polling Council model, education weighting, new 2PP methods, honest measures of uncertainty and 24 disclosure standards. The Australian Polling Council, set up after the election, commits its members (YouGov, Essential, Ipsos and others) to publishing method details.

What it means for our forecast

We take half the average past miss off each party's polls only if that correction would have improved past forecasts, with each election predicted without knowing its result. In Australia the test fails. The correction was closer at two of seven elections and further out at five, mostly because 2025 reversed the Labor lean that 2019 seemed to confirm. So we do not correct Australian polls for 2019. We use its size instead: each party's simulated national error has the spread of the past record (about 1.9 points for Labor and the Coalition), which keeps a 2019-sized miss in play in every forecast. We also build our own 2PP from first preferences and the last election's preference flows, not from the pollsters' 2PP. Pollster-level records are on our Australian pollster ratings.

Sources

How to cite

Polling Forecast (2026). The miracle that was a sampling problem: Australia's 2019 polling failure. Retrieved 2 November 2026 from https://www.pollingforecast.com/research/missed-au-2019. Live figures change as new polls arrive; the date above is when the text was last revised.