Most polls underestimated Labor. How did they get it wrong?

As opinion polls firmed against Peter Dutton in the final weeks of the election campaign, Labor and Liberal operatives warned the numbers could be wrong, just like in 2019.

In the end, they were wrong, but in the other direction.

Labor's national two-party preferred vote share is 55 per cent on current counting and expected to settle at around 54, according to Poll Bludger's William Bowe.

YouGov's MRP, RedBridge, Resolve and Essential landed on 53. Newspoll, which did Labor's internal polling, sat on 52.5, and YouGov's national poll on 52, all low but within the margin of error.

But disastrously for the Dutton campaign, its own internal pollster, Freshwater, was the most inaccurate, putting it on 51 — a result that would have seen Labor go backwards on election night.

As late as the last week of the campaign, multiple sources told the ABC that Liberal internal polling was showing it could pick up Labor seats like Gorton in outer Melbourne.

And as competitors detected a shift in Queensland in the last fortnight, the Liberal internals showed no cause for alarm, giving Dutton the confidence he could win his own seat of Dickson and send Labor into minority.

Instead, his campaign was sleepwalking into the worst loss in modern Liberal history.

Sampling challenges

The most complicated and time-consuming task for pollsters is deciding who to poll.

The difficulty is not just an appropriate sample size, but a representative sample — a poll of 20,000 people could be less accurate than a poll of 1,000 if those selected give a skewed picture of the Australian voting public.

For this reason, pollsters ensure their samples are balanced by age, gender and other characteristics, including prior voting history.

When they can't achieve balance in who they survey, they correct the results by "weighting" the survey responses — giving extra weight to the answers of people who are under-represented in the sample.

That imperfect process can account for polling differences.

After the error of 2019, pollsters including YouGov, Ipsos, Essential and SEC Newgate signed a voluntary code of conduct, agreeing to publish their methodology.

Freshwater is not part of this group, but its pollster, Michael Turner, penned an opinion piece in the Australian Financial Review, which publishes Freshwater polls, seeking to explain why he "missed Labor's landslide".

Mr Turner said Freshwater overestimated how many previous Labor voters who had voted "No" in the Voice referendum would shift to the Coalition, and how many minor party voters would preference the Coalition, and that its "fieldwork" finished too early to pick up a very late swing.

Freshwater was not alone in controlling for referendum votes.

YouGov, which publishes its methodology, also ensures appropriate sampling of "Yes" and "No" voters, but was able to correctly predict that Labor would substantially increase its majority.

And most published polls detected a swing in Labor's direction in the final two or three weeks of the campaign, including tracking polls that included Queensland seats.

How to poll the young?

For all pollsters, another challenge is finding enough young people to make a sample representative without relying too heavily on weighting, a challenge which may explain any bias towards the Coalition.

Polling methodology makes a significant difference to this, with young people far less likely to participate in phone calls, whether conducted robotically or by real people, and slightly more likely to fill out online polls.

The need to verify users makes these online "panels" more expensive to do properly, leaving smaller pollsters reliant on phone calls vulnerable to samples that skew old, a particular problem in cities.

YouGov pollster Paul Smith said panel curation was central to his company's success.

"Getting the younger demographic is particularly challenging, that's why we invest significant funds in having a panel that can provide the full demographic range of Australians so we can provide high-quality results," he said.