THE United States political world was stunned in 2016 when Donald Trump won despite pre-election polls showing him behind, and in 2020 Joe Biden's winning margin wound up slimmer than polling suggested.
Have pollsters learned enough from their mistakes to be more accurate this year?
With just days to go, Democrat Kamala Harris and Trump seem locked neck and neck. But if polls are once again underestimating the Republican vote, the ex-president could well be ahead.
The heart of the problem hasn't changed since Trump's dramatic 2016 upset, experts say, with a chunk of his electorate refusing to respond to polls.
"We haven't found a silver bullet," Courtney Kennedy, vice- president of methods and innovation at the Pew Research Center, said.
In the meantime, each polling firm is taking its own steps to try and rectify the issue.
In 2020, many people when reached by phone "would just simply yell at us 'Trump!' and hang up", said Don Levy, director of the Siena College Research Institute, which conducts highly regarded polling for the New York Times.
In an attempt to better account for such voters, Siena now registers responses from people who hang up, even if they don't answer all of their questions.
It also tries several times to reach people who don't pick up the first time, Levy said.
"If we call them the third or fourth time, we get more of them, so more callbacks, more potential Trump voters in the sample."
Pew, on the other hand, Kennedy said, now allows respondents to answer online or by phone. "We get very different types of people participating those different ways."
Once responses have been collected, pollsters must decide how to weight the data.
If a particular group is under-represented in the data compared to their proportion of the general population — say for example Republicans from rural areas — those responses are usually given greater weight in the final calculation.
Levy said the New York Times/ Siena poll performs weighting based on what they think the 2024 electorate will look like, using previous elections as a reference point, along with a mix of other factors.
Joshua Clinton, a political science professor at Vanderbilt University, expressed caution about weighting purely based on previous elections.
"Suppose there's more enthusiasm among Democrats than Republicans... then weighting to 2020 will basically undercut the support among Democrats in your poll."
It's a "no-win situation", said Clinton, with pollsters able to find out if their weighting was correct only until after actual voter data comes in.
"We're weighting to a universe that doesn't yet exist," said Levy.
With so much focus on Trump voters possibly missing from polls, Clinton warns that on the flip-side, Democratic support may also be underestimated.
After 2016 and 2020, "you may be tempted to conclude that the polls are always going to understate Republicans... but that's not true", he said.
In Michigan, a key state in this year's presidential election, polls underestimated Democratic support in the 2022 midterm elections, he said.
"Who knows what's going to happen in 2024?"
Siena College's Levy also raised the possibility of a "shy Harris voter", someone who does not want to publicly disclose their support for the vice-president, possibly because they live among staunch Trump supporters.
When asked if polls could be over-correcting for missing Trump supporters, Pew's Ken-nedy expressed scepticism.
"I've seen enough data to make it very clear that it's really hard for surveys to reach enough Trump supporters and to do a really complete job getting that right."
With the contest so tight and well within the margin of error, "we really should just tell ourselves it's tied, and not pretend that we can get stuff right in the decimal place", she said.
The writer is from AFP