Following Republican president-elect Donald Trump's resounding victory in this month's US election, some pollsters are scrambling to understand why their surveys once again underestimated his support among American voters.
Polls before the November 5th vote had shown Mr Trump trailing Democratic vice president Kamala Harris by 1 percentage point, according to an average of dozens of national opinion polls compiled by 538, a website on data analysis for politics and sports.
With vote tallies nearly finalised, Mr Trump led Ms Harris nationally by 2 percentage points: 50 per cent to 48 per cent.
That undershoot of 3 points follows polls in 2020 that underestimated Mr Trump's national margin by 4 points and in 2016 that underestimated his performance by 2 points, according to 538's averages.
Those results are all largely within pollsters' normal margins of error -- a measure of statistical precision influenced by a variety of factors including the number of respondents to a poll -- and the narrow divide of the American electorate makes getting the call right especially difficult.
But the fact that national presidential opinion polls have persistently been off in the same direction is driving suspicion among some pollsters and polling experts that surveys are having trouble getting responses from Mr Trump supporters, leading them to underestimate his strength.
It's hard to escape the idea that there's a challenge in reaching Trump voters.
"It's hard to escape the idea that there's a challenge in reaching Trump voters," said Charles Franklin, a pollster at Marquette Law School in Milwaukee who will be on a task force assessing pre-election polls for the American Association for Public Opinion Research, the main professional organisation of experts on US polling and survey research.
Even in the battleground states of Arizona, Georgia, Nevada and North Carolina - where the surveys were largely correct in pointing to Mr Trump victories on November 5th - the former president's margin on election day was about 1 to 3 percentage points higher than it was in the polls.
Rust Belt
In the Rust Belt swing states of Pennsylvania, Wisconsin and Michigan - where Mr Trump was behind in the surveys but won the vote - the polling averages similarly underestimated him by 2 to 3 points.
Vanderbilt University political scientist Josh Clinton, who led a task force analysing how surveys performed in the 2020 election cycle, said he was worried that the apparent mismeasurement could amplify its possible cause: public distrust of polling and of political institutions writ large.
"It delegitimises the whole operation in ways I think are pretty damaging," said Mr Clinton, one of four leading US pollsters and polling experts interviewed by Reuters who have participated in past American Association for Public Opinion Research postmortems or are due to participate in this year's exercise.
Despite missing the mark, some polling experts stress that getting within a few points of the result remains a remarkable feat and improving their record in the future will be difficult for pollsters. In addition to being less likely to answer their phones, the huge increase in volume of spam emails and texts over the past decades has made Americans less receptive to unsolicited contact in general, they said.
"This is a very different world from what we were dealing with 20 years ago or 30 years ago," said Christopher Wlezien, a political scientist at the University of Texas at Austin who participated in 2016 and 2020 presidential polling postmortems.
Trump effect
Polling experts told Reuters they remain unsure of why polls underestimate Mr Trump, who has railed against polls as biased against him throughout his political career.
"Pollsters, DC pundits, and the media always underestimated President Trump and his historic coalition of supporters," said Trump transition spokeswoman Karoline Leavitt. "The only poll that matters happens on Election Day."
Ms Harris' campaign did not respond to a request for comment.
The American Association for Public Opinion Research postmortem following the 2020 election considered the possibility that polls underestimated Mr Trump because his supporters responded less to surveys, consistent with his brand as a political outsider. But the pollsters and experts on the panel didn't draw a firm conclusion.
The idea is expected to be among those discussed by the 2024 taskforce, according to interviews with two pollsters who say they plan to attend.
Polling experts and pollsters also say it is possible that pollsters are also introducing errors when they process their surveys' raw numbers to give more weight to the responses coming from people they consider most likely to actually vote.
Fine-tuning
Measuring support for Mr Trump seems to be a particular problem for pollsters.
Public opinion polls did not similarly underestimate Republican strength heading into the 2018 and 2022 congressional midterm elections when Mr Trump wasn't on the ballot, said Mr Franklin.
That suggests something different could be happening in presidential elections, when the electorate includes more people who only vote occasionally.
Close to two-thirds of eligible voters cast ballots in this year's election, compared with just under half in the 2022 congressional elections, according to an analysis by the University of Florida's Election Lab.
Mr Franklin, who runs Marquette's Wisconsin poll, tried harder this year to reach Trump voters. Instead of dividing the state into five regions to sample from by randomly dialling people's phone numbers as it had in the past, Marquette's poll this year divided the state more finely into 90 regions and followed up with potential respondents first by email and then by phone, all in an effort to keep response rates higher across both Republican and Democratic towns and cities.
Mr Franklin said those changes may be why his poll underestimated Mr Trump's margin by about 2 points this year, compared to a 4-point miss in 2020.
The survey non-response problem also predates Mr Trump's rise as an anti-establishment icon.
Response rates have been falling for decades among large non-political surveys like those conducted by the Census Bureau. "We are dealing with a major non-response challenge," said Chris Jackson, head of US public opinion research at Ipsos.
