Pollsters are taking drastic measures including scrapping phone calls after repeatedly predicting presidential elections incorrectly.
In 2016, just a handful of polls predicted that voters would send Donald Trump to the White House.
Respected statisticians gave the Republican a 1 per cent chance of becoming the commander-in-chief – just days before he did exactly that. Virtually the only question was what Hillary Clinton’s margin of victory would be.
Eight years later, and with just days to go, Kamala Harris’ team will be kept up at night by the thought that pollsters are once again underestimating the level of support for Trump.
As of Friday, things were on a knife edge with Ms Harris leading in the majority of polls by a whisker.
It comes as, in the lead-up to the race, pollsters were furiously updating their methods after coming under fire for inaccurately predicting the last two presidential elections.
“If you go back 10 years, most pollsters were conducting surveys by telephone,” Scott Keeter of Pew Research told the Telegraph.
“That was very accurate. The polling in 2004, 2008 and 2012 was very good. After that, it became increasingly expensive to do it.
“Telephone polling has virtually disappeared. There has been an explosion in “opt-in” polling, where people volunteer and that has proved to be less accurate.
“Pollsters have faced the challenge of trying to reach people who distrust institutions and are more likely to vote for Trump.
“We are in a period of experimentation; we are in a position of the Wild West. There has been a rapid growth in the number of pollsters out there.”
Trying to redress the balance, some pollsters are now asking people how they voted in the past.
“That is a controversial practice, but it is an effort to get over the problem of the Trump voters who may not participate,” said Mr Keeter
Another problem with telephone polling has been the recent surge in scam robocalls, which has led to many people using call blockers.
“By 2022, we would dial 100 numbers in order to get one person to pick up,” Charles Franklin of the Marquette Law School Poll told Politico.
“That was just driving costs way, way up. And it wasn’t people refusing to do interviews. It was not picking up so that we couldn’t get the foot in the door.”
The issue means that more participants are being approached by text, post and online, with pollsters convinced the hybrid approach will prove more accurate.
Rather than choosing panelists at random, firms carrying out the polling are using questionnaires to create – or at least so they hope – a more representative cross-section of voters.
Philip van Scheltiga, of Redfield & Wilton Strategies, the pollsters used by The Telegraph, said great care is taken in the selection of participants.
“We have completely changed the way we recruit our respondents, finding them where they are, on everyday apps that they use,” he said.
“The hope is that by being even more expansive in our outreach, we can reach a more representative sample before we get into weightings and such.
“If you don’t have good recruitment, there’s not much you can do by weighting the data to fix that. Bad data in is bad data out.
“At the end of the day, people shouldn’t expect polls to be perfectly precise.”
Not representative
According to Prof Steven Smith of Washington University in St Louis, there have also been changes in how the results are weighted.
This entails crunching the numbers and tweaking the results to create a more accurate snapshot of likely voters.
“It’s inevitable that a sample of respondents will not be representative of the population. Weighting is a question of what is relevant,” he told The Telegraph.
“In general, there are just a few demographic characteristics which are used. One thing lacking in 2016 was weighting for education.
“Also missing were people who were not responding. There is reason to believe in 2016, many Trump supporters were suspicious and refused to answer.
“There doesn’t seem to be substantial under-polling of working-class white males as there was in 2016. The weighting issue has resolved itself.
“This election seems to be hinging on turnout rather than candidate choice. There seem to be relatively few voters who have not made up their mind.”
Data from polling aggregator FiveThirtyEight shows that in 2016, just one nationwide poll gave Trump the victory in the race in – and even then it was by a margin of just 1 per cent.
On polling day, the Princeton Election Consortium gave Mrs Clinton a 99 per cent chance of beating her Republican rival, and predicted she would claim 312 electoral votes.
Hours later, she ended up with just 227 electoral college votes. It was her rival who ended up walking through the doors of the Oval Office.
And although Joe Biden did win the election in 2020, his margin of victory was far tighter than had been expected and the polls were the least accurate in four decades.
Polling averages put the Democrat eight points ahead of Trump – some even gave him a double-digit lead. In the end, Mr Biden won the election with tiny margins of victory in a handful of swing states.
Then in the mid-term elections, many pollsters predicted a sweeping Republican victory which failed to materialise.
A study by Pew Research found that 17 per cent of pollsters were using several different techniques this year to compile their samples, compared with only 2 per cent in 2016.
After the 2020 election, more than a third changed their approach to choosing participants.
How accurate the polls will be is a matter of debate, and remains to be seen.
