The pollsters blew it in 2020. Will they be wrong again in 2024?

Surveys had indicated that Joe Biden was closing strong against then-President Donald Trump. He led by a comfortable 8.4 percentage points in the final Fivethirtyeight.com average of national polls just before the election, and by 7.2 points in the RealClearPolitics polling average.

Biden ended up winning the national vote by less than 4.5 points—a lead that barely let him eke out victory in the Electoral College. If polls are missing the mark this year by the same magnitude, the narrow leads for Vice President Kamala Harris in many national averages today would actually be leads for Trump.

Pollsters have spent the years since 2020 experimenting with ways to induce hard-to-reach voters to participate in surveys and testing statistical techniques intended to improve accuracy. But expert opinion is mixed on whether polling is in for a repeat of 2020, which the professional association of pollsters called the most inaccurate performance in 40 years. New developments, such as the shift of Black and Latino voters toward Trump and the proliferation of online surveys, are creating potential sources of additional error.

“We are headed for more disaster," said Jon Krosnick, a Stanford University political scientist. Among other problems, he believes many of the newer, online surveys are using unproven sampling methods.

Courtney Kennedy of the Pew Research Center has tracked the changes pollsters made in recent years to address accuracy, and she is cautiously hopeful of improvement.

“I wish I could say the polling industry has done these three things to give us all very high confidence that polls won’t systematically underestimate Trump’s support again," said Kennedy, who supervises survey design and data science at Pew. “But that’s not the case. Pollsters have tried very hard to correct for the error, but there’s no silver bullet."

The industry’s analysis of 2020 polling, conducted by a select committee of the American Association for Public Opinion Research, was full of dour news. Pollsters had understated Republican support not only in the presidential race but in elections for Senate and governor. No method of polling—by phone or online—emerged as the most reliable, leaving few clues for improving surveys. Pollsters had fixed problems that skewed some results in 2016, such as talking with too few working-class, white voters, Trump’s most supportive group. But new sources of error had apparently turned up.

In the aggregate, the panel said, polls overstated support for Biden by 3.9 percentage points in the national vote in the final two weeks of the campaign. That was a departure from 2016, when national polls were among the most accurate in 80 years but state-level polls failed to detect signs of Trump’s eventual victory in the Electoral College.

In 2020, the final, pre-election Wall Street Journal poll of the presidential race, conducted at the time with NBC News, found Biden leading Trump by 10 percentage points. The result came close to indicating actual Biden’s level of support—he won 51.25% of the national vote, compared with 52% in the survey—but understated Trump’s ultimate 47% share by about 5 percentage points.

A postelection analysis found that the poll understated the share of white voters and older voters who ultimately cast ballots, among other missteps, and it overstated Biden’s support in urban areas compared with the actual results. NBC’s pollsters have taken steps to address those issues. The Journal now conducts its polls separately from the network.

Pollsters are eager to point out the limits of their own craft. While surveys this year are revealing important, durable results—resistance to Biden seeking a second term, for example, was high, and core Democratic groups were unenthused by his candidacy—no poll can detect whether voters on an Election Day at some point in the future will favor one candidate or another in a state that will be decided by a small fraction of the electorate.

Moreover, 2020 presented some unusual conditions. Covid prompted some states to make it easier to vote. Early voting became more popular, and turnout hit a historic high. “All of these were confounding factors in measuring ballot support," said Bill McInturff, a Republican pollster works on NBC News polls.

Pollsters after 2020 realized that they needed to work harder to contact hard-to-reach voters and ensure the right mix of respondents. Even if they had enough Republicans in their surveys, the GOP voters who were most excited about Trump might also be the ones least likely to take polls. While pollsters worked to include more working-class, white voters, the voters in that group might include too many who work in offices and too few who work in fields such as construction, which could skew results.

Now, many pollsters use multiple techniques to reach voters, not just through landline phones and cellphones but also by texting voters and inviting them to take a survey online or with a live interviewer. Some groups, including Pew, are turning to an old form of communication—reaching some respondents by mail and inviting them to fill out surveys on paper if they don’t want to go online.

All pollsters statistically adjust the sample of voters in their surveys in an effort to make sure the sample matches the demographic mix of the voters in a state or in the country, depending on the poll, or the mix of adults as recorded by the census. Pew has found that adjusting a sample on a larger number of variables can improve results, and some pollsters have begun doing so.

One adjustment that has gained traction is adjusting a sample based on “recalled vote"—whether a voter backed Trump or Biden in 2020. Some pollsters believe this helps assure that support for Trump isn’t understated, while others say it introduces new forms of error, as people don’t accurately recall or report a vote they made several years earlier.

“It’s the single easiest thing you can do to try to true-up the partisan mix of a sample," and it can be effective, said Kennedy. But the results vary by election, she said. “I would say it’s got a spotty track record."

Scott Tranter, data science director at Decision Desk HQ, a political data firm, said some of this year’s trend-breaking developments, such as the shift of Latino and Black voters toward Trump, could cause new problems. Even if pollsters detect the shift in voting preferences among those voters, determining how many will actually cast ballots remains difficult. Similarly, a surge in suburban women who support abortion rights turning out to back Harris, should that materialize, might be hard for a pollster to foresee.

As a result, Tranter says he expects polling error to be about the same magnitude as in the recent past—but he doesn’t know if it will understate support for Trump, or for the Democratic nominee, as polls did in 2012, during then-President Barack Obama’s re-election campaign.

“It’s like a roulette wheel," Tranter said. “Just because it showed up red the last two times doesn’t mean it’s going to show up red this time."

Pollsters are still debating whether Trump himself exerts a unique force on voters that pollsters haven’t yet navigated. While state polls were off-base in 2016 and national polls erred in 2020, polling was considered generally accurate in the midterm elections of 2022, when Trump wasn’t on the ballot.

Some in the survey field have thought that Trump supporters are less trustful of civic institutions and so less likely to take surveys, while others might be wary of telling pollsters their true choice of candidate. Kennedy said there are signs in some recent polls that Trump voters in fact aren’t masking their voting intention, at least this year: People registered as Republicans have been as likely as registered Democrats, if not more so, to take some surveys.

“He’s become so normalized," she said of Trump, “so I’m hoping that the measurement of people’s voting intention has become more accurate."

Write to Aaron Zitner at aaron.zitner@wsj.com