A steady stream of polls in recent weeks, as well as high-profile polling averages, suggested the presidential election was more likely to be a decisive victory for Democratic nominee Joe Biden than a squeaker, with wide margins forecast in critical swing states.
Although millions of votes are still being counted, it appears many of those polls were significantly off the mark, just as they were four years ago, when President Donald Trump surprised the nation by eking out a narrow victory.
“There’s no question the polls missed (again),” Patrick Murray, director of the reputed polling group at Monmouth University, wrote in a Twitter post Wednesday morning. “But we won’t know by how much until all votes are counted (including estimates of rejected ballots). Then we will reassess.”
What could have led to such a significant error in prominent polls, such as the average by the closely followed FiveThirtyEight website that had Biden winning the popular vote by a comfortable 8.4 percentage points? FiveThirtyEight also had Biden winning Michigan by 7.9 percentage points, Wisconsin by 8.4, and Pennsylvania by 4.7. They also forecast that Biden would win — though in much tighter races — Florida, North Carolina, and Georgia.
As of Wednesday afternoon, however, Biden was leading the popular vote by about 2.2 percentage points, though that margin of victory was expected to grow as more votes are counted in large Democratic states such as California and New York. Biden led Trump in battleground states by even slimmer margins, including 0.6 percent in Wisconsin and 1.2 percent in Michigan.
With 20 percent of the vote still to be counted in Pennsylvania, Biden was trailing the president by less than 5 percent, though that margin was expected to narrow sharply as more absentee ballots were counted. Biden was also in far tighter races with Trump in North Carolina and Georgia, while losing by more than 3 percent in Florida.
“Polling within states for presidential races is notoriously difficult,” said Jesse H. Rhodes, chair of the political science department at the University of Massachusetts in Amherst.
He noted that most polls typically have margins of error of 3-5 percentage points, and that pollsters must make an array of assumptions in collecting responses, such as who to consider likely voters. Polls are also increasingly difficult to conduct in the age of cell phones, when respondents are less likely to answer calls and pollsters often rely on online panels to approximate random samples.
“Given the challenges of recruiting and weighting likely voters within states, we might expect differences between anticipated outcomes … and actual outcomes,” Rhodes said. “Given the difficulties of predicting outcomes at the state level, it is essential to take any given poll within a battleground state (particularly one that appears to give a candidate a huge advantage) with a very large grain of salt.”
Others were more cautious about finding fault in the polls before all the votes are counted.
“It would be a mistake to overgeneralize about polling misses when the election is playing out to a specific script that puts much of the Democratic vote still uncounted,” said Joshua Dyck, director of the Center for Public Opinion at the University of Massachusetts in Lowell.
For example, if Biden ultimately wins the popular vote by closer to 3 points, as seems likely, the polls would be within a margin of error “not exceptional by historic standards,” he said.
“The discussion will fall heavily on why we see highly accurate characterizations in states like New Hampshire and Arizona and larger misses in states like Florida, Texas, and Ohio,” Dyck said.
Part of the explanation for the polling error might include the historic turnout, unexpected support for Trump by Latinx voters, and the number of rejected absentee ballots, given a surge of mail-in voting.
“A full examination of what went wrong with polls this year is going to take a while,” said Doug Schwartz, director of polling at Quinnipiac University. “After the 2016 election, it took six months for the American Association for Public Opinion Research to release their findings about polling errors. I would expect a full evaluation of 2020 to take at least as long.”
Some pollsters defended their work.
David Paleologos, director of the Suffolk University Political Research Center who conducts polling for The Boston Globe, said his group got it mostly right, with seemingly accurate forecasts of the presidential outcome in New Hampshire, Arizona, Minnesota, and Florida. He acknowledged their polling appears to have been off in Pennsylvania and in the popular vote, though additional vote counts could bring those projections closer to the mark.
“This might be an election where the vote counting catches up to the polling, as opposed to an election where the exit polling catches up to the polling,” he said.
Polling errors were even more notable in some races for the Senate. FiveThirtyEight had predicted that Democrats had a 75 percent chance of winning control from Republicans, but on Wednesday afternoon the GOP looked likely to retain their majority.
Among those races with significant polling errors was in Maine, where most pollsters predicted that Democratic challenger Sara Gideon would oust Senator Susan Collins. Collins was leading by more than seven points, with more than 80 percent of the votes counted, and Gideon had conceded.
Daniel M. Shea, chair of the government department at Colby College, said his polls — which predicted Gideon would win by 3 percentage points — failed in part because there were more “split-ticket” voters than expected, or those who chose Biden and Collins.
Their polling also underestimated the large margins Collins would rack up among blue-collar workers in rural parts of the state.
“Polls capture attitudes and opinions, and election day captures behavior,” he said. “We just weren’t capturing the enthusiasm of conservative voters in the age of Trump. We’re not the only ones to do that.”