Pollsters Changed How They Do Things After Messing Up In 2016. They Still Screwed Up In 2020.

Here’s what we know so far.

An illustration depicting polling arrows which have missed an archery target

There have been a lot of voting surprises this week — enough for some pundits to declare that the polling industry is a “wreck” that should be “blown up.” Some things definitely went wrong, but what exactly? Let’s try to make sense of it.

How bad was polling this year?

It wasn’t great. Much like in 2016, this year’s polls consistently and substantially underestimated support for Donald Trump. In Florida, for instance, most pollsters thought Joe Biden held a small but persistent advantage over Trump. But the president appears to have won Florida by more than 3 percentage points — an even wider margin than he had four years ago.

In her US Senate race, Republican Susan Collins won reelection by a substantial margin, even as the polls consistently favored her challenger, Sara Gideon.

It’s too early to know exactly how off polling has been in 2020. If it’s as bad as some say, the consequences could be dire. “The political polling profession is done,” longtime Republican pollster Frank Luntz told Axios. “It is devastating for my industry.”

But some of the numbers that looked terrible for pollsters on election night have gradually improved since then. In Pennsylvania, for example, Biden was losing to Trump by nearly 8 points early Wednesday afternoon, despite polls that had given him a clear edge. But Democrats disproportionately voted via mail-in ballots, which the state was slower to tally. By early Thursday afternoon, Biden was losing by less than 2 points — with hundreds of thousands of votes left to be counted. (Update: By Friday morning, Biden had taken the lead.)

One bright spot in the polling is Georgia, which appears will be decided by the thinnest of margins. Polls there had flipped back and forth between tiny leads for both candidates.

“In the final accounting,” FiveThirtyEight’s Nate Silver predicted Wednesday evening, “the polls will have done mediocrely, but not terribly.”

What the heck happened in Florida?

Florida provided the first big surprise of the night, with Decision Desk HQ calling the state in the president’s favor just after 8 p.m. ET. As of Thursday afternoon, with the vast majority of votes counted, it appears Trump outperformed polls by about 6%.

Miami-Dade County was the biggest disappointment for the Biden camp. In 2016, Hillary Clinton beat Trump by 29 points in that large, diverse Democratic stronghold. This year, however, Biden still won the county but only took it by roughly 7 points.

Election analysts have pegged this massive shift to the high turnout for Latino voters who broke for Trump in large numbers. Polls have a consistently hard time properly accounting for Florida’s unique Latino population.

Pollsters typically interview a sample of registered or likely voters in a state and then carefully “weight” the results, assigning greater or smaller statistical value to different interviewees to better match the population.

Lee M. Miringoff, director of the Marist College Institute for Public Opinion, said that while polls try to properly weight the Latino demographic, they won’t always adjust for more specific communities, such as Cuban Americans, who have historically leaned Republican. Analysts say Trump’s frequent attempts to color Biden as a socialist may have resonated with Cuban American voters and others in South Florida.

Miringoff said he was aware of Biden’s weaknesses among Florida’s Latino voters. In early September, a Marist poll showed Biden had 46% support among Florida Latinos to Trump’s 50% — a community that, at least according to exit polls in 2016, Clinton won heavily.

Did pollsters mess up in the Upper Midwest again?

Yes. In 2016, the polls were infamously wrong about the so-called blue wall states in the Upper Midwest: Minnesota, Wisconsin, and Michigan. The polls had expected Clinton to win each of these traditional Democratic strongholds by more than 4 percentage points, but they became extremely competitive. Clinton held on to Minnesota by just 1.5 percentage points and lost both Wisconsin and Michigan.

Pollsters appear to have misread this region again. Wisconsin and Michigan have, as predicted, both been called for Biden, but the numbers were way off from the polls suggested — perhaps by an even larger margin this time.

FiveThirtyEight’s average of polls showed Biden up by 8 points in both Wisconsin and Michigan. Both those races ended up being nail-biters — so close in Wisconsin, in fact, that the Trump campaign has said it wants to request a recount.

After 2016, pollsters studied what had gone wrong. Their biggest takeaway: They had failed to account properly for voters’ education levels, particularly in the Midwest. Voters with college degrees, the postmortems found, were more likely to support Clinton and also more likely to respond to polls in the first place. But the pollsters hadn’t factored that phenomenon into their numbers. As a result, the polls looked better for Clinton than they should have. Since then, most pollsters have started adjusting responses for education, similar to how they adjust based on a respondent’s ethnicity.

The 2020 results, however, suggest that this fix wasn’t all that was needed.

“I don't think it was wrongheaded,” said Sean Trende, an elections analyst for RealClearPolitics. “I think you have to weight by education. I just think people are a little too flippant assuming that was what the whole thing was.”

Is the problem just “shy Trumpers”?

After 2016, the “shy Trump voter” emerged as a popular explanation of why pollsters had overestimated Clinton’s numbers. The idea was that some voters might have been reluctant to admit that they were supporting a controversial candidate. But there’s not much evidence to support the theory.

An alternative theory is that pollsters simply aren’t reaching enough Trump voters.

Trende said partisan attitudes toward polls and the media play into the dynamic; he thinks the most ardent Trump supporters won't even talk to pollsters, skewing the data toward Republican respondents who are more moderate. “If there's a segment of population that just doesn't want to talk to you, there's nothing you can do about that,” Trende said.

Miringoff said this was a cop-out. "That sounds like, my god, there's this army of Trump people that aren't talking to us," he said. “What proportion of them would have to be doing that?”

So, what happens next?

While the debate over the value of polling has already begun, experts are urging restraint while the final votes are being tallied.

“Anyone who feels let down or whiplashed should take a deep breath,” columnist David Byler wrote for the Washington Post.

A full autopsy of 2020’s polling won’t be possible until the complete results come in. Pollsters are eager to see exactly where the discrepancies occurred and whether they fall within the polls’ margins of error.

But in the end, any such autopsy may matter less than public perception. “We just have to understand the limitations of these things,” Miringoff said. “They look more precise than they are.”

Skip to footer