Going into Election Day, forecasters predicted a Joe Biden victory, but cautioned that Donald Trump still had a chance. Experts warned that the process of counting mail-in votes could take days or even weeks to finish, and that early vote totals in the Rust Belt states might start off looking heavily pro-Trump but then shift blue as absentee ballots were counted. We heard that Trump planned to declare himself the winner on election night and call for ballot counting to halt in states where he was in the lead.
On Wednesday afternoon, as I write this column, Biden is on track for a close victory, but Trump still has a chance. The results hinge on a handful of swing states that may not finish counting votes until the end of the week; in the Rust Belt, Trump’s early leads look to be morphing into narrow Biden victories as absentee ballots get counted. Meanwhile, Trump has indeed declared that “as far as I’m concerned, we already have won it.” In other words, everything is turning out just as we’d been told. So why does it all feel so surprising?
The answer begins in Florida. Heading into Tuesday, The New York Times announced it would be reviving its notorious election needle, but only for the three states that were expected to count most of their votes quickly and report detailed statistics on who voted where, and by what method: Florida, North Carolina, and Georgia. FiveThirtyEight’s poll averages gave Biden an advantage in those states, in percentage points, of 2.5, 1.8, and 1.2, respectively. All practically toss-ups, but heading into the evening it seemed reasonable to guess that Biden would win at least one of these key states, and that we’d know it before bedtime.
This did not happen. Florida, where Trump won by 1.2 points in 2016, was the first to report results—and they were stunning. The Times needle swung dramatically to the right as results came in, indicating a near-guaranteed Trump victory. While Biden did well in some parts of the state, Trump increased his 2016 margins in the heavily Cuban stronghold of Miami-Dade County. The Times started predicting a 4-point Trump victory; as of now, with most of the votes counted, that lead is closer to 3 percent, suggesting that the FiveThirtyEight average had been off by 5.5 points.
The other Times needles, feeding on Florida’s data, reacted accordingly, predicting similarly decisive wins for Trump in Georgia and North Carolina. But as the night went on, it became clear that the results in Florida—perhaps the nation’s most demographically and politically idiosyncratic swing state—both over- and under-predicted the scope of the national polling error. In Georgia and North Carolina, as more votes got counted, the races tightened; as of now, Trump is on track to win North Carolina by just over 1 percentage point, while in Georgia, a batch of outstanding ballots from the Atlanta area could actually deliver the state narrowly to Biden.
>
Elsewhere in the country, however, the polls downplayed Trump’s support even more flagrantly than they did in Florida. Biden led the averages by 8.4 points in Wisconsin and 7.9 points in Michigan. As of now, with all votes counted in Wisconsin and nearly all in Michigan, he is up in those states by just 0.6 and 1.2 points, respectively. In Ohio, which Trump won easily last time, polls showed Biden within less than a point. He’s currently down 8. These errors are even larger than the ones from 2016, as you can see if you scroll down a bit here.
Polls are not and never have been perfect, and state polls are typically worse than national ones because it’s harder to build representative samples with smaller subgroups. This year’s apparent errors may also shrink once all the votes are counted. “Some of these outcomes are still moving targets,” said Courtney Kennedy, director of survey research at Pew Research Center. “Michigan, Pennsylvania, Arizona, Nevada—I think it’s important that we pump the brakes for the next few days and let that play out, because it’s very likely that those vote outcomes are going to shift a little bit more toward where the polling was.” Nate Silver of FiveThirtyEight predicts that, once all the votes have been counted, this year’s national polling error will settle around 3 percent, with some states ending up better than predicted for Trump and others worse—just like 2016.
But performing “only” as badly as in 2016 is hardly a victory for an industry that has spent the past four years trying to correct for this exact type of mistake. The consensus on 2016 was that pollsters erred because they hadn’t weighted their samples by education level. That supposedly explained how they ended up missing a huge split between white voters with and without college degrees. This time around, most good pollsters thought they would account for the missing Trump supporters by making sure to properly weight non-college white voters.
So what went wrong this time? One problem, as the Miami-Dade case suggests, is that polls failed to predict Trump’s improvement among Hispanic voters, most notably in Florida and Texas and, perhaps, Nevada. This is more an unforeseen 2020 issue than a rerun of 2016. Solomon Messing, chief scientist at the liberal digital organization ACRONYM, thinks it might have more to do with Covid-era campaign choices than polling issues per se, as Biden’s campaign was much less inclined than Trump’s to do in-person canvassing.
“In-person voter registration was almost non-existent in populations that traditionally vote for Democrats, not so for Republicans,” he said via email. Republicans also put more effort into door-to-door voter persuasion efforts, while Democrats mainly relied on digital contact.
In the Midwest, however, the problem appears to be more similar to the one from the last election: Polls systematically undercounted Trump supporters. Weighting by education simply didn’t work.
Brian Schaffner, a political scientist and pollster at Tufts, suspects the problem is that the more someone supports Trump, the more that person distrusts the media and academia and the less likely they are to respond to a poll. This is not, by the way, the same thing as the “shy Trump voter” hypothesis, which suggests that Trump supporters lie to pollsters about their voting intentions. There’s no evidence that that happens. Rather, the issue would be that Trump supporters are declining to talk to pollsters in the first place.
>
Schaffner tried to avoid this problem in his own polls by weighting samples according to people’s actual vote history in 2016. But he now suspects that the distrust among Trump’s strongest supporters led him to oversample wavering Trump fans and undersample the diehards.
“A lot of the polls I’ve been working on, we’ve been showing Trump losing something like 8 percent of the people who voted for him in 2016, with them basically saying they were going to vote for Biden,” he said. “But if you look at the results, it’s hard to imagine that’s actually what happened. Basically, we had a group of Trump voters that were more likely to defect than the actual population of Trump voters were. If we’re overstating how many of those Trump supporters are switching, that’s going to have a big effect on the final margin in polls.”
This poses a rather sticky challenge that isn’t going away any time soon. It’s also not limited to the US. Karlyn Bowman, an expert on public opinion research at the American Enterprise Institute, pointed out that polls underestimated support for Brexit in the United Kingdom and populist candidates in Australia. If it turns out that support for right-wing populism is independently correlated with survey response rates, then it could be impossible for pollsters to get an accurate count no matter how they weight the demographics.
All that said, please put the pitchforks down and stop looking up Nate Silver’s home address. Election forecasters warned us that the election could be tight. They reminded us that polling errors happen, and noted that Biden’s lead in the polls was so comfortable, it could conceivably withstand even a big one. This appears to be exactly what happened.
We know at this point that polls consistently underrate support for Trump and his brand of belligerent, mendacious, cultural-grievance-based politics. It would be nice to know why that happens. But by far the more pressing question is how that support got to be so high to begin with.