Don’t Sweat the Polls

Why the 2020 election won’t be a 2016 sequel

An illustration of a feedback loop being broken.
The Atlantic

“I want to feel hopeful about Joe Biden’s chances this year, but I just can’t,” my neighbor confessed to me, as we stood in line outside a coffee shop. What had begun as pleasant conversation—dogs, the temperature, clouds—had been pulled, through the vortex known as Late October in an Election Year, into an airing of political anxieties. “I’m still so afraid that 2016 is going to happen again and Trump is going to win,” she said.

Based on the sample size of my life, every Democrat feels this way. Yes, they’ll preface, the polls look all right for Biden. But four years ago, they looked good for Hillary Clinton too. And so, they fear, the horror film of 2016 is about to get its sequel.

There is a small chance that their fears will come true. But for the past few weeks, I’ve been stockpiling all of the quantitative reasons why the 2020 election is really, truly different from 2016, from new polling methodologies to fewer undecided voters. As always, do not allow any level of optimism (or pessimism) to guide your decision to vote. Just vote.

1. In 2016, the pollsters totally whiffed on the Great Lakes states. In 2020, they’ve changed their methods.

National polls weren’t more off in 2016 than in previous years. The problem happened at the state level. Whereas state polls underestimated Barack Obama’s support by about three points in 2012, they underestimated Trump’s support by more than five points in 2016, the largest error so far this century. The most important reason, according to a postmortem from the American Association for Public Opinion Research, was that state polls undercounted non-college-educated voters, who turned out in droves for Trump.

Here’s how that happened. Most polls are weighted surveys. That means a pollster collects a bunch of responses and then weights, or adjusts, the answers by age, gender, and political orientation so that the final count closely resembles the American electorate. For instance, if the sample is 60 percent male, the pollster will want to give the women’s responses more weight, because women actually vote more than men.

In 2016, many pollsters failed to adjust for the fact that college-educated Americans are typically more likely to respond to surveys. Another way to say this is that pollsters “under-sampled” non-college-educated voters. At the same time, the electorate split sharply along the “diploma divide” to give Trump an advantage among non-college-educated voters. In short, state pollsters made a huge, obvious mistake: Their surveys failed to account for 2016’s most important demographic phenomenon.

The good thing about huge, obvious mistakes is that they’re huge and obvious. Practically every high-quality state pollster acknowledged the non-college-educated-voter problem and committed to weighting their 2020 polls by education. The Pew Research Center now weights by education within racial groups. The Marist College and NBC News/Wall Street Journal polls now also weight by geography, in part because college-educated voters are more likely to live in urban and suburban areas.

Does this mean that the state polls in 2020 are guaranteed to be perfect? Absolutely not. In fact, they’ll almost certainly be wrong again. (They’re never exactly right.) But the polls almost certainly won’t undercount the pro-Trump non-college-educated vote by the same margin, given how many pollsters adjusted their methodologies specifically to avoid making the same mistake in consecutive presidential elections.

2. In 2016, a ton of undecided voters broke late for Trump. In 2020, most of those voters have already decided.

Two weeks before the 2016 election, FiveThirtyEights Nate Silver noted that 15 percent of voters still hadn’t made up their minds, which was roughly three times more than the number of undecideds in 2012. This statistic was the Chekhov’s gun of the last election: the ominous presence foreshadowing a final-act surprise. “One of the reasons why our models still give Trump an outside chance at victory,” Silver wrote four years ago, was that Trump could eke it out “by winning almost all of the undecided and third-party voters.”

And so he did. Exit polls and post-election surveys found that undecided voters broke strongly for the president. About one in seven voters in key swing states decided in the final week, and they broke for Trump by about 30 points in Wisconsin and 17 points in Florida and Pennsylvania, spelling disaster for Clinton.

But 2020 doesn’t have the same capacity for last-minute Democratic horror, because there aren’t nearly as many undecided voters. Fewer undecided voters means less volatility and a smaller chance of last-minute surprises that actually move votes.

The relative lack of undecided voters suggests another positive difference for Biden. In 2016, voters disliked both candidates, which is why so many were persuadable in late October. In 2020, voters dislike Trump, and actually like Biden—certainly more than the last Democratic nominee. Biden’s national polling has consistently been about four to six points higher than Clinton’s. His net favorability rating is 17 percentage points higher than Clinton’s was on Election Day. In short, many of 2016's undecideds have decided in 2020 to vote for Biden.

3. In 2016, we had the mother of all October surprises. In 2020, we have the most stable race in decades.

Biden’s lead is larger and more stable than Clinton’s lead was in 2016. In fact, by one measure, it’s more stable than any presidential nominee’s lead in more than 30 years.

In every election going back to the 1980s, the loser was, at some point, ahead in mainstream polls or in the average of polls. In the summer of 1988, Michael Dukakis led George H. W. Bush by double digits. In the spring of 1992, both Ross Perot and Bush were leading Bill Clinton. In January 1996, Bob Dole held a narrow lead over Clinton in Gallup polls. In September 2000, Al Gore surged ahead of George W. Bush. In August 2004, John Kerry led Bush. In September 2008, John McCain led Obama. In October 2012, Mitt Romney inched ahead of Obama. And in 2016, Hillary Clinton’s lead over Trump pogo-sticked all year—from up 10 in March to tied in May, to up six in June, to tied in July.

The 2020 election has been totally different. Biden, who is currently up about eight points in the FiveThirtyEight and RealClearPolitics averages, has led by at least four points since October of last year. Through everything—the primaries and the pandemic; 4 percent unemployment and 9 percent unemployment; the Democratic National Convention and the Republican National Convention—Biden has led Trump by a moderate to wide margin, and Trump’s support has never exceeded 46 percent in polling averages.

People who can remember only the 2016 election are anchoring their expectations to a historically bonkers election. The Comey letter, released on October 28, likely moved the electorate several points toward Trump. In the final weeks of the election, careful poll analysts could see Clinton’s support melting in white working-class districts. But in 2020, that just isn’t happening.

4. In 2016, district-level polls indicated a last-minute Democratic collapse. In 2020, they indicate Democratic strength.

In early November 2016, several careful polling analysts started sounding the alarm for Hillary Clinton in the upper Midwest.

Six days before the election, The Atlantic’s Ron Brownstein noted that “Clinton has not visited Wisconsin since April, and appeared just twice in Michigan from June through October.” By abandoning these key states, she was acting “like a general who has sent out a large expeditionary force and left modest forces to defend their homeland.”

Five days before the election, Dave Wasserman at the Cook Political Report tweeted a poll from upstate New York that found Trump ahead by 14 points in a district where Obama and Romney had tied four years earlier. It suggested that Clinton’s support among white working-class districts was collapsing at the worst possible time. “Five days from Election Day, it’s clear who has the momentum,” he wrote. “And it’s not Hillary Clinton. This thing is close.”

This year, those congressional polls are telling a different story. Rather than illuminating surprising weaknesses for Biden, they’re reaffirming his strengths. In some cases, the district polls are pointing to an even larger Biden blowout than the national or state surveys. Most important, Trump isn’t getting anywhere near his 2016 margins in Michigan and Pennsylvania, Wasserman observed. Four years ago, there was a quiet “Trump! Trump! Trump!” alarm going off that only congressional polling analysts could hear. This year, they’re listening closely—but no Trump alarm is sounding.

5. In 2016, there wasn’t a global pandemic. In 2020, there is a global pandemic.

It’s been four years of one “shocking but not surprising” thing after another. But this year’s October surprises have been—unshockingly, unsurprisingly—all about the plague. The president’s COVID-19 diagnosis, which overlapped with a disastrous first-debate performance, buoyed Biden at the moment when Trump needed to stage a comeback. An autumn surge of nationwide cases refocused the national media’s attention on the pandemic, which the public believes Trump has mishandled.

No one can say for sure who will win the election. If undecideds in key states such as Pennsylvania break hard for Trump, and the president benefits from another large polling error—or smallish errors in the right places—he could eke out another victory. Or he could fight Biden to a tie and then win the ensuing legal battles that discount Democratic mail ballots. But that is not the most likely outcome.

Biden holds a solid and steady lead over the incumbent president, while the pandemic is becoming more, not less, of a story as the country heads into the final days of voting. When the president complained, once again, on Monday about the news media’s pandemic obsession, his critique usefully crystallized his campaign’s biggest problem: “COVID, COVID, COVID.” The most important difference between 2016 and 2020 isn’t about polling methodology or the opposing candidate. It’s this: Four years ago, Trump ran on the vague promise of success, and this year he’s running on a clear record of failure. Judging by the polls, Americans have noticed.