Trump is leading the polls, but there's plenty of time for Biden to catch up

Early polls are weak predictors, and they're especially noisy this year.

March 13, 2024, 5:23 PM

If the 2024 general election were held tomorrow, President Joe Biden would probably lose to former President Donald Trump. That's because, by 538's averaging, Biden trails in every major swing state — not to mention in national polls, too. Things only get worse for him the closer you look: The incumbent president is currently viewed more unfavorably than his predecessor, for instance, and polls show few voters are giving Biden credit for an improving economy. They trust Trump more to handle several issues Americans rate as the most important to their vote, including immigration, jobs/the economy and foreign policy. Voters are likelier to say conditions were better under the previous president and that the current executive has made their lives worse.

That is a tough position for any politician to start their general election campaign in. But so far, the Biden campaign's strategy has been to publicly dismiss the polls rather than confront their bad numbers head-on. For instance, at the opening ceremony of a campaign field office in Manchester, New Hampshire, Biden disparaged polls as inaccurate and stuck using old-fashioned technology.

Biden is wrong about that. Despite the challenges facing the industry, polls remain as accurate as ever. But that doesn't mean that polls taken eight months before Election Day are reliable predictors of the final outcome. My analysis of polling from past elections reveals that there is plenty of precedent for a candidate like Biden to gain enough ground to win the election. Moreover, I find there is more uncertainty in the polls now than in previous cycles, further casting doubt on the early gloomy prognostication about Biden's odds.

How predictive are early polls?

First, I wanted to know how much I could trust these early horse-race polls, so I booted up 538's spreadsheet of historical polling data. This dataset has national polls going back to 1944 and state-level polls going back to the 1950s (though state polling really ramps up in the '70s). We can compare the averages of past polls taken around this time of year with final election results to get an informed guess at how much polls can change over the course of a year.

There have been, in technical speak, some whoppers of early poll misses. At this point (237 days until the election) in the 1980 election cycle, for instance, incumbent Democratic President Jimmy Carter was ahead of Republican Ronald Reagan by 14 percentage points. But on Election Day, Carter actually lost by 10 points. That's a 24-point error.

Such a miss is not uncommon in the history of pre-election public opinion polls. In 1992, the early polls were off by 19 points; in 1972, by 12; and in 2000, by 8. Over our whole dataset, polls from mid-March have missed the final outcome of the presidential national popular vote by an average of about 8 points.

It is not surprising that early polls have a high amount of uncertainty predicting eventual election results. In many presidential election years, parties had not even selected their nominees by mid-March, making it difficult for voters to think about a choice they would make about hypothetical candidates eight months in the future. Many voters also simply aren't paying attention to the election yet.

One wrinkle in this story, however, is political polarization. As voters have become more partisan, their votes have become more predictable even from afar. This may make early polls more reliable too. At this point in the 2020 election cycle, polls were off by less than 1 point from Biden's final vote margin over Trump. And in 2012, polls this early overestimated President Barack Obama's margin over Mitt Romney by just 3.5 points. On the other hand, at this point in 2016, Hillary Clinton was up by 8.5 points in the polls, making for a 6.5-point miss.

So while the average error of early polls has dropped over time, some big misses still happen. Those misses also carry more weight since our polarized elections are decided by much narrower margins. A miss of 5 points in a 52-48 election, for example, is just as consequential as a miss of 10 points in a 55-45 election.

Our polling average is exhibiting higher uncertainty than normal

At his campaign stop in New Hampshire on Monday, Biden also said, "polling has kind of changed a lot too. It's not nearly as accurate. It's not nearly as capable as it was before because you've got to make 6 zillion calls to get one person on their cell phone."

While these remarks overstate the challenge for pollsters — many polls are conducted online now, for example — Biden is right that it has gotten harder to pull off a good poll over time. Response rates are at record lows, which drives up costs and requires pollsters to make more statistical adjustments to get their samples to more closely match the demographics of the U.S. population.

And even with all those adjustments, pollsters still may not arrive at a representative sample of the population. It's possible that they adjust for the wrong factors, for instance, and end up with a sample that is too Republican- or Democratic-leaning anyway. That's what happened in the 2020 election: Even polls that had the correct proportion of Republicans and Democrats, or 2016 Biden and Trump voters, ended up underestimating the number of 2020 Trump supporters within those buckets.

Pollsters call this extra error "non-sampling" error — that is, error that goes beyond the uncertainty inherent in taking a random sample of the population. The total amount of this extra error is usually about 3 points beyond the published margin of sampling error for a poll. That means the total uncertainty for a poll of about 1,000 people could be 6 or 7 points.

But in an early version of our general election polling average, I'm finding there is about 4.3 points of extra non-sampling error in horse-race polls today. And that's for the vote share for either party — it's about double that for the margin between the candidates. This means it would not be unusual at all to get a poll one day with Biden leading by 5 and another poll the next with Trump up 5.

While all this extra error means individual polls are noisy, the averages — contrary to Biden's suggestion — are not inherently biased from extra noise. And the averages paint the most damning picture for him. Biden is currently polling behind Trump in all the key swing states. That includes a nearly 10-point deficit in Florida, 5 points in Arizona, 4 points in Michigan and Wisconsin and about half a point in Pennsylvania. The president leads by small margins in Minnesota, Virginia and New Hampshire.

Assuming these averages are spot-on, if the election were held today, Trump would win the Electoral College handily with 312 electoral votes to Biden's 226.

Betting on bias

Of course, the election will not be held today, and it's foolish to assume the polls will nail the final results exactly. Every year, the polls are off by a few points. And usually, the polls across different states and regions all miss in roughly the same direction.* So it's likely that the current polls are either overestimating or underestimating Biden by a few points too.

Some of the president's detractors have pointed out that polls in the last two presidential elections have underestimated support for Trump. They suggest polls are likely to do the same again this year. But polls have not traditionally overestimated the same party reliably over time. In 1976, for example, polls underestimated Carter's margin of victory over President Gerald Ford. But in 1980 they overestimated Carter by 3.2 points — an error rivaling that of the 2016 election. And after slightly underestimating Obama in 2012, polls famously overestimated Clinton in 2016. Trump again beat expectations in 2020 by an even bigger margin.

And while polls do not usually overestimate the same party three cycles in a row, it has happened before. By dismissing the current polls out of hand, Biden and his campaign are turning a blind eye (at least publicly) to the possibility that polls will underestimate support for Trump again this year. In a recent interview with the New Yorker's Evan Osnos, a senior Biden staffer said, "Polling is broken. You can't figure out how to get someone on the phone. I think the only person who calls me on my landline is Joe Biden." But polling is not "broken," and most polls are not even conducted by landline anymore. In the 685 2024 general election polls 538 has gathered so far, at least 593 were conducted at least partly over the internet, seven more (for a total of 600) relied on text messages and the rest were completed by a mix of live phone calls to cell phones and landlines (83 polls) or automated calls to landlines alone (two polls).

As history has shown, betting on bias in the polls is not usually a safe bet. Neither is counting on the polls to move in your direction. The Biden campaign, of course, knows this, and it has eight months to change the minds of the American people. Given the empirical uncertainty in early general election polling, and the tendency for bias to jump around from cycle to cycle, there is a distinct chance they do so. But before you can dig yourself out of a hole, first you have to acknowledge that you're in one.


*Working on 538's upcoming election forecast, I've found that the best statistical assumption is that polls share about 30 percent of the error across the country, about 40 percent across states in their region, and the remaining 30 percent is specific to each individual state.