When Jack Dozier ’27 was working with the Democratic Coordinated Campaign in Virginia this summer, he spent much of his lunch break looking at polls. He had plenty to look at—the presidential debate between Biden and Trump had taken place just weeks before. But Dozier’s boss didn’t approve of his lunch break pastime.
“My boss would always say, ‘They’re not right, right now—these are incorrect polls. It’s early and people are going to change their minds. It’s a bunch of numbers that aren’t going to change the outcome that much until we get a lot closer.”
The 2024 election, predicted by polls to be a toss-up with Harris leading the popular vote within the margin of error, ended with Trump winning both the electoral college and the popular vote on election night. The results, which defied public expectations of a closer contest with prolonged vote counts and potential recounts in swing states, brought the issue of polling inaccuracies to the forefront of post-election discussion. This issue has deep roots and a turbulent recent history.
***
While Dozier feels polls are a good way to gain a better perspective of “what is going on in the world,” he believes it’s important not to rely too much on them, especially after the 2016 election. In 2016, polls projected that Hillary Clinton would win the race, only for Trump to take the election with 304 electoral votes. While it caused some to lose faith in polling entirely, others have alternative explanations for the inaccurate predictions.
“The issue in 2016 was that there was a very big educational divide in how people voted,” said Kevin DeLuca, an assistant professor of political science at Yale University. “If you were a white voter with a college degree, you were much more likely to support Hillary Clinton in 2016, and white voters without a college degree were much less likely. So if you weight on education, especially for white voters, that fixed the polling errors in 2016.”
Matt Taglia, the senior director of Emerson Polling, concurred but also saw 2016 as a coverage issue.
“In some states like Wisconsin and Michigan, there just wasn’t much polling at all,” Taglia said. “Most people knew they would be close, and they were battleground states. They weren’t on anybody’s radar because there were only a few released polls. It was more of a coverage issue––coverage in terms of just not enough polling and actual media coverage as well.”
Milan Singh ’26 is the director of the Yale Youth Poll, which surveyed 5,500 registered voters this October with a focus on young voters. He thought the polls in 2016, which underestimated Trump by approximately 1.3%, were “wrong by the normal amount.” The RealClearPolitics average predicted Clinton to win the popular vote by 3.2%, and the final margin was 2.1%. However, he said, “how the error was interpreted was out of proportion.”
“It just happened to be that in that case, that one point polling error was enough to put Trump over the top in key swing states,” he said. “People were so confident that Trump would lose that it came as something of a shocker.”
But the story did not end there. When pollsters came into the 2020 election with renewed strategies from 2016, their polls once again fell short, overestimating Biden’s performance over Trump’s. While Biden won the election, the race was much closer than the polls expected. In fact, polls in 2020 showed the largest error since the 1980 election, when Ronald Reagan defeated incumbent Jimmy Carter.
Why another failure? There are a lot of explanations. Singh described David Shor’s theory that “liberals listened to lockdown orders, stayed at home, and were more likely to answer random phone calls.” Thus, Democratic voters were overrepresented in the samples. Singh finds this theory most convincing.
DeLuca, meanwhile, thinks the polling miss in 2020 remains a mystery.
“Even if you control for all these demographic characteristics––even if white non-college voters who answered surveys were more likely to support Biden than white non-college voters who didn’t answer surveys––you still miss a bunch of people,” he said. “A lot of people think 2020 was just such a weird year because there was a pandemic and the election was about Covid.”
After the 2022 midterm election, pollsters had reason to feel more confident. According to 538, pollsters predicted the election more accurately than any election since 1998. This result seemed to support the “shy Trump supporter” theory, which posits that polls consistently underestimate Trump in presidential elections because a significant number of Trump voters are unwilling to admit their voting plans to pollsters. This theory, unique to elections where Trump is on the ticket, could explain why the midterm election polls were more accurate.
Both Taglia and DeLuca reject this theory.
“I think a lot of people had the wrong turnout model,” Taglia said about the 2016 and 2020 elections. “They didn’t expect that all of these white voters without college degrees would show up because they’re low-propensity [voters.]”
Deluca added that it would have been ungrounded to expect Trump to outperform the polls again in 2024.
Pollsters made their best efforts to account for Trump’s electoral power before the 2024 election. In fact, some experts like DeLuca wondered “whether pollsters have overcompensated for missing Trump voters in the past,” causing polls to be biased towards Trump.
For example, in an effort to mitigate their past underestimation of Trump’s performance, many pollsters, like Emerson, started weighting on recall vote in addition to factors like age, race, and gender. This means pollsters consider how voters say they voted in past elections when they tailor their numbers. This method is generally avoided due to voters’ tendencies to forget who they voted for in past elections and their likelihood of saying that they voted for the winners. Thus, the polling results are more likely to shift towards the results of the past election and fail to consider “real changes that are happening in the new election,” according to DeLuca. Some pollsters also started weighting by party identification in an attempt to avoid undersampling Republicans.
“If you’re worried that you’re underestimating Trump voters and you need to weight to make your poll a little more Trump-friendly, then you weight by party ID,” DeLuca said. “Usually, you’re not supposed to do this because it makes your poll results less accurate.”
Despite pollsters’ efforts to accurately sample voters’ support for Trump, even through using unconventional methods, they again underestimated his support in 2024.
“We can’t discount the possibility that Trump being on the ballot makes polling a little bit trickier,” Singh said. “Three times in a row is hard to rule out. At the same time, three times in a row is not so many instances that we can definitely say [something.] I think it’s important to separate 2020 from 2016 and 2024 just because of Covid.”
Taglia thinks that in any case, Trump-specific theories are not likely to be relevant for future elections, and pollsters would have to adjust again to a world without Trump.
“Going forward, since Trump himself has said he will not run for elected office again after this, everything will change again,” he said. “Some of these low-propensity voters who he activated aren’t necessarily going to turn out for these other Republican candidates. We see other Republican candidates and Senate campaigns underperform Trump even when he is on the ballot. So it’s going to be a problem for Republicans. It’s going to be a problem for all of polling going forward, too.”
***
Polling technology has evolved dramatically in recent years. According to Taglia, since young people are less likely to answer the phone, pollsters have become more reliant on online panels, in which voters sign up to share their opinions, sometimes with financial incentives.
“There are concerns over whether these folks are representative, but frankly, it’s the only way that we’re going to reach people under the age of 30,” Taglia said.
As the use of landline phones has declined, many pollsters have shifted from random-digit dialing to email and text messaging to collect responses. Taglia said Emerson Polling sent two to three million text messages over the course of this election cycle. Dozier, who is also the deputy director of Yale Youth Poll, said that pollsters are still trying to “find the right fit” with these new methods.
“It takes a long time,” he said. “I’m hopeful that we’re almost there because we’re getting closer. It’s certainly a lot better than it was in 2016.”
Determining which voters will turn out is a consistent challenge for pollsters. According to DeLuca, pollsters strive to only include answers from respondents who actually cast a vote on election day.
“It’s very well known that 18-year-olds, even if they say they’re going to vote, a lot of the time, don’t vote. There’s just a pattern you see,” he said. “And if you’re a senior citizen and you say you’re going to vote, you’re very likely to actually vote.”
Most pollsters directly ask respondents whether they plan to vote. Taglia said that before Labor Day, Emerson Polling keeps all registered voters in their sample. After Labor Day, when people are more likely to have made up their minds on their vote, they narrow down their sample.
“We take their word for it, right?” he said. “We ask them how likely they are to turn out. If they’re very likely or they’ve already voted, we keep them. If they say anything other than very likely or that they’ve already voted, then they’re terminated.”
The emergence of partisan pollsters drew concern about the accuracy of polling this election. However, DeLuca said that aggregators like 538, Nate Silver, and Split-ticket take pollsters’ biases into consideration when creating polling averages. Taglia thinks these aggregators need to do a better job of removing data from “suspect” pollsters.
“There are some [pollsters] where they definitely are suspect, and the aggregators aren’t doing anything,” he said. “Not every poll deserves a public release and publicity. There are some low-quality polls out there that shouldn’t be a part of the public’s conversation.”
According to DeLuca, pollsters with stronger accuracy in the past and more transparent methodology receive higher weight in these averages, increasing the reliability of the aggregators. But even the most reputable pollsters are sometimes wide off the mark.
One notable outlier poll in the 2024 election was “the Iowa poll” conducted by J. Ann Selzer, one of the most highly rated pollsters in the country. Released four days before the election, the poll found that Harris was leading by 3% among likely voters in Iowa, a reliably Republican state. Although recognized as an outlier––Trump did end up carrying the state by 13%––the poll increased hopes for a Harris victory among Democratic circles, not least because of Selzer’s reputation. Singh said Selzer’s poll, which showed Harris’s strength with Midwestern white voters, particularly seniors and women, was one of the reasons why he thought Harris would win.
Zach Pan, the treasurer of Yale College Democrats who worked in a congressional campaign in Iowa when he was 17, celebrated with his friends when the Selzer poll came out.
“Ann Selzer is a legend,” he said. “I was texting my friends from the campaign. They were celebrating. We thought we’d win Iowa or at least come very close in Iowa because we couldn’t fathom Selzer being that wrong.”
***
Despite all of its imperfections, the demand for polling persists. According to DeLuca and Taglia, polling is not a tool precise enough to offer certainty about who will win an election, especially a close race like 2024. However, movements in these numbers throughout the election cycle demonstrate changes in voters’ minds. Significant fluctuations during the 2024 race included Biden’s diminished support after his debate with Trump and Harris’s rise to a sizable, three-point lead over Trump after entering the race in August.
“You’ll never know [until election night] if you’re underestimating Trump supporters the whole time, but the relative movement would still give an idea of what’s going on,” DeLuca said.
Singh also added that quantitative tools like polls, and qualitative tools like focus groups work together to create a more complete picture of public opinion. While polls can reach more people, there’s a limit to how much nuance they can pick up from voters. Conversely, while focus groups can gain deeper insight into voters, the media can only reach a certain number of people due to limited time and resources.
“Think of [polls] as those old maps of America where they haven’t fully filled out the West,” he said. “Then you have the in-person conversations and you fill in the details.”
DeLuca said polling can be more useful outside the context of horse races, such as when determining which policies are popular with voters and which candidate voters trust more with specific issues like the economy, immigration, and abortion. These results are essential for campaigns in choosing their strategy as they target different groups of voters.
“I’m sure polling is part of why Trump changed his tune on abortion,” Singh said, “because he knows that his previous stance, based on polls, wasn’t really doing well.”
Dozier added that Trump won the 2024 race partly because he focused on the economy, which polls consistently indicated was voters’ top concern.
Campaigns use polls to mobilize volunteers as well. According to DeLuca, when campaigns are slightly behind in their internal polls, they might release them to supporters to ask for donations and support. When they are ahead in internal polls, they might release them to show they have momentum and inspire optimism.
“A lot of this is fundraising-based or trying to build a narrative about the campaign,” he said.
As an organizer, Dozier empathizes with this galvanizing effect of poll numbers, regardless of whether his candidate is leading or lagging.
“When it would show Harris was ahead, I would say, ‘We still got to keep working,’” he said. “When it started showing a little bit more worrying things, I said, ‘That’s more reason to get on the doors.’”
Polls also impact voter turnout, especially in high-stakes races.
“I think it’s a good thing for the public to know that they should go out and vote, especially in these states where it can turn the election,” Taglia said. “I wish we had more local polling so the folks could see how much it matters to go out and vote on the local level, even if you’re in a non-competitive state.”
Additionally, Taglia said it’s important that polls set reasonable expectations for elections, “especially after 2020 and the controversies surrounding the vote counting.” In 2020, Trump claimed multiple cases of voter fraud in battleground states, which have been deemed baseless by courts, election officials, and governors.
“Polls are not perfect, but they’re almost always better than your gut instinct,” Singh said.
The entertainment value of polling is hard to ignore as well, especially for the so-called “political junkie” class. DeLuca compared it to sports fans watching every game, following their players and their stats, and talking about sports with other fans. The media has responded to that demand by creating forecast and prediction models, with each news outlet having their own apparatus.
“I think the appeal of polling is that, in some ways, it’s like a crystal ball,” Pan said. “And everyone wants to know what’s going to happen in the future. So there’s a psychological element that attracts people to polling. That being said, I think all polls have to be taken with a grain of salt, especially when you’re a month out from the election. You’re given a snapshot of the present, not a crystal ball of the future.”
The media also frequently uses polls for eye-catching headlines and stories. Over the summer, Dozier saw an outlier poll about a local race in Virginia that seemed biased towards a Republican candidate. When he printed out the poll and read it page to page, he soon realized the sample was primarily Republican, despite Virginia being a reliably blue state, and it was unweighted.
“It did not hold water. It had crazy samples,” he said. “In my opinion, it was not the strongest poll, and headlines just ran with it because it caught [viewers’] attention.”
***
According to 538, the average poll conducted throughout the last three weeks before the 2024 election missed the margin of the election by 2.94%. By comparison, polls for the 2016 and 2020 elections had an average error close to 4.7%. In addition, the polls this year essentially projected a 50-50 race in the electoral college, with Nate Silver’s model predicting a Harris victory in 40,012 simulations out of 80,000, and the final popular vote polling average slightly favored Harris, who led by 48.6% to 47.6%. However, Trump won the electoral college with 312 votes and led the popular vote count by 50.1% to 48.3%, outperforming polling averages.
“One takeaway from 2024 is that the polling error was roughly normal size,” Singh said, “but it was uniform in that it was a consistent, small underestimation of Trump, and that mattered.”
While Dozier, who expected the race to go either way, recognized the critical ramifications of polling error, he doesn’t think it’s time to abandon trust in polling. He hopes polling will keep evolving to better reflect the electorate.
“Every year, we’ll have more and more changes, and eventually we’ll hit that perfect model, but then the world will change as well,” he said. “There will be that moment where they line up, and at that moment, you need to figure out how to keep it on track, how to make them parallel, how to make them move the same way, because polling is such a crucial tool.”
Whether polls are used to set expectations, guide campaigns, mobilize volunteers, or simply satisfy entertainment value, they occupy a unique place in the American political landscape. Pollsters continue to improve their techniques, reflecting on past mistakes and keeping up with technological change. For the public, there’s nothing close to a replacement. In a rapidly changing world, the art of polling serves both as a mirror and a map, reflecting society’s values while guiding its direction. And it’s not going anywhere.