Most Polls Underestimated Trump. Here’s Why Ours Didn’t
CNN’s senior data reporter, Harry Enten, on Tuesday observed that since 1972 no party’s presidential candidates have outperformed opinion polls more than twice consecutively in key battleground states. Donald Trump broke that streak. The polling industry has underestimated his performance three times in a row.
Polls missed Mr. Trump’s vote share by 4.3 percentage points nationally in 2016. The 2020 election provided a fig leaf, as surveys correctly predicted that Joe Biden was ahead—but they fell short of Mr. Trump’s total by 3.4 points on average. This year, the fig leaf blew away. Polls’ underestimation of Mr. Trump will be close to 2016 levels. The former president has proved to be kryptonite for pollsters again, with Trump voters underrepresented in their samples.
But you likely didn’t hear about underrepresentation during the campaign. Some political pundits claimed—based in part on Democratic overperformance in 2022— that the industry had overcorrected for its past errors and was overestimating Mr. Trump’s support.
While most of the polling industry was wrong, we got it right. Our final election forecast was the most accurate of all the major pollsters and aggregators. We forecast the widest Trump margin in the Electoral College and were one of only two in a list of 10 major pollsters that predicted a Trump victory. Our final national poll, on Nov. 1, had Mr. Trump leading Kamala Harris 49% to 46%. It appears that in the final reckoning, Mr. Trump will finish with about 50% of the vote.
How did we avoid the trap that most of our colleagues fell into repeatedly? We think there are three main reasons other pollsters keep missing the mark: Mr. Trump turns out low-propensity voters, he has a unique appeal to harder-to-reach nonwhite voters, and voters who favor him often don’t like to admit that they do.
Phone polling alone, which in 2024 has a 2% response rate, isn’t going to reach low-propensity voters or politically disengaged nonwhite men. So relying solely on randomdigit dialing, as Ann Selzer did in Iowa, is wrongheaded. Our statistical analysis shows that the group most likely to talk to pollsters on the phone for 20 minutes during the workday is exactly the group Ms. Selzer overestimated in a poll that showed Ms. Harris ahead by 3 points in the Hawkeye State—liberal older white women. Some wrongly took their concerns, including about abortion, to be the story of the election.
Online-only polls fall into a different trap. They pick up younger voters, more engaged voters, and those who work from home. Those groups are more likely to be Democrats. That’s not the way to reach enough people who don’t trust polls or bluecollar Trump voters who don’t al--ways have the time to talk on the phone with pollsters.
To survey these people, we used a mixed-methods approach including phone calls, text messages, and online and in-app polling. In-app polling surveys voters as they shop or play games on their phone. It meets
voters on their own terms while they’re going about their daily lives. Those we reach using the in-app method are more likely to be younger nonwhite men. This approach allowed us to predict correctly that 1 in 3 nonwhite voters would back Mr. Trump.
Our second tactic was to keep those who hadn’t voted before in the sample. While other firms scaled this group down, we kept open the possibility that they would turn out to vote. In our final sample, 17% didn’t vote in 2020—and among that group, Mr. Trump had a 20-point lead.
Polling necessarily involves subjective judgments, such as the decision to leave nonvoters as they were in our sample—without weighting them up or down—or to treat Ms. Harris as if she wasn’t an incumbent in our model. Every pollster has to make these calls. But ours were informed by the other side of our business: in-person focus groups and structured 90-minute interviews with voters.
In these interviews, we got beneath the surface and past the socially desirable face a voter may present when clicking an online form or answering a question over the phone. In the first 10 minutes of our conversations with Hispanic men, they often said they were backing Ms. Harris. By 90 minutes in, as we built trust, they admitted they were going to back Mr. Trump and shared more of their worldview: frustrations about the border, the economy and the Democrats’ abandoning traditional family values. We spotted that despite the assertions of the Harris campaign, undecided voters weren’t breaking late in any particular direction, and that many former nonvoters in rural areas were going to cast their ballots for Mr. Trump.
Polling is broken, but it doesn’t have to be. The industry needs to reinvent itself and approach voters on their own terms using the same communication methods they use in 2024, not those of 2000. For too long, the industry has taken voters as they have found them rather than seeking to really understand how they think.
Messrs. Lubbock and Johnson are co-founders of J.L. Partners, a polling firm.
