For the third consecutive presidential election, polls underestimated Donald Trump’s support. Yet 2024 also saw a significant improvement in polling accuracy. While polling averages before Election Day did predict that Kamala Harris would win the national popular vote, their average error of less than 3 percentage points was substantially lower than the nearly 5-point misses in 2016 and 2020.
In competitive states, where the stakes are highest, polls came within 2.2 points of the actual results on average. This means these polls were more accurate than polls in any presidential election since 2012.
Why did pollsters do a better job this year? Our analysis suggests that more aggressive weighting techniques helped reduce the biases that had led polls to miss the mark more dramatically in the two prior presidential elections. The polls still underestimated Trump’s support, but the improvement indicates that pollsters have made progress – and may be able to improve poll accuracy for future election cycles.
The polling landscape is constantly changing
As we described in our analysis published here on Good Authority in October, very few state polls back in 2016 adjusted their data to account for education levels. Many experts later blamed this oversight for the large polling errors in swing states that year. As the graph below shows, just 17% of polls in competitive states conducted in the last three weeks of the 2016 campaign made adjustments to ensure that their samples reflected the educational attainment of the state’s electorate. But in this past cycle, more than three-quarters of state polls factored in educational levels.
In our last post, we also noted another significant shift. This year’s national polls were much more likely to weight responses for party affiliation and past voting behavior – that is, which presidential candidate respondents recalled voting for in 2020. In 2016, 22% of polls weighted on at least one of these factors – but almost 70% of polls did so in 2024. Two-thirds of state polls incorporated these weights in 2024, mainly by weighting to past vote choice.
Have these changes made polling more accurate?
While the pro-Democratic bias in national polls (the extent to which the polls overstated support for Harris) was still about 3 percentage points, the figure below shows that the inaccuracies would have been much worse without these weighting practices.These estimates come from a regression model where we also control for other factors such as a poll’s sample size, how the sample was reached (online, phone, or a mix of survey approaches), and population type (likely voters or registered voters).
Weighting to a respondent’s education level emerged as the most effective adjustment, reducing Democratic bias by almost 1.1 percentage points.
Weighting on income levels and urban/suburban/rural distinctions also showed promise, reducing bias by 0.39 and 0.64 points, respectively. One caveat here is that relatively few pollsters actually adjusted for these variables, so we would need more data before we could be confident about how this weighting affects polling accuracy. And while the urban/suburban/rural distinction had some impact, regional weighting (adjusting for geographic regions of the state or in America, depending on whether the poll is a national or state poll) actually increased Democratic bias slightly (+0.38).
Despite the dramatic increase in the use of weights for past presidential vote and partisanship, these factors showed surprisingly modest effects on overall accuracy in 2024. Weighting for past vote decreased Democratic bias by 0.39 points, while weighting to party made essentially no difference in reducing Democratic bias. Notably, state-level polls benefited much more than national polls from weighting to past vote. This may be due to the fact that demographic weights are sometimes less influential in states with more homogeneous populations, an issue that weighting to past vote can help solve.
Ultimately, our analysis suggests that a poll of likely voters weighted on educational attainment and past vote would reduce the pro-Democratic bias by about half, compared to a poll that did not weight on those factors.
Increased weighting helped pollsters avoid big misses
Fewer polls in 2024 also missed the actual results by a wide margin, compared to the 2016 polls. In 2016, three of the 23 national polls fielded in the last few weeks of the campaign were off by more than 6 points. A late October 2016 ABC News poll for instance, showed Hillary Clinton up by 12 points over Donald Trump. In 2024, the biggest miss was a Big Village poll showing Kamala Harris up by about 7 points.
At the state level, this pattern of improved accuracy is even more pronounced. In 2016, more than one-fifth of polls in competitive states were off by more than 6 points, and five of these polls were off by more than 10 points. In 2024, such big misses happened only about half as frequently. And while five polls in competitive states missed by more than 10 points in 2016, that happened just once in 2024.
In general, the polls in this election cycle produced estimates more similar to each other than we witnessed in 2016. Some experts, including Nate Silver, have suggested that this is because pollsters were selectively releasing polls that were close to the polling averages, a phenomenon called “herding.” But it’s also possible that the smaller number of outliers is because pollsters now weight on more factors that are better predictors of people’s voting behavior (like education and past vote choice). Consequently, weighting serves to reduce the random error in the polling estimates. All of these moves help produce polls that look more similar to each other.
Pollsters need to continue to adapt to challenges
This increased accuracy in 2024 doesn’t mean that there isn’t more room for improvement. After all, the polls overstated Americans’ support for the Democratic candidate for the third consecutive presidential election. Our analysis suggests several promising directions for pollsters to explore, including weighting survey responses based on respondents’ income levels and their geographic setting (urban, suburban, or rural). While few pollsters did this in 2024, those that did saw a lower level of pro-Democratic bias.
Looking ahead to 2026 and beyond, there’s every reason to believe polls will become even more accurate. Pollsters have shown they can successfully adjust their methods to capture changing voter behavior. The data from 2024 clearly demonstrate that polls are far from irreparably broken. In fact, U.S. polling has improved significantly since the reckoning that followed the major misses of 2016.
