Many posts have discussed pollsters' challenges, including the plunge in response rates. This year, the challenges are even greater. In an polarized era, partisan nonresponse bias is a big problem.
There are two factors that help explain why it’s possible for the polls to be right, even when Mr. Trump’s supporters are less likely to respond.
One is turnout. There’s a lot of evidence suggesting that Mr. Trump’s supporters aren’t just less likely to respond to surveys, but also less likely to vote in midterm elections. This would help explain the basic pattern of polling error during the Trump era: Polls have a pretty good record in midterms but not in presidential elections.
The second factor is a technique called “weighting,” in which pollsters make statistical adjustments to give more weight to voters from groups underrepresented in the survey responses. This gives pollsters a chance to produce reasonable results, simply by giving more weight to registered Republicans, self-identified Republicans or the voters who say they voted for Mr. Trump who did respond to the survey.
Weighting is not a panacea. That’s because nonresponse bias can run much deeper than what can be observed and corrected in the data. A poll might be weighted to have the right number of Trump supporters but still have the wrong kind of Trump supporters — and particularly the kind of Trump supporter who might be voting Democratic.
Imagine, for instance, if the Trump voters who responded to a pollster were likelier to be the kind who supported Nikki Haley in the 2024 primary. A pollster might then have too many of the kinds of Trump supporters who would readily defect to Democrats. The Post believed it had this kind of issue in Kansas: Even after weighting, the results still seemed too good for Democrats.
Despite its imperfections, weighting has kept the survey industry afloat. In elections like 2018 or 2022, it was sufficient. For many pollsters in 2024, it probably prevented another misfire like those in 2016 or 2020. As far as I can tell, many public pollsters — maybe even most public pollsters — would have shown Kamala Harris with a clear lead over Mr. Trump in 2024 if they had used the weighting techniques that were common back in 2018 or 2020. Instead, they’ve adopted heavy-handed weighting techniques that can bludgeon even the worst samples back into the ballpark.
You can see why it’s hard to make any assumptions about the polling error this cycle in today’s Times/Siena polls. This summer, we adopted a new set of weighting techniques to improve our polling. When we re-evaluated our 2024 polls with this new weighting approach, we saw that our polls would have been about 1.6 points better on average. The difference felt meaningful, but it was fundamentally modest.
...
What does all of this mean for 2026? Well, you can choose your own adventure. More than anything, it illustrates how hard it is to predict how the polls might err.
Maybe it implies that we’ve adopted weighting techniques that can handle the problems that caused the Times/Siena poll to overestimate Democratic strength in the past, especially in white working-class states. If so, maybe we’re dead-on.
Or perhaps one could argue that the Times/Siena polls will underestimate Democrats. Our polls were accurate in 2018 and 2022, after all, so maybe one should trust the old approach in a midterm election. If so, maybe the Democrats will do even better than our polls suggest.
Or maybe the extraordinarily blue results on our old methodology and the increased effect of weighting suggest that nonresponse bias is worse than ever. If so, polls could seriously underestimate Republicans once again. There’s no reason to assume that 2016 or 2020 is the worst case scenario.
We’ll find out soon enough.
