By Chris Barnes, President, Escalent
Photo by Mick Haupt on Unsplash
The 2026 U.S. midterm elections saw pollsters get the story right and the numbers wrong. The majority of polling from prominent pollsters said Democrats were headed for strong midterm results, and they were. But in race after race, the final averages understated the size of the wave, in some of the most-watched contests by nine points or more.
That gap will extend talk of broken polls. The truer lesson is quieter: measuring opinion has become genuinely hard, and even a flawless poll is a photograph of a moment already gone.
The degree of difficulty for polling is increasing with many of the same challenges market research faces: people are harder to contact, fewer people participate in surveys, fraud is rampant in online polling, and more. Add unwelcome additions, like “fake polls” released by rogue organizations, and it is clear that polling is demonstrably more difficult than the days of random digit dialing to quickly get 1,000 respondents.
What the polls missed
Nearly every major average underrated Democrats. An early estimate puts their national House margin near 13 points, against a final polling average of about 9. In the marquee state races, the misses were bigger (AP counts as of Nov. 4):
Georgia's average named the right winner and still missed by more than nine points. Calling the winner is not the same as measuring the electorate.
Compare 2024, the last time polls were said to have “blown it.” The final averages showed Kamala Harris narrowly ahead in Wisconsin and Michigan and, by 538's count, Pennsylvania too. Trump won all three, but those leads were a point or less.
In 2024, no swing-state average missed by 3 points, and every error underrated Republicans. Two years later, several of the most-watched races were missed by more than 9, nearly all underrating Democrats. Pollsters spent eight years learning to find the voters they kept missing. Then the error switched sides and grew. Giant swings in the composition and behavior of the electorate make turnout models less stable.
Barrier One: Almost nobody picks up
A typical Pew phone survey completed interviews with 36% of sampled households in 1997, 9% in 2012 and 6% in 2018. Robocalls and spam filters helped push Pew itself mostly online.
Low response alone doesn't doom a poll. The danger comes when willingness to answer is tied to politics. If one party's voters grow more or less willing to talk to pollsters, demographic weighting can't fully fix it, because the bias hides inside groups that look balanced on paper. The American Association for Public Opinion Research (AAPOR) concluded nonresponse drove at least part of the 2020 miss but could not pin down who was missing. You cannot interview the people who refuse to be interviewed.
Barrier Two: Polling an electorate that doesn't exist yet
A perfect sample of adults is still the wrong sample. An election poll must describe the people who will vote, a group that doesn't exist until the polls close, and people are poor forecasters of their own behavior.
That guess carries enormous weight. In October, YouGov showed that one of its surveys could put Democrats ahead by anywhere from 6 to 16 points depending only on how likely voters were defined; it judged 8 to 14 defensible. If the count holds near 13, the answer sat inside that window but well above the roughly 9-point consensus.
Midterms make the guess harder: lower turnout, mid-decade redistricting, shifting voting rules and, this year, the fewest incumbents on the ballot since 1992.
Barrier Three: Every fix is a bet on the last election
After underrating Republicans in three straight presidential elections, pollsters adapted. By 2024, almost 70% of national polls were weighted by party or past vote, which Pew found made them modestly more accurate.
But every correction is a bet. Weighting to the 2024 vote assumes the next electorate resembles the last. In a midterm where one side surges and the other stays home, that anchor can drag a poll back toward an electorate that no longer exists. Whether that happened this year is now being tested. Correct too hard and you overshoot; don't, and you repeat the miss.
The field has also gotten noisier. Cheap opt-in polls have multiplied, bot and bogus respondents have crept in, and some "synthetic surveys" never ask real humans at all.
The clock doesn't stop when the poll does
Even a flawless poll measures the past. Final polls are fielded days before the election, while campaigns, news and undecided voters keep moving. In 2016, about 13% of voters in Wisconsin, Florida and Pennsylvania decided in the final week and broke heavily for Trump, after nearly all the Wisconsin polling was done. Early voting adds a twist: some ballots are cast weeks before the last poll, others after it.
There is an honest caveat here. "Polls are snapshots" can become an all-purpose excuse. A miss as large and as uniform as 2026's, running the same direction from Texas to Maine, looks at least partly like a measurement problem. Only studies matching respondents to voter records can settle it.
Why it still has to be done well
Polls steer where parties and donors spend, which candidates get attention and which races look competitive. A race that appears hopeless can be starved of money and actually become hopeless. The errors ripple outward too: prediction markets largely missed the same upsets the polls did.
Surveys are also how the public tells leaders what they think about far more than horse races. Each visible miss makes it easier for politicians to wave away numbers they dislike.
Expecting the right things from polls
The margin of error printed under a poll covers only sampling luck, not refusals, turnout guesses or late shifts. A study of more than 4,000 state polls found real-world error ran about twice the stated margins, with polls of the same race sharing common errors. So treat a three-point lead as a close race, trust averages over single polls, ask whether real people answered, and remember that close races tend to move together.
In 2026, the polls told us the national mood and which way the election leaned. They could not say precisely how far, because no one can fully measure an electorate still deciding whether to show up. That isn't a scandal. It is what photographing a moving crowd looks like, and the answer is better cameras and realistic expectations, not throwing the camera away.
About the Author
Christopher Barnes is president of Escalent, a research and advisory firm. He has more than 35 years of experience in market and public opinion research, including co-founding the Center for Survey Research and Analysis at the University of Connecticut. His studies have appeared in The Wall Street Journal, USA Today, The New York Times, and Time cover stories.