How to read a poll without being misled

Summary: A poll measures what a sample of people say about a question, and it comes with 3 facts attached: A margin of error, meaning how far its numbers can be off because only a sample was asked; The dates it was taken and the group it asked. Most polls that mislead readers were reported accurately, because the headline kept the lead and dropped those 3 facts. Put them back before you believe the number.

Polls rarely mislead people. Headlines about polls do it every week. A poll arrives with a margin of error, a set of field dates, meaning the days on which it was conducted, and a definition of who was asked. A headline has room for 1 number, so it keeps "Candidate A leads by 2" and drops the 3 facts that say what the 2 is worth.

The margin of error applies to each number, not to the gap

The margin of error is the first fact to put back, and most readers put it back wrong. Pew Research Center has an explainer on the subject that makes 2 points which rarely reach a headline. First, a stated margin of plus or minus 3 points applies to each candidate's number on its own. The gap between the 2 candidates can be off by about 6, because one number could be 3 points too high while the other is 3 points too low. Second, "plus or minus 3 at 95% confidence" means that if the same poll were run 100 times, about 95 of the results would land within 3 points of the true figure. 5 would not, and nothing in a poll's published results can tell you whether the one you're reading is among the 95 or the 5.

Domenico Montanaro, a political correspondent at NPR, the US public radio network, put the practical version in a guide to the November 3 midterm polls, published on October 13, 2026. With a 3-point margin of error, a candidate shown 2 points ahead "could have a 5-point lead or be losing by a point." His conclusion: "When someone has a 2-point lead, that's not a lead at all." Apply Pew's doubling to the gap and that candidate could be about 8 points ahead or about 4 points behind.

The margin also covers only 1 kind of error, the chance that a random sample differs from the whole population. It says nothing about the people who refused to answer, the people the pollster couldn't reach, or a turnout model, the pollster's guess at who will actually vote, that was wrong. Pew's 2024 guide to election polling puts it bluntly: "The real margin of error is often about double the one reported." Larger samples shrink the margin, but slowly. A typical sample of about 1,000 people carries a stated margin of about 3 points, and Pew's July 2026 midterm survey, which asked 3,554 people, only brought its margin down to 1.9. That is why almost no poll you'll see gets much below a 2-point margin, and why "leads by 2" is rarely a lead in any of them.

Who was asked changes the answer

The second fact a headline drops is the group. Polls of "adults," "registered voters," and "likely voters" measure 3 different populations: Everyone over 18; The people on the voter rolls, many of whom won't vote; And the subset a pollster judges will actually cast a ballot, usually based on their voting history and whether they say they will vote. In a midterm year, when turnout is lower, the 3 can give very different answers to the same question. Pew's October 8, 2026 report surveyed 3,917 registered voters from September 28 to October 4, and 44% said they would back a Democrat for the US House of Representatives while 36% said a Republican. A headline would say "Democrats lead by 8." But Pew noted in the same report that "many registered voters will not turn out to vote this fall," so the 8-point lead belongs to a group that includes people who won't be in the election, and the headline never said so.

On October 5, Douglas Rivers of YouGov, a British online polling company, showed how much the choice of group matters. He sorted registered voters by how likely they were to vote. Among those who said they had already voted or would definitely vote, Democrats led by 13 points. Among the rest, Republicans led by 15. Same month, same question, and the answer swings by 28 points depending on which group you count. The direction of the swing changes over time, too: Rivers notes that between 1994 and 2010, in polls by Gallup, the American polling company, Republicans did 4 to 11 points better among likely voters than among registered voters, and in 2026 Democrats are the ones who do better among the people most likely to vote. So if a likely-voter poll shows Democrats ahead by 13 one week and a registered-voter poll shows them ahead by 8 the next, the race hasn't moved 5 points. The 2 polls counted different people, and a headline that reports the second one as a slide for the Democrats has invented a change.

How polls are made now

The third fact is the method, and it has changed more in a decade than in the 50 years before. Pew's 2023 study of national pollsters found that 61% used methods in 2022 that differed from what they used in 2016. The biggest shift is in how pollsters find people. The traditional method picks people at random, for example by dialing random phone numbers. The newer method, an opt-in online sample, recruits people who have signed up to take surveys, often for small rewards. The number of national pollsters using opt-in samples rose from 10 in 2012 to 47 in 2020. Opt-in polls are much cheaper, but because the respondents chose themselves, the pollster has to do more adjusting afterward to make the sample look like the country. By 2022, only 10% of pollsters still used live telephone calls alone.

The reason is that almost nobody picks up. Pew's typical telephone response rate fell to 6% in 2018, and the cumulative response rate on its 2026 panel, counting everyone invited since recruitment, is 3%. Pew's review of the research still found that "response rates are an unreliable metric of accuracy," because a small sample that's been weighted, meaning adjusted so its mix of ages, education levels, and other traits matches the population, can beat a large one that hasn't. The low rate does mean that every poll now depends on choices the pollster made about which people to count and how much, and those choices should be written down where you can find them. Montanaro's rule covers it: "If you can't find a methodology, that's a red flag."

Polls were more accurate in 2024 than their reputation

Readers who distrust polls usually cite 2016 and 2020, and the record for those years is bad. A task force of the American Association for Public Opinion Research, the professional body for pollsters, in a report released October 29, 2025, measured 611 polls from the final 2 weeks of the 2024 campaign. The average error on the margin between the 2 parties was 3.3 points, down from 5.3 points in 2020 and 5.2 in 2016. The task force called 2024's state polls the most accurate for any presidential cycle since 1944.

The same report names the problem that remains. In 2016, 2020, and 2024, the polls gave Democrats slightly more support than they received on election day: By 3.1 points in 2016, 4.6 in 2020, and 2.7 in 2024. Pollsters have tried different fixes, such as calling by phone instead of polling online, or adjusting their samples in new ways, and the task force found that none reliably did better than the others: "Most major methodological choices showed little relationship to error size." Nobody yet knows exactly why the polls lean the way they do, so no pollster can promise you a poll without that lean.

Put the 2 findings together and 2024 reads differently from how most people remember it. Most swing-state polls showed a race inside the margin of sampling error, the task force says, and the result was close. The polls said "too close to call," and they were right. The readers who came away certain of a winner got that certainty from headlines, averages rounded to a leader, and forecasts.

The case against reading polls at all

The strongest objection is that none of this helps on the day. The average poll in 2024 was still 3 points off, the error has pointed the same way for 3 cycles running, and the task force itself couldn't say which method avoids it. A reader can't correct for a bias the professionals can't locate, and the horse-race coverage that polls feed, the reporting that treats an election as a contest over who is ahead, tells you nothing about what a candidate would do in office. If you accept that argument, you shouldn't read polls at all.

We'd concede the first part. A poll can't tell you who will win a close race, and the systematic error, the lean toward Democrats that has shown up in 3 elections running, is real and unsolved. But a poll doesn't need to name the winner to be useful. Even in 2020, when the average poll was 5.3 points off, a race the polls showed at 20 points apart was still safe, and a race they showed at 2 points apart was still close. The polls got the closeness of each race right even when they missed the winner, and in 2024 they got both right in most states. Closeness and movement are the information a poll exists to give you, and nothing else measures public opinion as carefully. The headlines turned a lead inside the margin of error into "X leads," and the forecasts turned a toss-up into a percentage. The fix is to read the poll instead of the headline about it.

How to read a poll in 6 steps

Each step checks something printed in the poll's own release.

1. Find the margin of error, and double it before judging a lead. If the gap between 2 candidates is smaller than twice the stated margin, read the race as tied.

2. Find the field dates, the days the poll was taken. Pew's August 2024 guide notes that shifts measured right after a major event are "sometimes short-lived," so a number taken before one tells you about before.

3. Find the population: Adults, registered voters, or likely voters. Compare a poll only with others of the same kind. In October 2026 the gap between registered and likely voters was larger than most candidates' leads.

4. Find the sample size and the method statement. About 1,000 respondents gives a margin near 3 points. If the release names no sample, no mode (how people were reached, such as by phone or online), and no weighting, there is no poll to read, and Montanaro's red flag applies.

5. Find who paid for it. A sponsor's name appears in the release. The polling association's task force noted that Republican-affiliated pollsters may tend to release Republican-friendly numbers in most elections, which made them look accurate in 2024 and wouldn't in a year the average poll leaned the other way. A campaign releases a poll when the number helps the campaign, so weigh it accordingly.

6. Read several, not one. A single poll with a surprising number may be an outlier, and the average of the last 10 is harder to fool. FiveThirtyEight, the poll-averaging website that ABC News owned, was closed by Disney in March 2025, so you may have to assemble the average yourself, and Montanaro's warning that state-level polling "is not very good" applies to state averages too.

None of these checks will tell you whether the poll in your hand is one of the 5 in 100 that land outside their own margin, or which way this year's systematic error runs. They will tell you whether a reported lead is a lead. On November 3, when the midterm results come in, a reader who doubled the margin on every "leads by 2" headline this fall read those races as ties, and the result in those races, whichever way it goes, is the one the polls were reporting all along.