How Accurate Are Polls?
The real error figures, a correct account of 2016, 2020 and 2024, and an honest line between a rigorous poll and an informal one.
- Error: 2 to 3 points
- 30,000 polls reviewed
- 2016, 2020 and 2024
- What your poll can say
Make a poll here, free. No account needed, and you get a link you can send anywhere.
Polling 1 million people every day since 2003
What a normal polling error looks like
Polls miss. The question is by how much, and whether the miss was bigger than the poll admitted.
The yardstick is error on the margin: a poll putting a candidate 3 points ahead who then wins by 1 has an error of 2 points. Average that across a cycle and you can compare one election with the next. Three task force reports from the American Association for Public Opinion Research (AAPOR) supply the figures for US presidential polls fielded in the closing fortnight.
AAPOR task force reports on 2016, 2020 and 2024 pre-election polling
Two things fall out. Misses of 2 to 5 points on the margin are the normal working condition of election polling, not a scandal. And state polls are consistently worse than national ones, which matters in a country that elects its president state by state.
Across every office, the 2024 report gave one number per cycle: among 611 polls of presidential, Senate, governor and House races in the final two weeks, average absolute error on the two party margin was 3.3 points in 2024, 5.3 in 2020 and 5.2 in 2016.
The margin of error you see quoted is not the poll's total error. A typical poll of 1,000 people carries a sampling margin of about plus or minus 3 points, covering only the luck of who got sampled.
It ignores noncoverage, nonresponse and measurement error. Pew Research Center is blunt: the real margin of error is often about double the one reported.
So a poll at 47 to 44 with a 3 point margin of error does not mean a 3 point lead, give or take 3. It means these two are close and the poll cannot say who leads.
Why "the polls were wrong" is usually a misreading
Four different mistakes get filed under the same complaint. Only one is a polling failure.
Treating the margin of error as the whole error
Sampling error is one of four error sources and often not the largest. A poll can be flawlessly executed and still land outside its stated margin, because the people it could not reach differed from those it did. An election is also a single draw, so honest uncertainty misses sometimes.
Applying the margin of error to the lead
A margin of error applies to each candidate's share, not the gap between them. Uncertainty on a lead is roughly twice that on a single number: a poll of 1,200 showing 35 to 34 is consistent with anything from a 7 point lead one way to 5 points the other.
Reading a national poll as an electoral prediction
2016 is the cleanest case. National polls put Clinton about 3 points ahead and she won the popular vote by 2.1, an average absolute error of 2.2 points and among the most accurate polling since 1936. The presidency is not decided by that number: state polls were off by 5.1 points and named the wrong winner in Pennsylvania, Michigan and Wisconsin.
Reading a poll as a forecast
A pollster measures opinion now. A forecaster models an outcome later, adding assumptions about turnout and how state errors move together. In 2016 two modelers put Clinton's chance of winning at 99 percent. That was a forecast, not a poll. Andrew Kohut's line, quoted in AAPOR's 2016 report: "I'm not a handicapper, I'm a measurer."
Candidate on 47Candidate on 44
The direction of recent US misses is a real finding, though, not noise. AAPOR's 2016 report found that whether polls miss toward Republicans or Democrats in a given year was historically close to a coin flip, citing 2000 and 2012 as years when Democrats were underestimated. The 2024 task force had to update that: three consecutive presidential cycles have now underestimated Republicans, which it calls historically rare.
2016, 2020 and 2024, in the right direction
Each cycle is misremembered in a different way. Here is what the official post mortems found.
| Cycle | National polls | State polls | Who the polls flattered |
|---|---|---|---|
| 2016 | 2.2 points | 5.1 points | Clinton |
| 2020 | 4.5 points | 5.1 points | Biden |
| 2024 | 2.6 points | 3.0 points | Harris |
2016: the national polls were fine, the state polls were not. AAPOR found national polls among the most accurate at estimating the popular vote margin since 1936: 2.2 points of average absolute error, 1.3 signed toward Clinton. State polls ran 5.1 absolute and 3.0 signed toward Clinton, and named the wrong winner in three states that together decided the Electoral College. The best supported causes were late movement, with about 13 percent of voters in Wisconsin, Florida and Pennsylvania deciding in the final week and breaking heavily for Trump, and failure to weight by education, which mattered because education tracked vote choice closely and better educated people answer surveys more readily. Tests for a "shy Trump" effect found nothing.
2020: the polls named the right winner and still had their worst year in four decades. This is the part most often stated backwards. AAPOR's 2020 task force found the highest national popular vote error in 40 years and the highest state level error in at least 20. Polls in the final two weeks missed the national margin by 4.5 points and the state margin by 5.1, and the error ran toward Biden: a signed 3.9 points nationally, 4.3 in the states. Senate and governor races averaged 6.0 points too favorable to Democrats.
The 2016 fix had already been applied. Of the state presidential polls that disclosed their weighting, 92 percent adjusted for education, and the task force found no evidence that education weighting caused the error. It also ruled out late deciders, shy Trump voters, demographic composition and differential turnout, and could not name a cause. Its leading candidate was unit nonresponse, a systematic difference between the people who took the polls and those who did not, covered on our nonresponse bias page.
2024: the best of the three, with the same lean. The 2024 task force, published in October 2025, found public polls more accurate than in 2020 or 2016. National presidential polls missed by 2.6 points and state polls by 3.0, the most accurate state polling of any presidential cycle since 1944. The lean persisted: Democratic margins overstated by 2.7 points across all offices. The report found no evidence of herding and no methodological recipe that reliably lowered error; the groups polls handled worst were Republicans in Republican leaning areas, Hispanic voters, and people who had not voted in 2020.
In 2020 the polls named the winner and posted their worst error in 40 years. In 2016 they were about as accurate as they have ever been and the election still shocked people. Calling the winner and being accurate are two different tests.
Are polls getting worse?
The best evidence says no, and it is not close.
Will Jennings and Christopher Wlezien assembled more than 30,000 national polls from 351 general elections in 45 countries between 1942 and 2017, published in Nature Human Behaviour. Their conclusion, in their own words: contrary to conventional wisdom, the recent performance of polls has not been outside the ordinary. Writing for Democratic Audit, Sean Swan sized the non change: polling error in the 2015 to 2017 period ran at 2.6 points, about 0.2 points above the whole 1942 to 2013 span.
What has changed is how hard people are to reach. Response rates to Pew telephone polls fell for two decades and have stabilized around 9 percent. That sounds fatal and is not: Pew found the average absolute difference between its telephone estimates and government benchmarks was 2.7 points in 2016 against 2.8 in 2012. Response rate is an unreliable indicator of bias, the central point of our nonresponse bias page.
The other change is who is polling. The number of organizations running national election polls more than doubled between 2000 and 2022, driven mostly by cheap opt-in sampling, and Pew finds nonprobability surveys can carry errors twice as large on average. The pool has widened and thinned, even though a well made poll is as good as it ever was. One signal does track accuracy: organizations in transparency programs such as the AAPOR Transparency Initiative and the Roper Center archive have less error on average.
What a rigorous poll can claim, and what yours can
This site makes informal polls. That is a useful tool with a real boundary, worth stating plainly.
A rigorous poll rests on five things: a stated target population, a sample drawn so everyone in it had a known chance of selection, weighting to independent benchmarks, wording tested for bias, and a published methodology. Remove one and the poll stops estimating a population and starts describing whoever answered.
An informal poll, which is what a link poll, a website widget or a social media poll is, has none of them. No frame, so no defined population. No weighting, so a group that over responds stays over represented. And anyone can share the link, so a motivated group can decide the outcome.
| Claim | Rigorous poll | Informal or link poll |
|---|---|---|
| "X percent of Americans think ..." | Yes, within a stated error | No |
| "X percent of readers who answered think ..." | Yes | Yes |
| "Our team prefers Thursday" | Yes | Yes, if you asked everyone |
| A number with a margin of error attached | Yes | No |
| A trend, rerun the same way on the same audience | Yes | Usually, with care |
The gap does not close just because a panel is professionally run. Pew compared opt-in polls against its own probability panel and found one reporting that 20 percent of adults under 30 agreed the Holocaust is a myth, against 3 percent on the panel. In a check question, 12 percent of opt-in respondents under 30 claimed a license to operate a nuclear submarine, a share that in reality rounds to zero.
Deciding inside a group
When you can reach everyone who counts, a poll stops being a sample and becomes a vote. Choosing a meeting time is a census of a small population, the one case where an informal poll is as valid as any other. A scheduling poll is the clearest example.
Fully valid
Measuring your own audience
If the population you care about is the people who read this page, the respondents are drawn from it. Report it as exactly that: the views of readers who answered, with the count shown.
Valid, if labeled
Engagement and pretesting
A poll holds attention, and running two wordings past a small audience shows which one confuses people. Fun poll questions and website feedback polls earn their keep this way.
Fine, if labeled
- Ask a list you control rather than broadcasting a link, so you know who was invited.
- Always publish the number of responses, and never quote a percentage without it.
- Name who you asked in the same sentence as the result. "Of 214 subscribers who answered" is honest; "of Americans" is not.
- Watch subgroup counts. A poll of 300 reporting what managers think may be reporting 19 people.
- Match the scale to the question, and do not average one with no fixed intervals. See levels of measurement.
- Keep the wording neutral. Survey question examples and feedback question examples help; for longer answers a form beats a poll, and how to make a poll covers the mechanics.
How to read a poll in two minutes
Six questions. If a poll will not answer them, that is the answer.
Who was sampled, and from what list?
Probability sample from a defined frame, opt-in panel, or a link anyone could click? This decides more than anything else whether the number means something about a population.
How many people, and how many in the group being quoted?
A national poll of 1,000 might hold 150 adults under 30. The headline margin of error does not apply to that subgroup, and plenty of viral findings are subgroup findings.
What was it weighted to?
Census benchmarks for age, sex, race, region and education are standard; weighting on party identification or past vote is a real choice with real consequences. What matters is that the pollster says which were made.
When was it in the field?
Error is most predictable in the closing week or two. A poll taken months out measures a different thing from one taken three days out.
What exactly was asked?
Wording moves results by more than most polling errors do. Pew found 68 percent favored "taking military action in Iraq to end Saddam Hussein's rule" in 2003, but 43 percent when the question added "even if it meant that U.S. forces might suffer thousands of casualties."
Whose poll is it, and do others agree?
House effects are real. The 2024 AAPOR report found Republican affiliated pollsters slightly more accurate that year, then noted the likely reason: releasing more Republican friendly numbers helps when the average poll leans Democratic and hurts when it does not. A poll disagreeing with all the others is usually a house effect, not news.
One last question, which is not about the poll: is this a poll or a forecast? A poll reports what a sample said. A forecast turns polls into a probability using assumptions you cannot see. Only one of them is a measurement.
Common questions
How accurate are polls, in one number?
Across 611 US polls fielded in the final two weeks of the 2024 campaign, average absolute error on the two party margin was 3.3 percentage points, against 5.3 in 2020 and 5.2 in 2016. In 2024, national presidential polls missed by 2.6 points and state polls by 3.0.
What does a margin of error actually cover?
Only sampling error: the luck of which people were selected. A typical poll of 1,000 people carries a margin of about plus or minus 3 points on each share. It ignores noncoverage, nonresponse and measurement error, and Pew's summary is that the real margin of error is often about double the one reported.
Were the 2016 polls wrong?
The national polls were not. AAPOR found them among the most accurate at estimating the popular vote margin since 1936: 2.2 points of average absolute error, correctly showing Clinton ahead in a popular vote she won by 2.1 points. The state polls were wrong, by 5.1 points on average, and named the wrong winner in Pennsylvania, Michigan and Wisconsin. That is the error that mattered.
Did the 2020 polls overstate or understate Biden?
They overstated him, and this is very commonly stated backwards. AAPOR's 2020 task force found polls in the final two weeks too favorable to Biden by a signed 3.9 points nationally and 4.3 points in the states. That was the largest national polling error in 40 years, worse than 2016 rather than better.
Why was 2020 worse than 2016 if pollsters had fixed the education problem?
Because education weighting was not the 2020 problem. Of the state presidential polls that disclosed their weighting, 92 percent adjusted for education, and the task force found no evidence that failing to do so caused the error. It also ruled out late deciders, shy Trump voters, demographics and differential turnout. The leading candidate is unit nonresponse, covered on our nonresponse bias page.
How accurate were the 2024 polls?
Better than either cycle before. AAPOR's 2024 task force, published in October 2025, found average absolute error of 3.3 points across all offices, with national presidential polls missing by 2.6 points and state polls by 3.0, the most accurate state polling since 1944. Democratic margins were still overstated, by 2.7 points.
Do polls have a partisan bias?
Historically no, recently yes in one direction. AAPOR's 2016 report concluded that whether polls miss toward Republicans or Democrats in a given year was close to a coin flip, citing 2000 and 2012 as years when Democrats were underestimated.
But the last three presidential cycles all leaned the same way, overstating Democratic margins by 3.1 points in 2016, 4.6 in 2020 and 2.7 in 2024. Midterms did not: 2018 overstated Democrats by 0.1 and 2022 overstated Republicans by 0.6.
Are polls getting less accurate over time?
The largest study says no. Jennings and Wlezien analyzed more than 30,000 national polls from 351 general elections in 45 countries between 1942 and 2017 and found the recent performance of polls has not been outside the ordinary. One summary puts contemporary error at 2.6 points, about 0.2 above the long run level.
Does a low response rate mean a poll is biased?
Not by itself. Response rates to Pew telephone polls fell for two decades and have stabilized around 9 percent, yet Pew found the average absolute difference between its telephone estimates and government benchmarks was 2.7 points in 2016 against 2.8 in 2012. What matters is whether nonrespondents differ from respondents on the thing measured, covered on our nonresponse bias page.
How many responses do I need for my poll to be accurate?
Sample size buys precision, not representativeness. About 1,000 responses gets you roughly plus or minus 3 points of sampling error, and going higher shrinks that slowly. If the sample is self selected, no number of responses fixes it: a link poll with 50,000 responses is still a description of the people who clicked.
Can I trust an online poll I saw on a website or on social media?
As a measure of the population, no. It has no sampling frame, no weighting and no way to know who saw it and passed, and anyone can share the link to recruit a favorable crowd. Even funded opt-in panels struggle: Pew found one reporting 20 percent of adults under 30 agreeing the Holocaust is a myth, against 3 percent on its probability panel.
Sources
- Jennings, W. and Wlezien, C. (2018). Election polling errors across time and space. Nature Human Behaviour 2, 276-283.
- AAPOR Ad Hoc Committee on 2016 Election Polling (2017). An Evaluation of 2016 Election Polls in the United States. American Association for Public Opinion Research.
- Clinton, J. et al. (2021). Task Force on 2020 Pre-Election Polling: An Evaluation of the 2020 General Election Polls, Executive Summary. American Association for Public Opinion Research.
- AAPOR (2025). Task Force on 2024 Pre-Election Polling: An Evaluation of the 2024 General Election Polls, Executive Summary. American Association for Public Opinion Research.
- Pew Research Center (2024). Key things to know about U.S. election polling in 2024.
- Pew Research Center (2017). What Low Response Rates Mean for Telephone Surveys.
- Pew Research Center (2024). Online opt-in polls can produce misleading results, especially for young people and Hispanic adults.
- Pew Research Center. Writing Survey Questions. Methods section.
- Swan, S. (2018). The limitations of opinion polls and why this matters for political decision making. Democratic Audit, London School of Economics.
Related guides
Ready to ask?
One question, a handful of answers, a link you can paste anywhere. No account, no limits on votes.
Make a free poll