How Do Opinion Polls Measure Public Opinion? | How It Works

What is the short answer?

An opinion poll questions a small group of people to estimate what a much larger group thinks. The key is how that group is chosen. If everyone has a known chance of being picked, about 1,000 people can describe a whole country to within a few points.

Pollsters then weight the answers to match known facts about the population, and report a margin of error. That margin covers only the luck of the draw. Who answers, how questions are worded and who turns out to vote can add more error.

Definition: The population is the whole group a poll describes, such as all adults in a country. The sample is the part actually asked. In a probability sample, every member of the population has a known chance of selection.

How can about 1,000 people speak for millions?

A cook can judge a pot of soup from one spoonful, if the pot is well stirred. Random selection does the stirring. What matters is how the sample is drawn and how large it is. Penn State's statistics notes say the margin of error is the same whether the population is 50,000 or 5 billion.

Pew Research Center notes that about 1,000 is now considered an adequate sample. In address-based sampling, a probability method, households drawn at random from the US postal delivery file are invited by mail.

An opt-in panel differs. People join through online adverts or reward websites, so nobody knows their chance of inclusion. In a 2023 test on 28 benchmark measures, Pew found an average error of 5.8 points for opt-in samples, against 2.6 points for probability-based panels.

Size does not rescue a badly drawn sample. In a 2018 study, Pew found virtually no difference in accuracy between opt-in surveys of 8,000 and 2,000 interviews.

What does the margin of error measure?

The margin of error shows how far a result could drift just because of who happened to be picked. Pew explains plus or minus 3 points at the 95% confidence level this way: run the same survey 100 times, and about 95 results would land within 3 points of the true value.

Pew notes that a simple random sample of 1,067 people has a margin of about 3 points on a candidate's support. The margin shrinks with the square root of the sample size, so four times the people only halves it.

The margin on the gap between two candidates is generally about twice the margin on each share. It also covers sampling error only: Pew and the American Association for Public Opinion Research (AAPOR) suggest a poll's real error is often about twice the reported margin. Weighting also widens the margin, an effect called the design effect.

How does weighting correct a lopsided sample?

Even a random sample comes back uneven, because some groups answer more readily than others. Weighting fixes the mix: each person's answers count for more or less until the sample matches known population totals.

A common method, raking, adjusts the weights one variable at a time, such as age, sex, education and region, and repeats until every target matches. In the United States, Pew takes its demographic targets from the Census Bureau's American Community Survey.

Weighting corrects only the traits it uses. If the young people who answer differ from those who do not, weighting by age leaves that gap. In Pew's 2018 study, choosing the right variables mattered more than the method, yet even the best adjustments could not remove most of the bias in opt-in samples.

What else can push a poll off course?

Coverage error arises when the list a sample is drawn from leaves people out. Non-response bias arises when those who answer differ from those who do not. AAPOR warns that with low response rates, even small differences in who responds can bias a poll.

Question wording matters too. In January 2003, Pew asked about military action in Iraq to end Saddam Hussein's rule, and 68% favoured it. When the question added that US forces might suffer thousands of casualties, 43% did. Question order can shift answers, because earlier questions set the context.

Social desirability bias is the pull towards the answer that sounds good. Pew notes it can be greater when an interviewer is present.

Election polls also need to know who will vote. AAPOR notes that actual turnout is generally lower than people report, so pollsters use a likely-voter screen: questions or a model that estimate who will vote.

How do exit polls and poll averages differ from a single poll?

An exit poll interviews people just after they vote, so it need not guess who will turn out. In the United States in 2016, Edison Research ran the exit poll for the National Election Pool of networks and the Associated Press.

It planned about 85,000 interviews at nearly 1,000 polling places, sampled from more than 110,000, plus phone interviews with early and postal voters.

Some countries restrict exit polls. Section 126A of India's Representation of the People Act, 1951 bars anyone from conducting an exit poll and publishing its result during a period the Election Commission notifies. Breaking it can bring up to two years in prison, a fine or both. For the 2019 Lok Sabha election, the window ran from 11 April to 19 May, so that people yet to vote would not be swayed.

A poll aggregator combines many polls into an average. AAPOR says averages are likely to have a little less error than single polls, because different surveys make different errors. An average cannot cancel an error most polls share.

What went wrong in 1936 and 2015?

In 1936, the Literary Digest magazine in the United States mailed about 10 million postcard ballots to names taken mainly from car registrations and telephone books. More than two million came back, pointing to a landslide for Alf Landon. Franklin D. Roosevelt won 60.8% of the vote and 523 electoral votes to 8.

The lists leaned towards wealthier voters, and those who replied may have differed from those who did not. George Gallup, using more systematic, scientific sampling, predicted that Roosevelt would win handily.

At the United Kingdom's 2015 general election, the final polls averaged 33.6% each for the Conservatives and Labour, the British Polling Council recorded; the result in Great Britain was 37.8% and 31.2%.

An inquiry chaired by Professor Patrick Sturgis reported on 31 March 2016 that the primary cause was unrepresentative samples. They over-represented Labour supporters and under-represented Conservative ones, and weighting did not fix this to any notable degree.

What did AAPOR find after the 2016 and 2020 US elections?

AAPOR's 2016 task force, reporting in May 2017, found national polls performed well but battleground state polls often missed. State polls would have done much better had they weighted for education, since less-educated voters were less likely to respond. Voters who decided very late leaned towards Trump.

In 2020, both national and state polls tended to overestimate Biden's support. The task force, reporting in July 2021, found some evidence suggesting that Trump supporters may have been a little less likely to answer polls. AAPOR adds that public polls were more accurate in 2024 than in 2020 or 2016.

Pew tested what errors of the 2020 size would do to questions on issues. Its 2021 analysis found they would change those measures by less than 1 point on average.

How does this work with real numbers?

Both examples use illustrative numbers.

How does the margin of error change from 1,000 to 4,000 people?

For a share p in a simple random sample of n people, the 95% margin is about 1.96 × √(p × q ÷ n), where q is the share not giving that answer, so p + q = 1. It is largest when p = 0.5.

Table: Illustrative margins of error at p = 0.5

Four times the sample gives half the margin, because √4 = 2. If the first poll puts Candidate A on 52% and B on 48%, each share carries about 3.1 points. The 4-point lead carries about twice that, roughly 6.2, so the poll cannot say who is ahead.

How does weighting change a headline figure?

A town's adults are 30% aged 18 to 34 and 70% aged 35 or over. A poll of 1,000 reaches 150 younger and 850 older people, and asks if they would use a new night bus.

Table: Illustrative weighting by age

Unweighted, 345 of 1,000 say yes: 34.5%. Weighted, the younger yeses count as 180 and the older as 210, giving 390 of 1,000, or 39%. As a check, 0.30 × 60% + 0.70 × 30% = 39%.

Where can the student go deeper in World 101?

This page follows one opinion poll, from the drawing of its sample to the careful reading of its result. The World 101 subject Statistics and Data covers the wider field, from how data are collected and summarised to how evidence is weighed, and how careless or selective numbers can mislead.

Three related subjects take these ideas further. Probability, Risk and Uncertainty explains chance and confidence, the ideas behind every margin of error and confidence level. Elections and Voting covers how votes are cast and turned into results, the outcomes that election polls try to anticipate. Psychometrics studies how carefully built questions can measure attitudes, and why wording and order change the answers people give.

Sources

• D. Cohn, "Sample Surveys and the 1940 Census", Pew Research Center, 2 April 2012.
• "The Literary Digest Poll" (1936), randomservices.org, read 7 November 2026.
• F. Newport, "75 Years Ago, the First Gallup Poll", Gallup, 20 October 2010.
• Penn State Department of Statistics, STAT 100 notes, Lesson 2, read 7 November 2026.
• A. Mercer, "5 key things to know about the margin of error in election polls", Pew, 8 September 2016.
• A. Mercer and A. Lau, "Comparing Two Types of Online Survey Samples", Pew, 7 September 2023.
• A. Mercer, A. Lau and C. Kennedy, "For Weighting Online Opt-In Samples, What Matters Most?" and its chapter "How different weighting methods work", Pew, 26 January 2018.
• US Census Bureau, "About the ACS", revised 19 February 2026.
• Pew Research Center, "Writing Survey Questions" (January 2003 example), read 7 November 2026.
• S. Keeter and C. Kennedy, "Key things to know about U.S. election polling in 2024", Pew, 28 August 2024.
• AAPOR, "Polling Accuracy" and "Election Polling Resources", read 7 November 2026; Reports page (2016 evaluation, 4 May 2017; 2020 evaluation, 19 July 2021).
• D. DeSilver, "Just how does the general election exit poll work, anyway?", Pew, 2 November 2016.
• Election Commission of India, notifications under Section 126A of the R.P. Act, 1951, of 7 April 2019 and 24 March 2021, and letter of 8 April 2019.
• British Polling Council, "General Election: 7 May 2015", 8 May 2015, and "BPC Inquiry Report", 31 March 2016.
• P. Sturgis and others, "Report of the Inquiry into the 2015 British general election opinion polls", NCRM, 2016 (abstract).
• S. Keeter and others, "What 2020's Election Poll Errors Tell Us About the Accuracy of Issue Polling", Pew, 2 March 2021.

Common errors and misconceptions

• Misconception: A bigger sample always makes a poll more accurate. Correct: Size cuts only random error; Pew found opt-in surveys of 8,000 barely more accurate than ones of 2,000.
• Misconception: A large country needs a sample in proportion to its population. Correct: The margin depends on the sample's size and design, not the population's size.
• Misconception: A 4-point lead with a 3-point margin is a clear lead. Correct: The margin on the gap between two candidates is about twice the margin on each share.
• Misconception: The margin of error covers every way a poll can go wrong. Correct: It covers sampling error only; non-response, coverage, wording and turnout estimates add more.
• Misconception: Weighting can make any sample representative. Correct: It corrects only the traits it uses; in 2015, weighting did not cure unrepresentative UK samples.

Key takeaways

• With probability sampling, about 1,000 people can describe millions; accuracy depends on how the sample is drawn, not the population's size.
• The 95% margin for 1,000 people at 50% is about 3 points; it halves when the sample quadruples and roughly doubles for a lead.
• Weighting matches a sample to known totals such as age, sex, region and education, but cannot correct hidden differences in who answers.
• Non-response, coverage gaps, question wording and order, social desirability and turnout estimates all add error beyond the reported margin.
• Exit polls interview people who have just voted; India bars conducting and publishing them during periods its Election Commission notifies.
• Analyses traced the 1936 and 2015 misses to unrepresentative samples; AAPOR linked 2016 state errors to education weighting and late deciders.

Test yourself

Why can a well-drawn sample of 1,000 describe a whole country?

The margin of error depends on how the sample is drawn and how large it is, not on the population's size.

What happens to the margin of error when a sample grows from 1,000 to 4,000?

It halves, from about 3.10 to 1.55 points at p = 0.5, because it shrinks with the square root of the sample size.

In the weighting example, why did the headline rise from 34.5% to 39%?

Young people, who backed the night bus more, were under-represented; a weight of 2.00 each restored their 30% share.

How does a probability sample differ from an opt-in panel?

In a probability sample everyone has a known chance of selection; an opt-in panel takes volunteers, whose chances are unknown.

What did the inquiry into the 2015 UK polls find?

The primary cause was unrepresentative samples, which over-represented Labour supporters and under-represented Conservative ones.

Why did the Literary Digest's huge 1936 poll fail?

Its names came mainly from car registrations and telephone books, which leaned towards wealthier voters, and those who replied may have differed.