How to read a public opinion poll before reporting it

What to take away

• Start with who was eligible to answer, not the headline result.
• Read the full question and any preceding prompt.
• Treat subgroup estimates as less precise unless shown otherwise.
• Margin of sampling error does not cover every survey error.
• Field dates matter when opinion may change quickly.

A poll is a measurement made under defined conditions. The reported percentage depends on who could be selected, who responded, how the question was asked, when it was asked, and how responses were weighted.

Use the steps below before writing a lead or comparing the result with another poll.

Photo and credit

The screen shows market software, not a public opinion poll. It illustrates why a plotted line needs a defined source, axis, and measure.

Step 1: identify the sponsor and fieldwork organization

The sponsor paid for or commissioned the poll. The fieldwork organization collected or managed responses. Record both. Sponsorship does not invalidate a poll, but readers should know who asked the question and who ran the study. The umbrella rules sit in the guide to reporting statistics and polls.

Look for a methods statement rather than relying on a press release.

Step 2: define the population

Possible populations include all adults, registered voters, likely voters, residents of one area, customers, or members of a profession. Results for one population cannot be relabeled as the view of another.

Likely-voter findings depend on a model or screen. Ask how respondents qualified and whether the rule changed from earlier releases.

Step 3: inspect sample and recruitment

Record sample size, source, selection method, recruitment, response mode, languages, and whether the sample was probability-based or nonprobability-based. A large opt-in sample is not automatically more representative than a smaller probability sample.

Pew Research Center's Methods 101 explainer on random sampling describes random sampling as a concept that undergirds all probability-based survey research, and the series exists to make survey methods accessible to people who are not statisticians or pollsters. It explains the design principle; it does not grade any individual poll.

Ask how people without internet or phone access could be included. Check whether panel members were surveyed before on the same topic.

Step 4: read the questionnaire in order

Question wording can change the concept being measured. "Support," "approve," and "prefer" are not interchangeable. Read answer options, instructions, preceding questions, and whether choices were rotated.

Do not shorten a question in a way that removes a condition. If respondents were told about cost, timing, or a tradeoff, include that information in the article.

Step 5: record field dates and events

A poll collected over five days reflects responses during that span. A major event on day three may affect later respondents. Do not describe the result as a reaction to an event that occurred after most interviews. Time discipline of this kind is the core of a news timeline as well.

Use the end of fieldwork, not the publication date, when placing polls on a chronology.

Step 6: examine weighting

Pollsters may weight responses to align the sample with known population characteristics. Record the weighting variables and source benchmarks. Heavy weights can increase variance, which should be reflected in uncertainty calculations when the design documentation provides it.

AAPOR's standards page lists disclosure items such as sample size, sampling error, weighting attributes, full question wording, mode, target population, construction, and recruitment. The AAPOR standards and ethics overview supports requesting those method details. Compliance with disclosure standards does not guarantee that a particular estimate is correct.

Step 7: interpret uncertainty

Report the confidence level and margin of sampling error when applicable. Check whether the stated margin is for the full sample only. Smaller subgroups normally have wider uncertainty.

Do not announce a leader when the difference is too small to distinguish under the poll's design and stated uncertainty. "Candidate A leads 48 to 46" can overstate a two-point gap if uncertainty around the difference is wider. The poll estimates comparison keeps results, margins, averages, and forecasts from trading places.

Also consider nonresponse, coverage, measurement, processing, and model error. These are not all contained in the familiar plus-or-minus figure.

Step 8: compare cautiously

Compare polls only after matching population, wording, mode, field period, and treatment of undecided respondents. A trend within one consistent series is often easier to interpret than a difference between unrelated polls. Every comparison also passes the definition and period items in the developing story context checklist.

For an average, read the inclusion, adjustment, and weighting method. Do not assume averaging removes shared bias.

Publication box

Include a compact method note:

• sponsor and fieldwork organization;
• target population and sample size;
• recruitment and mode;
• field dates;
• weighting summary;
• question wording or source document;
• uncertainty statement;
• notable limits.

Common questions

Does a larger sample always mean a better poll?

No. Selection, coverage, response, weighting, measurement, and model quality also matter.

Is an online poll scientific?

It can be, depending on recruitment and design. An open website vote where anyone chooses to participate does not represent a defined population without strong additional assumptions.

Why are subgroup results unstable?

They use fewer observations and may have different design effects. Read subgroup sample sizes and uncertainty rather than applying the full-sample margin.

Can a poll predict an election result?

It measures opinion under its design and field dates. Election outcomes also depend on later opinion, turnout, candidate choice, and administration of the vote.