How we poll, in full

Principles

Integrity Polling exists because polling’s credibility problem is, at root, a verifiability problem. Readers cannot tell a careful poll from a careless one because the information needed to tell them apart is not published. We publish it. These five commitments apply to every public release without exception:

1. Full methodology with every release. Field dates, sample size, mode, sampling frame, weighting targets, response rate and margin of error, published alongside the numbers. Never on request. Never in a footnote.
2. Verbatim question wording, in the order asked. Every question exactly as respondents heard it, in sequence, including rotations and screeners. If the wording could have moved the answer, you can see for yourself.
3. Downloadable crosstabs, every time. A CSV of every public release broken out by age, party, region, gender and education, with unweighted sample sizes. Not a PDF screenshot. A file you can open.
4. Every ballot-qualified candidate, named. If a candidate is on the ballot, they are in our question and in our report by name. Third-party and independent voters are not rounded into "other."
5. Public self-grading after every election. Within 30 days of certification we publish how far off we were, graded against our own margin of error, with the misses given the same prominence as the hits.

Everything that follows describes how we meet those commitments in practice. Where our methods differ from industry convention, we say so and say why.

Sampling approach

We draw probability samples wherever a usable frame exists. For election polls, the frame is the state voter file: every registered voter, with age, party registration (where recorded), address and vote history. We stratify the draw by region, age band, party registration and turnout history so that the sample starts close to the population rather than relying on weighting to get it there.

For general-population polls, the frame is a probability-based online panel recruited by address-based sampling, with telephone completion or a provided device for households without internet access. We do not use opt-in online panels for public releases. Where a client-commissioned study uses one, the release says so in the first line of the methodology box, and we do not describe its results as representative.

Sample sizes are set from the subgroup estimates we intend to report, not from the topline. A statewide poll that will report results among independents needs enough independents to say something about them; the target n for the smallest reported subgroup is 250, and we do not report crosstab columns below 100 unweighted interviews.

Modes of contact

No single mode reaches everyone, and each mode’s non-response bias is different. We field election polls in a mix of text-to-web, live telephone (cell and landline) and panel completes, and we publish the mix. We examine results by mode before release; where a mode produces results that diverge from the others by more than its own margin of error after weighting, we report that in the release notes.

Interviews are offered in English and Spanish. Interviewers are live, not automated; we do not use IVR (“robo-polls”) because they cannot legally reach cell phones and systematically miss younger and non-white adults.

Weighting philosophy

Weighting corrects for who did not answer. Every poll today depends on it, and the choice of weighting targets is the single most consequential decision a pollster makes. Ours are published with every release. The standard set is:

• Demographics: age, sex, race and ethnicity, education, region (and metro status for national samples), to Census Bureau benchmarks.
• Politics: party registration (where the voter file records it) or party identification averaged over our last three waves, and recalled vote in the most recent presidential election, to the certified result.
• Engagement: a civic-engagement item (contacted an official, attended a public meeting, or volunteered in the past year) to an external benchmark. This is unusual. We do it because the people who answer surveys are more engaged than the people who do not, and demographic weights do not fix that.

We use iterative proportional fitting (raking) and trim weights at 5. We report the design effect from weighting in every methodology box and fold it into the stated margin of error. We do not weight to the results of other polls, to a candidate’s expected share, or to anything we would be embarrassed to publish.

If you can see our weighting targets, you can tell whether a result is coming from the respondents or from the weights. That is the point of publishing them.

Defining likely voters

Who counts as a “likely voter” moves toplines, sometimes by more than the stated margin of error, and the definition is rarely published. Ours is: registered voters who say they will definitely or probably vote, plus those who say the chances are “50-50” and whose voter-file record shows participation in two of the last three general elections. The screener wording appears verbatim in every election release.

We publish the registered-voter result alongside the likely-voter result in the crosstab file so readers can see how much the screen moved the numbers. We do not use a turnout model to re-weight likely voters toward a projected electorate; we report who said they would vote and let the Scorecard show whether that was right.

Question design

Questions are written to the following rules, and the rules are checked by a second researcher before fielding:

• One idea per question. No double-barreled items.
• Balanced stems. “Do you support or oppose,” never “Do you support.” Response scales are symmetric, and the order of the scale is rotated across the sample.
• Definitions before judgments. When a question concerns a policy most people have not heard of, we describe it neutrally first, and we have the description reviewed by people who favor and oppose the policy. Their edits are logged in the questionnaire download.
• Candidates and institutions rotated. Lists are presented in randomized order, except where a fixed position is methodologically necessary (for example, “neither” read last), and we say so.
• No message testing in public polls. A public release never contains a question written to see whether an argument moves opinion. That is commissioned research, and it is not published as public opinion.
• Verbatim publication. Every question appears in the release exactly as asked, in sequence, including screeners and interviewer instructions. Questions we fielded and did not report are listed too.

The all-candidates commitment

Every candidate who has qualified for the general-election ballot as of the first day of fielding is read by name, in rotated order, in the ballot question. “Another candidate” and “undecided” are accepted but never read. Each named candidate’s share is reported in the topline chart, in every crosstab, and in the CSV.

We do not collapse minor-party and independent candidates into “other.” We do not ask a two-way follow-up that pushes their supporters toward a major-party candidate and report that as the headline. If an undecided-leaner question is asked, it is reported separately and labeled as such.

After the election, every named candidate gets a line on the Scorecard with the signed error between our final poll and the certified result. If naming minor-party candidates produces numbers that run high, as the literature suggests they sometimes do, that error is published with the rest. The remedy for a measurement that tends to run high is to publish it and then publish how high it ran. It is not to stop measuring.

How we report uncertainty

Every release states the margin of sampling error at the 95 percent confidence level for the full sample, including the design effect from weighting. We then apply three rules in the write-up:

1. The margin applies to each share, not to the gap. The margin on the difference between two candidates is roughly 1.9 times the margin on a single share. A lead smaller than that is described as “within the range where a different sample could show a tie,” not as a lead and not as a tie.
2. Subgroups carry their own margins. Every crosstab column shows its unweighted n. We do not describe a subgroup difference as meaningful unless it exceeds the margin for that subgroup.
3. Sampling error is not the only error. Each methodology box says so, names the other sources (non-response, screening, wording, timing), and links here. We do not pretend the margin of error is the margin of total error, and the Scorecard grades us on the total.

For the Trust Tracker, each index is published with its standard error, and quarter-to-quarter changes smaller than two standard errors are described as “no significant change.”

AAPOR Transparency Initiative alignment

The American Association for Public Opinion Research’s Transparency Initiative asks member organizations to disclose a defined set of methodological items with every public release. We built our release template so that every one of those items is on the release page itself, not available on request. We will apply for Transparency Initiative membership upon our first public release.

Beyond the Initiative’s requirements, we publish the full crosstabs as a machine-readable file, the questions we asked but did not report, and the post-election grade.

Corrections and versioning

Errors happen. When one is found in a published release, the release is corrected in place, a dated correction note is added at the top of the page describing what changed and why, and the CSV is reissued with a version suffix. The original file remains available. We do not silently edit. Corrections are announced on the same channels as the original release.

This page is versioned. Changes to the standard are listed at the bottom with the date, and a change never applies retroactively to a release published under an earlier version. Substantive critiques of our question wording or weighting are published alongside the release they concern, with our response, at corrections@integritypolling.com.

What we won’t do

• Publish a public poll funded by a candidate, campaign, party or political committee. Commissioned political work is private and is labeled as such.
• Release a poll without its methodology, question wording and crosstabs. If a release cannot meet all five commitments, it does not go out.
• Hold back a result because it is inconvenient to a client, a reader or ourselves.
• Describe an opt-in sample as representative.
• Report a horse race without naming every ballot-qualified candidate.
• Grade our own work in private.

Revision history. Version 1.0, 2026: initial standard. Questions about this document go to hello@integritypolling.com.