Methodology

Crosstab runs opt-in online polls. That sentence carries real limits, and we'd rather state them plainly than dress our numbers in borrowed authority.

What these polls are

Respondents find our polls through search, social sharing, and email. They choose to take part. That makes every poll here a convenience sample: a large, fast, honest read on the people who answered, not a probability sample of American adults. No sample size fixes self-selection, so we never describe results as representative of the general population, and we never attach a “margin of error” in the textbook sense.

What the numbers mean

• Percentages describe our respondents, with 95% confidence intervals (Wilson score) that quantify within-sample precision only.
• Group breakdowns hide any cell with fewer than 30 responses, shown as “not enough responses yet” rather than noise dressed as signal.
• Early polls show raw counts until 100 votes, and unlock breakdowns at 500.
• One vote per person is enforced by browser identity plus rate limiting. It raises the cost of ballot-stuffing; it cannot make it impossible, and we monitor for anomalous voting patterns and exclude flagged votes from published numbers.
• Nobody is paid to answer. Paid opt-in panels attract people and software that answer for the reward; the research on bogus respondents is about that incentive, and we do not create it. What remains is the incentive to tilt a public result, and automated answering is now cheap. We record how long each response took, flag responses that are implausibly fast or answer every question the same way, and report the flagged share alongside the results they were removed from.

Screening

Every night, and before any analysis, we screen responses. Nothing is screened at the moment you answer, and the rules would rather miss a bad response than set aside an honest one.

• Single polls: 5 or more votes on one poll from one network within 10 minutes are set aside together and taken out of every number, including the group breakdowns.
• The flagship poll: completions faster than a floor (the fastest 2% of takers, and never under 60 seconds); completions that click the same side of the screen for every question in 4 or more topics, which our left-right randomization makes visible; and earlier results from someone who retook the poll. Set-aside results still show their taker the result; they just never shape anyone else's percentiles.

So far: 0 of 30 poll votes set aside (0%), and 2 of 9 flagship completions (2 superseded retakes).

How questions are written

Our editorial line isn't left or right. It's that most polls are written badly. Every question here follows the same standard: one proposition per question, no leading or loaded wording, response options that span real opinion, tradeoffs stated with their costs, and an honest “not sure.” Where wording itself is contested, we sometimes run wording experiments: half of respondents see one phrasing, half another, and we publish the gap as a finding.

Who answers

We ask each respondent a few demographic questions (age, gender, party, voter registration, education) to power the group breakdowns. Every question has “prefer not to say.” Polls are open to adults 18 and over. Demographic answers are tied to an email address only when a respondent chooses to keep their results and confirms; otherwise the two live in separate systems with no link between them.

Respondents arrive from wherever the polls are shared, and a burst from one community can tilt a result. Links we share carry a label naming where they went (a newsletter, a forum), and each respondent keeps the label of the link they first arrived by, so we can see where answers came from and check whether one source is moving a number.

How our respondents compare with probability polls

Some of our questions use a major pollster's exact words, so our opt-in respondents can be set beside a probability sample. The gap is our house effect on that question: how far our respondents sit from the national number, reported whatever direction it runs.

Where this is headed

As the respondent base grows we'll publish census-benchmark weighting alongside raw numbers, weighted on age, gender, education, region, and party, because the evidence is that Republicans are now slightly less likely than Democrats to take part in polls. Weighting is not a cure: on some questions it changes nothing, and on a few it makes the number worse, so the unweighted figure always stays visible. We'll also publish calibration comparisons against probability-sample polls on identically worded questions, and the validation data for the flagship poll's dimensions. When we miss, we'll say so and show our work.