Why Polls Disagree
Two pollsters survey the same race in the same week and publish different leaders. Neither is necessarily wrong. Polling involves a chain of judgment calls, and each one can move the numbers.
The baseline: sampling noise
Before any methodological choices enter, two polls of the same race will differ simply because each interviews a different random sample. That alone produces spread of a few points between honest, competent polls. Disagreement only needs explaining when it is larger or more persistent than chance would produce — and that is where the choices below come in.
Mode effects
How a pollster reaches people shapes who answers. Live telephone interviews, online panels, text-message surveys, and automated calls each reach somewhat different slices of the public — by age, by how online they are, by willingness to talk to a stranger. People can also answer differently when speaking to a human than when tapping a screen. Two firms using different modes can both execute well and still see different electorates.
Likely-voter screens
A poll of all adults, of registered voters, and of “likely voters” are three different surveys. Deciding who counts as likely to vote is a judgment call: some firms ask respondents directly, some use vote history, some build statistical models. Because the people screened out often lean differently from those kept in, the screen itself can shift the topline — especially in lower-turnout midterm elections, where guessing who shows up is hardest.
Weighting choices
No raw sample looks exactly like the electorate, so every pollster adjusts — weighting respondents so the sample matches known population benchmarks like age, sex, education, race, and region. Firms differ on which benchmarks to use and how hard to correct, and contested choices (such as weighting on past vote or party) can move results by several points. Weighting is necessary; it is also a place where reasonable experts diverge.
House effects
Add those choices up and many firms show a persistent lean relative to the polling field — consistently a bit better for one party than the average of everyone else. This is called a house effect. It is usually a byproduct of method, not intent, and a firm with a house effect can still be informative — its movement over time often tracks real change even when its level sits high or low. Our average doesn't apply house-effect corrections, because doing so requires modeling assumptions we prefer to avoid; we say so plainly on the methodology page.
Timing
Polls are conducted over field windows of days, sometimes a week or more. Two polls “released” the same day may have interviewed people at different times, straddling a debate, a major news story, or a campaign event. When opinion is moving, part of the disagreement is simply that the polls measured different moments.
What to do with disagreement
Treat it as information, not malfunction. A wide spread across pollsters is itself a signal that the race is hard to measure — and a reason to hold conclusions loosely. Averaging many polls smooths over individual choices, which is why we publish a weighted average rather than chasing single polls. But an average can't fix a miss the whole field shares; how we account for that is covered in How We Handle Uncertainty.