Survey weighting

Survey weighting assigns each respondent a numeric weight so that, once applied, the weighted sample's characteristics, age, gender, region, education, and so on, match known totals for the target population, correcting for groups that were over- or under-represented in who actually responded.

Common methods include post-stratification, weighting to match known totals within a few categories, and raking, or iterative proportional fitting, which adjusts to match several marginal totals at once when the full joint distribution across categories is unknown. Weights are typically capped so a handful of respondents cannot dominate the results. This differs from designing a good sampling frame up front: weighting corrects an imperfect sample after the fact, while frame design tries to prevent imbalance before data collection even begins.

Nearly all public opinion and election polls are weighted, because no survey, even one built on a random sampling frame, perfectly mirrors the population once real-world response rates and nonresponse bias are accounted for. More advanced approaches, such as multilevel regression and poststratification, extend the same idea to produce estimates for small subgroups. The core pitfall is that weighting can only correct for the variables it is built on; a sample skewed on a trait not included in the weighting scheme, such as political engagement, stays skewed no matter how well the included demographics are balanced.