Weighting and Adjustment

Why Weighting Is Necessary

No survey sample perfectly mirrors the population. Young people are harder to reach by phone. White college graduates are more likely to participate in online panels. Politically engaged people are more likely to complete surveys. Without correction, these patterns would make every poll biased.

Weighting adjusts for these imbalances by giving more weight to underrepresented groups and less weight to overrepresented ones. If a poll's sample is 40% college graduates but the population is 33%, each college graduate's response is weighted down (multiplied by roughly 0.8) and each non-college respondent is weighted up. The goal is a sample that, after weighting, matches the population on key demographics: age, race, education, gender, and often geographic region and past vote.

Worked Example: Weighting to Targets

Imagine a simplified sample of 1,000 respondents split by education, compared to the known population (census) targets:

Each weight is the population target divided by the sample share. College respondents count as 0.83 of a person; non-college respondents count as 1.12. Check the math: 400 x 0.83 = 332 effective college respondents (33%) and 600 x 1.12 = 672 (67%). The weighted sample now matches the population.

Now suppose non-college voters favor Candidate A 55-45 and college voters favor B 55-45. Unweighted, the sample understates A because it has too many college respondents. After weighting, A's share rises by roughly a point. Real pollsters 'rake' across several variables at once (age, race, education, region, sometimes 2020 vote), iterating until every margin matches its target simultaneously. That is exactly the education adjustment many pollsters were missing in 2016.

The Challenges of Weighting

Weighting works well when the variables you weight on predict the thing you are trying to measure. If young and old people have different political preferences, and your sample underrepresents young people, age-weighting corrects the bias. But if the bias is along a dimension you are not weighting on, the correction fails.

This is exactly what happened in 2016. Many state pollsters weighted for standard demographics but not for education level. Non-college-educated white voters, who strongly favored Trump, were underrepresented in survey samples. After 2016, most pollsters added education as a weighting variable, which improved accuracy in 2018. Yet in 2020 polls still overstated Democratic support, showing that the bias had shifted to a dimension pollsters had not fully solved. The lesson is general: there is always a risk that the next election's bias will be along a dimension that pollsters have not thought to weight on.