Survey weighting explained: raking to population targets
Published November 16, 2026 · 7 minute read
No survey sample is a perfect miniature of the population. Weighting gives each response a weight so that the weighted sample matches what we know about the population, which reduces bias in the results.
Why weight?
Some groups respond more readily than others: in most resident and customer surveys, older people and women are over-represented and young men under-represented. If those groups answer differently, the raw results are biased. Weighting corrects for differences on the variables you weight by.
How raking works
Raking only needs each variable’s population distribution, not the full cross-classification, which is why it is the standard method when targets come from different sources.
- Choose the variables you know population figures for, such as age group, gender and ward (from a census or customer database).
- Start every response with a weight of 1.
- Adjust the weights so the sample matches the first variable (for example multiply everyone aged 18 to 34 by target share ÷ sample share).
- Do the same for the next variable, then the next. Each step disturbs the earlier ones slightly.
- Repeat the cycle until all variables match their targets closely (usually within a few iterations).
Choosing targets
- Weight on variables that relate both to response and to the answers (age, gender and area are the usual starting point).
- Use reliable, current population figures for the same population your survey covers.
- Merge small categories: weighting on a cell with five responses makes those five people count far too much.
Check the weights
- Look at the minimum and maximum weight; trim extremes (for example to between 0.2 and 5) if a few responses dominate.
- Report the design effect and effective sample size, and use the effective sample size for margins of error.
- Compare weighted and unweighted results: large differences tell you the sample was unbalanced and the weighting is doing real work.
Limits
Weighting only corrects for the variables you weight on. If people who respond differ from those who do not in ways not captured by those variables, some bias remains. Good sampling and reminders reduce how much the weights have to do.