A poll can interview exactly 1,000 people and still contain too many graduates, too few older people or an unusual balance of previous party supporters. Weighting changes the contribution each respondent makes to the result, using information about the population the survey is intended to represent.
Weighting can improve a sample without making it perfectly representative. The adjustment depends on the characteristics measured, the targets chosen and the people who actually took part.
A calculation you can follow
Consider an illustrative population split equally between two groups. A sample contains 600 people from the first group and 400 from the second. To give each group equal influence, respondents in the first receive a weight of 500/600, or about 0.833, and those in the second receive 500/400, or 1.25.
Now give Party A 40% support in the first group and 20% in the second. Its unweighted result is 320 supporters out of 1,000, or 32%. After adjustment, its weighted support is 200 plus 100 out of a weighted total of 1,000: 30%.
No answers were changed. The composition of the calculation changed. Our guide to reading polling tables explains how the weighted and unweighted bases expose this distinction.
Which targets are used?
Age, education, region and past voting behaviour are common possibilities, but organisations choose different combinations and sources. Targets can be interlocked, such as age within sex, rather than adjusted only as separate totals.
Ipsos’s explanation of its 2025 methodology identifies a range of demographic and other variables. Published methods should be read at their stated date: a list used for one election is not necessarily the list being used today.
Past vote adds useful political information, but someone’s recollection can differ from their actual behaviour. An adjustment which treats misremembered answers as accurate risks creating a new imbalance. A pollster’s treatment of these answers belongs alongside its handling of undecided respondents when judging the final estimate.
Why matching a few totals is not enough
A sample can match the population’s age and education totals while still containing an unusual mixture within those categories. For example, the younger graduates interviewed might be much more politically engaged than younger graduates generally.
Weighting cannot directly correct a characteristic which was never measured or whose population distribution is unknown. Research by Hartman and Huang examines sensitivity to missing adjustment variables; their work helps explain why a tidy demographic table is not proof that every relevant imbalance has been removed.
Precision and very uneven weights
When a few respondents carry much more influence than the rest, the nominal interview count can overstate the information available. A widely used diagnostic is Kish’s effective sample size: (sum of weights)² divided by the sum of squared weights.
For illustration, two equally weighted people give an effective size of two. Weights of one and three give an effective size of 16/10, or 1.6. This diagnostic isolates unequal weighting; it is not a complete measure of every design effect or a substitute for a model’s uncertainty estimates.
Comparing adjusted polls
Do not judge a poll by whether weighting helped or hurt a preferred party. Ask whether the targets suit the population, whether the organisation explains changes and whether its approach performs credibly against evidence.
Different weighting choices can contribute to pollster house effects. They also affect how sampling uncertainty should be interpreted. A bigger unweighted sample, on its own, settles neither question.
Sources and further reading
- Ipsos: sampling and weighting changes
- Hartman and Huang: limits of adjustment for observed characteristics
- British Polling Council: published weighting information
Reviewed 7 October 2026. Worked examples are illustrative unless an election or poll is identified.