Home Statistical Dictionary Raking
Raking
An iterative weighting procedure that adjusts survey weights to match known marginal population totals for two or more categorical variables separately, without requiring their full joint (cross-classified) distribution to be known -- also called iterative proportional fitting.
In Plain English
Raking solves a practical problem with post-stratification: sometimes you know the population's totals for each variable separately, say, the true breakdown by age and, separately, the true breakdown by education level, but you don't know the full joint breakdown (how many people of each specific age fall into each specific education category), which is exactly what post-stratification would need. Raking gets around this by iterating, adjust the weights to match the age totals, then adjust again to match the education totals (which will have thrown the age totals slightly off), then re-adjust for age again, and so on back and forth, cycling through each variable in turn until the weights settle down and approximately satisfy all the marginal totals simultaneously.
Definition
Raking, formally iterative proportional fitting, adjusts survey weights to match known marginal population totals for two or more categorical variables separately, rather than requiring their joint cross-classified totals as post-stratification does; the algorithm alternates, adjusting weights to match the first variable's marginal totals, then adjusting to match the second variable's marginal totals (which slightly disturbs the fit to the first), cycling repeatedly through each variable's margins in turn until the weights converge to values that approximately satisfy all the specified marginal constraints simultaneously. Raking is especially useful when the full joint distribution across several variables is unknown or would require sample cells too sparse to weight reliably, but only known marginal totals for each variable are available, it is a specific, iterative instance of the broader calibration estimation framework, converging to the maximum-entropy solution consistent with the specified marginal constraints, and is the standard technique used to weight many opt-in and probability-based public opinion polls to known census demographic margins.
Properties
- Raking's key practical advantage over post-stratification is that it only needs marginal population totals for each variable separately, not their full joint cross-classification, this makes it usable in situations, common in practice, where census or administrative data provide reliable marginal totals for age, sex, education, and region separately, but not a reliable joint table across all of them at once.
- The raking algorithm is not guaranteed to converge to a unique solution in every configuration, and can behave poorly (producing extreme weights) if the sample's composition is very different from the population's on the variables being raked, in practice raking weights are commonly capped or trimmed to prevent a small number of respondents from receiving disproportionately extreme influence.
- Raking is a standard, near-universal weighting step in modern opinion polling, especially for online opt-in panels, which typically have no controlled probability sampling design to rely on at all, raking to known census margins (age, sex, education, region, sometimes party identification or past vote) is often the primary tool used to make such panels resemble the target population on those measured dimensions.
At a Glance