Statistics Glossary → Domain (subpopulation) estimation
Domain (subpopulation) estimationconcept
Estimating a mean, proportion or total for a subpopulation (domain), such as women or a particular region, from survey data collected with a complex sample design.
The obvious route is to filter the data to the subgroup of interest and analyse only those observations. In a stratified or clustered sample, however, the subgroup's size is a random quantity that was not fixed in advance; filtering can drop entire strata or clusters and breaks the design information. Standard errors are then computed wrongly (usually too small). Correct domain estimation keeps the whole sample; observations outside the domain are handled so that they contribute nothing to the estimate (their values are treated as zero), and the variance is computed over the full design.
For each domain the estimate, design-based standard error, confidence interval, coefficient of variation (CV) and design effect (DEFF) can be reported, and differences between domains are tested with contrasts that use the same design information. When a domain's sample is very small, direct estimates become unstable and the CV rises; model-based small-area estimation methods (e.g. Fay-Herriot) then come into play, but these require different assumptions. Although the idea resembles 'subgroup analysis' in regression, it calls for a calculation specific to survey sampling theory.
An everyday example
In a nationwide survey of 10,000 households you want to estimate the average income of rural households only. If you copy the rural households into a separate file and analyse them, part of the information on which strata and clusters the sample came from is lost, and the standard error looks smaller than it really is. Domain estimation produces an estimate for rural households alone while keeping the design information of all households.
In MerQur
It is offered through the 'Subgroup (by)' parameter in the Weighted Mean, Weighted Frequency and Weighted Total analyses; comparisons between domains are given by the 'Differences between domains (contrasts)' extra output.
Analyses where this concept appears
Weighted Mean (Survey Means)Weighted Frequency (Survey Freq)Weighted Total (Survey Total)Weighted Linear Regression (Survey Regression)Weighted Logistic Regression (Survey Logistic Regression)Survey-PHREG (Weighted Cox)