MerQuR+Statistics Glossary

Statistics Glossary → Primary sampling unit (PSU) and cluster sampling

Primary sampling unit (PSU) and cluster samplingconcept

In multistage sampling, the units selected at the first stage (e.g. neighbourhoods, schools); because observations from the same unit resemble each other, variance is computed over these units.

In national surveys, sampling individuals directly at random is expensive; clusters such as neighbourhoods, villages or schools are selected first (the primary sampling units), and households or individuals are then taken within them. People in the same cluster share common conditions and so resemble one another (intraclass correlation), which reduces the independent information each new observation adds. As a result, standard errors under cluster sampling are usually larger than in a simple random sample of the same size; this inflation is summarized by the design effect (DEFF).

Design-based variance estimation (Taylor linearization) uses PSU totals and treats the first stage as if sampled with replacement (the 'ultimate cluster' approach); the variance of later stages is not modelled separately. The degrees of freedom depend not on the number of observations but on the number of PSUs minus the number of strata, so a small number of PSUs gives wide confidence intervals despite thousands of observations. If a stratum contains only one PSU, its variance contribution cannot be computed; it is either treated as contributing nothing or centred at the overall average. Ignoring the PSU information makes standard errors look smaller than they are.

An everyday example

A health ministry measures childhood vaccination coverage by selecting 30 neighbourhoods and visiting 20 households in each. In some neighbourhoods a clinic is nearby and coverage is uniformly high; in others it is uniformly low. The 600 households do not carry as much information as 600 independently chosen ones; with a design effect of 2.5, the sample is worth about 240 independent households, and the standard error grows accordingly.

In MerQur

Available as 'Cluster / PSU' and 'Single-PSU stratum' (No contribution / Centered at the average) in weighted mean, weighted frequency, weighted total, and weighted linear and logistic regression. In Survey-PHREG the score residuals of observations in the same PSU are summed to compute the variance.

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)

Related concepts

Stratified sampling (design stratum)Design effect (DEFF)