MerQuR+Statistics Glossary

Statistics Glossary → Cluster-robust (clustered) standard errors

Cluster-robust (clustered) standard errorsconcept

Sandwich-type standard errors that allow observations within the same cluster (school, patient, firm) to be dependent.

Classical standard errors assume that all observations are independent. When pupils in the same school or repeated measurements on the same person resemble each other, this assumption fails and the standard errors usually come out too small. The cluster-robust approach first sums the scores (or residual contributions) within each cluster and then builds a sandwich variance that treats clusters as the independent units; no assumption about the form of the within-cluster correlation is needed.

It does not change the coefficient estimates, only their uncertainty. It relies on a large number of clusters: with few clusters (roughly fewer than 30-50) the standard errors can still be too small, which is why small-sample corrections such as G/(G − 1) and a t distribution with G − 1 degrees of freedom are used. Unlike mixed models, which model the clustering explicitly, it corrects for dependence only at the inference stage.

An everyday example

The effect of an education programme is studied with 2000 pupils in 40 schools. The programme is delivered at school level, so pupils in the same school have similar outcomes. Treating pupils as independent gives a standard error of 0.8 points for the programme effect, while clustering by school gives 2.1 points; the first calculation overstated the precision.

In MerQur

'Cluster variable (cluster-robust SE)' in Logistic Regression; the Clustered choice of 'Standard errors' and the 'Cluster column' in Instrumental Variables Regression; the unit- or time-clustered choices of 'Standard errors' in Panel Data Regression; 'SE clustered by unit' in Difference-in-Differences; 'Cluster (person) column - robust SE' in Conditional Logit; 'Cluster (robust SE)' in Cox Regression and 'Within-subject robust SE' in Time-dependent Cox Regression; the required 'Cluster' in Cluster-Robust Cox; the sandwich SE with the group (cluster) variable in GEE; and, for matching in Propensity Score Analysis, the SE clustered by matched set.

Analyses where this concept appears

Logistic RegressionGeneralized Estimating Equations (GEE)Panel Data RegressionInstrumental Variables Regression (2SLS)Propensity Score AnalysisDifference-in-Differences (DiD)Conditional Logit (Choice Model)Cox Regression (Proportional Hazards Model)Time-dependent Cox RegressionCluster-Robust Cox (Marginal and Frailty Models)

Related concepts

Robust (sandwich) standard errors (HC0-HC3)GEE working correlation and marginal models