MerQuR+Statistics Glossary

Statistics Glossary → GEE working correlation and marginal models

GEE working correlation and marginal modelsconcept

An approach for repeated or clustered data that approximates within-cluster dependence with a working correlation matrix and estimates population-averaged (marginal) effects.

Generalized estimating equations (GEE) extend the GLM to correlated observations from the same person or cluster. The analyst supplies a 'working' assumption about the form of the within-cluster correlation: independence (no correlation), exchangeable (all pairs equally correlated), AR(1) (correlation decaying with the time gap) or unstructured. The coefficients are found from estimating equations that use this structure, while the standard errors are computed with the sandwich estimator.

GEE's key property is that the coefficients and sandwich standard errors remain consistent even if the working correlation is chosen wrongly; a good choice only improves efficiency (smaller standard errors). The coefficients have a marginal (population-averaged) interpretation: they describe how the population mean changes, not how 'an average individual' changes, and they differ numerically from the subject-specific (conditional) coefficients of mixed models, especially with a logit link. Because GEE is not likelihood-based, QIC is used instead of AIC, and missing data are assumed to be missing completely at random; with few clusters the sandwich standard errors are too small and corrections such as Mancl-DeRouen are needed.

An everyday example

A school health study follows tooth decay each year among pupils in 40 schools. Pupils in the same school share similar drinking water and canteen conditions, so their outcomes are dependent. An exchangeable-correlation GEE estimates that a fluoride toothpaste programme reduces the odds of decay by 30% on average across the whole pupil population, with a standard error that accounts for within-school similarity.

In MerQur

The basis of the Generalized Estimating Equations analysis; within-cluster dependence is modelled with a working correlation matrix and coefficients are reported with sandwich standard errors. The cluster-robust (marginal) Cox model applies the same marginal approach to survival data.

Analyses where this concept appears

Generalized Estimating Equations (GEE)Cluster-Robust Cox (Marginal and Frailty Models)

Related concepts

Covariance (correlation) structure (UN, VC, AR(1), exchangeable)Cluster-robust (clustered) standard errors