Generalised Estimating Equations (GEE)

Description

A regression method for clustered or repeated outcomes that estimates population-average effects with a working correlation and robust sandwich standard errors.

When the same user, account or team contributes many observations, ordinary regression understates uncertainty. GEE fits a marginal model for the mean (any generalised linear model) and solves estimating equations that include a working correlation matrix, usually exchangeable or AR(1). Standard errors come from the sandwich estimator, which stays consistent even if the working correlation is wrong. The key decision is the target of inference. GEE coefficients describe the average effect across the population; a random-effects (mixed) model describes the effect for a given individual. For linear and Poisson models the two coincide, but for logistic and probit models population-average slopes are attenuated towards zero, more so as within-cluster correlation grows. For binary data the admissible correlation is bounded by the probabilities, so an exchangeable structure can be impossible with large clusters.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. Identify the clusters, the covariates that vary within clusters and those that are constant within them.
  2. Decide whether the question is about population averages (GEE) or within-individual change (mixed model).
  3. Choose a working correlation that reflects the design: exchangeable for members of a group, AR(1) for repeated measures over time.
  4. Fit the model by iteratively reweighted least squares and use the sandwich variance for Wald tests and confidence intervals.
  5. For binary outcomes check that the estimated correlation lies within admissible bounds and expect slopes smaller than a random-intercept logit would give.
  6. Build the model with the usual selection, scale and interaction checks, and compare with a random-effects fit if subject-specific interpretation is also needed.

Sources

  • Applied Logistic Regression (Hosmer & Lemeshow, 2000)
  • Mixed Models: Theory and Applications (Demidenko, 2004)
  • Beyond t Test and ANOVA (Yu et al., 2022)
  • Bayesian Statistical Modelling (Congdon, 2006)

Also known as

  • GEE
  • Population-averaged model
  • Population-averaged logistic model
  • Marginal model
  • Liang-Zeger GEE
  • Sandwich variance
  • Marginal versus subject-specific model

Categories

  • Analytics & Behaviour
  • Statistics & Modelling