Generalized Estimating Equations (GEE)

Core Idea

Generalized Estimating Equations (GEE) provide population-average estimates for correlated data without specifying the full likelihood of the data. Unlike mixed-effects models, which model subject-specific effects and require distributional assumptions on random effects, GEE requires only a correct specification of the mean model and a "working" correlation structure (which need not be correct). The sandwich (robust) variance estimator provides valid standard errors even when the working correlation is misspecified, making GEE remarkably robust. GEE estimates marginal (population-average) effects — the effect of a predictor averaged over all subjects — which differs from the conditional (subject-specific) effects estimated by mixed-effects models, particularly for nonlinear models like logistic regression.

Explainer

Mixed-effects models and GEE both handle correlated data, but they approach the problem from different philosophical starting points. Mixed-effects models specify a complete probability model — the distribution of the response conditional on fixed and random effects — and estimate both population parameters and subject-specific deviations. GEE takes a more modest approach: it specifies only the mean model (how the expected response relates to predictors) and the variance function, then uses a "working" correlation matrix to account for within-cluster correlation. No full likelihood is specified, and no random effects distribution is assumed.

The working correlation in GEE is a pragmatic tool, not a belief about the data. You specify a structure (exchangeable, autoregressive, unstructured) that you think approximates the true correlation. GEE then uses this structure to form the estimating equations. If you get the correlation right, you gain efficiency (smaller standard errors). If you get it wrong, the point estimates are still consistent, and the sandwich (robust) variance estimator corrects the standard errors by using the empirical covariance of the residuals. This double protection — consistency regardless of the working correlation, plus robust standard errors — makes GEE extremely popular in practice.

The critical conceptual distinction is between marginal and conditional effects. In a linear model, these are the same: the population-average slope equals the individual-level slope. In nonlinear models (logistic regression), they differ. The GEE marginal odds ratio compares the average log-odds between populations with different covariate values. The mixed-effects conditional odds ratio compares log-odds within an individual when the covariate changes. Because averaging a nonlinear function produces a different result than applying the function to the average, the marginal effect is attenuated (closer to null) relative to the conditional effect. Neither is wrong — they answer different questions. If you want to know the effect of a policy change on a population's disease rate, GEE gives you the right answer. If you want to predict an individual patient's risk, mixed-effects models are more appropriate.

GEE's main limitation is its reliance on large numbers of independent clusters. The sandwich estimator derives its properties from averaging across many clusters, and with fewer than 40 clusters, it can substantially underestimate standard errors. Small-sample bias corrections (e.g., the Mancl-DeRouen or Fay-Graubard adjustments) help but do not fully resolve the problem. Additionally, GEE handles missing data poorly — it assumes data are missing completely at random (MCAR), whereas mixed-effects models are valid under the weaker missing at random (MAR) assumption. When dropout or missing data patterns are related to the outcome, inverse-probability-weighted GEE or pattern-mixture models may be needed.