Generalized Estimating Equations

Generalized Estimating Equations: Concept Architecture

Orientation and learning roadmap

This page follows repeated or clustered outcomes from the study design to a population-level regression estimate. It then shows how within-cluster correlation enters estimation, illustrates the interpretation with hospital outcomes, and identifies the assumptions that matter for confidence intervals.

Choose the outcome and independent clusters

Start by identifying the independent units, such as patients with repeated observations or clinics containing multiple patients. Measurements within one unit may be correlated, so treating every observation as an independent replicate can distort the estimated precision of a treatment effect.

Specify the outcome distribution and link for the marginal mean, along with treatment, time, and prespecified covariates. For example, a logit link can model the average probability of a hospital readmission across the target population, while an identity link can model a mean cost difference when its assumptions and predictions are appropriate. The original Liang and Zeger longitudinal-data paper develops estimating equations for this marginal regression setting.

How the estimating equation uses correlation

The mean model relates each observation's expected outcome to its covariates. A working correlation describes how outcomes within the same cluster tend to move together and helps weight their contributions to the regression estimate.

Writing $Y_i$ for cluster $i$'s outcome vector, $\boldsymbol\mu_i(\boldsymbol\beta)$ for its modelled marginal means, $D_i=\partial\boldsymbol\mu_i/\partial\boldsymbol\beta^{\mathsf T}$, and $V_i$ for its working covariance matrix, the estimate solves:

$$U(\boldsymbol\beta)=\sum_{i=1}^{m}D_i^{\mathsf T}V_i^{-1} \{Y_i-\boldsymbol\mu_i(\boldsymbol\beta)\}=\boldsymbol0.$$

Common working choices include independence, exchangeable correlation across all within-cluster pairs, and correlation that decays across successive measurements. A correctly specified marginal mean and independent clusters support large-sample consistency under appropriate regularity conditions even if the working correlation is misspecified; efficiency may suffer, and uncertainty should use a suitable cluster-robust sandwich variance estimate. Neither the choice of correlation nor the robust variance fixes confounding or an incorrectly specified mean model.

Worked readmission interpretation

Imagine 100 patients per strategy, each observed at two planned follow-ups, and suppose the observed readmission proportions are approximately 20% under an intervention and 30% under usual care at each visit. An illustrative marginal risk difference is $0.20-0.30=-0.10$, or ten fewer readmissions per 100 patient-visits on average across those observed visits.

Because each patient contributes two measurements, the 400 visit records are not 400 independent patients. A patient-clustered GEE can estimate a covariate-adjusted marginal contrast and account for within-patient dependence in its uncertainty; the observed proportions alone do not determine its adjusted coefficient or confidence interval. If time and treatment interact, report the visit-specific contrasts rather than forcing one constant effect across both visits.

Marginal effect versus subject-specific effect

GEE coefficients describe the regression mean across the population represented by the model and analysis weights. A mixed-effects model instead commonly introduces subject or cluster-specific random effects, so its conditional coefficients answer a different question.

For binary outcomes with a nonlinear link, a population-averaged odds ratio and a conditional odds ratio generally differ even when both models describe the same underlying data. Choose the target estimand before comparing models or translating a result into a decision model. In health economic work, an absolute predicted risk difference or a standardised marginal prediction may be easier to apply to population event counts than an unqualified odds ratio.

Precision, missingness, and study design

Cluster-robust variance estimates rely on an adequate number of independent clusters for their usual large-sample approximation. With few randomised clinics, standard sandwich standard errors can be too small, so a justified small-sample correction and appropriate inference method are important.

The small-cluster GEE simulation study examines this finite-sample issue. Standard unweighted GEE also does not make informative dropout harmless: assess why outcomes are missing and consider methods such as inverse-probability weighting or imputation under their stated assumptions. In a cluster-randomised study, the clinic is the independent unit for variance estimation even if many patients and visits are measured within it.

Use and validation in economic analysis

GEE can estimate average changes in utilisation, clinical events, quality-of-life measures, or costs when observations repeat or cluster. Map the link-scale coefficient to the quantity needed by the economic model, such as an adjusted mean difference or population-level predicted probability, and keep the relevant follow-up period and target population clear.

Report the number and size of independent clusters, outcome family, link, working correlation, missing-data strategy, variance estimator, and any prespecified effect modification. Inspect fitted means, residual patterns, influential clusters, and sensitivity to reasonable correlation and missingness choices. Correlation-aware uncertainty improves inference about a specified association; causal interpretation still depends on the study design and identification assumptions.