Generalized Estimating Equations (GEE)

Generalized Estimating Equations (GEE) is a statistical method for analyzing correlated or clustered data, such as repeated measurements on the same individuals over time, observations nested within groups (schools, clinics, families), or multi site studies. GEE provides population averaged (marginal) estimates of the relationship between predictors and outcomes, accounting for the within cluster correlation without requiring full specification of the correlation structure.

Much of the data collected in health, social, and behavioral research has a correlated structure. Patients measured at multiple time points produce repeated measures data. Students nested within classrooms and classrooms nested within schools produce clustered data. Ignoring this correlation and analyzing the data as if all observations were independent usually leads to underestimated standard errors, inflated Type I error rates, and incorrect inferences. GEE provides a practical way to handle these correlations while producing population level estimates that are readily interpretable. It is widely used in epidemiology, public health, medicine, and the social sciences for both longitudinal and cross sectional clustered data.

GEE extends the generalized linear model to handle correlated data. The analyst specifies three components: a link function (connecting the expected value of the outcome to the linear predictor, such as a logit link for binary outcomes or an identity link for continuous outcomes), a variance function (describing the relationship between the mean and the variance of the outcome, based on the assumed distribution family), and a working correlation structure (describing the assumed pattern of correlation among observations within a cluster).

Common working correlation structures include independent (assumes no within cluster correlation), exchangeable (assumes all pairs of observations within a cluster have the same correlation), autoregressive (assumes correlation decreases with the time lag between observations), and unstructured (estimates a separate correlation for each pair of observations). The choice of working correlation structure affects the efficiency of the estimates but, importantly, GEE estimates are consistent (converge to the correct value) even if the working correlation structure is misspecified, provided the mean model is correctly specified. The consistency comes from the estimating equations themselves; valid standard errors under a misspecified working correlation come from the sandwich (robust or empirical) variance estimator, which estimates the variance from the observed within cluster residuals rather than trusting the working correlation, and which relies on a large number of clusters.

GEE produces population averaged estimates, which answer questions about how the average response changes across levels of the predictor in the population. This contrasts with subject specific estimates (produced by mixed effects models), which describe how the response changes within an individual. For linear models, population averaged and subject specific estimates are the same, but for non linear models such as logistic regression, they generally differ and answer different questions.

GEE is appropriate when the research question concerns population level associations rather than individual level trajectories, when the data are clustered or have repeated measures, when the outcome is continuous, binary, count based, or from another distribution in the exponential family, and when the analyst wants estimates that are resistant to misspecification of the correlation structure. It is particularly popular in epidemiological cohort studies with repeated outcome measurements and in multi site clinical studies.

A common error is using GEE with a very small number of clusters (fewer than about 30 to 40, a common rule of thumb), which can produce biased standard errors even with the sandwich estimator. Small sample corrections (such as those by Mancl and DeRouen or Fay and Graubard) should be applied in these situations. Another mistake is choosing a working correlation structure that is extremely different from the true structure, which reduces efficiency (wider confidence intervals) even though consistency is maintained. Some researchers do not report the working correlation structure or the number of clusters, making the analysis difficult to evaluate. Confusing population averaged estimates with subject specific estimates is another common error; the two answer different questions for non linear models. Finally, GEE assumes that data are missing completely at random (MCAR); when data are missing at random (MAR), weighted GEE or multiple imputation approaches may be needed.

A longitudinal study follows 800 patients with diabetes at 25 primary care clinics, measuring HbA1c levels every 6 months for 3 years (7 measurements per patient). The researchers use GEE with an exchangeable correlation structure and identity link to assess whether a clinic based educational intervention reduces HbA1c over time. The analysis, with sandwich standard errors and a Mancl and DeRouen small sample correction (25 clinics is fewer than the roughly 30 to 40 clusters often recommended for the sandwich estimator), shows that intervention patients have a mean HbA1c 0.4 percentage points lower than control patients at 3 years (95% CI: 0.2 to 0.6 percentage points, p less than 0.001), adjusting for baseline HbA1c, age, and diabetes duration. Treating each clinic as a cluster, the sandwich standard errors account for both the clustering within clinics and the correlation among repeated measures within patients, even though a single exchangeable correlation only approximates that two level pattern.