Generalized Estimating Equations (GEE)

GEE handles regression for correlated/clustered data — longitudinal studies, repeated measures, or clustered designs where observations within a subject/cluster are not independent.


Step 1 — Provide your data

Each row is one observation:

  • Subject ID — cluster identifier
  • Time Point — when observed
  • Outcome — dependent variable
  • Covariates — independent variables

Supports Excel Import, Sample Data Generator, and Manual Entry.


Step 2 — Configure the model

Outcome family:

Family Use when
Gaussian Continuous outcomes
Binomial Binary outcomes (0/1)
Poisson Count outcomes

Link function:

Link Typical pairing
Identity Gaussian
Logit Binomial
Log Poisson

Working correlation structure:

Structure Assumption
Independence No within-cluster correlation
Exchangeable Uniform correlation (0.3)
AR(1) Decay with time distance (0.5^k)
Unstructured All pairwise estimated

Step 3 — Results

  • Coefficients table — Estimate, SE, Z-value, p-value, 95% CI per variable
  • Working Correlation Matrix
  • Model fit — Quasi-likelihood, AIC, BIC
  • Convergence status — converged, iterations, final tolerance
  • Model summary — subjects, observations, time points, family/link/correlation
  • Interpretation — adaptive plain-language summary

A UniversalChatBot is available for discussion.


Statistical methods used

Model

g(E[Y_ij]) = beta0 + beta_time * t_ij + sum(beta_k * X_ijk)

Estimation — Iterative least squares, max 20 iterations.

Standard errors — Conservative estimates accounting for within-cluster correlation.

Confidence intervals and p-values (Wald)

95% CI: beta +/- 1.96 * SE, Z = beta / SE, p = 2*(1 - Phi(|Z|))

Quasi-likelihood

Family Formula
Gaussian -0.5 * sum((y - mu)^2)
Binomial sum(y*log(mu) + (1-y)*log(1-mu))
Poisson sum(y*log(mu) - mu)

Key property: GEE produces consistent population-average estimates even when the working correlation structure is misspecified.