A1991
Title: Accelerated generalized estimating equations via inexact quasi-Newton preconditioned conjugate gradient
Authors: Tom Chen - Harvard Pilgrim Health Care and Harvard Medical School (United States) [presenting]
Abstract: Generalized estimating equations (GEE) are a standard tool for marginal regression with clustered or longitudinal data, but the classical Fisher-scoring implementation scales poorly when within-cluster sizes are large, as each outer iteration requires multiple dense linear solves per cluster. An inexact quasi-Newton framework recasts GEE as a nonlinear root-finding problem. The outer iteration is driven by an L-BFGS update in root-finding form, while inner cluster-level solves use preconditioned conjugate gradient with adaptive forcing and structured preconditioners tailored to the working correlation. The framework accommodates flexible additive working correlations, including classical, kernel-based, and basis-expansion components, with a block-coordinate update of the regression and correlation parameters. Simulation studies indicate sizeable computational gains over standard implementations as cluster size grows, with negligible loss of statistical accuracy. The methods are implemented in an open-source R/Rcpp package.
