Resampling methods for longitudinal data analysis

Abstract

Resampling methods for the regression parameter estimates in GEE models for longitudinal data analysis are proposed in this thesis. First type of method is smooth bootstrap, a random perturbation to the estimating algorithm, which provides a simple way to produce bootstrapped copies of parameter estimates. Two versions of such smooth bootstrap methods are provided, one is robust to the misspecification of the correlation structure and the other one is model-based. The second type of resampling schemes proposed is based on the studentized estimating function, and one first-term corrected studentized estimating function derived from the Edgeworth expansion, which can gain higher order distribution approximation. Simple perturbation methods are discussed for general parameter estimation. Specific methods with less computation are proposed for constructing the confidence intervals. In simulation studies, the proposed resampling methods generally outperform competitive estimators in terms of variance estimation and confidence interval construction.