Weighting Approaches for Longitudinal Data with Time-Dependent Cluster Sizes
Generalized estimating equations (GEEs) are commonly used in the modelling of correlated data. However, in the presence of informative cluster sizes, estimates obtained using GEEs may be biased. In order to correct for this bias a weighted GEE may be used. Previous research has extended the use of weighted GEEs to a longitudinal setting but requires that cluster sizes remain constant over time. In this thesis, two new weighting schemes are investigated to allow for valid parameter estimation in a longitudinal setting where cluster sizes are informative and may change over time. Specifically, this thesis considers weighting by the inverse of the time-dependent cluster size, and by the total number of observations for a given cluster. Through Monte Carlo simulation, the performance of traditional GEEs, GEEs under previously proposed weighting schemes, and these two new models are compared. Results of these studies show that weighting by the total number of observations results in unbiased parameter estimates with excellent coverage.