|
Abstract:
|
In a typical longitudinal comparative clinical trial, for a variety of reasons, some subjects discontinue the study before the end of their planned follow-up period. Restricted maximum likelihood (REML) is commonly used to analyze incomplete longitudinal data. REML can be highly inefficient in the presence of non-normal data; as such, many turn to generalized estimating equations (GEE) and weighted-GEE when normality is untenable. As an alternative, we propose methodology that combines multiple imputation of the missing values with non-parametric (i.e., rank-based) methods. We study Type I error, power, bias, mean square error, width and coverage of 95% confidence intervals to show the robustness and power advantages of our proposed method relative to REML and (W)GEE. A numerical example and simulation results are used to illustrate the methodology and support our conclusions.
|
