| General linear mixed models provide a useful method for analyzing repeated measure data encountered in bioequivalence studies. Because of its flexibility in modeling the complicated correlation structure among observations from the same subject, it presents a modeling framework for the assessment of individual bioequivalence under a wide range of replicated crossover designs.;The concept of individual bioequivalence has been studied extensively in the literature since Anderson and Hauck (1990) published a seminal article in which they proposed a statistical linear mixed-effects model to study drug interchangeability. A number of publications have described various approaches to the assessment of individual bioequivalence (Sheiner 1992; Schall & Luus 1993; Hsuan & Holder 1993; etc.). All these approaches, including that proposed in the recent Food and Drug Administration (FDA) draft guidance (FDA 1997, 1999), used the linear mixed-effects model as the basis of statistical estimation and testing.;In spite of its widespread use, the linear mixed-effects model has limitations. As a special member of general linear mixed models, the linear mixed-effects model provides only one specific way to model the correlation structure among the multiple observations of the same subject. In this dissertation, we demonstrate that the mixed-effects model implicitly assumes a correlation structure that is not always adequate for real data from bioequivalence studies. We propose a class of general linear mixed models that facilitates valid statistical assessment of individual bioequivalence (EBE) under a wide range of replicated crossover designs. We demonstrate that this class of general linear mixed models has wider applicability than the mixed-effects model. We analyze two FDA published data sets and perform simulation studies to illustrate these points. We propose three different IBE testing procedures: the first is based on the restricted maximum likelihood estimate for the variance components and the asymptotic theory; the second is an extension of the small sample approach proposed by Hyslop, Hsuan and Holder (Statistics in Medicine, In Press); the third is based on the method of moment estimator and the bootstrap procedure. We perform simulation studies to investigate the adequacy of these models, and to compare the merits of these testing procedures. |
