Statistical Models for Identification of Treatment-sensitive Subgroups Based on Longitudinal Outcomes in Clinical Trials

Abstract

In randomized clinical trials, an average effect of a treatment is often evaluated over all patients enrolled. But in the era of personalized medicine, there is an increasing interest in identification of patients who may benefit from or be sensitive to a specific type of treatment. Various statistical methods have been proposed recently in the literature to help clinical researchers to identify treatment-sensitive subgroups of patients based on baseline covariates or biomarkers and assess the difference in the treatment effect between different subgroups of patients. Time to an event, such as overall or disease-free survival, is the clinical outcome used to identify treatment-sensitive subgroups in majority of these methods proposed in the literature. Longitudinal outcomes, such as measurements on the quality of life of patients at different time-points, are also important outcomes collected in clinical trials. There are, however, very few statistical approaches available which can be used to identify treatment-sensitive subgroups based on longitudinal outcomes. All current methods proposed in the literature require a subjective definition of subgroups. This thesis is devoted to the development of statistical methods for identification of treatment-sensitive subgroups based on longitudinal outcomes. Three new statistical models are introduced based on three different clinical scenarios. Specifically, a threshold linear mixed model is introduced when the longitudinal outcomes are assumed to follow a normal distribution and there is only a single continuous covariate available to define the subgroups. When the longitudinal outcomes may subject to potential floor and ceiling effects but there is still a single continuous covariate available, a threshold mixed-effects Tobit model is introduced. Finally, when there are multiple covariates available, a generalized single-index linear threshold model is introduced by assuming the marginal distribution of the longitudinal outcomes is in an exponential family. Statistical procedures are proposed to make statistical inferences on the parameters in these models, which can be used to identify and assess treatment-sensitive subgroups. All of the proposed models and inference procedures are assessed through simulation studies, as well as applications to the analysis of data from randomized clinical trials.