Flexible modeling tools for hierarchical and incomplete data, with applications in comet assays and clinical trials

Abstract

In this thesis, some modeling issues and design aspects that arise in toxicological studies and clinical trials are addressed. The text is structured in two parts. The first part of the thesis is motivated by a comet assay, a toxicological study design to assess DNA damage, which has been a standard tool in the pharmaceutical industry for the assessment of the safety of potential new drugs. In this part, a flexible modeling tool is proposed addressing the different modeling issues and an estimation technique is explored. In particular, various models, i.e., a flexible model for hierarchically clustered and overdispersed outcomes, the multivariate extension through a joint modeling technique, mixture models, zero-inflated models, and the use of Gaussian variational approximation are discussed in this part. The second part is related to incomplete data in clinical trials. It is very common practice for patients to drop out from a study due to different reasons. Such missing data, in general, have a potential to affect/distort inferences drawn. Some trials allow for data-driven adaptation when the dropout rate is high. The second part of the thesis focuses on the impact of such data driven adaptations in some characteristics. In particular, the type I error rate associated with dose group switching is assessed when the primary analysis is in terms of a longitudinal outcome. The error rate is assessed through a simulation study, inspired by a clinical trial in Alzheimer’s disease.