Parametric and Non-Parametric Models in Health Research: Analysis and Design

Abstract

Personalized medicine is emerging in both clinical practice and clinical trials. “Precision” medicine not only promises improved safety and efficacy but also lowered cost in clinical practice and clinical trials. In 2015, President Obama launched the Precision Medicine Initiative. This initiative requires close collaboration among clinicians, researchers, and biostatisticians. Enrichment design is an important strategy for increasing study efficiency in personalized medicine. Enrichment clinical trial designs involve identifying high-risk patients and choosing patients most likely to respond to treatment. In this dissertation, we have developed and applied parametric and non-parametric models to the following specific problems: 1) identifying high risk patients using Classification and Regression Trees (CART) model; 2) using Bayesian distributional approach and finite mixture normal model to improve trial efficiency in a rare endpoint scenario; 3) using dynamic linear normal model in enrichment trial designs with ordinal risk subgroups. The topics we discussed in this dissertation form a self-contained system within the enrichment clinical trial design structure. Identifying high risk patients and efficient statistical models are two major components in enrichment designs. However, the application of the models we discussed is far beyond the scope in this dissertation. Using CART to identify high risk subpopulations can overcome the incapacity of logistic regression models in revealing unknown interaction effects. A distributional approach using finite mixture normal model provides a flexible model design to fit strongly skewed data. Dynamic linear normal model in the enrichment trial design shows to be more efficient and robust compared to previously studied designs because it locally smoothes the trend. All these methods help us to accurately identify target patients and treat patients efficiently.