COY ARJATE ADruSTED RESPONSE ADAPTIVE RANDOMIZATION DESIGNS WITH TARGETED LEARNING

Since Food and Drug Administration (FDA) issued the very first draft guidance in 2010, adaptive designs have been considered to be one of the most promising approaches to make drug development more efficient and less costly. Two approaches, covariate-adjusted response-adaptive (CARA) randomization and adaptive seamless phase II/III design (ASD) have garnered growing attention recently. However, most of the CARA designs are based on parametric models and suffered from model misspecification. The research of incorporating CARA into ASD is also limited and whether type I error rate can be controlled has not been answered. In this dissertation, a new family of CARA emphasizing on efficiency and ethics using targeted maximum likelihood estimators (TMLE) was proposed to address public health questions as well as tackle the issue of restrictive modeling assumptions. Moreover, the combination of ASD and CARA using TMLE was studied under different scenarios. The asymptotic properties of these approaches were provided and proved rigorously. The simulation studies were carried out to check the concept further. The operating characteristics revealed that all of the proposed approaches have well-controlled type I error rates around the nominal level, increases in power and advantage in other ethical aspects.