Dissertation Defense: Qiaohui Zhou

Candidate Name: Qiaohui Zhou

Major: Biostatistics

Thesis Advisor: Ming Tan, Ph.D.

Title: Estimand Framework in Clinical Research

This thesis addresses key challenges in the design and analysis of clinical trials, focusing on baseline covariate imbalance, causal inference for studies with binary outcomes, and the relationship between intercurrent events (ICEs) and mediation analysis. We propose a joint chi-square test for assessing baseline covariate imbalance, offering a more comprehensive evaluation than individual tests and potentially improving trial result validity. We introduce an enhanced doubly robust estimator (eDRE) for binary outcomes in causal inference, using semiparametric models with nonparametric monotone link functions for propensity score and outcome models. This approach further enhances robustness of the traditional parametric doubly robust estimators. We present an iterative algorithm for parameter estimation and establish the estimator’s asymptotic properties. Simulations demonstrate our eDRE method’s superior performance compared to inverse probability score weighting and naive estimators across various scenarios. We explore the connection between ICE and mediation analyses within the estimand framework, clarifying distinctions and similarities between them. Causal diagrams illustrate strategies for handling ICEs, and we develop statistical models for both ICE and mediation analyses, highlighting their interconnections and differences. Simulation studies are performed to illustrate the two types of post-trial events under various data generation settings, including scenarios combining ICE and mediator features. Our work contributes more robust methods for assessing baseline imbalance, estimating causal effects with binary outcomes, and understanding ICE-mediation relationships. These advancements have significant implications for improving clinical trial design, analysis, and interpretation, enhancing medical research reproducibility. Future research directions include extending these methods to different types of trial designs and exploring real-world clinical applications.