Presented by Lindsey Turner
Ph.D. Candidate in Biostatistics
Ph.D. Adviser: Dr. Thomas Murray

In randomized controlled trials, ordinal outcomes provide a more comprehensive summary than binary outcomes and are often more statistically efficient. However, there is a lack of transparent ways of summarizing the treatment effect across all levels of an ordinal outcome without relying on the proportional odds assumption. This dissertation develops novel estimands for summarizing the treatment effect on ordinal outcomes, Bayesian models for their estimation, and adaptive trial features based on these innovations. First, we propose summarizing the overall treatment effect by taking the geometric weighted mean of the threshold specific odds ratio from a Bayesian partial proportional odds model using weights based on the incidence of each level of the ordinal outcome. This novel approach aligns with the estimand framework while providing appealing efficiency properties relative to a proportional odds approach. We extend this weighting scheme to also include clinical relevance weights and propose a novel stratified partial proportional odds model. This approach offers flexibility to define certain levels of the ordinal outcome as more clinically relevant and provides a novel framework for conducting sensitivity analyses with different utility values. Finally, we describe a framework for early futility analyses in platform trials based on a summary of the treatment effect on an intermediate ordinal outcome that incorporates prognostic information between this intermediate outcome and the primary outcome which requires substantially longer follow-up to observe. This method builds off of the utility-based weighting scheme and embeds this estimand into an adaptive platform design to provide more reliable early futility decisions. The methods developed in this dissertation broaden the toolkit for analyzing ordinal outcomes in clinical trials, and provide transparency and control over the influence of the ordinal outcome levels, which may have widely-varying clinical importance, on treatment effect evaluations.