DOI
https://doi.org/10.25772/48FC-PM67
Defense Date
2020
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Biostatistics
First Advisor
Victoria C. Garcia
Second Advisor
Adam P. Sima
Abstract
Optimal response-adaptive randomization (RAR) is used in clinical trials (CTs) to balance ethics and power by randomizing patients according to the unknown population parameters of an assumed parametric response distribution selected prior to data collection. If the distribution of the observed response data is misspecified, the resulting randomization may produce unanticipated trial characteristics. This dissertation introduces flexible estimators of the optimal RAR ratio that adjust for such misspecification.
When observed response data were continuous in nature, sample moments were estimated from a continuous fit of the empirical CDF. These sample moments then served as the operands for the first flexible estimator of the RAR ratio presented in this work. The second estimator utilized a weighted-average approach. An array of RAR ratios corresponding to a set of candidate response distributions were estimated from the observed response data. These were then weighted by that distribution’s fit to the observed response data, and the resulting products were summed. These methods produced the desired randomization results in simulated CTs. However, they were computationally expensive, required large lead-ins, and did not consistently outperform the conventional normality assumption, even in small samples.
Where a number of optimal RAR design exist for CTs with continuous outcomes, there is a paucity of the same for CTs with survival outcomes. Thus, convention for such trials is the optimal RAR framework constructed by Zhang and Rosenberger (ZR RAR), which has limitations. It relies upon the distributional specification of survival times and uses proxy metrics to define the RAR ratio, where only two cases (exponential and Weibull) have been addressed. It was hypothesized that the ZR RAR convention may be highly susceptible to distributional misspecification for these reasons. Thus, its behavior was examined under increasingly severe deviations from correct specification in the present work.
A novel RAR ratio based on the cumulative hazard (H-RAR) was then developed using the ZR optimal RAR framework. Randomization results produced by H-RAR were compared to those obtained from (a) correctly-specified ZR RAR, (b) misspecified ZR RAR using well-behaved survival times, and (c) severely misspecified ZR RAR using change-point survival times. H-RAR recreated the correctly-specified ZR RAR and outperformed ZR RAR when misspecified. These results are promising and demonstrate a computationally straightforward and accurate estimator of the optimal RAR ratio for CTs with survival outcomes.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
8-6-2020