Preparing, Executing, and Designing Complex Clinical Trials Using Simulations

The work in this dissertation provides applied methods in designing, preparing, and executing complex clinical trial designs with the use of simulations. These topics include designing and implementing a multi-site comparative effectiveness Bayesian response adaptive randomized trial; continued work to detail the execution of this study and assess if any treatment outcome drift should have been accounted for; and lastly, using data-driven simulations to compare a proposed statistical model with an alternative model choice. Response adaptive randomization and Bayesian methods have grown rapidly, but these contribute to more complex study designs. Many challenging statistical issues arise when designing, conducting, and analyzing adaptive trials. Bayesian methodology can address some of the complexities as it was developed to deal with data as they come in by updating prior information. Even with the rapid growth of Bayesian methods within adaptive trials, the literature detailing the implementation and conduction of these complex designs is lacking. To address this, we first detail the study design and implementation of a multi-site Bayesian comparative effectiveness response adaptive randomized trial that combined the safety and efficacy of treatments into a utility function. We detail an in-house computation software program we developed using REDCap, SAS ®, and R, allowing us to collect study data, perform interim analyses by extracting and analyzing the most up-to-date data, and adjust the new allocation sequence for the subsequent participants with minimal interruption to the study sites. We then continued the work after trial completion to provide literature on the conduction of this design detailing results at each interim analysis. We show how the tolerability and efficacy measures, combined into a utility function, drove the respective sample size for treatments, with more participants receiving the better-performing treatments. A concern in adaptive designs is the potential for biased estimates resulting in inflated type I error or decreased power due to treatment outcome time trends or drift. When originally planned, our primary analysis did not specify an adjustment for response drift, so a post-hoc analysis with a time-adjusted model was fit using the Bayesian Time Machine. This method allows researchers to account for any temporal time trends in modeling or, in our case, perform a post-hoc analysis to determine if the time-adjustment should have been included in our primary analysis model. The time-adjusted model did not change the results from the primary analysis in our four-arm study. Lastly, we used data-driven simulation studies to compare two statistical analysis methods for a longitudinal, group-randomized trial design. Considerations must be taken when designing group-randomized clinical trials due to the hierarchical data structure, and longitudinal designs have an added layer of nesting, adding more complexity. Simulation studies are performed to compare the operating characteristics and validate statistical models for these, many using simulations from known distributions under set assumptions. Our manuscript aims to use previous study data to compare statistical methods through data-driven simulations, allowing the data to drive the assumptions of the models. Modeling % weight change at 24 months, we use bootstrapping to compare the empirical power and type I error rate for the proposed longitudinal mixed-effects model against a baseline adjusted model at a single time point. It has been shown that the longitudinal model could have an inflated type I error under certain assumptions. The type I error rates for our proposed model did not result in an inflated type I error rate, and the power of the models for varying effect sizes was comparable. This implies that the longitudinal model is appropriate to use for the prospective study.

Grantor dc:publisher

University of Kansas

Year dc:date.issued

2022

Author dc:creator

• Brown, Alexandra

Advisor dc:contributor.advisor

• Gajewski, Byron

• Biostatistics
• adaptive
• Bayesian
• cluster-randomized
• longitudinal
• response adaptive

Statement dc:rights

• Copyright held by the author.

Language dc:language.iso

en

Dc Identifier Other

http://dissertations.umi.com/ku:18753

OAI identifier oai:identifier

oai:kuscholarworks.ku.edu:1808/36429