DOI
https://doi.org/10.25772/9RB0-1G73
Defense Date
2020
Document Type
Dissertation
Degree Name
Doctor of Philosophy
Department
Biostatistics
First Advisor
Dr. Roy Sabo
Abstract
Sequential, multiple assignment, randomized trials (SMARTs) allow investigators to develop and compare experimental treatment regimens in which individuals are successively randomized to different treatments based on some set of predetermined rules. The rules used to make decisions on when and how to switch treatments are based on a chosen set of tailoring variables. Although not always true, intermediate response is commonly used as the primary tailoring variable as it is often predictive of future treatment success. As such, successful implementation depends on identifying patients who respond to treatment, though in some situations such mechanisms may not exist. Further, patient-level covariates may affect intermediate response and should be taken into account as a secondary tailoring variables in order to achieve the development of more deeply tailored trials. Our goal was to develop a method for probabilistically assigning responder statuses to subjects completing the first stage of a SMART supplemented with information provided through pilot data. Several approaches for estimating and updating these probabilities are discussed, including a combination of cluster analysis and discriminant analysis as well as mixture-model approaches. Extensions of these approaches that adjust for patient level covariates were also explored. In all cases the estimated responder probabilities were used to allocate patients between responder and non-responder classifications, allowing patients to continue onto the second phase of treatment. Approaches were evaluated using simulation studies and compared in terms of misclassification error rates and resulting bias in the stage-two outcomes under a variety of situations. Recommendations are made for which methods perform best under different sampling and population settings. Overall, both the cluster/discriminant and finite mixture modeling techniques provide a way of operationalizing a clinical decision support tool for determining responder statuses sequentially in a SMAR-like trials when no other valid methods are available.
Rights
© Keighly Bradbrook
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
8-5-2020
Included in
Applied Mathematics Commons, Applied Statistics Commons, Biostatistics Commons, Medicine and Health Sciences Commons