Bayesian adaptive randomization with covariate-adjustment signature design
The conduct of a traditional, equally randomized clinical trial may sacrifice patient benefit by equally assigning patients to different treatment arms even though the trial's accumulating results indicate that one particular treatment is more promising. In addition, a randomized trial that uses broad eligibility criteria often overlooks the effective agents and fails to identify the subgroup of patients that is most sensitive to the treatment. In this research project, we propose the Bayesian adaptive randomization design with covariate adjustment (ARCA) to preferentially allocate more patients to the more effective treatment arm. We compare the performance of ARCA with other trial designs, including equal randomization and response-adaptive randomization. To accommodate a loose enrollment criteria, we implement Freidlin and Simon's adaptive signature feature, which allows us to identify the treatment-sensitive patients in the middle of the trial and then to test the treatment effect in the subset of the selected treatment-sensitive patients among the remaining patients. For trials that collect enormous amounts of patient information, we explore the use of the machine learning voting method to select potentially important variables for use in covariate-adjusted adaptive randomization as well as to develop a classifier to identify the treatment-sensitive subgroup. Future work will evaluate the least absolute shrinkage and selection operator (LASSO) and smoothly clipped absolute deviation (SCAD) methods in identifying an effective variable selection method.
Qiao, Wei, "Bayesian adaptive randomization with covariate-adjustment signature design" (2014). Texas Medical Center Dissertations (via ProQuest). AAI3689983.