Files
Abstract
Statistical methods for evaluating predictive biomarkers' clinical utility, reproducibilityand sample size for specific study designs have been slow to develop in response to the surge of need. In this dissertation, we propose three statistical methodologies: one is develop a metric using Bayesian decision theoretic framework; the second is propose a sample size estimation method (SWIRL); and the third is develop a reproducibility metric .A metric which measures the decrease in the expected event rate as a result of predictive biomarker guided therapy is proposed using Bayesian decision theoretic framework for a count clinical end point. Since Phase II data are usually small, maximum likelihood based estimates are biased and inefficient. This new metric, which also incorporates clinician inputs in the form of a prior however, is informative in making a go-no-go decision and the study design to choose for Phase III studies. Using toy simulation and a simulation conducted to mimic asthma clinical trial study, we show the robustness of the method under different scenarios.Sample size estimation methods that match the study design and the metric under consideration are key in predictive biomarker clinical utility evaluation process. In this dissertation we propose a sample size estimation method, Squared Width Inversion Regression Line (SWIRL). The SWIRL method is used to estimate a sample size n such that the 95% confidence interval width of the metric under consideration is smaller than a user defined length (Wtarg). This is the first sample size method developed for estimating this target predictive parameter.During assay development and validation processes, an original clinically validated assay is required to be modified for a number of different reasons. However, such medication invalidates the previous biomarker-outcome association studies and would force researcher to re-run the previous studies under the modified biomarker. This is time consuming and expensive. Here, we propose a reproducibility metric _r which measures the impact of assay modification on patient outcome. A combination of both novel equations and simulations were used to estimate _r and the associated 95% confidence interval.