Robustness and Efficiency in Modern Biostatistical Studies
There is a need to develop robust and efficient versions of commonly used biostatistical methods in order to maintain valid statistical inference while remaining comprehensible to other investigators and recognizable to regulatory agencies. In this dissertation, we will explore the intersection of robustness and efficiency in realistic and complex biostatistical settings through two clinically-motivated examples. The first relates to the validation trial of a novel bleed severity classification system, where a kappa statistic was used to assess inter- and intra-rater reliability. We investigate how clustering and within-category item non-exchangeability seen in the trial affect the analytic variance estimate of kappa, and propose a new variance estimator to correct for variance inflation. We also explore how such a trial could be made more efficient with the use of group sequential design methodology. We show that implementation of traditional boundary estimation methods on kappa yield approximately correct operating characteristics, and discuss study design considerations particular to kappa. The second example relates to the discovery of novel biomarkers in Alzheimer’s disease. New biomarkers are often explored within observational study frameworks and while there is no artificial intervention being tested, efficient procedures – such as group sequential design – are still desirable to assess the viability of potentially large numbers of prospective biomarker candidates using scarce or difficult-to-obtain samples. We examine how failure to incorporate confounding mechanisms when designing group sequential observational studies affects operating characteristics, and develop algorithms and processes to correct them.