Presenter: Mohammad Reza Fahimi
Supervisory Committee: Osvaldo Espin-Garcia (Supervisor), Aya Mitani (Supervisor), Victoria Landsman
Chair: Tony Panzarella
Date and Time: Thursday, August 1, 2024, 10:20-12:30pm EST
Location: 155 College Street, Health Sciences Building, Room 507
Abstract: Two-phase sampling designs offer a cost-effective approach for studies where measuring key variables is expensive. This research focuses on identifying optimal sampling strategies for two-phase studies with ordinal outcomes, an area not extensively studied. We develop a semiparametric maximum likelihood framework for vector generalized linear models focusing on cumulative, adjacent-category, and stopping-ratio logits. The expectation-maximization (EM) algorithm along with the Louis’s method are used for obtaining more accurate parameter estimates and covariance matrix, respectively. We examine balanced sampling strategies with and without a genetic algorithm (GA) that minimizes the asymptotic variance of the parameter of interest to identify optimal designs. Simulations reveal that optimal designs, particularly those using a GA with the EM algorithm, significantly reduce bias and improve efficiency. The practical application of these methods is illustrated using biopsy data on breast cancer, demonstrating the potential for improved precision in parameter estimation with a small Phase 2 subsample.