Have a Question?
Ask the Graduate
College at our new
Doctoral Dissertation Announcement
Candidate: Jason W. Bodnar
Doctor of Philosophy
Department: Educational Leadership, Research, and Technology
Title: Bayesian Item Response Theory: Statistical Inference and Power Analysis
Dr. Brooks Applegate, Chair
Dr. Chris L. Coryn
Dr. Jessaca Spybrook
Dr. Jung-Chao Wang
Date: Monday, October 3, 2011 10:00 a.m. to noon.
3310 Sangren Hall
The regulatory pharmaceutical approval process is flawed in that industry clinical trials (ICTs) are always powered for efficacy and rarely powered for safety. The key safety parameter is the adverse event (AE). This practice may result in efficacious products with confounded safety. An ICT’s ability to be powered for detecting AE trends may improve patient safety. Therefore, this dissertation’s purpose was to determine if power analysis resulted in feasible sample sizes for substantiating AE hypotheses.
AEs were modeled with three Bayesian 2PL IRT models. The unidimensional latent trait, transfusion-related AE, was modeled as a patient predisposition for experiencing an AE. Parametric and nonparametric inference and power analysis approaches were derived for paired IRFs. Analysis was based on 1,000 bivariate binomial simulations of nine AE types for n=30 and 250 patients. 2-PL, 2-PL EX, and 2-PL MEX adhered to the multiple chain assumption. Parameter estimators were stationary after 25,000 and 15,000 Gibbs samplers (GS), respectively, for 2-PL and 2-PL EX, and serial autocorrelation was removed. Simulation results revealed that the 2-PL EX demonstrated reasonable model fit based on linear trapezoid and spline approximations of the exact area under paired IRFs. Bootstrap, jackknife, and partial batch approaches were used for parametric and non-parametric inference. Optimal results occurred for the nonparametric bootstrap approach on the spline approximation. Superiority was expectedly not achieved. Equivalence (delta=10%) was not statistically substantiated for n=30, but was for n=250. Coverage was achieved for all inference. The superiority IRT approach required 933 patients for 95% confidence and 80% power different from existing methods, which required a minimum of 174,451 patients. The equivalence IRT approach required 60 patients for delta=10%. Existing method required a minimum of 95 patients.
Simulations resulted in comparable IRFs. Inference correctly characterized relationships between IRFs. IRT sample sizes were smaller than existing methods, and were expectedly larger for superiority. Powering a study to differentiate comparable groups typically requires enormous sample sizes. Equivalence was a viable solution to superiority. The next step is to have the IRT inference incorporated into the ICT safety investigation of all pharmaceutical products.