Bayesian Analysis for Assessing Diagnostic Test Accuracy and Optimal Design in a Cardiology Clinical Trial

• Activity Number: 381
• Type: Contributed
• Date/Time: Wednesday, August 11, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Manuela Buzoianu*+ and Joseph B. Kadane
• Companies: Carnegie Mellon University and Carnegie Mellon University
• Keywords: verification bias ; missing data ; MCMC algorithms ; utility function ; optimal design ; stepwise sampling

Abstract

Evaluating new diagnostic tests is necessary when the gold standard, which provides definitive verification of a disease, has a risk of mortality. This is the motivation for the current cardiology trial, in which a noninvasive test for coronary artery disease diagnosis has to be evaluated against the gold standard coronary angiography. For this, we studied the test sensitivity and specificity. When true disease status is missing for many patients, the estimates of the test parameters are likely to be biased. We performed a Bayesian analysis to correct for verification bias, which is the bias resulting from this missing data mechanism, and compare it with a likelihood-based approach. An important implication of this analysis is to make decisions about future patients so that the test properties estimates can be improved. Thus, the paper also discusses designing experiments, in which new patients are carefully selected for disease verification. Several experiments are possible and the design choice is regarded as a decision problem, which is based on experiment utility maximization. Simulation-based methods and stepwise sampling algorithms are developed to obtain the optimal design.