Design of ROC studies in diagnosis and screening.

摘要

The complexity of the parameter set and the need for disease verification present significant challenges in the design of ROC studies with rating data in both diagnostic and screening populations.; The first part of this thesis examines currently used approximations for determining the required sample size in the estimation of a single ROC curve from rating data. Analytic results and computer simulations show that, by ignoring aspects of the ROC model, such as the number of categories in the data, these approximations may lead to unduly low estimates of the required sample size in practice. As a consequence, simulation studies may well be required to estimate the necessary sample sizes even in the relatively simple case of estimating the ROC curve of a single test.; The second part of the thesis presents a Bayesian approach to sample size selection for the estimation of an ROC curve. In contrast to the approximate solutions, the Bayesian approach makes full use of the ROC model, including the number of categories and the cut-points in the latent scale. The Bayesian approach also allows the investigator to account for the uncertainty regarding the assumptions about parameters of the model that are necessary in the sample size calculations.; The findings from the first two parts of the thesis inform the methodology of the third and most extensive part, which considers the problem of efficient design of ROC studies. The basic approach here is the use of two-phase designs, which permit estimation and comparison of ROC curves without requiring verification of all study participants. The utility of two-phase designs is higher in studies with low prevalence of the disease, such as evaluations of the diagnostic accuracy of screening modalities. We propose Bayesian methods for the design and analysis of two-phase ROC studies and compare them to maximum likelihood methods. The presentation illustrates the use of both types of methods first in studies designed to assess a single test and second in studies designed to compare the performance of two alternative tests. In practice, the cost and difficulty of verification of true disease status tend to be considerably higher than the cost of screening. In such settings, the two-phase designs are shown achieve efficiencies in cost over single-phase designs while maintaining the required level of precision of estimates.