The proposed research will investigate and generate methods for the evaluation of multiple correlated diagnostic test results. The biostatistical methods themselves are applicable to diverse fields. This application will focus on analysis of receiver operating characteristic (ROC) curves generated from diagnostic imaging studies. ROC analysis is used to compare the capacities of imaging systems (i.e., combination of imaging modality and reader) to discriminate between actually positive and actually negative cases. Diagnostic imaging studies commonly use designs in which each case is imaged with multiple modalities and each image is interpreted by several radiologists independently, to increase the power of comparisons of discrimination capacities across modalities and to learn about variation in discrimination capacity across readers. The discrimination capacity of an imaging system can also depend on characteristics of cases. The overall goal of this research is to create innovative statistical methods for evaluating medical diagnostic tests that will be practical for use in the real world. First, multivariate ordinal regression models will be developed that allow discrimination capacity to depend on characteristics of cases while allowing inferences to generalize to populations of readers (random effects). Second, multivariate ordinal regression methods that can accommodate the types of missing data that arise in repeated measurements studies will be developed. Finally, issues of study design and performance of multivariate ordinal regression models in studies with small to moderate sample sizes will be investigated. The ordinal regression models used in this research will employ a multiplicative predicator such that both the height and summetry of the ROC curve can depend on characteristics of cases and/or readers. The proposed research requires difficult innovations in biostatistical methodology for the analysis of multivariate ordinal categorical data. The development of new methods and generation of new ideas and insights will have a significant impact in both the medical and the biostatistical communities on research in methods for evaluating diagnostic tests.
| Toledano, Alicia Y (2003) Three methods for analysing correlated ROC curves: a comparison in real data sets from multi-reader, multi-case studies with a factorial design. Stat Med 22:2919-33 |
