A comprehensive study of the theory and application to biomedical research of Receiver Operating Characteristic (ROC) curves has continued. ROC analysis is used to compare two diagnostic or laboratory tests when the data are ordinal categories or continuous variables. A study of the applicability of Lomax distribution to ROC curves has continued. A paper with Dr. Ratnaparkhi on the maximum likelihood estimation of Lomax models and their iterative fit via computer is under revision. An important practical area of ROC analysis recently studied is the incorporation of covariate information. A collaboration joint with Drs. J. Norman, D. Levy and J. Bailey has used age and body mass index (habitus) in a linear regression model to improve fuzzy ROC performance of an ECG test in prediction of left ventricular hypertrophy. The study of the relationship between ROC curves and artificial neural networks has commenced. ROC curves are being studied in the evaluation of the performance of individual artificial neural networks (ANN). Also, a study has begun to use some measure based on ROC curves such as area or some measure of tradeoff of different errors and risks to optimize the performance of the ANN. A manuscript with M. Zweig (CC) is in preparation concerning the fundamental role of ROC plots and analyses in the evaluation of laboratory tests in clinical chemistry and pathology. A study has begun of randomization tests, transformations, and statistical inference. A special topics session on methodology for the evaluation of diagnostic and laboratory tests and its applicablility to biomedicine has been organized for the National Annual Statistical Meetings in Boston in August, 1992. Also work has begun on a special oral and written presentation for the NIH Biostatistics Symposium to be held in January, 1993.
