Latent Class Analysis of Diagnostic Agreement


Yen-Ling Kuo
2009-12-04  12:30 - 14:30
Room 405, Mathematics Research Center Building (ori. New Math. Bldg.)

Assess diagnostician agreement is often encountered in medical research. (See "The reliability of clinical methods, data and judgments" in New England Journal of Medicine Vol 293 pp642-646, 695-701, 1975). In this talk, I review the paper entitled " Latent Class Analysis of Diagnostic Agreement" by John S. Uebersax (1990). We first describe methods of probability modeling to analyze rater agreement. Statistical techniques for analyzing agreement data are described to address questions such as how many opinions are required to make a medical diagnosis with necessary accuracy. Focus is on two related techniques, which differ in assumptions about disease subtypes and associated differences among cases in their ability to be correctly diagnosed: (1) latent class agreement analysis; and (2) latent trait agreement analysis. Specifically, these methods make it possible to determine from the opinion of panels of diagnosticians in an agreement study the following: the probable accuracy of an individual diagnosis; the probability of disease presence or absence given unanimous or conflicting opinions by several diagnosticians; and how many opinions should be required to make the diagnosis. In addition, we present the EM algorithm which resolves difficult in estimating parameters with traditional MLE method. Moreover, we’ll apply the latent class methods to examples of medical agreement data with tuberculosis patients and find out that the data is poorly fitted by two-class. But three-class models will greatly improve the fitting.