GEE approaches to marginal regression models for medical diagnostic tests

The evaluation of a new medical diagnostic test may focus on two different scientific questions: (1) The new test may replace an existing one because of lower cost or higher validity. A related question would be the selection of the 'best' test(s) from a bundle of new or established measurements. (2) The new test may be used supplementary to other new or established procedures. In a recent publication, Leisenring and co-workers (Stat Med 1997; 16:1263-1281) developed a general marginal regression model for comparisons of diagnostic tests focussing on question (1), i.e. on the selection of the 'best' procedure. They applied the GEE approach of Liang and Zeger (Biometrika 1987; 73:13-22) to adjust for the correlation of data as a nuisance parameter. Using the general framework provided by Leisenring et al., we extend their approach and apply the GEE methodology to question (2), i.e. to the investigation of which of several diagnostic tests should be used supplementary to each other. We analyse data from a longitudinal study concerning pathogenesis, diagnosis and long-term course of the eye disease glaucoma. We find a dependence of the correlation structure of several diagnostic measurements on the severity of the disease. This result may be useful in clinical applications as regards the selection of subsets of diagnostic measurements in individual diagnostic processes but also in investigations concerning the relationship of the pathogenic process and the rationales of the different diagnostic procedures.