Estimation of Sensitivity and Specificity for Clustered Data

The most commonly used measurements for evaluating the accuracy of diagnostic tests in binary data are sensitivity andspecificity. Sensitivity refers to the ability of a test to detect patients with some specific disease. Specificity describes how well testabnormality is restricted to those persons who have the disease in question. The variance estimators of these measurements areusually obtained using a binomial estimator, although there are several other methods. However, under certain conditions more thanone observation can be taken from a subject and analyses are performed on these observations. This type of data structure is calledÒclustered dataÓ. For calculating sensitivity, specificity and related variances in clustered data, conventional statistical methods withthe assumption that all observations are independent will not be valid. The most commonly used methods for clustered data areÒratio estimatorÓ, Òwithin-cluster correlation estimatorÓ and Òweighted estimatorÓ. In this paper, for calculating diagnosticmeasurements and their variances, the above-mentioned methods are described. In the example of magnetic resonance angiography,which is used for the diagnosis of renal artery stenosis, the results of a binomial estimator are compared with those of the other 3methods. Although there were no significant differences between sensitivity estimates, variance estimates obtained from thebinomial estimator were higher than the other estimates