The performance of a diagnostic test is often expressed in terms of sensitivity and specificity compared with the reference standard. Calculations of sensitivity and specificity commonly involve multiple observations per patient, which implies that the data are clustered. Whether analysis of sensitivity and specificity per patient or using multiple observations per patient is preferable depends on the clinical context and consequences. The purpose of this article was to discuss and illustrate the most common statistical methods that calculate sensitivity and specificity of clustered data, adjusting for the possible correlation between observations within each patient. This tutorial presents and illustrates the following methods: (a) analysis at different levels ignoring correlation, (b) variance adjustment, (c) logistic random-effects models, and (d) generalized estimating equations. The choice of method and the level of reporting should correspond with the clinical decision problem. If multiple observations per patient are relevant to the clinical decision problem, the potential correlation between observations should be explored and taken into account in the statistical analysis.
Supplemental material: http://radiology.rsna.org/lookup/suppl/doi:10.1148/radiol.12120509/-/DC1.
© RSNA, 2012.