AbstractReceiver operating characteristic (ROC) analysis plots true positive rates over false positive rates to describe the discriminatory power of a test to differentiate between two specifiable populations (that is ‘normal’ from ‘abnormal’) using a continuous dichotomous threshold. This assumes that the test separates cases into observed ‘normal’ or ‘abnormal’. However, in practice many tests have some uninterpretable results as an inherent feature of the test itself, whether independent or dependent on the sample populations. This paper defines a method to describe the ability of tests to discriminate between specifiable populations when uninterpretable results occur non‐informatively about disease status. A mixed model modified ROC curve is developed. Formulae to estimate the area under the modified ROC curve are given. Comparisons of conventional with the mixed model ROC analysis are shown in mathematical formulae as well as simulated experiments. An example of diagnosis of spinal fracture in renal transplant patients is presented. Copyright © 2002 John Wiley & Sons, Ltd.