Current advances in technology stimulate development of new medical diagnostic tests for identifying disease status. Performance of a diagnostic test is often evaluated by comparison to a currently accepted gold standard and measured by sensitivity (proportion of truly diseased patients who test positive for the disease) and specificity (proportion of truly non-diseased patients who test negative). When collecting data for a diagnostic test evaluation, one often finds that not all patients undergoing the test have their disease status verified with the gold standard because of ethical reasons. Naive estimates of sensitivity and specificity based only on the patients with verified disease can be severely biased. This bias is called a verification bias or a work-up bias. In this talk the problem is framed as a missing data problem. We discuss various models and assumptions which provide point estimates of sensitivity and specificity. In addition, we suggest assumption free Test Ignorance Region (TIR) encompassing all combinations of sensitivity and specificity values compatible with the observed data. We argue that the TIR should be a routine reporting tool for a first step in evaluation of a diagnostic test when the gold standard is partially missing. This way the information included in data and "information" induced by assumptions can be clearly separated. The methodology is illustrated with several real data examples.
Return to Biostatistics Working Group