Analyzing paired diagnostic studies by estimating the expected benefit

When the efficacy of a new medical drug is compared against that of an established competitor in a randomized controlled trial, the difference in patient-relevant outcomes, such as mortality, is usually measured directly. In diagnostic research, however, the impact of diagnostic procedures is of an indirect nature as test results do influence downstream clinical decisions, but test performance (as characterized by sensitivity, specificity, and the predictive values of a procedure) is, at best, only a surrogate endpoint for patient outcome and does not necessarily translate into it. Not many randomized controlled trials have been conducted so far in diagnostic research, and, hence, we need alternative approaches to close the gap between test characteristics and patient outcomes. Several informal approaches have been suggested in order to close this gap, and decision modeling has been advocated as a means of obtaining formal approaches. Recently, the expected benefit has been proposed as a quantity that allows a simple formal approach, and we take up this suggestion in this paper. We regard the expected benefit as an estimation problem and consider two approaches to statistical inference. Moreover, using data from a previously published study, we illustrate the possible insights to be gained from the application of formal inference techniques to determine the expected benefit.