Design of a study to improve accuracy in reading mammograms.
This paper is concerned with the design and analysis of mammography reading studies. In particular we consider studies aimed at evaluating interventions to improve the accuracy with which mammograms are read. A simple randomized design is suggested in which a relatively large group of readers read sets of mammograms before and after an intervention phase. We propose solutions to three difficult statistical issues that arise in the context of such studies: (i) the choice of primary outcome measure; (ii) the data analysis technique to be employed; and (iii) the methodology for calculating sample sizes for readers and images to be read. First, we argue in favor of using sensitivity and specificity as the primary outcome measures rather than receiver operating characteristic (ROC) curves in mammography studies, although the latter are considered state of the art for many types of radiology reading studies. We argue that sensitivity and specificity are more clinically relevant and conceptually more straightforward than ROC curves. Second, we suggest a bivariate approach to data analysis for evaluating intervention effects on sensitivity and specificity. This accommodates the correlations inherent between these measures and allows for estimation of joint effects on them. Finally we propose a method for power calculations that uses computer simulation techniques. Simple formulas for sample size calculations are not available in part because variability in accuracy amongst readers and variation in difficulty among images introduce complexity into power calculations. The simulation method that we propose accommodates such complexity and is easy to implement. The methodology was motivated by a study funded by the Department of Defense to evaluate the potential efficacy of an educational intervention. In the context of this study we illustrate the steps involved in power calculations and apply the data analytic techniques to the sort of data expected to result from this study. Though the proposed methods were motivated by this particular study, the statistical considerations are relevant more broadly in mammography and indeed in other types of radiologic imaging studies. Standards for the conduct of radiologic reading studies are not yet well developed, as they are for randomized clinical trials and for case-control studies. We hope that the discussion in this paper will add to the dialogue necessary for development of such standards.