ESTIMATING DIAGNOSTIC-TEST ACCURACY USING A FUZZY GOLD STANDARD

This study uses Monte Carlo methods to analyze the consequences of hav ing a criterion standard (''gold standard'') that contains some error when analyzing the accuracy of a diagnostic test using ROC curves. Two phenomena emerge: 1) When diagnostic test errors are statistically in dependent from inaccurate (''fuzzy'') gold standard (FGS) errors, esti mated test accuracy declines. 2) When the test and the FGS have statis tically dependent errors, test accuracy can become overstated. Two met hods are proposed to eliminate the first of these errors, exploring th e risk of exacerbating the second. Both require a probabilistic (rathe r than binary) gold-standard statement (e.g., probability that each ca se is abnormal). The more promising of these, the ''two-truth'' method , selectively eliminates those cases where the gold standard is most a mbiguous (probability near 0.5). When diagnostic test and FGS errors a re independent, this approach can eliminate much of the downward bias caused by FGS error, without meaningful risk of overstating test accur acy. When the test and FGS have dependent errors, the resultant upward bias can cause test accuracy to be overstated, in the most extreme ca ses, even before the offsetting ''two-truth'' approach is employed.