van der Schouw Y T, Straatman H, Verbeek A L
Department of Medical Informatics and Epidemiology, University of Nijmegen, The Netherlands.
Med Decis Making. 1994 Oct-Dec;14(4):374-81. doi: 10.1177/0272989X9401400408.
Many measures, including sensitivity and specificity, predictive values, and likelihood ratios, are available for the assessment of diagnostic tests. A drawback of the use of these measures is that continuous test results are often dichotomized, with consequent loss of information. Receiver operating characteristic (ROC) curves do not depend on discrimination thresholds, and therefore the area under the ROC curve (AUC) is one of the preferred measures. Although quantitative test results are often presented dichotomized, it would be convenient still to be able to estimate the ROC curve and the AUC. The authors present equations for such estimates when only one pair of a true- and a false-positive rate is given, for inherently logistically and normally distributed data. Illustrative empirical data are provided for both distributions. In contradiction to earlier reports, the authors also show that differential disease verification may skew the ROC curve. The ROC curve is thus not invariant to selection bias.
许多用于评估诊断试验的指标，包括灵敏度和特异度、预测值以及似然比等。使用这些指标的一个缺点是，连续的试验结果常常被二分法处理，从而导致信息丢失。受试者工作特征（ROC）曲线不依赖于判别阈值，因此ROC曲线下面积（AUC）是首选指标之一。尽管定量试验结果常常以二分法呈现，但能够估计ROC曲线和AUC仍然会很方便。作者给出了在仅给出一对真阳性率和假阳性率时，针对固有逻辑分布和正态分布数据进行此类估计的方程。针对这两种分布都提供了说明性的经验数据。与早期报告相反，作者还表明，不同的疾病验证方式可能会使ROC曲线产生偏差。因此，ROC曲线对于选择偏倚并非不变。
