A Macro For Getting More Out Of Your ROC Curve

Abstract

As a part of clinical decision making, a continuous measure reflective of a patient’s condition (e.g. laboratory test result) is often dichotomized (or “cut”) into two groups (e.g., abnormal and normal) for predictive or screening purposes. The decision of where to select the cut-off point is often governed by a reasonable trade-off between sensitivity and specificity for a particular test. A common method to help find this balance is to plot sensitivity versus (1-specificity) as a ROC” (Receiver Operator Characteristic) curve. Although a useful tool, the ROC curve rarely displays the individual cut-off values. Because of this limitation, the user cannot visualize the impact of varying sensitivity and specificity against the cut-off values of the prediction variable. The macro presented in this paper creates a single graph that, for a given prediction variable and a binary outcome (“true condition”), simultaneously displays the following: sensitivity, specificity, Youden Index, and various userdefined measures of misclassification error against cut-off values of the prediction variable. The graph is a simple but highly informative visual tool that provides the user with greater functionality than a standard ROC curve. This macro is intended for an intermediate SAS user with PC SAS (SAS/STAT® and SAS/GRAPH®) version 8.2 or 9.1 capabilities.