Note: See also Example of ROC Curves.
Suppose you have an x value that is a diagnostic measurement and you want to determine a threshold value of x that indicates the following:
• A condition exists if the x value is greater than the threshold.
• A condition does not exist if the x value is less than the threshold.
For example, you could measure a blood component level as a diagnostic test to predict a type of cancer. Now consider the diagnostic test as you vary the threshold and thus cause more or fewer false positives and false negatives. You then plot those rates. The ideal is to have a very narrow range of x criterion values that best divides true negatives and true positives. The Receiver Operating Characteristic (ROC) curve shows how rapidly this transition happens. The goal of the ROC curve is to have diagnostics that maximize the area under the curve.
Two standard definitions used in medicine are as follows:
• Sensitivity, the probability that a given x value (a test or measure) correctly predicts an existing condition. For a given x, the probability of incorrectly predicting the existence of a condition is 1 – sensitivity.
• Specificity, the probability that a test correctly predicts that a condition does not exist.
A ROC curve is a plot of sensitivity by (1 – specificity) for each value of x. The area under the ROC curve is a common index used to summarize the information contained in the curve.
When you do a simple logistic regression with a binary outcome, there is a platform option to request a ROC curve for that analysis. After selecting the ROC Curve option, you must specify which level to use as the positive response.
If a test predicted perfectly, it would have a value above which the entire abnormal population would fall and below which all normal values would fall. It would be perfectly sensitive and then pass through the point (0,1) on the grid. The closer the ROC curve comes to this ideal point, the better its discriminating ability. A test with no predictive ability produces a curve that follows the diagonal of the grid (DeLong et al. 1988).
The ROC curve is a graphical representation of the relationship between false-positive and true-positive rates. A standard way to evaluate the relationship is with the area under the curve, shown below the plot in the report. In the plot, a yellow line is drawn at a 45-degree angle tangent to the ROC Curve. This marks a good cutoff point under the assumption that false negatives and false positives have similar costs.