GraphPad Statistics Guide

Key concepts: Receiver-operating characteristic (ROC) curves

Key concepts: Receiver-operating characteristic (ROC) curves

Previous topic Next topic No expanding text in this topic  

Key concepts: Receiver-operating characteristic (ROC) curves

Previous topic Next topic JavaScript is required for expanding text JavaScript is required for the print function Mail us feedback on this topic!  

When evaluating a diagnostic test, it is often difficult to determine the threshold laboratory value that separates a clinical diagnosis of “normal” from one of “abnormal.”

If you set a high threshold value (with the assumption that the test value increases with disease severity), you may miss some individuals with low test values or mild forms of the disease. The sensitivity, the fraction of people who have the disease that will be correctly identified with a positive test, will be low. Few of the positive tests will be false positives, but many of the negative tests will be false negatives.

If you set a low threshold, you will catch most individuals with the disease, but you may mistakenly diagnose many normal individuals as “abnormal.” The specificity, the fraction of people who don't have the disease who are correctly identified with a negative test, will be low. Few of the negative tests will be false negatives, but many of the positive tests will be false positives.

You can have higher sensitivity or higher specificity, but not both (unless you develop a better diagnostic test).

A receiver operating characteristic (ROC) curve helps you visualize and understand the tradeoff between high sensitivity and high specificity when discriminating between clinically normal and clinically abnormal laboratory values.

Which is the best combination of sensitivity and specificity?  It depends on the circumstances. In some cases, you'll prefer more sensitivity at the expense of specificity. In other cases, just the opposite. Prism cannot help with those value judgments.

Why the odd name? Receiver operating characteristic curves were developed during World War II, within the context of determining if a blip on a radar screen represented a ship or an extraneous noise. The radar-receiver operators used this method to set the threshold for military action.

ROC curves can also be used as part of the presentation of the results of logistic regression. Prism does not do logistic regression so does not prepare this kind of ROC curve.

The review by Berrar (1) is excellent both for understanding ROC curves and for appreciating some of their pitfalls.

ROC curves can also be used as a way to display results from multiple logistic regression. This is not something Prism can do.

 

1.Berrar D, Flach P. Caveats and pitfalls of ROC analysis in clinical microarray research (and how to avoid them). Brief Bioinform. Oxford University Press; 2011 Mar 21;13(1):bbr008–97.