Statistics Glossary → Youden's J index (optimal cut-off)
Youden's J index (optimal cut-off)concept
A measure, J = sensitivity + specificity - 1, that combines a diagnostic test's or classifier's sensitivity and specificity in one number, and the cut-off that maximizes it.
J ranges from 0 (the test is no better than a coin toss) to 1 (perfect discrimination). On the ROC curve, J for each threshold is the vertical distance from that point of the curve to the diagonal; the threshold that maximizes J is often used as the 'optimal' cut-off, giving equal importance to sensitivity and specificity. In ecology the same formula is known as the True Skill Statistic (TSS) or the Hanssen-Kuipers score.
The Youden criterion weights false negatives and false positives equally and ignores disease prevalence. In practice the two errors usually have different costs (e.g. in screening, missing a patient is more costly); the weighted Youden criterion adjusts this balance using a cost ratio and prevalence. A cut-off chosen on the same data gives optimistic performance and should be validated on another data set. The point closest to the (0,1) corner of the ROC curve usually gives a similar, but not always identical, threshold.
An everyday example
A laboratory examines how well blood glucose separates people with and without diabetes. At 110 mg/dL, sensitivity is 90% and specificity 70% (J = 0.60); at 126 mg/dL, sensitivity is 80% and specificity 88% (J = 0.68). The 126 mg/dL threshold, with the highest J, is chosen as the best balance that treats the two errors as equally important.
In MerQur
One of the choices of the 'Cut-off criterion' option in ROC curve analysis; for the weighted Youden, the 'Cost ratio (FN/FP)' is used. The True Skill Statistic analysis gives the same formula as TSS and includes a Youden choice under 'Threshold criterion'; Confusion Matrix Metrics reports Youden's J.
Analyses where this concept appears
ROC CurveTrue Skill Statistic (TSS)Confusion Matrix Metrics
Related concepts
Classification threshold (cut-off)Sensitivity and specificity