How this instrument works
Youden's index (also written J or Youden's J statistic) combines a test's two most important error-avoidance properties, sensitivity and specificity, into a single figure: J = sensitivity + specificity − 1. Statistician W. J. Youden introduced it in 1950 as a way to rate a diagnostic test's overall discriminating power with one number instead of two, so that different tests, or different cutoff thresholds for the same test, could be ranked directly against each other.
The scale is built to be intuitive: J ranges from -1 up to +1. A useless test that performs no better than chance scores J = 0 — for instance a test with 70% sensitivity and 30% specificity nets to zero, correctly signalling that it adds no real diagnostic information despite looking reasonable on sensitivity alone. A perfect test that never misses a positive and never false-alarms on a negative scores J = 1, the maximum. Negative values indicate a test performing worse than chance, effectively inverted.
Youden's index is especially prominent in ROC (receiver operating characteristic) curve analysis, where a test's cutoff threshold can be dialled up or down to trade sensitivity for specificity. The threshold that maximises J is a commonly used rule of thumb for picking the 'optimal' operating point on the curve — the point where the combined benefit of both correct-classification rates is greatest.
- Enter the count of correctly flagged positive cases into True positives (TP).
- Enter the count of negative cases wrongly flagged as positive into False positives (FP).
- Enter the count of correctly cleared negative cases into True negatives (TN).
- Enter the count of positive cases wrongly cleared as negative into False negatives (FN).
- Read Youden's J index — sensitivity plus specificity minus one, ranging from -1 (worse than chance) through 0 (no better than chance) to +1 (perfect).
- Both TP+FN and TN+FP must be nonzero — the instrument needs at least one actual positive and one actual negative case in the matrix to compute sensitivity and specificity in the first place.
Worked example — the same 95-case confusion matrix
Enter TP = 50, FP = 10, TN = 30, FN = 5. Sensitivity is 50/(50+5) = 50/55 = 0.909091 and specificity is 30/(30+10) = 30/40 = 0.75 — the same two figures verified independently on this site's sensitivity and specificity instruments. Youden's index reads J = 0.909091 + 0.75 − 1 = 0.659091.
That combined score of roughly 0.66 sits well above 0 (no discriminating power) and reasonably close to the maximum of 1 (a perfect test), summarising in one figure that this particular test does a solidly good — though not flawless — job of separating positive cases from negative ones.
Questions
What does a Youden's index of 0 mean?
J = 0 means the test provides no diagnostic value at all — its sensitivity and specificity trade off exactly to net zero, which happens for any test that performs no better than chance at separating the two classes (for example, a coin flip, or a test where sensitivity and specificity sum to exactly 1). A higher J means more real discriminating power; J = 1 is the theoretical maximum, achieved only by a perfect test.
Why combine sensitivity and specificity instead of looking at them separately?
Sensitivity and specificity usually trade off against each other as a test's decision threshold changes — tightening the criteria to catch more positives (raising sensitivity) usually lets in more false alarms (lowering specificity), and vice versa. Youden's J gives a single number that rewards a test only when both rates are simultaneously good, which is useful for ranking tests or thresholds without having to compare two figures at once.
How is Youden's index used to pick an ROC curve cutoff?
As a diagnostic test's decision threshold is swept from one extreme to the other, sensitivity and specificity each change, tracing out the ROC curve. Computing J at every threshold and picking the threshold where J is largest is a standard, simple rule of thumb for choosing the 'optimal' cutoff — the point balancing the two error types most favourably, though the right threshold in practice also depends on the relative real-world cost of a false positive versus a false negative.
Can Youden's index be negative?
Yes. If sensitivity plus specificity is less than 1 — meaning the test is doing worse than chance at telling the two classes apart — J comes out negative, down to a minimum of -1 for a test that gets every single case wrong. In practice this usually signals that a test's positive and negative labels have been swapped, or that the classifier has genuinely learned to anti-correlate with the true outcome.
Is Youden's index the same as balanced accuracy?
They're closely related but not identical. Balanced accuracy is the average of sensitivity and specificity, (sensitivity + specificity) / 2, and is scaled from 0 to 1. Youden's J is that same sum minus 1, scaled from -1 to 1 so that 0 represents 'no better than chance' rather than 0.5. The two measures rank tests identically; they differ only in where the 'chance' baseline sits on the scale.