Youden (1950) Index for Rating Diagnostic Tests

Summary

Youden proposes a single-number index — now universally called Youden's $J$ — for rating a binary diagnostic test from its $2\times2$ table of results on known-diseased and control groups. The index is $J = \text{sensitivity} + \text{specificity} - 1$: the average of the two group-specific success rates, or equivalently the sum of the correctly-classified proportions minus one. It ranges from 0 (a worthless test that yields the same positive rate in both groups) to 1 (a perfect test with no false positives or negatives), and — crucially — is independent of the relative and absolute sizes of the two groups, making it a clean summary for comparing tests and detecting improvements.

Key Claims

• The need. No test discriminates perfectly, so one needs an objective index both to rate tests and to decide whether two tests (or a modification of one) really differ in discriminating power.
• Construction. For the diseased group, charge the correctly-classified proportion $a/(a+b)$ against the misclassified $b/(a+b)$, leaving $(a-b)/(a+b)$ as the group's success measure (zero when $a=b$, i.e. no discrimination). The analogous control-group measure is $(d-c)/(c+d)$. Averaging the two gives $$J=\tfrac12\!\left[\frac{a-b}{a+b}+\frac{d-c}{c+d}\right]=\frac{a}{a+b}+\frac{d}{c+d}-1=\text{sensitivity}+\text{specificity}-1.$$
• Equal weighting. The average form assumes false positives are as undesirable as false negatives; Youden is explicit that the relative seriousness of the two error types is not a statistical matter.
• Properties. (1) $J\in[0,1]$; (2) $J=0$ whenever the test gives the same positive rate in both groups (worthless regardless of that rate); (3) $J=1$ only when both false positives and false negatives are absent; (4)–(5) $J$ is invariant to the relative and absolute sizes of the diseased and control groups.
• Caveats he flags. Whether controls are truly disease-free, and confounding (e.g. a test responsive to age rather than disease) — problems "no statistical protection" fixes except by matching the control group to the diseased group in all respects but disease.

Concepts Introduced or Extended

• Diagnostic Accuracy

Entities Mentioned

• W. J. Youden

Quotes

"The purpose of an index of performance is to reduce a table of data ... into one figure that will adequately characterize the diagnostic test."

"The index has the value zero whenever a diagnostic test gives the same proportion of positives for both diseased and control groups regardless of what that proportion is. Such a test is obviously worthless."

My Take

A three-page classic that gave clinical statistics one of its most durable summary numbers, and its logic still repays reading: $J=\text{sens}+\text{spec}-1$ is exactly the vertical distance from the operating point to the chance diagonal, which is why the "Youden index point" — the ROC threshold maximizing $J$ — remains the default cutpoint-selection rule decades later. The two features Youden stresses are the ones people still rely on: prevalence-invariance (unlike accuracy or predictive values, $J$ doesn't move with the diseased/control mix) and the equal-error-cost assumption baked into the simple average, which he is refreshingly candid is a value judgment, not statistics. For the wiki it is the elementary, gold-standard counterpart to the harder no-gold-standard latent-class estimation of sensitivity/specificity (Albert-Dodd, Vacek): $J$ is what you compute once you have sensitivity and specificity, whatever route produced them.