Youden Index Calculator

Youden’s index, often called the Jaccard index, is a single number describing how well a diagnostic test distinguishes people with a condition from people without it. It is written J = sensitivity + specificity − 1 and takes values from −1 to 1.

W. J. Youden proposed it in 1950 as a way to summarise a test’s performance without being fooled by the class balance. Because sensitivity and specificity are proportions within their own groups, the index stays the same whether the condition is common or rare.

Start with the 2×2 table. Sensitivity is the share of people who have the condition that the test catches, so it is TP ÷ (TP + FN). Specificity is the share of people without the condition that it correctly clears, TN ÷ (TN + FP).

Add the two proportions and subtract one. Both sit between 0 and 1, so their sum sits between 0 and 2, and the index lands between −1 and 1.

The reading follows from that range. J = 0 is the point where the test is uninformative, J = 1 is perfect separation, and J < 0 means the test’s errors are systematic enough that reversing them would score better.

Maximising J over all available cut-offs is the practical use: the threshold chosen this way gives the smallest total of false negatives and false positives, weighted equally.

J is a summary of two rates, not a probability. An index of 0.7 does not mean the test is 70% accurate, and it is not a p-value or a risk.

The band is a rule of thumb from Youden’s own paper, useful for comparison between studies but not a regulatory threshold. Context still decides whether “good” is good enough for the decision being made.

Balanced accuracy is the average of the two rates and sits between 0.5 and 1. It is the same information as J on a rescaled axis, which makes it easier to read next to 50%.

Prevalence changes accuracy but not sensitivity or specificity. If your accuracy figure disagrees with the index, the difference is almost always the case mix in your sample.

Likelihood ratios translate the index into evidence: the positive ratio says how much a positive result raises the odds, and the negative ratio how much a negative result lowers them.

• The cut-off or threshold. J is defined at one operating point, and a different threshold moves both sensitivity and specificity in opposite directions.
• How the validation sample was chosen. Prevalence, spectrum of disease and referral patterns all shift the rates for the same underlying test.
• Sample size. The index is a ratio of counts, so a small study produces a noisy value. Confidence intervals are essential before comparing two tests.
• The scoring rule. Dichotomising a continuous score at its best cut-off flatters the index, because the threshold was chosen on the same data used to measure it.

• Comparing two diagnostic tests on the same population, since the index is independent of how common the condition is.
• Selecting a threshold for a screening or alerting rule, where the cost of a missed case and a false alarm are treated as equal.
• Reporting test performance in a paper or protocol, where a single number summarises discrimination alongside a confidence interval.
• Quality auditing of an existing rule, checking whether the current cut-off is still the best available.
• Training and teaching, because the index makes the difference between a prevalence-dependent measure and a prevalence-free one obvious.

• The index assumes the counts came from a validation sample that resembles the population the test will be used in. A shifted spectrum of disease changes both rates.
• It weights false negatives and false positives equally. Where one error is far more costly, weighted scores or a net benefit analysis are more defensible.
• It is a point estimate with no uncertainty attached. Always report a confidence interval, which widens sharply with small samples.
• It measures discrimination only, not calibration. A test can separate the groups perfectly while its risk estimates are badly biased.
• The threshold is usually chosen on the same data used to compute the index, which inflates it. Validate on fresh data before trusting the value.