Specificity (True Negative Rate)

Specificity is the mirror of sensitivity, read down the other column: of everyone who does not have the condition, what share did the test correctly clear? Like sensitivity it is a property of the test measured against a reference standard, and like sensitivity it is blind to how common the condition is.

Yerushalmy named this one too, in the same 1947 chest-radiograph paper, and the same correction applies — he supplied the term, not the concept. It is worth pausing on what he was actually studying, because it is unusually relevant: he was measuring how often two competent radiologists disagreed with one another about the same film. The pair of terms was born in the observation that the reference standard is itself made of fallible readers, which is a caution the terms have since shed. Where no true gold standard exists, the honest measure is agreement rather than accuracy, and that is Cohen's kappa.

The companion aid is SpPin: a highly Specific test, when Positive, rules in. This is where specificity earns its keep, and it is the number that matters most in screening. Sensitivity governs how many cases you find; specificity governs how many healthy people you frighten, biopsy and follow up. Because a screening population is overwhelmingly healthy, a small deficit in specificity is multiplied by a very large denominator, and 1 % of a million is ten thousand people.

The trade-off between the two is not a matter of judgement but of where a threshold sits. Any test that outputs a continuous quantity — a titre, a densitometry score, a marker concentration — becomes a yes-or-no test only when a cut-off is imposed, and moving that cut-off buys sensitivity with specificity at an exchange rate the ROC curve draws out. There is no setting that improves both. Improving both requires a better test, not a better cut-off.