Sensitivity and Specificity

Sensitivity is the proportion of genuine cases a test flags — its true positive rate. Specificity is the proportion of non-cases it correctly clears. Both are properties of the test, measured against a reference standard; establishing that standard is the hard part of validating anything at all, and it is the Ground Truth the numbers depend on.

They are conditioned the wrong way round for a reader. Sensitivity is the chance of a positive result given the condition; someone holding a positive result wants the chance of the condition given the result, which depends on how common it is in the population tested — the pretest probability.

The arithmetic is unforgiving. Take 99% sensitivity and 95% specificity, applied where the condition is present in one person per thousand. Among 100,000 people, 100 have it and 99 are flagged; of the 99,900 who do not, 5% — 4,995 people — are flagged anyway. A positive result is correct about 2% of the time. Nothing is wrong with the test; the missing quantity is the base rate, and its absence is where the impressive-sounding 99% does its work through the Anchoring Effect.

A negative is treated even more carelessly. On an imperfect test it lowers the probability without eliminating it, so "the test was negative, therefore it isn't that" is an Exhaustive Claim the instrument cannot support. Hence the standard shape: a sensitive test to rule out, a specific one to rule in, confirmation on the positives.

Most tests have a threshold, and moving it trades one error for the other; where to place it depends on the cost of a missed case against an unnecessary follow-up. Optimising a screening programme for cases detected rather than outcomes is Goodhart's Law with clinical consequences, and a false negative is the purest form of Silent Failure.