A test that always says positive catches every case and is useless
Imagine a bogus kit that reads positive for everyone. It never misses a sick person, so its sensitivity is a perfect 100 percent. It also flags every healthy person. That thought experiment shows why diagnostic tests are judged by two numbers, sensitivity and specificity, terms coined by biostatistician Jacob Yerushalmy in 1947.
Sensitivity asks: of the people who truly have a condition, what share does the test flag? Specificity asks the mirror question: of those who truly do not, what share does it clear? Every result lands in one of four boxes. Sick people correctly flagged are true positives, and those wrongly cleared are false negatives. Healthy people wrongly flagged are false positives, and those correctly cleared are true negatives. If 100 people known to have a disease are tested and 43 come up positive, sensitivity is 43 percent; if 96 of 100 healthy people test negative, specificity is 96 percent.
The two usually trade off, because most tests rely on a cut-off. Slide the threshold one way and fewer sick people slip through, but more healthy ones get flagged; slide it the other way and the reverse happens. In one illustration with 40 results on each side of the line, shifting the cut-off swaps the figures between about 91.4 and 82.2 percent. High sensitivity matters most when missing a case is costly; high specificity matters when a false alarm brings extra testing, expense, stigma or anxiety.
Neither number depends on how common a condition is; both are properties of the test itself. What does depend on prevalence is the predictive value, the chance that a given result is right. That is why the popular mnemonics SPPIN and SNNOUT, which claim a specific test rules conditions in and a sensitive one rules them out, are called inferentially misleading, though SNNOUT holds up somewhat when a condition is very rare in those tested.
Laboratories use the same words differently: analytical sensitivity means the smallest detectable amount of a substance. Statisticians combine the diagnostic pair into informedness, sensitivity plus specificity minus one, where zero means performance no better than chance.
Source: Sensitivity and specificity
