Sensitivity and Specificity

#Two measures with different denominators

Sensitivity is the proportion of people with a defined target condition whose test is positive. Specificity is the proportion without that condition whose test is negative. A false negative is a negative result in someone with the condition; a false positive is a positive result in someone without it. The diagram shows these four categories, not the performance of a particular test.

Sensitivity is true positives divided by true positives plus false negatives. Specificity is true negatives divided by true negatives plus false positives. These are different groups. A study sensitivity of 90% does not mean that 90% of positive results identify the condition: that is a question about positive predictive value.

Evidence: FDA: reporting diagnostic test performance / CDC: diagnostic sensitivity and specificity

#The comparison must establish the target condition

A diagnostic accuracy study compares the test being evaluated with a reference standard: the best available method for deciding whether the target condition is present. That method can combine tests and clinical information, and may itself have limitations. Read what condition was defined, how the reference standard worked and who received it.

FDA distinguishes this from comparison with another test that is not a reference standard. Such a study can report positive and negative agreement rather than directly establish sensitivity and specificity. Two tests can agree and both be wrong. If only selected participants receive the reference standard, the analysis must address possible verification bias.

Evidence: FDA: reporting diagnostic test performance / STARD 2015: explanation and elaboration

#Thresholds and participants change the estimate

A test with a continuous measurement may use a cut-off to label results positive or negative. Its sensitivity and specificity belong to that threshold, not to the instrument in the abstract. STARD asks authors to explain the cut-off and distinguish one chosen in advance from one selected after looking at the results. Choosing the best-looking cut-off in the same data can give an overly optimistic estimate.

The study population also matters. A study of obvious severe cases and very healthy controls may omit the harder cases seen in practice. STARD explains that accuracy can vary between patient groups, settings and prior-testing pathways. The estimates should therefore be read with the recruitment method, disease spectrum and intended use.

Evidence: STARD 2015: explanation and elaboration / FDA: reporting diagnostic test performance

#Read uncertainty and incomplete results together

Sensitivity and specificity from a sample are estimates. Look for the counts behind the percentages and their confidence intervals, not only a headline number. A larger sample may reduce sampling uncertainty, but cannot by itself remove systematic bias in recruitment, testing or the reference standard.

Indeterminate results, failed tests and missing results matter too. Excluding them can distort the apparent performance when their occurrence is related to the condition. STARD asks authors to report how these cases were handled. Neither sensitivity nor specificity replaces the other; neither by itself supplies a personal diagnosis. MedlinePlus explains that lab results must be considered with symptoms, history and other information.

Evidence: FDA: reporting diagnostic test performance / STARD 2015: diagnostic accuracy reporting / STARD 2015: explanation and elaboration / MedlinePlus: understanding lab results

Source note

The sections above were checked against the linked sources. No clinical review has been performed. This is general research education, not a clinical guideline.