Predictive Values and Prevalence

#Start with the result, not the condition

Positive predictive value (PPV) is the proportion of positive results that are true positives: true positives divided by true positives plus false positives. Negative predictive value (NPV) is the proportion of negative results that are true negatives: true negatives divided by true negatives plus false negatives.

These reverse the starting point used by sensitivity and specificity. Sensitivity starts with people who have the condition, while PPV starts with people whose results are positive. Do not substitute one for the other when reading a statement such as "the test is 90% sensitive." A positive or negative result is not automatically a certain diagnosis.

Evidence: Shreffler and Huecker: diagnostic accuracy measures / FDA: reporting diagnostic test performance / MedlinePlus: understanding lab results

#A hypothetical population shows the base-rate effect

Consider a hypothetical test with 90% sensitivity and 95% specificity, applied to 10,000 people. Suppose 100 have the target condition (1% prevalence). Of those 100, 90 test positive and 10 test negative. Of the other 9,900, 495 test positive and 9,405 test negative. There are 585 positive results in total. PPV is 90/585, about 15.4%; NPV is 9,405/9,415, about 99.9%.

Now suppose 1,000 of 10,000 have the condition (10% prevalence), holding sensitivity and specificity fixed for this illustration. There are 900 true positives, 100 false negatives, 450 false positives and 8,550 true negatives. PPV is 900/1,350, about 66.7%; NPV is 8,550/8,650, about 98.8%. These are calculated examples, not measured performance of any real test or estimates for an individual.

Evidence: Shreffler and Huecker: diagnostic accuracy measures / CDC: diagnostic sensitivity and specificity

#Check which population the numbers describe

With sensitivity and specificity held fixed, a lower prevalence reduces PPV and raises NPV; a higher prevalence does the reverse. The example isolates that arithmetic. In real settings the accuracy measures themselves can change with the kinds of people tested, the threshold, the testing process and prior testing.

A study deliberately enriched with people known to have the condition need not have the same prevalence as routine testing. Its observed predictive values should not be carried straight into another setting. STARD asks for participant selection, settings, dates and participant characteristics so readers can judge whether the study fits their question.

Evidence: Shreffler and Huecker: diagnostic accuracy measures / STARD 2015: explanation and elaboration / FDA: reporting diagnostic test performance

#Population measures are not personal conclusions

The prevalence in a tested group is not automatically the likelihood for one person before testing. Clinical history, symptoms and the reason for testing help define that starting likelihood. MedlinePlus says a result is only part of the picture and may need other information or further tests for interpretation.

When reading a study, check its reference standard, confidence intervals, missing and indeterminate results, and whether its participants match the proposed use. A high predictive value does not by itself prove that a testing pathway improves health outcomes or makes its burdens acceptable. This page explains the measures; it does not tell anyone to start, repeat or stop a test.

Evidence: Shreffler and Huecker: diagnostic accuracy measures / MedlinePlus: understanding lab results / STARD 2015: diagnostic accuracy reporting / STARD 2015: explanation and elaboration

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.