Statistics Glossary → Positive and negative predictive values (PPV, NPV)
Positive and negative predictive values (PPV, NPV)concept
The probability that someone is truly positive when a diagnostic test or classifier says 'positive' (PPV), and truly negative when it says 'negative' (NPV).
Sensitivity and specificity describe a test's performance among people whose true status is known: how many of the sick it catches and how many of the healthy it correctly clears. PPV and NPV turn the question around and reflect the viewpoint of someone who knows the test result: PPV = TP / (TP + FP), NPV = TN / (TN + FN). For this reason they usually carry more direct meaning for clinical decisions.
Unlike sensitivity and specificity, PPV and NPV depend strongly on how common the condition is (prevalence). The same test can give a high PPV in a hospital clinic where prevalence is high and a low PPV in a general screening programme where prevalence is low, because when healthy people are numerous even a small false-positive rate produces many false alarms. PPV and NPV found in one study therefore cannot be carried over directly to populations with a different prevalence; they must be recomputed for the new prevalence with Bayes' theorem. If the sample comes from a case-control design (where prevalence was fixed artificially), PPV and NPV should not be interpreted directly.
An everyday example
A test with 90% sensitivity and 95% specificity is used in a community where 10 in every 1000 people have the disease. The test catches 9 of the 10 sick people but also wrongly flags about 50 of the 990 healthy ones. Of the roughly 59 people who test positive, only 9 are actually ill: the PPV is about 15%. Almost everyone who tests negative, on the other hand, is healthy (NPV about 99.9%).
In MerQur
In the ROC Curve, the 'PPV/NPV at the threshold and CI' extra output gives them at the chosen threshold with Wilson confidence intervals. Confusion Matrix Metrics gives PPV and NPV with Wilson intervals, and Logistic Regression gives them in the classification table output.
Analyses where this concept appears
Logistic RegressionROC CurveConfusion Matrix Metrics