Prior probability (the pretest best guess) affects predictive values of diagnostic tests.

Abstract

Authors who publish evaluations of dichotomous (yes/no) diagnostic tests often include the predictive values of their test at a single prior probability (eg, the prevalence of the target disease within the evaluation data set). The objectives of this technical note are to demonstrate why single-probability predictive values are misleading and to show a better way to display positive predictive values (PPV) and negative predictive values (NPV) for a newly evaluated test. Secondly, this technical note will show readers how to calculate predictive values from only sensitivity and specificity for any desired prior probability. As prior probability increases from 0% to 100%, PPV increases from 0% to 100%, but NPV goes in the opposite direction (drops from 100% to 0%). Because prior probabilities vary so greatly across situations, predictive values should be provided in publications for the full range of potential prior probabilities (if provided at all). This is easily done with a 2-curve graph displaying the predictive values (y-axis) against the prior probability (x-axis).