SEARCH

SEARCH BY CITATION

Reply: —I appreciate Peirce and Gerkin's comments and welcome the opportunity to respond.

Peirce and Gerkin's distinction between probability and likelihood is unnecessary when applied to diagnostic testing (the subject of my paper), because unlike measures of therapeutic effect, diagnostic testing admits only 2 possible hypotheses: the proposed diagnosis of interest is either present or it is absent. These hypotheses are not random, and therefore the likelihood ratio (LR) reduces to only 2 possible probabilities: the probability of a finding if the disease is present divided by that of the same finding if the disease is absent.

“Odds” have always appealed to epidemiologists and statisticians because of their attractive mathematical properties, but whether clinicians use “probability” or “odds” to describe the likelihood of a particular hypothesis is only a matter of personal taste. Most of us are more comfortable using “probability.” Similarly, LRs are nothing more than diagnostic weights, describing how much a particular finding argues for or against a particular hypothesis (and thus changes “probability” or “odds”). The beauty of LRs is that they place multiple findings (or tests) on the same scale, thus allowing clinicians to quickly compare them and identify the most discriminatory tests for (or against) a particular diagnosis and discover the clinical setting in which a finding is most discriminatory.

The only limit mentioned in my paper is that pretest probability must be between 10% and 90% for my method to work. Therefore, my method should not be used to combine large LRs (e.g., LR = 10) with pretest probabilities less than 10% (e.g., screening tests) nor should it be used to combine tiny LRs (e.g., LR = 0.1) with pretest probabilities near 100% (a clinically unlikely scenario). For virtually all clinical problems of symptomatic patients, however, pretest probabilities do range between 10% and 90%, and if clinicians unfamiliar with odds desire to gauge how much the value of a particular LR changes this pretest probability, the method described in my paper accomplishes this simply and accurately.—Steve McGee, MD,University of Washington, Seattle, Wash.