Since its initial invention by the French worker Milian in 1901, the bleeding time has been put forward as a clinically useful test in three contexts: diagnosis (particularly of platelet disorders), prediction of clinically important bleeding, and assessment of the adequacy of various forms of therapy. Attempting a complete review of the published experience with this test, we assessed 862 articles. Original bleeding time data appeared in 664 of these articles, from which we tabulated 1083 distinct studies in humans. ROC analysis, which characterizes the sensitivity and specificity of the test, was applied in every instance in which published data were adequate (34 studies). ROCs from 27 studies of the bleeding time in association with aspirin ingestion reveal high variability in the ability of the bleeding time to detect aspirin intake, and provide evidence against claims that recently devised bleeding time methods have improved discriminatory ability based on improved reproducibility. Two ROCs from surgical studies, in which the bleeding time was used to try to predict abnormal bleeding, were statistically indistinguishable from that of a completely noninformative test. In ROCs from five studies of abnormal bleeding in uremia, the test performed approximately the same as the platelet count or hematocrit (taken singly); in one of these studies, prothrombin consumption was determined and was a better predictor of bleeding than bleeding time, hematocrit, or platelet count. In the settings of renal biopsy (one study) and massive transfusion (one study), data allowed estimation of predictive value: in no instance was there evidence that the bleeding time significantly altered a priori estimates (based on prevalence) of the risk of bleeding. Linear regression analysis was applied to data from 23 studies relating platelet count to bleeding time, to assess published claims that the bleeding time and platelet count follow a predictively useful linear relationship. In 22 of 23 instances, the inverse relationship between bleeding time and platelet count was associated with broad statistical scatter, making it impossible to predict precisely one variable given the other. The pathophysiology of an abnormal bleeding time remains poorly understood. The bleeding time is affected by a large number of diseases, drugs, physiologic factors, test conditions, and therapeutic actions, not all of them platelet-related. The test is likely to remain widely used for the diagnosis of inherited disorders of platelet function, such as von Willebrand's syndrome, despite the lack of clear criteria for its use in this context.(ABSTRACT TRUNCATED AT 400 WORDS)