Predictions by housestaff physicians of whether or not patients would return for follow-up visits and of the amount of prescribed medication they would take were compared to measured compliance for 187 patients discharged from the medical service. Although physicians were able to predict visit compliance better than by chance alone, at best they could accurately predict only 35% of the noncompliers and one half of their predictions of noncompliance were incorrect. In predicting medication compliance, less than one half of physician predictions correctly discriminated between compliant and noncompliant patients and three fourths of their predictions of noncompliance were inaccurate. Because of physician limitations in this important clinical area, medical education should expand efforts to develop physician skills in diagnosing and managing sociobehavioral aspects of illness, and efforts to improve quality of care cannot ignore these "nontechnical" factors.