Pairs of hypothetical medical and non-medical problems were given to 44 pediatric residents at three levels of hospital training. Each problem was designed to detect a specific heuristic-based bias in making diagnoses. Discounting, disregarding base rate, and over-confidence in contextually embedded redundant information were more evident on medical than on non-medical problems. In particular, a greater number of third-year residents disregarded base-rate information than did first- and second-year residents on medical but not on non-medical problems. On medical problems, a greater number of first-year residents expressed greater confidence in redundant information that was contextually embedded than in information that was presented in a listed format. Over one-third of the residents confused prospective and retrospective probabilities; three-fourths showed evidence of augmentation; virtually all residents expressed greater confidence in a diagnosis based on redundant rather than on non-redundant listed information. These latter effects were consistent across training level and occurred on both medical and non-medical problems. The results are discussed in terms of prototype theory and the nature of medical training.