Three categorization experiments were run in which subjects saw a set of stimuli, varying on four continuous dimensions. Subjects first categorized the stimuli and tried to predict some of the dimensions given the values of others. The first experiment used Iris-like stimuli based on the descriptions of Fisher (1936). It showed that having subjects categorize the stimuli was essential to being able to perform the prediction task and that merely observing the stimuli was not sufficient. It also indicated that subjects could use within-category as well as between-category correlations for predictions. The second and third experiments used stimuli with artificial variations of values. Subjects processed categories that had different within-category correlations. Subjects' behavior could be predicted as a combination of sensitivity to within-category correlation and bias about the sign of the correlations. These results were fit to the rational model of categorization (Anderson, 1991) and an exemplar model (Nosofsky, 1988).