Predictions from uncertain categorizations

Abstract

Eleven experiments investigated how categorization influences feature prediction. Subjects were provided with sets of categorized exemplars, which they used to make predictions about properties of new exemplars. Because the categories were provided for subjects, this method allowed a test of categorization and prediction processes, bypassing initial concept formation and memory. The experiments tested a Bayesian rule of prediction according to which (1) predictions of an object's features are based on information from multiple categories, and (2) features are treated as independent of one another. With one exception, the studies found evidence against both of these claims. Subjects did not generally alter their predictions as a function of information outside the most likely "target" category. In addition, feature relations had reliable effects on these predictions. We discuss the implications of these results for understanding how categories are used in drawing inferences.