The foundations of adult reasoning about probabilities are found in children's reasoning about frequencies. Adult probabilistic reasoning is impaired by heuristics based on typicality or representativeness, but development of these heuristics in childhood has not been studied. Previously, typicality has only been shown to enhance children's reasoning. In Experiment 1, however, children in Grades 3, 5, and 7 overwhelmingly judged subclasses to be larger than inclusive classes when the subclasses were typical of a given scenario. In Experiment 2 children explained these incorrect responses by reference to the frequency of the subclass in the context or by reference to the method by which the subclass is generated, in accordance with the representativeness heuristic. The same children explained correct responses by describing the inclusion relationship. Even when asked to select the most frequent class from three alternatives including typical and atypical subclasses and an inclusive class, children frequently selected the incorrect typical subclass. Experiments 3 and 4 investigated whether children could be taught to reason logically and to suppress the use of the representativeness heuristic. Both immediately, and 10 days after a brief training, children responded correctly far more often. Furthermore, children's explanations of their answers when they were untrained emphasized the typicality of the representative subclass, but after training they emphasized logical relationships. The results are attributed to a representativeness heuristic that is acquired early, and overcome through experience and training in logical thinking. © 1991.