The perception of risks (e.g., diseases, accidents, and natural hazards) is investigated using a multitask, multimodel approach. We studied the proximities among 18 risks induced by three tasks: judgment of similarity, conditional prediction, and dimensional evaluation. The comparative judgments (similarity and prediction) were reasonably close, but the dimensional evaluation did not correlate highly with either similarity or prediction. Similarity judgments and conditional predictions appear to be represented best by tree models, which are based on discrete features, whereas the dimensional evaluations are better explained by spatial models, such as multidimensional scaling and factor analysis. We discuss the implications of these results for the study of mental representations and for the analysis of risk perception.Much work in cognitive psychology is aimed at constructing formal representations of specific domains of knowledge or behavior. These representations are commonly constructed on the basis of observed data using some appropriate statistical, geometric, or computer model. Formal representations of psychological structures are generally incomplete because the data usually reflect only limited aspects of the process under study and because the assumptions that underlie the representations are approximate at best. The use of reaction time, error rate, or verbal protocols, for example, provides only a limited view of human reasoning. Analogously, the use of hierarchical clustering or multidimensional scaling to represent some semantic domain may exclude significant aspects of the data or impose extraneous features that are not present in the data. Although there are no general methods for avoiding errors of omission or commission caused by the selection of tasks and models, these errors may sometimes be reduced by the use of a multitask, multimodel