This article discusses limitations inherent in using unilevel behavioral descriptors as the basis for the construction of psychiatric classifications. As an alternative we consider the advantages of the polythetic classification strategy as suggested by Sokal and Sneath (1963) in their description of a “numerical taxonomic” approach. They argue for the development of naturally occurring, empirically defined classes rather than categories based on single domains of behaviors such as psychiatric symptoms. Typically the research into a topic such as schizophrenia starts with a behaviorally defined class of individuals who show schizophrenic symptoms, contrasted with a nonschizophrenic control group. These groups are then tested for aberrations in other systems of functioning (blood, neurophysiological, hormonal, attentional functioning, and the like). The approach offered here requires a classificatory strategy that treats data from all levels of functioning as equivalent in forming the classes of individuals rather than as secondary to the defining symptomatic attributes. Particular emphasis is given to the choice of measures with sufficiently broad and sensitive features to find natural classes. A set of standards is given for the selection of a test battery that might be used in a polythetic approach to the investigation of schizophrenia. Finally, the utility and limits of statistical clustering techniques for handling multilevel data banks are discussed.