A procedure is presented for finding clusters of individuals in situations where every individual is classified with respect to each of a set of attributes, which may have ordered states (interval or ordinal) or nonordered states. The procedure is based on a probabilistic measure of similarity generated entirely from the given data. It can use but does not require measurement data, permits mixtures of kinds of attributes, and has been successfully used to reveal nonindependence of attributes as well as to cluster for taxonomic purposes. It appears to give results with measurement data nearly identical to those of methods that require such data.