SUMMARY A new clustering method for binary data is proposed. The method is based on the new concept of homogeneity within a set of two or more operational taxonomic 1lnits. It should entirely replace the previous procedure of calculating a similarity coeHicient, in which the results need to be worked out by a second, logically entirely dilarerent, method of cluster analysis. Homogeneity ill this sense may be considered as a generalized measure of similarity. Moleover, a probability value is associated with every possible cluster and only statistically significant clusters are considered. Probability being a scalar quantity, the method allows two-dimensional graphic representation of the results without loss or distortion of information as in the methods previously proposed. 1. Introductioll A large number of basically similar techniques have been developed in the last 20 years under the name of 'numerical taxonomy'. The purpose of numerical taxonomy can be briefly defined as the construction of objective clusters of units by means of a quantitative measure of their affinity. Its name comes from the fact that the first methods were proposed for, and essentially applied to, the biological classification. However, non-numerical taxonomists still doubt the possibility of producing a purely numericallybased classification if a priori assumptions on the relative primitiveness and Oll the relative values of the characters employed need to be made. We shall not discuss this point here, but we emphasize that numerical taxonomic methods present a very powerful multiple comparison instrument insufficiently known or used in several disciplines, ranging from ecology to archaeology and from biostratigraphy to psychiatry. Sneath and Sokal (1973) give a comprehensive review of the available methods and of their published applications. A numerical taxonomic analysis aims to construct clusters of particularly similar objects among a set of objects to be classified (termed as OTtJs, 'operational taxonomic units'). This is usually done in two different steps. First an L x L symmetric similarity matrix is calculated by means of a similarity coefficient, to measure the similarity between every possible pair among the L OTtJs to be studied. Similarity being homologous to distance, the similarity matrix will occupy an (L- 1)-dimensional hyperspace. For sets including more than a few OTtJs, a similarity matrix obviously becomes hard to read and impossible to understand. For this reason, a second procedure, called cluster analysis, is employed to allow a two-dimensional graphical representation of the