A clustering method is presented which is appropriate when a mixture of multivariate distributions is used to model the data. The procedure assumes that the measurements made on several independent characteristics follow probability distributions of different nature (e.g., normal, Poisson, multinomial). Likelihood methods are used to both estimate the parameters of the models and perform the cluster analysis. The procedure is applicable when the data are subject to censoring. A data set on the concentrations of volatile halocarbon contaminant in Lake Ontario is used to illustrate the methods of this paper. ¿ American Geophysical Union 1993