Experience gained over the last decade with multivariate treatment of chemical data for provenance determination has enabled archaeometers to focus on efficient, reliable procedures for forming and refining compositional groups. At the same time, the sophistication and flexibility of commercially available computer programs for statistical analysis have increased substantially. These two trends have converged to the extent that many provenance studies can now use commercial statistical packages (such as SPSS, BMDP, and SAS) to aid in data interpretation. A scheme for the use of such packages is presented in this paper, with specific reference to methodological issues. The scheme was developed to treat chemical data on 417 samples analysed by neutron activation in a study of prehistoric pottery trade in the Aegean. The major part of the data analysis was accomplished with two multivariate programs from the BMDP statistical software package: centroid cluster analysis and stepwise discriminant analysis. These programs were used interactively, first to eliminate samples with extreme compositions ( outliers ) at individual sites, next to form tentative groups of similar composition, and finally to refine these groups by the testing (and possible reassignment) of every sample. A grey area of sample <i>association</i> but not <i>attribution</i> was defined, allowing the true attributions to be made with more confidence. This distinction and other unusual features of the strategy for group formation are described in detail and illustrated with examples. -- AATA