A Note on Proximity Measures and Cluster Analysis

Clustering techniques and related approaches to numerical classification are beginning to receive a fair amount of attention by marketing researchers. Three articles on the subject [2, 9, 11] have already appeared in JMR, and a variety of marketing studies using clustering procedures have been reported in working papers. One of the principal problems in applying cluster analysis is the choice of what proximity measure to use in summarizing the similarity (or dissimilarity) of profile pairs. Morrison [10] discussed some problems associated with using a Euclidean distance measure in the space of original variables, a point also made by Overall [12] in the psychological literature. This article shows some of the interrelationships among various measures that have been suggested for summarizing pairwise proximities and to demonstrate that clustering results are not invariant over these alternative measures. Despite the arguments for using one measure in preference to another, we believe that no dominant proximity measure currently exists, given such high variation in the researcher's objectives [5]. The ten proximity measures used in this comparative study follow: