An empirical demonstration of the problem of cluster dissimilarity from different clustering methods in a single sample

Abstract

This study utilized the four most commonly employed clustering techniques (CLINK, SLINK, UPGMA, and Ward's) to illustrate the dissimilarity of cluster group membership (based upon short-form MMPI scale scores and a measure of alcohol dependency) between partitions in a sample of 113 impaired driving offenders. Results, examined with the Rand index of cluster comparison, demonstrated that cluster group membership can be so different between alternative clustering methods as to equal chance assignment. Cautions are given with regard to the use of cluster analysis for other than exploratory work. In particular, psychologists are cautioned against attempting to use cluster analysis based upon personality inventory scores (which can never be wholly reliable or discrete) for patient classification.