Cluster Analysis: A Brief Discussion of Some of the Problems

Abstract

The availability of computer packages for classification has led to many psychiatrists (and others) using the methods on their data without much thought as to the problems involved. This can lead to results and conclusions which, to say the least, are suspect. Much more work needs to be done on the clustering techniques in use, such as applying them to sets of data of known structure, to investigate which of them are `good' and which `bad' methods (at the moment I would suggest that most methods are nearer the `bad' than the `good' end of the scale). Meanwhile if an experimenter decides to use any one method he will do well to bear in mind the remarks of the previous paragraph before coming to any firm conclusions.