IDENTIFYING CLUSTER OVERLAP WITH NORMIX POPULATION MEMBERSHIP PROBABILITIES

Prior research has suggested that the NORMIX clustering procedure performs poorly under conditions of cluster overlap. In each previous study, however, the criterion of clustering accuracy specified that an entity be assigned to the single population from which it was sampled. NORMIX population membership probabilities were used for this purpose by assigning entities to the single population of highest probability. It is argued in this research that, when populations overlap, a more appropriate criterion of accuracy would recognize that entities which fall in the overlapping regions of populations either belong to multiple populations or they belong to a single population whose identity is indeterminate. It is further argued that NORMIX probabilities can be useful in identifying the entities from such overlapping regions. The results of a Monte Carlo test support this argument.