Dissimilarity Analysis: a new Technique of Hierarchical Sub-division

Abstract

IN hierarchical classifications each sub-group may be formed from the splitting into two parts of a larger group, or alternatively from the union of two smaller groups. Both these procedures are repetitive, and in either case ‘false’ decisions (arising from the statistical variability of the data) made in the early stages of the analysis will distort its subsequent course. For this reason, divisive methods, which start with the whole sample, are in general safer than agglomerative methods. In the past, one attraction of agglomerative methods has been their flexibility; any two sub-groups could be considered for possible combination. With divisive methods, in all but the simplest cases some restriction is necessary on the possible subdivisions considered, since there are 2ⁿ⁻¹ − 1 ways of dividing n individuals into two groups. One procedure¹ is to consider only monothetic subdivisions, that is, those definable in terms of the possession or lack of a single attribute by the individuals concerned.