Sensitivity of the Complete-Link Clustering Technique to Missing Individuals

The effect of missing individuals upon the stability of the partition hierarchy yielded by the complete-link clustering technique was investigated. Empirical distributions of indices measuring the goodness-of-fit between partition hierarchies based on full and reduced sets of objects were obtained for data sets in which one, two, or three individuals were missing. The obtained results, based upon an object set obtained from an educational setting and four artificial sets, showed that the goodness-of-fit was a function of the number of missing individuals and the structure of the inter-object proximity matrix. In addition, a “key object” phenomenon was noted where the removal of certain individuals can result in low agreement between the hierarchies based upon the full and reduced set of objects. The results suggest that the complete-link clustering technique should be used with some caution, when less than the full set of objects is analyzed.