Missing Data in Tryon's Cluster Analysis

Abstract

To explore the effects of different ways of treating missing data on the results of cluster analysis, six cluster analyses were calculated varying the total amount of missing data per participant and the way in which the missing data were treated. Variations in the amount of missing data (10%, 25%, and 50%) made little difference in the cluster structure, although there were some inversions of the order in which dimensions appeared. Variations in treatment of missing data (pairwise deletion versus conversions of blanks to zeros) also made little difference, although the appearance of one dimension among only those methods which converted blanks to zeros implies that treating blanks as missing may discard meaningful variance.