A Study of the Beta-Flexible Clustering Method.

The beta-flexible clustering method has exhibited good recovery performance in the few comparative simulation studies that have included the procedure. Further, the beta-flexible technique has not been examined in a systematic manner over the range of parameter values that can be specified for the method. The present report was designed to provide a more complete study of the recovery characteristics of the method. Artificial data were generated with a wide range of cluster configurations. Each data set was subjected to a series of different techniques for the introduction of error. The results indicate that the beta-flexible method with β=-.25 or .2 generally produces recovery rates that are competitive with the group average and Ward's (1963) technique. When outliers are present in the data, values in the range -.7 ≤ β ≤ -.4 are needed. The beta-flexible method showed a more stable or robust performance pattern than the two competing procedures across the differing error conditions. Finally, results on the impact of reduced coverage levels as originally proposed by Edelbrock (1979) are reinterpreted in the light of an improved measure of recovery of true cluster structure.