Increasing the Statistical Power of Empirically Derived Taxonomies in Criminal Justice Research - Final Report

• NCJ Number: 90476
• Author(s): C Edelbrock
• Date Published: 1982
• Length: 116 pages
• Grant Number(s): 81-IJ-CX-0059
• Agencies: NIJ-Sponsored
• Publication Type: Research (Applied/Empirical)

Annotation

Computer evaluations of inverse factor analysis, nonhierarchical clustering technique, and centroid clustering were completed on 20 mulitvariate normal mixtures. A strong relationship was found between level of coverage and both accuracy of clustering solutions and the statistical power of cluster-based classifications.

Abstract

For each of the three clustering procedures, 100-percent coverage resulted in less than optimal accuracy in recovering underlying populations from the computer-generated mixtures. For all three methods, the accuracy of clustering solutions was substantially increased by leaving 11-25 percent of the subjects unclassified. Results suggest that accuracy of clustering solutions can be increased in the range of 55-85-percent coverage. For all the methods tested, increasing coverage about 85 percent had deleterious effects on clustering accuracy. Graphs, tables, and 75 references are included. Technical data are appended. (Author abstract modified)