NCJ Number
84969
Date Published
1981
Length
50 pages
Annotation
The paper describes a method of finding patterns in binary data and reports the results of experiments evaluating the procedure with artifical data.
Abstract
An alternative to conventional factor analysis of binary data is cluster analysis, involving the section of an index suitable for binary measures and an increase in the homogeneity of subsets among these measures through a process clustering together variables measuring similar properties. Experiments reveal that the clustering method works well but that it makes little difference on what index of association it is based. With realistic data, success rates of correctly identifying clusters were as high as 100 percent. This rate of success can be degraded by making the data less reliable and by shifting the criterion of success to one that includes correct identification of the number of cluster elements. By incorporating an objective rule of the number of such elements (the end of the cluster is the point at which the cluster-consistency index takes the largest drop), the success rate is reduced but not significantly. Data tables and about 15 references are included.