More on Choosing #Clusters in General
(not just k-means (fusion plot etc in chapter))
• Some researchers do their cluster analysis and then to demonstrate that the resulting
clusters are “significantly” different, they run a (one-way) anova and voila, show the F is
– Well duh! The cluster analysis’s objective was to find groups that were maximally
• Take a look at Milligan & Cooper (1985). They compared some 30 methods of trying to
determine the proper #clusters. They found 3 criteria that produced good results: a
pseudo F (Calinski & Harabasz 1974), a J statistic (Duda & Hart 1973), and CCC, the cubic
clustering criterion. The 1st and 3rd of these are displayed in SAS (Proc Cluster).
• For example, the pseudo F:
pseudoF 
C 1
N C
N=#observations (sample size)
C=#clusters (at a particular level of the clustering hierarchy)
Look at the eqn: it’s basically MSbetween/MSwithin
so larger is better, and of course, need to factor in that it should get better w >C
If multivariate normal, distributed F on p(C-1) & p(N-C) df (where p=#vars),
And can compare F across # C’s to find optimal C
More on Choosing #Clusters in General
• References
– Breckenridge, James N. (2000), “Validating Cluster Analysis: Consistent Replication
and Symmetry,” Multivariate Behavioral Research, 35 (2), 261-285.
– Calinski, R. B. and J. Harabasz (1974), “A Dendrite Method for Cluster Analysis,”
Communications in Statistics, 3, 1-27.
– Krolak-Schwerdt, Sabine and Thomas Eckes (1992), “A Graph Theoretic Criterion for
Determining the Number of Clusters in a Data Set,” Multivariate Behavioral
Research, 27 (4), 541-565.
– Milligan, Glenn W. and Martha C. Cooper (1985), “An Examination of Procedures for
Determining the Number of Clusters in a Data Set,” Psychometrika, 50, 159-179.
– Steinley, Douglas and Michael J. Brusco (2011), “Choosing the Number of Clusters in
K-Means Clustering,” Psychological Methods, 16 (3), 285-297.

similar documents