How to decide how many cluster groups to retain?

[Amy Mittner]

Hi,

Is there any statistical test, for which there is a stata command or user-written software, to guide the choice of how many clusters groups I should retain after cluster analysis? With factor analysis, there is at least the eigenvalue, that can give you an idea how many factors to retain.

After I have sent my manuscript to a journal, the reviewer required me to justify why I opted for this exact number of cluster groups. He was not happy with my response, that I chose x cluster groups merely after viewing the corresponding dendrogram.

Would appreciate any help. Thanks.

[Carlo Lazzaro]

Amy:
The reviewer is probably under the impression that your choice was sample (rather than theory)-driven.
Does the literature in your research field report some previous example on the number of clusters chosen for a research topic similar to yours?

[Amy Mittner]

Dear Carlo,

thanks for your remark. The data I am working on has never been studied before. As such, the cluster analysis was performed to explore rather than confirm.

Is there any metric, akin to the eigenvalue in factor analysis, or better any statistical tests outputting an "optimal" number of cluster groups to help decide how many groups of cluster to retain?

[Carlo Lazzaro]

Amy:
thanks for providing further details.
The following thread http://stats.stackexchange.com/quest...er-of-clusters seems promising.

[Brendan Halpin]

The cluster stop commands offer a number of stopping rules, in particular the Calinski-Harabasz and Duda-Hart indices. These are widely used to determine the optimal number of clusters.