Important question on cluster analysis and market data research...

Due to a lack of rules, the only recommendation that can be given concerning sample sizes and variable numbers is to critically question if the dimensionality is not too high for the number of cases to be grouped. One hint can be deducted from literature on latent class analysis, where similar dimensionality problems occur. Formann (1984) suggests the minimal sample size to include no less than 2k

cases (k = number of variables), preferably 5*2k.

The number of clusters problem is as old as clustering itself (Thorndike, 1953). Clearly, the number of clusters chosen a priori most strongly influences the solution. Different approaches have been suggested to tackle the problem (Milligan, 1981; Milligan and Cooper, 1985; Dimitriadou, Dolnicar and Weingessel, 2002 for internal index comparison and Mazanec and Strasser, 2000 for an explorative two step procedure), but no single superior solution emerges.

These two may help:

search Formann [1984] on

http://ro.uow.edu.au/cgi/viewcontent.cgi?article=1286&context=commpapers

Pg 243 and 263 of a search on "Chapter 9 Cluster Analysis"

Best.

Thank you. I knew that number of observed objects must be higher than the number of variable by which the objects are classified. I found the above equation but without any explaination and source