Re: st: "testing" a cluster analysis

[Steven Samuels <ssamuels@albany.edu>]

I agree with Nick's point about the distinct outcomes.  I would list  
all combinations of 2, 3, 4 ,5, 6, & 7 variables (start with 7). This  
is easily done after -contract- and -fillin-.  Changing the order of  
the variables in the list might be revealing. This approach will also  
show you associations: what responses almost always or almost never  
appear together. Validation of any "clusters" that you detect would  
require that you set aside a validation sample.

Latent class analysis is another approach that might be considered.

Steve

On Feb 7, 2007, at 9:59 AM, Nick Cox wrote:

I think Ron�n and Ken made excellent points.

I am also queasy about this for quite a different
reason. As I understand it, you have a discrete
outcome space with 2^7 = 128 distinct outcomes.
I am not clear that this lends itself to cluster
analysis, nor would calculating means be what springs
to my mind as natural.

In principle, you lose no information by tabulating the
frequencies of those 128 composite outcomes and sorting
the table.


*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/