One approach might be to -collapse- the data, thus ignoring all the fine
structure of missingness. A simple default -collapse- would produce
means of 152 "statements". For that a formal cluster analysis would seem
pointless as you have a distribution that could be tabulated and plotted
to see all the variability.
A variant on that would be to -collapse- to two or more summary
statistics and then combine datasets to get a composite dataset
structure of
summary statistic list * statement list
The real motivation behind cluster analysis is often unclear to me and
it's certainly not clear to me in this case, so better advice could
depend on a finer specification of the scientific problem here.
Nick
Walter R. Paczkowski, Ph.D. wrote on Wed 29 Jul 2009 11:08:34 -0400
A client has a dataset from a survey in which consumers were shown a
randomly selected set of 25 needs statements from a total of 152
statements. Each consumer saw only 25. The client want to cluster the
152 needs statements (i.e., 152 variables). Since the 25 were selected
at random, this should be a Missing Completely at Random problem. But
with each consumer responding to only 25, each record will have 127
missing values. I assume that Stata's clustering routines will do
list-wise deletion so there should be no data available for clustering.
Does anyone have any ideas how to handle this? Any suggestions? Can a
similarity matrix still be created (how?) with so many missing data points?
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
