🚀 go-pugleaf

I have an interesting problem for which I would appreciate some
statistical backup from the literature.  To date my arguments have been
primarily "gut feeling" responses to the client in question and I would
really like to be able to put a more concrete argument together.

THE PROBLEM:

A researcher came by with a set of 125 obs and 50 vars.  We did a bit
of work and the results were published, mainly descriptive/summary.
The researcher returned with a letter from a "leading light" in the
area of research.  This LL wrote that she has two other colleagues
with datasets of 45, 47 and 17 obs each.  The datasets are from entirely
different parts of north america.  The LL wants my researcher to send
her dataset down so that the four raw datsets may be combined into one
dataset of 234 obs and run a factor analysis on the combned dataset.
I have advised them that this goes against the idea of a random sample
representativeness and a lot of other common sense arguments.  I have
also mentioned that doing factor analysis on 50 vars with a 5:1 obs/var
ratio is somewhat questionable as well.  Factor analysis on the
individual datasets is out of the question.

In talking with a number of social science researchers, I have found
that this practice of combining datasets is not uncommon.  This worries
me and I think is a result of a stats/social science schism.  My
primary goal here is to get more theoretical justifications for either
side of the argument.  Any comments or suggestions would be helpful.

Jeff Sloan
Statistical Consultant
Manitoba Nursing Research Institute
SLOAN@UOFMCC
(204)474-9080