Joseph's answer may be what you are looking for, but here as often I am
puzzled by the tacit assumption that there is always a technical fix for
a technical question regardless of the problem-situation and what the
researcher is trying to do.
Instead of asking "what do I do?" why not ask yourself whether feeding a
mix of variables to a factor analysis is a reasonable thing to do given
your scientific or practical objectives. As you don't state the latter,
it is difficult even for those people who may have similar objectives or
experiences to comment.
Otherwise put, Stata doesn't in any sense know that your variables are
on different measurement levels; that's largely in the researcher's
head.
Similarly, Stata doesn't know that you are "assuming" approximate
normality.
Nick
[email protected]
Joseph Coveney
You can use -polychoric-, a user-written command available from SSC that
handles
datasets like yours. There's a chance that the matrix will not be at
least
positive semidefinite, but you can use the -forcepsd- option of
-factormat- for
that.
Stefan Duke wrote:
I have question concerning factor analysis on variables with different
measurement levels.
The questionnaire consists of binary and ordinal variables. If I would
have just binary variables, I would use the tetrachoric correlation
coefficients. For the ordinal I assume approx. normality and then use
the ordinary factor analysis capability.
But what do I do when I have both variables? Is it an option to
construct the variance-covariance matrix by hand? And what do I take
for the correlation between binary and ordinal?
Maybe is there a model class which takes care of that, that yields
similar outcomes as factor analysis but can deal with such kind of
data (e.g. correspondence analysis).
*
* 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/
