🚀 go-pugleaf

Steve Blinkhorn cites John Collins as asking (I didn't catch the
original posting):

> what's wrong with factoring a matrix of dichotomous data?

One way to look at it is to ask the question "Do I think the underlying
variables here are really dichotomous, or are they continuous variables
that are measured with dichotomous indicators?"  It is fairly easy to
construct examples of the latter situation.  For example, suppose
a researcher administered a battery of personality scales, but the
results on each scale were converted to dichtomies.  The argument
would be that if you wanted your factor analysis solution to reveal
the true factor structure of personality traits, you would be better
off using a method that takes into account bias introduced by the
dichotomization.  However, most methods for doing this require
additional assumptions--e.g., that the underlying continuous
variables have a multivariate normal distribution, or at least
close to that.  Obviously that assumption is dubious in many
cases.  I don't know if anyone knows what happens to results if
the assumption is violated.

--John Uebersax
  Bowman Gray Medical School
  Winston-Salem NC (home of the fattest robins in the U.S.A.)
  uebersax@phs.bgsm.wfu.edu