Whether one calculates xb prior to forming the logit is a matter of
preference. If calculation of this intermediate value is not desired, then
the formula for the logit can be applied directly to the data. In the case
of a binary outcome, creating xb as a separate step may be considered
unnecessary. In the context of a multinomial logit with multiple outcomes,
however, it's probably a good idea to calculate either exp(xb) or xb as a
first step, since this reduces the total number of calculations performed.
There is no mathematical difference between the approaches illustrated by
Nick Cox and Alan Feiveson. (Note that there is a typo in the formula for
the variable p in Alan Feiveson's example -- you want exp(z) in both the
numerator and the denominator.) I suspect they were merely providing more
rather than less information. Many individuals benefit from answers that
show all the details and/or proceed in readily comprehensible steps. I know
I do.
As for using variables to store the coefficient values, I'm sure Alan was
not implying there are variables created automatically. He was, in fact,
arguing against the use of variables even though he realizes it is possible
to do so. Why on earth would you want to put coefficients in as variables?
I can think of one reason. Suppose you wanted to compare predicted
probabilities across strata or subsamples, but you also had different
coefficient estimates for these strata. Using variables *might* be more
convenient.
-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu]On Behalf Of Daniel Müller
Sent: Friday, November 22, 2002 1:18 AM
To: statalist@hsphsun2.harvard.edu
Subject: st: AW: RE: simulating data ( => logit case)