st: Re: Predicted values where a subgroup of variables are held constant

["Eric Uslaner" <euslaner@gvpt.umd.edu>]

Not a good idea.  My colleague Mike Hamner, together with Ozan Kalkan (a recent Ph.D.) have a paper under review in which they show that using mean values rather than actual values can lead to nonsensical results.  This is particularly critical in the case of dummy variables such as gender, where the mean is nonsensical (half man, half woman).  But it can also lead to predictions for cases that may not exist in the real world: The NY Times recently sought to find who is the happiest person and they found that it is a man who is 65 years or older, is both Asian-American and Jewish, and who owns his own business.  Well, they actually found one person who fits this profile, but undoubtedly there aren't too many.

You are far better off using the "natural" values.  I wrote do files that handle this for probit/logit and mlogit rather easily.  You can e-mail me for them.




*
*   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/