multinomial logit can be estimated as a series of independent logits
(unless you induce some parametric dependence). so the same issues apply.
the min number of observations per category depends only on whether you
wind up with perfect discrimination. in that case, the whole model falls
apart and its not identified (the constant term goes to infinity, and so
the other parameters can take on any value and you get the same
likelihood).
Gary
On Thu, 10 Feb 2005, Yanbo Wang wrote:
Hi folks,
Sorry for troubling you guys. Hopefully somebody here can give me some advice on
the method of multinominal logistic model. Your help will be deeply
appreciated. Best, Yanbo
Question: What is the minimum number of observations required per category given
n independent variables? Is the rule the same in MLR as it is for a logistic
regression model (~20 observations per independent variable) for the entire MLR
(i.e., all of the nominal categories combined) or, since each nominal category
is a separate equation, does that imply that the number of observations per
category be at least 20*(the number of independent variables).
_______________________________________________
gov2001-l mailing list
gov2001-l(a)lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l