Minimum sample size for multinomial logistic regression

[Bruce Weaver]

Can anyone point me to guidelines regarding the minimum sample size
required for multinomial logistic regression? Does the rule of 15 or
20 still apply, as it does for binary logistic regression (i.e., one
needs at least 15 or 20 'events' per parameter)? Hosmer & Lemeshow's
discuss this issue briefly (2nd Ed., pp. 346-47), but it's not clear
to me whether their advice* extends to models with more than 2
categories for the outcome variable.

And of course, I am talking about the minimum sample size needed to
avoid over-fitting the model, not the sample size required to achieve
some desired level of power.

Thanks,
Bruce

* They give a rule of 10 events per parameter, based on the findings
of Peduzzi et al (1996). But to be fair about it, they advise
caution, and suggest that this is only a guideline that may need to be
modified depending on the nature of the variables, whether there are
interactions, etc.

--
Bruce Weaver
bwe...@lakeheadu.ca
http://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM."