[R] multinom(nnet) analogy for biglm package?

[Yeh, Richard C]

Dear Prof. Ripley,

Many thanks for your reply, especially during the holiday season!

> From: Prof Brian Ripley [mailto:ripley at stats.ox.ac.uk] 
> You seem to believe a multinomial logistic regression is a GLM: it is
not.

With one line, you've struck the heart of one of my problems.

As a follow-up question, would you (or anyone else) care to comment on
the feasibility of using R to perform multinomial logistic regression on
a large data set?

The scale of the problem is similar to the one you and Fei Chen treated:
instead of insurance policies, I am considering residential mortgages.
I read that you and Fei Chen used a generalized linear model, which I
understand you are saying does not apply to my approach.  But, it's not
obvious to me why I cannot just guess some coefficients, partition the
large data file into digestible excerpts, score each excerpt with a
log-likelihood function, combine the scores and gradients, and iterate
the guess.  Is that totally different from what you described in
"Statistical Computing and Databases" (Proc. DSC 2003)?

Thanks again for helping me as I consider how to approach my problem!

Richard

212-933-3305 / richard.c.yeh at bankofamerica.com


NOTICE TO RECIPIENTS: Any information contained in or attach...{{dropped}}