[R] negative binomial latent class regression in package flexmix

[Carson Farmer]

Hello list,

Has anyone had any luck creating an M-step driver for negative
binomial regression for use with package flexmix? I've had a look
here: http://cran.r-project.org/web/packages/flexmix/vignettes/flexmix-intro.pdf
as well as poking around in the flexmix source, but I haven't had much
luck getting anything to work. I can't figure out how to a) come up
with an initial estimate of theta, as I don't think this really
belongs in the M-step, and b) how to properly implement the M-step so
that refit() doesn't give me:
Error in .local(object, ...) :
  not implemented yet for restricted parameters.
(which I think is due to the additional theta parameter)? I realize
this is a bit of a vague question, but my hope is that there is
someone out there who has experience working with NB in a latent class
setting that can help me out here. Frankly I'm pretty stumped on this
one, and I don't think posting any code in this case will enlighten
any one!

Thanks in advance for any tips or pointers.

P.S. I'm not married to using flexmix, though I do like the
implementation and extensibility so far.

-- 
Carson J. Q. Farmer
ISSP Doctoral Fellow
National Centre for Geocomputation
National University of Ireland, Maynooth,
http://www.carsonfarmer.com/