Skip to content

Neg Binomial In GEE

4 messages · SamiC, Ben Bolker

#
Hi, I want to fit a GEE with a negative binomial distribution.  I have uesd
already a poisson glm and then neg binommial to deal with alot of
dispersion.  In my neg binomial residuals i have some patterns so i have
implemented a GEE, but only with a poisson family as i couldnt with neg
binomial.  However the residual patterns in fact look worse here.  When i
try and put neg binomial family it wants a value of theta??  I am using the
geepack package.

Thanks

--
View this message in context: http://r.789695.n4.nabble.com/Neg-Binomial-In-GEE-tp3627924p3627924.html
Sent from the R help mailing list archive at Nabble.com.
#
SamiC <s.cox.10 <at> aberdeen.ac.uk> writes:
Give us a (small, simple) reproducible example please?

  What do you mean by "some patterns"?
  What do you mean by "worse" (in terms of the residuals)?

  It is indeed the case that if you use family=negative.binomial
(from the MASS package) that you need to specify theta.  You could
try running GEE fits in a loop or an optimizer with a range of theta values and
selecting the one that maximizes some goodness-of-fit statistic
(this is what MASS::glm.nb does).

  I would suggest looking into Zuur et al's book on mixed models
in ecology to see if there is anything useful there.

  Ben Bolker
#
Thanks,
I have been using zuurs book but it only goes as far as poisson and binomial
GEE's.  Initially I fitted a glm with poisson and this was over dispersed. 
Then moved to binomial, but residual patterns are not great (ie variance). 
Looks like some spatial correlation.  However, in the GEE with poisson it
looks like i am stil having issues with over dispersion.


Also I am getting convergence errors once I have built the model so far. 
Not sure what to do with this either as I have all ready reduced variables
as much as possible and still havent finished model selection (ie. with AIC
and anova test).  

Thanks

--
View this message in context: http://r.789695.n4.nabble.com/Neg-Binomial-In-GEE-tp3627924p3628613.html
Sent from the R help mailing list archive at Nabble.com.
#
SamiC <s.cox.10 <at> aberdeen.ac.uk> writes:
How about using family=quasibinomial?  It has a different variance
structure from NB (var = phi*mu rather than var = mu*(1+phi*mu), and
NB is sometimes preferred because it has a slightly stronger foundation --
the NB parameterized as mu*(1+phi*mu) is in the exponential family
if phi is fixed -- but this is not so important if you are using
GEE anyway.
Well, if you have reduced the variables "as much as possible"
you may simply have a problem with not enough/poor quality data.
Sometimes that happens and you just have to simplify your model more
than you would like.

  Remember that if you are going to do model selection (throwing
away variables on the basis of some form of model goodness-of-fit)
then you should *not* make inferences on the basis of the parameters
in the selected model -- e.g. see Harrell's _Regression Modeling
Strategies_.

  Ben Bolker