Helene <msc0235> asks:
> -----Original Message-----
> From: msc0235@geo.ed.ac.uk [mailto:msc0235@geo.ed.ac.uk]
> Sent: Saturday, June 22, 2002 1:54 AM
> To: s-news@lists.biostat.wustl.edu
> Subject: [S] mixed effect models
>
> Dear all,
>
> I would like to analyse a dataset of parasite abundance - this follows a
> negative binomial distribution, and comes in a hierarchical fomat
> (repeated
> parasite samples(10 samples) per habitat type (3 types) per host-abundance
> block (4 blocks)). Are mixed effect models possible in this case (given
> that
> it is count data, not really continuous data), or is another method more
> advisable?
[WNV] The negative binomial distribution can be thought of as
perhaps the simplest non-trivial case of a mixed-effects Poisson model. If
all you want to do is fit Negative Binomial models with fixed effects for
habitat type and blocks then you could use glm.nb( ) from the MASS library.
I think I would try this first.
If you genuinely want Negative Binomial models with some additional
random effects (to wit normally distributed components of the linear
predictor) life gets more interesting. If you already know the negative
binomial parameter theta (as used in glm.nb) then you can set up a glm
family with that fixed value of theta using negative.binomial(theta =
mytheta) and use glmmPQL to fit a mixed model. If you want to estimate
theta along with additional variance componets, you have a nice research
problem on your hands, I suspect. One possibility would be to make all
random components normal and fit a Poisson glmm with glmmPQL. [In other
words, if the ball is going in slightly the wrong direction - gently nudge
the goal posts! :-)]
Bill Venables.
> Thanks a lot!
>
>
> Helene
>
> --------------------------------------------------------------------
> This message was distributed by s-news@lists.biostat.wustl.edu. To
> unsubscribe send e-mail to s-news-request@lists.biostat.wustl.edu with
> the BODY of the message: unsubscribe s-news
|