[R-sig-ME] glmm with a tweedie distribution

[Ben Bolker]

Rubén Roa <rroa at ...> writes:

>  I wouldn't call it ad hoc. The power parameter p in the variance
> function that defines the Tweedie family of exponential
> distributions, v(mu)=phi*mu^p, can be estimated via profile
> likelihood, and then the maximum profile likelihood estimate of the
> p parameter can be inserted in the glmm, essentially estimating the
> glmm by an estimated likelihood. So there are two stages of
> approximation but the approximation methods are not ad hoc, they are
> pretty much mainstream approximation methods to complex
> likelihoods. Here is a pseudo code using the tweedie package and
> glmmPQL from MASS (plus msm). For a published application you can
> see Tascheri, Saavedra-Nievas, Roa-Ureta. 2010. Statistical models
> to standardize catch rates in the multi-species trawl fishery for
> Patagonian grenadier (Macruronus magellanicus) off Southern
> Chile. Fisheries Research 105:200-214.

  For what it's worth, the upcoming/development version of 
lme4 (now on R-forge) should work with custom family arguments,
so this approach *should* be possible with glmer as well as
glmmPQL ... (but I would also definitely give a thumbs-up
to cplm, which looks quite powerful).

  Anyone who wants to do some http://glmm.wikidot.com/faq - editing
is welcome ...