[R-sig-ME] profile and/or bootstrapped CIs for GAMM random effects
Ben Bolker
bbolker at gmail.com
Thu Sep 3 22:14:44 CEST 2015
John Morrongiello <john.morrongiello at ...> writes:
[snip snip snip]
> In regards to the Gamma/log-link model: I decided to go down this
> path as my response variable is distance from 'home' (longdist)
> which is strictly positive and continuous. I realise I could
> log-transform the data, but thought it good to just use an
> appropriate distribution. I'm happy to be convinced to go down the
> transformation path as it would help with other response variables
> that are continuous but include zeros (here I've had to split the
> analysis into two steps: 1) a binomial GLMM for zero or >0 distance;
> 2) a gamma GLMM for all distances >0)
I think assuming log-normality (which is essentially what you're
doing when you log-transform and fit a LMM -- the only thing you have
to be careful about is comparing (log)likelihoods with models on
non-transformed data, since you have changed the scale of the data) is
not on average any better or worse (more or less correct) than
assuming a Gamma distribution. There is some work on comparisons/
reciprocal robustness (referenced in McCullagh and Nelder, I don't
remember the ref. right now).
Unfortunately transformation won't help you with the zeros in
the response, as they will be 'illegal' for the log-transformation
as for the Gamma likelihood. You'll still have to do the two-stage
model or something else a little bit fancier than a plain mixed model.
More information about the R-sig-mixed-models
mailing list