Hi,

I am trying to fit growth functions on several levels of a fish population (global, by habitat, by size at maturation) by non-linear mixed effects models (since the age/length data originates from repeated measurements, i.e. back-calculations). Following Burnham & Anderson, I'd first like to select a model (e.g. Bertalannfy, Gompertz, Log...) that fits the data well (based on AIC, BIC, LL, RSE, whatever...).

1. On which level would the best model be selected?

Certainly one would start with the global dataset to select a model and assuming that the general process describing growth is the same, this should apply to lower levels as well. But here's the catch: I am working with eel (very high plasticity in growth) and having looked at my data from all directions, I can say that for different habitats actually different functions are supported. And not by a small margin!

I can think of 2 approaches:

a) Stick with the model first selected (ignoring this, I think quite interesting information).

b) don't go for the global dataset but rather look for a model that is, at least essentially, supported on all of the lower levels (again there is a catch: I have 7 habitats, that can be done. But it gets messy when getting to the lowest level with ~5 size calsses that would mean finding a model that is supported by 5 x 7 = 35 subsets of the data. Which leads to the next question:

2. If working with nlme, is it appropriate to select the model based on another, computaionally less intensive method?

I came across a paper doing essentially the same and (though only using the global dataset) they used optimum least square (OLS) to identify the most parsimonious model. Though tempting for the considerably smaller effort, I do not believe this is very reasonable since it ignores autocorrellation (repeated measures) and random effects for the individuals in the model selection (and at least for my data results from ols and nlme would not be the same).

Maybe there is a way inbetween by performing model selection based on generalized non-linear least squares since this is essentially nlme without the random effects (results confirm that estimates are much more similar to nlme). To get correct estimates of variance and more sound statistics, random effects could be picked up after the model selection (i.e. select model based on gnls, perform all analysis with this model in nlme). The problem with the levels remains, but at least its much less effort since gnls is a lot faster to fit...