Barker model convergence

[mconner]

I have been using the Barker model for large (3000-17000 records) Chinook salmon data sets. Seasonal survial is 0.1-0.7, recapture rates are almost zero, but resight rates vary seasonally from 0.05-0.45. Typically there are 4 groups (sites) and 6-8 time periods, per cohort (each cohort run separately).

I want to use some additive models for some of the parameters [e.g., R(g + t)], but I have found that the when I use the logit link I get very different AICc values than when I use the sin link. For exapmle, for the model:

S(g*t) p(t) r(t) R(g*t) R'(t) F(t) F'(.), the deltaAICc between the sin link and logit link is 265 (yes, two hundred and sixty five).

As expected, I get more estimates with the sin link - lots of logit estimates are 1.0 with SE=0.

When code this model using an identity link, and feed it the betas (transformed to logit) from the sin link, the AICc values jive. When I use the alternative optimization method, the sin model is still 4 deltaAICc units less (and some estimates are quite different).

So, while I can use the sin link for multiplicative models, I am stuck with what to do for additive models. I can provide initial values for some of the betas, but not all. I guess I could average across groups and use those as initial estimates, but really - isn't there a better way? (I've tried estimates from simpler models, but not much difference).

My big question is, why are the AICc values so different?

And, would you use additive models given the poor performance of the logit link in this situation?

Thanks for any advice on Barker model performance. Cheers, Mary