I have a set of CJS models that I am testing, a few of which have some inestimable parameters.
I recently acquired a new computer which makes it feasible to try the alternate optimization method (simulated annealing) and so I re-ran the entire set of models using this approach.
It worked wonders for parameter estimation, but the AIC order is different than the same set of models run under normal optimization (I did correct the number of parameters in the models that had inestimable parameters when run with normal optimization).
Is this normal, and if so, which method should I trust? My gut tells me that the simulated annealing approach should be better, but hopefully there's a more concrete answer than that.
Thanks very much.
Cheers,
Andy