Transition Probabilities in CJS Models

[Robertal]

I'm trying to figure out how to incorporate transition probabilities (from one stage class to another) in CJS models. I'm using the transition probabilities in subsequent Matrix models, yet need to match up this with my survival analyses. I have a size-based stage model, so some fish stay within a class for 2-3 years. Do I need to go to a multi-strata model? or can I incorporate these transition probabilities into the design matrix?

Any help would be great. Robert

[cooch]

I'm trying to figure out how to incorporate transition probabilities (from one stage class to another) in CJS models. I'm using the transition probabilities in subsequent Matrix models, yet need to match up this with my survival analyses. I have a size-based stage model, so some fish stay within a class for 2-3 years. Do I need to go to a multi-strata model? or can I incorporate these transition probabilities into the design matrix?

Any help would be great. Robert

Use a MS approach. However, I can tell you from a conceptual standpoint that for lefkovitch matrices, where the probability of a size-class transition may depend on how long you've been in the stage already, simple M-S models have some limitations. You can 'trick' memory models into handling some of the complications, and there are a few tricks you can try using invidivual covariates, or 'clever' constraints using cumulative logit link, but MS models in MARK remain simple first-order Markov, and a projection model that is higher order or conditional Markov aren't readily handled by MARK (or anything else currently available that does most of the work for you - although the M-SURGE crowd might have some new approaches that I'm not aware of - M-SURGE was written from the ground up to treat everything as a variant of a MS model).

I won't even get started on the general 'arbitrariness' of size classes. There are some very nifty new kernel based approaches that get around the general problem (see whole slew of papers by Steve Ellner), but they're not easily mapped to any estimation framework I know of yet.

So, MS approach for now - with caveats.