Unobserved stated in MSCRD or MS models with > 2 states

[lbmcnew]

I realize this thread is a bit old, but I was wondering if someone could send me an example of an MSCRD or MSORD analysis with more than one unobservable state (Kristine?). I expect the .inp to be the ms standard ‘A00AB00B’ (for the case with 2 observable states) but how are multiple unobservable states specified? I’m thinking about a specific case where each observable state has a corresponding unobservable one?

Thanks, lb

[gstauffer]

lb,

I think the key is to remember that unobservable states are defined in MARK (or whatever software you are using), not in the encounter histories (i.e., a zero in an encounter history denotes non-detection, not a specific state). Unobservable states are any states that you define in MARK that do not appear in any of the encounter histories.
For example, suppose you have individuals observable in three different states, A, B, or C, and you have reason to expect that there are three corresponding unobservable states, which you might call X, Y, and Z (but these codes of course do not appear in the encounter histories). You might set up a project in MARK and specify states A, B, C, X, Y, and Z. Because X, Y, and Z never appear in the encounter histories, you would now have defined them as 3 unobservable states. MARK will create corresponding PIMS for all the transitions. If some transitions are not biologically possible, you would set them to zero. For example, Psi (XY) might not be possible if individuals must first transition through B before they can get from X to Y.
I do have an MSORD analysis with 3 unobservable states (coded for RMark) if you think that would be useful to you. But I think once you understand how to set up the states in MARK, you should be on the right track.

Glenn

[lbmcnew]

Glenn,

Thanks for the response. This is exactly how I defined the structure for a simulated data set last week, but was having convergence issues even under the simplest of models with constraints on unrealistic transitions. At this point, I'm unsure why, but I'll experiment a bit with different data simulations before bothering you for your example analysis.

Thanks again,
lb

[jlaake]

If the data are simulated, if you haven't already use the simulated parameter values as starting values. If it still doesn't converge then you may have a problem with the simulation or the model specification.

--jeff