transition to the unobservable state

[kthall]

Hello all-

I'm using a multi-strata model to get tree snail dispersal rates. My field site is a square grid of continuous quadrats, with monthly movement recorded between quads. I code the encounter/movement history at each interval as A (same tree), B (new tree), and then of course U (outside the grid and unobservable). I "reset" all B locations back to A with an additional short interval following each actual survey- but that is another topic altogether!

Example history: AA00A0BA0B0A

Question: I have a time-varying individual covariate (snail location) beta that will greatly influence the transition probability from A->U. This probability is equal to another transitional probability that IS observable for a certain value of this covariate. Example, at the corners of my site, A->U (movement to a tree outside the grid) is as equally probable as A->B (movement to a tree within the grid). Also, when this covariate equals 0 (indicating a location in the center of the grid), the A->U probability is near 0 as well.

I know the unobservable state causes lots of problems since you can never actually get a "U" into the encounter history, and was hoping that being able to code the additional information I have accordingly would help minimize that. Much thanks in advance for any insight on how to do this constraint in the design matrix!

-Kevin

[kthall]

to clarify, the example history above would be coded for encounter as:

AAAA0A0AAA0ABAAA0ABA0AAA, with p and psi B->A both =1 every other interval, just to do the "reset" and obtain a generic dispersal estimate independent of previous movements