Multistate models, radiotelemetry and MLogit

[Andrea Bowling]

I am using a multistate model to analyze monthly movements of radio-tagged birds among 3 geographic states (1, 2, and 3).
I am only interested in movement therefore I put "." (i.e., p = 0) after the last time a bird was encountered and fixed survival to 1 (but still estimate p). So the Encounter Histories look like:
/*1*/ 00000001101020202......... 1;
/*2*/ 00023001.................. 1;

Questions:
1-Is this a valid approach to estimate transition probabilities among states? My concern is that for each individual the "true detection" for the last occasion is now always 1. I do not know if that is a problem or not.

2-If this is a problem, should I replace the last occasion by a "." ? This way the "true detection" for the last occasion for each bird is now not necessarily 1.

3-Are there any problems or special considerations for using MLogit with individual and "real" covariates (for example: distance between sites or rainfall) (my guess is that there is no problem with "real" covariates, but I am not sure about individual covariates; I was unable to find anything about this in the "Gentle Intro")

I appreciate any advice and help. Thank you in advance!

Andrea

[Bill Kendall]

Andrea,

I haven't used the dot notation yet, but my understanding is that if all dots are at the end of a history it is equivalent to censoring the animal after its last capture. This is equivalent to the the usual history but with '-1' as the capture frequency. Assuming all this:

1) Your approach is legitimate, even estimating p. Essentially you are capturing the bird in that last time period but not rereleasing it. If every animal were censored on the same occasion then a problem might arise.

3) Covariates with transition probabilities can be tricky. Since they have to sum to 1.0, and when you use the MLOGIT link, by predicting one with a covariate you are affecting the others. Gary had some recent practical experience with this problem. I believe I remember what he did, which was to put the same covariate on each alternative, but maybe he should pipe in. However, here is an approach you might try, especially if you feel the covariate only affects one transition type. Run the model without covariates and with the MLOGIT transformation. Then, using those results as initial values, rerun the model with the covariate but without the MLOGIT transformation. Hopefully the psi's will remain well behaved and not add to >1.0.

Good luck!

[Andrea Bowling]

Thank you Bill. I will start working on that right now!

Andrea