Modeling survival in a telemetry study with "missing" birds

[evan.adams]

All,

Here's some background of the study (and myself), I’m working on a project that is estimating survival on a bird species using telemetry (not my first choice, but that’s what we have to deal with). I have a decent amount of experience with MARK (various typical survival analyses and I’ve taken Jim Nichols’ modeling course at U of Florida) but this analysis is veering into some uncharted waters for me. The transmitters we are using only have about 21 d of battery life (small birds) so these are very short term survival estimates. Normally I would just use a known-fate analysis and be done with this, but we had a problem with a significant number of transmitter signals going “missing.” We also we having a hard time recovering transmitters in certain habitats as the transmitter signal doesn't project well once under water or mud. Using a known fate model this would lead to us having to censor a large portion of our data (25%) and lead to spurious results. After doing some research I discovered that you could model such kinds of telemetry data by splitting the data set into two component parts: mark-recapture and a ring recovery (ala Burnham’s paper in 1993 and Chapter 10 via Ch. 16 in the MARK book). This specific analysis is new to me and I'm hoping that I made the right decision on getting away from known-fate measurements - can I get any feedback on that decision?

So assuming that this is the right model type for this study, I've already gotten the data into MARK and I have some questions regarding this analysis in particular. For example, the model always estimates F as 1 and while I’m sure that’s due to the fact that we really aren’t getting ring recoveries outside of our sample area, I was worried that something was going awry in the modeling. Also, because transmitter battery life would have such a big impact on detectability and thus survival, I wanted to model each individual over the exact same time scale (28 d, to represent the maximum lifespan of a transmitter) to standardize for the artificial means by which we were estimating survival. The data aren't ragged, though that might not matter so much since this isn't known-fate. Initially I was thinking that I would model everything by calendar day, but that lead to some birds have a lot more zeroes after the transmitter failed than others based purely on capture date, which is unfair to those birds. So I just varied detectability by time and threw capture date in as a correlate in the model. Unfortunately we catch around an average of 3-4 birds per day (150 birds total), so I don’t really want to do a true cohort model (it would dramatically increase the number of parameters).

I'd appreciate any thoughts or suggestions on this analysis. Thanks for listening!

Evan

[Jochen]

Evan,
the menaing of the parameters in the model, and thus the answer to your question, depends on the data you have. Without detailed information there is at least a more general point: if a parameter is called "fidelity" this does not necessarily mean that it really measures fidelity, again depending on the data. Supposing that the parameters of interest are S, p, and r, there may even be reasons to fix F at 1.
I didn't really understand the issue with time. Why not use calendar days and an "age"-dependent p to model transmitter life?
Best
Jochen