bias in lambda due to permanent trap response

[Eric Janney]

I have a problem concerning bias in lambda estimates using Pradel model due to permanent trap response. I thought I would post my situation to see if anybody had any suggestions or ideas. I am interested in estimating population rate of change and recruitment for two endangered fishes in a large lake in Oregon. We began capturing and tagging fish annually using PIT tags in 1995. In the past we relied on capturing fish in nets as our primary marking and recapture method; however despite tagging large numbers of fish, recapture probabilities were extremely low (.03 - .10). In 2005 and 2006 we incorporated remote underwater PIT sensing equipment at strategic locations to improve p's. The new remote PIT tag detections systems greatly improved our p's (around .4 - .; however, they have also created a severe permanent trap response. This arises because unmarked fish have a very low probability of being captured, marked, and released; however, once they are marked they have a very high probability of being detected in the remote gear the following year. In a sense they become extremely trap happy. This new technology has greatly enhanced our ability to estimate and model survival, but I'm not sure I can incorporate it into Pradel models without introducing severe bias. Any suggestions?

[gwhite]

Eric:
The Pradel model makes the critical assumption that initial captures and recaptures are identical, because there is no other way to estimate initial capture rates. So, for your data, you have violated this assumption. As you guessed, your PIT tag detectors help you estimate survival, but cannot help you estimate recruitment, and hence lambda.
One idea to try is to use a time-varying individual covariate to model whether a recapture came from a PIT tag detection (value = 1), or a true recapture (value = 0) that would be the same whether a fish was previously captured or not. By using the time-varying individual covariate, you get to include the additional captures to assist with estimating phi, but not for initial captures (since they would never have a value = 1). I've not simulated such an approach, but minimal thinking about the idea seems to me that it would work.

Gary