Modeling detectability prior to modeling occupancy

[jbauder]

I am analyzing a 3-year occupancy monitoring data set for the eastern indigo snake. I have 40 sites surveyed four times for three years for 480 surveys total. I am using a single-species multi-year model to test various hypotheses about how different factors influence both detection rate and occupancy. I am particularly interested in understanding how different factors affect detection rate so that we can use that information to increase our detection rates in the future. I have six variables with which to model detection rate and ten variables with which to model occupancy.
I had considered using a step-wise process where I hold occupancy constant and fit my a priori models for detection rate, then select the AIC best model for detection and use that term to model p while I then fit my a priori models for occupancy. However, I am concerned that I will either have high model uncertainty for p (and therefore many "AIC best" models) or by holding occupancy constant I am missing some interactions between my sampling and site covariates. This led me to consider fitting all of my models for p to each model for occupancy. But with six potential models for p and ten for occupancy that leads to a large number of models.
I was wondering if there are any general guidelines you would recommend following for when modeling both detection rate and occupancy simultaneously and if the "step-wise" approaches seen in the literature are really the best way to go?

Thank you,
Javan

[jhines]

I'd suggest starting with a model with initial occupancy constant, colonization and extinction year-specific, then find the best model(s) for detection.

If you're interested in how occupancy changes over time with respect to covariates, then you'll want to use one of the alternate parameterizations, where colonization or extinction (one or the other) is not estimated, but computed from the other parameters. You could then find the best model for annual occupancy, using the best model for detection. If there are a few models for detection which were reasonably close to the 'best' one in AIC, I'd repeat the search for the best occupancy model with each of those detection models.

Jim

[Eurycea]

Javan,

This article might be of interest to you:

http://link.springer.com/article/10.100 ... 010-0598-5

Paul F. Doherty, Gary C. White, Kenneth P. Burnham. 2010. Comparison of model building and selection strategies. J Ornith

Nate