Trap availability approach

[Kirby Mills]

Hi all,

Along similar lines to the question asked earlier today, I am trying to figure out the best way to approach model setup for years of data where we have no information on trap availability. This is specifically for very trap happy foxes that are likely to have multiple traps in their home range that they can access. 

Over our ~30 year dataset, we have some years with specific trap-night availability (great), some years with only the proportion of trap-nights available across all traps in a grid (not great), and some years with no information on availability at all (oh no). We are using yearly sessions, with grids (+ buffers) within a year aggregated into one state-space. 

We have been debating the best way to deal with this inconsistency, with a couple ideas:

1. For years with only grid-level availability, randomly assign 'unavailable' trap-nights across non-capture trap-nights within each grid, based on the proportion of available trap-nights.

2. Ignore availability altogether, leaving all non-capture nights as 'available' (either for all years or only years with no information) 

But generally we are having trouble wrapping our heads around how these approaches will impact the detection probability estimates and overall density estimates. Will it change the predicted activity center locations or population size estimates? Would changing the availability approach between years have unexpected effects? I'd love to get your thoughts on possible approaches or any considerations we should think about. 

Thanks in advance for any insights you can provide!

Kirby Mills

[Daniel Linden]

Hi Kirby, this is quite a challenging problem.  I think a general strategy would be to turn this challenge into a sensitivity analysis and essentially conduct the density estimation using several approaches.  At the end, the findings that are consistent across approaches are the ones most useful for inference (i.e., patterns in density, etc.).

Unfortunately, without a simulation it would be difficult to assess how exactly this lack of information would affect your density estimates.  The spatial patterns in availability will influence the effective detection probabilities and the realized density estimates.  Whether the induced noise is enough to degrade any important signals is tough to tell (though the sensitivity analysis could shed light).

Are these single-catch traps?  That is a general problem in SCR modeling with few straightforward solutions, though it has not stopped folks from simply listing the caveats and fitting models anyway.