Open klg-2016 opened 1 year ago
Nice progress, Katie!!
Is there a particular aspect of the results that you're questioning, or just the NaN parameter estimates? NaN parameter estimates certainly indicate an issue with the model, and a warning that should be heeded. My guess is that you may be over-parameterizing the model, given the small camera sample size (I think that may be the bigger issue than the small number of detections, although it's probably some combination of the two).
What happens if you drop detect.obscured entirely? It's not my favorite covariate, as I didn't quantify it in a very systematic way...
As for which models to try, I'm curious for a bit of justification about how you have gone about enumerating candidate models thus far. Maybe it's a bit of "copy what I did in Rota and then add new stuff" approach? But, for example, I would think that including activity pattern as a detection intensity predictor would be wise from the get-go, given that you know that these species are strongly nocturnal.
Interesting about the f1/f2/f12 covariates and fit, as I think that the spatial component of the model is effectively the same as the Rota. Hmm.
This gets into a more philosophical question about the purpose of model selection—I tend to prefer a hypothesis-driven approach (like you did in your thesis) to a kitchen-sink approach. We had included some covariates (ground cover and detection obscured) in all of our models as a default, but maybe it makes sense to revisit this, given that we are adding complexity to the model. Perhaps activity pattern and ground cover become the new default detection intensity predictors?
Before getting too into the weeds of model selection—have you figured out how to actually look at the results (direction/magnitude of coefficient effects) in addition to the AIC? I think that may be more informative than looking at AIC—in other words, are these models actually doing what you expect them to do, and do they concur with what you found in the Rota model? That, to me, is the next important question to answer.
Again, really amazing work in getting all of this up and running. I'm quite impressed.
Thanks for your response! I hope your travel is off to a good start :)
Responding to your points:
my question/request was both specific about the NaN and more general about the results so far. That's what I was thinking might be the issue leading to the NaNs as well. I'll work without detect.obscured next week and see what happens.
my justification was basically starting with the natural parameter covariates from my ox thesis, and then working with similar combinations of detection intensity covariates as Arielle did in her paper. I think I may need to take a step back and think through what the different covariate options available to us mean if they're part of the top model. I'm struggling a bit in my thinking given that we don't have a strong assumption for dominant/subordinate but it's effectively baked in to the model. I thought the same thing about the f1, f2, and f12 parameters.
I will work with that idea--activity pattern and ground cover as default
I have not figured out how to look at the results yet, I'll see what I can work on there as I come back to the models next week as well.
Thank you :) Have a great trip!
If you have time, can you check out the results from the models I ran with a scaled detect.obscured (v2 on the spreadsheet I shared)? The final two threw a warning for some of the parameter estimates (they're NaN). I'm not sure if that's just because I'm asking too much of the model with the number of detections we have overall.
I am waiting to hear back from Arielle about excluding f12 and latent sp 2 (coyote in her case) occupancy; I'd appreciate any thoughts you have on initial results/any other combos to try. If Arielle responds with how to exclude f12, I'll run those models as well, and I'll run all the ones I have with the reverse dominant/subordinate (that's just going to take a bit to go through and change the spp in the code). I'm a little concerned that including the natural parameter covariates (f1, f2, and f12) doesn't really improve model fit relative to the null model, but those covariates were the best from the Rota model with my thesis. Thank you!