unstable mask.check value in Exponential detection function

[ndu...@gmail.com]

Hi Murray & everyone

I'm trying to fit model with exponential detection function, the models were fitted with varies buffer from 10000-25000 m (for half normal 15000 was good enough). "mask.check" showed unstable log likelihoods and spacing, even though I used 25000 m. Should I use "suggest.buffer"? or It was not enough buffer width?

                   model                 detectfn npar    logLik           AIC          AICc       dAICc     AICcwt

fit.Expo        g0~1 sigma~1     exponential      2 -164.5110   333.022    333.451   0.000     0.9091

fit.null          g0~1 sigma~1     halfnormal        2 -166.8136   337.627    338.056   4.605     0.0909

I expected exponential detection in disturbed forest for bears, and half normal for undisturbed forest.

Thanks,

Dusit

[Murray Efford]

Hi Dusit!

We don't have much to go on... In general the exponential has a longer tail, so you might need a wider buffer. But 25 km is a long way, even for a bear... Are you sampling over such a long period that the sample includes home-range shifts or transients?

The real question is whether it matters (i.e. affects the density estimate): suggest.buffer and mask.check only give clues. Have you looked at esa.plot? That gives a quick view of how the estimated density changes as you increase buffer width.

I doubt anyone can predict the effect of habitat on shape of detection function, but I may be wrong.

That's my 2 cents worth.
Murray

[ndu...@gmail.com]

Hi Murray

I don't think long period cause home-range-shifts, because only 6 months as same as the other sites. But, real problem might be extreme illegal rosewood cutting drive bears moving (I guess). Hundreds of people inside the forest during surveys. I did esa.plot too and it showed that 15000 m was good enough. 

I used baited that why I think detection function may change with combination of habitat.  

Thanks,

Dusit

[Murray Efford]

Hmm. The idea that people stirred the bears up and changed their behaviour does seem very plausible. I can see that the detection function could easily be _different_ in the two habitats (at least different in scale i.e. sigma, and perhaps g0/lambda0). I just wasn't sure why you would expect one shape (halfnormal) in one habitat and another particular shape (exponential) in the other. But maybe that's not what you meant.
Murray

[David Borchers]

Hi Dusit

LIke Murray, I suspect that the shape of the detection function is not really predictable in advance. And depending on your detector density, you may not have much ability to distinguish between shapes on the basis of the capture data (higher detector density gives greater ability). You do however, seem to have good reason to suspect that animal ranges are different in the two habitats, and it these circumstances my starting point would be to use a detection function in which both g0/lambda0 and sigma depend on the mask covariate "habitat type" (or whatever you call it).

You can't use suggest.buffer() etc. with this kind of model, but you can quite easily look into the effect of buffer size by just looping through a series of fits in R, using a model like that above, and increasing buffer size on each fit. If you save the detection function parameter estimates, density and esa estimates inside the loop, you can then construct plots after the loop that show how these estimates change with buffer size and at what buffer size they (hopefully) stop changing much. (If you are not sure how to program this, I can help - or Murray may already have code to hand that does it.)

One thing to bear in mind: I suspect that you might end up needing quite a high-resolution mask (i.e. lots of mask points) because a suitable buffer for the animals with big ranges might give a very big mask for animals with small ranges and if you don't have enough grid points in the mask, it is concievable that you end up with so few grid points per range of the small-range animals that you get too crude an estimate of the small-range animals' esa. (The space that animals actually occupy is continuous - the mask is a discrete approximation to this space and if it is too discrete it can be a poor approximation.)

David

[Murray Efford]

Hi again
David has a point about the mask possibly getting too coarse as you go to extreme buffer widths (if you don't increase nx or hold spacing constant in make.mask). That at least is an explanation for instability in the mask.check likelihoods.

I suspect that Dusit's habitat patches were discrete study areas (sessions), so it would be simple to handle the habitat variation as session-dependent. 'secr' doesn't normally allow detection parameters (g0/lambda0/sigma) to depend on mask covariates, although there is a trick in the next release that effectively allows sigma to depend on mask covariates. Watch this space.
Murray

[ndu...@gmail.com]

Hi David and Murray

Now, I understand the problem of detection function shape.

Good point on number of mask points. I will check on this. Last time, I hold spacing constant at 900 m based on mask.check. 

I will check your new release. This might help me explain why sigma was significant different among sites by model it as mask covariates.

Thanks,

Dusit