I have been trying to estimate the density of a population of Fishing Cats in a coastal area. I am running the null model with the hazard rate detection function (which had more support than the exponential and halfnormal null models), and those accounting for behavioral effects on g0. The code used for model fitting is as follows:-

- Code: Select all
`fcsouth1<-read.capthist(captfile=captfile1, trapfile=trapfile1, detector="count")`

fcsouth_HR_default<-secr.fit(fcsouth1, detectfn="HR", mask=mask2_south, trace=FALSE)

fcsouth_HR_g0_b<-secr.fit(fcsouth1, detectfn="HR", mask=mask2_south, model=g0~b, trace=FALSE)

fcsouth_HR_g0_bk<-secr.fit(fcsouth1, detectfn="HR", mask=mask2_south, model=g0~bk, trace=FALSE)

The AIC results for the 3 models are as follows:

model detectfn npar logLik AIC AICc dAICc AICcwt

fcsouth_HR_default D~1 g0~1 sigma~1 z~1 hazard rate 4 -113.2849 234.570 236.170 0.000 1

fcsouth_HR_g0_bk D~1 g0~bk sigma~1 z~1 hazard rate 5 -364.2528 738.506 741.006 504.836 0

fcsouth_HR_g0_b D~1 g0~b sigma~1 z~1 hazard rate 5 -366.6017 743.203 745.703 509.533 0

Warning message:

In AIC.secr(fcsouth_HR_default, fcsouth_HR_g0_b, fcsouth_HR_g0_bk) :

models not compatible for AIC

The deltaAIC values between the null models and those incorporating behavioral effects are very high. As my dataset includes few spatial and temporal recaptures of fishing cat individuals, I was expecting relatively strong support for the models incorporating behavioral effects (more because of inter-trap distance being fairly large in comparison to likely fishing cat home range sizes).

I know that AIC comparisons should not be used between models incorporating a group structure and those which do not. However, I did not try to incorporate sex or any other kind of group structure in my data (principally because only 6 out of 30 fishing cats in the dataset could be unambiguously sexed and only 1 of these was a male, so I felt that even hcov models might struggle to converge/generate poor precision in estimates). It appears that the null models are not comparable to those incorporating behavioral effects on detection parameters. Although these two types of models form two categories within which comparisons are possible.

I did a little more digging using the AICcompatible() function and obtained the following:

- Code: Select all
`list1<-secrlist(fcsouth_HR_default, fcsouth_HR_g0_b, fcsouth_HR_g0_bk)`

AICcompatible(list1)

data CL groups hcov binomN

FALSE TRUE TRUE TRUE TRUE

It appears that the two classes of models are operating on different sets of data, although I have no idea why that would be the case.

I am not particularly experienced with respect to mark-recapture analyses and this is my first post in this forum, so I apologise for any inadvertent errors in my post.

Thanks in advance,

Soumya Banerjee