Two-species model errors

[maddiez]

Hi,
We are running Single-season Two-species models for bats in Presence with 22 sites and 4 nights. We have 5 species so to cross all of them with each other once results in 10 projects. Without putting in any site-level or sample-level parameters, we ran the models with phi-delta parameterization. For 8/10 models we get realistic values. For the other 2/10 we get no values back, only errors like "IND" and "QNB". We have run these over and over by inputting in different formats. There is nothing abnormal about the data that would make these 2 pairs work incorrectly while the others work correctly.

Could someone point me in the right direction or suggest a fix for these 2 models? Will putting in landscape and weather parameters help? Thank you, much appreciated!

-Maddie

[jhines]

Hi Maddie,

Please send me the most recent backup zipfiles from the project folder(s) which are not working and I'll try to find the problem.

Jim (jhines@usgs.gov)

[jhines]

Hi Maddie,

The problem you're having with those data-sets has to do with the parameterization. The phi/delta paramaterization has the parameters, psiA, psiB, and phi. PsiA and psiB are constrained to be between zero and one, while phi is constrained to be >0. In the likelihood function, the parameter, psiAB must be computed and the formula for it is: psiAB = psiA*psiB*phi. In the optimization process, some combinations of values for psiA, psiB and phi result in an invalid value of psiAB (psiAB not between zero and one). This can cause the optimization to fail and is what happened for those species. This is more likely to happen with sparse data, which is also the case here (22 sites).

The solution is to use the other parameterization (psiA, psiBA, psiBa), where the parameters are all constrained to be between zero and one. Presence will compute the species interaction factor, phi, as a derived parameter. The only drawback to this parameterization is that you can't model phi as a function
of covariates. I don't think that is an issue here as you barely have enough data to estimate the parameters as constants anyway.

Jim

[maddiez]

Hi Jim,

Thank you so much for your quick reply! That seemed to do the trick.

Best,
Maddie