Fixing parameters: Concerns

[arjun]

Hi

I noticed that Program PRESENCE allows for fixing parameters. I want to know the caveats of using this option. Is it correct (from the analytical perspective) to fix any parameter to an expected value? Specifically, I have the following problem:

I'm trying to run multi-state occupancy models for large carnivores. My detection matrix has 175 sites and 12 temporal replicates. The two "states" correspond to breeding and non-breeding animals. I am using the default parameterization in PRESENCE (PSI, R, p1, p2, delta). I have 5 covariates each for PSI and R, and 2 covariates for p1,p2 and delta. The issue is, even if I try including a single covariate for any of these, the model fails to converge. So here is what I did: I used the standard single season single species model and estimated PSI and p. Based on this, I was able to get an estimate of PSI (after modelling with 5 covariate combinations). The PSI here is essentially the same PSI (in terms of interpretation) in the multi-state model. I then fixed the value of PSI in the multi-state model with the estimated PSI from the previous step. And this worked fine. I was able to model R with 5 covariate combinations and the p1, p2 and delta with 2 covariate combinations.

1) Is there anything wrong with this approach?
2) If this method is incorrect, then is there an alternative method that I can try?
3) If this approach is fine: since I'm fixing PSI, is ok to still estimate p1 in the multi-state model ?

Thanks in advance!

Cheers
Arjun

[jhines]

The 'Fix parameters' option in PRESENCE is usually used to fix parameters to zero or one in cases where either you know the value of the parameter (eg., no survey done, so p=0), or there is insufficient data to estimate a parameter. The main problem with fixing a parameter (Psi) to a specific value is that you would be pretending that occupancy is a known constant, instead of an estimated parameter with a variance. This would artificially reduce the variances of the other parameter estimates.

Did you try entering initial values for the parameters, instead of letting PRESENCE use the default starting values? I'd suggest using the final beta estimates from your single-state model for your multi-state model.