unrealistic Psi-hat estimates

[sgagne]

Dear list members,

I am using RMark to model the occupancy of seven anuran species in 21 ponds. In particular, I am interested in the effects of surrounding urbanization age on anuran species occupancy and have included this variable as well as a few local habitat descriptors as covariates. However, in some instances, I seem to be getting weird estimates of Psi-hat (i.e. Psi-hat = 1 or 0) and very large parameter estimates for my covariates (with very small SE). For example, the null model for the Western chorus frog has a Psi-hat = 1 but this species only occurred in 4 of my 21 ponds. Also, the models for which Psi-hat = 1 or 0 have low AIC values (~45). Any ideas what might be going on?

P.S. I double-checked that RMark is reading my encounter histories correctly and it is.

Thanks,

Sara

[jlaake]

Sara-

There is little anyone can tell you with the information you have given. If you want, send me the data and R script you are using for the species you mentioned in which the null model estimated Psi=1.

--jeff

[jlaake]

Sara sent me her data. I believe the problem is very sparse data. The one example she mentioned had 21 sites and 6 occasions. Of the 126 possible values only 4 had a value of 1 - each of 4 sites was observed occupied on one occasion. I don't have a lot of experience with occupancy models but presumably it can't discern a reasonable value for p with this little data and estimates p=0 so Psi=1. Others with more experience want to comment?

--jeff

[jlaake]

As a follow-on. With that example the estimate of p=0.03. So the probability of a site being occupied and not being seen =(1-.03)^6 = .83. Thus it would only expect 1 in 6 occupied sites to be observed as occupied. It had 21 sites and 4 were observed as occupied. Thus the best prediction for Psi=1.

To show how little information there are in the data, I added a single 1 to one of the 4 capture histories with an observation and the model converges to Psi=0.41. My guess would be that occupancy modelling would follow the rule of thumb as for c-r models that the higher the p the better and p>.2 is good lower bound to target, although that will also depend on number of occasions.

--jeff

[darryl]

The methods are statistical, not magical (sorry if that sounds harsh). If you have very sparse data then you're going to be limited in what you'll be able to estimate. One option might be to share information across species by analyzing their data simultaneously if you expect them to have a similar p ro psi for example.

Cheers
Darryl