working with a boundary issue

[rachelselwyn]

Hi,
I have a question about how to work with models where a beta parameter is encountering boundary issues.

I am using several covariates, including 'habitat type' which is made up of 6 categorical covariates. One of the habitat types (LG) has had no sightings of this species over the whole study period. Because of this, when I run a simple psi(habitat)p(.) model I get a Beta estimate of -29.9 for LG (SE=670,000). I am also getting a non-convergence warning with convergence happening to 3.38 sig dig.

The big problem this is creating is that when I run more complex models; psi(habitat+time)p(.) and psi(habitat+time+fruit), I am now getting VC warnings and problems.
psi(habitat), psi(habitat+time) and psi(habitat+time+fruit) are my top three models (between 0 and 1.34 delta AIC), next best model has a delta AIC of 8.43.

One of the recommendations for this was to try using a different habitat type as the 'standard' category. This does help but as I am also running occupancy models for other species with this data set with the same set of covariates, I am wanting to keep things consistent between them. This new 'standard' habitat type does not work for all unfortunately...

Another thread discussed fixing the beta parameter, but it seemed unclear if this was recommended as a solution or not. viewtopic.php?f=11&t=1521&p=4309&hilit=fix+parameter&sid=7e8a9f7b8e8987ca939cf7721bb5c101#p4309

My questions are; A) should I fix the beta parameter for LG habitat to be -30? Would this potentially help with the more complex models?
B) How do you actually fix a beta parameter? I can't find this information anywhere. I am trying to work in RPresence but I would settle for PRESENCE if that is the only way.

C)Do I need to just throw out the more complex models entirely?

Thanks,
-Rachel

[jhines]

Hi Rachel,

The beta estimate of 29.9 seems reasonable since there were no detections for that habitat type. The large standard error is normal. The standard error for the real parameter psi for that group should be zero. How do the other estimates and standard errors look for that model?

Do you have p(.) for the more complex models as well? If so, then perhaps the problem is with sparse data and not due to the LG group. It's hard to say without seeing the data.

Instead of changing the intercept group, you could try the formula in RPresence which gives each group their own intercept: psi(-1+habitat). The "-1" tells R to omit the common intercept in the design matrix and make each group have it's own intercept.

There is no way to fix a beta parameter in Presence/Rpresence.

I need to see the data before I could advise on whether the more complex models are possible. Feel free to send it to me if you'd like some suggestions.

Jim (jhines@usgs.gov)