Multi-state model Var/Covar errors

[cheryl]

Dear all,

I am running single-season multi-state models with my dataset. I am getting the variance/covariance error on ALL models. This still occurs after I have fixed the unestimatable parameters to 0’s and 1’s. Now all of my untransformed estimates are between -3 and 3, but with huge se’s and still the cover/var error msg.

FYI. I have 169 sites and 9 surveys. The 1st survey is of suitable habitat, the 2nd-9th surveys are live-trapping. My states are:
0- no sign (i.e. hab. not suitable)
1- potential sign or positive sign present (detected for species A or B -can’t tell difference).
2- Occupied by Species B (detected by live-trapping)
We use habitat sign surveys as a first pass. We trap all plots where potential sign found and some random plots where no potential sign found. There were no animals captured where no potential sign was found (p1~1).

I have tried the analysis both in the habitat suitability framework (with all p’s fixed at 1 and interpretation of delta) and with estimates of p1 and p2. For these analyses, I set p21 to 0- since there is no probability of trapping(detecting) an animal during a habitat survey.

Also, is it possible to get transformed parameter(site) estimates for delta?

Any help is appreciated. I’d be happy to send the Presence files.

Thank you,
Cheryl

[darryl]

Hi Cheryl,
Even though you've fixed the parameters, have you checked that you've deleted all the corresponding beta parameters out? What about models with no covariates? Can you ever observe a '1' with the live trapping, or for surveys 2-9 can you only observe either a 0 or a 2?

Cheers
Darryl

[darryl]

In the insane hope that someone might search the forum before posting a similar question, here's the solution to Cheryls problem that we sorted out off-list. Note that these issues apply to all models, not just the multi-state.

1. When you fix a real parameter in PRESENCE, in the corresponding row of the design matrix make sure it's all 0's (strictly speaking you don't have to depending on the types of constraints you're using for the particular model, but this is the safest approach). This means that if you fix all parameters of a certain type (as in Cheryls case) you actually have to delete all columns from the design matrix. Whether you have to adjust the parameter count after fixing a parameter (to get AIC's correct etc) depends on whether you're fixing the value for biological reasons, or analytic ones because estimated values are on the boundary of allowable values so you have to fix parameters to get standard errors out.

2. If you have a categorical covariate with m categories, if you've included an intercept term (as is often recommended) you should only include m-1 dummy (or indicator) variables in your design matrix. Otherwise the beta parameters are not uniquely identifiable (the model is over parameterised)

Hope this helps others
Cheers
Darryl