c-hat for spatial dependence model

[sunita]

My data set consists of 277 sites with 2-8 spatial replicates. When I run the single-species model with spatial dependence using covariates, I get a c-hat of 0. The number of bootstraps I used for model assessment was 10000. The Probability of test statistic >= observed is 0.9 for all the spatial dependence models where I have used co-variates. From what I understand, this indicates that none of the covariates used explains my data well which is surprising as all the covariates have been chosen based on the biology of the species.

When I run the models without the spatial dependence, I get a c-hat closer to 1 (for teh global model I get a c-hat = 1.13 and p = 0.17).

Can someone tell me under what circumstances c-hat becomes 0?

In the book - Program Mark - a gentle introduction (chpt5 pg6), it is recommended that if c-hat < 1, then set c-hat = 1. Does this recommendation hold good when c-hat=0? or are there any alternatives ?
Thanks in advance
Sunita

[darryl]

Firstly, I'm not sure whether Jim Hines has modified the code for the model fit test such that it's working properly for the 'spatial dependence' model. If your c-hat value is exactly 0 then I'd be suspicious as it's calculated as the observed test statistics / average of the bootstrapped test statistics. To get a value of 0, either your test statistic is 0 (which could only happen if your model is a perfect fit to the data) or if the average test statistic is many orders of magnitude greater than the observed value (so it's actually near zero).

Second, if your c-hat is <<1, then sometimes that's the result of over-fitting your data, you've got some unnecessary parameters in the model or are modelling some spurious patterns in the data. Try removing some of the covariates and repeat the procedure.

Cheers
Darryl

[jhines]

After seeing this post, I checked the code for the spatial-dependence gof test. I found an error which was introduced when I added the modification to allow survey-specific theta's. I've corrected this and the test seems to work.

If you're still getting zero for the test-statistic or c-hat, I'd like to see the data file.

Jim