single season- covariate analysis help

[Deepak]

Dear all,
I have been running single season model for 40 different trails and included canopy cover % as one of the covariate. When I modeled site occupancy changes with canopy cover in the results I got a warning message. ********************** WARNING ***************************
Variance-covariance matrix has not been computed successfully.
Ignore matrix values in the below output.
--------------------------------------------------------------

In contrast when I normalized canopy% covariates this warning didn’t appear for any of the results. Is it mandatory that one should always normalize the covariates before saving for the analysis?

I also have 1, O (presence or absence) data for other covariates like disturbance, water body & grass marsh. Here the problem is slightly different I also get an other warning message along with it:

Numerical convergence may not have been reached.
Parameter estimates converged to approximately
2.69 significant digits.

Moreover in some case there is no warning message however the estimated occupancy is 1.000, which eventually becomes the best in selected models.

What’s causing the problem?
Can anyone clarify it for me?

Thanks in advance,
Deepak

[darryl]

Hi Deepak,
I hesitate to do this, but RTFM; well actually, read the FAQ for this forum and there's info there on those 2 warnings. Standardizing covariates isn't always necessary, but it can often help particularly when you're covariate values are outside of +/- 10.

I'd be nervous of a model that gives an estimate of psi of 1.0 if you've got sparse data (relatively few 1's compared to 0's), or a pretty complicated model structure. If the estimate from this model is quite different from that given by other models, you could discard the model if you don't think that estimate is biologically reasonable provided you think it's just that that particular model isn't' working very well with your data set. Although you really have to think hard about the reason for discarding it as you don't want to be seen as doing so just because you don't like the estimate.
Cheers
Darryl

[Deepak]

Dear Darryl,
Thanks for the info. As you pointed out its the problem with more zero's but less 1's, that's how my study animals occupancy is! low. In most of the models the detection probability is 0.2 - 0.29 not more than that and the models are not very different from the others except the psi value becomes high.

Thanks,
Deepak