Spacial Correlation Models

[Ana Carolina]

Hi people,

I have studied road impact in large and medium-size mammals. I have a survey dataset that involves 20 transects perpendicular to the road and in each one I have 12 sampling points. I surveyed the sites 5 times during 4 days each sampling station. I research about how the mammals are distributed in adjacent areas to road. I want know if the road is a determinant factor for distribution them. I am using the occupancy analysis for this! I suspect that my sampling points have dependence spacial because sometimes they localize in trails and they are relatively next. I tried run the spacial correlation models (Hines et al 2010), but I am not right if I have good results and models. My better models show that road is a important factor to occupancy, but the confidence interval is enormous (Occupancy: 0.9982; Standard error: 0.0174; Confidence interval: 0.0000 – 1.0000). I read a lot the paper Hines et al 2010, but I didn´t find an explanation for these results. I suspect that is because of my data are very sparse (I have many zeros). Is it possible? Could you tell other explanation for these results? I would like if you give me some tips!
Thanks a lot!
Ana.

[jhines]

Hi Ana,

The standard error of the 'real' estimates is computed using the 'profile likelihood', which means that the lower confidence limit is computed by subtracting 1.96*X from the untransformed parameter (where 'X' is the standard error of the untransformed parameter), and transforming it using the inverse-logit function (exp(b)/(1+exp(b))). The upper confidence limit is computed by adding 1.96*X to the untransformed parameter and transforming it. So, I suspect that the untransformed parameter associated with occupancy is a large number with a large standard error, giving an estimate near 1.0 and a large confidence interval.

Usually, when an estimated parameter is near zero or one, the standard error approaches zero. However, the logit-link function will sometimes cause numerical problems due to trying to take the exponential of large numbers, giving a larger than expected variance and standard error. You might try re-running the model with better starting values in hopes of getting a smaller standard error for the untransformed parameter.

If that doesn't give satisfactory results, I'd be happy to look over your data (just email me the most recent zipfile in your project folder).

Jim

[Ana Carolina]

Jim,

Thank you very much for the tips!
I sent you my project folders.

Regards,
Ana

[darryl]

Hi Ana,
What does each row in your data file represent? Is it a transect or a sampling point? Note that the 'spatial correlation' model is to account for a lack of independence in the repeat surveys (ie the detection process) rather than spatial correlation in occupancy between points.

Note that the wide confidence interval is a result of the logit-link function, and while mathematically correct, it doesn't really make sense because it's disagreeing with the small standard error, which is indicating the probability is actually estimated quite precisely. I'd ignore the confidence interval in this case. Like Jim said, this is caused by the probability being so close to 1 (and can also occur close to 0).

Cheers
Darryl

[Ana Carolina]

Darryl,
Thank you for your explanations!
But I have some doubts.

First, each row in my data represents a sampling point, ok?

I didn't understand as the "spatial correlation model" is to account for a lack of independence in the repeat surveys rather than spatial correlation in occupancy between points. I thought that the lack of independence in surveys was resolved when I used the sinlge-season models.

Can you explain better for me?

Then, do you think that my results are reliable despite wide confidence interval?
Do you think that I can use these results? Or no?

Thanks a lot.
Your tips are being very useful.

Cheers,
Ana.