covariate selection prior to model building

[m.e.]

Hi all--
I conducted 3 repeat visits during one season of a waterbird, and collected several habitat covariates. I’m interested in doing an RSPF while incorporating imperfect detection, but am unsure as to how to proceed in the initial covariate selection stage. If I do this prior to model building in PRESENCE (i.e. using a different statistical package such as STATA), my concern is that the importance of the variables may change when detection is incorporated (as I’ve seen in MacKenzie’s Pronghorn Antelope example). Can anyone shed some light on this? thanks

[dhewitt]

I don't know what RSPF means, and regardless of that I don't understand why you need to do any pre-modeling selection of covariates. Presuming you selected the covariates to measure for a good reason and you have ideas about how they relate to occupancy and detectability, build the occupancy models to assess the relationships and then put them all in a set for model selection. With only several covariates, this can't be too unwieldy. If a covariate has no effect, model selection should tell you.

If your pre-screening of covariates involves models like logistic regression then the pre-screening is not a good idea, for the reason you state. And if you take the typical approach above, it's not necessary. If the pre-screening is some sort of ordination to reduce the number of redundant covariates, that may be fine.

[darryl]

RSPF = Resource Selection Probability Function

[m.e.]

Thank you for the comments. So variable assumptions (i.e. linearity in the logit) are not applicable here? The reason I ask is because in some (but not all) of my models, I’m either getting the dreaded variance-covariance warning, or extremely large SE estimates, depending on the combination of covariates for both detection and occupancy (with no absolute pattern). I think the problem may lie with a detection categorical covariate, which I originally had coded as 0 (non-breeding sites, n=33) and 1 (breeding sites, n=10). I’ve changed the reference category (although I don’t know if this matters with only two categories) and also have used 1’s and 2’s instead of 0’s and 1’s. The problem now is, sometimes the model works with a different coding and sometimes it still gives me the var-covar warning.

I only have 43 sites, with up to 3 surveys per site, so I’m wondering if this is perhaps the issue…? Also, all of my breeding sites are “present” sites (no absences) so might this contribute to the issue?

[dhewitt]

Quick answer for your last questions:

If all sites have a 1 in the history for breeding, the occupancy is 1 and is thus a boundary estimate. That would throw the var-cov warning I believe.

With only 43 sites and not even 3 surveys (usually the bare minimum) at all sites, you are data-limited to a high degree.

Not sure about the rest of the estimation stuff in Presence.