Simplifying covariates when sample size is small

[jmd]

Hi all-
I'm working on a project examining post-fire recovery of small mammals. We have presence-absence data for 25 3-day trap sessions (over quite a few years), but only on 16 sites, and with quite a few missed trap sessions for some sites. We are running multi-season occupancy models and finding that even when limiting the number of covariates included, models often have difficulty estimating SE.

We plan to try running some models in which we drop or combine sites that have a lot of missing data, but we'd also like to try simplifying some of our covariates, if possible. For example, one of our covariates is a principal component value (continuous) for vegetation on each site that varies with trap session. Would we be better off using a categorical covariate describing vegetation at the beginning of the study at each site and then comparing models including this covariate to time-only models, or models with a vegetation x time interaction? Is this actually a simpler approach more likely to produce models that happily estimate SE?

Any thoughts appreciated!

[darryl]

Obviously your first step should be to try very simple models without covariates. If you're having problems estimating SE's from those models then adding covariates are unlikely to improve the situation. Given you have 16 sites, it maybe that some of your estimates are tending to be very close to 0 or 1 in some years which may be causing you an problem. Note that you can also have covariates whose value changes through time, but the effect of that covariate is actually consistent so it would actually be an interaction.
Cheers
Darryl