[R-sig-Geo] Best approach to account for sp. auto in models with count data?

[Jenn Barrett]

Hi everyone,

I've been searching all over the Internet, and through the literature, and can't seem to find the answer to this question - hopefully someone here can help.

I have a dataset that consists of counts of birds (6 different species) within circular plots. The goal of our study is to examine the relationship between:

1) species richness and habitat features;
2) the abundance (zero-truncated) of each species and habitat features; and
3) the presence/absence of each species and habitat features.

The count data has a negative binomial distribution. Moran’s I correlograms of bird presence by plot indicated spatial autocorrelation in all species groups, and spatial autocorrelation (positive) was also present in the residuals of the nb.glm. We therefore wish to account for spatial autocorrelation in our models; however, I'm a little stuck on how to do this for raw counts. For 3) above, I'm using an autologistic model (i.e., I'm including a distance-weighted autocovariate in the regression equation); however, I've read that an auto-poisson (or, I'm assuming an auto-negbin) model can only account for negative spatial autocorrelation (not positive). Also, while I have used spatial error models in the past on continuous, normally distributed data, my impression from the literature is that they are not meant for count data - so this doesn't seem like a good option either.

Many articles that I've read which modeled richness or counts while accounting for spatial autocorrelation seem to simply transform the response variable (either log or sqrt), and then apply auto-Gaussian methods (e.g., AR, SAR or CAR). Is this the norm? Or is there some way to model the raw (i.e., non-transformed) counts?

Thanks!

Cheers,
Jenn