I'm working with non-negative count data at the county level, with my dependent variable being the number of farmers with agricultural insurance (a hypothetical example here). I'm using a negative binomial regression (over-dispersed count variable) with theoretically relevant predictors and control variables to account for county size.

I'm examining the residuals of the models, and found some spatial dependence after running Moran's I test (I: 0.054, p < 0.01). In my social sciences field, some have suggested it's not worth pursuing additional testing if Moran's I score is very low even if the p-value is below the established significance threshold (see here). Other published studies in my field have started off with considerably higher Moran's I scores (eg., ~ 0.40, p < 0.01), which decrease after using a spatial error regression and are deemed acceptable, even if significant (eg., 0.059, p < 0.05). I suppose here the assumption is while, in a strict sense, the independence of the residuals is violated, the impact was judged to be small, or least come down a lot after the use of a spatial error regression. This is similar to my case.

I'm wanting a second opinion on taking this route. In addition, I'm wondering if anyone can guide me showing me a sample code on running a spatial autoregressive model with a negative binominal structure. That would be incrediblely helpful to me.