Good Group 1 GoF, Poor Group 2 GoF

[Emily Blackwell]

Hi everyone,

I've been working with spOccupancy to model occupancy from a large multi-city camera trapping project. An issue I've been having with my models is poor GoF as calculated by the ppcOcc() function. I seem to have made some progress and found a detection formula that yields decent community and species level Bayesian p-values when calculated across group 1 (site), but the group 2 (replicate) Bayesian p-values are still very poor. I'm not really sure how to interpret this discrepancy, much less try to improve it. For further context, these results are from a model with a detection formula that includes random effects for city and site and an effect for impervious surface cover. The occupancy formula is ~1. Any insights would be appreciated!

[Jeffrey Doser]

Hi Emily,

Thanks for the note. The interpretation of the different GoFs can be tricky at times, but the problem likely has to do with the lack of any occupancy covariates. Is this model part of a larger model comparison analysis? If so, I personally would not worry about looking at the GoFs for all candidate models because they are likely to just show up as not fitting well, particularly for simpler ones that assume occupancy and/or detection are constant. Instead, if doing model comparison, I would do the comparisons, I would determine the top-performing model and then do GoF assessment for that model. In your case here, you are assuming occupancy is constant across all sites in your data set if fitting a non-spatial model. If fitting a spatial model, the only thing allowing spatial variation in occupancy would be the spatial random effects. In either case, the results you are finding are not too surprising and instead are indicating that there is likely spatial variation in occupancy probability that your null occupancy model is not accounting for. This provides support for including more covariates and/or flexibility in the occupancy portion of the model. So, the likely way to remedy this is to include covariates that you think impact occupancy probability in the occ.formula.

Hope that helps,

Jeff