Analyses increasingly used gridded data at a high spatial resolution, for example in research on climate and conflict. Using such data as the unit of analysis can be helpful in characterizing local geographic variation relevant to the response. However, given that many features tend to be spatially correlated the effective N will be much smaller than the apparent N, and it is questionable whether current practices of controlling for first order spatial lags will be sufficient to address the consequences of spatial dependence. This paper examines the adequacy of common approaches, and proposes some simple sampling based alternative model approaches and diagnostics based on concepts from geostatiscal analysis.
