Thanks all. Yeah I’m considering some of the methods you’ve shared. I think my interest is a very fundamental one – any of us who have considered the problem of spatial autocorrelation before usually use a “nearest neighbors” approach, basically for simplicity’s sake. But not all “spillover phenomena” are necessarily primarily based on geographic proximity. In COVID’s case, the car-based trips may account for much more epidemiological interaction than local pedestrian trips. So one of the ways to think about Safegraph data is as a “mobility-weighted neighbor matrix” that can be used to replace the “nearest neighbor matrix” in spatial autocorrelation. And the idea of creating contiguous clusters would be similar, if you were doing something like a stepped wedge RCT and wanted to ex ante define groups that minimized interference.