> Spatial statistics aren't the same as regular statistics<p>I've always been frustrated with the gap between statistics and spatial statistics. For example, some of the methodologies with conducting hot-spot analysis is somewhat misleading, especially to uninformed geospatial analysts. For example, Esri [0] implements this first by conducting geospatial aggregation, then calculating z-scores based on Gaussian assumptions, then generates a corresponding "p-value" to extract "statistically significant areas" that are coined "hot spots". At that point, an analyst typically color-codes those p-values showing regions with low p-values as "extreme" areas of interest. I'm really curious if there's any empirical or anecdotal research that validates this methodology.<p>There are some attempts to try and normalize sampled data. Location Quotient [1] (and Standardized Location Quotient), for example, compares a local measure to a global measure. However, this too has Gaussian assumptions and doesn't properly account for variance in the data.<p>I would definitely love to see a hierarchical Bayesian spatial model that takes into account a geospatial prior (such as the overall density of tweets) allowing you to solve for the posterior of cluster centers. Has anyone seen this done before?<p>[0] <a href="http://resources.arcgis.com/en/help/main/10.1/index.html#//005p00000010000000" rel="nofollow">http://resources.arcgis.com/en/help/main/10.1/index.html#//0...</a><p>[1] <a href="http://www.bea.gov/faq/index.cfm?faq_id=478" rel="nofollow">http://www.bea.gov/faq/index.cfm?faq_id=478</a>