There are many different ways by which we can model a cluster analysis, and depending on which software you use, there may exist ways to tes and adjust for the existence of spatial autocorrelation or not. In clinical surveillance studies, SaTScan is a very popular software that is used by over 100 countries, and it does not allow any adjusting for spatial autocorrelation. Moran's I is used widely by environmental scientists and by social scientists and by others. There are also Bayesian models and algorithms that allow the incorporation of spatial autocorrelation. My questions focuses on what do you do about spatial autocorrelation when you use data for a cluster analysis on some geographic regions?
I think there may be a little confusion about the use and motivation (and necessary technique) for different types of analysis. Spatial autocorrelation is certainly an important issue and you are correct that it is handled well in many social science and other studies. By this I mean that the presence of spatial autocorrelation is detected and if necessary the model framework can be modified to incorporate it. (This would typically be in regression type situations.) Cluster analysis has a different agenda. You WANT to find clusters (if they are there). The correlation between other spatial observations is what we are trying to determine. I think the SaTScan package is specialized towards this later problem and so it is understandable that they have no particular need to do spatial autocorrelation types of assessment or adjustment. It does however adjust for the underlying spatial inhomogeneity of a background population.
Take a look at Bailey and Gatrell's book on exploratory spatial data analysis.
- Badges
- Science topic
- Similar topics
- Political Science