Treatment of Spatial Autocorrelation in Geocoded Crime Data

In modelling spatial data, when measurements at one location are influenced by the measurements at neighbouring or nearby locations, we say that spatial autocorrelation is present. This violates the assumption of statistically independent observations commonly applied in standard regression analysis. We first examine the basic theory and methods used to analyse spatial lattice data (i.e., data aggregated to regions as opposed to observations at discrete points). Next, we focus on the two basic forms of simultaneous autoregressive models, termed the spatial lag and spatial error models. Emphasis is placed on how the spatial effects are incorporated into the model. Finally, we implement the spatial models in a study of urban neighbourhood crime rates in the Canadian city of Montreal. We examine the differences in the models obtained with and without modelling the spatial effects