It pleases this geographer to see the dissemination of the newer developments of spatial models in marketing to a broader audience. Marketing is inherently spatial. More often than not, key decision makers ignore the impact of space, or handle it too simplistically to aid decision making. Space as expressed in terms of distance, location, contiguity, networks, or even spacetime metrics, influences the behaviours of consumers and business in many different ways.
Bronnenberg outlines how many of these areas are modelled in marketing today. Generally these models capture the spatial influences directly as independent variables in the model itself or as spatial 'weights' designed to model the spatial similarity among observations. In the latter case, the weights are designed as measurable components of the model error term. Both these approaches are designed to improve model performance (i.e. explanation in some cases, prediction in all cases) by modeling the heterogeneity among observations that are correlated with distance metrics. Yang and Allenby [1] build a hierarchical Bayes choice model using patterns of both space and social networks. It is particularly intriguing because it goes beyond using pure spatial contiguity as a metric of heterogeneity.
In human society, a process of spatial congregation and segregation occurs. For a variety of reasons, people, businesses, and other activities have a tendency to congregate with others like themselves with whom they share common interests, activities, needs, desires, etc. This tendency to congregate also leads to separation from others in space. The segregation of people, businesses, and activities coexists with the congregation process because not everyone can be everywhere simultaneously. This leads to the spatial patterns of human activity studied by geographers and marketers alike. While I have not described the concepts of spatial congregation and segregation, the point I make is that human activities in space will display a high degree of spatial autocorrelation because of these processes. As such, any model that attempts to account for respondent heterogeneity through the use of spatial variables, or weights, will predict better than models that do not.
The use of spatial autocorrelation models to improve prediction, either in time, across space, or in the instances of missing data are important improvements to current models in marketing. However, I caution the user of spatial models that the improved predictability should not come