Title: Parametric and non-parametric Bayesian approaches to spatial modeling of crime in Philadelphia
Authors: Shane Jensen - The Wharton School of the University of Pennsylvania (United States) [presenting]
Abstract: Urban data analysis has been recently improved through publicly available high resolution data, allowing us to empirically investigate urban design principles of the past half century. We will focus on one particular direction: spatial-temporal modeling of the change in crime over the past decade in the city of Philadelphia. We will show that Bayesian parametric spatial models can improve the accuracy of crime predictions (compared to simpler methods) by inducing both global and local shrinkage between proximal neighborhoods. However, there is a need for even more sophisticated crime models that take into account the geography of the city and find larger regions that share similar trends in crime over time, which motivates nonparametric approaches that can clusters of neighborhoods. Conventional Bayesian nonparametric clustering priors, such as the Dirichlet process, do not naturally handle spatial data. Thus, we will explore different strategies for introducing spatial cohesion into our nonparametric spatial clustering models.