DATE AND TIME: Friday, March 2, 2018, 2:10 p.m.
PLACE: Snedecor 1109
SPEAKER: Nicholas Clark
Self-Exciting Spatio-Temporal Models for Count Data
The modeling of spatio-temporal count data often assumes conditional independence given a correlated latent Gaussian structure. However, in the modeling of the spread of violence and crime, a latent structure does not allow for possibility of repeat or near-repeat victimization. In this presentation we will introduce the spatially correlated integer generalized heteroskedastic (SPINGARCH) model that can be used for spatially and temporally correlated count data allowing for repeat victimization. We will demonstrate how this model arises naturally from straightforward assumptions on how violence or crime evolves over space and time. We will further show how this model allows for a wide range temporal and spatial correlation while only minimally impacting the variance to mean ratio, fixing an issue with the well-known INGARCH model. We will also demonstrate how efficient Bayesian inference can be performed using off-the-shelf Bayesian software and show how the model performs analyzing burglaries in the south side of Chicago.