Visualization and Modeling for Crime Data Indexed by Road Segments

This research develops crime hotspot analysis and visualization methodology that use street segments as the basic study unit. This incorporates the distance between points along a polyline rather than the standard Euclidean distance and has some distinct advantages over past methods. For each crime, this method creates a weight according to its distance from each road segment of its surrounding block. To create the hotspot visualization map, crime counts are smoothed over road segments based on the distance to nearest segments and the angle at which nearest roads meet at intersections. Crime data from the City of Alexandria, VA Police Department and San Francisco, CA (available at data.sfgov.org) are considered here using a combination of conventional ArcGIS and R graphics.;I assume that demographic variables related to crime in large areas are still relevant to crime rates at the local level and seek to make use of the most spatially detailed data accessible. Decennial demographic variables at the block level for 2010 from the U.S. Census are associated with road segments by assigning the available values to the surrounding segments of each block. These variables include age, gender, population, and housing for both locations. Variables also considered are police calls for service, housing prices, elevation and speed limits.;I discuss/compare area crime counts with polyline crime counts using (zero-inflated) Poisson and Negative Binomial regression with crime-related covariates, as well as MCMC Poisson-Gamma Conditional Autoregressive (CAR) model in CrimeStat IV and a localized CAR model in R using distances between segments as weights. Conditional variable importance is measured using conditional random forest modeling to see which of the covariates are the most important predictors of crime and to decide which variables are the most appropriate to consider for visualization. Principal components are also used to create independent linear combinations of predictor variables. While most visualization approaches for street segments have emphasized one variable at a time, this research uses a 3 x 3 grid of maps using DPnet to highlight each grouping of road segments associated with classes based on two covariates. This multivariate visualization will allow us to explore multiple variables at a time and their patterns along a road network.