Spatial variation of bus stop crime response to changes in the surrounding environment and transit level of service

Abstract

Increasing concerns over the significant number of crimes occurring at and around bus stops have motivated the analysis of intervening factors to propose solutions to this problem. However, models found so far overlook important features of crime data: spatial dependence and spatial heterogeneity. In addition, crime predictor data has shown to be multicollinear in previous studies. To tackle these issues, this paper analyzes mobile phone thefts and robberies at 19,329 bus stops in São Paulo (Brazil) based on Geographically Weighted Regression (GWR), using components retained from a Principal Component Analysis (PCA) as explanatory variables. A comparison is carried out between GWR and a non-spatial Transformed Linear Regression (TLR), and a Negative Binomial Regression (NBR) with uncorrelated predictors. Effects on crime from 9 PCs, representing central areas, bus transit level of service, transport infrastructure, land use and sociodemographic features, were proven to have high spatial variability. Changes in the surrounding environment can cause higher or lower increases in mobile phone thefts and robberies at stops according to their spatial location. Results showed that GWR performs better than NBR and TLR in predicting bus stop crime, thus compensating for a loss of information associated with PCA. In addition, GWR was able to completely incorporate the spatial dependence found in the non-spatial model, which covered the nearest 96 neighboring stops. The proposed method can effectively detect critical bus stops and regions, contributing to crime prevention. GWR coupled with PCA can also help identify the best locations to install new bus stops.