This project studies the spatial distribution and patterns of cannabis-related crimes in Chicago, focusing on understanding the potential influence of selection bias in the data. By employing advanced spatial analytics and feature engineering techniques, the study aims to develop a comprehensive spatial risk model that captures the localized clustering of these offenses. The analysis incorporates various risk factors, such as the presence of vacant and abandoned buildings, non-functional street lights, and proximity to public parks and police stations. Through the use of grid cell aggregation and the calculation of nearest neighbor distances, the project creates a nuanced exposure relationship across space. The study employs Local Moran's I statistic to examine the spatial autocorrelation of cannabis crime counts and introduces novel spatial process variables to capture significant clusters. Cross-validated Poisson regression models are used to assess the generalizability of the findings, comparing the performance of models with and without spatial process features. The project also investigates the model's generalizability across different neighborhood contexts, considering racial demographics. Finally, the spatial risk model's predictive performance is compared to the traditional kernel density hotspot mapping approach used by police departments. The recommendations highlight the need for caution in implementing the algorithm due to potential biases in the data and the importance of further scrutinizing the spatial processes of crimes before deployment.
.png)