Exploring the usefulness of the INLA model in predicting levels of crimes in the City of Johannesburg, South Africa

Abstract

Abstract Crime prediction serves as a valuable tool for deriving insightful information that can inform policy decisions at both operational and strategic tiers. This information can be used to optimize resource allocation and personnel management for crime prevention. Traditionally, the Poisson model has been the widely used model for crime prediction. However, recent statistical advancements introduce Integrated Nested Laplace Approximations (INLA) as a promising alternative for spatial and temporal data analysis. This study focusses on crime prediction using the INLA model. Specifically, the first-order autoregressive model under the INLA modelling framework is employed on longitudinal data for crime predictions in different regions of the City of Johannesburg, South Africa. The model parameters and hyperparameters considering space and time are estimated through the INLA model. In this work, the suitability and performance of the INLA model for crime prediction is assessed, which effectively captures spatial and temporal patterns. This study contributes to research by first introducing a novel approach for South African crime prediction, secondly developing a model using no demographic information other than clustering attributes as an exogenous variable, thirdly quantifying prediction uncertainty, and finally addressing data scarcity through demonstrating how INLA can provide reliable crime predictions, where conventional methods are limited. Based on our findings, the INLA model accurately ranked areas by crime levels, achieving a 29% Mean Absolute Percentage Error (MAPE) and 0.8 R-Squared value for crime predictions.