On how population-at-risk estimates influence crime rates and predictions: Comparing residential and ambient-like estimates

The present study compares different residential and ambient-like population estimates across two key applications in spatiotemporal crime analysis: (1) the calculation of crime rates, and (2) the prediction of monthly micro-geographic crime risks using machine learning. Using data from Ghent, Belgium across three crime types (i.e., aggressive theft, battery incidents and bicycle theft), we compared traditional administrative residential counts with mobile phone data and various alternative population estimates (i.e., GHS, WorldPop, ENACT and LandScan). Results show that mobile phone counts provided the most robust proxy for the true population-at-risk, demonstrating stronger associations with crime risk and resulting in different crime rates and improved crime prediction performance measures. However, openly accessible alternatives such as GHS and WorldPop redistributed residential estimates performed comparably well, likely due to their incorporation of features that indirectly capture human activity patterns. The findings support crime opportunity theories and highlight that both the nature of population data and its spatiotemporal resolution may substantially influence crime analysis outcomes. For crime prevention, population estimates that directly or indirectly account for human mobility patterns provide avenues to improve crime risk estimation and resource allocation, with freely available alternatives offering cost-effective solutions when fine-grained data are inaccessible.