Analytical Study of Predictive Policing Algorithmic Bias in Nigerian Urban Crime Data

Algorithmic bias in predictive policing systems trained on historically biased police data can perpetuate systemic inequities, and examining the potential for and dimensions of such bias in Nigerian urban crime data represents an important ethical data science research priority. This study analytically examined the risk of algorithmic bias in predictive policing models trained on Lagos State crime incident data. A sample of 28,000 incident records from Lagos State Police Command covering 2019 to 2022 was obtained under a research agreement. Bias analysis examined whether predictive hotspot models trained on this data systematically over-predicted crime risk in low-income communities controlling for actual crime prevalence. Spatial regression and counterfactual fairness testing were applied. Results showed that community socioeconomic status was a significant predictor of model-assigned risk independent of observed crime frequency (beta = -0.41 for income index, p < 0.001), consistent with differential enforcement bias amplification. Low-income communities in Mushin and Surulere had risk scores 2.3 times higher than comparable high-income communities with equivalent crime rates. Fairness metrics including equalised odds and demographic parity showed significant violations across socioeconomic strata. The study fills an important analytical gap in responsible AI for Nigerian law enforcement contexts and recommends NPF adopt algorithmic fairness auditing before deploying any predictive policing tool, bias correction post-processing methods, and community oversight panels for any AI-assisted policing system deployed in Lagos. Keywords: algorithmic bias, predictive policing, Lagos, fairness in AI, crime data analytics