Beyond Victimisation: Modelling Crime Reporting Behaviour Using Machine Learning Techniques in South Africa

Crime reporting is essential for justice administration; nevertheless, it is still highly under-reported throughout the world, especially in developing nations. In South Africa, victimization studies show consistently high levels of under-reporting of crimes, reducing the credibility of crime statistics. In this research, the data from the Governance, Public Safety and Justice Survey (2023-2024) representing national sample of 21,121 respondents were analysed employing a quantitative research design, specifically multivariate logistic regression coupled with several machine learning algorithms (Random Forest, XGBoost, & Neural Networks). The models were tested using ROC-AUC, precision, recall, F1-score and Brier score metrics, while their interpretability was analysed using SHAP (Shapley Additive Explanations). The crime reporting rates were low (~4.4%). According to logistic regression, significant predictors of reporting were type of crime (OR = 2.45), trust in police (OR = 1.82) and perceived effectiveness (OR = 1.56). Machine learning algorithms outperformed linear regression with best model being XGBoost (AUC = 0.796). The SHAP analysis showed that previous contacts with police are the key predictor of reporting. Crime reporting is largely institutional rather than socio-demographic issue. Machine learning methods demonstrate better predictive ability and shed light on complex behavioural dynamics, thus providing valuable insights for public health and criminal justice policy.