Improving Recidivism Prediction and Causal Inference: From Regression and SEM to Instrumental-Variable Models

Risk assessment tools in criminal justice are typically evaluated on predictive performance, with much less attention to whether the risk factors they include have causal effects on recidivism. This study compares classical prediction models—probit regression and latent‐variable structural equation models (SEM)—with instrumental‐variable (IV) extensions to examine both prediction and causal inference in the context of a validated risk protocol (RisCanvi) used in Catalonia. Guided by a large automated meta-analysis of over 300,000 publications on violence and recidivism, we selected four core criminogenic domains—antisocial traits, antisocial history and peers, work/social instability, and substance use—and modelled 3-year recidivism in a cohort of individuals assessed with RisCanvi between 2015 and 2023. Classical SEM achieved the highest discrimination (AUC ≈ .74) but poor calibration, whereas probit models were well calibrated but had very low sensitivity. IV-SEM preserved discrimination while substantially increasing the estimated effects of criminogenic needs and achieving the highest sensitivity, at the cost of worse calibration. IV-probit produced calibrated probabilities and modest gains in discrimination over classical probit. These findings illustrate a methodological “division of labour”: classical models provide transparent, well-calibrated risk estimates for routine decisions, while IV-based models offer stronger evidence on which criminogenic needs are true levers for change, informing the design and targeting of interventions and policy in recidivism reduction.