Predicting Environmental Violations: A Cross-Method Framework Integrating Parametric and Non-Parametric Approaches

Prior research has mostly relied on linear and parametric models while explaining environmental non-compliance, but they have a limited capacity to capture non-linear and asymmetric effects of elements affecting firms' environmental compliance. We integrate a random effects logit model with advanced machine learning methods on a longitudinal firm-year panel dataset of US publicly traded firms from 2000 to 2024 in order to address this gap. After comparing the random effects logit model estimates with XGBoost metrics and SHAP mean values, we find that firms' prior environmental compliance behaviour is a major predictor of future environmental penalty violations. The random effects logit model confirms the cumulative penalty count as a significant positive predictor, while XGBoost assigns it the highest gain and SHAP importance scores. The marginal contribution of the cumulative penalty count in the SHAP analyses suggests its diminishing marginal effects, which are hard to identify using the logit model alone. Cross-method comparison further reveals that governance and executive incentive variables are weak predictors of future environmental violations. The findings reveal the trajectory-dependent nature of environmental non-compliance, with enforcement outcomes strongly associated with behavioural trajectories. This study offers a methodological contribution and demonstrates the value of cross-method triangulation in resolving inconsistencies across parametric approaches by integrating parametric analysis and machine learning methods.