context
Studying accident patterns around toll removal
The project investigated changes in severe and fatal accident typology around the removal of motorway tolls in Catalonia.
problem
Avoiding leakage and misleading headline accuracy
The pre/post target was highly imbalanced and corridor definitions could leak the outcome into the model. A high headline accuracy would be misleading.
approach
Imbalance-aware evaluation with fixed thresholds
I engineered corridor and infrastructure features, fixed the decision threshold before test evaluation and compared models with imbalance-aware metrics. Grouped permutation importance separated explanatory blocks instead of over-reading individual correlated features.
results
Predictive structure without causal overclaiming
The selected balanced Random Forest achieved 0.6287 macro F1, 0.7206 ROC AUC and 0.3457 PR AUC. Road infrastructure and context formed the strongest explanatory block; toll/corridor features added a smaller measurable signal.
limitations
What observational data cannot establish
The analysis is observational. It identifies predictive structure, not a causal effect of toll removal. A causal design would need exposure data, traffic volumes and a credible comparison group.