Auditing iRAP’s ViDA Risk Engine: A Two-Stage Surrogate Learning and Orthogonalized Heterogeneity Framework for Modelled Road Safety

Road safety studies commonly use machine learning to predict crashes or to estimate crash-based treatment effects. This study instead audits the modelled fatal-and-serious-injury (FSI) risk produced by the iRAP ViDA risk engine. We analyse 147,466 segments (100 m each) from 12 surveys grouped into four European reporting groups. In Stage 1, gradient-boosted trees reproduce the engine’s risk surface under road-grouped cross-validation(R2 ≈ 0.92 with flows and survey identifiers), and Shapley-based attributions identify which coded attributes drive modelled risk at 396 hotspots (top-three segments per road). In Stage 2, a causal-forest double machine learning estimator adjusts for 38 covariates to estimate segment-level conditional contrasts between modelled risk and six retrofittable treatments across all eligible segments. Simple absolute and relative reduction thresholds translate these associations into 1170 association-based candidate upgrades. On 321 over-lapping hotspots, the candidate upgrades show moderate agreement with iRAP’s Safer Roads Investment Plan (Recall = 0.77; Precision = 0.66; Cohen’s κ = 0.40). All results are conditional associations on a calibrated risk engine whose totals are anchored to project- or network-level fatality totals or fatality estimates used in calibration, not causal effects on observed crashes.