Quantifying Transferability in Road-Safety Digital Twins via Mechanism–Exposure Decomposition

Conference poster, 2026

Transferring safety insights across jurisdictions is a core challenge for road-safety digital twins, yet no interpretable framework exists to quantify when and why such transfer succeeds or fails. We propose a two-stage, interpretable framework for assessing cross-domain transferability in road-safety digital twins. First, Causal Invariance Screening (CIS) uses sample-size-independent effect-size criteria to partition crash variables into an environment-stable mechanistic set  (environment-invariant) and an exposure complement  (scene-composition). Second, four symmetric, model-free indicators quantify pairwise transferability: exposure overlap (COI), mechanism divergence (CMD), risk-gradient preservation (RGP), and high-risk alignment (OHRA). We validate the framework through a three-tier design of increasing domain divergence using three crash databases (US FARS, US CRSS, and UK STATS19): temporal splits within the US (positive control), cross-database comparison within the US (medium difficulty), and US–UK cross-country comparison (negative control). External validation across three model families (logistic regression, random forest, XGBoost) confirms that indicator scores align with cross-domain prediction degradation independent of model choice. CMD correctly orders all domain pairs by transfer difficulty, including a non-monotonic result where cross-country CRSS–UK transfers better than within-country FARS–CRSS due to shared sampling-frame structure. However,  alone is insufficient for deployment: a CRSS–UK model achieves  yet  due to 3.3% source prevalence. These results show that mechanism transferability and operational deploy ability are distinct: even when rank-based transfer remains high, low source severe prevalence can make severe-case detection unusable without recalibration.