Toward an Audit-Informed Training Effectiveness Index for Construction Safety: A Bayesian Hierarchical Model Using Objective Performance Proxies

Authors

• Duc A. Nguyen, Hanoi University of Civil Engineering
• Hong-Ha Le, Hanoi University of Civil EngineeringFollow

Abstract

This paper proposes a Bayesian hierarchical model for a Training Effectiveness Index (TEI) inferred from four objective performance proxies: inspection non-conformities, permit-to-work violations, near-miss events, and corrective-action closure time. Applied to data from ten subcontractors over six quarters in Hanoi, Vietnam, seven MAP optimizations all converged to log-posterior = −594.94 and σα = 1.316 (95% CI: 1.193–1.439) confirms identifiable between-subcontractor heterogeneity. LOSO cross-validation (ELPD = −685.0) confirms hyperparameter transferability. Near-miss counts are near-redundant (r = 0.9996 with three-proxy model; max difference 2.8 points). Comparison with PCA and z-score baselines yields r = 0.913. Key limitations - cells-to-parameters ratio of 1.94 and negligible period effects (1.4 TEI points) - are quantitatively disclosed. Results are a proof-of-concept requiring multi-project replication.

Keywords

construction safety; training effectiveness; Bayesian hierarchical model; latent variable model; leave-one-out cross-validation; non-conformities; corrective action

Recommended Citation

Nguyen, Duc A. and Le, Hong-Ha (2026) "Toward an Audit-Informed Training Effectiveness Index for Construction Safety: A Bayesian Hierarchical Model Using Objective Performance Proxies," CIB Conferences: Vol. 2 Article 51.
DOI: https://doi.org/10.7771/3067-4883.2209