A Bayesian spatio–temporal INLA framework for network-wide safety performance functions: Application to Abu Dhabi road networks

Rapid urban growth in Gulf Cooperation Council (GCC) countries has increased traffic exposure and crash risk, creating the need for locally calibrated safety models. This study develops a Bayesian spatio-temporal framework to estimate Safety Performance Functions (SPFs) across more than 57,000 road segments in Abu Dhabi. Police-reported injury crashes from 2013–2020 are modelled using a Negative Binomial specification within the INLA framework, incorporating roadway characteristics, exposure measures, continuous spatial dependence, and an annual effect. The results show a clear safety hierarchy: junction links (intersections and roundabouts) experience several times the crash frequency of mid-block segments, which all sit below the network average, while higher traffic volume, free-flow speed, and number of lanes are each associated with elevated crash frequency. The relationship between volume-to-capacity ratio and crash frequency is broadly increasing, with crash risk elevated under denser and over-saturated conditions. The annual effect shows a reduction in 2020 consistent with COVID-19 mobility restrictions. Using a continuous spatial field (SPDE), this study identifies a spatial influence radius of about 2–3 km, a locally calibrated screening reference for corridor bundling and control-site selection that complements, rather than replaces, existing buffer techniques. The framework delivers interpretable, locally calibrated safety parameters that support transparent and cost-effective investment decisions in rapidly motorising metropolitan networks.