Development of Safety Performance Functions in the Piemonte region

A key focus of road safety research is understanding how the operational and geometric characteristics of roads, segments, and junctions relate to the frequency and severity of crashes. This relationship is formalised by the Safety Performance Function (SPF), which helps explain crash mechanisms and supports the design of targeted strategies to reduce crash frequency and severity. This thesis contributes to this effort by developing data-driven SPFs for the Piemonte region that account for traffic exposure and infrastructure factors. The analyses focused on rural highways and used data from three official sources: (1) the national crash dataset (2012–2022) provided by ISTAT, and (2–3) the BDTRE datasets of road segments and nodes supplied by Regione Piemonte. Duplicate or incomplete records have been deleted, and each incident has been linked to the relevant segment or node. This process produced two new datasets containing annual crash counts for both segments and intersections. Outliers were removed to limit their influence on parameter estimates while preserving underlying trends, preparing the datasets for modelling. Several count data models were estimated, including Poisson (P) and negative binomial (NB) regressions, which are widely used in crash frequency analysis. Due to excess zeros, zero-inflated Poisson (ZIP) and zero-inflated negative binomial (ZINB) models were also tested, with a logit model for the zero-count state. A provincial-level random intercept captured systematic differences due to territorial heterogeneity. Model performance was compared using the Akaike and Bayesian information criteria, with the latter preferred for large samples. Basic models were first developed using average annual daily traffic (AADT) and segment length, and maximum and minimum AADT for nodes. Results showed that the ZINB model with a provincial random intercept best captured overdispersion and excess zeros. Advanced SPFs incorporated additional traffic and infrastructure variables, including AADT, segment length, public transport presence, heavy vehicle rate, and road tortuosity. The probability of zero inflation was found to depend on the logarithm of segment length. For nodes, a standard NB model (without zero inflation) was most appropriate, as zero counts were less prevalent than for segments. The key variables for the final node model were AADT, altitude, and the presence of public transport and traffic control systems such as traffic lights. Although Italian crash data include only injury and fatal collisions (excluding property-damage-only events), this study shows that issues related to excess zeros and limited comparability with international SPFs can be effectively addressed. Standard SPF forms from the Highway Safety Manual proved unsuitable in this context, but the application of zero-inflated regression models with random effects overcame these limitations and produced robust estimates. Overall, the study demonstrates the potential for developing regional SPFs in Piemonte that are both methodologically sound and practically valuable for road safety management.