This project aimed to model and predict the severity of traffic crashes across 22 local areas in Mazandaran province, Iran. The goal was to improve understanding of spatial crash patterns and inform safer traffic policies by applying Bayesian spatial models to injury severity data.
**Statistical Tools:** R (INLA, MCMCpack), Bayesian hierarchical modeling
**Models:** Structured Additive Regression (STAR), Multinomial Logit
Techniques: Multinomial-Poisson Transformation, Conditional Autoregressive (CAR) spatial effects
Analyzed crash data from 2016–2018 categorized into **three severity levels**: no injury, minor injury, and serious injury.
Selected **speed, driver age, and time of day** as covariates.
Built a **Bayesian STAR multinomial model** to capture:
Non-linear effects (e.g. of driver age)
Spatial dependencies across 22 regions
Applied **INLA** as a faster alternative to MCMC for approximate Bayesian inference.
Addressed the identifiability problem in multinomial models by comparing:
**Corner-point constraint** (fixing one category as baseline)
**Sum-to-zero constraint** (treating categories symmetrically)
Compared model performance against the **fractional split multinomial model**.
The **sum-to-zero constraint** produced more stable and interpretable estimates than the corner-point constraint.
**Speed** had a significant positive effect on serious injuries — reinforcing the known link between high speeds and crash severity.
**Driver age** showed a non-linear effect:
Young drivers (around 25) were linked to more severe crashes.
Older, more experienced drivers (ages 45–60) were more likely to be involved in no-injury crashes.
Spatial patterns showed **clustering of crash severity** — particularly in **tourism-heavy and rural regions**.