Bayesian Railway Safety Analysis
Date:
Developed a reproducible Bayesian analysis of significant railway accidents across 27 European countries from 2010–2024, using a balanced panel of 405 country-year observations. (GitHub Link)
- Built and compared Poisson, Negative Binomial, and hierarchical Negative Binomial models, using train-km as an exposure offset to model accident rates rather than raw counts.
- Validated the modeling pipeline through prior and posterior predictive checks, MCMC diagnostics, parameter-recovery experiments, DIC comparison, sensitivity analysis, and frequentist benchmarking.
- Found an estimated 3.3% annual decline in significant accident rates, equivalent to approximately a 37% reduction over the study period, while identifying substantial persistent cross-country heterogeneity.
- Implemented the workflow in Python, R, JAGS, rjags, coda, ggplot2, and glmmTMB, with structured outputs for results, tables, figures, reporting, and reproducibility.
Technologies: Python, R, JAGS, Bayesian hierarchical modeling, MCMC, count-data regression, statistical diagnostics, data visualization