Hierarchical Bayesian Analysis of Aviation Safety
A hierarchical Bayesian study of fatality rates across aircraft manufacturers, showing partial pooling beats complete pooling by ~1,944 ELPD.
A study of fatality-rate patterns across aircraft manufacturers, fitting hierarchical and non-hierarchical Bayesian models in R with brms (Stan backend) to 1,766 aviation crashes spanning 40 manufacturers from 1976 to 2024.
The question
Are some manufacturers genuinely safer than others, or do apparent differences vanish once you account for how few crashes each manufacturer contributes? A naive per-manufacturer rate is noisy; complete pooling hides real variation. Hierarchical partial pooling sits between the two.
What the models showed
The hierarchical partial-pooling model outperformed complete pooling by 1,944 ELPD points, confirming significant manufacturer-level variation in fatality rates that persists despite industry standardisation. The temporal trend, by contrast, was minimal.