Hierarchical Bayesian estimation of a probabilistic fire occurrence model considering the effects of size, occupancy, and time-of-day

Evaluating ignition probability is essential for assessing fire risk in buildings. Previous studies have typically analyzed factors such as building size and occupancy using aggregated fire statistics. However, this approach restricts applicability to categories with sufficient fire incident data. In this study, we constructed a non-aggregated ignition probability model by integrating Geographical Information System (GIS) data and Fire Incident Reports (FIR) data for Tokyo in 2016. Logistic regression was employed as the basic framework, with three models compared through the stepwise introduction of explanatory variables: (A) a model with only floor area, (B) a model additionally incorporating occupancy as a random effect, and (C) a model further introducing time-of-day effects through fixed effects and a conditional autoregressive (CAR) structure. Across all models, floor area consistently showed strong explanatory power, with ignition probability increasing nonlinearly as floor area expanded. Model (B) quantified differences in risk across occupancies alongside the effect of size, while Model (C) did not yield significant time-of-day effects. These findings provide a quantitative basis for rational evaluation of building fire risk.