Combining spatial and sociodemographic regression techniques to predict residential fire counts at the census tract level

This work examines different spatial and sociodemographic models for predicting residential fire counts at the census tract level 118 U.S. departments across 25 states. The give five-year forecasts of 3392 tracts which contain over 13 million residents in total. All described this paper train on incident data from National Fire Incident Reporting System (NFIRS) interval 2006–2011 (inclusive) are evaluated based their ability to predict that occurred 2012–2016. Two strictly considered- a simple “count” model serves as baseline all other utilizes kernel density estimation (KDE) with statistically optimized bandwidths. Using American Community Survey (ACS), an examination effects demographic housing factors risk is presented. suggest per person generally higher attributes corresponding socioeconomic disadvantage such low median incomes small fractions college degrees. These trends inform design Bayesian hierarchical Poisson regression model, shown make predictions 9% lower root mean squared error (RMSE) relative base model. A then conducted residuals regression, results 15% RMSE improvement compared conditional autoregressive (CAR) incorporates information directly into regression. Although CAR model's point estimate (7% than model), it allows generation probabilistic gives spatially-informed statistical estimates variables. highlights utility geocoded well machine learning techniques can utilize these datasets improved predictions.