Structural Housing Features, Exposure Context, and Wildfire Damage Severity in California's Wildland–Urban Interface

Abstract California CAL FIRE Damage Inspection (DINS) records from fire-impacted areas (analytic N = 89,826; 298 incidents) do not support a simple “code materials always help” story once exposure context sits in the model. A pre-specified confirmatory set of nine literature terms, with Holm correction on those nine rather than on 56 kitchen-sink dummies, keeps tile roofing (odds ratio [OR] = 0.55), enclosed eaves (OR = 0.74), wood siding (OR = 0.57), local canopy, Wildfire Risk to Communities (WRC) Direct exposure (OR = 1.76), and recorded defense (OR = 0.62) after incident-clustered standard errors. Metal roofing is an honest null (OR = 1.08). Neighborhood density’s clustered interval (OR = 1.38, 0.99–1.92) is compatible with no effect to a large one and does not travel from South to North. Combustible siding destruction rises from 61% to 96% across density tertiles, while wood’s protective-direction association strengthens in dense neighborhoods. An extended-binary CatBoost ensemble is kept for those density-gated siding bends and two residual phenotypes—packed low-canopy survivors (n = 289, three fires) and sparse hardened losses (n = 379)—not as a second odds-ratio table. Weather enters as a three-variable principal component analysis (PCA) with elevation held out as a geographic control. Incident-held-out leave-one-fire-out quadratic weighted kappa (QWK) is ≈ 0.48; random-split QWK ≈ 0.73–0.77 is in-fire discrimination only. McFadden R² = 0.1738. Associations are post-inspection and under exposure context—not measured ember flux and not a retrofit prescription.