The Influence of Housing, Parcel, and Neighborhood Characteristics on Housing Survival in the Marshall Fire

This study develops random forest models to interrogate housing survival in the 2021 Marshall Fire, investigating the role of housing, parcel, and neighborhood characteristics. This grass fire affected suburban communities and destroyed more than 1,000 houses. The authors compiled a data set consisting of all the destroyed houses, along with damaged and standing neighboring houses. After removing houses with insufficient data, 1055 impacted houses were used to develop models for each of the three impacted jurisdictions and for the full data set of 1055 houses. In addition, model versions were developed that use only the subset of predictor characteristics available pre-fire. The pre-fire model results showed that the five housing characteristics that resulted in the largest increase in mean square error (MSE) when randomly permutated were predominantly neighborhood and parcel level characteristics. All predictors resulting in percent increases in MSE of 15% or greater were neighborhood level characteristics. Parcel and community characteristics encompass 78%, 86%, 100%, and 80% of predictors resulting in percent increase in MSE greater than 5% for the models developed for Louisville, Superior, Unincorporated Boulder County, and all data, respectively. Additionally, the identification of the most important parameters showed that the majority of the most impactful variables were not within the homeowner’s control. This conclusion demonstrates the importance of neighborhood and community characteristics on housing survival that are controlled by the jurisdiction, especially in a home rule state where building codes and planning may differ across jurisdictional boundaries. We found little reduction in model accuracy (%-change in balanced accuracy under 12%), when only pre-fire variables were considered. Taken together, these results suggest a crucial role for jurisdiction or community mitigation of risk in WUI areas.