Predicting earthquake evacuation: an interpretable machine learning approach

The engineering of sustainable and resilient evacuation systems requires behavioural models that identify who is likely to evacuate and which barriers constrain protective action. Using stated-preference survey data from 725 residents in the New Madrid Seismic Zone, this study compares mixed logit, extreme gradient boosting (XGBoost), and Bayesian Additive Regression Trees (BART) for earthquake evacuation decisions. The sample showed a high stated intention to evacuate (79.3%), indicating both strong perceived risk and a potential for hypothetical bias. XGBoost achieved the best predictive performance (AUC = 0.808, accuracy = 76.1%, F1-score = 0.828), followed by BART (AUC = 0.792) and mixed logit (AUC = 0.744). The main innovation is a triangulated framework that combines econometric interpretation, interpretable machine learning through SHAP, and Bayesian uncertainty quantification. Across all models, budget availability, dwelling type, and information-seeking behaviour emerged as the most influential predictors. These results provide practical evidence for socially sustainable evacuation planning, including targeted financial assistance, communication strategies for single-family residential areas, and smart information systems that support equitable disaster response.