Mobile-phone mobility data are increasingly used to quantify the economic impact of natural disasters on local businesses by estimating counterfactual trajectories of customer visits. Bayesian structural time series (BSTS) models provide a flexible framework for such counterfactuals, but their performance depends critically on the choice of covariates used to construct synthetic controls. Existing disaster applications typically rely on ad-hoc or heuristic covariate choices, such as selecting a single highly correlated control series or simple brand-level aggregates. In this paper, we develop and evaluate an automatic covariate selection framework for BSTS models applied to mobility-based business visit data. Our approach constructs a rich pool of potential covariates from control businesses and auxiliary time series, and uses spike-and-slab priors to perform Bayesian variable selection within the state-space model. We further incorporate grouping structures at the business-category level to share statistical strength while avoiding overfitting. Using mobility traces of visits to retail and service businesses before and after a major hurricane, we show that the proposed method substantially improves pre-disaster predictive accuracy compared to heuristic covariate strategies. Improved counterfactuals yield more stable and interpretable estimates of disaster impact, both at the level of individual businesses and across categories and regions. Our results demonstrate that principled covariate selection is essential for reliable mobility-based impact assessment, and provide a general modeling template for future studies using BSTS in disaster, policy, and intervention analysis.
Mobile-phone mobility data are increasingly used to quantify the economic impact of natural disasters on local businesses by estimating counterfactual trajectories of customer visits. Bayesian structural time series (BSTS) models provide a flexible framework for such counterfactuals, but their performance depends critically on the choice of covariates used to construct synthetic controls. Existing disaster applications typically rely on ad-hoc or heuristic covariate choices, such as selecting a single highly correlated control series or simple brand-level aggregates. In this paper, we develop and evaluate an automatic covariate selection framework for BSTS models applied to mobility-based business visit data. Our approach constructs a rich pool of potential covariates from control businesses and auxiliary time series, and uses spike-and-slab priors to perform Bayesian variable selection within the state-space model. We further incorporate grouping structures at the business-category level to share statistical strength while avoiding overfitting. Using mobility traces of visits to retail and service businesses before and after a major hurricane, we show that the proposed method substantially improves pre-disaster predictive accuracy compared to heuristic covariate strategies. Improved counterfactuals yield more stable and interpretable estimates of disaster impact, both at the level of individual businesses and across categories and regions. Our results demonstrate that principled covariate selection is essential for reliable mobility-based impact assessment, and provide a general modeling template for future studies using BSTS in disaster, policy, and intervention analysis.
