Using mobile phone data for quantifying large-scale household-level disaster recovery

Natural disasters often result in evacuations, travel disruptions, power outages, school closures, and closures of other facilities affecting the ability of individuals to maintain their typical daily patterns. Visits to home and work follow regular patterns that may be interrupted due to a natural hazard. These disruptions impact productivity and well-being. However, there does not currently exist a way to estimate how long individuals' home and work routines were interrupted at the geographic scale of a large natural hazard. Surveys provide useful information, but only for small samples of the affected population. With surveys alone, we cannot model and understand the extent of recovery time across a large set of households with diverse experiences of the disruption. This lack of a method to assess widespread household-scale recovery of normal daily patterns is the key gap we address in this paper. We develop an approach to use location-based services data from smartphones to capture patterns in visits to home and work and deviations from those patterns that may indicate disruption and recovery while maintaining anonymity. We introduce a Bayesian belief network-based anomaly detection method to identify household-level lack of recovery and demonstrate this approach for Hurricane Irma. Our results show the proportion of users experiencing an anomalous period and the average length of recovery, validated against the limited available survey results. These large-scale data-driven results on household recovery contribute to further analysis on the impacts of the hazard and social vulnerability on recovery at the scale of individual homes and workplaces. • Location data from cell phones can reveal recovery of routines following a hazard. • Bayesian belief networks can identify anomalous behaviors in location visit data. • More Floridians faced disruptions in appearances at work than home after Irma. • On average, recovery took longer for home appearances than work appearances. • Data-driven approaches can supplement surveys for evaluating hazard recovery.