Political Storms: Tracking Hurricane Evacuation Behavior Using Smartphone Data

Abstract

The 2017 season devastated the U.S. gulf coast with two of the worst hurricanes in history: Harvey (107 deaths, $125B in damages) and Irma (134 deaths, $50B in damages). Despite extensive warnings, most affected residents did not evacuate their homes before the storms hit, complicating rescue and recovery efforts. Combining a large GPS dataset for 2.7 million smartphone users in Florida and Texas with U.S. Census demographic data and 2016 U.S. Presidential election precinct-level results, we empirically examine evacuation behavior. A difference-in-differences analysis demonstrates that Trump/Clinton vote share strongly predicts evacuation rates, but only after the emergence of conservative-media dismissals of warnings in September 2017, just before Irma made landfall in Florida. Following this viral hurricane trutherism, we estimate that Trump-voting Florida residents were 10-11% less likely to evacuate Irma than Clinton-voters (34% vs. 45%) after controlling for key demographic and geographic covariates, highlighting one consequence of political polarization. This effect size is similar in magnitude to that of an official watch. We confirm the causal impact of advisories using a spatial regression-discontinuity design that compares evacuation rates for residents living just on opposite sides of county boundaries who received differential alerts. A watch causally increases rapid evacuations (within 24 hours) by 6 percentage-points compared to no watch, and by 4 percentage-points compared to a tropical storm watch.