Continuous-Time Information Design for Hurricane Evacuation: Disclosure, Congestion, and Optimal Phasing under Model Uncertainty

Abstract

We study continuous-time information design for emergency evacuation, where an Emergency Management Agency (the Stackelberg leader) steers strategic evacuation zones via two levers: public advisory precision (information design) and a tiered release schedule. The latent storm is a jump-diffusion process with publicly observed rapid-intensification epochs tracked by an exact finite-dimensional belief filter. Zones play a capacity-constrained congestion game on shared corridors with belief-weighted hazard exposure. The running cost couples beliefs to a convex congestion externality, making disclosure double-edged: sharper information reduces false-alarm departures but synchronizes genuine ones, and convex congestion penalizes that synchronization. We prove that: (i) the followers' game admits a potential reduction to a convex control problem; (ii) the leader's distributionally robust relative-entropy problem is characterized by an Isaacs equation whose value is the unique viscosity solution, with verification valid for non-smooth bang-bang feedback; and (iii) without transfers, the leader's first-order condition retains an equilibrium-response term, positioning optimal information design as a second-best congestion toll. Structurally, we show that a staggered evacuation order dominates simultaneous advisories; phased evacuation emerges endogenously as optimal information design. Furthermore, public-signal precision is sign-ambiguous due to an informational Braess effect, where vague advisories are optimal unless complemented by a staggered order. Calibrated to Hurricane Rita using NHC archives, TxDOT capacities, and HRRC surveys, the model reproduces the observed gridlock along the Interstate 45 (I-45) evacuation corridor in Texas. The optimal policy removes essentially all in-transit congestion exposure, reducing social cost by 89%, while staggered disclosure alone yields a 70% reduction.