Optimal COVID-19 Containment Strategies: Evidence Across Multiple Mathematical Models

Abstract

Since March 2020, numerous models have been developed to support policymakers in understanding, forecasting, and controlling the COVID-19 pandemic. Differences in data, assumptions, and underlying theory, coupled with unknowns about a novel virus, led these models to generate divergent forecasts and proposed responses. A policymaker using a single model is left to wonder if their decision is truly of high quality or if they are being misled by the idiosyncrasies of the selected model. In addition, many COVID-19 optimization models are cast as optimal control problems with abstract decision variables and frequent changes to policy, so translating the optimal solution to implementable actions is not straightforward.

We propose a multi-model optimization (MMO) framework that identifies policies that perform well across structurally distinct models, and we apply this to design 12-month COVID-19 containment strategies. Our approach differs from the existing literature in two important aspects. First, we optimize using multiple state-of-the-art forecasting models currently in use. Second, we intentionally draw feasible intervention levels from each state’s own past and current responses, making it easy to implement the proposed policy.

We find that a policy based on a single model can perform badly (cost increases of 100% or more) when models are misspecified, and that the MMO policy significantly diminishes the impact of model uncertainties. We propose optimal containment policies for all 50 US states over a one-year period and find that the optimal policy can vary significantly by state. We also study the impacts of virus variants and lockdown fatigue.