Study population and setting
This paper provides forecasts of the COVID-19 epidemic trajectory in each US State and demands on the healthcare system from March through July 2020. The authors used a statistical approach to estimate the expected daily number of deaths in each state by assuming that the daily death rate would roughly follow a normal distribution. To project the peak mortality level and the day when it would occur, the authors extrapolated from the relationship between mortality and the timing of key non-pharmaceutical interventions (e.g., lockdowns) from Wuhan, China. Assumptions about ICU utilization and age-specific death rates were based on data from Italy, China, Korea, and the US. Projected death rates are then used to estimate hospital service utilization using an individual-level microsimulation model, which attempts to describe how macro-level policies affect individuals.
Summary of Main Findings
Authors estimated the peak of the US epidemic to occur in the second week of April 2020, during which time excess demand from COVID-19 was expected to be: 64,175 (95% uncertainty interval [UI]: 7,977-251,059) total beds; 17,380 (95% UI: 2,432-57,955) ICU beds; 19,481 (95% UI: 9,767-39,674) total ventilators. The projected date of peak excess health system demand varied across states from mid-April through May. In total, the authors estimated that there will be 81,114 (95% UI: 38,242-162,106) COVID-19 deaths from mid-March 2020 to mid-July 2020 in the US, with less than 10 deaths per day expected between May 31 and June 6.
This model provides estimates across all US states and includes healthcare utilization forecasts, which are critical for planning and decision-making. Results for each state are accessible through an online visualization tool.
Uncertainty in model estimates should increase the further out in time a model attempts to make predictions (e.g., months out); however, in this study, uncertainty shrinks, which suggests that there are many uncaptured elements. The authors rely on a purely statistical modeling approach with no epidemiological basis, and the assumed a priori shape of the mortality curve prevents the model from being able to capture complex and diverse disease dynamics experienced across states. The forecasts are largely driven by timing and intensity of social distancing measures, and do not account for adherence to social distancing measures, social norms, or underlying health characteristics of the population beyond age. Authors assume any US state implementing three of four interventions (school closures, closing non-essential services, shelter-in-place, and major travel restrictions) will see an epidemic trajectory similar to that reported in Wuhan, China. This is unlikely considering multiple additional interventions implemented in China, such as mandatory masks in public. Finally, the authors do not present any data on how well the model predicts (e.g., such as through running it using previously observed outcomes).
This study presents the first set of estimates of predicted health service utilization and deaths due to COVID-19 for each US state, assuming social distancing is maintained throughout the epidemic.