Aggregating Expert Opinion on COVID-19

Community-wide transmission of SARS-CoV-2 was reported in the US on February 26th. Since the virus was declared a pandemic by the WHO on March 11th, it has spread rapidly throughout the US and has caused more than 1.1M confirmed cases and 62K deaths. Forecasts provide public health officials at the state and federal level actionable information that can be used to mitigate the impact of an outbreak in advance. But forecasts from computational models often depend heavily on noisy surveillance data and take a large amount of work to build. Without guidance from forecasts, public health must rely on sparse past experiences to inform intervention. An expert consensus forecast, synthesizing both objective and subjective information, can be developed quickly and make predictions during the early phase of an outbreak, when computational models are still being built. In addition, an expert judgment model can address changing public health needs by asking different questions without much overhead. Each week from February 17th (early in the outbreak) to present we asked experts to predict the trajectory of the COVID-19 outbreak in the US. More than 40 experts made predictions that provided early information to public health officials. Over two months, experts made 4 predictions of the number of deaths in the US by the end of 2020 and gave median predictions between 150,000 to 250,000 deaths. Corresponding 80% prediction intervals of the number of deaths were generated and the minimum lower bound was 19.0K and maximum upper bound was 1.2M. Experts also estimate that, as of Apr 27th, there are over 14.5 million total SARS-CoV-2 cases present in the US with an 80% prediction interval of 4.8M to 28.1M. US national-level hospitalizations, experts predict, will most likely peak between April and June. Expert’s estimates agree with other computational model forecasts, and expert’s 80% prediction intervals covered the true number of confirmed cases in the US on nine out of ten evaluable predictions. An expert consensus has little overhead and can be deployed early in an outbreak to provide decision makers insight into an evolving disease. After computational models are built, experts can examine a broad array of questions to complement computational forecasts during a global catastrophe such as the novel coronavirus.