SOME 80 DAYS have now passed since the first death from covid-19 occurred on America’s shores. Since then over 90,000 people in the country have succumbed to the virus. That toll is greater than America’s combat deaths in Korea, Vietnam, Afghanistan and Iraq combined. Whereas governments do not release their forecasts of how many people will be killed in wars, predictions of covid-19 deaths in America have been published and are widely scrutinised.
As in the fog of war, early epidemiological projections have been subject to the largest errors. One of the first prominent institutions to issue a long-run forecast for covid-19 was the Institute for Health Metrics and Evaluation (IHME) at the University of Washington. On March 26th it predicted that a total of 81,000 people were likely to die from the epidemic in America. Though they were uncertain about the precise toll—estimates ranged from 38,000 to 162,000—their model was confident that the virus would be done by July. The IHME’s projections fluctuated wildly afterwards. A forecast issued in mid-April predicted just 60,000 deaths.
Despite the misgivings of many experts, the IHME’s model was widely cited. Donald Trump’s administration used the forecasts as part of its evidence of when best to ease lockdowns. Meanwhile, the IHME’s method raised concerns among fellow health experts because it used an unorthodox “curve-fitting” approach rather than a conventional SEIR method. SEIR has been the basis for most epidemiological models for 100 years, and estimates the share of a population who are susceptible (S); exposed (E); infected (I); recovered or dead (R). The IHME began incorporating results from a SEIR model in early May.
Two other long-run models have also entered the forecasting fray. The first, from the Los Alamos National Laboratory (LANL), is a SEIR-like model. It has tended to be more pessimistic than other forecasters. Youyang Gu, a young graduate from the Massachusetts Institute of Technology, built another after seeing the IHME’s early missteps. Mr Gu’s model uses machine-learning algorithms to instruct a SEIR model that has recently been more accurate than forecasts from many established outfits. Mr Gu’s model, as well as those from the IHME and LANL, is one of a dozen individual forecasts that America’s Centres for Disease Control and Prevention uses to arrive at an “ensemble” estimate of covid-19 deaths.
Although long-run forecasts will always be subject to larger errors than short-run projections, they have been getting smaller. An analysis by The Economist finds that all models succumbed to large forecast errors in their early stages. Across the three models, projections made on April 12th for the death toll two weeks ahead had an average absolute error of 17%. Those made at the beginning of May for the next two weeks had an average error of just 4%.
There are several reasons why forecasts got off to a bad start. First, the early models suffered from an absence of good historical numbers. The first models extrapolated patterns from the virus’s path through Wuhan and Italy, where the transmission rate was lower than in America. Second, the stabilisation of the death rate, which is now falling gradually, has made forecasting easier. Finally, predicting the future behaviour of individuals, such as their adherence to lockdowns, is guesswork at best. Some new models are attempting to fill this gap with real-time location information.
With better data and improved models, the forecasts made today should prove more accurate—either by luck or design—than earlier ones. Yet future estimates remain widely divergent. The IHME expects total deaths to stabilise in July at 140,000, whereas Mr Gu expects the virus to persist into August, at which time there will have been nearly 200,000 deaths. Either total would mean that covid-19 would have exacted a truly terrible toll. ■
Sources: Covid19-Projections.com; Institute for Health Metrics and Evaluation; Los Alamos National Laboratory; New York Times; Reich Lab, University of Massachusetts Amherst;The Economist
This article appeared in the Graphic detail section of the print edition under the headline “A terrible toll”