How mortality predictions in a leading coronavirus model dropped over time

A body is removed from a refrigeration truck that is serving as a temporary morgue at the Brooklyn Hospital Center. (Bryan R. Smith/AFP/Getty Images)

By now, there were supposed to have been more than 1,000 people dead in Alabama.

Well, to be fair, there were supposed to have been between 607 and 1,667 people in the state who had died of covid-19, the disease caused by the novel coronavirus. Somewhere in that range, according to a model of the virus’s spread created earlier this month by the Institute for Health Metrics and Evaluation at the University of Washington.

Happily, that estimate was wrong. As of Monday, according to data from Johns Hopkins University, only about 100 Alabamians have died of the disease.

Modeling, as White House coronavirus task force member Anthony S. Fauci would remind us, is a forward-looking estimate. It is an educated guess, based on the available data, used to make decisions about resource allocation and strategies. As data come in, the model should and will change.

As the IHME model did. Its initial estimates of how the virus would affect the United States were based on the effects of social distancing measures in Wuhan, the city in China where the virus first emerged. When more data became available from Italy and Spain, that was included in the model and the estimated death toll nationally was reduced. More data from the United States pushed the estimates lower still last week. In an update this week, they were revised back upward a bit, with an estimated 69,000 deaths nationally by Aug. 1 — assuming social distancing measures are maintained through May and, well, that new data don’t point somewhere else.

Alabama was one of the model’s most problematic estimates. Comparing the highest number of deaths predicted in four iterations of the model released this month (considering the mean predicted figure, not the upper end of the model’s estimate range), states are on average at about half the number of coronavirus deaths as the model suggested.

Nationally, the estimate was a bit better. The April 6 iteration predicted 31,000 deaths nationally by April 13 vs. the 23,500 actually included in the toll.

But Alabama isn’t alone. The interactive below shows the mean predicted death toll in four iterations of the IHME model, comparing them to what was actually recorded. Clicking a state will show a large version of the graph. Lines are relative to the peak predicted death toll by July 1, shown at upper right.

Your browser cannot display this graph.

Some states, like Hawaii and North Dakota, have stubbornly — and happily — seen low death tolls, despite what the model predicted. Other states, like Washington and Missouri, have tracked with IHME’s predictions.

To a large extent, this is how the model is meant to work. It uses available data to predict what might happen and then revises its estimates as more data become available. That the estimates for the United States broadly and states like Alabama specifically were revised so significantly, though, has led to understandable criticism.

Speaking at a White House briefing last week, Fauci defended the process of modeling the outbreak.

“There is nothing in the literature about specifically what to expect,” he said, “and I think that’s why the models continue to modify themselves based on what actually has happened with social distancing and hand washing and all of the pieces that the American people are doing."

“So the models continue to evolve based on the impact,” he added a bit later.

Hopefully, the general trend will continue to be downward.