As both Tobias Hoonhout and Andy McCarthy pointed out yesterday, the University of Washington’s Institute for Health Metrics and Evaluation recently downgraded its predictions of how many coronavirus deaths the U.S. will see and how much hospital capacity the pandemic will take up. I just wanted to jump in with a cursory explanation of what’s going on.
The IHME model is very different from the Imperial College model that drew a lot of attention a few weeks back. The Imperial College model simulated how a virus spreads through a population based on a host of assumptions about how easily the disease jumps from one person to another and how often people come into contact with each other under various social-distancing rules. The IHME model, by contrast, simply looks at what has actually happened elsewhere in the world during this outbreak and uses that information to predict what will happen in the U.S. and individual states.
Yet the IHME model is still very much a work in progress — as the rest of the world is constantly churning out new data to add and the researchers are finding ways to improve the structure of the model itself. So, about those downgrades.
When the group released its new predictions, it also put out a lengthy explanation of what had changed, including this quite salient bit regarding deaths:
• At the time of our first release on March 26, the only location where the number of daily deaths had already peaked was Wuhan City. These data from Wuhan formed the basis of our estimation of the time from implementation of social distancing policies to the peak day of deaths. Since then, an additional seven locations in Italy and Spain with large coronavirus epidemics appear to have reached the peak number of daily deaths (see below): two in Spain (Castile-La Mancha and Madrid), and then five in Italy (Emilia-Romagna, Liguria, Lombardy, Piedmont, and Tuscany).
• With today’s update, we now estimate the time from implementation of social distancing policies to the peak of daily deaths using all eight locations where the number of daily deaths appears to be peaking or to have peaked. The time from implementation of social distancing to the peak of the epidemic in the Italy and Spain location is shorter than what was observed in Wuhan. As a result, in several states in the US, today’s updates show an earlier predicted date of peak daily deaths, even though at the national level the change is not very pronounced.
And this section on hospital capacity:
First, we have been able to include more up-to-date data for estimating ratios of hospital admission to deaths. These ratios inform model parameters that are used to predict need for hospital beds, ICU beds, and ventilators. In previous releases of our estimates, our ratios were informed by a CDC report with information on early COVID-19 cases in the US – from February 12 to March 16 – and those patients’ outcomes. Based on these data (509 admissions divided by 46 deaths), our overall ratio was 11.1 hospital admissions per COVID-19 death.
Over the last few days, we have been able to incorporate data sources, including data provided by state governments, on a substantially larger sample: 16,352 hospital admissions and 2,908 deaths related to COVID-19. This allowed us to estimate state-specific ratios where data were available on at least 50 deaths from COVID-19, using random-effects meta-analysis.
Our estimates released today use the state-specific ratios noted below and for those states without data, the pooled ratio of 7.1 hospitalizations per death (95% CI 4.0 to 12.7). These lower ratios of admissions to deaths result in predicted peak hospital resource use – total beds, ICU beds, and invasive ventilators – that is lower than previously estimated.
There was a ton we didn’t know at the beginning of this process, and the gaps had to be filled with sketchy data and guesswork. As new information comes in, good data can replace bad, and the predictions improve.
That means the original model might turn out to be off by a huge number. But it doesn’t necessarily mean there was a radically better way to go about this. You have to make do with the data you have, and sometimes those data ain’t great. There’s still oodles of uncertainty here, as illustrated by the huge gray bands around IHME’s estimates — meaning that, no matter what, we’re taking a risk of under- or overreacting. That’s life when a brand-new, lethal virus emerges from the animal kingdom to prey on humanity.
If you want to dive deeper into the guts of the model, see this academic paper and (if you’re into computer code) this GitHub page.