New Baseline Coronavirus Model

A long nine days ago, I posted about The Paper from Imperial College London that predicted the US would face 2.2 million COVID deaths if nobody changed their behavior at all. The Paper used a detailed SIR model to simulate how infectious diseases spread through schools, workplaces, homes, and communities in US cities. The headline of 2.2 million deaths was a worst-case scenario and a reference point for how many deaths we can avoid by implementing different interventions. Although The Paper simulated detailed spread across the United States, it didn’t include results at any finer level than the whole country.

There is a new paper from the University of Washington that takes a very different approach – and an additional week of coronavirus data – to forecast the course of the virus in all 50 states. Their forecasts include the demand for general hospital beds and ICU beds. They have data on the number of each kind of bed in all states, so we can see which states will likely exceed their capacity and by how much. They have visualized their results here.

The new paper doesn’t simulate the mechanics of how the virus spreads. Instead, they fit the growth curves to those seen in other countries before and after they implemented different interventions, with small adjustments for age demographics. You might immediately object that the US is different in lots of ways. In a sense, you’d be right. But this approach – assuming that outcomes for new events will be similar to outcomes for similar events in the past and then adjusting only slightly for whatever makes the new situation different – is the approach the superforecasters take. They’ve found that most people, including domain experts, exaggerate how special each situation is.

The new paper considers four interventions:

1. stay-at-home orders
2. school closures
3. non-essential business closures
4. severe travel limitations

Their simulations include when states have implemented each of these interventions. Crucially, they assume that all 50 states will implement all four interventions by April 2. That assumption is buried in the paper and I don’t see it anywhere on their visualizations page. It is almost certainly too rosy an assumption. I fear that people will look at the forecasts, think they don’t look that bad, and decide their states don’t need to do anything. The travel limitation assumption is especially rosy. Not a single state has met their definition of ‘severe travel limitations’ so it strikes me as bizarre to assume that all 50 will do it in the next week.

I’ve picked out a few of the states I’ve been following most closely. I’ll be comparing to these numbers to set expectations as new data comes out.