Coronavirus: What's Next For U.S.?

March 23, 2020: COVID-19 Cases and deaths are shooting up like a rocket ship in US and Europe right now. At the same time, the scope of the Coronavirus in the US is likely much broader than case counts are showing. I estimate about 300,000 total US infections as of today versus less than 50,000 cases confirmed. The good news is that the rate of growth is likely slower than what case counts are showing because our testing is catching up with the symptomatic cases. Further good news is that I expect us to see infections slowing. According to the model, cases will begin to slow by a meaningful degree in about two weeks.

As lockdowns and social distancing interact with the timing of the virus incubation and contagious period, expect new cases to level off in a little more than two weeks. The previous experience in China, South Korea, The Diamond Princess Cruise and Italy show lockdowns and social distancing will slow the virus. The bad news is that we have a lot more asymptomatic people that have the potential to spread the virus without even knowing it. The lockdown will keep many asymptomatic people from unwittingly spreading the virus.

Let's take a close look at Italy and then look at my model for the US. At the end of this article, I'll briefly discuss a data effort related to rebuilding the economy, and share how you can get involved.

Italy's lockdown is showing signs of success, but you wouldn't know it if you are following the headline death toll. The BBC yesterday led with this chart (see below). But as I've shown with the data I got and analyzed from China, deaths lag by an average of 21 days of exposure. Therefore, you won't see the rate of deaths decline for at least three weeks after lockdown. It has only been about two weeks. Exposure-to-symptoms-to-confirmed test lags by a little over a week, so let's look at cases to gauge the effectiveness of the lock-down.

As we can see from the chart below, yesterday's figures show a drop in the new case count. Take a look at a seven day growth rate by taking today's number divided by the number of cases seven days earlier. Average the previous seven days for a more stable figure and you'll see today's (3/22) growth rate is about half of the growth rate a week prior (3/15). And 3/15 is about half what it was a week prior. It was 118% then 60% and now 33%. That is good news. South Korea had a similar decline after lock-down -- albeit they brought the rate down much faster. Two weeks ago, they were at 38%. A week ago 2.1%. Today 1.3%.

To better spot this decrease in the rate of growth, use the log scale.

Other good news from Italy is that we have a better measurement of how many asymptomatic cases there are, and how broad measurement can stop (or at least significantly slow) the spread.

A small town in Italy tested EVERYONE. They initially found 50 to 70 percent were asymptomatic. This matches asymptomatic levels found from the Diamond Princess Cruise. I've been using these asymptomatic levels in my modeling, as you'll see in a moment. But what is more important about this story is the people that were asymptomatic were informed that they could spread the disease, so they cooperated with the quarantine. Now, Newsweek is reporting that the town has successfully stopped the spread. This shows the power of broad testing and people acting to protect others. Even though the virus had spread to 3 percent of the town, the testing and subsequent interventions succeeded in arresting the spread of the virus.

Turning to the US, about 1/3 of the country is on Lockdown. However, the US government was slow to ramp up testing. At the start of the month, the US reported only 61 cases to WHO. But, based on my model, there were a lot more infections than the 61 reported by our feeble testing system. My model estimates we were north of 15,000 infections (17,335 as of March 2, based on the model). Based on the model, US cases continued to climb at a rate more than doubling each week. Recently a combination of social distancing and governments imposed lockdowns in different parts of our country are slowing the virus spread.

This week, the model estimates there are 120,955 symptomatic infections. But, since about 60% of cases are asymptomatic (based on data from the closed population study in Italy and Diamond Princess Cruise), the model estimates 302,387 total infections as of this week. It is likely that only those symptomatic will get tested, which understates the risk. These asymptomatic infections are why the lockdown is important to slow the spread of the virus.

The model forecasts we'd begin to see growth in hospitalizations in the range of four to five thousand per week for the next two weeks, and then see a significant slowing in the growth at end of March. That said, the estimate of hospitalizations is the area of the model with the least confidence because the health of Americans and the US health system is dissimilar to South Korea and the Diamond Princess Cruise, which provides the basis to estimate the percent of infections requiring hospitalization and the percent of infections requiring critical care.

To be clear, my model is full of assumptions based on the data I've been able to gather. I'm making the model available so you can adjust the assumptions. Make a copy, change the model drivers and see what you think about where we are at in the US and what to expect over the next six weeks. For my friends in Europe, you can replicate the model for your country using the "template" tab. You will need a count of deaths and age details to replicate the approach used here.

Assumptions (all of which you can adjust):

1. Lag time & Age Adjusted IFR: Lag time from exposure to death is a median of 21 days. The IFR is based on an age adjusted data reported by CNN on the first 100 deaths in the US, about half of which reported the age and the index of death by age I calculated from China and their CCDC report. The analysis is detailed in my academic article on calculating the IFR based on the Diamond Princess Cruise.
2. Asymptomatic cases are averaged at 60% (mid point of range from Italy, and in the range of data observed in the Diamond Princess Cruise). The model assumes that someone that is asymptomatic does not get tested. Only people that are symptomatic get tested and show up in our case numbers.
3. Of those that have symptoms, the model assumes that 89% are mild and 11% result in some form of hospitalizations. This is roughly in line with the data from the Diamond Princess Cruise after adjusting for age. On the DPC, 23% were hospitalized as of March 16, but with their age distribution, they were just over 2x more likely to get hospitalized compared to the overall population. It appears that higher exposure to those with the virus results in severe outcomes. Because of the close quarters of the DPC, it may be that the hospitalization rate skew higher than it will be for the US as a whole. It should be appreciated that if exposure level is correlated with the severity of COVID-19, medical personal are at a high risk.
4. The model assumes that the "real" weekly rate of spread is increasing at a rate per week that is different pre and post lock-down. Pre is assumed to be 230% and post follows the rates from Italy and then South Korea, declining to 60% in the first week, 33% in second week and 2% for third week. This weekly growth assumption is hard to get right. If you choose to change it, be aware that weekly growth is not the same as R0. R0 has to be converted to a weekly rate to be used in this model. You can also play a "What-if" scenario and leave the growth rate at 230% or raise it and see what happens to total cases and hospitalizations.

A way to test the validity of the model is to check the how well the model accurately predicts the total number of hospitalized cases. The hospitalization number is a key consideration because hospital capacity is the constraint driving a lot of decision making. Here is a tweet from my Governor which points out the limited number of critical care beds available in the state. I haven't found a solid data source on total number of COVID-19 hospitalizations in the US. I feel like the model needs work to calibrate it against hospitalizations. I used the hospitalization rate based on the older skewing Diamond Princess Cruise (average age = 58) adjusted for by age. If you have a good source of cases to hospitalizations, please share so we can compare and adjust the model accordingly.

You can also look at the symptomatic infection estimate from the model compared to known cases and decide if it looks reasonable to you. As of today, the model estimates a total of about 120,000 symptomatic infections and there are about 33,272 confirmed cases as of yesterday -- which implies, with our approximately 250,000 tests, we have caught about 30% of symptomatic cases. Keep in mind the prospect that there are more than twice as many asymptomatic cases. Lockdown is helpful in keeping asymptomatic cases from unintentionally spreading the virus.

It is important to recognize that the cases are not evenly distributed across the US. As in China and S. Korea, infections are concentrated regionally. Therefore, to fully appreciate the capacity constraints in our medical system, the data needs to be examined locally. New York City, like Wuhan, has a disproportionate number of infections and will hit medical capacity limits sooner than other areas.

The model projects out to the end of April and you should take the forecast with caution because the hospitalization rate and critical care rates could turn out differently in the US. In fact, every one of the factors listed in the model could turn out differently in the US. My intention is to give you a tool where you can adjust assumptions and draw your own conclusions.

I don't have conviction that I've got all the assumptions right for the US in this particular model. One of the things I wanted to include in the model is a curve for hospital release, but I couldn't find great data on this. China released people relatively quickly, but Japan still has over 100 Diamond Princess Cruise patients in the hospital six weeks after their infections started. If someone has a good median days for hospitalization-to-release along with standard deviation, we could show how hospital capacity improves as patients recover and are discharged.

You are encouraged to make a copy of the model, make your changes and post a comment on what you found, your data sources and your conclusions. Tells us where you think we will land in terms of cases as of the end of April.

A few final notes:

1. I am seeing more people ask the question, "Won't we develop heard immunity so shouldn't we just let the virus run its course and not wreck the economy?" It takes somewhere in the range of 70% of the population to have had the disease and recovered from it. We are at less than one-tenth of one percent right now. Play with the assumptions of the rate of spread to crank them up for the next few weeks and you can see how devastating it would be in terms of hospitalizations. I don't think the math works to gain herd immunity without crushing our medical system and infecting and killing a meaningful portion of our medical system in the process.
2. I do think we may get some help from warmer weather based on data analyzed from China, and from a cursory analysis I did of cases by latitude. This is not in the model as it will likely have a benefit in late April and thereafter. I also think we will see a resurgence come flu season in the fall unless we do extensive testing to find asymptomatic people as was done in the Italian village. If we see a resurgence, it will start building to concerning numbers right around Christmas time. Herd Immunity will not be a viable solution. Expanding testing and implementing tracing protocols, potentially with GPS as done in China, should be considered. It would be a lot easier to contain it come fall if we are doing a lot of testing and can quarantine those that test positive and those they encountered that also test positive. In the US, we do not have a sophisticated system to test, trace and contain. We need to build it. We may need test, trace, contain system sooner if we don't get help from warmer weather. We will stop a lot instances of the virus with this lockdown and social distancing time, but not everyone is social distancing and the virus has made a lot of copies of itself. Some people may be continuing to pass it on to others.
3. The economy is a mess. I've started working on the economic implications. I am interested in how we can bring the business community together to crowd source consumer spending estimates so we can have a more complete picture of how our economy is responding to stimulus. I am seeing some great data coming from the marketing research community. My hope is this collaborative effort will help business and government make better decisions for a faster recovery. If you want to stay tuned or help with this effort, follow me on Twitter and send me a message @rexbriggs. I'll share a data effort related to rebuilding the economy, and you can decide if you can help.