Recent Covid-19 Fatality Forecasts are Far Too Dire
With the sharp increase in positive tests being reported recently, we are starting to see some really extreme predictions about the death rates over the next few weeks–including people predicting “hundreds of thousands of deaths.” Such predictions are extreme and unfounded. This is not going to happen.
As I observed in my estimation book, “Subject matter experts tend to use simple (or even primitive) estimation strategies even when their expertise in the subject matter they’re estimating is high.” Unfortunately we’re seeing some real-time examples of that with medical professionals who are good at medicine (I hope) but who apparently do not understand estimation at all.
Estimating Deaths from Infections and the Fatality Rate
If we knew the number of infections, and we knew the infection fatality rate (IFR), it would be very simple to forecast the number of deaths from the number of infections. Throughout the pandemic, however, we have been estimating both the number of infections and the IFR, and in many cases doing that badly.
IFR is in the Neighborhood of 0.5%
I’ll explain how I arrived at this conclusion in another article. The super quick explanation is that this number can be calculated, at least approximately, from the antibody tests that New York State conducted in late April, combined with information about New York state’s fatalities from Covid-19. Once we know the approximate IFR, we can bootstrap methods for estimating the number of infections from the number of deaths, and that helps us zero in on the relationship between positive tests and infections.
Positive Tests Do Not Equal Infections
Positive tests and infections are not the same thing. Positive tests are a useful proxy for the number of infections, but they are not the same. This simple concept seems to thwart many of the estimates that are created.
The relationship between positive tests and infections has been shifting over time. Early in the pandemic it appeared that the ratio of infections to positive tests was somewhere between 10:1 and 20:1. Now the ratio appears to be somewhere in the range of 3:1 to 5:1.
The cornerstone of my estimation approaches has been tracking the changes in Naive CFR, as an indicator of the ratio of infections to positive tests. Naive CFR is calculated as today’s deaths divided by positive tests 13 days earlier. Because deaths lag positive tests, positive tests are a useful leading indicator of future deaths.
I continue to be amazed at the huge inaccuracy of many highly publicized short-term fatality estimates. If you simply took the average naive CFR for the past 7 days and used it to estimate deaths for the next 14 days, based on 13-day-lagging positive tests, you’d create more accurate estimates for deaths over the next 14 days than the vast majority of estimates that have been published.
For most of my estimates I’ve used the strong correlation between positive test percentage (positivity) and naive CFR. Because we know test positivity when test data is reported, it has been useful as a leading indicator of the death rate, and I’ve used recent positivity along with the number of positive tests to estimate near-term future death rates.
But Naive CFR has dropped prodigiously since the beginning of the pandemic. It started out over 50%, and the past few days it’s dropped as low as 1.4%. As Naive CFR has dropped, positivity has become less useful as a basis for estimating.
The main factor that has changed is that the average age of people being tested has been trending downward. The percentage of people who test positive that are under age 50 is also strongly correlated with naive CFR. The CDC reports that information, but not in a very timely way. It lags about 2 weeks.
Updated Estimation Methodology
The bottom line on the two estimation methods is that I’m not sure either is as reliable as positivity was for the past few months, so I’m reverting back to using multiple estimation models.
Method 1 – This is similar to the method I’ve been using. It is a curve-fit model that is based on (a) positive test percentage (positivity) and (b) the trend in average age being tested. Both of these factors have high RSQ values vs. Naive CFR, but as the average age trends downward, part (a) of the model becomes less accurate and part (b) becomes more accurate. The main difference from past estimates is thus that (b) is weighted much more heavily. The main input to this model is positive test cases from 13 days earlier than the date of estimated deaths.
Method 2 – This method is based on a straight trending of Naive CFR. I take the history of recent Naive CFRs, calculate a trend line, and project that forward. I then apply the estimates of future Naive CFR’s, day by day, to the positive tests from 13 days earlier.
The composite model is just the average of the two methods, fine-tuned on a day-by-day basis to account for a history of day-of-week fluctuations in reporting.
How My Estimates For the Next Two Weeks Compare to Other Estimates
There were 9,079 deaths in the most recent 14 day period.
My Method 1 estimates 12,300 deaths over the next 14 days. This is a 35% increase over the preceding 14-day period.
My Method 2 estimates 6,100 deaths over the next 14 days. This is a 33% decrease over the preceding 14-day period.
My composite total estimate is 9,200. This is a 1% increase–basically break even.
The situation is still serious. People are still getting infected in large numbers, and we will see more deaths. The death curve is essentially flat, but not at the level where we would like it to stabilize.
That said, I don’t see any possible way the death rate will increase to the extreme degree that some people are predicting, at least not based on what we’ve seen over the past two weeks in terms of new positive tests. Further out than two weeks, anything could happen. But over the next two weeks, we’re going to remain in territory that has become all-too familiar.