Top news, reports and insights for today:
- Despite state plans to re-open, U.S. deaths and cases remain stable with a longer term downward trend. Five states match or set record high deaths
On Monday, U.S. deaths among lab-confirmed COVID-19 patients rose 5% to 37,446 and new lab-confirmed cases rose 3% to over 776,000. Absolute numbers of deaths and new cases were lower than Sunday, but the differences were within the range of our confidence. We do see a trend toward declining cases since April 16 so our confidence is growing that the epidemic is slowing. The decline in deaths began April 14-15 and has been more “choppy”. As always, the big picture can mask important trends occurring in states and regions. Five states matched or exceeded daily deaths. This included one hard-hit state (Connecticut) where 204 deaths were reported. The remaining 4 states have comparably low numbers of total fatalities, but deserve our attention as potential future hot spots. This includes West Virginia, Arkansas, Nebraska and North Dakota. Three-day growth in deaths has been strongest in Pennsylvania (+59%), West Virginia (+61%), and North Dakota (+42%).
What it means? We know that our case and death numbers are prone to both random and systematic errors. Despite that, we get caught up at looking at the numbers each day and forgetting this fact. We are seeing absolute declines in deaths and cases over the last week. This is good news, but we must be careful to recognize that these numbers are estimates based on imperfect data. Longer term trends are more meaningful than day-to-day comparisons. I would hope to see at least a few more days of declines before concluding that we have reached peak.
- The data are lemons. How are we suppose to use them?
Two former student from Johns Hopkins sent me messages through social media over the past few days. Both are now working in public health and both had good questions. I thought I would report their questions and my answers in this space. The conversations have been edited for clarity and to protect their identities.
Question: Hi Dr. Glass, thanks for your postings! I am on call for [an urban health department] approving testing, and taking the data we receive and entering into our […] system. The data entry is inconsistent and the quality is far less valid. Question: what data source are you using to generate these predictions? Thank you for your instruction, and for [your willingness to address] the controversial questions.
Answer: No data set is perfect of course. For consistency reasons, I have been using the data compiled on Wikipedia every day. They get it from state health departments. At least the errors are consistent. By the way, I am not really making predictions that I know of; I’m trying to stick to showing the best data I can find while acknowledging the limitations. As far as your data go, data entry is always inconsistent. The data quality is “far less valid” than what? We have been spoiled by data that is already clean, dependable and orderly. Now, it’s pandemonium and the data we have sucks. We aren’t in Kansas any more Dorothy! But it’s an epidemic, this is our job! Time to buckle up and get to work with what we have. Some basic principals: 1) Squint when you look at the data. In other words, adjust your expectations. Don’t pretend the data is crisp and clean. When we squint at our data, we remind ourselves not to pretend the data reveals truth. 2) Harness the power of consistency. The GDP doesn’t measure the US economy very well at all. Fine, but looking at trends in GDP is much better than having a single snapshot measure that is better than GDP. 3) Don’t pretend your numbers are precise. Report ranges when ever you can based on confidence intervals. Pay as much attention to the standard errors as the numbers themselves. 4) Be very careful with the language and labels. Don’t say “COVID-19 cases” when you mean “Lab-confirmed COVID-19 cases” because you have a strong prior that your ascertainment rate is tiny. Don’t say “Coronavirus deaths” when you mean “Deaths among lab-confirmed COVID-19 cases” because you know that people who die before their test results come back aren’t being counted. Good luck. Remember, squint!
– exchange on social media, April 20
Question: Hi Tom. […] I have a question and would like to hear your thoughts… I’ve seen the modeling about the number of cases/fatalities we can expect in the coming weeks and months, and most of them don’t make sense to me. I remember from my MPH that long-term epi curves tend to be calculable with simple math. So this is what I get: So far, about 1 out of 1,000 people in the US have a positive test for covid-19. Assuming we are missing 80% of the cases, and assuming the original estimates that most (50-70%) will be infected over the next year, I don’t understand estimates that 100-200k deaths would be expected. Given the above assumptions, we would expect about 0.5% of the population has been infected to date and that a year from now at least 200X more will be. That puts case/death estimates much higher than current projections. Am I missing something? Are we expecting ‘physical separation’ to work that well? Or maybe I’m estimating figures on a longer time-scale.
Answer: You are right to be confused about this. If we had two numbers, we could forecast deaths over an 18-month timeframe with considerable, if not frightening accuracy. We would need solid estimates of the case fatality rate (CFR) and the attack rate (AR), otherwise known as the prevalence of infection in the population. Problem is, we don’t have anything close to good estimates of either yet. I talk about this in my briefing on Saturday. The Santa Clara County study is our first attempt, but that’s far from a perfect estimate (as the recent gnashing of teeth on twitter has shown). It’s critical we welcome each new study, recognize its limitations and toss it on the pile of accumulating wisdom. No one study will give us the “real” CFR or AR. By looking at the pattern of results across studies (including the German Heinsberg study, the USC study in Los Angeles, and other ongoing studies), we will start to hone in on a number we can believe. Based on data from Spain, that country has a mortality ascertainment rate for COVID-19 deaths of roughly half. WORLDOMETER says there have been 21,000 deaths in Spain, but there have probably been about twice that number, about 40K. We don’t have the data to estimate the ascertainment rate in the US, but let’s say its also about 50%. It’s probably at least that bad. So instead of 45,000 US COVID-19 deaths, we now stand at something closer to 90,000. Will the ascertainment rate go up or down going forward? Probably both. Pennsylvania [recently] started reporting “probable” deaths. Ascertainment rate goes up. But lots of politicians are going to be under fire for their state’s death counts, so some states will try to bury the numbers as China has been doing. Ascertainment rate goes down.
– Social media exchange with a former student on April 20
The point is, I don’t trust any mortality forecasts out there now and you shouldn’t either. Until we have credible estimates of CFR and AR, it will just be G-I-G-O. As to your question, the mortality projections you speak of come from Chris Murray and the U. of Washington group. Those estimates have been changing rapidly in recent days. It’s fair to say that their models can fairly be criticized for using parameter estimates that weren’t yet solid. And they do generally assume that social distancing is powerful. The bottom line is this: the seroprevalence estimates are showing three fairly consistent things: 1) the attack rate is way bigger than our current testing captures (e.g., ascertainment rate is low), 2) the CFR is probably much closer to seasonal flu than would be expected by relying on crude death ratios, and 3) Herd immunity is a long way off. When we get better numbers for CFR and AR, just keep in mind that we are already 40,000+ deaths closer to whatever forecast you can live with than current data suggests. So, let’s add the ascertainment rate as a 3rd crucial number we need and don’t yet have. Best, Thomas
Thomas A. Glass, Ph.D. 