Weekend COVID-19 Briefing

Top news, reports and insights for today:

  1. Curated headline summaries for Wednesday:
  1. Top story: U.S. passes 13 million lab-confirmed cases after a million cases tallied each week in November (NBC News).
  2. A new model from CDC researchers, just accepted for publication, suggests the total number of COVID-19 infections in the U.S. may be 8 times higher than data shows, now reaching 100 million after accounting for the difficulty of counting all cases. Still, the report estimates that “approximately 84% of the U.S. population has not yet been infected and thus most of the country remains at risk” (NPR).
  3. Thanksgiving holiday is skewing the data due to backlogs in testing and reporting that makes it look like cases and deaths have peaked. There is strong evidence that this is an artifact of the holiday impacting the data and not of the pandemic slowing (The Atlantic).
  4. The U.S. Food and Drug Administration has approved the first at-home kit that tests for COVID-19. The device, called the Lucira COVID-19 All-In-One Test Kit, provides results in under 30 minutes by testing a self-collected nasal swab sample, and then a light on the display indicates whether the result is positive or negative (Lonely Planet).
  5. Researchers say COVID-19 immunity can last 6-8 months after infection. While this may be the best answer to this important question we have, the new study will first have to be carefully reviewed and published (Science Alert).
  6. New national poll says 27% of Americans say that the coronavirus pandemic is the biggest issue facing President-elect Joe Biden over the next four years. That’s the top percentage for any issue and shows that Republican claims that COVID-19 would fall off America’s radar screens after the election were incorrect (CNN).
  7. U.S. coronavirus hospitalizations have doubled in the last month, reaching a record high over 91,000 on Saturday. While case and death numbers are now unreliable due to Thanksgiving, the continuing rise in hospitalizations offers a truer picture of where the epidemic is going (Daily Mail, See Figure A).
Figure A from The COVID Tracking Project.
  1. U.S. jumps from 11 to 12 million lab-confirmed COVID-19 cases in record time (six days). Cases and deaths appear to slow due to holiday data disruption
     The rate of new cases continues to be greater than any of the previous peak periods. This in part comes from rising test numbers. Still, as headline 2 indicates, the numbers we see are generally only a rough estimate of the true number of cases, now estimated to be 8 times higher than the official count. The data shows the U.S. added the 12th million cases in just 6 days (See Figure B). In some data aggregators (see headline 1) we have already passed 13 million as I type. Looking at the daily numbers last week, it appears as though cases (Figure C) and deaths (Figure D) may be have reached a new peak. Both 7-day moving average curves appear to have inflected. However, as headline 3 makes clear, the reporting of tests, cases and deaths has turned sluggish due to the weekend and holiday slow downs in data processing. Staffing and supply shortages, heavy data entry backlogs, vacation days, staff sick leave, and Thanksgiving itself have all conspired to paint a picture of a peak that I strongly suspect is not actually there. The primary evidence for that assertion comes from the hospitalization data (Figure A) that shows no sign of the slowing despite the holiday. When other surveillance tools suffer, as they do during a national holiday, smart disease detectives trust the data stream that is most resistant to holiday lapses (hospitalizations).
     Instead of staring at the daily numbers, I spent the afternoon looking more big picture at the overall COVID-19 attack rates (or incidence rates) cumulatively by state. Remember when the incidence rates (lab-confirmed cases per 100,000 population) were 10-times higher in New York compared to any other state and it looked like COVID-19 was going to be a Northeast problem? Have a look at the rates now in Figure E. Now, the Northeast has the lowest overall rate of the 4 U.S. regions. Astonishingly, there are now 34 states with higher attack rates than New York, and four of those have rates that are more than twice as high: Iowa (7,171), North Dakota (10,227), South Dakota (8,344) and Wisconsin (6,541). In fact, all 13 Midwest states, 9 of 13 Southern states and 10 of 13 Western states are all higher than New York.
    Bottom line: The U.S. is essentially flying blind right now due to disruption in the timely and complete flow of information caused by the Thanksgiving break. Don’t trust the apparent peak in cases and deaths. Thanksgiving not only knocked our surveillance off kilter, but also, like Memorial day before, has sewn the seeds of still greater acceleration of transmission intensity that won’t become apparent until, well ehm, Christmas.
Figure B
Figure C.
Figure D.
Figure E.
  1. Original analysis: Did any of our prediction models accurately project where we are now a month ago?
     Does anybody remember back in the spring when we were arguing about which statistical forecasting model we should believe? It seems like most experts and commentators have stopped even paying attention. Not me. It’s a classic case of throwing out the baby with the bathwater. Sure, our projection models weren’t very good in March, but we had very little data to work with and a great many basic questions that were unanswered. The models should have gotten better since then. For a while they were multiplying like bunnies. Since the summer, new models have slowed and a bunch have disappeared. Survival of the fittest perhaps?
    A month ago, I decided to do my own “experiment” by selecting a handful of models that seemed to have been working well this Fall and evaluate them. The idea is simple. Four weeks ago, I graphed the predictions of candidate models for both weekly cases and deaths using the COVID-19 Forecast Hub website created by the Reich Lab at UMASS Amherst. I then took the predictions made by the various models that I had “photographed” a month ago, and overlaid what actually happened as of today (Figure F. for incident cases and Figure G. for incident deaths). Each figure has a purple arrow pointing to where we are now, but the graph itself was made at the end of Week 44 on October 29. The top three predictions are labeled #1 to #3 for both cases and deaths. The last two figures present what those models say will happen four weeks from now (Figure H. for cases, Figure I. for deaths). Below are my summary conclusions based on this analysis. Feel free to comment or send me email.
    1. None of these models is a perfect crystal ball (that we would not expect). Although, the 4-week projections are better than they were in the summer.
    2. For both cases and deaths, none of the 17 models tested predicted how bad things got; a month ago, every model under-predicted weekly deaths (-2,700 on average) and cases (-569,000 on average), in many cases drastically.
    3. There was a clear winner for cases and deaths, but it wasn’t the same model (the Discrete Dynamical Systems, Negative binomial dynamical systems model from Rahi Kalantari at University of Texas got, by far closest for cases; for deaths Dean Karlen’s model from the University of Victoria got the closest).
    4. Most of the models compared were crap.
    5. Relatedly, the so-called “ensemble model” that basically “averages” over all these models (Bright pink COVIDhub ensemble in Figures F and G) didn’t perform in the top 5 simply because so many of these models are complete rubbish.
    6. If you believe the 2 best models, 4 weeks from now, we will triple cases from 1.3 million to 3.3 million (DDS-NBDS) while deaths will double from 10,000 to over 22,000. It’s worth noting that just because these two models outperformed the others last month, there is no guarantee they will do so next month.
    7. Taken together, the models that made the most extreme projections were closest to being right, while the average projections came in too low. If that trend continues, the next month will be worse than common sense and most models suggest.
Figure F
Figure G
Figure H. Predicting 4 weeks from now: weekly cases
Figure I. Predicting 4 weeks from now: weekly deaths
  1. Quirky Qorner: Lock down is getting old for one comedian
     I found a funny video on CBS news by comedian Jim Gaffigan who says his Thanksgiving spent, like the rest of the year, with just his wife and kids, would, if set to music be Hotel California by the Eagles. If you are like me, it does feel that you can check out but never leave?

One thought on “Weekend COVID-19 Briefing

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s