A statistician friend (Dana Farber Institute testing) tells me CoVid19-Projections has been much more accurate than IMHE, and yesterday they put up their state-by-state projections from May to illustrate their accuracy.  It holds with what we have seen pretty well, and I like people who are openly willing to be graded in order to get things right.

I will be vacationing at a lakeside cottage until Friday, and so will not be commenting on all of your intelligent musings.

15 thoughts on “CoVid19-Projections”

  1. It’s an impressive piece of work; looks like it was done by one guy on his own, rather than by some well-funded organization.

    All these models need to be taken with truckloads of salt, of course: there are a *lot* of assumptions about things like incubation time, etc, and it’s not possible for a machine learning system to adjust for a whole lot of parameters over a relatively short time period. And there is also human behavior: to the extent things get notably worse in a particular area, at least some people will adjust their behavior in consequence.

    Interesting that his Twitter feed sounds less optimistic than I would have guessed from looking at the model results.

  2. Would like to know what the reasoning is behind the bounce a couple months out…

    Also, YTF the USA’s death toll is so much higher than all the others on the map. This seems improbable.

  3. OBM….USA projected total deaths (by Nov 1) is 669 per million. Europe as a whole is 346, BUT if you look at the major countries, with the exception of Germany, you get UK=732, Italy=592. Spain=618.

    I would guess that the bounce a couple of months out is due to the machine learning trying to adjust transmission rates in light of what is currently happening. I don’t know whether it adjusts case fatality rates or not, probably can be determine from his description of the algorithm. (or, for the truly ambitious, from the source code)

  4. I notice that they headline the number of deaths per day which is the number we care about most. Unfortunately for the validity of the model, the most important number is the number currently infected. If the model is to be valid, this is the number that will drive all the other categories. This is labeled as an estimate so we are back in the weeds of conjecture or less politely, guess.

    They mention that they are assuming that there are 5 times as many infected as reported. Why 5? I could guess 10 or 2 and there really isn’t much way to tell which is closer to right or any compelling reason that I can see to not use the actual numbers reported. This is a major assumption that seems to be glossed over.

    If you look at the application of the model to New York and Texas, the differences seem to be driven by the estimated number of already infected persons. This gives the conclusion that New York is looking at an additional 1,300 deaths while Texas still has 11,000 to endure. They put the total infected in New York at 18.7% and Texas at 5.3%. Yet slightly more people live outside New York City than in and the course of the epidemic outside the city has been closer to Texas than New York City. If their model is correct, it looks like the rest of the state has some suffering to do until they catch up with the city.

    You could take the data reported and your choice of spreadsheet and derive a polynomial to fit the data to date and simply straight line extrapolate using the last few weeks and arrive at an answer not much different what they have. They claim to have used machine learning without specifying what data they were learning from. If they were using subsets of the data which is a common technique, it isn’t much different than the curve fitting exercise above or from all the unsuccessful attempts to apply it to the stock market. It would, even if it worked, be useless for the next time, since when we would most need insight there is least data available.

    On a technical note: They chose to use a 7 day average for the deaths per day. This is probably exactly the wrong period to use since the data seems to show a strong periodicity of about 7 days. It probably doesn’t make a difference to their model but it does make it harder to visually interpret the data.

  5. I think this all goes back to the experiment of nature that was the cruise ship in January. 20% got infected over a month isolation. The passengers were typical cruise ship passengers, probably age 70 and fairly good health. The crew was probably similar to most cruise ship crews, Filipino and young.

    I think there were 14 deaths total, only one under 70.

    Here is another issue. Tamura said that he felt the sensitivity of PCR test was about 70% False negatives are a big factor. Still, the deaths were all in the older group and treatment at that time was still in early stages.

    My wife, who is 75 and has COPD, was very ill in June and had 5 negative Covid tests but she takes hydroxycholorquine for rheumatoid arthritis and may have had an aborted case. Her primary care internist and the hospitalist who cared for her both think so.

  6. Is there any single fact that is not in dispute? It may have been circulating for months before even the Chinese knew. Evey time we look, we seem to find evidence of earlier infections here. There was a lot of traffic between China and us. The last I heard was early December. We know that we don’t have any idea how many people have been actually infected. We know that we don’t know how many people have died from it and never will. I would bet that a lot of the hospitalizations are people that figured that was the quickest way to get tested and the hospitals are going along to try to make up revenue that they’ve lost. Notice how fast they are to correct any suggestion that they are running out of ICU beds which would close their operating rooms again.

    All of this changes the goal of any model from representing the real world to trying to match whatever numbers, arrived at by whatever criteria fits the desired narrative this week. I’d bet that New York went from reporting every death they possibly could as wuflu to reporting only those that they can’t avoid when it started to reflect badly on their pols.

    I think the real lesson is that next time is likely to be just as disorganized and confused. By the time the signal works its way through the different layers, it will be too late to control. We were lucky with SARS and MERS, they turned out to be self limited. That won’t prevent various people and organizations from offering to protect us in exchange for large sums of money. What are the odds that China or us, for that matter, would lock down a major potion of their country after a half dozen cases of something that looks like the flu but isn’t. I think investing in vaccine technology to shorten the interval between the appearance of a new bug and an effective vaccine is more likely to pay off.

    Even with the deliberate Chinese delay, the time line of this moved pretty quickly and it still wasn’t fast enough to catch it.

    We were told that this would continue until about as many people as were susceptible got it. I’m hoping that the number is closer to 20% than 80%, but it looks like they were right.

  7. I have more Neanderthal genes than most so I have been taking hydroxychloroquine since June when my wife was sick. I am suspicious of all the antagonism to HCQ. Remdesivir, which seems to be effective in severe cases, is $3500 per dose. HCQ is about $6. Hmmmm.

  8. I agree, the model is factoring in immunity from prior infection and about 20% of the population infected the growth pretty much stops. That would be consistent with the experience here in Connecticut. Areas with high death rates were usually in metro locations while close by suburbs had few or none. There was no real indication that lockdowns impacted the urban growth rates at all so it really seems that they got to some level of immunity then deaths declined rapidly. The suburbs and rural areas had almost no trend with time. With no local mass transit to speak of (no subways etc.) the two groups (urban/suburban) did not mix much. Once a reasonable fraction of city people had acquired immunity the walls were plugged. It may be that the urban areas had much higher infection rates (50%+) than the suburbs so the ensemble is about 20%.

    This would imply that vaccination should start in urban areas if there was no serious outbreak (or similar dense housing in factory towns and farm worker dorms.)

  9. There’s a big question as to the actual infection rate. It took some reporters in Florida to ask how different labs and hospitals were returning high 90% or even a perfect 100% of positive tests. When the “experts” looked, the proper number should have been 4-9%. These people can’t do math, they can’t even count.

    It will have to be a very good vaccine in order to give it to billions of people without killing or incapacitating more people that the disease itself. So far, only a few million have been infected. Organizing trials large enough to catch fairly rare side effects that multiplied by billions would make a disaster will take a good deal of time.

    Some of us remember the Swine Flu vaccine debacle. It wasn’t even that bad, with very few people that didn’t recover in time and not that many bad reactions to start with. According to this, about 1/100,000 cases of Guillain-Barre syndrome, about 450 total.

  10. MCS: “… hospitals were returning high 90% or even a perfect 100% of positive tests. When the “experts” looked, the proper number should have been 4-9%.”

    That is really easy to explain. The hospitals were using the same software developed to count Democrat votes.

  11. It’s a good thing cynicism isn’t considered communicable or we’d all have to go around with blindfolds, gags and ear plugs to flatten the curve.

  12. Mike K Says:
    I have more Neanderthal genes than most so I have been taking hydroxychloroquine since June when my wife was sick. I am suspicious of all the antagonism to HCQ. Remdesivir, which seems to be effective in severe cases, is $3500 per dose. HCQ is about $6. Hmmmm.

    FWIW, Doctor Pat Santy (moribund blogger “Dr. Sanity”) had this to say on FB 10 weeks back (early May):

    Pat Santy
    It’s also extremely inexpensive. I have prescribed it for my Lupus patients in the past and it’s been around for a long time so most of its side effects are well known. OTOH the drug Remdesivir is very expensive, but also shows some promise in the treatment of COVID-19.

    There’s also this piece:

    Yes, it’s the Post, but still. It’s citing doctors, not expressing a position.

Comments are closed.