I retract the main points of my blog post on covid-19 · 6:26am Mar 30th, 2020
Thanks to everyone who criticized my post on covid-19. I still think the analysis I made was well-justified, but it certainly wasn't clear, and didn't adequately deal with death rates.
I was rewriting the article to be more balanced and less conclusive when I was flummoxed by something no one pointed out, which should have been obvious from the beginning: The number of positive cases and deaths is already greater than predicted by the models.
The CDC thinks that the first 2 people in the US were infected 76 days ago, on 1/14. That corresponds to having a single case on about 1/11, hence 79 days of virus since US case zero.
If you go to the Epidemic Calculator and set the population to about 330 million and the time from incubation until death to 14 days, and then look at day 78, you see it predicts that we would by now have 124 total deaths out of 14,436 infections. We're way past those numbers, so either the Epidemic Calculator is buggy, or the parameters I've been questioning are too conservative.
The doubling times scientists have reported all would indicate a much, much faster growth, if they could be taken as doubling times--but I already pointed out that doubling times are a poor way of modeling a virus, so I shouldn't have modeled it that way in my own head. Shame on me.
It's still wrong to look at graphs of detected infections and think they show the rate at which the virus is spreading. It's still even true that the model "the virus is currently at equilibrium, and so we always detect the same fraction of infected cases no matter how many tests we put out" predicts the US 3/3-3/26 positives and deaths much, much better than any other model I'm aware of. R ~ .98, guys. I still don't understand why model parameters are based so much on studying outbreaks. And I still don't think anybody has come up with a plausible reason why the fraction of people tested who test positive remains almost constant.
But those points are moot. The models--at least, the one used by the Epidemic Calculator--are not overestimating the rate of spreading.
Sorry, everypony.
Is it possible that we don't have a neat exponential growth rate because of various forms of controls being implemented at different times? So the rate of growth has repeatedly changed rather than being anything consistent?
5231614 Probably, but that wouldn't rescue my idea that the growth rate was already leveling off. It would mean we'd had even higher-than-higher-than-expected growth rates in Jan. or February.
We did not only start with one case. International travel has shuffled so many people around that we may have had dozens or up to a hundred cases imported from China and other countries. Darned reality always buggers up perfectly good models.
I'll admit that I've not read the original post, but IIRC hasn't one of the issues with this pandemic been that testing kits for the virus are in short supply? Supply has been ramped up, but wouldn't that mean that the rate of tests being done is limited by the number of testing kits. I know that they were limited enough for both February and March that you were not to be tested for it even if you showed all the symptoms unless you could confirm contact with another exposed, infected individual, something which angered a lot of people but with tests so scarce, they had to make certain they were narrowing things down.
A limit to the number of tests available, increasing at a set rate, would probably account for a little of that, because even if one showed all the symptoms, they wouldn't have been tested without contact with someone already tested positive. That'd probably wreck the numbers a little.
No worries, my thought was primarily China’s numbers were completely off, but that the ratio of cases/deaths were accurate. Like the video I sent commented previously, ONE morgue in China handed out ~6500 urns in Wuhan. That’s twice the number of admitted deaths for the country in one facility. The numbers in Japan have also shot up as soon as the Olympics were postponed, and Iran is unwilling and unable to test properly. From those three countries alone downplaying this (not to mention all the illness in various militaries/governments, all of the projections are
) As some estimates put the number of asymptomatic at 50% of all cases, you need to roughly double the number of cases to get closer to the real number
Germany believes ~70% of the country will be infected, the U.K. said you will know at least one person that dies. California put an estimate of 56% infected, and the ~lancet~ (meh) university thought there were 100,000 Chinese infections back in January. The numbers are so off from lack of testing and extremely difficulty in qualifying for a test among a variety of things that the reality may be much worse than any are predicting.
This thing is aerolized, waterborne, bloodborne, can survive for 17 days on nonliving surfaces (Diamond Princess ship), can be asymptomatically spread for at maximum 27 days (admittedly an outlier) , and seems to have dormancy traits similar to cancers and autoimmune disorders. Plus (I know I’m ranting sry) many of the asymptomatics are at extremely high risk of heart attack and organ failure to the point that people have gone from perfectly fine asymptomatic to dead in as little as two hours.
Most of this is so unexpected for a Coronavirus that scientists worldwide refuse to believe the data until there is no other explanation for what’s happening
(Putting away soapbox)
5231620
Honestly, I'm willing to bet we did have more cases in the US than anyone currently knows about back in January and February. We weren't looking as hard for it back then.
Actually, my model looks like it's going to be pretty spot on: https://www.fimfiction.net/blog/894111/a-prediction-about-covid-19-why-im-scared-and-why-you-should-be-too
If I had to guess, I'd say this is a poor way to model the early stages of the virus's growth. During the early weeks of this year we likely had dozens or hundreds of infected people arrive from overseas -- people from China but who traveled through different countries, and thus evaded the travel ban. Although organic growth of the virus may have been very limited in the US at that time, which is what a model would predict, that slow growth was vastly overshadowed by the number of new cases just walking off airplanes.
The best metric I've found so far to model the number of cases is the number of deaths, because while we're certainly undercounting the number of cases, we're probably catching most of the deaths. The problem with this method is that it only tells you how many cases there were 18 days ago -- the mean time from initial infection to death. For example:
Yesterday, Michigan reported 800 new cases and 21 deaths. Right away those numbers should look incongruous, if you use the generally accepted fatality rate of 1%. Twenty-one deaths implies 2100 infections. Except, for those 21 people to have died today, they were on average infected (along with 2079 other people) 18 days ago. If Michigan only reported 800 new cases yesterday, but we can be reasonably sure they had 2100 new cases 18 days ago, it stands to reason that they are vastly undercounting the number of cases. Depending on how fast you believe the rate of infection is doubling, I would estimate that Michigan actually had between 16,000 and 134,000 new cases yesterday, with a much rougher estimate of 32,000 as the most likely number.
I believe the social distancing policies and closures we've put in place are slowly the virus dramatically, but I don't think we'll know how much until mid-April or so.
Testing is still limited, and deaths aren't getting counted on the federal level all that quickly, either. So like 5231669 said, the number of cases is significantly higher than the published figures reflect.
You're very smart. Also you're not afraid to admit when you change your mind. Those are great things.
There were two points I was going to make, but didn't want to pile on. But I will make them now.
First point: you may have been falling into the fallacy known as the fallacy fallacy. You identified a specific issue that may have altered the data, but identifying a problem with an argument doesn't mean the argument's conclusions are wrong, it just means that particular argument is flawed. There are many lines of evidence that suggest the growth is exponential that were not covered by your observation, not least of which is the fact that the growth has to be some sort of exponential (initially, at least; overall it's modeled better by a logistic curve with some tweaks to it) because that's how it works in theory because it spreads from the newly infected.
Second point: there are a lot of super-smart people who study this for a living who have been looking at it extensively, and I trust the way they characterize it because there is very little disagreement (among actual experts, not Republican pundits).
Thank God. For a minute there you were beginning to think like a Frenchman ("Yes it works in practice but does it work in theory?")
I’d like to know how this Epidemic calculator figures the numbers. Flu type viruses are notorious for being highly contagious, but with lower mortality. Of course, by merit of higher infection numbers, even a 2% mortality would roughly equal the current number of deaths to infect ratio.
Whatever that model is using must be under 2% though, less than 1% at least. Makes me want to see what logic went into the algorithm even more.
Also, remember the H1N1 swine flu panic? Those infection numbers reached the millions in the US, and is quite enlightening to how infectious flus are.
5231635
Yes, one of the problems with the current infected count is simply the lack of available testing; leading to a lower count reported than actual cases on the ground. And that's likely to stay a problem for quite some time.
It doesn't help that there appears to be a huge (by epidemiological standards, at least) percentage of Covid-19 infections are asymptomatic, or mildly symptomatic, but which are still highly contagious. It's not clear just what percentage fit this category, but it's definitely enough to complicate gathering accurate statistics.
5231644
While I can't comment on the majority of what you've said, the novel coronavirus most certainly cannot survive for 17 days on surfaces. Like all viruses, COVID-19 is not a living thing and cannot survive for long outside a host. It only survived three days in pristine laboratory conditions. What was found on that ship were traces, not viable samples.
5231644
Recall, though, that the BI et al. article I cited looked at data that had been gathered by tracing covid-19 cases forward, going to the households of people who'd had it and testing all the family members, and they found that the probability of getting the disease even if you were living in the same house or apartment as someone with the disease was only .112-.158. And the infection rate on the Diamond Princess was only about 17%. These numbers suggests that most people can resist covid-19.
No; that's in the Rocklöv et al. paper I cited, and it said that RNA from the virus was detected on surfaces after 17 days. That's to be expected; RNA doesn't just disappear when the virus dies. The virus dies when the RNA breaks into pieces. These people were certainly not trying to detect whole viruses; that would be absurdly expensive. They were undoubtedly taking samples and then running PCR on them against small fragments of the virus, or checking them with a microarray of virus RNA fragments.
5231716
You can follow the link and see the model. It's a bunch of differential equations.
Related to what Georg pointed out earlier 5231626, the model would give predictions nearly matching the positives detected now if we assume there were actually 7 people in the US with the virus on 1/11 rather than just 1. That seems quite likely to me. I'm only puzzled how we could be detecting all of the cases with such inadequate testing.
5231734
5231733
Okay, I retract my statement of the excessive survival rate outside a human host. I had not had time to look too deeply into it and trusted media sources for that
However, the problem with the low number of family members coming up positive may be due to a variety of factors. Primarily a lot of the initial test kits sent out by the US and now by China, were extremely faulty and provided an excessive number of false negatives. Data from Italy about the recent surge of younger individuals at the hospital and ICU also indicates that “younger” members of the family are simply holding back the virus to the point of indetectability similar to how (first thing that comes to mind) HIV can remain hidden in a persons bloodstream and only be found by checking for antibodies. Before overwhelmingly the immune response and causing ARDS or general symptoms to appear from the swelling this causes in the major organ system.
It generally took much longer for them to have enough replications for testing to be as accurate on them as with severe cases. Although with a currently estimated 96.6% survival rate that’s not surprising. Additionally, some individuals with accurate testing have gone positive, negative, negative, negative, positive. What remains unclear is if the person in question is recontracting a variant of the virus, or if the virus is only fully controllable through continuous treatment.
5231739
A) Because we still have a testing backlog. Supposedly California has 64,000 samples untested and is doing 2,000 tests a day
B) Because this virus acts different than previous viruses in that it seems to be contagious earlier, is asymptotic in a large fraction of the population (who may still be carriers), and generally acts like a tiny evil version of Discord in that some infected people won't shed any virus, and others are practical fountains. (Thus models are only vaguely accurate)
C) Because a valid positive gives you a good idea where to backtrack to find other cases in low-contact environments. Bob tests positive, and remembers that Fred was sneezing in his office a few days ago, so Fred gets tested, is positive, and everybody in the office is then lined up and swabbed, etc... If you just go out and pick a hundred people at random, you're going to get one, perhaps two positives if any, and since the test has a certain margin of error, you have to retest the positives to be *certain*
5231759 D) (thanks, by the way) Early tests had such a problem, both from the CDC and FDA fumbling their roles and from screwed-up testing supplies.
Legal Insurrection has an informative article on Dr. Fauchi's actual statement and how the media ran wild with it
Maybe I'm missing something, but isn't positives over total tests a completely meaningless ratio? These tests are not being done on an evenly distributed random sampling of the population. The tests are being done on people who are already suspected of having the virus. That's massive sampling bias, and thus the ratio should just be an indicator of how good doctors are at guessing whether someone has covid.
I saw a person on Facebook state they had symptoms precisely matching Covid-19 when they returned from China to the US in late November. Can't find the comment now or I would link it. If it was Covid-19 they had, then the models are wildly off-base on the start date.
5231739
Didn’t they grant the use of experimental testing early February? I wouldn’t doubt that there is a bit of estimation and extrapolation going on in the process. Especially if the test kits like what Abbott is putting out use, “molecular testing,” but get results in five to thirteen minutes. They describe it as, “identifying a small section of the virus' genome, then amplifying that portion until there's enough for detection.”
I could be wrong here, but isn't this on the short side? Everything I read said the asymptomatic-but-contagious incubation period could be up to 14 days, a mild case lasts two weeks after becoming symptomatic, and a severe case can last up to six weeks.
Unless you're just counting the pre-symptomatic period when people might unknowingly infect people, but I'm not sure that's accurate given that very mild cases can reportedly be mistaken for flu/allergies/colds (so especially early on might have been ignored, giving the virus up to a month of infecting time.)
5232044 The confusion here is that scientific papers (IIRC!) usually just say "incubation" to mean "the end of the incubation period", and say "infection" for "start of incubation period". But common usage tends to associate "incubation" with "start".
14 is a low number anyway. Other numbers given range up to 32. Also, all these figures are supposed to be averages for those who end up dying of the disease. Those who survive it have different disease trajectories.
5232025
Yes, and they concluded those February local tests didn't work. I don't know of anyplace where the resulting data was made available. The data I have for Feb.is all for testing done at the CDC.
What Abbott calls "molecular testing" is PCR. There's no estimation or extrapolation. If you get a positive result, you can be as close to 100% sure as you can get that a DNA sequence very close to the one you were looking for was in your sample. If you get a - result, you can be almost 100% sure that DNA sequence is not present in the sample. What you can't be sure about is
5231938
The ratio should reflect how good doctors are etc, what the prior odds of being infected are, and what the ratio of tests available to sick people wanting testing is.
If you have more tests than sick people, you can just test everybody, and then the ratio is the fraction of people with symptoms who have covid-19, which is also the prior odds of someone with symptoms being infected.
Say the prior odds are 1/2. If you have half as many tests as sick people, someone (let's call her Doctor D) picks half of the people to test. If D is perfect, he'll pick exactly the half with covid-19, and the fraction of positives would be 1. But if he's perfect, and the prior odds of infection are only 1/4, the fraction of positives will be 1/2.
IF we assume that the number of tests available is limited by production capacity, then it's very unlikely that graph of production over time will match the graph of virus over time.
IF the local testers are being smart, and predicting the course of the virus, and holding back test kits so as to use a number on each day corresponding to the number of expected cases on that day, THEN you might think you could get a constant fraction.
But you wouldn't, because the prior odds must keep going up if the virus is spreading, because most people with covid-19 develop covid-19 symptoms, while most people without it, don't. If on March 3, 100 people had covid-19 symptoms, but didn't have covid-19, it would be really weird if on March 25, 10,000 people had covid-19 symptoms but not covid-19. This is the key point which most people don't understand. Why does the number of people with covid symptoms but not covid explode just as fast as the number of people who have covid-19?
The only way to get the results seen is if the CDC has predicted the course of the virus accurately, and the states are all cooperating to give just the right number of tests on each day so that the fraction positive will stay at 20%. That is possible. But I haven't heard any indication that they do that.
5231669
Yeah, that sounds plausible.
That's a very good point. I think MIchigan is an outlier; CTP reported for 3/28 20,827 new positives and 463 new deaths. That's 2.2%.. But that 2.2% doesn't account for the lag time.
5231759
If the test was immunological, that's expected to happen sometime, because it's a quantitative test and somebody could have blots that are borderline +/-.
If they were PCR tests, the person was probably infected throughout, but shedding less virus for a time.
Do you know a source that gives details about how often this happens? I don't know why they would even retest someone who already tested positive.
5232448
Also, if it's a PCR test, WHY CAN'T WE SCALE UP A PCR TEST?! All you need is a PCR machine and the right templates, and you can use PCR to duplicate the templates! There are tens of thousands of labs in the US that could be churning out templates tomorrow if you gave them the template sequence today.
I'm going to guess the problem isn't making the test "kit" (the reagents for the PCR), but doing the PCR on the sample. The samples have live virus, so they don't want to send those just anywhere. So probably the test kit is actually doing the PCR? Probably each kit has a microfluidics PCR machine in it.
I'll put it as...
The physicians and epidemiologists have been sounding alarms on this one for months, and people didn't listen, and now it's going exactly as they predicted in terms of spread (Or if anything, worse than they predicted).
I first heard 100k-1 mil deaths about 3-4 weeks ago, when things were still relatively normal. I suspect that range has been revised considerably upwards since then, given the models I have had shared with me - I have a close family member whose a somewhat prominent physician, and the stuff they are saying is...not good.
5232447
That's a common fallacy I've seen people falling for: looking at a day's fatalities and new cases and trying to derive a death rate based on those numbers. But because of the time lag, the numbers will always overestimate the death rate and underestimate the number of cases. At least, until we get on the other side of the curve, when it will reverse.
5232450
CDC test kit is just three sets of primers (two for detecting two different segments of the viral RNA and a control primer set detecting a human gene, along with a positive sample containing a non-infections viral RNA sample to use as a positive control). The CDC kit initially had problems because somehow the kits had been contaminated so that negative control samples (running PCR with no sample added) were giving positive results (perhaps during manufacturing some of the positive sample had contaminated the primer samples).
After the CDC fixed the problem, labs ran into additional problems scaling the test because they were running out of the other reagents needed to test for the virus. First, there were shortages of kits to purify viral RNA from swabs (later the CDC approved additional kits from more manufacturers to help with the shortages) and later labs even began running into shortages of basic supplies like the swabs themselves. Testing has managed to scale up to day largely because the CDC and FDA has approved state labs and private companies to perform their own tests.
It's instructive to compare the US situation to what occurred in South Korea:
https://www.propublica.org/article/how-south-korea-scaled-coronavirus-testing-while-the-us-fell-dangerously-behind
Ironically, South Korea was able to dismantle burdensome regulations to unleash the power of the private sector, while the American response involved relying on a centralized federal agency (the CDC) to (mis)manage testing while preventing individual states and private companies from developing their own tests.
5232448
This was the first info I could find: https://www.npr.org/sections/goatsandsoda/2020/03/27/822407626/mystery-in-wuhan-recovered-coronavirus-patients-test-negative-then-positive but I’m guessing they are retesting certain groups of people and monitoring them in various countries to try and predict the long and very long terms effects of the virus. These people probably relapsed and China was trying to figure out how bad they screwed up and when they can send people back to work
5232480
Yeah I know someone in town who tested positive and then their entire family got sick but weren’t tested. The logic “Yeah you all probably have it, self quarantine and leave unless you get critical, we need the tests for other people. PS: we are only going to count the one person we tested in the official numbers”
5232448
5232504
Here's a figure from a case report monitoring some of the early COVID-19 cases in Singapore:
i.imgur.com/NexoAbd.png
https://www.nejm.org/doi/full/10.1056/NEJMc2003100
You can see that there are examples of negative tests followed by positive tests. The negative tests occurring between positive tests are likely false negatives, which can occur for any number of reasons. However, it is also worth noting that some research suggests that people can test positive after they are no longer infectious.
As an aside, figures like the one above from various scientific papers are one reason why I'd be suspicious of using total tests - positive cases as a metric for negative cases. It does seem like some patients are being tested daily in the hospital (e.g. one of the CDC's guidelines for ending self-isolation of infected cases is two negative tests 24 hours apart). To me, this is evidence that people are getting tested multiple times and if the total tests metric is not careful to avoid counting samples from the same person, then the cases/tests metric you've been analyzing is not useful.
One last observation:
Sure there's no such thing as centrifugal force, but the time to insist on this is not when everybody's turning green on the Mad Tea Party ride:
38.media.tumblr.com/b05f8d534d25d63e6315743351033433/tumblr_n8liwsl8lJ1s2wio8o2_500.gif
Don't ask Alice. I think she'll blow.
5232444
I don't know if it's done this way or not, but the criteria for testing may include "isn't showing symptoms, but has been in contact with a known infected person". The number of such people would increase proportionally to the number of infected.
5234195 Surprisingly, the CDC guidelines say not to do that. They say "NON-PRIORITY: Individuals without symptoms".
So, something interesting in regards to all of this: several public-health officials and epidemiologists have now come out and stated outright that our current models are definitely wrong in several ways, from the way we measure R to the way we've tried to measure rising cases. The current term I've seen thrown around is 'all models are wrong, but some are useful.'
Again, not to say the virus isn't a threat, but rather that our current modeling isn't clearly detailing that threat or the optimal response.