The first is that we have a strong mechanism story we can tell. Viruses take time to multiply. When the immune system detects a virus it responds. If your initial viral load is low your immune system gets a head start, so you do better.
The problem with this story is that it assumes that immune system detection time is not dependent on viral load, which seems highly unlikely. The more viral particles, the more likely they will be detected. How that interacts with viral load’s more obvious direct effects is complex and probably virus strain dependent.
The second category is the terrible outcomes in health care workers on the front lines. Those who are dealing with the crisis first hand are dealing with lots of intense exposures to the virus. When they do catch it, they are experiencing high death rates.
Your third category on analogy evidence from other viruses makes sense, with the single example from the other SARS coronavirus carrying more weight as it’s a much closer relation than smallpox or measles.
For example, 712 of 3700 people on DM became ill, which gives crude AR = 19.24 per cent.
(I think you meant DP for Diamond Princess)
This is a lower bound on the number infected. From what I understand, PCR viral detection peaks in the 80% to 90% range a few days after exposure, but then falls off to 20% or lower after about a week or two on average (but at some variable rate depending on immune interaction as you mention).
They didn’t test everyone quick or frequently enough such that the known case number is a tight bound on the true case number. From what we know on PCR false negative time curve, it seems likely that the true AR on DP is anywhere from 30% to as high as 60%.
If we model the PCR detection time curve as being age dependent (which seems reasonable), then that predicts that the AR was probably above 50% on the submarine and perhaps DP, it just wasn’t all detected. For the submarine in particular the population is probably skewed a bit younger/healthy and thus more mild/asymptomatic cases that fight it off quickly.
Most places in europe seem to be in or near mid sigmoid at this point (with the US probably not far behind), but it’s too early to tell whether that’s due to high AR and herd immunity or lower AR and social distancing.
The Kinsa data seems to suggest that the AR was on average only about on order flu AR, but perhaps higher in some hotspot cities (where the implied AR is perhaps larger than flu). That data also suggests social distancing & closures were effective, but there are some counterexamples (like miami ) where it seemed to peak naturally too early to explain by (late) social distancing.
There’s also the Japan mystery, which should have a high AR at this point. Right now the most likely explanations I can think of are either flu like severity/mortality that’s not that noticeable when you aren’t directly looking for it, or a strain difference.
Interesting—hopefully it’s not long until someone publishes a serology random sampling study.
Not surprised at symptomatic fraction of 50% - was already indicated by DP, Iceland, and other data.
One thing that is surprising/mysterious to me is how steady the PCR test positive% has been across space and time. When the sampling is of general populations outside hospitals, it’s ~1% in Iceland without changing much over time, and 2% in NBA players and 1% in expats flown home from china.
The test positive fraction for tests conducted by clinics/hospitals in the US and Iceland is steady at about ~10% and hasn’t fluctuated greatly over time.
Of course there are some places where it’s much higher like 30% on DP, but that’s an exceptional environment.
Now the typical PCR test of nasal/throat swab is only accurate for about a week or so after infection, so it’s more of a blurred measure of the infection derivative, but still it doesn’t look like there’s any recent exponential growth—suggesting it was in the past.
That’s interesting. Over the weekend I wrote a monte carlo simulation for the Iceland data incorporating a bunch of stuff including a lognormal fit to know median and mean time from confirmation to death. Going to write it up, but the TLDR: the posterior assigns most of it’s mass to the 0.2 to 0.4 range for reasonable settings. Want to do something similar for Diamond Princess and other places.
I expect the real IFR will vary of course based on age structure, cofactors (air pollution seems to be important, especially in Italy), and of course the rather larger differences in coroner reporting standards across jurisdictions and over time.
You can avoid alot of that by looking for excess mortality—which right now seems null in europe except for in Italy. But Spain has about the same cases and deaths per capita and no excess mortality.
For a really rough analysis, the overall IFR on the DP was probably about 1% (10 deaths / 1000 infections) after adjusting slightly for false negatives / missed tests.
All those deaths are 70+ age with an in IFR in that group ~2%. About 10% of the US population is in the 70+ bracket, so the projected IFR is ~0.2%. However about half the deaths were in the 80+ age bracket, and if you do a more fine grained binning it’s probably more like 0.15%, but it’s not a high precision estimate.
Assuming a similar age distribution for actual infections, this means a larger fraction of young people is coming down with severe disease.
Disease severity increases with age, and testing probability increases with severity and thus age (in most places). Thus the ratio p(tested | infection) is age skewed and typically much lower for younger ages.
After adjusting by dividing by age dependent p(tested | infection) you can correct that skew and you probably get something more similar to influenza hosp rate curve.
So again you aren’t comparing even remotely the same units and it’s important to realize that.
If you look at my estimate I’m already effectively predicting that their CFR will increase via predicting additional deaths. I think it makes more sense to predict future death outcomes in the current cohort of patients we are computing IFR rather than predicting future CFR changes based on how they changed in other countries and then back computing that into IFR.
The CFR can change over time not only because of delays in deaths vs stage of epidemic but also due to changes in testing strategy and or coverage, or even changes in coroner report standards or case counting standards (as happened at least once with china).
In terms of true number of infected, I’m predicting that SK has on the order of 100K to 200K cases and say 4K in Iceland, and I don’t find this up to ~50x difference very surprising. Firstly, it’s only about an 18 day difference in terms of first seed case at 25% daily growth.
SK’s first recorded case was much earlier in Jan 20 vs Feb 28 for Iceland. SK’s epidemic exploded quickly in a cult, Iceland’s arrived much later when they had the benefit of seeing the pandemic hit other countries—they are just quite different scenarios.
Source for a virus making threats?
It skewed the age structure toward a younger demographic. Were you aware of this or did you assume that the religious group is skewed toward old people like typical churches? I didn’t realize this up until like ten days ago, but the Christian cult was predominantly pretty young people!
Yes, I should have made this more clear—but it skewed it younger. Or at least that’s my explanation for their much higher than expected # cases in younger cohorts vs elsewhere. That should lower their CFR of course.
And about Iceland: Isn’t it really very clear that Iceland is weeks behind South Korea, and that Iceland’s numbers are therefore unrepresentatively low?
No this isn’t clear. Iceland’s case count entered a linear regime roughly 2 weeks ago—ie they do seem to have it under control (at least for now). Modeling one country as “X weeks behind” some other country is hazardous at best and also unnecessary as Iceland provides direct graphs on their daily #tests and #positive.
As for hospitalizations, I was comparing the age distribution of hospitalizations for flu and confirmed covid. I found that the ratio of 20-45:65+ hospitalizations for flu was 1:7, and that the same ratio for covid was 1:2. Assuming a similar age distribution for actual infections, this means a larger fraction of young people is coming down with severe disease.
Age distribution of estimated hospitalizations for flu? or confirmed? (It seems difficult to get the latter) Source?
It should also be noted that some all-cause death rates are coming out in North Italy, and the excess deaths over this time last year are 3x the confirmed covid death
Source? This is potentially interesting especially if it’s for a large region like all of Italy or North Italy (i’ve seen models which estimate excess influenza mortality in Italy) - but the smaller the region the more likely it’s due to chance or cherry-picking.
From Govenor Cuomo’s briefing:
Everything we do now (procure ventilators etc) is in preparation for possible apex (when curve hits the highest point)
Apex in New York is estimated in 14-21 days from now
172 new ICU admission in the last day, vs. 374 in the preceding day, may indicate a decline in the growth rate
A demand of 1000 ICU beds suggests about 300K infected in NY assuming influenza like IFR of ~0.1% and ICU mortality of ~30%, so this isn’t in disagreement. More likely if 1M are infected demand should be for ~3000 ICU beds.
There may or may not be a difference in mean ICU/ventilator length of stay—that isn’t something I’ve looked at yet. According to Cuomo C19 patients need ventilators for 11 to 21 days vs 3 to 4 days for all other causes. This paper indicates 6 to 17 days for H1N1 in 2009.
Are you saying that some significant fraction of NY hospitals are currently overcrowded with C19 patients right now? Or that one hospital is? What is the actual dataset source for “they are strapped for space”?
There seems to be good evidence for asymptomatic transmission—you’ve probably seen those papers, which indicate that tracking and isolating cases doesn’t work.
What does seem to work is social distancing.
Firstly, Iceland is not ‘randomly’ testing people. People are signing up to be tested voluntarily. That population is likely to contain a larger fraction of people who have reason to think they were exposed or feel sick. Thus 0.8% is an overestimate of the fraction of the population that has been infected.
Technically true—and this is why in the earlier version of this on my blog, I used the word ‘random-ish’.
Obviously the test is voluntary, but it’s also clearly designed to estimate prevalence:
″ This effort is intended to gather insight into the actual prevalence of the virus in the community, as most countries are most exclusively testing symptomatic individuals at this time,” said Thorolfur Guðnason, Iceland’s chief epidemiologist to Buzzfeed.
During this time of year less than 10% of the population has symptoms, so if it was a random sampling of only that subset, we would predict at most 400 cases, so we can reject that.
Nonetheless, I think this does justifying widening the prediction of #infections and moving the mean down a bit.
Secondly, the asymptomatic period is on average a week or so for those who develop symptoms, with hospitalization often occurring upwards of a week after symptoms, and death often occurring more than 2 weeks after symptoms. .. .This thing is damn infectious and still expanding, it is not anywhere near a steady state anywhere
Did you actually look at the Iceland data? They entered a linear regime (midpoint of the sigmoid) about 10 days ago, which defeats the brunt of this argument. Additionally the vast majority of the cases were discovered through normal testing after symptoms present, so subtract a week from your timeframe. And finally I already did attempt to predict future deaths based on ICU. I also considered adding another predicted death from the # in hospital now, but it’s unclear if that is distinct from ICU or not.
Ultimately though only time will tell, but I find it unlikely they are going to get up to dozens of deaths without also growing case count.
According to links in the above writings, 0.5% of flu cases in the 20-45 age group result in hospitalization compared to 10% in the over 65 age group, and taking population into account that results in ~7x as many flu over-60 hospitalizations than 20-45. Current American test results, however, have ~2x the over-65 covid hospitalizations as 20-45 hospitalizations.
The hospitalization rate that matters is (hosp | infected), not (hosp | tested). You are comparing the estimated (hosp | infected) curve of influenza to the (hosp | tested) curve of COVID-19, which is a unit mismatch. For that comparison to be meaningful you need to first correct for age-specific (tested | infected) ratio.
And data is indicating that surviving ICU stays for this disease are ~3x as long as ICU stays for flu.
South Korea is unusual in that the outbreak there is best understood as two separate outbreaks: an initial outbreak in a strange highly interconnected cult, and then the outbreak in the general population. They ended up testing everyone in the cult, but their testing strategy in the general population seems more limited, similar to other countries. So the testing of several hundred thousand cult members pushed both their CFR and test positive fraction lower than it otherwise would be, and rather obviously skewed their case age structure.
Nonetheless they have tested far less of their population than Iceland (about 5X less as of 3⁄20 according to ourworldindata), so if the ratio of infections/cases is 4x to 5x in Iceland it seems reasonable that it’s 10x to 20x in SK.
Naively if ICU fatality is ~30%, and we worst-case assume those all become deaths absent ventilators, that suggests about 3X higher deaths sans ventilators. However in reality we would/will probably just quickly produce more ventilators, start sharing ventilators, jury-rigging C-pack machines into ventilators, etc.
Perhaps this isn’t clear enough from the title (but should be clear from the post), that the similarity I”m discussing is in terms of outcomes given illness: IFR and IHR.
Absent controls and behavioral changes, I agree that it seems likely that considerably more than 1% of the population would be infected. Seasonal flu infects perhaps 10%. It’s clear at this point that C19 is often asymptomatic/mild especially in younger people, and I recall some potential bio explanations like pre-existing partial immunity through cross reactive antigens. On the DP we know about 30% were infected and it could be higher—perhaps 50%, but that population is half retirees. So from this evidence alone my estimate is somewhere between 10% to 50% would be infected absent any behavioral changes.
However social distancing appears to have already be crushing fever prevalence in the US.
For the contagious part—I guess what really matters is what % of the population it could infect, and how fast that could occur. But most of the world has gone into social isolation, which at least in the US appears to already have been highly successful.
The kinsa thermometer dataset is quite interesting and worthy of it’s own post. If you look at places that didn’t do much social isolation in time, like Miami, it appears that the answers may be that it causes fevers in about the same % of the population as the flu does, and cycles through the population in perhaps half the time-frame (viruses move through cities faster in general).