Hacker News Reader

Comment by svara | Hacker News Reader

This model is nice in principle, but its results depend crucially on what is ultimately an educated guess of the reporting rate of symptomatic cases.

This parameter can be found experimentally by testing for antibodies against the virus in a random sample of the population. This has not been done so far. As long as that's true, CFR estimates will be quite uncertain.

Also really interesting in this context is how Germany is looking more and more like a huge outlier in terms of fatality. Almost 1000 cases and 0 deaths so far. Sure, that may still change, but the outbreak has been going on for a while there already.

View on HN

Or look at S Korea, 7313 cases 50 deaths, .68% CFR. Would like to see the age group breakdown.

If I had to guess why CFR is so much lower in S Korea, it's because they tend to test more widely, ie. don't only test when patients show up at the hospital with severe symptoms. Or in other words they are closer to the real CFR.

> Or look at S Korea, 7313 cases 50 deaths, .68% CFR. Would like to see the age group breakdown.

As of 2020-03-08 0:00 AM UTC+9 [1]:

              Patients        Deaths
    Age       (% of total)    (% in group)
    
    80+        212  (3.0%)      14  (6.6%)
    70--79     409  (5.7%)      18  (4.4%)
    60--69     878 (12.3%)      11  (1.3%)
    50--59    1349 (18.9%)       5  (0.4%)
    40--49     975 (13.7%)       1  (0.1%)
    30--39     760 (10.7%)       1  (0.1%)
    20--29    2133 (29.9%)       0  (0.0%)
    10--19     360  (5.0%)       0  (0.0%)
    0--9        58  (0.8%)       0  (0.0%)
    
    Total     7134              50  (0.7%)

[1] https://is.cdc.go.kr/upload_comm/syview/doc.html?fn=15836441... (p. 10)

Most of Korea's cases were discovered since late February and have not had a chance to either recover or die (i.e., the numerator is right-censored). Expect the CFR to rise there.

You are right. If you want some other statistics, as of 2020-03-07 there were 59 patients [1] that are under respiratory support or oxygen therapy and thus considered "severe".

[1] https://www.yna.co.kr/view/AKR20200307046000017

Or not. What about all the infected people who haven't had symptoms serious enough to warrant seeing a doctor and getting tested positive for Coronavirus?

That is a constant factor on the denominator, which is lower in Korea than elsewhere due to the availability of 10-minute drive-through tests. The factor on the numerator will be much higher at the start of an epidemic and go down to 1 when it completes.

The estimated average time to die for those who die from COVID-19 is 28 days. 28 days ago, Korea had fewer than 50 cases total.

How is the number of mild, untested cases a constant factor? If less serious patients start to get tested more frequently, this number will change.

> If less serious patients start to get tested more frequently, this number will change.

If everyone were tested, the denominator goes up by some constant factor. In Korea, people are already being tested at the first sign of cold symptoms.

The vaccines will change all these numbers.

Great data, thanks for taking the time to format it cleanly.

South Korea proactively tested people so their numbers are more skewed towards being the infection fatality rate (case fatality rate are those tested and cobfredtobe sick, ie sick enough to go to the doctor, get tested and be a confirmed case). Where as IFR is the estimate of everyone infected)

Also the demographics of those tested are skewed to the 20–40s age range as the member of a reclusive church was sick and didn’t say anything. They the proactively tested everyone they were in contact with at the church and the demographics of that church skew younger.

Does their test measure antibodies (meaning you were at some point exposed) or does it measure an active infection, or does it measure an active symptomatic infection?

I believe these are PCR tests, so they test for presence of the virus.

Afaik antibody tests do not yet exist but are in development

Yes South Korea are testing huge numbers. Worth watching closely.

Misleading. You can’t look at CFR with the denominator being total number of cases, at least in the exponential growth phase. It will massively understate the actual CFR.

But you can compare this ratio between countries that have similar numbers of cases to get a hint at relative detection efficacy, assuming constant CFR.

Germany's outbreak has not been going on for a while. They had 79 cases one week ago (https://web.archive.org/web/20200301001656/https://www.world...), and the number was 16 two weeks ago (https://web.archive.org/web/20200223003733/https://www.world...), and 14 four weeks ago (https://web.archive.org/web/20200210100748/https://www.world...). It takes weeks to die after getting infected.

That's not all there is to it, since the number of people infected grew at a similar pace in the early phase of an outbreak in different countries.

For example, the number of people infected in France and Germany is rising in near perfect synchrony, but France has 11 deaths now. Similarly, growth in the early phase of the South Korea outbreak looks nearly identical to Germany and France (with a ca. 2 week delay), and they had 10 deaths when they had 1000 infected.

Germany is looking increasingly anomalous every day.

There has to be a difference with respect to detection, unless it's something more out there like viral strain differences.

Actually, to the point of viral strain differences, I did come across this paper indicating exactly that idea— that there are two strains. I am not a scientist, so I cannot evaluate the veracity of the claims, but did find the work interesting and compelling: https://academic.oup.com/nsr/advance-article/doi/10.1093/nsr...

This was posted here a few days ago, and there were plenty of criticisms of it: https://news.ycombinator.com/item?id=22499152

79 cases last week to 1000 today doesn't indicate that there's been a new outbreak this week; it indicates that there's been an outbreak for a while and they're just now noticing. Those 79 people alone couldn't have infected 1000.

That, and with Christmas still in living memory the endangered age bracket is effectively quarantined anyways, not expecting the the next quarterly visit from children and grandchildren before the end of the month. Germany is a country of lonely old people.

I also saw this and found it odd. Assuming p=0.01, N=1000, k=0, the binomial probability is 0.00004317. At p=0.02, 0.03 it's even lower.

not following closely, but unfortunately i don't think the data is good enough to take a hard-stats point of view of significance - take for example the rest home cluster in washington state which would hugely skew the results on a 1k sample size.

South Korea are testing huge numbers of people, 10,000 per day, so the stats from there are instructive

But can you know that Germany is an outlier? Germany likely has a higher reporting rate than China so its numbers should be closer to the truth.

Germany is not testing nearly enough people. There are lots of reports of people with legitimate concerns that did not find a way to get tested. Northern italy is like 2 hours by car and each weekend thousands drive there for a short trip.

> Germany is not testing nearly enough people

Source? I’ve heard exactly the opposite - Germany is testing the whole cluster as soon as infection is found. That explains the low mortality.

In Germany they are tracking down clusters. As a random person you won't get tested, unless you visited a risky region, had contact with an infected person or your doctor orders a test for you. As a German, this strategy makes sense to me.

Considering the shortage of reagentia for these tests, that definitely makes sense.

Belgium only does tests on people who have recently been in regions at risk (China, Italy) or who have symptoms of pneumonia.

200 infected to date, zero deaths (so far).

Yep, that’s my understanding too and it makes perfect sense.

The low mortality is likely due to the recent introduction of the virus into the community there. It takes weeks to go from infection to ICU.

That’s incorrect, Germany is very much like France in terms of number of cases and their duration.

Something needs to explain the difference though between 939:0 in Germany and, e.g., 949:11 in France. Maybe the diagnoses in Germany are closer in time to infections? That would still point to a higher reporting rate in Germany vs. other countries.

France has had a continuous epidemic since early on, but Germany successfully contained the first outbreak (thanks to one phone call from China at the right time), then it was all quiet for a while and only now Germany is catching up in case count, at a very high speed. This means that Frances, with a comparable total, has far more cases that are old enough to have already played out negatively.

Germany and France have had very parallel number of cases: see https://github.com/CSSEGISandData/COVID-19/blob/master/csse_...

The first case in Germany was 3 days after the first in France. They reached 4 cases the same day. When France hit 11 cases, Germany already had 13 and for 15 days they both only got 1 to 3 more cases in total. They both began increasing a lot on the 26th of February.

You’re assuming that the reporting in Germany is the same as the reporting in France.

While they reached 4 cases the same day, that doesn’t mean that they had the same number of unknown infections. Let’s say France had 10x as many unknown infections than Germany. How would that impact the disparity in death count? And is this even a large enough sample size for that disparity to not just be explained to a high degree by mere chance?

I haven't followed the French situation closely, are the late February cases in France clearly unrelated to early February cases or could they be the result of a phase of undetected spread?

That's not true, case numbers in France and Germany are highly correlated. What you are saying would imply a substantial difference in growth rate between the two countries, which is easily disproved by looking at the data.

Exactly my point ;)

Reply on news.ycombinator.com