AlanSubie4Life
Efficiency Obsessed Member
Another factor I didn't mention before is that because the numerator is smaller than it
Really what I'm saying is that of course you get close to the right result there, since both numbers have reasonably settled. But you would have got a very pessimistic result with a 14-day lag if your case number had fallen on the exponential part of the curve. For example, March 7th and Feb 22nd. You get: 48/436 = 11%
What you're trying to do here is estimate CFR given an epidemic during the full exponential growth period. You can't do that if you're trying to test your method on a curve that isn't even close to exponential (for the cases number specifically - the death number doesn't have to fall on the exponential for the most part, though there will be some error, due to distribution of death delays).
I agree the Iceland data is too noisy to tell, and besides, their testing method is such that you'd generally expect lower CFRs.
Specifically in regards to the US data, using your 14-day method:
April 1st: 5102
March 18th: 9197
That gives you: 55% projected CFR for the US.
If you use a 6-day lag (April 1st, Mar 26th):
US: 5102 / 104125 = 4.9% (Probably still thrown off by insufficient test capacity resulting in a very low denominator...)
Germany: 931 / 43938 = 2.1%
South Korea (Mar 7th, Mar 1st) 48/3173 = 1.5%
Iceland: 2/802 = 0.25% (I think thrown off by the test method).
Just picking arbitrarily a couple countries where the outbreaks are not so severe that they overload the system leading to under-testing 6 days ago, and where they might be competent:
Israel:
6-day: 26/ 2693 = 0.9%
14-day: 26/433 = 6%
Canada:
6-day: 114/4043 = 2.8%
14-day: 114/727 = 16%
Switzerland:
6-day: 488/11811 = 4.1%
14-day: 488/3115 = 16%
In the end, there are so many sources of error with this I think it's probably futile, though. I think all you can do is take the current deaths (assuming they are a good count) and use that to back calculate approximately how many cases existed x number of days ago. It's harder to determine the number of cases today, because it depends a lot on the growth rate in the particular country in question.
No, we should not ignore this at all, this is actually the best example, since using the 14-day lag, one could see that coming already.
Really what I'm saying is that of course you get close to the right result there, since both numbers have reasonably settled. But you would have got a very pessimistic result with a 14-day lag if your case number had fallen on the exponential part of the curve. For example, March 7th and Feb 22nd. You get: 48/436 = 11%
What you're trying to do here is estimate CFR given an epidemic during the full exponential growth period. You can't do that if you're trying to test your method on a curve that isn't even close to exponential (for the cases number specifically - the death number doesn't have to fall on the exponential for the most part, though there will be some error, due to distribution of death delays).
I agree the Iceland data is too noisy to tell, and besides, their testing method is such that you'd generally expect lower CFRs.
Specifically in regards to the US data, using your 14-day method:
April 1st: 5102
March 18th: 9197
That gives you: 55% projected CFR for the US.
If you use a 6-day lag (April 1st, Mar 26th):
US: 5102 / 104125 = 4.9% (Probably still thrown off by insufficient test capacity resulting in a very low denominator...)
Germany: 931 / 43938 = 2.1%
South Korea (Mar 7th, Mar 1st) 48/3173 = 1.5%
Iceland: 2/802 = 0.25% (I think thrown off by the test method).
Just picking arbitrarily a couple countries where the outbreaks are not so severe that they overload the system leading to under-testing 6 days ago, and where they might be competent:
Israel:
6-day: 26/ 2693 = 0.9%
14-day: 26/433 = 6%
Canada:
6-day: 114/4043 = 2.8%
14-day: 114/727 = 16%
Switzerland:
6-day: 488/11811 = 4.1%
14-day: 488/3115 = 16%
In the end, there are so many sources of error with this I think it's probably futile, though. I think all you can do is take the current deaths (assuming they are a good count) and use that to back calculate approximately how many cases existed x number of days ago. It's harder to determine the number of cases today, because it depends a lot on the growth rate in the particular country in question.