4 ms·

The standard way is to count how many deaths there were in excess of what would be expected in a normal year. Yes, but you can only do that in aggregate over a

by tremon 6y ago

The standard way is to count how many deaths there were in excess of what would be expected in a normal year. Yes, but you can only do that in aggregate over a larger period, you can't use it as a basis for daily statistics, not in the least because not all deaths are recorded the same day they occur. The daily death tallies are based on confirmed cases only, in pretty much every country. That means underreporting of cases is the norm, not the exception [0]. I must say those graphs are a bit hard to interpret, but if I understand correctly: a gray bulge above the dark gray dotted line implies excess deaths not officially attributed to Covid-19. From those figures, it seems only Denmark and Germany are accurately reporting Covid deaths, and Belgium, France and the US may be overreporting the death tally (although it can also mean that not all deaths in the past weeks have been recorded yet in the official figures). [0] https://www.theguardian.com/world/ng-interactive/2020/may/29/excess-deaths-uk-has-one-highest-levels-europe https://www.theguardian.com/world/ng-interactive/2020/may/29...

jtbayly 6y ago
The excess death numbers in the US are from the CDC, and their numbers are incomplete unless you go back about 8 weeks. I don’t know why people keep creating graphs and charts from their incomplete data! It paints a wrong picture every single time.
- toast0 6y ago
  Creating a graph with only data from 8 weeks ago would be an accurate graph, but it would tell us what we should have been doing 8 weeks ago, not what we should be doing now. Of course, death is already a lagging indicator, so really this would tell us what we should have done 10 weeks ago. It's just not an actionable graph, because it's too delayed. Of course, using incomplete data may not tell us anything either. It's also not an actionable graph, because there's too much uncertainty.
dredmorbius 6y ago
Covid-specific mortalities, and moreso testing, give trends indications. Even here there are lags and noise; testing trails infection by 1-2 weeks, mortality lags testing by about 19 days (call it three weeks), and actual infection/exposure by about five weeks, aas a very rough average. All available data are seen through the rear-view mirror and tell us where we were not where we are. Those who work in such scenarios are aware of this, and other limitations: test coverage and limitations, sampling, selection criteria, reporting lags, comorbidity biases, etc. Most of the rapidly-available data are subject to greater bias. Much of the more accurate data are subject to greater lags. But with sufficient history we can use the latter to adjust the former (pending final calibration measurements) to improve the clarity of current assessments.

by tremon 6y ago

jtbayly 6y ago

The excess death numbers in the US are from the CDC, and their numbers are incomplete unless you go back about 8 weeks. I don’t know why people keep creating graphs and charts from their incomplete data! It paints a wrong picture every single time.

toast0 6y ago
Creating a graph with only data from 8 weeks ago would be an accurate graph, but it would tell us what we should have been doing 8 weeks ago, not what we should be doing now. Of course, death is already a lagging indicator, so really this would tell us what we should have done 10 weeks ago. It's just not an actionable graph, because it's too delayed. Of course, using incomplete data may not tell us anything either. It's also not an actionable graph, because there's too much uncertainty.

dredmorbius 6y ago

Covid-specific mortalities, and moreso testing, give trends indications. Even here there are lags and noise; testing trails infection by 1-2 weeks, mortality lags testing by about 19 days (call it three weeks), and actual infection/exposure by about five weeks, aas a very rough average. All available data are seen through the rear-view mirror and tell us where we were not where we are. Those who work in such scenarios are aware of this, and other limitations: test coverage and limitations, sampling, selection criteria, reporting lags, comorbidity biases, etc. Most of the rapidly-available data are subject to greater bias. Much of the more accurate data are subject to greater lags. But with sufficient history we can use the latter to adjust the former (pending final calibration measurements) to improve the clarity of current assessments.