Read on Twitter
I know the term "EXCESS DEATHS" gets mentioned quite a bit since death is the most obvious adverse outcome associated with this pandemic.
There was even an MMWR on Oct 23 dedicated to this, which spoke to ~300,000 excess deaths this year.
Brief
1/ https://www.cdc.gov/mmwr/volumes/69/wr/mm6942e2.htm#contribAff
There was even an MMWR on Oct 23 dedicated to this, which spoke to ~300,000 excess deaths this year.
Brief
1/ https://www.cdc.gov/mmwr/volumes/69/wr/mm6942e2.htm#contribAff
The data reside at the site 
, updated weekly.
CDC offers several datasets and Tableau workbooks to download, in addition to the technical details on the process.
Many things we could get into, but let's do the simplest - all-cause excess deaths.
2/
https://www.cdc.gov/nchs/nvss/vsrr/covid19/excess_deaths.htm#techNotes
, updated weekly.CDC offers several datasets and Tableau workbooks to download, in addition to the technical details on the process.
Many things we could get into, but let's do the simplest - all-cause excess deaths.
2/
https://www.cdc.gov/nchs/nvss/vsrr/covid19/excess_deaths.htm#techNotes
All-cause means we include any death.
It's the easiest because you don't have to worry about misclassification regarding cause of death.
The goal is to compare the # of deaths that occur each week in 2020 to what we would have EXPECTED based on the past 5 or so yrs.
3/
It's the easiest because you don't have to worry about misclassification regarding cause of death.
The goal is to compare the # of deaths that occur each week in 2020 to what we would have EXPECTED based on the past 5 or so yrs.
3/
If what we OBSERVE (actual deaths) is more than what we EXPECTED, then those are EXCESS deaths.
If what we OBSERVE (actual deaths) is not more than what we EXPECTED, there are NO EXCESS deaths.
4/
If what we OBSERVE (actual deaths) is not more than what we EXPECTED, there are NO EXCESS deaths.
4/
It is important to note that the approach used by CDC (Farrington surveillance algorithm) is only designed to detect whether we are seeing higher than expected deaths.
IT is NOT DESIGNED to detect fewer deaths than expected...this is due to reporting lag (will get to that).
5/
IT is NOT DESIGNED to detect fewer deaths than expected...this is due to reporting lag (will get to that).
5/
This begs the question. What constitutes the # of deaths we EXPECT?
Death data are taken from previous yrs (e.g., 2015-2020).
A model is used to estimate what we might expect in 2020.
The model accounts for the fact that deaths ebb and flow during the year (seasonality).
6/
Death data are taken from previous yrs (e.g., 2015-2020).
A model is used to estimate what we might expect in 2020.
The model accounts for the fact that deaths ebb and flow during the year (seasonality).
6/
For each week in 2020, 2 estimates are generated:
(1) the average # of deaths we should EXPECT
(2) the upper bound # of deaths we should EXPECT
These numbers form the thresholds for expected deaths.
We will now compare what we observed to see if there is an excess.
7/
(1) the average # of deaths we should EXPECT
(2) the upper bound # of deaths we should EXPECT
These numbers form the thresholds for expected deaths.
We will now compare what we observed to see if there is an excess.
7/
This is a simple figure, all ages combined, using the most recent CDC dataset.
This is where we compare the deaths observed to the range of what the model said we should expect in 2020.
Notice I've grayed out the more recent weeks (more on why later).
Y-axis is deaths.
8/
This is where we compare the deaths observed to the range of what the model said we should expect in 2020.
Notice I've grayed out the more recent weeks (more on why later).
Y-axis is deaths.
8/
Simply put, if the # of deaths we observed is higher than the upper bound of what we should expect, then the difference between those represent the lower estimate of excess deaths.
I've tried to shade that in yellow.
9/
I've tried to shade that in yellow.
9/
Please note that we could also compare what we observe, not to the upper bound, but to the average expected deaths. This is the higher estimate of excess deaths.
So, now we have a lower and higher estimate, each week, for the # of excess deaths.
10/
So, now we have a lower and higher estimate, each week, for the # of excess deaths.
10/
Why a lower and higher estimate of excess deaths?
Why not just one estimate?
Well, because these are based on models with statistical uncertainty - having a range of plausible estimates represents that uncertainty better...even if it drives us crazy
11/
Why not just one estimate?
Well, because these are based on models with statistical uncertainty - having a range of plausible estimates represents that uncertainty better...even if it drives us crazy
11/
We can then distill this down to what's meaningful here: a focus on "excess all-cause deaths" (Y-axis) during each week.
I've labeled the numbers based on the higher estimate of excess deaths, but one could do the same for the lower estimate.
12/
I've labeled the numbers based on the higher estimate of excess deaths, but one could do the same for the lower estimate.
12/
A little bit of the weeds.
When we compare observed to expected, the "observed" is weighted to attempt to adjust for incomplete reporting of deaths (the reporting lag).
This lag can be considerable is recent weeks.
13/
When we compare observed to expected, the "observed" is weighted to attempt to adjust for incomplete reporting of deaths (the reporting lag).
This lag can be considerable is recent weeks.
13/
This "adjustment" can be seen below.
Notice on
how the red line (adjusted for under-reporting) deviates from the actual # of deaths reported more and more in recent weeks.
Or look at the
to see the difference in the adjusted and unadjusted estimates over time.
14/
Notice on
how the red line (adjusted for under-reporting) deviates from the actual # of deaths reported more and more in recent weeks.Or look at the
to see the difference in the adjusted and unadjusted estimates over time.14/
Even with this adjustment, each week that CDC releases new data, the observed numbers of deaths (adjusted and unadjusted) continue to increase for recent weeks, just like in the animation for FL deaths below.
What does this mean?
15/
What does this mean?
15/
Conclusions one makes in RECENT WEEKS as to:
(1) whether the observed deaths exceed the threshold of expected deaths
(2) how many excess deaths there may be
May be misleading.
16/
(1) whether the observed deaths exceed the threshold of expected deaths
(2) how many excess deaths there may be
May be misleading.
16/
This is why I gray out the right side of the figure below, just like we gray out the right side of any "deaths by date of death curve".
And why other commonly used "deaths by date of death" curves SHOULD BE grayed out or with a disclaimer.
17/
And why other commonly used "deaths by date of death" curves SHOULD BE grayed out or with a disclaimer.
17/
There are so many other issues to consider, like excess deaths by age, race/ethnicity, jurisdiction, etc...
But this is hopefully a simple primer to understand what CDC is trying to do with their weekly estimates.
18/
But this is hopefully a simple primer to understand what CDC is trying to do with their weekly estimates.
18/
TL;DR
(1) excess deaths = the extent to which deaths that occur in 2020 are higher than what we expected
(2) what we expect is estimated with a statistical model
(3) Be careful in interpreting excess deaths during the most recent 3-4 weeks (at least), can be misleading
19/END
(1) excess deaths = the extent to which deaths that occur in 2020 are higher than what we expected
(2) what we expect is estimated with a statistical model
(3) Be careful in interpreting excess deaths during the most recent 3-4 weeks (at least), can be misleading
19/END
