Read on Twitter
A 
about case-fatality rates (CFR) in FL
I've seen a lot lately about CFR decreasing over time.
Even if we restrict to Mar-Oct cases (to remove egregious lagged reporting), for 50y+ we see the chart
I use two time periods for simplicity's sake.
1. Mar-Jun
2. Jul-Oct
1/
about case-fatality rates (CFR) in FLI've seen a lot lately about CFR decreasing over time.
Even if we restrict to Mar-Oct cases (to remove egregious lagged reporting), for 50y+ we see the chart
I use two time periods for simplicity's sake.
1. Mar-Jun
2. Jul-Oct
1/
But, we know testing in early months was not very expansive, resulting in identifying cases that were more likely to have a more severe "experience" with COVID-19.
As testing becomes more available, you get more "cases" who may have no or minor symptoms (
risk of death).
2/
As testing becomes more available, you get more "cases" who may have no or minor symptoms (
risk of death).2/
Although the data in the caseline file are much less than perfect for identifying hosps, ED visits, etc., I illustrate this by classifying cases into 3 mutually exclusive groups:
1. Hospitalized
2. ED visit (but no hosp)
3. No hosp or ED visit
& I do this for each age group.
3/
1. Hospitalized
2. ED visit (but no hosp)
3. No hosp or ED visit
& I do this for each age group.
3/
Now when we look at the CFRs in Mar-Jun and compare them to Jul-Oct:
1. Remember the analysis that lumped all case types together showed universal decreases (on LEFT)
2. But, when we stratify by age & case type, the story is more nuanced (on RIGHT) - some
some
4/
1. Remember the analysis that lumped all case types together showed universal decreases (on LEFT)
2. But, when we stratify by age & case type, the story is more nuanced (on RIGHT) - some
some 4/
Let's focus more on 1 age group to illustrate: 80-84 years
CFR in Mar-Jun vs. Jul-Oct
Overall
- all cases: 23.0% vs. 17.3%
Case "subtypes"
- hosp cases: 39.1% vs. 39.5%
- ED cases: 12.8% vs. 16.3%
- No hosp or ED: 9.0% vs. 9.2%
5/
CFR in Mar-Jun vs. Jul-Oct
Overall
- all cases: 23.0% vs. 17.3%
Case "subtypes"
- hosp cases: 39.1% vs. 39.5%
- ED cases: 12.8% vs. 16.3%
- No hosp or ED: 9.0% vs. 9.2%
5/
WAIT 
How can the overall CFR be nearly 6% LOWER in Jul-Oct than it was in Mar-Jun...
...while also being HIGHER in every case subgroup (hosp, ED, and no hosp or ED)?
The answer
CONFOUNDING
6/
How can the overall CFR be nearly 6% LOWER in Jul-Oct than it was in Mar-Jun...
...while also being HIGHER in every case subgroup (hosp, ED, and no hosp or ED)?
The answer
CONFOUNDING6/
In Mar-Jun, when testing was less pronounced, only 49.2% of all cases in those 80-84y had NO indication of hosp or ED visit.
That to 70% in Jul-Oct when testing expanded.
So, testing picked up more LOWER-SEVERITY cases in Jul-Oct (who are less likely to die)
7/
That
to 70% in Jul-Oct when testing expanded.So, testing picked up more LOWER-SEVERITY cases in Jul-Oct (who are less likely to die)
7/
This happens when comparing adverse outcomes between 2 hospitals, hosp A that cares for the sickest patients & hosp B that cares for those w/ less severe disease.
Even if the hosp A has worse outcomes, it's not necessarily b/c care is worse, it's patients were just sicker.
8/
Even if the hosp A has worse outcomes, it's not necessarily b/c care is worse, it's patients were just sicker.
8/
So, what is usually done is to "risk adjust" - statistically assuming that the two hospitals saw a similar mix of patients, and then we compare the rate of adverse outcomes.
A simple way to do this is "direct standardization"
Let's try it for our data on 80-84y olds.
9/
A simple way to do this is "direct standardization"
Let's try it for our data on 80-84y olds.
9/
If we assume that testing in Mar-Jun and Jul-Oct was identifying cases with the same severity mix (reflected by likelihood of hosp or ED visit)...
...then there is not much of a difference in CFR betwee nthe 2 time periods.
Calcs are below, but prob too much detail.
10/
...then there is not much of a difference in CFR betwee nthe 2 time periods.
Calcs are below, but prob too much detail.
10/
If you do this process - using direct standardization to assume that testing is picking up a similar "mix" of cases - you get the results below when comparing adjusted CFRs in Mar-Jun with Jul-Oct.
11/
11/
NOTE: This is clearly a flawed assessment b/c it leverages hospitalization and ED visit indicators in the caseline file that are incomplete at best.
But, it's all we (the public) have available to us.
I also consider only two time periods to simplify the explanation.
12/
But, it's all we (the public) have available to us.
I also consider only two time periods to simplify the explanation.
12/
Moreover, although I exclude cases from Nov & Dec, the death reporting lag can be so pronounced that even deaths from Oct & Sep (even Aug) are incomplete.
13/
13/
So despite being a superficial analysis...
I think one cannot compare CFRs from different time periods without considering how expansive testing was/is and the degree to which testing is picking up (or missing) a higher % of people with no or extremely mild symptoms.
14/
I think one cannot compare CFRs from different time periods without considering how expansive testing was/is and the degree to which testing is picking up (or missing) a higher % of people with no or extremely mild symptoms.
14/
Ideally, we'd estimate the infection fatality ratio (IFR) in FL, including all people who had the virus, whether or not they've tested +
People have estimated it nationally, but it's hard to know definitively b/c it's hard to know how many + people testing is missing.
15/
People have estimated it nationally, but it's hard to know definitively b/c it's hard to know how many + people testing is missing.
15/
TL;DR
- Age-specific CFR in FL has been going
- This is at least due, in part, to testing "finding" a larger % of sicker cases (er risk of death) in early months
- The lag btw when a person died & when it is reported makes recent trends in deaths difficult to assess
16/
- Age-specific CFR in FL has been going
- This is at least due, in part, to testing "finding" a larger % of sicker cases (
er risk of death) in early months- The lag btw when a person died & when it is reported makes recent trends in deaths difficult to assess
16/
