Thread by @zorinaq

I dusted off my COVID-19 model (that predicted the Florida July wave) & applied it to Sweden

After today's data update from the Swedish Public Health Agency (FHM) I confidently forecast Sweden will surpass the peak of 100 COVID deaths/day they had in April

Hard to believe?

1/n

Specifically: by 25 December we will see Sweden has recorded 100 deaths/day around 11 December

(due to reporting delays, it takes up to 2 weeks past a given date to have a complete count of deaths on this date: )

2/n

My model is formally described in outbreak.flashpub.io/pub/method-of-…

It predicts deaths from cases alone, but let me explain in layman's terms how it works...

3/n

The key insight that made the model so accurate so far is that the Infection Fatality Ratio of COVID-19 is highly dependent on age:

Eg. the probability of dying for a 60-year-old is 10 times higher than for a 40-year-old

4/n

So a huge spike in cases among 20-40-year-olds will not significantly inflate deaths

This is in part why the lag between a rise in cases and a rise in deaths can be many weeks

5/n

FHM only gives the cumulative number of cases per 10-year age group since the beginning of the pandemic in their data file (arcgis.com/sharing/rest/c…)

This is not enough to forecast deaths

6/n

We need the age of cases 𝗼𝘃𝗲𝗿 𝘁𝗶𝗺𝗲.

The only way to get this is to make archives of the FHM data over time. Adam Altmejd maintains such an archive: github.com/adamaltmejd/co…

7/n

These archives give us cases by age group over time.

Note how despite cases rising well above their April peak for many weeks, it's only 𝘁𝗵𝗶𝘀 week that cases among ages 80+ reached the same level as April, see also this heatmap by @jocami_ca

8/n

We then multiply the known age-stratified Infection Fatality Ratio of COVID-19 with the number of cases by age group

This gives us the estimated number of deaths ~3 weeks from today, as the mean infection-to-death time is 22.9 days (see pg. 4 of static-content.springer.com/esm/art%3A10.1…)

9/n

It's important to not confuse IFR & CFR

CFR has decreased 𝘀𝗶𝗴𝗻𝗶𝗳𝗶𝗰𝗮𝗻𝘁𝗹𝘆 over time but it's an artifact of the improving case ascertainment rate

IFR has remained relatively constant over time—studies suggest just slight 10-20% decline:

10/n

My source for the age-stratified IFR is the geometric mid-point of 13 independent studies: github.com/mbevand/covid1…

11/n

With all that taken into account, we find that the forecast actually underestimates deaths

This is because not all cases are detected. We need to adjust FHM data to account for undetected cases. IOW we need to find the case ascertainment rate.

12/n

Through trial and error I found the case ascertainment rate was about 50% a month ago, but has been declining to about 35%, meaning Sweden now detects 1 in 3 cases.

35% isn't great but is supported by reports of heavily strained testing capacity:

13/n

So with a 35% case ascertainment rate + the IFR mid-point of 13 studies, I confidently project:

Sweden will hit 100 COVID deaths/day around 11 December. This will surpass the 7-day MA of 98.7 deaths/day of 17 April

14/n

ourworldindata.org/coronavirus-da…

It is sad that the state epidemiologist at FHM apparently does not see this sharp rise in deaths coming.

How can me, a nobody, see it, but not an expert?

15/n
dn.se/sverige/tegnel…

Of note: even if cases were to drop to ZERO tomorrow, Sweden is still expected to hit 100 deaths/day before Christmas

My model only infers deaths from past cases. It does not assume case growth must continue

16/n

Regarding strained testing capacity: FHM confirmed it in their status report for week 46, published today:

«The capacity to test for covid-19 continues to increase, but the demand for testing in several regions is currently greater than the supply»

folkhalsomyndigheten.se/globalassets/s…

17/n