This forecast update (2020-07-26) has some major changes:
1/N
I scrapped model 1 (redundant with model 2) and model 3 (was the least accurate of the 5 models).
All my curves are now charted with a simple moving average instead of *centered* moving average. Less confusing.
I now produce a BEST GUESS estimate, pretty tight interval.
2/N
Notice how has an extremely wide confidence interval (dotted pink lines): by August 12 it estimates between 70 and 250 deaths/day. Whereas my best guess interval is 150-200.
So how does the best guess estimate work?
3/N
The best guess estimate is based on the age-stratified CFR from model 5, that is the Case Fatality Ratios calculated on Florida cases.
I believe model 5 most closely predicts the rate of increase/decrease of daily deaths. At least experimentally I've seen that, see chart:
4/N
However model 5 slightly underestimates deaths for reasons I've explained before. I have not yet found the time to fix it properly, so for now my best guess is a quick hack that should still produce decent results
Also my forecast can predict 18 days ahead of time. I used to lose a day (predicting 17 days) because I ignored the last day of data as it contains incomplete data. Now I've committed a proper workaround (github.com/mbevand/florid…)
6/N
Finally, I've also made my charts colorblind friendly, using various linestyles (dashed, dotted, etc)
7/N
Let's point out numerous obvious flaws in his analysis, shall we?
1/n
Firstly, Berenson chooses to ignore Asian countries. Why? Because geography is "the most important factor in how hard Covid hits a country"
In other words: "countries who did better than Sweden don't count, because, well, I sAy So"
Such well-reasoned logic. Much wow 🤣
2/n
Secondly, he claims geography is the most important factor, but IGNORES ALL the countries geographically close to Sweden: Finland, Norway, Denmark (none of them are charted)
I modeled excess deaths per capita by age group, for each US state, since the beginning of the pandemic.
I believe this is the first time an analysis of this type has been done BY AGE GROUP for each state. This removes the need to do "age-adjustment" to compare states.
1/n
Based on NYT's mask survey across United States counties, plot deaths/capita recorded during the survey & up to 30 days later, along with each county's mask wearing score
Result:
Linear regression (Y log-transformed) R²=0.144 in red:
There are hundreds of factors affecting the dependent variable (deaths). Ignoring ALL of these factors, looking at mask usage only, and still finding R²=0.144 is pretty cool/unexpected
Confounders abound!: people who wear masks often are likely doing more social-distancing. Etc.
Methodology behind the chart:
Two data sources:
- mask wearing survey github.com/nytimes/covid-…
- COVID deaths (and population) by US county as per JHU CSSE
His charts seem to claim that nothing works. Locking down doesn't work, masking doesn't work, vaccination doesn't work, your printer doesn't wo—wait scratch that one
1/n
One iota of critical thinking is all you need to expose numerous errors in his charts:
Error #1 — Case ascertainment rate bias:
A country may detect 1 in 2 cases, while another 1 in 4. We say the case ascertainment rate is respectively 50%, and 25%.
This variance in case ascertainment rate alone is enough to put half of @ianmSC's charts where they belong: in the trash🗑️
Real Science™ looks at covid deaths—not cases—to compare the severity of the pandemic across different regions. This avoids case ascertainment rate bias.
I compiled a list—as exhaustive as possible—of all peer-reviewed & published research articles that evaluate the effectiveness of nonpharmaceutical interventions, specifically lockdowns on COVID-19
➡️Papers finding NPIs effective outnumber, by 8 to 1, those finding the opposite
Criteria for inclusion in the list:
1-Be a RESEARCH ARTICLE (data, methods, results). Commentaries, opinion pieces, etc, do not qualify
2-Be PEER-REVIEWED & PUBLISHED among the 26,000 titles in Scopus
3-Be EXPLICIT. No secondhand interpretation of the data
Regarding criterion #3: the authors must explicitly state in the text whether their results suggest NPIs are effective or not
Their exact words have been peer-reviewed & published. Your interpretation of figures or data tables has not.
2. President of Burundi died on 8 June 2020. The cause of death was given officially as "cardiac arrest" by the Burundian government, but is suspected to be COVID: economist.com/middle-east-an…
His wife was flown to Kenya and hospitalized for COVID a week before his death.