Thread by @Sondreus

On @TheEconomist's cover this week: our model of the true death toll from covid-19 (1/23) economist.com/briefing/2021/…

In the accompanying briefing, I show that while the official death toll is 3.3m, the pandemic has probably claimed 10m lives so far (viz: @MatterofMatt, interactive: @martgnz, edits: @Eaterofsun; 2/23)

This work was the result of months of data collection and modelling, but based on a simple fact: not all who die are counted. To learn how I got there, read on (3/23)

Since I joined The Economist in February 2020, most of my work has been about the pandemic (4/23)

My very first article looked at how democracies and autocracies have historically fared differently in response to epidemics (5/23) economist.com/graphic-detail…

In early March, I argued that covid-19 was spreading undetected, using data on tourism flows and case-fatality rates to show it was massively under-reported (viz: @MatterofMatt; 6/23) economist.com/graphic-detail…

...and using data on 64,000 geo-tagged images on Instagram, that by mid-March it was already likely to be in 125 countries (viz: @_rospearce; 7/23) economist.com/graphic-detail…

By September, I estimated the true number of infected to be between 500-730m, based on 279 surveys of covid-19 antibodies around the world (viz: @MatterofMatt briefing with @JamesFransham @J_CD_T @natashaloder @SlaveaChankova 8/23)

At the time, only 32m cases were reported, making my estimate 20 times higher. Weeks later, the WHO would give a near-identical multiplier (9/23) economist.com/briefing/2020/…

I would later create tailor-made infection models for several countries, including the United States, estimating infections by week and age (10/23) economist.com/graphic-detail…

I took a deep dive into who dies because of covid-19. I first showed that based on people’s age alone, you would expect the disease to be ten times more deadly in Italy than in Uganda (11/23) economist.com/graphic-detail…

Then, with help from academic and industry experts, I was given access to a massive database of covid-19 patients. I used it to build a machine-learning... (12/23)

...model, predicting hospitalisation and death probability based on age, gender and more than 20 underlying conditions (viz: @MatterofMatt interactive: @martgnz edit: @DanRosenheck; 13/23) economist.com/graphic-detail…

All of this work would come together for this week's cover briefing: how quickly and undetected the disease spreads, the large official undercounts, and variation in who dies (14/23)

Based on work by @J_CD_T, our in-house expert on excess mortality, I created an enormous database, with countries' excess deaths when known and 121 other indicators (15/23)

I used this to find patterns between excess deaths and other data, such as demography, test positivity rates, the presence of antibodies to covid-19 in the population, and mobility (16/23)

Coding it all up was not trivial: to mention just one example, I calculated distance-weighted averages for ten indicators. With data on 210 countries and territories and 72 weeks of data, that meant over 150,000 calculations (17/23)

At this point, it is not just enough to know how to write a program to do something: you must know how write a program that does it really fast (18/23)

Altogether, the model code is 106,144 characters long, or 2,383 lines of code, including comments to help other people see and understand how it works (transparency is important) (19/23)

All this work allowed me to estimate excess deaths that were unknown. Official data suggest 3.3m have died so far: I found the most likely number to be 10m, with a likely range of 7-13m (20/23)

Those who want to read our methodology can do so here, with brilliant charts from @MatterofMatt: (21/23) economist.com/graphic-detail…

The resulting briefing and editorial contain an urgent warning: while the rich world been hit hard, most deaths have in poorer regions, and countries there remain vulnerable. More vaccines are desperately needed, and the rich world can do more to help (22/23)

For all my projects, I have uploaded full replication code and data on Github (unless I legally or ethically can't). You can find the code and data behind our excess death model here: (23/23) github.com/TheEconomist/c…

Amazing cover art: @AndreaUcini