This repository contains the replication code and data for The Economist's excess deaths model, used to estimate excess deaths due to the covid-19 pandemic.

Before running the R scripts, please install all the dependencies listed in this file.

Keep in mind that you need to use the development version of agtboost, which has been rewritten to load machine learning ensembles faster (our approach requires loading 200 of these).

To install it, first install the devtools package:

install.packages('devtools')

And then install the development version from GitHub:

devtools::install_github("Blunde1/agtboost/R-package")

To update the model dynamically on a daily basis, go to the scripts folder and run 0_excess_deaths_global_estimates_autoupdater.R. This will generate update excess deaths estimates for every country and territory from Jan 1st 2020 until the present.

To replicate the model and export estimated excess deaths for a locality, please run the scripts 1, 2, and 3, in the scripts folder. As the model draws most of its data dynamically, you can also use these scripts to generate updated estimates and models as time passes.

• To read about what the model shows, see our Briefing: Counting the dead.
• To understand how we constructed it, see our Methodology: How we estimated the true death toll of the pandemic..
• To see all the improvements we have made since, as well as all our estimates, updated daily, see our Interactive: The pandemic’s true death toll

A special thanks to all our sources and to those who have made the data to create these estimates available. We list all our sources in our methodology. Within script 1, the source for each variable is also given as the data is loaded, with the exception of our sources for excess deaths data, which we detail in on our free-to-read excess deaths tracker. The gradient booster implementation used to fit the models is aGTBoost, detailed here.

Calculating excess deaths for the entire world over multiple years is both complex and imprecise. We welcome any suggestions on how to improve the model, be it data, algorithm, or logic. If you have one, please open an issue.

The Economist would also like to acknowledge the many people who have helped us refine the model so far, be it through discussions, facilitating data access, or offering coding assistance. A special thanks to Ariel Karlinsky, Philip Schellekens, Oliver Watson, Lukas Appelhans, Berent Å. S. Lunde, Gideon Wakefield, Johannes Hunger, Carol D'Souza, Yun Wei, Mehran Hosseini, Samantha Dolan, Mollie Van Gordon, Rahul Arora and Austin Teda Atmaja.