The Economist's excess deaths model
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.
Getting started
Before running the R scripts, please install all the dependencies listed in this file.
Keep in mind that you need to use a modified development version of agtboost, which has been rewritten to load machine learning ensembles faster (our approach requires loading 210 of these).
To install it, first install the devtools package:
install.packages('devtools')And then install the development version from GitHub:
devtools::install_github("sondreus/agtboost/R-package")Running the models
To update the model dynamically on a daily basis, run scripts/0_excess_deaths_global_estimates_autoupdater.R (from the main directory). 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
Interpreting estimates
Estimating excess deaths for every country every day since the pandemic began is a complex and difficult task. Rather than being overly confident in a single number, limited data means that we can often only give a very very wide range of plausible values. Focusing on central estimates in such cases would be misleading: unless ranges are very narrow, the 95% range should be reported when possible. The ranges assume that the conditions for bootstrap confidence intervals are met. Please see our tracker page and methodology for more information.
New variants
The Omicron variant, first detected in southern Africa in November 2021, appears to have characteristics that are different to earlier versions of sars-cov-2. Where this variant is now dominant, this change makes estimates uncertain beyond the ranges indicated. Other new variants may do the same. As more data is incorporated from places where new variants are dominant, predictions improve.
Non-reporting countries
Turkmenistan and the Democratic People's Republic of Korea have not reported any covid-19 figures since the start of the pandemic. They also have not published all-cause mortality data. Exports of estimates for the Democratic People's Republic of Korea have been temporarily disabled as it now issues contradictory data: reporting a significant outbreak through its state media, but zero confirmed covid-19 cases/deaths to the WHO.
Acknowledgements
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 as well as on GitHub. 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 Ã…. S. Lunde, Gideon Wakefield, Johannes Hunger, Carol D'Souza, Yun Wei, Mehran Hosseini, Samantha Dolan, Mollie Van Gordon, Rahul Arora, Austin Teda Atmaja, Dirk Eddelbuettel and Tom Wenseleers.
All coding and data collection to construct these models (and make them update dynamically) was done by Sondre Ulvund Solstad. Should you have any questions about them after reading the methodology, please open an issue or contact him at [email protected].
Suggested citation
The Economist and Solstad, S. (corresponding author), 2021. The pandemic’s true death toll. [online] The Economist. Available at: https://www.economist.com/graphic-detail/coronavirus-excess-deaths-estimates [Accessed ---]. First published in the article "Counting the dead", The Economist, issue 20, 2021.
(See also our citation file)