Estimating Covid-19 mortality from official statistics on all-cause mortality
Analysis by
Andrea Galeotti (https://andreagaleottiblog.wordpress.com/)
Sebastian Hohmann (https://sebastianhohmann.github.io/)
Paolo Surico (https://sites.google.com/site/paolosurico/)
Additional contributors
Luis Fonseca (https://luispfonseca.com/) for the analysis of Portugal
Riccardo Trezzi for the analysis of Italy
Thanks
Indrek Mossov for pointing us to the Swedish data
Marino van Zelst (https://www.marinovanzelst.com/) for pointing us to the Dutch data
Special thanks to Mark Rabkin (https://twitter.com/mrabkin) for pointing us to the Spanish data and helping us understand their limitations
Objective
This project aims at estimating Covid-19 mortality from official statistics on total deaths. We aim to constantly update the analysis as new data becomes available. We also aim to add robustness checks and alternative estimates in an effort to accurately determine Covid-19 deaths. The results of the analysis are regularly updated in this article.
We would like to replicate the analysis for as many countries as possible. If you have data or you know where we can access data in a country, please share them with us. We are happy to replicate the analysis, share the results with you, and update the analysis so that other people can access this information. If you have the data and you want to perform the analysis yourself, in what follows you will find all material you need. However, we would appreciate if you share with us the results, with a summary of how you conducted the analysis. In particular, with explicit reference to changes in the codes or methodology. This is to assure that results are comparable across countries.
Methodology
We estimate Covid-19 mortality for several countries. We always use two inputs:
(i) Official numbers for total mortality (not just Covid-19) in a country.
(ii) Official numbers for total Covid-19 mortaliy.
We carry out the following steps:
1) Taking as the starting point the day of the year before the first Covid-19 death, we compute, for each of the years 2015-2020, the cumulative total of deaths.
2) Using the cumulative totals, we compute daily growth rates for each year 2015-2019.
3) We take the average of the growth rates of the previous five years. (An earlier version of our work had reversed steps 2) and 3), that is rather than computing growth for each of the past five years separately and then averaging, we first averaged the cumulative totals and computed the daily growth from this. This change alters our results only very marginally. Please see the commit history of this repository for the earlier versions).
4) We fix the number of people that died on the day before the first Covid-19 death was recorded in 2020 and extrapolate this number forward using the average growth rate computed in 3). This yields what we call the "counterfactual" for 2020.
5) We add the official number of Covid-19 deaths from (ii) to the counterfactual created in 4).
6) The difference between the actually observed total number of deaths in (i) and the counterfactual in 4) is our estimate for Covid-19 mortality. The difference between (i) and 5) are excess deaths in 2020 that are not designated official Covid-19 deaths.
Please see the folders for each country for examples on how this method is implemented in python.
WARNING ON INTERPRETATIONS OF THIS ANALYSIS. It is crucial to appreciate that in the vast majority of countries the number of officially recorded Covid-19 deaths provided by the government necessarily reflects mostly (if not only) deaths that occur in hospitals (and similar structures). A main reason is that those are the places where governments are concentrating their efforts and thus most tests have been so far conducted. As such, it is very hard for any government to keep track of all Covid-19 deaths and thus produce in real-time an accurate aggregate estimate of the actual number of total deaths most likely associated with Covid-19. The analysis in the article and the codes that we are making publicly available at in this repository develop a very simple and naive tool. Our work is not meant to replace rigorous epidemiological modelling of the number of excess deaths due to Covid-19 and we are not epidemiologists ourselves. Rather, this naive tool aims to provide a simple and transparent rule-of-thumb to appreciate how unusual the months of March and April 2020 might be. We welcome criticism and discussion of our naive approach and results. To aide in this, we are making all aspects of our analysis public.