Table 1.
Main features of the construction of the compared evaluations of excess deaths.
| eLife | Lancet | Economist | WHO | Levitt | |
|---|---|---|---|---|---|
| Reference period years | 2015–2019 | 2010 (or earliest available)-February 2020 | Unclear, not mentioned | 2015–2019 (countries with monthly historical data); 2000–2019 (country with annual historical data) | 2017–2019 |
| Modeling of reference period | Linear fit | Ensemble of 6 models (weighted): 4 using splines with different placement of the last knot, one Poisson, and one taking 2019 only | Machine learning. Mix of boosted Gradient, Random Forest and Bootstrapping. | Sum of an annual trend (thin-plated spline) and a within-year seasonal variation (cyclic cubic spline) | Static average |
| Exclusions | Heat waves | Heat waves | Unclear, not mentioned | Not mentioned | No |
| Time unit of modeling data | Weekly (preferred) or monthly or quarterly | Weekly or monthly | Weekly for most, some monthly | Monthly | Weekly |
| Pandemic time period covered in the original publication/release | Varies per country, mostly 2020 to mid-2021, exact start in 2020 depends on availability of weekly (week 10), monthly (March), or quarterly (January) data | 2020–2021 (acknowledged potential problem with late registration for last weeks/months) | 2020 to late 2021 | 2020–2021 (had also released early estimates for 2020) | 2020–2021 |
| Pandemic time period covered in the current comparative analysis | 2020–2021 | 2020–2021 | 2020–2021 | 2020–2021 | 2020–2021 |
| Source of data for all-cause mortality | Human Mortality Database, others | World Mortality Database, Human Mortality Database, European Statistical Office | World Mortality Database, Human Mortality Database, others | Eurostat, Human Mortality Database, World Mortality Database | Human Mortality Database |
| Source of data for COVID-19 deaths used in original paper | Johns Hopkins | Apparently Johns Hopkins (although too high for Spain and UK) | Unclear | Not used | Johns Hopkins |
| Source of data for COVID-19 deaths used in the current comparative analysis | Johns Hopkins | Johns Hopkins | Johns Hopkins | Johns Hopkins | |
| Age adjustment | No | No (authors stated that they may adjust for age in future work) | No | Yes (excess deaths summed across 7 age strata) | Yes (excess deaths summed across 5 age strata), also done without age-adjustment |
| Gender adjustment in calculations | No | No | No | Yes | No |
| Any other adjustment | No | Under-registration corrected for countries with <95% death registration | Probably no (unclear) | No | No |
| Eligibility criteria for countries modeled directly | Weekly, monthly or quarterly data available for at least one pre-pandemic year and for pandemic period | Weekly or monthly data available for any pre-pandemic years and for pandemic period | Data availability (unclear about details) | Data availability (Age and sex specific death for 2020 aggregated to 5-year age bands), excluding the countries that have experienced conflict, small population numbers, incomplete deaths and/or erratic/implausible age-pattern | Weekly data available in Human Mortality Database from 2017 onwards |
| Number of countries modeled directly | 103 in the publication. 77 with data to December 2021 | 74 countries and territories in the publication | 78 countries apparently had mortality data, but it seems that all countries were included in the machine learning | 50 | 36 |
| Eligibility criteria for countries inferred from the directly modeled countries | None | All countries considered | Unclear | All countries. All data for 2021 were inferred | None |
| Number of countries inferred from the directly modeled countries | None | Remaining world | Remaining world | Remaining world | None |
| How were they inferred? | Not applicable | LASSO regression, selected 15 covariates related to pandemic (e.g. seroprevalence) and to background population health metrics (e.g. Healthcare Access and Quality Index) | Machine learning as above; totally impossible to reproduce based on thinly presented information, 121 indicators considered | K-mean clustering. Countries are divided into 5 clusters with different values of, Human Development Index Mean age at death, Crude excess rate | Not Applicable |