Recent Gains in Life Expectancy Reversed by the COVID-19 Pandemic

The COVID-19 pandemic halted longevity improvements and mortality reductions at older ages in 2020. Life expectancy at birth - a widely-used indicator of population health - declined from 2019 to 2020 in 24 out of 26 countries for which high-quality vital statistics are presently available, including most European countries, Chile and the USA. Males in the USA and Bulgaria experienced the largest losses in life expectancy at birth during 2020 (2.1 and 1.6 years respectively), but staggering reductions of more than an entire year were documented in eleven countries for males, and seven among females; a magnitude of loss not witnessed since WW-II in many countries. Reductions were mostly attributable to increased mortality above age 60 and to official COVID-19 deaths.


MAIN TEXT
More than two million lives are estimated to have been lost due to COVID-19 around the world in 2020 (Dong et al., 2020). This estimate -while staggering -masks the uneven impact of the pandemic across different countries and demographic characteristics like age and sex (Pifarré i Arolas et al., 2021), as well as its impact on population health and longevity (Trias-Llimós et al., 2020). Moreover, variations in testing capacity coupled with definitional inconsistencies in counting COVID-19 deaths make the true global toll of COVID-19 infections difficult to estimate with accuracy (Karlinsky and Kobak, 2021). To address these measurement challenges, significant efforts have been directed at the harmonisation and analysis of allcause mortality data. A widely-used approach to quantify the burden of the pandemic using all-cause mortality is through the analysis of excess mortality, defined as the number of deaths observed during the pandemic above a baseline of recent trends (Pifarré i Arolas et al., 2021;Aburto et al., 2021). Here we go beyond excess deaths and country-specific analyses and focus on the pressing issue of revealing the impacts of the pandemic on life expectancy in a cross-national perspective.
Life expectancy at birth is the most widely-used metric of population health and longevity. It refers to the average number of years a synthetic cohort of newborns would live if they were to experience the death rates observed in a given period throughout their lifespan. While the indicator does not describe a cohort's actual life course (Goldstein and Lee, 2020) and should not be interpreted as the expected lifespan of any individual (see Supplemental text 2), it summarizes the mortality profile of a population in a given period (Andrasfay and Goldman, 2021;Aburto et al., 2020). As life expectancy is unaffected by the size and age structure of a population, it is the preferred indicator for comparisons over time within and across countries (Andrasfay and Goldman, 2021). Life expectancy can also be calculated as conditional on surviving to a given age, e.g. 60, where it refers to remaining life expectancy from age 60. The study of life expectancy in the context of the COVID-19 pandemic matters because it enables us to compare the cumulative impacts of the pandemic against recent trends across different countries using a standardized indicator that is routinely monitored to capture differences in mortality.
Prior to the pandemic, life expectancy at birth typically increased almost monotonically in most countries over the 20th and 21st centuries (United Nations, 2021). In recent decades, improvements were predominantly driven by gains made at older ages (65+) in high income countries (Aburto et al., 2020), although significant cross-country heterogeneity persists. This heterogeneity has become more prominent since 2010-while some countries in Eastern Europe and the Baltics experienced significant gains in life expectancy in the past decade (Aburto and van Raalte, 2018), others witnessed noticeable slowdowns in the pace of improvements, and in some cases, stalls or even temporary reversal (Ho and Hendi, 2018). For example, life expectancy in the USA (Mehta et al., 2020) and England and Wales saw only limited gains in the last decade (Leon et al., 2019). These atypical trends have been linked to slower improvements in old-age mortality, and increases in working-age death rates (Leon et al., 2019;Ho, 2013).
In a context where trajectories of life expectancy progress became more varied, the COVID-19 pandemic triggered a global mortality crisis posing additional challenges on population health. Deaths from COVID-19 vary by sex, with higher case-fatality rates among older age groups (Levin et al., 2020); precisely those that have accounted for mortality improvements in recent years. The pandemic also indirectly affected mortality from other causes of death. Emerging evidence has highlighted the impacts of delayed treatments or avoidance of care-seeking for cancers or cardiovascular diseases (Hanna et al., 2020;Wu et al., 2021) resulting in increasing mortality from these conditions, while lockdowns may have reduced the number of deaths due to accidents (Calderon-Anyosa and Kaufman, 2021).
This study is the first to use an unprecedented collection of demographic data from 26 countries, repre-  . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 2019 and 2019-2020 by broad age groups and sex (finer age grouping in Supplementary Fig. S1). Between . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 3, 2021.   2015-2019, mortality improvements above age 60 consistently contributed to increasing life expectancy across countries. In several countries, life expectancy gains above age 80 surpassed those in the age group 60-79. Improvements below age 60 contributed less, but progress was still observed in many countries, especially in Eastern Europe, with the exception of the USA and Scotland where life expectancy decreased due to worsening mortality at these ages.
The COVID-19 pandemic led to sharp declines in life expectancy (see Fig. 2), predominantly due to elevated mortality in the older age groups (see Fig. 3). Increased death rates above age 60 contributed the most to life expectancy declines across all countries between 2019 and 2020. While for females across most countries increased death rates at ages 80+ contributed most to life expectancy losses, among females in Eastern Europe worsening mortality at ages 60-79 contributed more to the decline. Among males elevated death rates at ages 60-79 contributed most to life expectancy losses across most countries, but the impact of mortality below age 60 was noticeable in many countries, especially those suffering the biggest losses in life expectancy. For example, males below age 60 in the USA contributed the most to decreasing life expectancy between 2019 and 2020.
We were able to quantify the contribution of registered COVID-19 deaths to life expectancy reductions . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.  for 13 countries (Fig. 4) for which more reliable age and sex-disaggregated data on COVID-19 deaths were available (see Methods). However, we acknowledge issues linked to the potential misclassification of COVID-19 deaths such that these estimates may not fully capture the true direct impact of COVID-19 (see Supplementary text 1). Despite these caveats, we estimated that deaths reported as COVID-19 in the COVerAGE database  explain the majority of life expectancy losses observed. Both females and males in Chile, England Wales, Belgium, Spain and Slovenia experienced losses of more than a whole year of life expectancy at birth due to COVID-19 official deaths, explaining most of the overall loss between 2019 and 2020. To put this into perspective, it took on average 5.6 years for these countries to achieve a one-year increase in life expectancy recently (Barbieri et al., 2015), a progress wiped out over the course of 2020 by COVID-19. In some contexts, such as Denmark and Chile, we note that life expectancy losses due to COVID-19 deaths were larger than total life expectancy losses, as increased mortality due to COVID-19 was offset by mortality reductions among non-COVID causes.
Our analyses show that the pandemic exacted a striking toll on population health in 2020, reversing progress in life expectancy from recent years in many countries. Emerging evidence from some developing countries (such as Brazil and Mexico) which have been devastated by the pandemic (Karlinsky and Kobak, 2021) suggests that life expectancy losses may be even larger in these populations. Similarly, losses in life expectancy are likely to vary substantially between subgroups within countries, with recent research from the US showing socially disadvantaged groups experiencing losses three times higher than those reported here at the national level (Andrasfay and Goldman, 2021; Arias et al., 2021). However, a lack of data currently limits direct and more dis-aggregated comparisons, but these are urgently needed to understand the full mortality impacts of the pandemic. With further data, researchers can examine the causes of these uneven mortality impacts of the pandemic within and across countries, including differences in underlying co-morbidities, health care systems, and non-pharmaceutical interventions. While COVID-19 might be seen as a transient shock to life expectancy, the evidence of potential long-term morbidity due to long-COVID and impacts of delayed care for other illnesses (Hanna et al., 2020;Wu et al., 2021), as well as health effects and widening inequalities stemming from the social and economic disruption of the pandemic (Bambra . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Data
All-cause death counts were retrieved from the Short Term Mortality Fluctuations (STMF) data series within the Human Mortality Database (HMD) (Max Planck Institute for Demographic Research, 2021).
Additional all-cause mortality data were retrieved for the USA from the CDC (Centers for Disease Control and Prevention, 2021) and for England Wales from the ONS (Office for National Statistics, 2021) to supplement the STMF data. STMF provides high-quality weekly death counts for 38 countries (at the time of writing) in both a harmonized and original format, but the completeness of these data varies by country. We processed the original input files and selected 24 countries for further analysis. Countries with weeks missing from the 2020 death tabulation or with coarse age grouping schemes were dropped from the analysis (see Table S1). Death counts for the USA and England Wales across the years 2015 to 2019 were only reported in extremely coarse age groups by the STMF, but were available in fine grouping for 2020.
For both countries we therefore utilized age-specific annual death counts as reported by the CDC (Centers for Disease Control and Prevention, 2021) and the ONS (Office for National Statistics, 2021) for the years prior to 2020. All-cause mortality data were supplemented with COVID-19 deaths from the COVerAGE-DB database for a subset of 13 countries with sufficient quality in their reported COVID-19 deaths and with at least 50 weeks of information . COVerAGE-DB provides information about COVID-19 confirmed cases and deaths to confirmed cases as reported by statistical agencies for over 100 countries, in a standardized format with harmonized age groups. Population estimates were retrieved from the United Nations for the years 2015-2020; and for the year 2020 we used age-specific mid-year population projections from World Population Prospects 2019 (United Nations, 2021). For England Wales, Northern Ireland and Scotland, mid-year population estimates from HMD for the year 2018 were projected for 2019 and 2020 using stable population assumptions (see Supplementary Fig. S2). We opted against official projections because the WPP does not disaggregate the UK into its constituent countries, and the oldest age group in ONS projections is 90+.

Methods
All-cause mortality data are available in irregular age groups for different countries. We harmonized data from STMF, CDC, ONS and COVerAGE-DB using a penalized composite link model (PCLM) which estimated death counts in single years of age from 0 to 110 from the grouped data (Rizzi et al., 2015;Pascariu et al.). The PCLM model was fitted independently to each country, sex, and year combination, and the smoothing parameters were estimated via a Bayesian Information Criteria based grid search. Prior to ungrouping we summed the weekly death counts from the STMF data into annual death counts, also taking into account deaths reported in unknown weeks. In cases where the age grouping changed from one week to the next, we first summed all the deaths within a stratum and year belonging to the same age grouping scheme, applied PCLM ungrouping separately for each scheme and then summed the ungrouped deaths into annual counts by single years of age.
Person years of exposure were approximated by estimated or projected mid-year population counts and used in the denominator of the age-specific death rates. Exposure estimates were adjusted for leapweeks as most countries in the STMF data report deaths using the ISO week date calendar (International Organization for Standardization, 2021). The length of a year in the ISO week calendar is either 371 days in . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 3, 2021. ; https://doi.org /10.1101/2021.03.02.21252772 doi: medRxiv preprint a leap-week year such as 2020, or 364 days in a regular year such as 2019. Thus, the longer reporting interval for leap-week years would, other things equal, increase observed death counts by a factor of 371/364=1.9 over regular years. To counter this spurious effect we adjusted the exposures by the factor 371/364 in ISO week date leap years. This adjustment has been made for all countries analyzed in this paper with the exception of annual deaths prior to 2020 for the US and England Wales, which have been reported over a Gregorian year by the CDC and ONS sources.

Life table construction
Life tables for all 26 countries by sex for the period 2015-2020 were constructed following a piece-wise constant hazard model using all-cause mortality by single age with the last age interval grouping deaths at ages 85+, consistent with standard demographic techniques (Preston et al., 2001). Our choice of grouping deaths above age 85+ is common practice in demographic research, and it ensures comparability across countries, especially because some countries used this last age group to report deaths for 2020.  Fig. S4). Most countries did not show deviations across sources, with the exception of Sweden and Austria, where the STMF underestimates death counts varying from 2% and 4% for females and males, respectively. We further calculated life expectancy using the death counts from both sources and WPP population estimates as exposures and found that the effect on life expectancy is negligible (less than 1%). For the USA, we also compared our life expectancy estimates with those based on provisional data up to June 2020 (Arias et al., 2021). Our results show a sharper decrease of life expectancy than those based on mid-year estimates, in line with including the remaining months of 2020 that also saw persistent levels of excess mortality.

Decomposition of life expectancy by age and cause of death
In order to disentangle age-specific effects and to quantify the impact of COVID-19 deaths on changes in life expectancy, we used the linear integral decomposition method (Horiuchi et al., 2008), a state-of-the-art method that allows us to decompose the difference of two values of life expectancy by age and cause of death, which has been implemented previously for this type of analysis (Aburto and van Raalte, 2018).
This methodology assumes that causes of death are exhaustive and independent. This assumption may not be realistic in the context of COVID-19, as the pandemic may have indirectly affected other causes of death. However, previous evidence suggests that the net effect of interactions between causes of death are negligible in decomposition analysis (Beltrán-Sánchez et al., 2008). Moreover, our results are likely to be unaffected by this assumption since most COVID-19 deaths are excess deaths. In addition, in most countries the contribution of COVID-19 deaths to changes in life expectancy can be interpreted as a lower bound due to late registration and testing strategies for COVID-19. We decompose yearly changes in life . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 3, 2021. ; https://doi.org /10.1101/10. /2021 expectancy by age and deaths registered as COVID-19 versus the rest of deaths for each country by age and sex. This methodology has the advantage that the resulting estimates are additive and the total change in a given period, such as 2015-2019, is the sum of yearly estimates from 2015 to 2019. As a sensitivity check we replicated our results with step-wise decomposition method (Andreev et al., 2002). Results from this sensitivity check show that most of the changes were a result of increased mortality at older ages as our method reported.

Limitations
There are several limitations of our approach. While most of the countries included in our analysis have high-quality systems of death registration and usually report timely and complete death counts (Barbieri et al., 2015), there are likely issues related to delays in death registration in 2020. We deal with these potential issues by harmonizing and smoothing death counts with a penalized composite link model which has been shown to provide reliable estimates of mortality for all countries in our study (Rizzi et al., 2015), and by adjusting the person-years exposed to the risk of death depending on the full-weeks reported in the STMF. In most STMF countries, deaths that take place in a given week are covered by more than 90% by the following two weeks (STMF metadata). Therefore, we expect the percentage of missing deaths in the last weeks of 2020 to be low. Moreover, according to the SMTF metadata, some countries (such as Austria, England and Wales, Germany, and the Netherlands) employ extrapolation procedures to correct for incompleteness to adjust for missing deaths counts in the preceding 1-to-3 weeks. We recognise nonetheless that late and/or under registration may affect our estimates by underestimating losses in 2020.
Therefore our results should be interpreted as estimates of the lower bound of life expectancy reductions in 2020. For COVID-19 official mortality figures, there are likely to be inaccuracies due to differences in classification and testing practices across countries. While we are unable to precisely measure the extent to which these inaccuracies may affect our estimates to make adjustments (i.e., since misclassification can occur for different reasons), we err on the side of caution and interpret our results as a lower bound of the impact of COVID-19 deaths on changes in life expectancy.
The magnitude of life expectancy losses can also be influenced by migration due to especially in countries where persistent out-migration flows are not tracked accurately (via population registers) between censuses. For example, in Eastern European countries with high levels of out-migration in their populations, increased mortality could have occurred as a result of elevated return migration due to the COVID-19 pandemic. Nevertheless, we anticipate that the magnitude of these effects is likely to be small in size due to the broadly younger age profiles of migrants and previous evidence that suggests that mortality at younger ages is more likely to be affected by international migration (Rogers and Castro, 1981). In 2020, as increased mortality occurred mostly at older ages, the effect of migration on our estimates is likely to be low. Despite these limitations, we used the most reliable data for all countries included in the study to gauge the contributions of age and COVID-19 to changes in life expectancy across countries, sex, and over time.

Interpretation of life expectancy
Life expectancy is a widely-used but often misinterpreted indicator of a population's mortality profile that is computed using a mathematical model. The model employed to calculate life expectancy is known as life table. As with any model it is based on assumptions. Life expectancy at birth is the average number . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted March 3, 2021. ; https://doi.org /10.1101/10. /2021 of years a synthetic cohort of newborns would live if along the course of their lives they were to experience the age-specific death rates observed in a given time or period. This indicator is thus often referred to as 'period life expectancy', as it simulates and summarizes the implications of a current period. This is in contrast to a cohort measure that would track the longevity of a group of people born at the same time.
The key advantage of period life expectancy, which results in its routine use and monitoring, is that it is an age-standardized indicator that allows comparisons between populations of different sizes and age structures, and across space and over time. Period life expectancy is not to be interpreted as a projection nor a forecast, and neither does it describe a cohort's actual life course. It is purely a summary measure of the age-specific mortality profile of a population. Data and materials availability: The replication files for this paper include customised functionality written in the R statistical programming language (R. Core Team, 2013). The code, and all harmonized input and output data pertaining to our analysis, is hosted both on Zenodo (a general-purpose open-access repository developed under the European OpenAIRE program and operated by CERN) at https://zenodo.org/record/4556982, and on GitHub (https://github.com/oxforddemsci/ex2020). Table S1: Data quality of raw all-cause and Covid-19 death data. Note: Source data retrieved 2021-02-26 from HMD-STMF, CoverAge-DB, UK-ONS, and US-CDC. All cause death counts from STMF unless stated otherwise in the notes. Countries noted in italics were excluded from analysis. Countries in bold are part of COVID-19 analysis. Number of age groups refers to at least 90% of the weeks in the input data for a given year. 'Miss' denotes missing weeks in 2020.
. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 3, 2021. ; https://doi.org /10.1101/10. /2021 Losses|Gains in life expectancy at birth, years . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 3, 2021. ; https://doi.org /10.1101/2021.03.02.21252772 doi: medRxiv preprint  . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 3, 2021.  2015'16 '17 '18 '19 2015'16 '17 '18 '19 2015'16 '17 '18 '19 2015'16 '17 '18 '19 2015'16 '17 '18 '19 2015'16 '17 '18 '19 2015'16 '17 '18 '19  year total deaths | Our divided by HMD Female Male Figure S4: Comparison of death counts between the Short Term Mortality Fluctuations dataset and the Human Mortality Database.
. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted March 3, 2021. ; https://doi.org /10.1101/10. /2021