Aging and COVID-19 mortality: A demographic perspective

Demographers were among the first to emphasize the importance of age in explaining the spread and impact on mortality of COVID-19. Yet, the relationship between age and COVID-19 is not fully understood. This study aims to answer the following questions: Do COVID-19 deaths follow the Gompertz force of mortality? How does the relationship between age and COVID-19 mortality compare with other major causes of death? Does this relationship vary between countries? Using US vital Statistics, COVID-19 mortality doubling time is compared with 68 major cause of death categories. COVID-19 fatality is similarly compared across 33 countries using COVerAGE-DB, a harmonized dataset of confirmed COVID-19 cases and deaths. Several findings are supported by the evidence. First, COVID-19 mortality increasing exponentially with age at a rate near the median of aging-related causes of death, as well as pneumonia and influenza. Second, COVID-19 mortality levels in the US are currently 4 to 8 times higher than pneumonia and influenza across the adult age range. Third, the relationship between COVID-19 fatality and age varies considerably across high-income countries. Fourth, these regularities can be utilized for indirect estimation of COVID-19 deaths by age. In conclusion, the relationship between COVID-19 mortality and age resembles the population rate of aging. Between-country variation in the Gompertz slope of COVID-19 fatality may point to differences in underlying population health, standards of clinical care, and data quality.


Introduction
There is mounting clinical evidence that the risk of dying from Coronavirus disease 2019  is associated with age (e.g., Zhou et al., 2020). Clinicians noted early on that COVID-19 fatality rises steadily with age; unlike other respiratory diseases, it did not present the typical U-shaped curve of heightened risk among both infants and older adults (Raoult et al., 2020). The World Health Organization (2020) and the US Centers for Disease Control and Prevention (2020a) have both defined older adults-individuals aged 60 and 65 and older, respectively-as a vulnerable group. In response, as the COVID-19 pandemic broke out, governments worldwide recommended or instructed older adults to remain in social isolation for a period lasting weeks or months (Brooke & Jackson, 2020). These simplified accounts of the relationship between COVID-19 and age may have contributed to the portrayal of COVID-19 as a "disease of the elderly," in the US and elsewhere, which some have argued was a key factor in the failure to contain the outbreak (Germani et al., 2020;Nowroozpoor et al., 2020). Nearly one year into the pandemic, there is now sufficient demographic data to establish key facts about the relationship between age and COVID-19 mortality.
Demographers were among the first to point out the importance of age in explaining COVID-19's spread and impact on mortality (e.g., Dowd et al., 2020;Dudel et al., 2020;Kashnitsky & Aburto, 2020). Yet, the relationship between age and COVID-19 is not fully understood. Do COVID-19 deaths follow the Gompertz force of mortality? How does the relationship between COVID-19 and age compare with other major causes of death? Does this relationship vary between countries? Can these regularities be utilized for indirect estimation of COVID-19 deaths by age? This study aims to establish these basic demographic facts using US vital statistics and data on confirmed COVID-19 cases and deaths in 33 OECD and high-income countries.
Several findings are supported by the empirical evidence. First, like many agingrelated causes of death, COVID-19 mortality follows the Gompertz law and rises exponentially with age. Second, it progresses with age at a rate similar to the median aging-related cause of death, and similar to seasonal mortality from pneumonia and influenza. COVID-19 mortality in the US, however, is currently 4 to 8 times higher across the adult age range than annual pneumonia and influenza mortality in the past 20 years. Third, the relationship between age and COVID-19 fatality varies considerably across OECD and high-income countries. Based on these descriptive findings, I suggest a simple method to apportion the total number of COVID-19 deaths by age, and apply it to Israel as a case in point.

Data and methods
Already in 1825, Benjamin Gompertz noted that beyond early adulthood, all-cause mortality rises exponentially with age (Missov et al., 2015): The increase in mortality rate with age, captured in the γ coefficient, has often been equated with the population rate of aging (Ricklefs & Scheuerlein, 2002). The rate of aging can also be expressed as the mortality rate doubling time (MRDT), the average number of years in which the mortality rate doubles: (2) = ln (2) While the interpretation of the Gompertz slope as the rate of aging has been criticized for conceptual reasons (de Grey, 2005), the Gompertz hazard function remains a useful model for senescent mortality (Gavrilov & Gavrilova, 2019) as well as for the relationship of specific causes of death with age (Juckett & Rosenberg, 1993). A larger Gompertz slope coefficient implies, if the model fits well, a steeper progression of mortality with age. Conversely, causes of death that progress rapidly with age will have low mortality rate doubling time, whereas those that progress slowly with age will have high doubling time.
First, using US vital statistics data from 2018 (Centers for Disease Control and Prevention, 2020b), I estimated the mortality rate doubling time of 113 underlying cause of death categories for ages 30 and over. I retained in the analysis only those causes of death for which the Gompertz model was a good fit (R 2 > 0.95), resulting in 68 major causes of death, and compared them with the COVID-19 mortality doubling time estimated from provisional US deaths counts (National Center for Health Statistics, 2020b).
Second, I compared COVID-19 age-specific mortality rates with those of Pneumonia and Influenza (P&I) between 1999 and 2018. Adjusting the exposure time is critical when comparing mortality rates because COVID-19 is an ongoing pandemic (Heuveline & Tzen, 2020). For this reason, I adjusted the P&I annual mortality rates to the same calendar period as the COVID-19 provisional death countsepidemiological weeks 6 through 32-using historical data from the National Center for Health Statistics (2020b).
Third, I examined whether the US age pattern can be generalized to other countries, because prior research suggests that COVID-19 mortality may vary by geographic regions and viral strain (Brufsky, 2020;Wilson et al., 2020), among other factors. Drawing on COVerAGE-DB, a harmonized cross-national database of confirmed COVID-19 cases and deaths by age , the age-specific case-fatality rate-i.e., the number of deaths per confirmed COVID-19 cases in each age groupcould be estimated for 33 OCED and high-income countries, as currently defined by the World Bank (2020). The case-fatality doubling time (FRDT) was then estimated using equation (1), substituting case-fatality rates for mortality. Although the two measures are not directly comparable, their relationship with age will be similar if attack (infection) rates do not vary markedly by age. If the FRDT and MRDT differed, . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10. 1101 it might indicate that certain age groups are disproportionally likely to either become infected or to be diagnosed, a phenomenon known as surveillance bias (Haut, 2011).
Lastly, I demonstrate how the Gompertz model can be used for indirect estimation of COVID-19 deaths by age when only the total number of deaths and age-specific exposures (diagnosed cases) are known. The cumulative number of COVID-19 deaths by country is widely available (Dong et al., 2020), whereas age-specific counts are far more scarce, depend on what governmental institutions are willing to share, and are often incomplete . This was, for example, the case of Israel, who's Ministry of Health publicly released age-specific COVID-19 exposures but suppressed the number of deaths by age due to confidentiality concerns. By borrowing information from other countries, the Gompertz model provides a simple method for estimating the number of deaths by age.
Assuming that age-specific COVID-19 fatality rates follow equation (1), the number of deaths as a function of age is where c(x) and f(x) are, respectively, the number of diagnosed cases and fatality rate as functions of age x, and f0 is the baseline fatality rate. The expected number of deaths by age can then be estimated using equation (3) for an observed total number of deaths, D, and a specified slope parameter, ̂, solving for f0: Values for ̂ can be borrowed from a reference country or a distribution, with f0 serving as a rescaling parameter to match the observed total number of deaths. In essence, this method amounts to assuming that COVID-19 fatality rates follow a model age schedule, weighted by observed age-specific exposures.

How does the relationship between COVID-19 mortality and age compare with other causes of death?
The mortality rate doubling time of major cause-of-death categories is shown in Figure 1. Age is a strong predictor of each of those causes because poorly fitting models have been excluded. In 2018, the MRDT varied in the US from 3.2 years for Alzheimer's disease (more age dependent) to 21.1 years for Asthma (less age dependent), with a median of 7.3 years across all 68 cause of death categories, excluding COVID-19. For example, the MRDT was estimated at 5.3 years for chronic lower respiratory diseases; 6.6 years for acute myocardial infarction; 7.0 years for malignant neoplasms; 8.0 years for diabetes; and 10.9 years for kidney infections. The COVID-19 MRDT, based on US provisional data through August 2020, is estimated at 7.6 years. In other words, it is near the median doubling time for all . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10. 1101 aging-related causes of death. Moreover, the relationship between COVID-19 mortality and age is not markedly different from that of Pneumonia & Influenza, which in 2018 had an estimated MRDT of 7.7 years.
[ Figure 1 here] 3.2 How does COVID-19 mortality compare with Pneumonia & Influenza? Figure 2 compares the age-specific COVID-19 mortality with a 20-year series of P&I mortality. Since COVID-19 is an ongoing epidemic, annual P&I mortality rates have been adjusted to the equivalent exposure period based on historical seasonal data (i.e., the proportion of P&I deaths occurring between epidemiological weeks 6 through 32). From 1999 to 2018, the P&I mortality-rate doubling time averaged 7.4 years, ranging from 6.8 to 8.9 years over the period. In other words, the relationship between COVID-19 mortality and age is similar to the multiyear average P&I mortality; the proportional increase in COVID-19 mortality with age is well within P&I's inter-annual variation. However, Figure 2 also shows that COVID-19 absolute mortality levels are substantially higher than P&I-it is about 4 to 5 times higher in the age groups 25-34 and 75-84, and as much as 7.6 times higher for those aged 55-64 years.

How does the relationship between COVID-19 fatality and age vary across countries?
The third objective was to determine whether the relationship between COVID-19 mortality and age varies across countries. In the absence of comparable mortality data, age-specific fatality rates were compared across 33 OECD and high-income countries. The results are shown in Figure 3. In most countries the Gompertz model fit the data well (median R 2 was 0.96). The FRDT varied from 5.3 years in the Czech Republic to 13.7 years in Mexico, with a median FRDT of 6.7 years and a mean of 7.4 years. It is worth noting that the FRDT in the US was slightly higher than the MRDT (8.2 compared to 7.6 years, respectively). The findings therefore reveal considerable variation in the association between COVID-19 case-fatality and age between countries, and that in the US the FRDT is higher than the MRDT by 0.6 years.

Indirect estimation of COVID-19 deaths by age
COVID-19 mortality appears to follow the Gompertz law in most countries and this regularity can be utilized to estimate the expected number of deaths by age. Israel, as a case in point, is one of the countries in COVerAGE-DB that reported the number of confirmed COVID-19 cases, but not deaths, by age. Nevertheless, the total number of deaths is published daily by the Israel Ministry of Health. As of August 22, a total of 819 COVID-19 deaths have been reported (Israel Ministry of Health, 2020); only a handful of those, 4 according to media accounts, occurred below age 30. Based on . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10.1101 the number of exposures by age and the total number of deaths, the expected number of deaths by age can be estimated for a range of Gompertz slope values. The steeper the slope, the more peaked the distribution of deaths at older ages. Figure 4 depicts three such scenarios for the 10th, 50th, and 90th percentiles of the Gompertz slope estimated across 33 countries. The median scenario, for example, assumes that Israel had the same Gompertz slope as Belgium, with deaths rescaled to Israel's known total and weighted by its age-specific confirmed cases. The 10th and 90th percentiles correspond, respectively, to Uruguay and Austria, providing an interval estimate for the expected number of deaths at each age (a more realistic interval estimate might account in addition for sampling variability and stochasticity).

Discussion
It is hardly news that the risk of mortality increases with age-a relationship that has only strengthened throughout the epidemiologic transition (Salomon & Murray, 2002). COVID-19 mortality is no different, rising exponentially with age. That COVID-19 follows the Gompertz law of mortality, however, is in no way a given. The Spanish Flu, in sharp contrast, wrought havoc among 20 to 40 year-olds, whose mortality was substantially higher than their elders (Ma et al., 2011). This study finds that COVID-19's relationship with age is approximately log-linear over 30 and its slope sits near the median of aging-related causes of death. Furthermore, the relationship between age and COVID-19 mortality resembles that of annual pneumonia and influenza mortality over the past 20 years; COVID-19 mortality levels, however, are currently 4 to 8 times higher in the US across the adult age range.
Taking a comparative approach, the rate in which COVID-19 fatality progresses with age varies considerably across OCED and high-income countries. While the Gompertz model was a good fit in most countries (R 2 > 0.92 in 27 out of 33 countries), its slope ranged from 0.51 to 0.13. This variation may reflect differences in the prevalence of risk factors, standards of clinical care, readiness of healthcare systems, or data quality and reporting practices (Carinci, 2020;Pasquariello & Stranges, 2020). In the US, the only country for which fatality and mortality could be compared in this study, the case-fatality doubling times was slightly higher than the mortality doubling time (8.2 compared with 7.6 years). This discrepancy may point to differences in attack rates by age (or, alternatively, to surveillance bias), because mortality appears to rise more sharply with age than case-fatality.
Although chronological age is one of the most widely available demographic attributes, it is important to note, as other have pointed out, that biological aging mechanisms are more predictive of COVID-19 mortality (Mueller et al., 2020). Individuals age at different rates and chronological age often masks substantial variation in frailty (Mitnitski et al., 2002). Treating COVID-19 as a "disease of the . CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10.1101 elderly" may foster ageist or paternalistic views and fuel intergenerational tensions against the backdrop of a global health crisis (Ayalon et al., 2020). As with many aging-related causes of death, there is no natural age cutoff to distinguish between low and high risk of COVID-19 mortality. Any such cutoff would be invariably arbitrary and depend on societal norms regarding accepted levels of risk (Glynn, 2020). The findings in this study highlight the need to better understand the relationship between age and COVID-19 mortality, and why it varies so widely between countries.

Acknowledgments
The author thanks Alex Weinreb, Alyson van Raalte, Liat Ayalon, and My Hedlin for helpful comments.
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10.1101/2020.10.15.20213454 doi: medRxiv preprint  is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020.  is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10.1101  Note: Scenarios based on 10th, 50th, and 90th percentiles of the Gompertz slope estimated across 33 OECD and high-income countries.
. CC-BY 4.0 International license It is made available under a perpetuity.
is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted October 18, 2020. . https://doi.org/10.1101