The COVID-19 pandemic has had overwhelming global impacts with deleterious social, economic, and health consequences. To assess the COVID-19 death toll, researchers have estimated declines in 2020 life expectancy at birth (). When data are available only for COVID-19 deaths, but not for deaths from other causes, the risks of dying from COVID-19 are typically assumed to be independent of those from other causes. In this research note, we explore the soundness of this assumption using data from the United States and Brazil, the countries with the largest number of reported COVID-19 deaths. We use three methods: one estimates the difference between 2019 and 2020 life tables and therefore does not require the assumption of independence, and the other two assume independence to simulate scenarios in which COVID-19 mortality is added to 2019 death rates or is eliminated from 2020 rates. Our results reveal that COVID-19 is not independent of other causes of death. The assumption of independence can lead to either an overestimate (Brazil) or an underestimate (United States) of the decline in , depending on how the number of other reported causes of death changed in 2020.
More than 6.5 million deaths were attributed to COVID-19 worldwide by mid-November 2022, a number that is likely underestimated. The two countries with the largest number of reported COVID-19 deaths are Brazil and the United States, with almost 698,000 and more than 1.09 million deaths reported, respectively, as of early February 2023. The impact of COVID-19 on life expectancy at birth () in these countries has received widespread attention from the media: a decline of 1.3 years in Brazil and 1.8 years in the United States for 2020 (Castro et al. 2021; Murphy et al. 2021). Because overall COVID-19 mortality increased in 2021 in both countries, declines in compared with 2019, were even larger than in 2020 (Andrasfay and Goldman 2022; Arias et al. 2022; Castro et al. 2021).
To examine the impact of COVID-19 on mortality, researchers have used numbers of deaths and population by age to calculate declines in (the metric reported here) during a specified period of the pandemic (e.g., 2020 relative to 2019). Life expectancy has numerous advantages over other mortality measures, including its interpretability, comparability over time and place, and the fact that, unlike the death rate, it is unaffected by the age distribution. Estimates of during the pandemic for both the United States and Brazil have varied across studies primarily because the National Center for Health Statistics (NCHS) and the Ministry of Health (MoH), respectively, have periodically released updated and corrected information. However, as we demonstrate in the following, another important source of variation stems from differences in the methods used to calculate , often driven by limited data availability.
Government agencies in the United States and Brazil generally release mortality data after a substantial lag to ensure completeness and quality. However, in 2020, both the NCHS and MoH released information on COVID-19—but not other—deaths shortly after receipt, permitting researchers to calculate the impact of the pandemic on without data from all death records. From mid-2020 until November 2021, preliminary data for all causes of death (CoD) started to be released every two weeks in Brazil. In November 2020, the United States began releasing preliminary all-cause mortality data, but with irregular frequency. Estimates of the impact of COVID-19 on and years of life lost (YLL) based on data from only COVID-19 deaths required the assumption that the risks of dying from COVID-19 were independent of the risks of dying from other causes (i.e., the presence of COVID-19 did not alter the risks of dying from any other cause) (Chan et al. 2021). Although this assumption is unrealistic, it pervades the literature—both before the pandemic and currently—primarily because of challenges in directly assessing dependence among causes. Previous applications of the assumption of independence among CoD typically involved determining the consequences of the hypothetical elimination or deletion of a long-standing cause, frequently a chronic disease such as cancer (Beltrán-Sánchez et al. 2008; Ho 2013; Yashin et al. 1986). In contrast, the sudden arrival of COVID-19 provides a natural experiment that permits a direct assessment of the plausibility of the independence assumption, in this case between COVID-19 and other CoD, in the context of adding rather than eliminating a cause. We explore the consequences of this assumption using data for the United States and Brazil.
Deaths from all causes by month, age, sex, and CoD for Brazil were extracted from the Mortality Information System (MoH) for 2018–2020, and population projections by age and sex were drawn from the Brazilian Institute of Geography and Statistics (Castro et al. 2021). For the United States, corresponding data were obtained from CDC Wonder (U.S. Department of Health and Human Services et al. 2021).
The methods used to calculate the decline in are described in the online appendix. We used three approaches. First, we constructed period life tables (LT) for 2019 and 2020 that considered all CoD by sex and age group and calculated the difference between in 2019 and 2020 (LT19 – LT20). This method does not make assumptions about independence among causes or depend on the accurate classification of causes; however, the estimate would include reductions and increases in numbers of deaths not related to the COVID-19 pandemic. The other approaches simulate hypothetical scenarios in which COVID-19 mortality for 2020 is added to 2019 death rates (DT19) or is eliminated from 2020 rates (DT20); both assume independence between the risks of dying from COVID-19 and those from other causes (Chiang 1968). We estimated reductions in due to COVID-19 from the difference between the 2020 period LT and the DT20 estimates, and between the 2019 period LT and the DT19 estimates.
We performed all analyses in R version 4.0.0 (R Core Team 2021). We created data visualizations in R and Adobe Illustrator CS6.
On the basis of LT19 – LT20, the United States lost 1.51 years, compared with 1.41 years in Brazil. Estimates of the change in derived from DT19 and DT20 under the assumption of independence are larger than those from LT19 – LT20 for Brazil but smaller for the United States (Table 1). Estimates of changes in and are shown in the online appendix Table S1.
Between 2019 and 2020, age-specific mortality rates from all causes combined excluding COVID-19 increased for most age groups in the United States, with substantial rises in the young adult and middle-aged groups (Figure 1, panel a). In Brazil, however, rates increased only modestly in the young adult and middle-aged groups and declined at younger and older ages. The decline at younger ages in Brazil was mostly driven by a reduction in deaths due to influenza, pneumonia, and chronic and other lower respiratory diseases, likely a result of physical distancing and school closure.
Both the United States and Brazil experienced large increases in diabetes mortality between 2019 and 2020, considerably more than between 2018 and 2019 (Figure 1, panel b, and online appendix Figure S1). Both countries, especially Brazil, experienced declines between 2019 and 2020 in cancer mortality. In contrast, Brazil generally faced declines in the following causes between 2019 and 2020, whereas the United States generally saw increases: heart and cerebrovascular diseases; influenza, pneumonia, and other respiratory diseases; and external causes. Not all of these changes were necessarily a result of the pandemic (e.g., a potential secular decline in mortality), but most were likely due directly or indirectly to COVID-19.
Results show that DT19 and DT20 underestimated the decline in for the United States and overestimated the decline for Brazil relative to LT19 – LT20. As shown in Figure 1, COVID-19 is not independent of other CoD. Because mortality rates from other causes generally increased between 2019 and 2020 in the United States, the assumption of independence underestimates the overall change in mortality, whereas the opposite occurs in Brazil.
There are many potential reasons for the dependence between risks of dying from COVID-19 and those from other causes. Because COVID-19 fatality increases in the presence of comorbidities (e.g., cancer, heart disease, diabetes), rates from these chronic ailments might have decreased if patients succumbed to COVID-19, a phenomenon related to “harvesting” (Schwartz 2000). Death rates from some external causes, such as violence and travel-related accidents, may have declined because of reduced social and work activities. In contrast, mortality rates from non-COVID-19 causes may have increased because of worsening of comorbidities owing to the effects of COVID-19; delays in primary and preventative care and reduced disease management for non-COVID-19 conditions; inadequate care in clinics and hospitals due to shortages of equipment, staff, and space; increased mortality from other causes due to long COVID-19; higher rates of smoking, drinking, drug use, and poor nutrition; reduced exercise; and potential health consequences from job loss, financial difficulties, and reduced social ties (Dey and Davidson 2021; Griffin 2021).
Our results are likely affected by the quality of reported death information, and the quality may vary by country. Misidentification or miscoding of CoD—particularly between COVID-19 and other respiratory conditions prior to widespread testing for the virus—could have inflated or reduced numbers of COVID-19 deaths. Unclassified deaths could also have biased the estimates, although there has been little change in the relative size of this category between 2019 and 2020 (5.8% in Brazil for both years, and 1.1% and 1.0% in the United States in 2019 and 2020, respectively). Underreporting of deaths could have affected the age-specific pattern of both COVID-19 and non-COVID-19 mortality. Underreporting was estimated at 1.4% in 2019 for Brazil (Instituto Brasileiro de Geografia e Estatística (IBGE) 2019) and less than 1% in 2015–2018 for the United States (Karlinsky 2021) but could have increased during the pandemic. Both miscoding and underreporting may be selective by CoD, age, and sex.
When all CoD data were not yet available, several attempts to determine the overall effect of the COVID-19 pandemic on relied on the use of multiple decrement methods, which assume independence between COVID-19 and other CoD. These methods, however, have serious drawbacks. Our results underscore that the assumption of independence could either underestimate or overestimate the overall loss in . Therefore, once made available, the use of all-cause mortality data to obtain estimates of should be prioritized.
T.A. acknowledges funding support from the National Institute on Aging under award T32AG000037. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
M.C.C. and N.G. conceived the research. M.C.C., N.G., C.M.T., and T.A. defined the methodology. T.A., S.G., and S.K. wrote the codes for analysis. M.C.C., N.G., T.A., and C.M.T. interpreted the data. M.C.C., T.A., and S.G. conducted data curation. M.C.C. and S.K. produced all visualization. N.G. and M.C.C. wrote the first draft of the manuscript. All authors contributed to the interpretation of results and manuscript editing.
The data and code required to reproduce the results presented in this manuscript are available at https://github.com/mcastrolab/Covid-19_competing_risk.