Study Points to Massive Undercount of Early COVID Deaths in the U.S.

Official U.S. counts of COVID-19 deaths likely missed over 150,000 unrecognized mortalities during the pandemic’s first 2 years, a machine learning model study suggested.

By the newer estimate, there were 995,787 deaths caused by COVID from March 2020 to December 2021, nearly 20% more than the official U.S. tally of 840,251 deaths with COVID listed anywhere on a death certificate during that period (adjusted reporting ratio [ARR] 1.19, 95% uncertainty interval [UI] 1.18-1.19), according to Andrew Stokes, PhD, of Boston University, and colleagues.

Those unrecognized COVID deaths were more likely among older people, non-whites, lower-income people, and residents of certain regions of the country, the researchers reported in Science Advances, noting that “the U.S. death investigation system reported COVID-­19 deaths in a systematically inequitable way that hid the true extent of pandemic mortality and inequities.”

“If we improve death investigation systems and the quality of our national death surveillance data, we’ll be much better prepared to detect the next pandemic earlier and more fully,” Stokes told MedPage Today. “We can also apply these machine learning methods to understand other causes of death that have long been underestimated, such as suicide mortality, mortality from Alzheimer’s disease and dementia, diabetes, drug overdoses, and a host of other things.”

Prior research had relied on excess mortality models to estimate the number of unrecognized coronavirus deaths. Studies using those models, which compare observed all-cause deaths with expected prepandemic trends, estimated excess mortality was higher than reported COVID deaths by 14% to 38% in 2020 alone.

By one count, excess deaths from 2020 to 2021 reached a staggering 18.2 million worldwide, compared with the 5.9 million officially attributed to COVID-19.

However, those models can’t tease out which deaths were uncounted COVID deaths rather than indirect pandemic effects such as healthcare interruptions and delays, Stokes said. Excess mortality models’ use of highly aggregated data also makes detailed subgroup analyses difficult.

To sharpen estimates, he explained, the machine learning algorithm used the “silver standard” of correctly classified in-hospital COVID deaths (where SARS-CoV-2 testing was nearly universal during the pandemic’s early stages) to predict whether out-of-hospital deaths (where such testing wasn’t universal) were likely COVID-19 related.

Based on this approach, many of the predicted unrecognized COVID deaths outside of hospitals were attributed to causes such as Alzheimer’s disease, cardiovascular disease, or diabetes, Stokes noted.

The researchers found that unrecognized COVID deaths were most likely in the pandemic’s initial wave, during March to May 2020 (ARR 1.49, 95% UI 1.48-1.51), then fell in subsequent waves. January 2021 and April 2020 saw the largest numbers of predicted unrecognized COVID deaths, at 35,665 and 32,110, respectively.

States with the greatest likelihood of estimated unrecognized COVID deaths included Alabama (ARR 1.67), Oklahoma (ARR 1.51), and South Carolina (ARR 1.47). States with the largest absolute numbers of estimated unrecognized deaths were Texas (24,024), New York (23,005), California (11,613), Alabama (11,501), and Florida (7,718).

Hispanic ethnicity (ARR 1.31, 95% UI 1.30-1.32) was tied to more estimated unrecognized deaths, as was male sex (ARR 1.22, 95% UI 1.21-1.23) and older age: ages 65-74 years (ARR 1.21, 95% UI 1.21-1.22), ages 75-84 years (ARR 1.22, 95% UI 1.21-1.23).

Other sociodemographic factors associated with undercounted COVID deaths included:

• West South Central region residence (ARR 1.31, 95% CI 1.29-1.33)
• Middle Atlantic region residence (ARR 1.26, 95% UI 1.24-1.27)
• Less than a high school education (ARR 1.29, 95% UI 1.28-1.31)
• Counties in the lowest quintile of median household income (ARR 1.34, 95% UI 1.31-1.36)
• Counties with more residents reporting poor or fair health (ARR 1.30, 95% UI 1.28-1.33)

“The communities affected by the undercounting of COVID-­19 deaths could be interpreted as a pattern of structural racism, classism, and ableism in the death investigation system that warrants further research and policy attention,” Stokes and colleagues concluded.

Study limitations include the assumption that the machine learning model’s training data set of inpatient deaths assigned to COVID-19 was correct.

Please enable JavaScript to view the comments powered by Disqus.