Overview
- Published in Science Advances, the peer-reviewed analysis used machine learning on 5.7 million CDC death certificates for adults 25 and older from March 2020 through December 2021.
- The model estimates 155,536 likely unrecognized COVID deaths, lifting the two-year toll to about 995,787 versus 840,251 recorded, with the largest monthly shortfall in January 2021 at roughly 35,000.
- Missed fatalities were concentrated outside hospitals, especially at home, where the predicted toll was about 160% higher than what death certificates recorded.
- Undercounting disproportionately affected Hispanic, American Indian/Alaska Native, Asian and Black people, as well as those with lower incomes and less education, and was most pronounced in the South.
- States with the highest undercount ratios included Alabama, Oklahoma and South Carolina, while large absolute gaps appeared in populous states such as California (about 11,600), Texas and New York; the authors caution the model cannot prove COVID as the cause for each flagged death.