📄 public and global health

Large-scale assessment of socioeconomic, demographic and health system structures with US county excess mortality, 2020-2024

This large-scale ecological study of 2,745 US counties reveals that socioeconomic structures and demographics, rather than health system resources, were the primary drivers of excess mortality during the 2020–2021 pandemic years, with these correlations diminishing significantly in subsequent years.

Original authors: Levitt, M., Marten, B., Oren, G., Ioannidis, J.

Published 2026-10-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Levitt, M., Marten, B., Oren, G., Ioannidis, J.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

When a pandemic strikes, it does not land with equal force on every community. Some neighborhoods suffer devastating losses, while others, often just miles away, see far fewer deaths. This uneven pattern is not random; it is the result of deep, pre-existing conditions that shape how people live, work, and access care. Scientists call these conditions social determinants of health. They include factors like how much money a family earns, whether a person has a college degree, if they have health insurance, and the quality of the housing they occupy. For years, researchers have known that these factors matter, but they have usually looked at only a few at a time, trying to guess which single element was the most important. The question remained: if we could look at everything at once, what would the full picture reveal about why some places suffered more than others during the crisis?

A team of researchers set out to answer this by examining the entire landscape of American counties. They did not pick a handful of suspects to investigate; instead, they gathered a massive list of nearly three thousand different facts about every county in the United States. These facts ranged from the number of doctors and hospital beds to the percentage of residents who are unemployed, the share of the population living in poverty, and the racial makeup of the community. They then compared these facts against the number of excess deaths in each county from 2020 through 2024. Excess deaths are a specific measure that counts how many more people died in a given year than would have been expected based on the county's own history before the pandemic. By running these comparisons, the researchers could see which county characteristics were most tightly linked to the surge in deaths.

The results painted a clear and consistent picture. The strongest links to higher death rates were found in the social and economic fabric of a community, not in the medical infrastructure. Counties with higher rates of poverty, lower levels of education, and more people without health insurance saw significantly higher excess deaths. In fact, the researchers found that the share of people living in poverty was the single strongest predictor, followed closely by the percentage of adults with a college degree and the share of households earning high incomes. Even the strongest single variable, however, shared under 42% of the cross-county variation in excess death, meaning no single characteristic is close to deterministic. In contrast, the availability of medical resources, such as the number of doctors per person or the number of hospital beds, showed very weak connections to the death toll. The data suggested that having more doctors or hospital beds did not protect a community as effectively as having a population that was financially stable and educated.

This pattern was most intense during the first two years of the pandemic, 2020 and 2021, when vaccines were not yet available and the virus was most lethal. During this time, the gap between the most disadvantaged counties and the most advantaged ones was stark. The researchers created a single score that combined the strongest social factors to rank counties by their level of advantage. In the most disadvantaged fifth of the population, excess deaths were 2.0 times as high as in the most advantaged fifth (0.274 versus 0.136). As the pandemic moved into 2022, 2023, and 2024, the overall number of excess deaths dropped, and the link to these social factors weakened, though it did not disappear entirely. The study also checked its work by looking at the data in different ways, using different years for the background information and different statistical methods, and the main conclusion remained the same: the social environment was the dominant driver of the death toll.

The researchers were careful to note that their study looked at groups of people, not individuals, so it cannot prove that a specific person died because they were poor or lacked insurance. However, the sheer size of the data set and the consistency of the results suggest a powerful reality. The findings imply that the ability of a health system to save lives is deeply constrained by the conditions in which people live. A hospital cannot easily compensate for the dangers of poverty, lack of education, or the stress of financial insecurity. The study suggests that to truly understand and prevent future health crises, we must look beyond the hospital walls and address the fundamental economic and social inequalities that leave some communities more vulnerable than others. The data shows that the most effective way to protect a population may not be to build more beds, but to lift up the communities that are left behind.

Problem Statement
The COVID-19 pandemic revealed significant geographical variability in mortality across US counties, with existing literature frequently implicating poverty, race, lack of insurance, and poor healthcare access as correlates of adverse outcomes. However, previous analyses have been fragmented, typically examining only a handful of predictors. This fragmentation raises concerns that conclusions may hinge on specific analytical choices rather than robust underlying patterns. Furthermore, there is a need to comprehensively study epidemiological exposures at a large scale, extending the "omics" paradigm to social, environmental, and structural determinants, while simultaneously testing the robustness of findings against different data vintages, weighting schemes, and correlation metrics.

Methodology
The authors conducted a large-scale ecological assessment of 2,745 county-level variables against population-weighted, age-standardized excess mortality in US counties from 2020 to 2024.

Data Sources: Predictor variables were drawn from the Area Health Resources File (AHRF), utilizing raw data columns rather than pre-built composite indices. The outcome was the per-year excess-death proportion, calculated by comparing observed deaths to expected deaths derived from each county's own 2017–2019 age-specific mortality rates. Mortality data were recovered from CDC WONDER using a complement-query method to bypass suppression rules for small cell counts (1–9 deaths).
Variable Processing: The study applied a "latest-year retention" rule, selecting the most recent pre-pandemic value (median year 2018) for each variable. Variables were normalized per capita and rescaled. Missing predictor values were filled with the county mean.
Clustering: Variable names were cleaned and embedded using an MPNet sentence-transformer model. These embeddings were clustered using Ward's agglomerative hierarchical clustering, resulting in 120 semantic clusters, which were further aggregated into 11 "super-clusters" (e.g., Socioeconomic, Demographic, Health System).
Statistical Analysis: Weighted Pearson correlation coefficients (CC) were computed between each variable and the excess-death outcome for each year (2020–2024). Correlations were categorized as weak ( $|CC| \le 0.30$ ), moderate ( $0.30 < |CC| \le 0.45$ ), and strong ( $|CC| > 0.45$ ).
Robustness Checks: The study performed extensive sensitivity analyses, including:
- Using different variable vintages (data up to 2015 vs. 2023–2024).
- Varying population weighting (unweighted, square-root, 0.75 power, full population).
- Comparing Pearson vs. Spearman correlation coefficients.
- Permutation testing (10,000 permutations) to establish an empirical null distribution and control for false discoveries.

Key Results

Prevalence of Correlations: Overall, 24.1% of the 2,745 variables reached at least a moderate correlation level ( $|CC| > 0.30$ ) in at least one year, and 7.4% (202 variables) reached a strong correlation ( $|CC| > 0.45$ ).
Temporal Dynamics: The concentration of significant correlations was sharpest in 2020 (11.4%) and 2021 (14.8%). This signal diminished significantly in subsequent years (3.3% in 2022, 1.3% in 2023, and 4.1% in 2024), particularly in the under-65 population.
Dominant Factors: The strongest correlations were observed for socioeconomic and demographic variables. The top three were poverty share ( $+0.64$ in 2021), college attainment ($-0.62$ in 2021), and high income ($-0.61$ in 2021). Of the 656 variables with $|CC| > 0.30$ , the vast majority belonged to socioeconomic and demographic super-clusters.
Health System Factors: In contrast, only 10.4% of health system variables (physician counts, hospital capacity, etc.) reached the moderate correlation threshold, compared to 42% of socioeconomic/demographic variables. Only two health system variables reached a strong correlation ( $|CC| > 0.45$ ): "Number employed aged 16 and over" ($-0.50$) and "inactive female doctors" ($-0.46$).
Geographic Patterns: An "advantage index" constructed from the 200 socioeconomic and demographic variables reaching $|CC| > 0.45$ showed a monotonic gradient: the most disadvantaged counties had an excess death proportion of 0.274 in 2020–2021, compared to 0.136 in the most advantaged counties. This gradient persisted into 2022–2024, though at a reduced magnitude.
Robustness: The qualitative findings were highly robust. The leading correlates (poverty, insurance, education) remained consistent across different data vintages, weighting schemes, and correlation methods (Pearson vs. Spearman). While the magnitude of correlations scaled with the weight given to large counties, the ranking of leading variables and the contrast between socioeconomic and health system factors remained stable.

Significance and Claims
The paper claims to provide a comprehensive, "bird's-eye view" of the correlates of excess mortality, moving beyond the selection of a few specific predictors to screen an exhaustive set of nearly 3,000 variables.

Socioeconomic Gradient: The study concludes that the socioeconomic structure and demographics of a county are more strongly related to excess mortality than baseline health system supply or capacity. The authors suggest that socioeconomic disadvantage operates at a deeper root than what can be addressed by traditional hospital-level interventions.
Temporal Specificity: The strong association between socioeconomic factors and excess mortality was most prominent during the early phase of the pandemic (2020–2021), before widespread vaccination and the emergence of less lethal variants.
Ecological Nature: The authors explicitly caution against causal inference or individual-level inferences due to the ecological nature of the data (ecological fallacy). They frame their findings as hypotheses generated from ecological data rather than definitive evidence of individual risk mechanisms.
Utility: The authors suggest that these patterns are useful for preparedness planning, noting that ranking counties by hospital beds or physician supply does not effectively rank them by risk; instead, socioeconomic gradients offer a better predictor of excess death risk. They propose that such large-scale "external exposome" studies can guide targeted, specific investigations into the architecture of risk for major health outcomes.

More like this