The COVID-19 pandemic, caused by SARS-CoV-2, presented significant challenges in accurately assessing its true burden. Early in the pandemic, testing resources in the U.S. were limited, primarily focusing on individuals exhibiting moderate to severe symptoms. This approach inevitably led to a substantial underestimation of the total number of infections, as a significant portion of infected individuals were asymptomatic or mildly symptomatic. Existing confirmed case counts, therefore, provided an incomplete picture of the pandemic's spread. The study aimed to address this limitation by developing a robust method for estimating the total number of SARS-CoV-2 infections in the U.S., accounting for both incomplete testing and imperfect diagnostic accuracy. Accurate estimations are vital for effective public health strategies, resource allocation, and understanding the true dynamics of the virus transmission. Prior studies relying on mathematical models, while valuable, often faced challenges due to model complexity, data limitations, and sensitivity to assumptions regarding population structure and contact patterns. This study sought a more direct and data-driven approach using probabilistic bias analysis to correct for the inherent biases in the available data.

Existing studies estimating the burden of SARS-CoV-2 infection primarily used mathematical models (compartmental or agent-based). While these models attempted to incorporate various factors like social structure and interventions, they faced limitations in accuracy due to complexity and data scarcity. The sensitivity of model outputs to assumptions about population structure and contact patterns also posed a challenge, especially given the novelty of the virus. The authors highlighted the need for a method that directly addresses the biases in empirical data, offering a more reliable estimate of the actual infection burden without relying on potentially flawed modeling assumptions.

The study employed a semi-Bayesian probabilistic bias analysis to estimate the total number of SARS-CoV-2 infections in the U.S. between February 28 and April 18, 2020. This approach corrected the empirical confirmed case counts for biases arising from incomplete testing and imperfect test accuracy. The methodology involved defining prior distributions for several key parameters, including probabilities of testing positive among individuals with different symptom severities (moderate to severe vs. mild or no symptoms) and the sensitivity and specificity of the diagnostic tests. These prior distributions were informed by available evidence from various studies on SARS-CoV-2 testing and symptom presentation. The study used a Monte Carlo simulation approach to generate a distribution of estimated infections, accounting for the uncertainty inherent in the prior distributions. The simulation involved multiple iterations to generate a probability distribution of estimated infections, providing not just a point estimate but a range of plausible values, including a simulation interval (2.5th and 97.5th percentiles) to quantify the uncertainty. The model considered the number of confirmed cases, the number of individuals tested, and probabilities of testing positive for different symptom categories to estimate the number of untested infected individuals. Additionally, it adjusted for imperfect test accuracy by incorporating test sensitivity and specificity. Sensitivity analyses were conducted by exploring alternative plausible scenarios and prior distributions for critical parameters to assess the robustness of the findings. The analysis was conducted at both the national and state level, allowing for a geographical assessment of infection underestimation.

The study's key finding was a substantial underestimation of SARS-CoV-2 infections in the U.S. during the study period. By April 18, 2020, the estimated total number of infections was 6,454,951, approximately nine times higher than the 721,215 confirmed cases. This translates to an estimated 1.9% of the population being infected, versus the 0.2% represented by confirmed cases. The 95% simulation interval for the estimated infections indicated a range of 3 to 20 times higher than the confirmed cases. The analysis attributed 84% (simulation interval: 64–99%) of this difference to incomplete testing, and 16% (simulation interval: 0.3–36%) to imperfect test accuracy. State-level analyses revealed wide disparities in both confirmed cases and estimated infections, with significant regional variations. The degree of underestimation varied considerably across states, highlighting the importance of geographically targeted interventions. The maps show the high variation of COVID-19 cases among the U.S states. Sensitivity analyses, exploring alternative prior distributions, generally yielded robust results, confirming the reliability of the study's approach. Figures displayed testing rates, the comparison between confirmed cases and estimated infections, and a map illustrating the geographical distribution of both confirmed cases and estimated infections across U.S. states.

The findings strongly emphasize the need to adjust COVID-19 infection estimates to account for testing limitations and diagnostic accuracy. The study's strength lies in quantifying the respective contributions of incomplete testing and imperfect test accuracy to the underestimation, showing that incomplete testing was the major factor. The methods used are adaptable to other settings and scales, making them valuable for future pandemic preparedness and response. The estimates of true infection numbers at the state level differ substantially from the confirmed cases, indicating the inadequacy of relying solely on confirmed cases for public health decision-making. The consistency of findings under different prior distributions strengthens the robustness of the approach. However, the study acknowledges limitations in data availability, particularly regarding state-specific parameters, which could influence the precision of the estimates. The study’s data-driven approach addresses gaps from previous studies relying on modeling.

This study provides compelling evidence of a significant underestimation of SARS-CoV-2 infections in the U.S. early in the pandemic. The semi-Bayesian probabilistic bias analysis presented a robust method to correct for biases caused by incomplete testing and imperfect test accuracy. The findings underscore the urgent need for widespread testing to inform effective pandemic response strategies. Future research should explore further refinement of the methodology using more granular data and explore the implications of these findings for modeling pandemic spread and implementing public health policies.

The study acknowledges that its accuracy is dependent on the quality and availability of data. The prior distributions used for certain parameters were based on limited evidence, particularly concerning the testing probabilities for individuals with mild or no symptoms. Variations in testing capacity and guidelines across states could affect the accuracy of state-specific estimates. The lack of accurate age-stratified data on COVID-19 deaths hampered the precise estimation of infection fatality rates (IFRs).