TL;DR
This paper tackles the challenge of accurately estimating COVID-19 infections in the US, where official case counts are significantly underestimated due to testing limitations. By employing four distinct, complementary methods that leverage influenza surveillance, testing data, and mortality records, the authors provide a more robust picture of the pandemic's true burden, revealing that actual symptomatic cases were 5-50 times higher than reported in early April 2020.
What Problem Does This Paper Address?
Accurately tracking the spread of infectious diseases like COVID-19 is crucial for effective public health responses, resource allocation, and forecasting. However, the early stages of the COVID-19 pandemic in the United States were plagued by widespread testing shortages and varying testing capabilities across regions. This meant that official reported positive case counts were a significant undercount of the true number of infections. Furthermore, a substantial portion of COVID-19 cases are asymptomatic or paucisymptomatic, making them even less likely to be detected by testing. This paper aims to overcome these limitations by developing and applying multiple methods to estimate the cumulative incidence of symptomatic COVID-19, providing a more reliable assessment of the pandemic's true scale.
Methods
The authors propose four complementary approaches to estimate COVID-19 incidence:
- Divergence (div-Hist & div-Vir): These methods leverage existing influenza-like illness (ILI) surveillance systems. They work by establishing a "control" time series representing what ILI activity would have been without COVID-19, based on historical data (div-Hist) or influenza virology data (div-Vir). The unexpected increase in ILI above this control is then attributed to COVID-19.
- COVID Scaling: This approach combines ILI data with COVID-19 testing results. It extrapolates state-level positive test percentages to estimate the proportion of ILI visits that are due to COVID-19, and then scales this to the entire population.
- Mortality Mapping (mMAP): This method uses reported COVID-19 deaths to infer past true case counts. It employs a time series deconvolution technique, accounting for the time lag between symptom onset and death, and uses the symptomatic case fatality rate (sCFR) to estimate symptomatic cases. An adjusted version (mMAPadj) also incorporates excess influenza and pneumonia deaths to account for unreported COVID-19 deaths.
- Global Epidemic and Mobility Model (GLEAM): This is a sophisticated, individual-based, stochastic, and spatial epidemic modeling platform. It simulates the spread of COVID-19 using real-world data on population demographics, contact patterns, and mobility (including airline and commuting data). The model is calibrated using reported deaths and then adjusted to estimate symptomatic cases.
The paper also discusses "adjusted" versions of these methods, which incorporate newer information or account for factors like increased healthcare-seeking behavior or lower estimated fatality rates, aiming for more accurate estimates.
Datasets
The study utilizes several key datasets:
- Influenza-like Illness (ILI) Surveillance Data: Primarily from the CDC's US Outpatient Influenza-like Illness Surveillance Network (ILINet).
- COVID-19 Case and Death Counts: Sourced from the New York Times repository (for US states) and the Johns Hopkins University COVID-19 dashboard (for international validation).
- COVID-19 Testing Data: Daily time series of positive and negative test results obtained from the COVID Tracking Project.
- US Demographic Data: Age-stratified, state-level population numbers from 2018 US Census estimates.
- Mobility Data: Google's COVID-19 Community Mobility Reports and data from OAG and IATA databases for airline traffic.
- Influenza Virology Data: From the CDC virologic surveillance system.
Key Insights and Ideas
- The Power of Complementarity: No single data source or method is perfect. By using multiple, diverse approaches, the authors build confidence in their findings, as consistent conclusions emerge across methods with different assumptions and data inputs.
- ILI Surveillance as an Early Warning System: Even with limitations, ILI surveillance can provide valuable insights into COVID-19 incidence, especially in the early stages of an outbreak, by detecting deviations from expected patterns.
- Mortality as a More Reliable Indicator: COVID-19 deaths, while also subject to reporting delays and biases, can offer a less noisy estimate of cases compared to testing data, particularly for severely ill individuals who are more likely to be tested.
- The Scale of Underestimation: The most striking insight is the vast difference between reported cases and estimated true incidence, highlighting the significant challenge of tracking pandemics with limited testing.
Key Results
- Massive Underestimation: As of April 4, 2020, the authors' estimates indicated that the cumulative symptomatic COVID-19 cases in the US ranged from 2.3 to 4.8 million, with a potential upper limit of 7.6 million. This is 5 to 50 times higher than the approximately 311,000 officially reported cases at that time.
- National Estimates: Nationally, by May 16, 2020, estimated cumulative symptomatic incidence ranged from 4.9 to 10.1 million, compared to 1.5 million reported positive test counts.
- State-Level Variability: The degree of underestimation varied significantly by state, with some locations showing estimated case counts over 50 times higher than reported.
- Method Agreement: Despite differences in their approaches, the adjusted methods showed a notable clustering of estimates, increasing confidence in the overall findings.
Limitations and Future Work
- ILI-Based Method Limitations: The Divergence and COVID Scaling methods are dependent on the ILI definition (fever, cough, sore throat) and may miss symptomatic cases that don't meet these criteria. They are also limited by the reliability of ILI surveillance systems, which can be affected by factors like routine healthcare use changes and overwhelming hospital capacity.
- Testing Bias: The COVID Scaling method assumes uniform testing, which may be biased by factors like prior exposure, potentially leading to an overestimation of cases.
- Mortality Data Delays: The mMAP method is affected by delays in death reporting.
- Data Availability: For some states (e.g., Florida), ILI data was not available, limiting the applicability of certain methods.
- Future Work: The authors suggest expanding their approaches to include other data sources (e.g., Google searches, electronic health records) and leveraging serological testing for more accurate incidence estimation. They also highlight the potential of existing ILI surveillance systems for monitoring future pandemics with similar symptoms.
Summary updated