State-level tracking of COVID-19 in the United States

H. Juliette T. Unwin, Swapnil Mishra, Valerie C. Bradley, Axel Gandy, Thomas A. Mellan, Helen Coupland, Jonathan Ish-Horowicz, Michaela Vollmer, Charles Whittaker, Sarah Filippi, Xiaoyue Xi, Mélodie Monod, Oliver Ratmann, Michael Hutchinson, Fabian Valka, Harrison Zhu, Iwona Hawryluk, Philip Milton, Kylie E. C. Ainslie, Marc Baguelin, Adhiratha Boonyasiri, N Brazeau, Lorenzo Cattarino, Zulma M. Cucunubá, Gina Cuomo-Dannenburg, Ilaria Dorigatti, Oliver Eales, Jeffrey W. Eaton, Sabine van Elsland, Richard G. FitzJohn, Katy A. M. Gaythorpe, William H. Green, Wes Hinsley, Benjamin Jeffrey, Edward Knock, Daniel J. Laydon, John A. Lees, Gemma Nedjati‐Gilani, Pierre Nouvellet, Lucy Okell, Kris V. Parag, Igor Siveroni, Hayley A Thompson, Patrick Walker, Caroline E. Walters, Oliver J. Watson, Lilith K. Whittles, Azra C. Ghani, Neil M. Ferguson, Steven Riley, Christl A. Donnelly, Samir Bhatt, Seth FlaxmanmedRxiv 2020 · DOI

Abstract

As of 1st June 2020, the US Centers for Disease Control and Prevention reported 104,232 confirmed or probable COVID-19-related deaths in the US. This was more than twice the number of deaths reported in the next most severely impacted country. We jointly modelled the US epidemic at the state-level, using publicly available death data within a Bayesian hierarchical semi-mechanistic framework. For each state, we estimate the number of individuals that have been infected, the number of individuals that are currently infectious and the time-varying reproduction number (the average number of secondary infections caused by an infected person). We used changes in mobility to capture the impact that non-pharmaceutical interventions and other behaviour changes have on the rate of transmission of SARS-CoV-2. Nationally, we estimated 3.7% [3.4%-4.0%] of the population had been infected by 1st June 2020, with wide variation between states, and approximately 0.01% of the population was infectious. We also demonstrated that good model forecasts of deaths for the next 3 weeks with low error and good coverage of our credible intervals.

Introduction

In the early months of the COVID-19 pandemic, confirmed case counts were an unreliable metric for the true scale of the outbreak. Severe testing shortages and inconsistent state-level reporting meant that many infections went unrecorded. To provide a clearer picture for policy decisions, researchers developed a framework to estimate the "true" number of infections, the current infectious population, and the time-varying reproduction number ($R_t$). This approach accounts for the lag between infection and death and the impact of both formal interventions and voluntary behavioral changes.

Methodology

The study utilizes a Bayesian hierarchical semi-mechanistic framework to model the epidemic across all 50 U.S. states and the District of Columbia.

Latent Infection Process

The model treats infections as a latent (unobserved) process, estimated using a discrete renewal process. It back-calculates infections from observed death data using state-specific Infection Fatality Ratios (IFRs) adjusted for age distribution.

Modeling Transmission ($R_t$)

A key innovation is parameterizing $R_t$ as a function of mobility data (from Google Community Mobility Reports). To account for behavioral shifts not captured by physical movement—such as mask-wearing or social distancing—the model includes a second-order autoregressive [AR(2)] process. This allows the model to capture "residual" transmission changes that are decoupled from mobility.

Partial Pooling

The hierarchical structure enables "partial pooling," where information is shared across states. This strengthens estimates in regions with smaller outbreaks or noisier data while still respecting state-specific differences in epidemic timing and intervention impact.

Key Insights

The "Hidden" Epidemic and Heterogeneity

The national estimate of 3.7% cumulative infection masks significant regional variation. While states like New York and New Jersey reached attack rates of 15-16%, many others remained below 1%. This highlights that by June 2020, the vast majority of the U.S. population remained susceptible.

Mobility vs. Behavior

The inclusion of the AR(2) term revealed that mobility alone is insufficient to explain transmission dynamics. For instance, in some states, $R_t$ dropped faster than mobility changes would predict, suggesting early voluntary behavioral shifts. Conversely, in others, transmission spiked just before lockdowns, potentially due to increased social density during "panic buying" phases.

Testing Gaps

The model estimates the Infection Ascertainment Ratio (IAR)—the proportion of true infections that are reported as cases. Mean IAR varied significantly by state (from 43% to 74%), indicating that even as testing capacity grew, a substantial portion of the epidemic remained undetected.

Results

By June 1, 2020, the model estimated that approximately 12.1 million people (3.7% of the U.S. population) had been infected. However, the epidemic was far from contained in many regions.

Probability that the reproduction number $R_t$ is less than 1 across US states as of June 1, 2020.

As shown in the map above, $R_t$ was estimated to be below 1.0 in only 23 states. In more than half of the country, the probability of $R_t < 1$ was low, suggesting the epidemic was still expanding or plateauing.

Estimated infection rates and ascertainment ratios by state as of June 1, 2020.

The table above details the state-level heterogeneity in infection rates and the effectiveness of testing (IAR). Nationally, only about 0.01% of the population was estimated to be actively infectious on June 1.

Predictive Performance and Limitations

The model demonstrated strong predictive validity, achieving a 3-week median cumulative death Mean Absolute Percentage Error (MAPE) of 9.9%. This suggests the mechanistic assumptions regarding transmission and death lags were robust for short-term forecasting.

However, several limitations remain:

  • IFR Sensitivity: Estimates are highly dependent on the assumed IFR. Shifts in the age profile of the infected population (e.g., toward younger individuals) would alter these results.
  • Social Mixing: In the absence of U.S.-specific contact surveys, the model assumed mixing patterns similar to the UK.
  • Collinearity: The high correlation between different mobility metrics (e.g., residential vs. workplace) makes it difficult to isolate the exact impact of specific behaviors.

Future iterations of this work will likely require age-specific modeling to better capture the evolving demographics of the pandemic.

Summary updated

Abstract

Abstract As of 1st June 2020, the US Centers for Disease Control and Prevention reported 104,232 confirmed or probable COVID-19-related deaths in the US. This was more than twice the number of deaths reported in the next most severely impacted country. We jointly modelled the US epidemic at the state-level, using publicly available death data within a Bayesian hierarchical semi-mechanistic framework. For each state, we estimate the number of individuals that have been infected, the number of individuals that are currently infectious and the time-varying reproduction number (the average number of secondary infections caused by an infected person). We used changes in mobility to capture the impact that non-pharmaceutical interventions and other behaviour changes have on the rate of transmission of SARS-CoV-2. Nationally, we estimated 3.7% [3.4%-4.0%] of the population had been infected by 1st June 2020, with wide variation between states, and approximately 0.01% of the population was infectious. We also demonstrated that good model forecasts of deaths for the next 3 weeks with low error and good coverage of our credible intervals.
BibTeX
@misc{unwin2020state,
  author = {H. Juliette T. Unwin and Swapnil Mishra and Valerie C. Bradley and Axel Gandy and Thomas A. Mellan and Helen Coupland and Jonathan Ish-Horowicz and Michaela Vollmer and Charles Whittaker and Sarah Filippi and Xiaoyue Xi and Mélodie Monod and Oliver Ratmann and Michael Hutchinson and Fabian Valka and Harrison Zhu and Iwona Hawryluk and Philip Milton and Kylie E. C. Ainslie and Marc Baguelin and Adhiratha Boonyasiri and N Brazeau and Lorenzo Cattarino and Zulma M. Cucunubá and Gina Cuomo-Dannenburg and Ilaria Dorigatti and Oliver Eales and Jeffrey W. Eaton and Sabine van Elsland and Richard G. FitzJohn and Katy A. M. Gaythorpe and William H. Green and Wes Hinsley and Benjamin Jeffrey and Edward Knock and Daniel J. Laydon and John A. Lees and Gemma Nedjati‐Gilani and Pierre Nouvellet and Lucy Okell and Kris V. Parag and Igor Siveroni and Hayley A Thompson and Patrick Walker and Caroline E. Walters and Oliver J. Watson and Lilith K. Whittles and Azra C. Ghani and Neil M. Ferguson and Steven Riley and Christl A. Donnelly and Samir Bhatt and Seth Flaxman},
  title = {{State-level tracking of COVID-19 in the United States}},
  year = {2020},
  doi = {10.1101/2020.07.13.20152355},
  howpublished = {medRxiv},
}
Figure