The COVID-19 pandemic is a major threat to the physical, emotional and economic health of our country and the world. To achieve the relief we need, we must prepare for the pandemic’s potential consequences and recognize that it will not impact every locale the same way.
One tool that we have is to build statistical models that allow us to predict where the virus will disproportionately harm a community and put together programs and policies to minimize its impact.
Many models exist to predict the spread of diseases, including those based on estimating the growth through the population or extrapolating other countries’ experiences with the disease and applying it to different areas.
We built a model based on the SIR (susceptibility, infection, recovery) framework that predicts the growth of COVID-19 for each state and county in the U.S. Additionally, we estimated the impact on the healthcare delivery system for each region and each hospital in the country, adjusting for things such as historical capacity, service area and elective admission rates. We individually estimated the expected number of COVID-related admissions, intensive care unit (ICU) visits and patients needing ventilators for each hospital, county and state.
Our estimates are built on the current trend of COVID-19 and are updated nightly. Importantly, our model projects how the pandemic is likely to expand given the current situation in each region.
As empirical evidence emerges that suggests social distancing and other responses are changing the rate of growth of the pandemic, the models will reflect those results. The estimates should be understood as the projected impact of the pandemic given how the pandemic is currently spreading, not a prediction of how social distancing or other policies will impact the pandemic.
Because policies are likely to only have a discernible effect over time, the projections for this model are most likely to be accurate for the next two to three weeks; longer-term projections (from three weeks to six months) would only be valid if no additional impact on disease is observable.
In short, our model shows where we are headed, but if we don’t like the destination, there is still time to change it.
Statistical models simplify complex realities, helping us make sense of millions of discrete data points. As a society, we know some things about how diseases spread. A model allows us to apply what we know to make sense of a new event with mechanics similar to those of past events and predict what will happen in the (near) future under those assumptions.
The value of a model is determined by how well it helps us understand the world and make better-informed decisions. With the COVID-19 pandemic, models help us understand what may happen, particularly as we are in relatively early stages of the outbreak when data on the current situation are lacking and policy responses can be most effective. Further, models help us identify the likely impact of different policies or approaches, particularly as we balance the physical health risks with emotional and economic risks.
All statistical models use assumptions to simplify reality. Other models, including the well-known COVID-19 Projections from the Institute for Health Metrics and Evaluation at the University of Washington, will differ from ours because of the different perspectives and assumptions made by other researchers.
The National Hurricane Center lists numerous models used to predict the path and intensity of hurricanes, but the composite “spaghetti model” is most useful to planners and contains an overlay of multiple models. Still, the spaghetti model is greatly improved by incorporating well-thought-out individual modeling perspectives.
This diversity of perspectives, and the resulting diversity of models, is a good thing. We hope others will find our perspective useful, as we have seen great value in other perspectives.
Predicting the growth of COVID-19
Most public health decisions and nearly all medical care is carried out at the state or regional level. While many models have attempted to project national trends or describe a single state or market, we believe this more granular view of each state and county supports decision-makers with an important, relevant perspective as they make life-and-death decisions with incomplete data.
In our view, policymakers and healthcare leaders need actionable information at a granular level on the current course of COVID-19 in their regions, without speculation on the possible effects of specific policy proposals or over-reliance on assumptions that don’t fit those regions.
Our model is designed to provide just that by drawing on high-caliber existing data sources and methodologies wherever possible. We employ an SIR model that incorporates the growth of infections within each county and hospital in the country. We acknowledge the COVID-19 Hospital Impact Model for Epidemics (CHIME) project for a helpful SIR framework.
We use data on current confirmed cases by county reported by Johns Hopkins University. The projections calculate regional and hospital-level burden based on expected need from peer-reviewed (PDF) research and Centers for Disease Control and Prevention data on COVID-19 morbidity in the U.S., Census Bureau data on county age distributions, and CMS hospital cost report data on beds and utilization and hospital service areas.
We use these data to estimate the number of COVID-19 hospital and ICU cases each hospital and county would expect to see given their demographics and the current number of active COVID-19 cases, including an algorithm we developed to account for diversions from full hospitals and ICUs to others in the area not yet at capacity. Combining these estimated severe COVID-19 cases with expected normal utilization, minus elective admissions, gives expected bed and ICU utilization, which can be compared to hospital and ICU capacity.
A full methodology is available here.
A critical parameter in many SIR models like ours is the rate of growth of the disease—often expressed as the time for the disease prevalence to double. The faster the disease spreads, the higher the likelihood that patient need will ultimately outpace healthcare capacity. Some models estimate the doubling rate based on observations in other parts of the world, evidence from other diseases, or other approximations.
Rather than making a generic assumption about the doubling time of COVID-19, we estimate each county’s current doubling time based on observed cases over the most recent eight-day window. A useful analogy is to view our country as a fleet of large ships traveling quickly across open water.
Recently, we discovered that we were heading straight toward a large, rocky island. Each boat (reflecting different states and communities) made decisions at different times and at different degrees to change the direction of their individual ship—such as instituting social distancing policies, canceling school and closing nonessential businesses. With any large, fast-moving ship, moving the rudder does not instantly lead to a radical change in course. By estimating current doubling times, our models are deliberately designed to predict what would happen if the current course were maintained, not how we hope or assume the disease curve is going to turn.
The time from COVID-19 exposure to the expression of symptoms is generally understood to be about two to three weeks. Because of this relatively long time frame, we expect the impacts of decisions made today are unlikely to be measurable for several weeks. When empirical evidence suggests that the rate of growth within a region has slowed, our models will predict a different future from the one we predict today based on that new reality.
What to do with this information
We believe our projections are accurate for two to three weeks from the date the prediction was made, but we also expect (and hope) that the longer-term projections can be reduced if wise policies are instituted to slow the spread of the virus. We believe that the estimates for two to three weeks are relatively accurate and can be used by hospital administrators, public health officials and other policymakers to make plans to address the known expected growth of patients and cases.
Beyond that three-week time frame, our models are better viewed as what is likely to happen if no improvements are made. It is important to remember that the impact of any policy may be very different between regions and states.
We believe our projections can be useful for a variety of people:
- Hospital administrators can understand the likely two- to three-week impact on admissions to their hospitals and prepare for this. If projected demand would overwhelm the hospital, seek to increase capacity or work with local policymakers to slow the growth of the disease.
- Public health officials can learn which regions of the country are likely to be most negatively impacted and when and create targeted educational programs for at-risk communities, train healthcare providers and ensure adequate access to supplies in advance of the crisis.
- Policymakers can understand the current rate of growth of the pandemic, including the direction the state or county is headed in the short term (unlikely to be modified) and longer-term (can be modified). As policies are put in place, officials can monitor how those policies and programs change the projected growth of the disease, gain insight into the likely timing for the pandemic and design appropriate economic packages to address concerns.
- Individuals can become informed about their local community and the likely progression of the pandemic, including the time to resolution, given the current environment.
We firmly believe that making these data available can lead to more informed decision-making and help us bring relief to our national and global community.
There are many approaches to address the COVID-19 epidemic. We give just two recommendations related to data.
- We recommend that the country begin regular testing of random samples of the population.
A key to predicting—and changing—the future of the pandemic is to understand what is really happening. All models are entirely dependent upon the data and assumptions that underpin them, and we are operating in a fog of war where we have incomplete data. However, testing is something we, as a country, have the power to address. Improved data will help with planning and restarting the economy at the right time.
Random sampling of the population will significantly improve the data that feed all COVID-19 models, which will lead to a better understanding of, one, where we are, and, two, where we are likely going. We will gain a better understanding of the true prevalence of the disease, the rate of asymptomatic carriers and a better time frame of exposure to symptoms. As a side benefit, we will also be able to identify people who have antibodies to the disease and could return to work and restart portions of the economy.
- We recommend that each community prepare a “worst-case” scenario plan as it relates to hospital capacity.
As we observed in northern Italy and New York, regions can quickly become overrun with COVID-19 patients, including running out of hospital beds, ICU beds and ventilators. While we do not expect all communities will be overwhelmed, it behooves regions to make plans on how to scale up capacity, including plans on how to build temporary hospitals, increase beds within existing facilities and non-hospital settings and transport patients to different regions.
To know what a worst-case scenario may be, communities and hospital administrators must be willing to entertain the idea that this pandemic could grow out of control within their community, and these models can help with that planning.
The COVID-19 pandemic is incredibly disruptive to the health, emotional and economic well-being of our country and the world. The better it is understood, including an honest assessment of where we are and where we may, at this moment, be headed, will help us overcome this challenge. These models can support the efforts of those who are fighting this disease on the front lines and improve the decisions of policymakers.
We hope that dire predictions, including ours, will never become a reality. That hope will be realized to the extent that we act in time to turn this ship away from danger.
David Muhlestein, Ph.D., is chief strategy and chief research officer for Leavitt Partners. Robert Richards, Ph.D., is a data science manager at Leavitt Partners.