Reflects What is Actually Happening: First and foremost, many well-known COVID-19 prediction models are trying to predict what will happen based upon the future and unknown effectiveness of interventions. Our projections are based on growth curves of what is actually happening on the ground and updated daily. Thus, when states implement social distancing, shelter-in-place, or other measures intended to reduce the spread of the virus, our projections take these into account by updating accordingly as we receive new case count data The models reflect the direction we’re currently headed, not where we are hoped or assumed to be headed.
Adjusts for Hospital Service Area: Not all counties have hospitals, and sometimes patients cross county lines to go to different hospitals. We base our projections for counties and individual hospitals on hospitals’ service areas as defined in the CMS Hospital Service Area file, which are local healthcare markets for hospital care, rather than simply on where the hospital is located. This means we have scores even for counties that have no hospitals in their borders, and that we account for hospital patients from outside the hospital's county.
Adjusts for Hospital Baseline Capacity: Some hospitals were already near capacity before the COVID-19 crisis. We base our estimates of available capacity on hospital- and ICU-bed utilization rates derived from the 2018 CMS Hospital Cost Reports.
Eliminates Hospital Elective Procedures: Many hospitals can make additional beds available for COVID-related illness by postponing elective admissions. Based on a sample of all-payer claims and 100% of Medicare hospitalizations we estimate the percent of admissions in each hospital that are elective, allowing us to account for increased capacity over the baseline utilization.
We use the following data sources:
Our Methodology page contains a full description of how these data sources are used.
We update the Torch Insight COVID-19 Burden Scores and, our underlying COVID-19 case, counts from Johns Hopkins University daily.
Projecting into the future inherently becomes less certain the further ahead we look. Our model depends heavily on what is happening today and what has happened in recent days, and doesn't assume the impact of future interventions. We know that there will continue to be more changes made, and we will continually see the impact of increased (and, in the future, decreased) social distancing, but we don't predict the impact of future changes. Our model instead is based on the concept that we continue in our current trajectory.
The result is that there is a balance between having accurate estimates (i.e., the predictions are likely to come true) and predicting far into the future, which provides more time for planning. We have decided to limit our published predictions to 3 weeks as a balance between the two. We also update our predictions every day, so there is a rolling three-week window to observe, with the empirical impact of policy changes being captured and updated.
Yes, we have projections for up to 180 out from today. This data is not available to the general public. If you are interested in this data, please reach out to us at info.torchinsight@milliman.com.
There are countless approaches to building a model, with each model being dependent upon the assumptions that are built into it. One common model is to look at how COVID-19 has expanded and contracted in other areas (China, Italy, etc.) and use that past experience, with the assumption that the United States will see a similar outcome to estimate the future in the United States. A second approach requires people to guess the impact of known policies, such as how much social distancing is likely to reduce transmissions, and project forward based on those assumptions.
The approach that we have chosen to take is to measure the current trend of the progress of the disease and model how the disease-spread would occur if we follow on that trend. The limitation of this is that we do not capture the impact of policies (social distancing, sheltering in place, etc.) as they are being implemented. The strength, though, is that once those policies start to bear fruit, we empirically capture their measurable impact. By updating our model each day, we have a chance to see where each hospital, county, metropolitan area, state and the nation are headed, and as amelioration strategies prove effective, we will capture their measured impact.
We encourage leaders, policymakers, planners and decision makers to review multiple models, understand their different approaches, including their strengths, weaknesses and limitations, to make better-informed decisions.
If you'd like to learn more about our methodology, please read our full methods.
Most, if not all, hospitals and regions in the US have or will implement surge plans and other measures to handle increased hospitalizations due to COVID-19. We do not account for such measures explicitly in our estimates, with the exception of cancelling elective admissions. Thus, our estimates of capacity do not take surge capacity into account and should be interpreted as being based on “normal” capacity plus some increase for cancelled elective admissions. Increases in capacity from surge plans would be the same as raising the horizontal dashed lines in the charts (the indication of available bed capacity). The curved projections of cases are not affected by capacity and would not change with respect to surge planning.
We have attempted to use state-reported data to obtain hospitalization rates as close as possible to those observed by states. However, keep in mind the following:
Use the tool to understand the current trajectory of COVID-19 burden, rather than fixate on peak dates and predicted shortages that are many weeks, if not months away. These peak projections will likely change based on effective interventions. For those needing to make operational decisions about the next two or three weeks, such as hospital administrators, we believe that the COVID-19 Burden Scores can be used to predict bed utilization and shortages within that time frame.
To ensure the accuracy of projections shown on this site, we drop projections for any counties that we estimate to have fewer than 11 active cases (excludes estimated recovered cases and deaths). Counties with fewer than 11 active cases, along with their projections, are included in state and national roll-ups.
Please use and share these scores to help policy makers make smart decisions, keep patients safe, and prepare intelligently. Reselling, repackaging, or stealing the data and all that nasty jazz is definitely prohibited. See our Terms of Use for the legal jargon.
Data can currently be downloaded one region at a time. If you’re interested in getting more data at once, please reach out to us at info.torchinsight@milliman.com.
Access to underlying data requires a Torch Insight subscription. Please reach out to us at info.torchinsight@milliman.com to chat about getting access to a full Torch Insight license.
We’d love feedback on our methodology. Drop us a line at info.torchinsight@milliman.com and we’re respond as soon as we can.
The confidence intervals only apply to the upper and lower limits of the estimated variables—namely the number of hospital cases and ICU cases. We are not estimating a lower and upper limit on the predicted COVID-19 cases or on hospital and ICU capacity measures.
County-level data was only available beginning March 23, 2020. Due to how we calculate active cases, which accounts for recoveries and deaths, we have a drop in active cases on April 6, 2020.
We do not have an API feed available for the general public. If you are interested in accessing the API, please reach out to us at info.torchinsight@milliman.com.
Please contact us at info.torchinsight@milliman.com.
