(Updated through July 29)
(I have stopped updating this post, because my model is not as useful in predicting the course of the pandemic in Illinois counties, now that different states have taken very different approaches to managing the pandemic, with widely varying results).
Many researchers have created models with projections for COVID-19 cases, hospitalizations, intensive care unit stays, and/or deaths. Some of the most widely cited are models created by researchers at the University of Washington Institute for Health Metrics and Evaluation (https://covid19.healthdata.org/united-states-of-america), the University of Pennsylvania Perelman School of Medicine (https://penn-chime.phl.io/), the Los Alamos National Laboratory (https://covid-19.bsvgateway.org/), and a diverse team of epidemiologists and data scientists under a project called Covid Act Now (https://covidactnow.org/). Many of those models are excellent and have been utilized by federal, state, and local policymakers to help guide their responses to the pandemic. I do, however, have three concerns about many of those models. First, most models require the researchers to make some important assumptions about parameters such as the number of people to whom each infected person can transmit the virus, the number of days that a person is infectious, and the effects of various types of social distancing measures on transmission. Second, many models are basing their assumptions on data from other countries, such as China, Italy, and Spain, but the pandemic may play out differently in the United States than it has in other countries. And third, most models are designed to produce national estimates, which is challenging, because the pandemic is at different stages in different parts of the country. Due partly to these issues and partly to more basic methodological differences, current models are often producing very different projections from one another (https://projects.fivethirtyeight.com/covid-forecasts/).
I have created a model that relies on the fact that some parts of the country are further along the growth curve of the pandemic to produce COVID-19 case projections for Illinois counties. I use data only from the 48 contiguous states (and the District of Columbia), with the expectation that those states provide a better comparison for Illinois than foreign countries do. And I make no assumptions about the virus or the effects of social distancing on transmission of the virus; instead, the model’s projections are based only on the actual data from the 48 states.
My model is a multilevel nonlinear growth model using county-level data from the New York Times on cases reported each day in each county; I did not include data from New York City, both because New York City seems to be a unique case in terms of the pandemic’s spread and because data from New York City are reported only at the city level in the New York Times data, and not at the county level. The fixed portion of my model estimates the shape of the growth curve in total cases per 100,000 people using a fractional polynomial. The random portion of my model includes both random intercepts and random slopes at the state level and at the county level; the random slopes at both levels are also fractional polynomials. So, the overall shape of the growth curve is determined by a combination of national, state, and county trends; the model has allowed state and county trends to have more influence on the shape of the growth curve over time. The model used to also adjust for several demographic variables, but those variables have become less important over time as more days of case data have become available, so each county’s growth curve is now determined entirely by national, state, and county case data over time.
Of course, my model still has many of the usual limitations associated with these models. First, the number of reported cases is still believed to underestimate the actual number of COVID-19 infections in most counties, so relying on those data is inherently problematic. Second, there have been many reported clusters of COVID-19 infections in places such as nursing homes and prisons, which are virtually impossible to predict. Finally, there is typically a lot of uncertainty in making projections and my model is no exception to that rule. I do not include confidence intervals for my point estimates below because those intervals are so wide that they are of little value. With those limitations in mind, the table below includes the actual number of total cases in each Illinois county through July 29 and the projected number of total cases for each county through September 30.
The following chart shows the actual number of new daily cases reported in Illinois through July 29 and the number of new daily cases projected by my model after that date through September 30. My model projects that the number of new daily cases in Illinois will continue to increase through at least the end of September.
How do the results of my model compare to the results from other models? The University of Washington model does not provide publicly-available county-level estimates, but that model projects that the number of new daily cases in Illinois will decrease slightly through the end of September, before beginning to increase again (https://covid19.healthdata.org/united-states-of-america/illinois). The University of Pennsylvania model does not provide publicly available estimates for Illinois. The Los Alamos model also projects that the number of new daily cases in Illinois will decrease slightly through at least September 9 (https://covid-19.bsvgateway.org/). The Covid Act Now model provides estimates of projected hospitalizations for both the state of Illinois and individual counties in Illinois and projects that the number of hospitalizations in Illinois will decrease slightly through the end of August (https://covidactnow.org/us/il). So, all of those other models disagree with my model and project that, despite the increase in daily cases over the last couple of weeks, that new daily cases will instead decrease over the next month. I hope that those projections are correct, but, unfortunately, my model’s projection of a continuing increase in cases seems more realistic at this point.
Focusing on the downstate Illinois area, following are charts showing the actual and projected number of cases for several different health service areas. As for the state of Illinois as a whole, the growth curves for all of these health service areas are projected to continue to increase through at least the end of September.
I will continue to update this model occasionally. If you have any questions about the model, please feel free to contact me at email@example.com.