What is TRACE?
TRACE (Testing Responses through Agent-based Computational Epidemiology) is an agent-based computational model developed by a team from Brookings and Washington University in St. Louis, with the specific goal of providing insights into how policies that use testing and contract tracing might help contain the COVID-19 pandemic. It draws on the extensive body of evidence about both the current and past epidemics, and is also designed to manage a high degree of remaining uncertainty about some of the parameters it uses.
The COVID-19 pandemic has presented policymakers with an immediate set of difficult challenges. Due to its recent emergence and lack of precedent in the modern era, there is much that we don’t know about the virus itself or the ways in which people will behave in response. However, there is a need for both timely and thoughtful action to address the threat that the current pandemic represents not just to life and health, but also to the economic and social fabric of our nation and global community.
There are many avenues of important research that are being simultaneously pursued by ourselves and others who are working to understand and address this crisis. We have elected to first focus our efforts on exploring the interplay between three types of activities that are currently being considered, conducted, and adjusted: testing members of the population for active viral infections, tracing their close contacts, and asking individuals to engage in “social distancing.” There are many factors that comprise each of these efforts: who is tested, how accurate are tests being deployed, who is asked to engage in social distancing and when or how are they asked to do so.
The consequences of different combinations of activities across contexts is neither straightforward nor intuitive. Therefore, we turn to computational simulation modeling in order to provide insight into how policymakers might make decisions that can minimize the negative effects of the pandemic in the near future. The model is not intended to provide exact forecasts of epidemiological outcomes for particular policy scenarios. Rather, it is intended to be used as a “virtual policy laboratory” that can help compare the likely effects of different combinations of policies. We thus hope that it is a valuable part the toolset that policymakers have at their disposal to make difficult decisions about what pandemic responses to engage in and how to do so.
How does TRACE work?
TRACE simulates transmission of the COVID-19 virus between individuals and the progression of the infection within individuals. Each simulated individual (“agent”) in our model represents a member of a population that is generally representative of the United States (for computational tractability, we use a reduced population of 50,000 agents with all assumptions scaled appropriately). Agents have demographic attributes and infection states, which shape the set of other individuals with whom they have social contact in ways that might transmit the virus. This approach allows us to compare cumulative levels and timing of population infection rates across a number of different testing, contact tracing, and social distancing policy combinations. We summarize our model here, and describe it in complete detail (including mathematical equations, data sources, and computational implementation) in a forthcoming manuscript.
Infection progression in our simulations uses a variant of the classic “Susceptible-Exposed-Infectious-Recovered” epidemiological model that is intended to specifically represent COVID-19 (Figure 1).
Figure 1: COVID-19 “states” and possible “state transitions” in our model
Simulated individuals who have never experienced a COVID-19 infection start as “susceptible.” If contact with an infectious individual transmits the virus, then they become “exposed.” After a set incubation period, they become “infectious.” Some individuals have shorter latent periods and are infectious before they display symptoms (i.e. are “pre-symptomatic”), while others never display symptoms (i.e. are asymptomatic). We allow infectivity—that is, one’s ability to transmit the virus to others with whom they interact—to differ across both individuals and infectious type. This allows us to represent the presence of “super-spreaders” who are highly contagious as well as a lower likelihood of non-symptomatic individuals transmitting the disease (e.g. by coughing). Finally, when the infection has run its course, an individual is “recovered.” For the purposes of this model, we assume that anyone in the “recovered” state cannot be re-infected with COVID-19. The degree of immunity conferred by prior infection remains an open question in the scientific literature.
Contact between agents
On any given day, individuals interact with others in ways that might transmit the COVID-19 virus (Figure 2). We allow the number of these daily contacts to differ between individuals and to be related to individuals’ ages. The set of people with whom an individual spends time is also determined by age (e.g., a school-age child will have a much larger proportion of their daily interactions with other school-age children than with senior citizens). In addition, we categorize the settings in which these interactions occur. Specifically, we denote which contacts take place in schools, workplaces, or at home.
Figure 2: Individuals (circles) contacted by a single agent (larger, central circle) during a simulated day. Circle coloring indicates COVID-19 state (see Figure 1). Blue lines indicate contacts that occurred in the workplace.
Figure 3: In this animation, a small subset of the agent population (75 agents out of 50,000) is shown as the epidemic spreads through social contacts. Each dot is one agent; the size of the dot corresponds to the age of the agent (small dots are children, medium dots middle-aged adults, large dots over 65). The colors of the dots change through time to reflect disease state as in the previous two figures. Lines represent social contact between agents. Light colored lines represent contacts at home, dashed lined represent contacts at work, and dark lines represent contacts at school and other settings.
Output from TRACE
We simulate two primary outputs: the number of infections at any point in time (prevalence), and the cumulative infection rate over the course of the entire time period simulated. Both are shown in the Interactive Dashboard.
Because TRACE is an agent-based model, we can collect complete information on which agents contract COVID-19, from whom, and under what circumstances. We can also keep track of how many true cases of COVID are “missed” by testing, how many true contacts of infected agents are “missed” by contact tracing, and how much disruption to individual economic or social activities is caused by any particular policy via quarantine or social distancing. This information can be used to increase the efficiency of policy interventions and maximize epidemiological “success” while minimizing disruption to society and the economy.
Using TRACE to simulate policy options
We explore the following policies in our model:
- PCR Testing. Polymerase chain reaction (PCR) tests are designed to detect active viral infections. In our model, agents who are exposed or infectious will test positive with probability equal to 1 minus the test’s false positive rate. Individuals who test positive are quarantined for 14 days. We explore variations in
- Test quality (proportion of false negatives): 5% and 30%
- Daily Test volume (number of available tests per day): 250,000, 350,000, and 500,000
- Testing Strategy (who gets priority when tests are limited): symptomatic individuals, agents age 65 and over, the contacts of symptomatic individuals, or a random sample of the population
- Contact Tracing. A list of individuals who had contact with people who have received positive PCR tests is generated, and up to a limit defined by capacity these contacts are traced for follow-up. We explore variations in
- Tracing capacity (number of contacts who can be traced in a given day): 1 per 500 individuals, 1 per 50 individuals, and Unlimited.
- Follow-up: traced contacts are asked to self-quarantine for 14 days, or traced contacts received PCR tests of their own and symptomatic agents are asked to self-quarantine for 14 days
- Serologic Testing. Individuals who test positive for COVID-19 antibodies are deemed “recovered.” They will not be given PCR tests and will not be asked to self-isolate if they have been in contact with someone who has tested positive on a PCR test. As with PCR testing, we explore variations in
- Test quality (proportion of false positives): 10% and 30%
- Test volume (number of available tests per day): 0 and 500k
- Testing strategy (who is tested): population at random or only workforce-age agents
- Social Distancing. Portions of the agent population may be asked to refrain from interactions that occur in specific settings. We explore variations in such policies including: no social distancing, school closures, remote work requirements for some or all of the workforce, and “safer-at-home” orders (i.e. eliminate contacts that occur outside of work, home, or school).
Any policy combination from the settings above can be compared (across multiple stochastic simulations) to a “baseline”: this can be either a full social distancing scenario (safer-at-home + all schools closed + all work remote or closed) or no-intervention scenario in which the population interacts freely with no restrictions. The outcomes available for comparison are described above.
There are many “known unknowns” that apply to the COVID-19 pandemic in the United States. Therefore, we use our model to explore every policy combination above under a wide range of conditions that represent plausible ranges of:
- Disease transmission. Uncertainty remains regarding the basic reproductive number of the disease during sustained transmission in normal social contact patterns. We therefore consider scenarios in which disease transmission is more or less rapid:
- R0: 2.2 and 3.4
- Initial disease prevalence. Policymakers may make different choices about how far the level of infections detected must fall before considering relaxing social distancing measures; there may also be uncertainty in the measurement of cases at any point in time. In addition, there remains uncertainty about how many undetected cases may have already occurred, resulting in a background level of recovered individuals in the population. We vary both of these factors at the time the simulation begins (e.g. as any new policies come into effect):
- Initial infected: 0.1% and 0.6%
- Initial recovered: 0.1% and 15%
- Mandate adherence. We explore scenarios in which individuals are quarantined or requested to self-isolate. Whenever an agent is asked to quarantine or self-isolate, there is some probability that they will refuse and continue with their daily contacts as normal.
- Agents adhere to quarantine with probability 50%, 75%, or 90%.
In total, we explore 165,888 scenarios, varying the underlying epidemiological parameters and policy choices. We conduct 20 runs of the stochastic simulation for each scenario, for a total of 3.3 million simulations. A sampling of these runs can be explored through the Interactive Dashboard.
Limitations and Future Work
We have prioritized the elements included in this initial model and the combinations of scenarios we have explored because we believe that they are the best match between policymakers’ immediate needs and data that are currently available. However, many additional factors are not included in our initial model that can be incorporated in subsequent versions in order to gain a more detailed picture of potential policy impact. These include:
- Further exploration of how to effectively conduct and respond to serological testing (i.e. testing for antibodies that indicate that an individual has recovered from a COVID-19 infection and may have at least temporary immunity) to allow resumption of key economic and healthcare activity
- Simulation of fatalities from the disease, how these may depend on factors like age and underlying health conditions in our heterogeneous agent population, and how these might affect policy efforts or healthcare capacity.
- An explicit representation of how the health care system engages in testing and contact tracing (e.g. the formation of a “backlog”).
- The importation of new cases from outside the agent population, to evaluate which policies are effective at containing spread despite “new arrivals” from other locations. Our results already provide a partial answer to this by indicating which test-and-trace policy configurations can suppress outbreaks which begin with as high as .6% of the population infected.
- The potential for the disease to manifest much differently in pediatric cases, a topic for which there is rapidly changing emergent science.
We have designed our model to be highly extensible, which will allow us to address any of the above in future versions of our model as sufficient, relevant data becomes available.
Our model is also designed to allow customization of the general framework for specific settings such as a state, county, or city. We have begun working with policymakers and stakeholders in state and local government on these applications. For an early example, see TRACE-STL.
For more Information
For media or collaborative inquiries regarding TRACE, please contact: Shannon Meraw SMeraw@brookings.edu
The TRACE team includes:
Ross A. Hammond, Ph.D., Director, Center on Social Dynamics & Policy, Senior Fellow, Economic Studies, The Brookings Institution; Betty Bofinger Brown Distinguished Associate Professor, Public Health and Social Policy, The Brown School, Washington University in St. Louis; External Professor, The Santa Fe Institute
Joseph T. Ornstein, Ph.D., Postdoctoral Research Associate, The Brown School, Washington University in St. Louis
Matt Kasman, Ph.D., Assistant Research Director, Center on Social Dynamics & Policy, The Brookings Institution
Rob Purcell, Research Programmer, Center on Social Dynamics & Policy, The Brookings Institution
A scientific manuscript, full documentation, and computational code for the TRACE model is in preparation and we will update this page with a link to this information shortly.