key: cord-0149154-1acqzy0j
authors: Jones, Kasey; Hadley, Emily; Preiss, Sandy; Kery, Caroline; Baumgartner, Peter; Stoner, Marie; Rhea, Sarah
title: North Carolina COVID-19 Agent-Based Model Framework for Hospitalization Forecasting Overview, Design Concepts, and Details Protocol
date: 2021-06-08
journal: nan
DOI: nan
sha: 1003678a48e9a089dd01f174f52526984f73d703
doc_id: 149154
cord_uid: 1acqzy0j

This Overview, Design Concepts, and Details Protocol (ODD) provides a detailed description of an agent-based model (ABM) that was developed to simulate hospitalizations during the COVID-19 pandemic. Using the descriptions of submodels, provided parameters, and the links to data sources, modelers will be able to replicate the creation and results of this model.

An Overview, Design concepts, and Details (ODD) protocol provides a standardized way of describing an agent-based model (ABM) so that other scientists can implement the model, promoting open science and replicability. 1, 2 This ODD is closely tied to the model source code.

The information presented, including methodology, assumptions, and parameters, reflects the state of the model and model source code at the time of this publication. It is subject to change as coronavirus disease 2019 (COVID- 19) knowledge evolves and as model updates are made.

Although parameters are discussed throughout the ODD, a complete list of parameters and associated references is available in Appendix A.

The purpose of the model is to simulate demand for hospitalizations and to forecast capacity challenges for North Carolina (NC) hospitals during the pandemic of COVID-19, the disease caused by SARS-CoV-2. The ABM output provides information through near-term (e.g., next 30 days) forecasts of hospital capacities, statewide and by region in NC, to help inform pandemic response efforts. SARS-CoV-2 infection predictions that correspond to different levels of infection spread are used as input for the ABM. The ABM forecasts demand for intensive care unit (ICU) and non-ICU beds for agents with and without COVID-19 (i.e., non-COVID-19 agents hospitalized for other reasons). Stakeholders can use model output to prepare for varying COVID-19 scenarios in NC.

To validate the hospitalization forecasts, we evaluate the ABM by its ability to reproduce two patterns. Not applicable Everything outside of the other location nodes 1 ! The ABM, originally developed to explore healthcare-associated infection prevention interventions in NC, began as a regional healthcare system network model, including 10 modeled UNC Health STACHS. 6 Subsequently, the model was expanded to a statewide platform to include all licensed STACHs in NC (UNC STACHs and all other non-UNC STACHS) in addition to other licensed healthcare facilities (LTACHS, NHs). Here, we maintain the nomenclature developed for the original ABM (i.e., UNC STACHs and non-UNC STACHs). # Based on available patient-level discharge data from seven acute care hospitals of UNC Health during July 1, 2016-June 30, 2017. 6 We do not model agent movement among households. Agents located in the community node can be conceptualized to be anywhere in the community other than an STACH, LTACH, or NH.

Both agents and locations are essential in calculating the capacity of healthcare facilities across time.

Healthcare facility locations (i.e., all non-community nodes) have several state variables, including name, physical location (i.e., county), bed count, and a unique location identifier (Facility ID) that is used throughout the ABM (Table 2 ). Agents who move to one of these locations are given a specific bed within the healthcare facility. At that time, an agent becomes an attribute of the location and if the location is a hospital, is assigned a specific bed. STACH locations have beds designated as ICU or non-ICU and whether a ventilator for that bed is available. Not all NC counties have a modeled healthcare facility. Note that state variables like bed counts, names, and location IDs are not relevant to the community location, and the county attribute is only available for STACHs. 

The ABM is implemented with a 1-day time step, and there is no sense of daily time in the model. After model initialization, the ABM produces 30-day hospital demand forecasts.

Each time step in the model consists of two substeps, a life substep and a location substep. A life substep represents events causing death for agents and could result in an additional facility bed becoming available (e.g., agent death followed by a discharge from an STACH). A location substep represents events that could result in agents seeking healthcare or changing healthcare facilities (e.g., becoming infected with SARs-CoV-2, changing facility locations, or seeking care for the first time at a healthcare facility). All functions executed in the model for these substeps will randomize the agent order before executing the substep. -For agents in a non-community location, send the agent to the community to make the bed that they were occupying available.

-Add the agent to the list of agents to recreate (see Section: Submodels).

Location substep:

• Check which COVID-19 agents are set to recover. If they recover, perform a COVID-19 state change.

• Administer a location update (see Section: Submodels) for any agent with a LOS ending on the current day.

• Complete a COVID-19 update (outlined below).

• Administer a location update for any agent selected to leave the community.

• Administer a location update for any agent whose readmission date is for the current day.

To complete a COVID-19 update:

• Using the SEIR infection projections by county, select susceptible, community agents by county to be newly infected.

• Estimate the probability of being tested among newly infected agents. Note that this step does not happen on the first day (see Section: Initialization).

• Assign symptom severity (i.e., asymptomatic or mild to moderate, severe, critical) for newly infected agents.

• Determine hospitalization among newly infected agents.

• Hospitalize COVID-19 agents and assign a LOS.

Of all the design concepts available in the ODD protocol, we describe below those that are relevant to the current model.

To be included in future updates.

To be included in future updates.

No adaptive decisions are made at this time.

Not applicable at this time.

Agents do not learn or adapt over time.

The ABM does not have explicit or implicit prediction components at this time.

Agents do not know anything and, therefore, do not use this information.

Agents in the ABM only have mediated interactions. When an agent occupies a bed in a healthcare facility, other agents do not have access to that resource. This might cause agents to be turned away from an STACH. Because the number of SARS-CoV-2 infections is predetermined using a set of SEIR models, SARS-CoV-2 is not transmitted by agent interaction in the ABM.

To simulate random events that happen to each agent (i.e., death), the ABM compares the probability of events to values generated by a random number generator. If randomly selected, an agent may move location, develop an infection, or die. Random selection of STACH LOS and type of bed assigned (i.e., non-ICU or ICU) also impact stochasticity in the ABM. By setting a random seed, results can be reproduced. All agents that are selected for a specific function, such as agents going to facilities from the community, are randomized before an update takes place.

To be included in a future update.

For each model run, initialization of the agents is the same. In contrast, the projected SARS- For each row in the augmented version of the input synthetic population file, an agent is initiated with the variables provided in that row. Agents are then randomly assigned a concurrent conditions binary category (i.e., comorbidities) based on their age group. 10 Unless selected to start in a facility, all agents are initially set to start in the community.

All NH and LTACH facilities do not rely on model parameters for their initial capacity levels. At initialization, agents are assigned to these facilities until they reach 70% capacity (expert input).

The agents selected to start in these facilities are randomly selected based on a probability distribution formed by taking the relative distance between the facility and each county. Counties close to the facility will be pulled from more often. For NHs, agents must be ≥ 65 years of age to be selected.

Each STACH in the model is initiated with ICU and a non-ICU agent to match starting capacity percentages specified by either model parameters or by real-time capacity percentages for each hospital bed which is slightly higher than the statewide percent hospitalized (to make sure all needed beds are filled). If there is an unfilled SARS-CoV-2 bed needed for a hospital within 60 miles of an agent selected for hospitalization, they are assigned that bed. Hospitals are considered in the order of increasing distance from the agent's county center using the County to Facility Distances input data file.

Data collected to calibrate previous iterations of the ABM 5,6 are not described in this section. All references to data files can be found on the models repository 32 at the location specified. The repository contains all code and data necessary to reproduce each file.

Once it is determined that an agent is moving to another location, the probability of transferring from an agent's current location to another healthcare facility type is outlined in the location transitions file. Probabilities are based on an agent's home county and age group. Although unique transition probabilities are available at the individual STACH level, other location types (community, NH, and LTACH) have a single row of probabilities in this file. Transition probabilities must sum to 1 and are based on discharge data for each healthcare facility type. For example, the six transition probabilities for a UNC STACH are based on the total discharges that specific STACH had to each of the six location types (community, UNC, Small non-UNC, Large non-UNC, NH, LTACH). 7 Individual facility-level data were not available for NHs and

LTACHs. Therefore, we estimated the total number of transitions between healthcare facility types by assuming that the total discharges to NHs and LTACHs was equal to their total admissions. See Table 4 for a list of parameters to determine transitions between specific STACH types, which was used to create the location transitions file. The probability of leaving the community on any given day and being admitted into a healthcare facility is outlined in the community transitions file. These probabilities are based on an agent's home county and age and were determined by assessing how many people were admitted to each facility over a 1-year time frame. 3 Note that only agents ≥ 65 years of age can be admitted to a NH. This file contains a column with the daily probability that an agent (from a specific county and in a specific age group) would leave the community for admission to any healthcare facility.

In addition to this daily probability, there are five individual probabilities for being admitted to each (non-community) healthcare facility type. These five probabilities must sum to 1. To calculate these probabilities, we evaluated the total expected admissions to each healthcare facility type from the community for each age and county combination. 3, [7] [8] [9] 13 Facility Discharge Data (data/discharges/county_discharges.csv)

For each STACH type there is a file containing the number of agents that were discharged from each hospital by county. 7 These files are used to create the healthcare facility transition files.

Discharge data by county of residence and hospital only include counties that account for ≥ 1% of discharges.

For each county and each facility type (STACHs, NHs, large non-UNC STACH, and LTACHs), there exists a dictionary of distances between that county and all facilities. Distance was calculated using the centroid of the county and the facility addresses.

The NH LOS file contains the number of NH patients that were discharged from NHs for each possible length of stay between 0 and 2,000+. 14 These data are converted into a list of potential LOS values and are used when assigning the LOS for a newly admitted NH agent. 

There are several files containing facility information used as input into the ABM. These files are crosswalks between the Name of the facility and the actual facility ID. The hospital facility file also contains the number of non-ICU and ICU beds for STACHs.

This file contains the number of confirmed COVID-19 cases by county and by day. The most recent data available should be used. Instructions for downloading and cleaning these data are found in the SEIR directory of the public repo. 32

This file contains the number of occupied inpatient beds and ICU beds and counts of confirmed and probable inpatient and ICU COVID-19 patients by hospital. As access to these data is limited; alternative expert-informed defaults are included in the public repository version of the ABM.

The ABM has three submodels, one for each state variable: life, location, and COVID-19. We also outline the SEIR model that is run before the ABM begins in this section. If another outside model is used for creating SARS-CoV-2 projections, the final subsection of Section: Submodels can be skipped.

Death is an important component of the ABM. The model is parameterized to certain hospital capacity levels, and without death, these levels would not be maintained. Hospital beds would also not become available as often as they do in reality if we did not model death.

The daily probability of death occurring from natural causes was derived using CDC WONDER data 4 and is included in the input parameters file. Death probabilities are provided by age group.

To calibrate to the number of expected deaths in each facility type, 3 we included multipliers in the parameters file for each facility type. These multipliers can be used to increase or decrease the frequency of deaths in the model and were raised and lowered during calibration until the correct number of deaths occurred for each facility type.

We do not specifically model death from COVID-19 within the ABM at this time, although this is an addition we plan to make in the future.

Every 15 days in the model, the attributes of deceased agents are used to create new agents.

Recreating agents every 15 days allows the model to skip over this function until enough agents have died that we need to recreate them. All regenerated agents start in the community, do not have COVID-19, and have no additional location attributes (e.g., current LOS, readmission date). Each agent in the community has a daily probability of moving from the community. In the community movement step, the probability (assigned using the community transitions file) for each agent is compared to a random number to see if that agent will leave the community. If selected, a second random number is drawn to determine which type of healthcare facility the agent will move to (based on the location transitions file). Finally, based on the type of healthcare facility selected, the agent is assigned a specific healthcare facility ID based on their home county (using the healthcare facility transitions file).

Agents in healthcare facilities only leave that facility if their LOS ends or if they die. For agents whose LOS ends, we use their location transition probabilities to determine the healthcare facility type of their next destination. A random probability is generated, and this probability is compared to their transition probabilities. Most agents will move to the community upon discharge, but some are selected for transfer. Once a healthcare facility type is determined, if a non-community node is selected, we compare a second random number to the facility transitions to determine the exact facility ID. There is one hard-coded component that is added to this logic for NH agents. For agents who previously transitioned from a NH to an STACH, we assume ~80% will return to the previous NH when their STACH LOS ends. 15 We included this hardcoded component in the face of a general lack of available data on NH agent movement; this can be updated in the future with additional data. When, because of censored values, the aggregate hospital discharge data cannot be used, we apply different restrictions for movement according to the healthcare facility type, as described in more detail below.

• Small (<400 beds) non-UNC STACHs: Small non-UNC STACH discharge data are available for 99 NC counties. The ABM uses distributions created from these available discharge data 7 to randomly assign agents selected to move to a small non-UNC STACH.

Agents that are discharged from a small non-UNC STACH can be selected to move to another small non-UNC STACH, with assignment based on the distributions. Agents are not permitted to remain at the same small non-UNC STACH once their LOS is complete.

Rather, in this rare situation, if an agent is selected to transfer from a small non-UNC STACH to another small non-UNC STACH and that agent is from a county with discharge data available for only one small non-UNC STACH, the ABM randomly assigns the agent to another small non-UNC STACH.

•

Large non-UNC STACH discharge data are available for 74 NC counties. The ABM uses distributions created from these available discharge data 7 to randomly assign agents selected to move to a large non-UNC STACH. If an agent who is selected to move to a large non-UNC STACH is from a county with no large non-UNC STACH discharge data available, the ABM assigns the agent to a large non-UNC STACH based on a probability distribution that weights the size of available facilities with the relative distance to the agent's county. The relative importance of bed count versus distance are parameters given to the model. Similarly, if an agent is selected to move from a large non-UNC STACH to another large non-UNC STACH, the ABM uses the distributions, followed by the probabilistic approach, to assign the agent. In the rare situation that both these methods fail, the ABM randomly assigns the agent to a large non-UNC STACH.

• UNC STACHs: Agent movement to UNC STACHs is based on the patient-level discharge data from seven UNC Health acute care hospitals during June 30, 2016-July1, 2017, for each of the 10 modeled UNC STACHs which serve a 41-county catchment area. 6 Most of the agent movement to and from UNC STACHs is completed by agents whose home county is among the 41-county catchment area. The ABM uses distributions that we created from the available discharge data to select an agent's initial UNC STACH and inform its movement from one UNC STACH to the next (i.e., transfer). If an agent is selected to transfer from a UNC STACH to another UNC STACH and that agent is from a county with discharge data available for only one UNC STACH, the ABM randomly assigns the agent to one of the two largest UNC STACHs.

When an agent arrives at an STACH or LTACH, the agent is assigned a LOS based on a gamma distribution unique to the healthcare facility. We use patient-level data, available for 7 of the 10 UNC STACHs, to obtain STACH-specific LOS gamma distributions. For the remaining three UNC STACHs and the non-UNC STACHs, for which patient-level data were not available, we used aggregate discharge data to estimate the parameters of a gamma distribution. 8 When an agent arrives at a NH, the agent is assigned a LOS based on a list of possible LOS values for NH agents (see Section: Input Data).

Non-COVID-19-related readmission is unique to STACHs, and several agents will be randomly selected for readmission at the time of leaving an STACH. Thirty-day readmission is approximately 10% and is based on the patient-level UNC Health data. If selected for readmission, the agent is randomly assigned a readmission date between 1 and 30 days from the current day. It is possible that an agent will be readmitted during this 30-day window. If this occurs, their previously assigned readmission date is deleted.

The final way that an agent can move to a healthcare facility is when they are assigned COVID- finding an STACH to meet that need, as follows:

1. The agent will try their second choice STACH, based on the STACH probabilities for that agent.

2. The agent will try any other STACH with a catchment area that includes the agent's home county. 7 3. The agent will try any additional NC STACHs located within a 200-mile radius of the centroid of the agent's home county.

If the agent is turned away from their first choice STACH, that agent is added to a list of agents turned away. In this list the model maintains the date, location, and county of agents that were turned away during a model run. If the agent is turned away from all STACHs that they tried, that agent is added to a list of agents who were completely turned away during the model run.

Agents who are randomly selected to transfer from another healthcare facility will only try their first choice STACH. If this STACH is at capacity, the agent returns to the community. We based this assumption on the premise that a healthcare facility would not transfer an agent to an STACH that did not have a bed available.

The COVID-19 submodel is a smaller model within the larger ABM that simulates COVID-19

and monitors COVID-19 status among agents. This submodel is primarily used to estimate demand for non-ICU and ICU hospitalization among all agents, including COVID-19 agents.

The submodel also includes preliminary estimates for ventilators and demand for NH beds following hospitalization with COVID-19. In the future, these estimates could be extended to include items related to hospitalization, including demand for personal protective equipment and healthcare staffing needs.

The agents in the COVID-19 submodel are updated daily. Agents transition between COVID-19 disease states (i.e., susceptible, infected, recovered) and hospitalization.

The agents in the COVID-19 submodel are updated daily. Agents transition between COVID-19 disease states (i.e., susceptible, infectious, recovered) and hospitalization. 

Starting COVID-19 agents with the status RECOVERED are not eligible to be reinfected. This assumption can be updated as more information becomes available regarding reinfection.

On each day, the model selects agents to be infected with SARS-CoV-2. This step requires the count of the number of agents to be infected by NC county. Day 0 is unique, as described in Only agents that have a status of SUSCEPTIBLE and are located in the community in a given county are eligible for infection. We do not currently model SARS-CoV-2 infections originating in healthcare facilities, although this is an area for future work. In the rare case that, on any given day the SEIR model estimates more infections than there are eligible agents, all remaining eligible agents in a county will be infected. This assumption will be revisited in the future.

The probability of an eligible agent being assigned COVID-19 is also dependent on the age of the agent. These probabilities are derived from observed age data among COVID-19 reported cases in NC and are described in further detail in Appendix A. At the end of the Select Infected

Agents step, the appropriate number of agents are infected as per the estimates from the SEIR model and the distribution of ages among the newly infected agents is congruent with the reported NC data from positive tests as described in Appendix A. Parameters and assuming that the distribution also applies to the agents created using the reported case multiplier.

On Days 1-30, agents newly infected with SARS-CoV-2 are identified as either tested and reported or untested and unreported. We model testing status because we assume that tested and reported infections are more likely to be hospitalized than untested and unreported infections. A basis for this assumption is that infections with more severe symptoms that would lead to hospitalization are more likely to be tested (expert input). Note that an untested and unreported infection could become a tested and reported infection, particularly if an agent is hospitalized and subsequently tested in the STACH. This change in status is not currently tracked in the model.

The probability of being tested is dependent on the overall observed ratio of tested and reported cases to unreported infections. This parameter is described in further detail in Appendix A. The model includes functionality to modify this parameter by age if evidence suggest that the probability of getting tested differs by age group.

Testing status is not used on Day 0 because the number of hospitalizations is forced to match the observed reality on Day 0, rather than based on probability of hospitalization as is done for Days 1-30.

Every agent with an active infection is also assigned a symptom severity level. Each symptom severity level is associated with a level of hospitalization. These symptom severities are assigned the same day as the infection starts for that agent and persists throughout the duration of infection for that agent until the agent's day of recovery (described in Check for Recovery from COVID-19 Among Agents Hospitalized on the Previous Day). Further development will explore opportunities to transition between symptom severity levels.

The levels of symptom severity in the model include the following: patients hospitalized in NC per hospital occupancy data.

COVID-19 agents seek an STACH based on their symptom severity level. Agents with Severe symptoms will only seek a non-ICU bed and will attempt multiple STACHs as described in

Location until admitted or completely turned away. Agents with Critical symptoms will only

seek an ICU bed and will attempt admission at multiple STACHs as described in Location until admitted or completely turned away. A proportion of agents with Critical symptoms are also flagged as in need of a ventilator. This proportion is further described in Appendix A.

If a COVID-19 agent is admitted to an STACH, their LOS is assigned based on observed data (as described in Appendix A from COVID-19 Among Agents Hospitalized on the Previous Day). If a COVID-19 agent is completely turned away from all attempted STACHs, they are not assigned a LOS but, instead, assigned a COVID-19 recovery day based on the observed duration of the infection (see Appendix A).

The variable categories collected for each day of the forecast in each model run are presented in Table 5 . (Table 6 ). infections that have occurred. [16] [17] [18] [19] This process is completed for each day, , such that:

• The estimated infections are added to % .

• %&' is updated to reflect the number of exposed individuals required to achieve the level of new infections for day .

• % is updated based on the amount of people recovered from %&( . • % is equal to 1 minus the sum of the rest of the compartments.

Transition from recovered to susceptible is not modeled currently but can be revisited as more information becomes available about reinfection. If there have been no COVID-19 reported cases for a specific NC county, we set the proportion of infected individuals for the last day of available data to be (1 / county population).

Before the SEIR model begins, using the current day as day 0, we set the susceptible proportion equal to the percent susceptible parameter, taking or adding an equal proportion from the recovered proportion. This allows the SEIR to start with the expected proportion of susceptible agents, as the pandemic has already been ongoing. The estimates for the exposed and infectious compartment remain unaltered. For Day 1 going forward, we use the standard SEIR method to model transitions between each SEIR compartment. SEIR models use a Beta*S*I calculation to determine the proportion of newly exposed agents. 22 The beta parameter is based on the estimated ) value. For each county-level SEIR model, the input $ is updated using what we call a county correction.

For a model run, we are estimating the number of new infections for the next 30 days given the specified $ value. Because each NC county has a different COVID-19 reported case count and will respond to interventions in different ways, we introduced a county correction into the SEIR model. This correction is based on comparing the estimated growth rate for the state over the past 2 weeks, − $ , to the estimated $ for each individual county, $ ! . Counties that have had higher $ values are assumed to continue to have higher than average SARS-CoV-2 spread during the next 30 days. To arrive at the final $ used in the SEIR model for each individual county, we use $ = $ * $ / $ ! . SEIR models, however, do not use $ values.

To estimate the ) for a specific county (the ) is in a totally susceptible population), we take the input $ and divide it by the remaining proportion of susceptible individuals at time , the last date of available data.

The final estimates of the SEIR model produce the number of estimated new reported COVID-19 cases and total infections (reported and non-reported) for each county and day for 30 days after the start date. This output is used to drive new SARS-CoV-2 infections for the ABM.

The odd protocol for describing agent-based and other simulation models: A second update to improve clarity, replication, and structural realism

Modeling infectious diseases in healthcare network (mind-healthcare) framework for describing and reporting multidrug-resistant organism and healthcare-associated infections agent-based modeling methods

NC Department of Health and Human Services, Division of Health Services Regulation

Centers for Disease Control and Prevention. CDC WONDER mortality data

On calibrating a microsimulation of patient movement through a healthcare network

Creation of a geospatially explicit, agent-based model of a regional healthcare network with application to Clostridioides difficile infection

Sheps Center for Health Services Research, University of North Carolina at Chapel Hill. Short term acute care hospital discharge data -patient characteristics. Summary data for all hospitals

Sheps Center for Health Services Research, University of North Carolina at Chapel Hill. Patient county of residence by hospital

Division of Health Services Regulation

Quantifying transmission of Clostridium difficile within and outside healthcare settings

RTI 2017 synthetic population extract

Total population. 2012-2016 American community survey. Community facts. N.d

Modeling inpatient and outpatient antibiotic stewardship interventions to reduce the burden of Clostridioides difficile infection in a regional healthcare network

CMS 2016 national patient-level fee-forservice claims data for CMS beneficiaries provided by Centers for Disease Control and Prevention

The potential for interventions in a long-term acute care hospital to reduce transmission of carbapenem-resistant Enterobacteriaceae in affiliated healthcare facilities

ISDH release preliminary findings about impact of COVID-19 in

Substantial undocumented infection facilitates the rapid dissemination of novel coronavirus (sars-cov-2)

Cumulative incidence and diagnosis of SARS-CoV-2 infection in New York. medRxiv

CDC chief says coronavirus cases may be 10 times higher than reported

COVID-19 community research partnership -study results and data

Impact of non-pharmaceutical interventions (NPIs) to reduce COVID-19 mortality and healthcare demand

An Introduction to Infectious Disease Modelling

COVID-19 model

COVID-19 cases in the United States

University of Geneva. COVID-19 daily epidemic forecasting

Forecasting/_w_d371d176/_w_8ca1db26/?tab=jhu_us_pred&country=North%20Carolina

COVID-19 projections using machine learning

America's COVID Warning System

COVID-19 length of hospital stay: A systematic review and data synthesis

COVID-19 pandemic planning scenarios

Characteristics and Clinical Outcomes of Adult Patients Hospitalized with Covid-19 -Georgia

North Carolina COVID-19 Agent-based Model for Hospitalizations

Acknowledgements: We are grateful for the support and input from Susan Eversole, Stacy Endres-Dighe, Georgiy Bobashev, and Alex Giarrocco of RTI International and from our UNC Health collaborators, and our public health collaborators. This activity was based on a model originally developed through support from CDC's Modeling Infectious Disease in Healthcare (MInD-Healthcare) Network.

The parameter names are the same as the parameters specified in the parameters file used in model runs.

Description Value Source r0Reproduction number; the average number of agents who will become infected with SARS-CoV-2from one agent with SARS-CoV-2 infection 

Bayesian probabilities are used in the model to determine the probability of infection given age and the probability of hospitalization given testing status, age, and presence of comorbidities.Using a Bayesian approach allows for the conditional probabilities needed to account for the underlying variables of interest (testing status, age, presence of comorbidities). Model results can be disaggregated by the underlying variables that are used in the Bayesian analysis, with a more accurate interpretation than without the Bayesian approach, noting the limitations and assumptions of the Bayesian method.

The probability of infection given the age of the agent is important for creating a realistic distribution of agents among COVID-19 agents. This probability is used when selecting agents for infection (i.e., assignment to COVID-19 status). Each agent receives a probability of infection with SARS-CoV-2 given their age. The equation used to calculate this probability is:

The results of this equation are scaled so that the sum of P(infection | age) is 1, since the number of infections is specified by the SEIR equation. The effect of this scaling means that P(infection) does not impact the outcome.

• P(age | infection): Calculated using public dashboard. 33• P(infection): Calculated as the cumulative estimated infections for the model divided by the number of individuals in the synthetic population. This parameter changes for each scenario, although the scaling to 1 eliminates the impact of this parameter.• P(age): Distribution of age groups in the synthetic population.

The probability of hospitalization is related to age and presence of comorbidities, and that tested cases are more likely to seek hospitalization than unreported infections. 31 We use the same Bayesian equation to separately calculate the probability of hospitalization given age and comorbidities depending on testing status. Note that we do not currently account for any changes in testing status (i.e., an untested infection could receive a test after hospitalization).The Bayes equation is:

•