metadata-NotreDame-FRED.txt

team_name: University of Notre Dame
model_name: FRED
model_abbr: NotreDame-FRED
model_version: 2020-08-17
model_contributors: Guido Espana <guido.espana@nd.edu>, Sean Cavany <scavany@nd.edu>, Sean Moore <smoore15@nd.edu>, Alex Perkins <taperkins@nd.edu>
website_url: https://github.com/confunguido/FRED
license: BSD Simplified
team_model_designation: primary
methods: FRED is an agent-based model developed for influenza, and modified
  to represent COVID-19 transmission
  More informtaion: https://www.sciencedirect.com/science/article/pii/S1755436521000396
modeling_NPI: We fit the maximum proportion of people sheltering in place. We then use google mobility data to estimate trends in people sheltering in place. The maximum level of face-mask wearing compliance is included in the calibration (see below). The trend in mask adoption is fitted to google shopping trends. For future interventions, we assume no further NPIs - existing NPIs are held constant.
compliance_NPI: See above
contract_tracing: Not implemented
testing: We fit to testing data up until the end of the COVID-tracking project, but don't directly model the testing process. We also calibrate a parameter describing the ratio of (1-specificity) and sensitivity of the testing process (i.e. not of the test itself).
vaccine_efficacy_transmission: We assume vaccine affects susceptibility but not infectiousness. Vaccine efficacy varies by variant, but not by age. We assume the newly emerging variant has the same efficacy as Delta.
vaccine_efficacy_infection: non-delta: 76%, Delta: 72%
vaccine_efficacy_symptoms: non-delta: 90%, Delta: 85%
vaccine_efficacy_hospitalization: 96%
vaccine_efficacy_delay: 15 days following second dose
vaccine_hesitancy: We don't directly model hesistancy: we have an upper limit of 90% for vaccine uptake, but we then fit vaccine uptake to the past data on vaccine uptake and assume the same rate of increase in uptake continues in the projection. For the 5-11 age-bracket we assume the same upper limit as other age groups, and the same rate of increase in uptake as the 12-17 age group.
vaccine_immunity_duration: We assume no waning in vaccine-derived immunity on the time-scales of the simulation.
natural_immunity_duration: After 8 months natural immunity wanes to 87%
case_fatality_rate: Derived from below based on probability of symptoms by age
infection_fatality_rate: Age dependent, as given by Verity et al.
asymptomatics: Age-dependent, based on Davies et al.
age_groups: 10-year age bands up-to age 79 inclusive, then 80+
importations: For variants, we take data from covariants.org at national level, and estimate different scaling factors for each of the variant to match estimated prevalence by state. For Omicron, we assume that importations started early in November 2021. 
confidence_interval_method: We simulate the model 100 times, sampling from the posterior distribution of each of the calibrated parameters
calibration: We have a three stage fitting process. In the first stage (up until Fall 2020) we fit a range of parameters including transmissibility. In the second stage, Fall 2020 - Dec 2020: keep all previous parameters at final values in Fall 2020, and calibrate school capacity and masking parameters. in the Final stage, Dec 2020 - present: fix school parameters and calibrate transmissibility and immune escape of variants. For each stage, we simulate 3,000 particles, and sample according to the particle likelihood the particles in the next stage. We fit to the targets described in data_inputs below. 
spatial_structure: Locations and location types explicitly described. Agents move around these locations
team_funding: NSF RAPID grant
data_inputs: whole period: Deaths - JHU. until fall 2020: COVID-19 tracking project testing data (positives and negatives), seroprevalence by age in Indiana.
citation:
Espana G, Cavany S, Oidtman R, Barbera C, Costello A, Lerch A, Poterek M, Tran Q, Wieler A, Moore SM, Perkins TA. "Impacts of K-12 school reopening on the COVID-19 epidemic in Indiana, USA." Epidemics, 2021. https://www.sciencedirect.com/science/article/pii/S1755436521000396

methods_long: "We used an existing agent-based model, FRED (Framework for Reconstructing\
  \ Epidemic Dynamics) which was developed by the university of Pittsburgh in response\
  \ to the 2009 influenza pandemic (Grefenstette et al. 2013). We modified the model\
  \ parameters to represent the natural history of COVID-19 and calibrated a set of\
  \ parameters to reproduce current trends of deaths due to COVID-19 in the US. FRED\
  \ simulates the spread of the virus in a population by recreating interactions among\
  \ people on a daily basis. To accurately represent the population of each of the\
  \ states simulated, we used a synthetic population of the US that represents demographic\
  \ and geographic characteristics of the population in 2010 (Wheaton 2012). Each\
  \ human is modeled as an agent that visits a set of places defined by their activity\
  \ space. This activity space contains places such as households, schools, workplaces,\
  \ or neighborhood locations. Transmission of SARS-CoV-2 can occur when an infected\
  \ person visits the same location as a susceptible person on the same day. Numbers\
  \ of contacts per person are specific to each place type. For instance, school contacts\
  \ do not depend on the size of the school but the age of the agent. Infected agents\
  \ have a probability to stay at home if they develop symptoms. Those who do not\
  \ stay at home continue their daily activities. Public health interventions are\
  \ included in the model to represent the changes in agents\u2019 behavior in response\
  \ to an epidemic.  In this study, we limited the interventions to school closure\
  \ and shelter in place. Schools are closed on a specific date to represent state-level\
  \ guidelines (IHME COVID-19 health service utilization forecasting Team and Murray\
  \ 2020). In the case that schools are closed, students limit their activity space\
  \ to household or neighborhood locations. Shelter-in-place interventions reduce\
  \ each agent\u2019s activity space to their household at a compliance level from\
  \ 0-100%, which was estimated as part of the model calibration. Agents who do not\
  \ comply with the shelter-in-place orders continue with their daily routines. We\
  \ used state-level orders to determine the date at which people are advised to shelter\
  \ in place (IHME COVID-19 health service utilization forecasting Team and Murray\
  \ 2020).  See the methods file in our repo for more detailed methods and model limitations.\n\
  Updates 2020-05-11 - We updated our likelihood used in our calibration, switching\
  \ from a Poisson to a negative binomial distribution in order to allow for greater\
  \ over-dispersion in the distribution of deaths. - We changed the way that we calculate\
  \ the proportion of the population that is sheltering in place. We no longer use\
  \ the dates at which government advisories were issued. Instead, we now exclusively\
  \ use the google mobility data (https://www.google.com/covid19/mobility/) on journeys\
  \ made to residential locations. We calibrate the maximum compliance to shelter\
  \ in place, and use a generalized additive model fitted to the google data to forecast\
  \ movement in the coming 5 weeks.
  Updates 2021-09-13 - In order to run for the whole of the United States, we calibrate\
  \ the model to data from Indiana from the start of the pandemic until present. For\
  \ other states we use infection estimates from CovidEstim\
  \ (https://doi.org/10.1101/2020.06.17.20133983), a timeseries of the frequency of each variant\
  \ from https://covariants.org/, and vaccination data to initialize the infection and\
  \ vaccination histories of the agents in FRED. We then use the parameters from the\
  \ Indiana calibration to run the forecasts in FRED for each state.\
  \ For more details on this approach, see the methods\
  \ in https://www.sciencedirect.com/science/article/pii/S1755436521000396"