Studying COVID-19 transmission in US state prisons using an agent-based modelling approach: a simulation study.

OBJECTIVES: We aim to use an agent-based model to accurately predict the spread of COVID-19 within multiple US state prisons. DESIGN: We developed a semistochastic transmission model of COVID-19. SETTING: Five regional state-owned prisons within North Carolina. PARTICIPANTS: Several thousand incarcerated individuals. PRIMARY AND SECONDARY OUTCOME MEASURES: We measured (1) the observed and simulated average daily infection rate of COVID-19 for each prison studied in 30-day intervals, (2) the observed and simulated average daily recovery rate from COVID-19 for each prison studied in 30-day intervals, (3) the mean absolute percentage error (MAPE) of each prison's summary statistics and the simulated results and (4) the parameter estimates of key predictors used in the model. INTRODUCTION: The COVID-19 pandemic disparately affected incarcerated populations in the USA, with severe morbidity and infection rates across the country. In response, many predictive models were developed to help mitigate risk. However, these models did not feature the systemic factors of prisons, such as vaccination rates, populations and capacities (to determine overcrowding) and design and were not generalisable to other prisons. METHODS: An agent-based model that used geospatial contact networks and compartmental transmission dynamics was built to create predictive microsimulations that simulated COVID-19 outbreaks within five North Carolinian regional prisons between July 2020 and June 2021. The model used the characteristics of an outbreak's initial case size, a given facility's capacity and its incarcerated vaccination rate as additional parameters alongside traditional susceptible-exposed-infected-recovered transmission dynamics. By fitting the model to each prison's data using approximate Bayesian computation methods, we derived parameter estimates that reasonably modelled real-world results. These individualised estimates were then averaged to produce generalised parameter estimates for North Carolina state prisons overall. RESULTS: Our model had a mean average MAPE score of 23.0 across all facilities, meaning that it reasonably forecasted facilities' average daily positive and recovery rates of COVID-19. Our model estimated an average incarcerated vaccination rate of 54% across all prisons (with a 95% CI of ±0.12). In addition, the prisons of this study were estimated to be operating at 90% of their capacity on average (95% CI ±0.16). Given the high levels of COVID-19 observed in these prisons, which averaged over one-third positive tests on respective 1-day maxima, we conclude that vaccination levels were not sufficient in curbing COVID-19 outbreaks, and high occupancy levels likely exacerbated the spread of COVID-19 within prisons.In addition, data gaps in facilities without recorded daily testing resulted in poor spread predictions, demonstrating how important consistent data release practices are in incarcerated settings for accurate tracking and prediction of outbreaks. CONCLUSION: The findings of this study better quantify how spatial contact networks and facility-level characteristics unique to congregate living facilities can be used to predict infectious disease spread. Our approach also highlights the need for increased vaccination efforts and potential capacity reductions to mitigate COVID-19 transmission in prisons.

Duke Scholars

Author Lauren Brinkley-Rubinstein Population Health Sciences