Epidemiological models, parameters, and prediction of the COVID-19 outbreak in a southwest Georgia hotspot

Abstract

Background: We investigate epidemiological models, their parameters, and the models’ predictive performance of daily new cases during the first three weeks of the SARS-CoV-2 outbreak in a southwest Georgia hotspot. Methods: We fit stochastic versions of the classical SIR, a network-based SIR, an SEIR model with two different priors on the latent period rate parameter, and a simple doubling time model to SARS-CoV-2 patients’ date of hospital admission and length of symptoms data from the Phoebe Putney Health System. Results: The estimated basic reproductive numbers and 95% Bayesian credible intervals for each of the models were: 1.98 (1.63, 2.45) for the SIR model, 1.99 (1.62, 2.48) for the network-based SIR model, 2.04 (1.64, 2.68) for the SEIR model with noninformative prior, and 3.37 (2.42, 4.68) for the SEIR model with informative prior. The SIR and network-based SIR models performed similarly in terms of predicting new cases each day, with median absolute error (MAE) of 3 cases. They had better predictive performance compared to the SEIR (MAE=13) with noninformative prior and doubling time models (MAE=42). The SEIR with informative prior on the incubation rate had an MAE of 4. Conclusions: These results indicate a herd immunity of between 50% and 70% were necessary to prevent further spread. The simple doubling time model consistently overpredicts daily new cases. Using duration of symptoms data from the severe cases with SIR-type models led to improved real-time predictions of cumulative daily cases compared to the doubling time and SEIR model with noninformative prior. The latent period parameter for the SEIR model was not identifiable from the data, but good predictions were achieved with an informative prior on this parameter using results from the literature.