Why is it difficult to accurately predict the COVID-19 epidemic?

Abstract

Since the COVID-19 outbreak in Wuhan City in December of 2019, numerous model predictions on the COVID-19 epidemics in Wuhan and other parts of China have been reported. These model predictions have shown a wide range of variations. In our study, we demonstrate that nonidentifiability in model calibrations using the confirmed-case data is the main reason for such wide variations. Using the Akaike Information Criterion (AIC) for model selection, we show that an SIR model performs much better than an SEIR model in representing the information contained in the confirmed-case data. This indicates that predictions using more complex models may not be more reliable compared to using a simpler model. We present our model predictions for the COVID-19 epidemic in Wuhan after the lockdown and quarantine of the city on January 23, 2020. We also report our results of modeling the impacts of the strict quarantine measures undertaken in the city after February 7 on the time course of the epidemic, and modeling the potential of a second outbreak after the return-to-work in the city.

Keywords: Bayesian inference; COVID-19 epidemic in Wuhan; Model selection; Nonidentifiability; Peak time of epidemic; Quarantine; SIR and SEIR models.

Fig. 1

Transfer diagrams for an SIR and an SEIR model for COVID-19 in Wuhan.

Fig. 2

Linkage between transmission rate β and diagnosis rate ρ.

Fig. 3

Model projections using two likely β-ρ combinations, corresponding to two endpoints on the curve in Fig. 2 (b). Day 0 is January 21, 2020.

Fig. 4

Distributions of estimated peak time (a) and control reproduction number ${\mathcal{R}}_c$ (b) for COVID-19 epidemic in Wuhan after lockdown. The dashed lines represent the 95% prediction intervals for the time course of COVID-19 epidemic in Wuhan after lockdown (c) and (d). Day 0 in simulations is set at January 21, 2020.

Fig. 5

Model predictions of time courses of COVID-19 epidemic in Wuhan with three different ranges of diagnosis rate ρ: $\left(0.02,0.03\right)$, $\left(0.05,0.1\right)$, and $\left(0.2,1\right)$. Day 0 in the simulations is set at January 21, 2020.

Fig. 6

Predictions of the COVID-19 epidemic in Wuhan with more strict quarantine measures after February 7, 2020. Impacts of reductions in transmission rate β and increases in diagnosis rate ρ are shown in (a). Impacts of only reducing the transmission rate (b) or only increasing the diagnosis rate (c) are also shown for comparison purposes.

Fig. 7

Model predictions of time courses of COVID-19 epidemic in Wuhan with return to work on (a) February 24, (b) March 2, and (c) March 31, 2020.