Abstract
Traditional epidemiological models often struggle to account for the dynamic nature of a pandemic, where both disease transmission and case reporting rates fluctuate. This paper introduces a Bayesian state-space framework using mechanistic models (SIR/SEIR) that allows these parameters to evolve over time. By employing Particle Markov Chain Monte Carlo (PMCMC) sampling, the study reconstructs the early stages of the COVID-19 pandemic in New York and Tennessee, providing more accurate incidence estimates and robust short-term predictions.
Introduction: The Challenge of Non-Stationary Parameters
During the COVID-19 pandemic, two critical factors undermined standard "constant-rate" models: mitigation strategies (lockdowns, masks) and testing bottlenecks. If a model assumes a constant reporting rate while testing capacity is actually expanding, it may incorrectly attribute a rise in reported cases to increased viral spread rather than improved detection. This research addresses this bias by treating transmission ($\beta_t$) and reporting ($p_t$) as time-varying stochastic processes.
Methodology: Bayesian State-Space Framework
The authors utilize a hidden Markov model (HMM) where the true infection state is latent, and reported cases are noisy observations.
- Stochastic Evolution: Parameters follow Brownian motion via Stochastic Differential Equations, allowing the model to adapt to real-world changes like policy shifts.
- PMCMC Inference: To handle the high-dimensional, non-linear state space, the researchers use Particle Markov Chain Monte Carlo. This combines Sequential Monte Carlo for state estimation with MCMC for parameter sampling.
Model Performance and Data Reconstruction
The framework was tested on data from New York (early massive spike) and Tennessee (slower initial spread). The model successfully "learned" the underlying dynamics of both states, accurately reconstructing the observed incidence data.
As shown in the figure above, the model reconstructions (b and c) closely track the observed weekly cases (a). In New York, the model estimated that the true number of infections between March and May 2020 was significantly higher (~530,000) than official records (~384,000), highlighting the impact of early testing scarcity.
Dynamics of Disease Transmission
By allowing the reporting rate to absorb the noise of changing testing availability, the estimates for the transmission rate became clearer and more physically plausible.
The transmission rate estimates reflect the impact of public health interventions. In Tennessee (a), a dip in cases in April was partially attributed to a temporary drop in reporting rather than just a decrease in transmission, a nuance missed by simpler models. In New York (b), the transmission rate peaked early and then stabilized as mitigation measures took effect.
Conclusion and Predictive Power
The time-varying models demonstrated superior predictive power, with one-week-ahead forecasts consistently capturing the ground truth within 75% predictive intervals. By accounting for the "invisible" changes in reporting rates, this Bayesian approach provides a more reliable tool for public health decision-making during evolving crises.
Summary updated