Estimating dynamic transmission rates with a Black-Karasinski process in stochastic SIHR models using particle MCMC

TL;DR

This paper introduces a stochastic SIHR epidemic model with a dynamic transmission rate modeled by a Black-Karasinski process, employing particle MCMC for parameter estimation, and demonstrates its effectiveness on synthetic and real flu data.

Contribution

The paper presents a novel stochastic SIHR model with a Black-Karasinski process for transmission rates and applies particle MCMC for joint parameter estimation, improving epidemic modeling accuracy.

Findings

Accurate estimation of BK process parameters except the mean-reversion rate.

Estimation accuracy remains stable despite misspecification of the mean-reversion rate.

Model applied successfully to real flu hospitalization data, aligning with survey estimates.

Abstract

Compartmental models are effective in modeling the spread of infectious pathogens, but have remaining weaknesses in fitting to real datasets exhibiting stochastic effects. We propose a stochastic SIHR model with a dynamic transmission rate, where the rate is modeled by the Black-Karasinski (BK) process - a mean-reverting stochastic process with a stable equilibrium distribution, making it well-suited for modeling long-term epidemic dynamics. To generate sample paths of the BK process and estimate static parameters of the system, we employ particle Markov Chain Monte Carlo (pMCMC) methods due to their effectiveness in handling complex state-space models and jointly estimating parameters. We designed experiments on synthetic data to assess estimation accuracy and its impact on inferred transmission rates; all BK-process parameters were estimated accurately except the mean-reverting rate. We also assess the sensitivity of pMCMC to misspecification of the mean-reversion rate. Our results show that estimation accuracy remains stable across different mean-reversion rates, though smaller values increase error variance and complicate inference results. Finally, we apply our model to Arizona flu hospitalization data, finding that parameter estimates are consistent with published survey data.