Abstract
We introduce a new method for inference in stochastic epidemic models which uses recursive multinomial approximations to integrate over unobserved variables and thus circumvent likelihood intractability. The method is applicable to a class of discrete-time, finite-population compartmental models with partial, randomly under-reported or missing count observations. In contrast to state-of-the-art alternatives such as Approximate Bayesian Computation techniques, no forward simulation of the model is required and there are no tuning parameters. Evaluating the approximate marginal likelihood of model parameters is achieved through a computationally simple filtering recursion. The accuracy of the approximation is demonstrated through analysis of real and simulated data using a model of the 1995 Ebola outbreak in the Democratic Republic of Congo. We show how the method can be embedded within a Sequential Monte Carlo approach to estimating the time-varying reproduction number of COVID-19 in Wuhan, China.
Addressing Likelihood Intractability
Epidemiologists rely on compartmental models, such as the SEIR framework, to track the movement of individuals through Susceptible, Exposed, Infective, and Recovered states. While deterministic versions of these models are mathematically straightforward, they often fail to capture the inherent stochasticity—the "statistical noise"—of real-world outbreaks. This randomness is critical for understanding small-scale outbreaks or the early stages of a pandemic.
However, stochastic versions of these models present a significant challenge: calculating the "likelihood" (how well a model fits the data) requires summing over a massive number of possible hidden transmission paths. Current gold-standard solutions, such as Approximate Bayesian Computation (ABC) or Markov Chain Monte Carlo (MCMC), are computationally expensive, require heavy simulation, and often necessitate expert tuning of parameters to achieve convergence.
Methodology: Multinomial Filtering
The researchers introduce a Multinomial Filtering approach that bypasses the need for expensive simulations. Instead of generating thousands of possible trajectories, the method uses a recursive mathematical approximation:
- Prediction Step: The distribution of individuals in different compartments at the next time step is approximated using a multinomial distribution.
- Update Step: As new data (e.g., daily hospital counts) arrives, a "shifted multinomial" approximation updates the model’s state beliefs.
This approach offers several advantages: it requires no forward simulation, is entirely tuning-free (no proposal distributions or tolerance levels), and scales with the number of compartments rather than the total population size. This makes the method exceptionally fast, even for large urban populations.
Key Results and Validation
The method was rigorously tested against both synthetic and real-world datasets. In the analysis of the 1995 Ebola outbreak in the Democratic Republic of Congo, the multinomial filter recovered true parameters with high precision, matching the performance of much slower MCMC methods while identifying distinct infection rate scenarios that previous methods struggled to isolate.
As shown in the table above, the multinomial approximation (MLE via EM-algorithm) provides estimates very close to the true values and comparable to the more computationally intensive MCMC approaches.
Application to COVID-19 in Wuhan
The authors applied their stochastic multinomial filter to a 15-compartment model of the 2020 COVID-19 outbreak in Wuhan. By replacing the original deterministic core with their stochastic filter, they were able to estimate the time-varying reproduction number ($R_t$) with more realistic uncertainty bounds.
The results indicated that after travel restrictions were implemented, $R_t$ likely dropped closer to 1 than previously estimated. The wider credible intervals produced by this method more accurately reflect the inherent noise of real-world data compared to traditional ODE-based studies.
Conclusion
This research demonstrates that stochasticity matters, even in large populations. The multinomial filtering technique provides a computationally efficient, scalable, and accurate alternative to simulation-based inference. Future work may involve embedding this framework into larger Sequential Monte Carlo systems to handle complex, multi-city transmission dynamics.
Summary updated