TL;DR
This paper introduces Poisson Approximate Likelihood (PAL) methods for fast and consistent inference in complex epidemiological models. By approximating likelihoods with Poisson distributions, PALs offer a computationally efficient alternative to traditional methods, enabling more scalable analysis of large and heterogeneous populations.
What Problem Does This Paper Address?
Epidemiological modeling relies heavily on compartmental models to understand disease spread. However, fitting these models to data, especially complex ones with many compartments or large populations, becomes computationally prohibitive. Traditional methods often involve approximations that simplify the model or require computationally intensive simulations. This paper addresses the challenge of developing a fast, scalable, and theoretically sound inference method for complex, finite-population, stochastic compartmental models, particularly when dealing with large populations.
Methods
The core innovation is the Poisson Approximate Likelihood (PAL). Instead of directly calculating the exact likelihood (which is often intractable), PALs approximate it using Poisson distributions. This approximation is derived from the filtering equations of finite-population stochastic compartmental models. The key steps involve:
- Specifying a finite-population, stochastic, discrete-time compartmental model.
- Combining this with an observation model to derive discrete-time filtering equations.
- Recursively approximating these filtering equations using Poisson distributions to define the PAL.
- Leveraging the large population limit to establish the consistency of parameter estimators derived from the PAL.
This approach avoids computationally expensive simulations and relies on elementary arithmetic operations, making it significantly faster. The paper also discusses how PALs can be integrated with other techniques like Sequential Monte Carlo (SMC) to handle over-dispersion and how they can be used with automatic differentiation for efficient parameter estimation.
Datasets
The paper evaluates its methods on several real-world datasets:
- Influenza data: An age-structured model of influenza from a 1957 outbreak in Wales.
- Rotavirus data: A model of rotavirus infections in Germany across three age groups.
- Measles data: A meta-population model of measles in 40 UK cities.
- Boarding school influenza outbreak data: A SIR model of an influenza outbreak in a boarding school.
The paper also uses simulated data to demonstrate the performance of its methods.
Key Insights and Ideas
- Computational Efficiency: PALs offer a dramatic speedup compared to traditional methods, especially for complex models and large populations, by avoiding simulations and relying on simple arithmetic. This opens the door to analyzing more realistic and detailed epidemiological models.
- Theoretical Guarantees: The paper provides theoretical guarantees for the consistency of PAL estimators in the large population limit, which is a significant contribution for approximate likelihood methods.
- Scalability: The method scales well with population size, a crucial factor for real-world epidemiological modeling where populations can be very large.
- Flexibility: PALs can be applied to a broad class of partially observed stochastic compartmental models, including those with complex structures like age-stratification and meta-populations.
- Over-dispersion Handling: PALs can be combined with SMC methods to effectively handle over-dispersion, a common phenomenon in epidemiological data that can be missed by simpler models.
Key Results
- Consistency: The paper establishes theoretical consistency of maximum PAL estimators in the large-population regime.
- Speed: PALs are significantly faster than traditional methods, with computational costs independent of population size. For example, in the boarding school influenza outbreak, PAL evaluations were approximately 100 times faster than LNA likelihood evaluations.
- Model Fit: In the rotavirus example, PAL-based models with over-dispersion achieved substantially better AIC scores than simpler models, indicating a better fit to the data.
- Parameter Estimation: The methods were successfully applied to real-world datasets (influenza, rotavirus, measles), demonstrating their practical utility in estimating model parameters and understanding disease dynamics.
Limitations and Future Work
- Analytical Tractability: While PALs are analytically tractable for certain models, extending them to models with over-dispersion that cannot be analytically handled (e.g., beta-binomial or negative binomial distributions) might require further approximations.
- Optimization: The paper notes that special optimization techniques were not devised for parameter estimation, suggesting opportunities for further research in this area to improve computational efficiency.
- Covariates: While discrete covariates can be handled by extending the compartmental structure, covariates taking infinitely many distinct values might necessitate further approximations.
- Model Specification: The paper acknowledges that when the model is not correctly specified, the PAL estimates may converge to limits different from the true parameters, highlighting the importance of model choice.
Summary updated