Statistical implications of relaxing the homogeneous mixing assumption in time series Susceptible-Infectious-Removed models

Luis D. J. Martinez Lomeli, Michelle N. Ngo, Jon Wakefield, Babak Shahbaba, Vladimir N. MininarXiv 2021 · DOI · arXiv

TL;DR

Time-series Susceptible-Infectious-Removed (TSIR) models are popular in epidemiology for their computational speed, but they rely on a "homogeneous mixing" assumption that rarely holds in the real world. This paper demonstrates that common methods used to relax this assumption introduce significant statistical bias, often exaggerating seasonal fluctuations in disease transmission. By comparing TSIR to more rigorous Bayesian SIR models, the authors show that researchers should exercise extreme caution when interpreting parameters from TSIR models that incorporate non-homogeneous mixing.

What Problem Does This Paper Address?

Infectious disease modeling usually starts with the "law of mass action," which assumes that any susceptible person is equally likely to meet any infectious person (homogeneous mixing). To make models more realistic, researchers often "relax" this by adding a mixing parameter ($\alpha$), raising the number of infectious individuals to a power.

While standard stochastic SIR models are the gold standard for capturing outbreak dynamics, they are computationally "intractable"—calculating the likelihood of the data is incredibly difficult. TSIR models were developed as a fast, discrete-time approximation based on branching processes. However, the authors argue that the mathematical justification for TSIR breaks down the moment you move away from homogeneous mixing. This creates a gap: we are using a fast approximation (TSIR) in a scenario (non-homogeneous mixing) where it is no longer a valid approximation of the underlying biological process.

Methods

The authors compare three distinct Bayesian approaches to parameter inference:

  1. BayesTSIR: The standard approach using a negative binomial approximation for the infection process.
  2. BayesPureBirth: A model that uses exact transition probabilities but assumes no recoveries occur within the observation interval.
  3. BayesSIR: The most rigorous model, which uses the MultiBD R package to calculate highly accurate numerical approximations of the transition probabilities for the full birth-death process of an epidemic.

The team used Markov chain Monte Carlo (MCMC) methods to estimate the posterior distributions of the infection rate ($\beta$), the mixing parameter ($\alpha$), and the recovery rate ($\gamma$).

Datasets

  • Synthetic Data: The authors generated an ensemble of 629 stochastic SIR trajectories using the Gillespie algorithm. This allowed them to test the models against a "ground truth" where the parameters ($\alpha=0.8, \beta=0.0045, \gamma=1$) were known.
  • London Measles Dataset: A classic benchmark in epidemiology containing biweekly incidence records from 1944 to 1964. The authors focused on the pre-vaccination era (1944–1951) to study how transmission rates fluctuated with the school calendar.

Key Insights and Ideas

  • The Mathematical Mismatch: The negative binomial likelihood used in TSIR is derived from a linear birth process. When a mixing parameter $\alpha \neq 1$ is introduced, the process becomes non-linear. The authors show that the TSIR formula simply cannot be interpreted as a valid approximation of a non-homogeneous SIR model.
  • Exaggerated Seasonality: One of the most striking "aha" moments is that TSIR tends to over-estimate the drop in infection rates during periods of low activity (like school breaks). It creates a "louder" signal of seasonality than actually exists in the data.
  • Recovery Rates Matter: Unlike standard TSIR, which usually assumes a fixed recovery rate to match the observation window, the BayesSIR approach can successfully estimate the recovery rate ($\gamma$), providing a more nuanced view of the disease's natural history.

Key Results

  • Bias in Simulations: In the synthetic study, BayesSIR successfully recovered the true parameters. In contrast, BayesTSIR was systematically biased, producing the highest Root Mean Squared Error (RMSE) and the lowest "coverage" (its 95% credible intervals only contained the true value about 49% of the time).
  • Measles Analysis: When applied to the London measles data, BayesTSIR estimated a much sharper decrease in transmission during school holidays compared to BayesSIR.
  • Mixing Evidence: BayesSIR provided much stronger statistical evidence for non-homogeneous mixing ($\alpha < 1$) in the measles data, whereas the BayesTSIR credible intervals for $\alpha$ often included 1.0 (homogeneous mixing), potentially masking the true dynamics.

Limitations and Future Work

  • Computational Intensity: The "exact" BayesSIR method relies on matrix exponentiation and numerical approximations that become computationally prohibitive as population sizes and case counts grow into the millions. Currently, it is best suited for datasets with hundreds or thousands of cases.
  • Scaling Up: The authors suggest that for very large populations, researchers might still need to rely on other approximations like the Linear Noise Approximation (LNA) or particle filters, rather than the TSIR negative binomial shortcut.
  • Forecasting: While this paper focuses on inference (understanding the past), the authors note that future research should investigate how these TSIR biases impact forecasting (predicting the future) of epidemic waves.

Summary updated

Abstract

Infectious disease epidemiologists routinely fit stochastic epidemic models to time series data to elucidate infectious disease dynamics, evaluate interventions, and forecast epidemic trajectories. To improve computational tractability, many approximate stochastic models have been proposed. In this paper, we focus on one class of such approximations -- time series Susceptible-Infectious-Removed (TSIR) models. Infectious disease modeling often starts with a homogeneous mixing assumption, postulating that the rate of disease transmission is proportional to a product of the numbers of susceptible and infectious individuals. One popular way to relax this assumption proceeds by raising the number of susceptible and/or infectious individuals to some positive powers. We show that when this technique is used within the TSIR models they cannot be interpreted as approximate SIR models, which has important implications for TSIR-based statistical inference. Our simulation study shows that TSIR-based estimates of infection and mixing rates are systematically biased in the presence of non-homogeneous mixing, suggesting that caution is needed when interpreting TSIR model parameter estimates when this class of models relaxes the homogeneous mixing assumption.
BibTeX
@misc{lomeli2021statistical,
  author = {Luis D. J. Martinez Lomeli and Michelle N. Ngo and Jon Wakefield and Babak Shahbaba and Vladimir N. Minin},
  title = {{Statistical implications of relaxing the homogeneous mixing assumption in time series Susceptible-Infectious-Removed models}},
  year = {2021},
  eprint = {2112.03186},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2112.03186},
}
Figure