Optimal design of experimental epidemics

Dan Pagendam, P. K. PollettJournal of Statistical Planning and Inference 2012 · DOI

TL;DR

This paper tackles the challenge of designing efficient experiments to study disease spread in controlled settings. By employing advanced statistical methods like Gaussian diffusion approximations and the Cross-Entropy method, the authors develop a framework to determine optimal observation times, initial conditions, and replication strategies for SI, SIS, and partially observed SIR epidemic models, leading to more precise parameter estimates.

What Problem Does This Paper Address?

Studying disease transmission in controlled experimental epidemics is crucial for understanding disease dynamics and estimating key parameters like transmission and recovery rates. However, designing these experiments efficiently is a complex problem. Traditional approaches often rely on simplifying assumptions about continuous observation or equally spaced sampling, which can be impractical or lead to suboptimal data collection. Furthermore, for complex epidemic models, deriving analytical expressions for likelihoods and their associated Fisher information matrices (essential for optimal design) is often intractable. This paper addresses the need for a principled, data-driven approach to experimental design that can handle realistic, discretely observed epidemic models and provide optimal strategies for data collection.

Methods

The core of the paper's methodology lies in using Gaussian diffusion approximations to model the dynamics of density-dependent Markov processes (DDMPs), which are commonly used for epidemic modeling. This approximation allows for the derivation of a tractable likelihood function even for discretely observed processes, especially when the population size is large.

To find the optimal experimental designs, the authors adopt the D-optimality criterion, which aims to maximize the determinant of the Fisher information matrix. This criterion leads to designs that provide the most precise estimates of model parameters.

Since finding D-optimal designs for non-linear models is challenging and often results in locally optimal solutions, the authors employ the Cross-Entropy (CE) method, a stochastic optimization technique. This method iteratively generates candidate designs (defined by observation times, initial conditions, and number of replicates) and refines them based on their performance (measured by the D-optimality criterion) until convergence. The CE method is detailed in Appendix II and involves updating parameters of beta probability distributions used to generate design elements.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and simulation-based evaluation of experimental designs for epidemic models.

Key Insights and Ideas

  • Diffusion Approximations are Powerful: Gaussian diffusion approximations provide a unified and computationally efficient way to handle the likelihood of discretely observed epidemic models, even complex ones.
  • D-Optimality Guides Efficient Data Collection: The D-optimality criterion effectively guides the selection of observation times and initial conditions to maximize the precision of parameter estimates.
  • Cross-Entropy Method for Design Optimization: The CE method offers a robust and effective way to search for optimal experimental designs in a complex, multi-parameter space.
  • Initial Conditions Matter Significantly: The choice of initial number of infected individuals can dramatically impact the precision of parameter estimates, with certain initial values leading to a substantial loss of information.
  • Replication vs. Scale: For a fixed total population size, running multiple smaller replicate experiments can be more efficient than one large experiment, though the diffusion approximation may not accurately capture this for comparing different population sizes.
  • Robustness to Parameter Misspecification: D-optimal designs can be sensitive to errors in the assumed model parameters, particularly when parameters are understated, highlighting the importance of accurate prior information or robust design strategies.

Key Results

  • The paper demonstrates that D-optimal designs can significantly improve the accuracy and precision of maximum likelihood estimators compared to simpler, equidistant designs for SI, SIS, and poSIR epidemics.
  • For the SIS epidemic, the choice of initial proportion infected ($x_0$) has a dramatic effect on estimator precision, with poor performance observed when $x_0$ is close to the equilibrium value ($x_{eq}$).
  • The D-efficiency of designs for the poSIR epidemic is shown to be sensitive to the initial proportion infected ($x_0$), with good agreement between theoretical predictions from diffusion approximations and simulation-based results.
  • The study suggests that for a fixed total population, using a small number of replicate experiments is generally more efficient than a single large experiment, a finding supported by simulations.
  • Relative efficiency plots show that D-optimal designs can be sensitive to parameter misspecification, with the efficiency declining when parameters are understated.

Limitations and Future Work

  • Locally Optimal Designs: The D-optimal designs obtained are locally optimal, meaning they depend on the initial guess of the parameters. The paper acknowledges this and discusses robustness to parameter misspecification.
  • Large Population Assumption: The diffusion approximation relies on the system size (population size) being sufficiently large. The paper notes that this approximation may not be accurate when comparing experiments with different population sizes.
  • Computational Cost: While the CE method is efficient, finding optimal designs can still be computationally intensive, especially for high-dimensional models.
  • Future Work: The authors suggest that the methodology can be extended to a wider range of epidemic models and other density-dependent Markov processes, indicating potential applications in fields like experimental epidemiology and biological control. Further investigation into robust design strategies that are less sensitive to parameter misspecification would also be valuable.

Summary updated

BibTeX
@article{pagendam2012optimal,
  author = {Dan Pagendam and P. K. Pollett},
  title = {{Optimal design of experimental epidemics}},
  journal = {Journal of Statistical Planning and Inference},
  year = {2012},
  doi = {10.1016/j.jspi.2012.09.011},
}
Figure