Covid19 Reproduction Number: Credibility Intervals by Blockwise Proximal Monte Carlo Samplers

Gersende Fort, Barbara Pascal, Patrice Abry, Nelly Pustelnik

TL;DR

This paper introduces novel Monte Carlo sampling methods, specifically "Blockwise Proximal-Gradient" (PGdec and PGdual) samplers, to accurately estimate the COVID-19 reproduction number (R) and its associated credibility intervals. These methods are crucial for reliable pandemic monitoring, especially when dealing with noisy and incomplete data, by providing a measure of uncertainty around R estimates.

What Problem Does This Paper Address?

Monitoring the COVID-19 pandemic's intensity is critical for public health policy. The reproduction number (R) is a key metric for this, quantifying how many new infections arise from a single infected individual. While previous methods could estimate R, they often lacked the ability to provide "credibility intervals" – a range of plausible values for R. This uncertainty is vital, especially during rapid pandemic shifts, to inform decision-making. Existing methods also struggled with the poor quality of real-world COVID-19 data (outliers, missing values). This paper aims to bridge this gap by developing robust statistical models and efficient sampling techniques that can provide these crucial credibility intervals, even with imperfect data.

Methods

The core of this work lies in developing new statistical models and sampling algorithms.

  • Bayesian Framework: The problem is framed within a Bayesian context, where the reproduction number (R) and potential outlier indicators (O) are treated as random variables with a posterior distribution. This allows for the quantification of uncertainty.
  • Nonsmooth Priors: To handle data corruption (outliers, missing values), the model incorporates nonsmooth prior distributions. This makes the estimation robust but introduces mathematical challenges, as standard optimization and sampling methods struggle with nonsmooth functions.
  • Blockwise Proximal-Gradient Samplers (PGdec and PGdual): To address the nonsmoothness, the authors propose two novel Monte Carlo Markov Chain (MCMC) sampling algorithms: PGdec and PGdual. These algorithms are designed to efficiently explore the posterior distribution by leveraging the "proximal operator" – a mathematical tool for handling nonsmooth convex functions. They decompose the problem into blocks, allowing for more efficient updates.
  • Metropolis-Hastings Framework: Both PGdec and PGdual are implemented within a Metropolis-Hastings framework, which involves proposing new states and accepting or rejecting them based on a probability ratio. This ensures that the samples eventually converge to the true posterior distribution.
  • Gibbs Samplers: Extensions of PGdec and PGdual, called Gibbs PGdec and Gibbs PGdual, are also developed. These leverage the blockwise structure to sample from conditional distributions, often leading to improved efficiency.

Datasets

The methods are evaluated using real-world COVID-19 data:

  • Johns Hopkins University COVID-19 Data: Daily new infection counts from over 200 countries and territories, collected from January 1st, 2020, onwards. This is a standard and widely used dataset for COVID-19 research.

Key Insights and Ideas

  • Credibility Intervals are Essential: The paper strongly emphasizes that single-point estimates of R are insufficient for effective pandemic monitoring. Credibility intervals provide a critical measure of uncertainty, allowing for more nuanced interpretation and decision-making.
  • Nonsmoothness is a Feature, Not a Bug: The authors ingeniously use nonsmooth priors to achieve robustness against data corruption, turning a data quality issue into a modeling advantage.
  • Proximal Operators are Key for Nonsmoothness: The development of PGdec and PGdual highlights the power of proximal operators in efficiently sampling from complex, nonsmooth posterior distributions, which are common in inverse problems.
  • Blockwise Structure Enables Efficiency: Decomposing the problem into blocks and developing specialized samplers for these blocks (PGdec, PGdual) significantly improves computational efficiency and mixing properties of the MCMC chains.
  • Gibbs Samplers Offer Performance Gains: The Gibbs variants of the proposed samplers generally outperform their Metropolis-Hastings counterparts, suggesting that exploiting conditional distributions can lead to faster convergence.
  • Data Quality Drives Uncertainty: The size of the credibility intervals for R is directly linked to data quality; noisier data leads to wider intervals, reflecting greater uncertainty.

Key Results

  • Successful Credibility Interval Estimation: The proposed PGdual samplers (especially the Gibbs variant) successfully generate credible intervals for R and denoised infection counts (Z(D)).
  • Improved Robustness: The methods demonstrate robustness to data corruption, providing more realistic estimates of pandemic intensity compared to raw counts.
  • Performance Comparison: Extensive comparisons across twelve different sampling strategies (including Random Walk samplers) show that the Gibbs PGdual O sampler consistently offers the best performance in terms of mixing properties and efficiency, as measured by metrics like distance to MAP, ACF, and GR statistic.
  • Practical Applicability: The developed tools are shown to be effective on real-world data from various countries, providing smooth and interpretable estimates of pandemic evolution. The authors make their MATLAB code publicly available for reproducibility.

Limitations and Future Work

  • Hyperparameter Tuning: While the paper uses a single set of hyperparameters (AR and λ₀) for all countries, the authors note that automated and data-driven estimation of these hyperparameters is an area for future investigation.
  • Computational Cost: While the proposed methods are more efficient than naive approaches, MCMC sampling can still be computationally intensive, especially for very long time series or large numbers of countries.
  • Generalizability: While the methods are presented for COVID-19, their applicability to other epidemiological models or inverse problems with nonsmooth priors is a promising avenue for future work. The authors explicitly mention their methods are suitable for a broader range of applications, including image processing.

Summary updated

Abstract

Monitoring the Covid19 pandemic constitutes a critical societal stake that received considerable research efforts. The intensity of the pandemic on a given territory is efficiently measured by the reproduction number, quantifying the rate of growth of daily new infections. Recently, estimates for the time evolution of the reproduction number were produced using an inverse problem formulation with a nonsmooth functional minimization. While it was designed to be robust to the limited quality of the Covid19 data (outliers, missing counts), the procedure lacks the ability to output credibility interval based estimates. This remains a severe limitation for practical use in actual pandemic monitoring by epidemiologists that the present work aims to overcome by use of Monte Carlo sampling. After interpretation of the nonsmooth functional into a Bayesian framework, several sampling schemes are tailored to adjust the nonsmooth nature of the resulting posterior distribution. The originality of the devised algorithms stems from combining a Langevin Monte Carlo sampling scheme with Proximal operators. Performance of the new algorithms in producing relevant credibility intervals for the reproduction number estimates and denoised counts are compared. Assessment is conducted on real daily new infection counts made available by the Johns Hopkins University. The interest of the devised monitoring tools are illustrated on Covid19 data from several different countries.
BibTeX
@article{fort2023covid,
  author = {Gersende Fort and Barbara Pascal and Patrice Abry and Nelly Pustelnik},
  title = {{Covid19 Reproduction Number: Credibility Intervals by Blockwise Proximal Monte Carlo Samplers}},
  journal = {IEEE TSP},
  year = {2023},
  doi = {10.1109/tsp.2023.3247142},
}
Figure