Model-based and model-free characterization of epidemic outbreaks

Jonas Dehning, F. Paul Spitzner, Matthias Linden, Sebastian Mohr, João Pinheiro Neto, Johannes Zierenberg, Michael Wibral, Michael Wilczek, Viola PriesemannmedRxiv 2020 · DOI

Abstract

Here, we provide detailed background information for our work on Bayesian inference of change-points in the spread of SARS-CoV-2 and the effectiveness of non-pharmaceutical interventions (Dehning et al., Science, 2020). We outline the general background of Bayesian inference and of SIR-like models. We explain the assumptions that underlie model-based estimates of the reproduction number and compare them to the assumptions that underlie model-free estimates, such as used in the Robert-Koch Institute situation reports. We highlight effects that originate from the two estimation approaches, and how they may cause differences in the inferred reproduction number. Furthermore, we explore the challenges that originate from data availability — such as publication delays and inconsistent testing — and explain their impact on the time-course of inferred case numbers. Along with alternative data sources, this allowed us to cross-check and verify our previous results.

Introduction

During the early stages of the COVID-19 pandemic, the reproduction number ($R$) became a critical metric for public health policy. However, different institutions often reported conflicting values, leading to confusion. This paper investigates the underlying causes of these discrepancies, focusing on the mathematical artifacts of estimation methods and the complexities of real-world data integrity.

The fundamental problem is that infections are not directly observable. Instead, we rely on "reported cases," which are filtered through a series of biological and administrative delays, including incubation periods and testing lags.

Methods: Model-Free vs. Model-Based Estimation

The authors contrast two primary mathematical frameworks for estimating $R$:

  1. Model-Free Estimation: Common in public health reports (e.g., the Robert Koch Institute), these methods calculate $R$ as a simple ratio of new cases today compared to a generation ago (typically 4 days). While computationally straightforward, they assume a fixed generation time and ignore the underlying dynamics of disease spread.
  2. Model-Based Bayesian Inference: This approach utilizes SIR (Susceptible-Infectious-Recovered) or SEIR models. By treating $R$ as a time-dependent variable with "change points" (corresponding to interventions), the authors use Markov Chain Monte Carlo (MCMC) sampling to infer the most plausible causal parameters. This method explicitly models the convolution of infection dates with biological and reporting delays.
The chain of delays from infection to reporting, showing how incubation and reporting delays shift and smooth the observed case numbers.

Key Insights: The "Phantom" $R < 1$ Effect

A significant finding is the "phantom" $R < 1$ artifact. When the true reproduction number drops sharply (e.g., from 3.0 to 1.1), model-free estimators often show $R$ dipping below 1.0 temporarily. This occurs because the rapid shrinkage of the newly infected pool makes the ratio of cases appear as a decline, even if the virus is still in an exponential growth phase ($R > 1$).

Comparison of the true reproduction number (R) with various model-free and model-based estimators, highlighting the 'phantom' dip below 1 in model-free methods.

Furthermore, the paper highlights that reporting delays act as a "low-pass filter," smoothing out sharp transitions. Models that fail to account for this convolution will systematically underestimate the speed and effectiveness of public health interventions.

Data Integrity and Testing Capacity

The study also addresses the "testing artifact" argument—the claim that rising case numbers were merely a result of increased testing. By analyzing testing backlogs and positivity rates in Germany, the authors demonstrate that while testing volume did increase, the exponential growth of the virus was a physical reality that far outpaced testing expansion during the initial outbreak.

Results and Robustness

Using updated "symptom onset" data (the "epi curve"), which is more accurate but subject to significant reporting latency, the authors confirmed their previous findings. Three distinct change points in $R$ were necessary to explain the German data, aligning with the cancellation of large events, school/store closures, and the national lockdown.

Leave-One-Out (LOO) cross-validation confirmed that the three-change-point SIR model significantly outperformed simpler models, proving the shifts in infection rates were statistically significant. The total delay from infection to official reporting was quantified at approximately 10–11 days.

Limitations

While Bayesian inference identifies the most plausible causal explanation, it does not establish "strict interventional causality," as unmodeled factors like voluntary behavioral changes may coincide with official mandates. Additionally, the "epi curve" data suffers from a "tail" problem, requiring up to three weeks to stabilize, and the current models assume a spatially homogeneous population.

Summary updated

Abstract

Here, we provide detailed background information for our work on Bayesian inference of change-points in the spread of SARS-CoV-2 and the effectiveness of non-pharmaceutical interventions (Dehning et al., Science, 2020). We outline the general background of Bayesian inference and of SIR-like models. We explain the assumptions that underlie model-based estimates of the reproduction number and compare them to the assumptions that underlie model-free estimates, such as used in the Robert-Koch Institute situation reports. We highlight effects that originate from the two estimation approaches, and how they may cause differences in the inferred reproduction number. Furthermore, we explore the challenges that originate from data availability — such as publication delays and inconsistent testing — and explain their impact on the time-course of inferred case numbers. Along with alternative data sources, this allowed us to cross-check and verify our previous results.
BibTeX
@misc{dehning2020model,
  author = {Jonas Dehning and F. Paul Spitzner and Matthias Linden and Sebastian Mohr and João Pinheiro Neto and Johannes Zierenberg and Michael Wibral and Michael Wilczek and Viola Priesemann},
  title = {{Model-based and model-free characterization of epidemic outbreaks}},
  year = {2020},
  doi = {10.1101/2020.09.16.20187484},
  howpublished = {medRxiv},
}
Figure