Hypothesis Test on a Mixture Forward-Incubation-Time Epidemic Model With Application to COVID-19 Outbreak

Chunlin Wang, Pengfei Li, Yukun Liu, Xiao‐Hua Zhou, Jing Qin

TL;DR

This paper develops a rigorous statistical framework to test the complexity of epidemic incubation models. The authors provide the mathematical "proof of work" for a mixture model that combines incubation and forward-time distributions, offering a new Likelihood Ratio Test (LRT) to determine if a simple exponential distribution is sufficient or if a more complex model is required to accurately reflect disease outbreaks like COVID-19.

What Problem Does This Paper Address?

Determining the incubation period—the time from exposure to the first appearance of symptoms—is the "holy grail" for public health officials setting quarantine durations. However, getting clean data is notoriously difficult. Most patients can’t pinpoint the exact moment they were infected (recall bias), and data collected during an active outbreak is often "length-biased."

To solve this, researchers previously proposed a mixture forward–incubation-time epidemic model. This model accounts for two types of people: those infected exactly when they left a hotspot (like Wuhan) and those who were already infected and were partway through their incubation period.

The problem? This mixture model is mathematically "non-regular." Under the simplest scenario (the null hypothesis), some parameters effectively disappear, causing standard statistical tests to break down. Furthermore, it wasn't clear if the model’s parameters were "identifiable"—meaning, can we actually find a unique solution for the variables given the data?

Methods

The authors tackle these theoretical hurdles using several advanced statistical techniques:

  1. Identifiability Proofs: They established sufficient conditions to prove that parameters for Weibull, Gamma, and Lognormal distributions can be uniquely estimated within this mixture framework.
  2. Likelihood Ratio Test (LRT): They proposed an LRT to test the null hypothesis ($H_0$) that the incubation period follows a simple, homogeneous exponential distribution.
  3. Handling Non-regularity: Because a nuisance parameter (the mixture proportion $p$) only exists under the alternative hypothesis, the test statistic doesn't follow a standard Chi-square distribution. The authors derived the limiting distribution of the LRT as the supremum of a Gaussian process.
  4. Explicit Representation: To make the test practical for other researchers, they used polar transformations to provide an explicit representation of this complex distribution, allowing for rapid calculation of p-values via Monte Carlo methods.

Datasets

The study evaluates its model using:

  • Synthetic Data: Extensive simulations with sample sizes ranging from $n=100$ to $n=1000$ to check Type I error rates and statistical power.
  • Wuhan Outbreak Data: A real-world dataset consisting of 1,211 confirmed COVID-19 cases. These individuals left Wuhan between January 19 and January 23, 2020 (just before the lockdown), were asymptomatic upon departure, and developed symptoms later.

Key Insights and Ideas

  • The "Identifiability" Trap: The authors discovered that if the incubation period is assumed to be a Weibull distribution with a shape parameter of 1, the model becomes an exponential distribution, and the mixture proportion $p$ becomes unidentifiable. This highlights why checking model assumptions is critical before drawing clinical conclusions.
  • Beyond Simple Averages: The paper demonstrates that a simple exponential model (which assumes a constant "forgetting" rate of infection) is often too blunt an instrument for epidemic modeling.
  • Mathematical Elegance in Chaos: By representing the test statistic as a "chi-square process" over a specific geometric set (defined by angles $\Delta_1$ and $\Delta_2$), the authors turned a daunting theoretical problem into a computable one.

Key Results

  • Simulation Success: The proposed LRT showed "tight control" of Type I errors (meaning it rarely gives false positives) even with sample sizes as small as 100.
  • High Statistical Power: The test was highly effective at detecting when the true distribution was not exponential, with power increasing as the parameters moved further from the null hypothesis.
  • COVID-19 Application: When applied to the 1,211 cases from the Wuhan departure cohort, the LRT yielded a p-value of nearly zero. This provides overwhelming evidence that the incubation period of COVID-19 is not a simple exponential distribution, justifying the use of more complex mixture models for setting policy.

Limitations and Future Work

The primary limitation is that the model is parametric; it assumes the underlying distribution belongs to a specific family (like Weibull or Gamma). If the true biological distribution of a virus doesn't fit these shapes, the test results might be skewed.

The authors suggest that future research could explore non-parametric versions of this model, where no specific distribution shape is assumed. However, they note that proving "identifiability" in a non-parametric setting is a significantly harder mathematical mountain to climb. Additionally, the current model assumes no one was infected after leaving the hotspot, which may not always hold true in more porous geographical settings.

Summary updated

Abstract

The distribution of the incubation period of the novel coronavirus disease that emerged in 2019 (COVID-19) has crucial clinical implications for understanding this disease and devising effective disease-control measures. Qin et al. (2020) designed a cross-sectional and forward follow-up study to collect the duration times between a specific observation time and the onset of COVID-19 symptoms for a number of individuals. They further proposed a mixture forward-incubation-time epidemic model, which is a mixture of an incubationperiod distribution and a forward time distribution, to model the collected duration times and to estimate the incubation-period distribution of COVID-19. In this paper, we provide sufficient conditions for the identifiability of the unknown parameters in the mixture forward-incubation-time epidemic model when the incubation period follows a two-parameter distribution. Under the same setup, Statistica Sinica: Newly accepted Paper (accepted author-version subject to English editing) we propose a likelihood ratio test (LRT) for testing the null hypothesis that the mixture forward-incubation-time epidemic model is a homogeneous exponential distribution. The testing problem is non-regular because a nuisance parameter is present only under the alternative. We establish the limiting distribution of the LRT and identify an explicit representation for it. The limiting distribution of the LRT under a sequence of local alternatives is also obtained. Our simulation results indicate that the LRT has desirable type-I errors and powers, and we analyze a COVID-19 outbreak dataset from China to illustrate the usefulness of the LRT.
BibTeX
@inproceedings{wang2023hypothesis,
  author = {Chunlin Wang and Pengfei Li and Yukun Liu and Xiao‐Hua Zhou and Jing Qin},
  title = {{Hypothesis Test on a Mixture Forward-Incubation-Time Epidemic Model With Application to COVID-19 Outbreak}},
  booktitle = {Statistica},
  year = {2023},
  doi = {10.5705/ss.202020.0318},
}
Figure