A resampling-based test to detect person-to-person transmission of infectious disease

Yang Yang, Ira M. Longini, M. Elizabeth HalloranThe Annals of Applied Statistics 2007 · DOI

TL;DR

This paper introduces a novel resampling-based statistical test to detect person-to-person transmission of infectious diseases, particularly useful for emerging threats like avian influenza where early detection is critical. The proposed "refined permutation test" demonstrates comparable or superior performance to traditional asymptotic tests, especially in scenarios with limited data, offering a robust tool for public health surveillance.

What Problem Does This Paper Address?

The early detection of person-to-person transmission is paramount for controlling emerging infectious diseases and preventing pandemics. Traditional statistical methods for this purpose often rely on asymptotic theory, which can be unreliable or unavailable when dealing with small sample sizes, a common scenario in the initial stages of an outbreak (e.g., avian influenza). Furthermore, testing for the absence of transmission (i.e., a probability of zero) presents a "boundary problem" that complicates standard statistical approaches. This paper aims to fill this gap by developing and evaluating a resampling-based statistical test that is robust to small sample sizes and can handle these non-standard conditions, providing a more reliable tool for real-time surveillance.

Methods

The core of this paper is the development of two resampling-based hypothesis testing methods: a simple permutation test and a more sophisticated "refined permutation test." These methods are designed to test the null hypothesis that there is no person-to-person transmission ($H_0: p_1 = p_2 = 0$) against the alternative that there is ($H_1: p_1 > 0$ or $p_2 > 0$), where $p_1$ and $p_2$ represent transmission probabilities within and between households, respectively.

The simple permutation test works by reassigning observed symptom onset days and infection statuses among individuals. Under the null hypothesis, all such rearrangements are equally likely. The empirical distribution of a test statistic (specifically, a likelihood ratio statistic) calculated from these permutations approximates the null distribution.

The refined permutation test builds upon this by introducing a "refinement stage." This stage involves a more complex resampling procedure that considers a larger sampling space, allowing for more nuanced rearrangements of infection times while maintaining the likelihood under the null hypothesis. This refinement is crucial for improving the test's power, especially in scenarios with latent periods.

The paper also contrasts these resampling methods with a conventional asymptotic test (specifically, a likelihood ratio test) which, while theoretically applicable under certain conditions, struggles with the boundary problem and requires larger sample sizes.

Datasets

The paper does not use real-world datasets. Instead, it relies on extensive simulations to evaluate the performance of the proposed resampling tests. These simulations generate synthetic epidemic data based on a defined statistical model of disease transmission, allowing for controlled comparisons of type I error rates and statistical power under various parameter settings (e.g., different transmission probabilities, population sizes, and latent period durations).

Key Insights and Ideas

  • Resampling as a Robust Alternative: The paper strongly advocates for resampling methods (permutation tests) as a powerful and flexible alternative to asymptotic tests, especially when dealing with complex statistical problems like boundary hypotheses or small sample sizes.
  • Handling Boundary Problems: The refined permutation test effectively addresses the challenge of testing hypotheses at the boundary of parameter spaces (e.g., testing if a probability is exactly zero), a common issue in epidemiological modeling.
  • Power in Small Samples: A key takeaway is that even with a very small number of observed cases (as few as seven or eight), the refined permutation test can achieve decent statistical power to detect person-to-person transmission, provided the disease is moderately transmissible and the population is sufficiently large. This is highly relevant for early outbreak detection.
  • Beyond Asymptotics: The resampling approach circumvents the need for strict regularity conditions often required by asymptotic tests, making it applicable to a broader range of real-world scenarios where these conditions might not hold.

Key Results

  • Type I Error Preservation: Both the simple and refined permutation tests effectively preserve the type I error rate at the specified level (e.g., 0.05), meaning they rarely falsely detect transmission when none exists. The asymptotic test also performs well in this regard, sometimes being overly conservative.
  • Superior Power: The refined permutation test consistently demonstrates higher statistical power compared to both the simple permutation test and the asymptotic test, particularly for small sample sizes and when transmission probabilities are moderate to high.
  • Effectiveness with Limited Data: Simulations showed that the refined permutation test could achieve 80% power with a moderate $p_1$ (SAR$_1 = 0.14$) even with a total of only seven cases, highlighting its utility in early outbreak scenarios.
  • Empirical Formula for Power Prediction: The authors developed an empirical formula that explains 99% of the variation in power based on the number of index cases and total cases, providing a tool to predict power levels without extensive simulations.

Limitations and Future Work

  • Assumptions on Latent/Incubation Periods: The study assumes that latent and incubation periods are identical and that their distributions are known. While the authors note that other assumptions could be made, this simplification might affect inference.
  • Common Source Exposure Duration (S): The model assumes a fixed duration (S) for exposure to the common source. Determining this point in real epidemics can be challenging, and mis-specifying S can impact the test's performance. The paper suggests that assuming $S > T$ (where T is the observation period) yields the most conservative p-value.
  • Data Requirements: The method requires not only symptom onset times but also data on household membership. While considered reasonable, collecting this data might be a practical hurdle.
  • Generalizability of Clustering: The current model focuses on household clustering. The authors suggest that the methods could be generalized to include other clustering groups like schools and workplaces.
  • Real-time Analysis and Censoring: The paper acknowledges that complete adjustments for right-censoring of infection status might be important for real-time analysis, especially when the observation period is not significantly longer than the latent period.
  • Correlation of Tests: The authors note that multiple resampling tests applied to the same data set are highly correlated, which could make adjustments for multiple testing (like Bonferroni) overly conservative.

Future work could explore more sophisticated methods for estimating latent and infectious period distributions, refine the handling of common source exposure, and extend the framework to incorporate other social structures beyond households.

Summary updated

Abstract

Early detection of person-to-person transmission of emerging infectious diseases such as avian influenza is crucial for containing pandemics. We developed a simple permutation test and its refined version for this purpose. A simulation study shows that the refined permutation test is as powerful as or outcompetes the conventional test built on asymptotic theory, especially when the sample size is small. In addition, our resampling methods can be applied to a broad range of problems where an asymptotic test is not available or fails. We also found that decent statistical power could be attained with just a small number of cases, if the disease is moderately transmissible between humans.
BibTeX
@article{yang2007resampling,
  author = {Yang Yang and Ira M. Longini and M. Elizabeth Halloran},
  title = {{A resampling-based test to detect person-to-person transmission of infectious disease}},
  journal = {The Annals of Applied Statistics},
  year = {2007},
  doi = {10.1214/07-aoas105},
}
Figure