TL;DR
This paper introduces a novel two-stage sampling strategy to more efficiently estimate the prevalence of a pandemic, like COVID-19. By leveraging information about both population distribution and diagnosed cases, their method offers a more flexible and accurate approach than existing techniques, particularly in complex scenarios.
What Problem Does This Paper Address?
Estimating the true prevalence of a pandemic is crucial for effective public health responses. However, this is challenging because many infected individuals are asymptomatic, meaning the number of diagnosed cases doesn't reflect the actual infection rate. Traditional sampling methods often fail to fully utilize available information, such as the distribution of diagnosed cases, which could provide valuable prior insights. Furthermore, nationwide surveys are costly, limiting the sample size and potentially introducing bias. This paper aims to develop a more efficient and robust sampling strategy that addresses these limitations.
Methods
The authors propose a two-stage sampling strategy.
- Stage 1: Sampling Positions: This stage focuses on determining where to sample. Instead of relying on traditional methods like Markov Chain Monte Carlo (MCMC) or Sampling/Importance Resampling (SIR), which can struggle with complex or multimodal probability distributions, they employ the Global Likelihood Sampling (GLS) algorithm. GLS is designed to handle such complexities and is robust to the multimodality of sampling densities.
- Stage 2: Sample Allocation and Testing: Once sampling positions are determined, sample sizes are allocated to each position. The method then uses simple random sampling at these positions to collect data and estimate infection prevalence.
A key innovation is the utilization of both population density ($f_P$) and diagnosed case density ($f_D$) to inform the sampling process, aiming to improve efficiency. The strategy also incorporates an initial rough estimate of the infection density ($f_I$) using a convex combination of $f_P$ and $f_D$, denoted by $\gamma f_P + (1-\gamma) f_D$.
Datasets
The paper uses simulated data for its numerical experiments and a real-world dataset for a practical example.
- Simulated Data: Used in Sections 5.1 and 5.2 to evaluate the performance of the proposed method against other sampling techniques under various scenarios.
- Real-world Data: The practical example in Section 6 uses cumulative COVID-19 diagnosed cases from the 51 administrative districts of the USA, obtained from the Centers for Disease Control and Prevention (CDC) COVID Data Tracker.
Key Insights and Ideas
- Leveraging Prior Information is Key: The paper strongly emphasizes that incorporating information about diagnosed cases and population distribution significantly improves the efficiency of prevalence estimation, a step often overlooked or underutilized in existing methods.
- GLS for Complex Distributions: The choice of GLS for determining sampling positions is a critical methodological advancement, allowing the strategy to adapt to real-world scenarios where infection densities might be complex and multimodal, unlike traditional MCMC or SIR methods.
- Robustness through Initial Estimation: Using an initial rough estimate of infection density ($f_I$) based on population and diagnosed case densities provides a practical way to approximate optimal sampling parameters, making the method applicable even when the true infection density is unknown.
- Flexibility Over Stratification: The proposed method is more flexible than stratified sampling, as it doesn't require pre-defined strata and can adapt to complex, continuous distributions of population and cases.
Key Results
- Improved Efficiency: Numerical simulations (Section 5) demonstrate that the proposed GLS-based two-stage sampling method achieves a smaller standardized standard deviation (SSD) compared to SIR, MH, and stratified sampling, indicating lower variance in the estimator of total infections ($N_I$).
- Higher Confidence Interval Coverage: The method also shows a higher coverage rate for approximate 95% confidence intervals, suggesting more reliable estimation.
- Practical Example Success: In a practical example using US COVID-19 data (Section 6), the proposed method achieved a slightly smaller standard deviation and a higher confidence interval coverage rate (99.5%) compared to stratified sampling (95.0%).
- Robustness of $\gamma$: Simulations suggest that a robust setting for the parameter $\gamma$ (used in the initial estimation of $f_I$) is often near the midpoint of its possible range, minimizing the maximum standardized deviation across various true infection densities.
Limitations and Future Work
- Practical Implementation of Optimal Settings: The paper acknowledges that achieving the exact optimal settings for the sampling density ($\phi$) and allocation function ($v$) is unrealistic in practice because they depend on the unknown true infection density ($f_I$). The proposed method focuses on "nearly optimal" settings.
- Second Stage Simplification: For simplicity, the second stage of the proposed method uses simple random sampling. The authors suggest that other sampling methods, like stratified sampling, could be incorporated into the second stage with modifications to further improve efficiency and reduce selection bias.
- Cost Considerations: While the method is efficient, the paper briefly touches upon incorporating cost functions to optimize parameters beyond just minimizing variance, suggesting this as a direction for future work.
- Theoretical Optimization of $\gamma$: The paper notes that theoretically optimizing the setting of $\gamma$ is difficult, and numerical studies are used to provide recommendations.
Summary updated