Bootstrap Based Confidence Interval Estimation of \\[6pt]Quantiles for Current Status Data

Wei Chen, Fengling RenInternational Journal of Statistics and Probability 2021 · DOI

This paper, published in the International Journal of Statistics and Probability, introduces a more practical way to handle a common but tricky type of data found in medical and social science research.

TL;DR

The authors propose a new bootstrap-based method to estimate confidence intervals for quantiles in "current status" data. By using a model-based resampling approach, they bypass the need to estimate complex "nuisance parameters," resulting in a computationally efficient and stable tool for researchers.

What Problem Does This Paper Address?

In many studies, we don't know exactly when an event happened; we only know its "current status" at the time of an exam. For example, in a study on Rubella, researchers might only know if a person was already immune at the time of their blood test, not the exact date they became immune. This is known as Case 1 interval-censored data.

Statisticians use the Nonparametric Maximum Likelihood Estimator (NPMLE) to analyze this, but it has two major headaches:

  1. Slow Convergence: It converges at a "cube-root" rate ($n^{1/3}$), which is much slower than the standard square-root rate ($n^{1/2}$) most researchers are used to.
  2. Complex Distribution: The math describing its behavior involves complex processes (like Brownian motion). To build a confidence interval (CI) using traditional methods, you usually have to estimate "nuisance parameters"—specifically, the density of the event time and the observation time at a specific point. These estimates are notoriously unstable and computationally expensive.

Methods

The authors develop a model-based bootstrap approach that simplifies the process:

  1. Initial Estimate: First, calculate the NPMLE ($F_n$) from the original data.
  2. Resampling: Instead of a "naive" bootstrap (which fails in this mathematical setting), they generate new "bootstrap" status indicators based on the probabilities from their initial $F_n$ estimate.
  3. Variance Estimation: They use these bootstrap samples to estimate the variance of the quantile estimator.
  4. The "Aha" Moment: Rather than using the bootstrap distribution directly to set the CI, they use the bootstrap variance to solve for a scaling factor ($\gamma_n$). This factor is then plugged into a theoretical formula to construct the final confidence interval.

This method is "plug-and-play" because it doesn't require the researcher to manually tune smoothing parameters or estimate underlying density functions.

Datasets

The researchers evaluated their method using:

  • Extensive Simulations: They tested three different scenarios: an exponential distribution, a truncated exponential distribution, and an absolute normal distribution. They varied sample sizes from $n=50$ to $n=800$.
  • Real-World Data: A classic dataset of 230 Austrian males tested for Rubella immunization (Keiding et al., 1996).

Key Insights and Ideas

  • Avoiding the "Nuisance": The primary innovation is bypassing the direct estimation of probability densities. In statistics, "nuisance parameters" are things you have to calculate just to get to the result you actually care about; by avoiding them, the authors make the model much more robust.
  • Direct NPMLE: Unlike other recent methods that require "smoothing" the data first (which introduces its own set of problems, like choosing a bandwidth), this method works directly with the standard NPMLE.
  • Addressing Inconsistency: The paper acknowledges that the "naive" bootstrap is inconsistent for this type of data. Their specific model-based approach is designed to overcome this mathematical hurdle.

Key Results

  • Simulation Performance: The method showed strong "coverage probability"—meaning the 95% confidence intervals actually contained the true value about 95% of the time, especially when sample sizes were over 100.
  • Stability: The Standard Error Estimation (SEE) was consistently close to the empirical Sample Standard Deviation (SD), proving the method's reliability.
  • Rubella Analysis: When applied to the real-world Rubella data, the bootstrap method produced intervals that overlapped with existing (but more complex) methods, confirming its validity while being easier to implement.

Limitations and Future Work

  • Tail Behavior: The method struggles at the extreme ends of the data (the 10th or 90th percentiles). This is a common issue in survival analysis because there is simply less data available at the very beginning or very end of a timeline.
  • Small Samples: For very small datasets ($n < 200$), the slow cube-root convergence rate remains a challenge. The authors suggest that incorporating "smoothing" might help here, though it adds complexity.
  • Future Directions: The authors aim to extend this bootstrap approach to "Case 2" interval-censored data (where an event is known to happen between two specific observation times) and to develop "confidence bands" that cover the entire distribution curve at once.

Summary updated

Abstract

In this paper, we proposed a bootstrap approach to construct the confidence interval of quantiles for current status data, which is computationally simple and efficient without estimating nuisance parameters. The reasonability of the proposed method is verified by the well performance presented in the extensive simulation study. We also analyzed a real data set as illustration.
BibTeX
@article{chen2021bootstrap,
  author = {Wei Chen and Fengling Ren},
  title = {{Bootstrap Based Confidence Interval Estimation of \textbackslash\{\}\textbackslash\{\}[6pt]Quantiles for Current Status Data}},
  journal = {International Journal of Statistics and Probability},
  year = {2021},
  doi = {10.5539/ijsp.v10n5p38},
}
Figure