Quantile Inference with Multivariate Failure Time Data

Guosheng Yin, Jianwen Cai, Jinheum KimBiometrical 2003 · DOI

TL;DR

This paper introduces a new statistical method for estimating quantiles (like medians) from complex survival data where individuals can experience multiple types of events, and these events are correlated. This is crucial for accurately comparing survival experiences in real-world scenarios like clinical trials or genetic studies, where simple methods fail due to this complexity.

What Problem Does This Paper Address?

In survival analysis, understanding how long it takes for an event to occur is key. Quantiles, especially medians, are preferred over means for summarizing this "time to event" data because they are less sensitive to extreme values. However, much of the existing statistical machinery for survival data assumes that each individual experiences at most one event, and that these events are independent across individuals. This assumption breaks down in many real-world biomedical studies. For instance, a patient might experience multiple distinct health events (like a heart attack and a stroke), or multiple occurrences of the same event (like recurrent tumors). Furthermore, within a single individual or a cluster (like matched mice or paired eyes in a study), these different failure times can be correlated. Existing methods cannot handle this "multivariate failure time data" with correlated events, leading to potentially inaccurate conclusions. This paper aims to fill that gap by providing a robust method for quantile estimation in these complex scenarios.

Methods

The authors propose nonparametric procedures for estimating quantiles in the presence of multivariate and correlated failure times. The core of their approach involves:

  • Nonparametric Estimation: They avoid making strong assumptions about the underlying distribution of failure times.
  • Multivariate Normal Distribution: They demonstrate that their proposed quantile estimators asymptotically follow a multivariate normal distribution. This is a powerful result as it allows for statistical inference (like confidence intervals and hypothesis testing).
  • Variance-Covariance Estimation: A critical part of their method is estimating the variance-covariance matrix of these multivariate quantiles. They achieve this using two techniques:
  • Kernel Smoothing: This is used to estimate the underlying hazard functions, which are then used to derive the variance-covariance.
  • Bootstrap: This resampling technique is employed to estimate the variability and dependence structure of the quantile estimators, particularly useful for handling correlated data.

The theoretical foundation relies on weak convergence properties of related processes, specifically the Nelson-Aalen estimator for cumulative hazard functions.

Datasets

The paper uses two real-world datasets to illustrate their proposed methods:

  • Burn-Wound Infections Study: This study involves patients randomized into two groups for body cleansing, with events of interest being time to excision and time to antibiotic treatment.
  • Diabetic Retinopathy Study (DRS): This study examines paired data from patients' eyes, where one eye is treated and the other is untreated, with the event being severe visual loss.

The paper also includes extensive simulation studies using generated data to evaluate the finite sample performance of their methods under various scenarios (different distributions, censoring rates, and correlation levels).

Key Insights and Ideas

  • Correlation Matters: Ignoring the correlation between multiple failure times within an individual can lead to inflated Type I error rates, meaning you're more likely to falsely conclude there's a difference when there isn't.
  • Quantiles as Robust Summaries: The paper reinforces the value of quantiles (especially medians) as robust summary statistics for survival data, particularly when dealing with skewed distributions or potential outliers.
  • Bridging the Gap: The proposed method provides a much-needed tool for analyzing complex, correlated survival data, which is prevalent in many biomedical research settings.
  • Bootstrap for Dependence: The use of a cluster bootstrap (treating individuals as units) is highlighted as an effective way to preserve the correlation structure when estimating variability in dependent data.

Key Results

  • Asymptotic Normality: The proposed quantile estimators are shown to be asymptotically multivariate normally distributed, enabling statistical inference.
  • Good Performance in Simulations: Simulation studies indicate that both the kernel smoothing and bootstrap methods perform well in finite samples, providing accurate estimates of quantiles and their variability.
  • Real-World Applicability: The methods are successfully applied to the burn-wound infection and Diabetic Retinopathy Study datasets, yielding meaningful insights into treatment effects and survival experiences. For example, they found significant differences in lower quartiles of time to excision and antibiotic treatment in the burn study, and a significantly longer lower survival quartile for treated eyes in the DRS.
  • Variance Estimation Accuracy: Comparisons between standard deviation ("sd") and estimated standard error ("se"), and covariance ("cov") and estimated covariance ("ecov") in the simulation tables suggest that both proposed methods provide good estimates of the variability of the quantile estimators.

Limitations and Future Work

  • Bandwidth Sensitivity: The kernel smoothing method's performance is noted to be sensitive to the choice of bandwidth. The paper discusses trade-offs between bias and variance and mentions cross-validation and trial-and-error as methods for selection, but optimal bandwidth selection remains a practical challenge.
  • Non-existence of Quantiles: The paper acknowledges that quantiles might not exist in cases of heavy censoring, which could prevent estimation.
  • Generalization: The authors suggest that their proposal can be generalized to quantile testing with covariates in Cox proportional hazards models for multiple events, and to smoothing survival curves first before estimating quantiles.
  • Future Extensions: The work lays the groundwork for more complex models, including those with covariates and multiple event types.

Summary updated

Abstract

Abstract Quantiles, especially the medians, of survival times are often used as summary statistics to compare the survival experiences between different groups. Quantiles are robust against outliers and preferred over the mean. Multivariate failure time data often arise in biomedical research. For example, in clinical trials, each patient in the study may experience multiple events which may be of the same type or distinct types, while in family studies of genetic diseases or litter matched mice studies, failure times for subjects in the same cluster may be correlated. In this article, we propose nonparametric procedures for the estimation of quantiles with multivariate failure time data. We show that the proposed estimators asymptotically follow a multivariate normal distribution. The asymptotic variance‐covariance matrix of the estimated quantiles is estimated based on the kernel smoothing and bootstrap techniques. Simulation results show that the proposed estimators perform well in finite samples. The methods are illustrated with the burn‐wound infection data and the Diabetic Retinopathy Study (DRS) data.
BibTeX
@inproceedings{yin2003quantile,
  author = {Guosheng Yin and Jianwen Cai and Jinheum Kim},
  title = {{Quantile Inference with Multivariate Failure Time Data}},
  booktitle = {Biometrical},
  year = {2003},
  doi = {10.1002/bimj.200390036},
}
Figure