A Quantile Regression Model for Failure-Time Data with Time-Dependent\n Covariates

Malka Gorfine, Yair Goldberg, Ya’acov Ritov

Introduction

In clinical research and reliability engineering, survival analysis is the standard framework for modeling the time until an event of interest occurs. However, the factors influencing these "failure times" are rarely static. In medical contexts, a patient’s treatment status, physiological markers, or age at the time of an intervention often change throughout the observation period. These are known as time-dependent covariates.

Traditional survival models, such as the Cox Proportional Hazards model or the Accelerated Failure Time (AFT) model, often struggle to capture the nuanced ways these covariates impact survival across different segments of a population. This paper introduces a novel quantile regression framework designed to accommodate time-dependent covariates and right-censored data, providing a more granular view of survival dynamics.

The Limitations of Conventional Survival Models

The most widely used tool in survival analysis is the Cox Proportional Hazards model. Its primary limitation is the "proportional hazards" assumption, which posits that the ratio of risks between two groups remains constant over time. Similarly, the AFT model assumes that covariates act as a constant multiplier on the survival clock.

In reality, a treatment might be highly effective in preventing early mortality but have a diminishing or even negative effect on long-term survivors. Furthermore, survival curves often cross, a phenomenon that violates the assumptions of many standard models. Quantile regression addresses these issues by allowing researchers to model the impact of covariates at specific quantiles of the survival distribution (e.g., the 25th, 50th, or 75th percentile), rather than focusing solely on the average effect or a constant hazard ratio.

Methodology: Instrumental Variables and Doubly-Robust Estimation

The authors extend the AFT framework originally proposed by Robins and Tsiatis (1992) into a quantile-based approach. The methodology relies on two key pillars:

  1. Instrumental Variables (IV): To solve the estimation equations in the presence of time-dependent covariates, the model assumes the existence of an instrumental variable—a variable related to the covariates but independent of the failure time itself.
  2. Doubly-Robust Estimator: Following the logic of Robins and Rotnitzky (1992), the authors propose an "augmentation-based" estimator. This estimator is "doubly robust" because it remains consistent if either the model for the censoring process is correctly specified or the model for the underlying data distribution is correct. This provides a critical safety net for researchers working with complex, real-world datasets where modeling the censoring mechanism perfectly is often impossible.

Because the resulting estimating equations are discontinuous, the authors utilize a weighted bootstrap method for inference and a smooth approximation (sigmoid-like function) to facilitate computation.

Empirical Validation

Simulation Studies

The researchers conducted extensive simulations to evaluate the finite-sample properties of their estimators. Using synthetic datasets with stepwise time processes (simulating changes like drug dosage adjustments), the study demonstrated that the proposed estimator maintains low bias and high coverage rates (approximately 95%) across various sample sizes ($n=200$ to $1000$) and censoring levels as high as 40%.

The Stanford Heart Transplant Dataset

The utility of the model was further demonstrated using the classic Stanford heart transplant dataset, which tracks 99 patients. The analysis included time-dependent covariates such as transplant status, age at transplant, and tissue mismatch scores.

While traditional models suggested a general benefit for younger patients, the quantile regression model revealed a more complex story. For patients at the median survival time, the model showed that receiving a transplant at age 53 actually decreased post-transplant lifetime by 50% compared to the expected survival without the operation. This highlights how the effect of age and transplantation can vary significantly across different quantiles of the survival distribution—a nuance that simpler models fail to capture.

Conclusion

This paper provides a robust statistical framework for analyzing survival data when covariates change over time. By moving beyond the "average" effect, the proposed quantile regression model allows for a deeper understanding of how treatments and risks evolve. While the method requires careful consideration of instrumental variables and computational resources for the doubly-robust estimator, it offers a powerful alternative for biostatisticians and engineers dealing with the complexities of real-world failure-time data.

Summary updated

Abstract

Since survival data occur over time, often important covariates that we wish\nto consider also change over time. Such covariates are referred as\ntime-dependent covariates. Quantile regression offers flexible modeling of\nsurvival data by allowing the covariates to vary with quantiles. This paper\nprovides a novel quantile regression model accommodating time-dependent\ncovariates, for analyzing survival data subject to right censoring. Our simple\nestimation technique assumes the existence of instrumental variables. In\naddition, we present a doubly-robust estimator in the sense of Robins and\nRotnitzky (1992). The asymptotic properties of the estimators are rigorously\nstudied. Finite-sample properties are demonstrated by a simulation study. The\nutility of the proposed methodology is demonstrated using the Stanford heart\ntransplant dataset.\n
BibTeX
@inproceedings{gorfine2016quantile,
  author = {Malka Gorfine and Yair Goldberg and Ya’acov Ritov},
  title = {{A Quantile Regression Model for Failure-Time Data with Time-Dependent\textbackslash\{\}n Covariates}},
  booktitle = {Biostatistics},
  year = {2016},
  doi = {10.1093/biostatistics/kxw036},
}
Figure