Cure Rate Quantile Regression for Censored Data With a Survival Fraction

Yuanshan Wu, Guosheng YinJournal of the American Statistical Association 2013 · DOI

TL;DR

This paper introduces "cure rate quantile regression," a new statistical model for analyzing survival data where some individuals are effectively "cured" and will never experience the event of interest. This method provides a more nuanced understanding of how factors influence survival across the entire distribution of survival times, not just the average.

What Problem Does This Paper Address?

Traditional survival analysis models, like the Cox proportional hazards model, often focus on the average survival time or hazard rate. However, in many real-world scenarios, such as medical treatments or reliability studies, a portion of the population might be "cured" and thus immune to the event. This creates a "survival fraction" that can skew results if not properly accounted for. Standard models can struggle to capture the full picture when survival times are right-skewed due to this cured group. Existing cure rate models typically focus on mean-based regression, which doesn't fully leverage the information available in the entire survival distribution. This paper addresses this gap by proposing a method that can model the survival of susceptible individuals at various quantiles, offering a richer understanding of covariate effects.

Methods

The core of the paper is the development of cure rate quantile regression. This model combines two components:

  1. Censored Quantile Regression: This is used to model the survival times of the susceptible individuals. Unlike mean regression, it allows us to estimate how covariates affect specific quantiles (e.g., the 10th percentile, 50th percentile, 90th percentile) of survival times.
  2. Logistic Regression: This is used to model the probability of an individual being susceptible to the event (i.e., not cured).

The authors propose two estimation methods:

  • Iterative Approach: This method estimates both the cure rate parameters and the quantile regression parameters simultaneously through an iterative process. It aims to leverage all available information from the regression quantiles.
  • Nonparametric Kernel Smoothing Approach: This method separates the estimation of the cure rate parameters from the quantile regression parameters. It uses a nonparametric kernel smoothing technique, which can be more stable and less prone to convergence issues.

Both methods are built upon martingale-based estimating equations, a standard technique in survival analysis for handling censored data.

Datasets

The paper uses a real-world dataset for illustration:

  • Bone Marrow Transplantation (BMT) Study: This dataset, collected between 1985 and 1991, involves 1715 leukemia patients. The study aims to understand factors influencing time to cancer relapse or death, considering that some patients might be considered "cured" after a certain period due to the absence of graft-versus-host disease.

Key Insights and Ideas

  • Beyond the Average: The most significant insight is the move beyond average-based survival analysis. By modeling quantiles, we can see how factors affect survival differently for those who experience the event early versus those who survive much longer.
  • Cure Fraction Nuance: The model explicitly handles the "cure fraction," allowing for a more accurate assessment of covariate effects on the susceptible population without being confounded by the cured group.
  • Two Estimation Strategies: The development of two distinct estimation methods (iterative vs. nonparametric) provides flexibility. The nonparametric approach appears to offer better convergence properties, which is a common challenge in complex statistical modeling.
  • Identifiability is Key: The paper carefully addresses the crucial issue of identifiability in censored quantile regression, particularly for quantiles close to 1 (long-term survival), ensuring that the model parameters can be uniquely estimated.

Key Results

  • Simulation Studies: Extensive simulations demonstrated that both the iterative and nonparametric methods perform well in estimating parameters and providing accurate confidence intervals, even under various error distributions and with a significant cure fraction. The nonparametric method showed better convergence rates.
  • BMT Data Application: When applied to the BMT data, the cure rate quantile regression model revealed that factors like HLA matching and Karnofsky scores significantly impact survival, especially at higher quantiles (indicating longer survival times). This contrasts with traditional methods that provide a more global effect. For instance, donor type differences were more pronounced for patients surviving longer.

Limitations and Future Work

  • Dimension of Covariates: The paper notes that the nonparametric estimation of the survival function can deteriorate as the number of continuous covariates increases. For high-dimensional data, they suggest dimension reduction techniques or fitting covariates one by one.
  • Model Checking: Developing robust model checking procedures for this new model is acknowledged as an area for further research, as standard methods based on martingale residuals become more complex.
  • Extending Flexibility: The authors suggest future work could extend the model to partially linear models and incorporate mixtures of quantile-variant and -invariant effects to enhance flexibility.
  • Identifiability of High Quantiles: While addressed, the identifiability of regression quantiles very close to 1 remains a subtle issue, and the estimation is confined to quantile levels below a certain maximum (Tmax).

Summary updated

Abstract

Censored quantile regression offers a valuable complement to the traditional Cox proportional hazards model for survival analysis. Survival times tend to be right-skewed, particularly when there exists a substantial fraction of long-term survivors who are either cured or immune to the event of interest. For survival data with a cure possibility, we propose cure rate quantile regression under the common censoring scheme that survival times and censoring times are conditionally independent given the covariates. In a mixture formulation, we apply censored quantile regression to model the survival times of susceptible subjects and logistic regression to model the indicators of whether patients are susceptible. We develop two estimation methods using martingale-based equations: One approach fully uses all regression quantiles by iterating estimation between the cure rate and quantile regression parameters; and the other separates the two via a nonparametric kernel smoothing estimator. We establish the uniform consistency and weak convergence properties for the estimators obtained from both methods. The proposed model is evaluated through extensive simulation studies and illustrated with a bone marrow transplantation data example. Technical proofs of key theorems are given in Appendices A, B, and C, while those of lemmas and additional simulation studies on model misspecification and comparisons with other models are provided in the online Supplementary Materials A and B.
BibTeX
@article{wu2013cure,
  author = {Yuanshan Wu and Guosheng Yin},
  title = {{Cure Rate Quantile Regression for Censored Data With a Survival Fraction}},
  journal = {Journal of the American Statistical Association},
  year = {2013},
  doi = {10.1080/01621459.2013.837368},
}
Figure