Estimation/Imputation Strategies for Missing Data in Survival Analysis

Elodie Brunel, Fabienne Comte, Agathe Guilloux2013 · DOI

TL;DR

This paper tackles the challenge of estimating survival functions when the indicator of whether an event was observed (censored or not) is itself missing. The authors propose and analyze novel estimation and imputation strategies, demonstrating their effectiveness through theoretical results and numerical simulations.

What Problem Does This Paper Address?

In survival analysis, we often deal with data where the exact time of an event (like death or failure) is not always observed. Instead, we might only know that the event occurred before a certain time (right-censoring). This is typically indicated by a censoring indicator variable. However, this paper addresses a more complex scenario where this censoring indicator itself might be missing. This "missing censoring indicator" (MCI) problem complicates standard survival analysis techniques. The authors focus on the "Missing At Random" (MAR) context, where the missingness of the indicator depends on observed data but not on the missing value itself. They aim to develop robust methods for estimating the hazard rate and survival function under these conditions.

Methods

The paper proposes two main strategies for dealing with missing censoring indicators:

  • Estimation Strategies: These involve estimating the unknown nuisance functions (specifically, the probability of the censoring indicator being observed, $\pi(y)$, and the probability of the event being observed, $\zeta(y)$) using either:
  • Nonparametric Regression: Using localized bases (like trigonometric or histogram bases) to estimate these functions without strong parametric assumptions.
  • Parametric Logistic Regression: Assuming a logistic model for $\pi(y)$ and $\zeta(y)$ and using maximum likelihood estimation. For $\zeta(y)$, where the event indicator is also missing, they employ a procedure inspired by Rubin (1987) for simulating missing values. These estimated nuisance functions are then used to construct penalized contrast estimators for the hazard rate.

  • Imputation Strategy: This approach is inspired by stochastic regression imputation. Instead of directly estimating the nuisance functions, it involves imputing missing values for the censoring indicator. This is done by drawing random numbers from a Bernoulli distribution with a probability estimated using either nonparametric or logistic regression. The imputed data is then used in a contrast function to estimate the hazard rate.

The authors also introduce model selection criteria to choose the optimal complexity of the nonparametric models. Theoretical properties, including asymptotic results and error bounds, are derived for the proposed estimators.

Datasets

The paper does not introduce or use any specific real-world datasets. Instead, it relies on simulated data generated from three distinct models to evaluate the performance of the proposed methods. These models are designed to represent different scenarios of survival data with missing censoring indicators.

Key Insights and Ideas

  • The challenge of missing censoring indicators: The paper highlights that standard survival analysis methods are insufficient when the censoring indicator itself is missing, requiring specialized techniques.
  • MAR context is crucial: Focusing on the "Missing At Random" assumption allows for the development of statistically sound methods that leverage observed data to infer missing information.
  • Two-step approach: The core idea is to first estimate or impute the missing indicator information (via $\pi$ and $\zeta$) and then use this to estimate the hazard rate.
  • Nonparametric vs. Parametric: The paper systematically compares nonparametric and parametric (logistic) approaches for estimating the nuisance functions, providing insights into their relative strengths and weaknesses.
  • Imputation as a viable alternative: The imputation strategy, while conceptually different, shows comparable performance to estimation strategies, offering flexibility.
  • Impact of nuisance function estimation: The authors emphasize that the accuracy of estimating the nuisance functions ($\pi$ and $\zeta$) directly impacts the final hazard rate estimator's performance, especially when $\pi$ is in the denominator.

Key Results

  • Theoretical Guarantees: The paper provides theoretical results establishing error bounds for the proposed hazard rate estimators under various assumptions. These bounds show that the estimators can achieve good performance, with rates depending on the chosen strategy and model complexity.
  • Numerical Performance: Simulation studies across three models (Weibull-Logistic, Monotone, Non-monotone) demonstrate that:
  • The imputation strategy and the estimation strategy using $\zeta$ (FEST$_{n,1}$) generally perform better than the strategy involving $\pi$ (FEST$_{n,2}$).
  • The logit (parametric) approach for estimating nuisance functions often outperforms the nonparametric approach, especially for smaller sample sizes, provided the logit assumption is reasonably met.
  • Performance deteriorates with higher missing rates (45%) and larger sample sizes, particularly when the logit assumption is violated.
  • The strategy involving $\pi$ (FEST$_{n,2}$) consistently shows poorer performance, likely due to the error propagation from estimating $\pi$ when it's in the denominator.

Limitations and Future Work

  • Assumptions: The theoretical results rely on specific assumptions about the data and the estimators, such as the boundedness of the hazard rate and the existence of certain constants. The practical implications of violating these assumptions are not fully explored.
  • Model Selection Sensitivity: While model selection criteria are proposed, their robustness and performance across a wider range of scenarios could be further investigated.
  • Computational Cost: Nonparametric methods can be computationally intensive, especially for large datasets and complex bases. The paper does not extensively discuss computational efficiency.
  • Future Work: The authors suggest further empirical studies to investigate the performance of the logit approximation when the logit model is not perfectly satisfied and to understand how the quality of $\pi$ and $\zeta$ estimation influences the final hazard rate estimator. They also hint at adaptive nonparametric rates for $\zeta$ estimation.

Summary updated

Abstract

International audience
BibTeX
@misc{brunel2013estimation,
  author = {Elodie Brunel and Fabienne Comte and Agathe Guilloux},
  title = {{Estimation/Imputation Strategies for Missing Data in Survival Analysis}},
  year = {2013},
  doi = {10.1002/9781118826805.ch15},
}
Figure