Discrete-time Competing-Risks Regression with or without Penalization

Tomer Meir, Malka GorfineBiometrics 2023

TL;DR

This paper introduces a novel statistical method for analyzing discrete-time survival data with competing risks, addressing a common limitation in existing continuous-time focused models. The new approach simplifies estimation and seamlessly integrates with modern machine learning techniques like penalized regression and feature screening, leading to more accurate and efficient predictions, particularly in complex scenarios.

What Problem Does This Paper Address?

Survival analysis is crucial for understanding how long events take to occur, but many real-world datasets have "discrete" time-to-event data. This means time is measured in distinct units (like days or cycles) rather than continuously. Compounding this, individuals can face multiple types of events (competing risks) that prevent them from experiencing other events (e.g., a patient can be discharged, transferred, or pass away). Existing methods often struggle with discrete data, leading to biased results. This paper tackles this by developing a robust statistical framework specifically for discrete-time competing-risks data, making survival analysis more applicable to a wider range of real-world problems.

Methods

The core of the paper's contribution is a new two-step estimation procedure. It cleverly separates the estimation of two key components of the survival model: the baseline hazard parameters (how likely an event is at a given time, independent of other factors) and the regression coefficients (how covariates like age or treatment affect the event risk).

  1. Estimating Regression Coefficients ($\beta_j$): The authors leverage a conditional-logistic regression approach on an "expanded" dataset. This means each observation is broken down into multiple time intervals, allowing for standard, well-established Cox regression routines to efficiently estimate the covariate effects.
  2. Estimating Baseline Hazard Parameters ($\alpha_{jt}$): Once the regression coefficients are estimated, the paper proposes a simpler, single-dimensional optimization for the baseline hazard parameters. This is done by minimizing the squared difference between observed and expected event proportions at each time point.

This separation is key because it allows for the straightforward integration of penalized regression methods (like Lasso and Elastic Net) and feature screening techniques. These methods are essential for handling high-dimensional data (many potential predictors) and selecting the most relevant ones.

Datasets

The primary dataset used for evaluation is the MIMIC-IV (version 2.0) dataset. This publicly available dataset contains detailed information on intensive care unit (ICU) admissions. The authors analyze the length of stay (LOS) for 25,170 ICU admissions, considering three competing events: discharge to home, transfer to another medical facility, and in-hospital death. They also conducted extensive simulation studies across various settings (different sample sizes, numbers of covariates, and event types) to validate their proposed method and compare it against existing approaches.

Key Insights and Ideas

  • Discrete Time Matters: The paper strongly emphasizes that treating discrete time-to-event data as continuous can lead to significant biases, particularly in the estimation of baseline hazards. Their method directly addresses this by modeling time discretely.
  • Separation is Power: The novel two-step approach, by separating the estimation of baseline hazards and regression coefficients, unlocks the ability to easily incorporate powerful penalized regression and feature selection methods. This is a major advantage over previous methods that required simultaneous estimation of all parameters.
  • Efficiency Gains: The proposed method is not only more accurate but also computationally more efficient, especially as the number of discrete time points increases. This is crucial for analyzing large, complex datasets.
  • Bridging the Gap: The work effectively bridges the gap between traditional survival analysis and modern machine learning techniques, making advanced predictive modeling more accessible for discrete-time competing-risks data.

Key Results

  • Improved Accuracy: Simulation studies showed that the proposed method yields practically unbiased estimates for baseline hazards, outperforming naive continuous-time approaches that exhibited substantial biases.
  • Comparable Performance to Existing Methods: The proposed method demonstrated similar bias and standard error performance to the established Lee et al. (2018) method in simulations, but with significant advantages in computational efficiency and ease of integration with penalized methods.
  • Effective Feature Selection: When applied with Lasso regularization, the method successfully reduced the number of selected predictors without sacrificing predictive performance (measured by AUC), as demonstrated in the MIMIC dataset analysis.
  • Computational Advantage: The proposed method showed a clear advantage in computation time as the number of discrete time points increased, and it was more robust to memory limitations compared to the Lee et al. method.
  • Real-World Application: The analysis of ICU patient length of stay using the MIMIC dataset provided valuable insights into factors influencing discharge, transfer, and death, showcasing the practical utility of the method.

Limitations and Future Work

  • Link Function: The current software implementation primarily uses the logit link function. While the proposed approach is valid for log and log-log link functions, its applicability to other link functions might require further investigation.
  • Computational Cost for High Dimensions: While more efficient than some alternatives, the computational cost can still be a factor for extremely high-dimensional data, especially when combined with extensive cross-validation for tuning parameters.
  • Future Work: The authors suggest extending the method to handle time-dependent covariates and coefficients more broadly, and exploring its application to other types of discrete-time survival data. The availability of the Python package PyDTS facilitates further research and application of their method.

Summary updated

Abstract

Many studies employ the analysis of time-to-event data that incorporates competing risks and right censoring. Most methods and software packages are geared towards analyzing data that comes from a continuous failure time distribution. However, failure-time data may sometimes be discrete either because time is inherently discrete or due to imprecise measurement. This paper introduces a new estimation procedure for discrete-time survival analysis with competing events. The proposed approach offers a major key advantage over existing procedures and allows for straightforward integration and application of widely used regularized regression and screening-features methods. We illustrate the benefits of our proposed approach by a comprehensive simulation study. Additionally, we showcase the utility of the proposed procedure by estimating a survival model for the length of stay of patients hospitalized in the intensive care unit, considering three competing events: discharge to home, transfer to another medical facility, and in-hospital death. A Python package, PyDTS, is available for applying the proposed method with additional features.
BibTeX
@inproceedings{meir2023discrete,
  author = {Tomer Meir and Malka Gorfine},
  title = {{Discrete-time Competing-Risks Regression with or without Penalization}},
  booktitle = {Biometrics},
  year = {2023},
}
Figure