sox: Structured Learning in Time-Dependent Cox Models

Yi Lian, Guanbo Wang, Archer Y. Yang, Mireille E. Schnitzer, Robert W. Platt, Rui Wang, Marc Dorais, Sylvie PerreaultOpenalex Unknown 2023 · DOI · OpenAlex

Abstract: Efficient procedures for fitting and cross-validating the structurally-regularized time-dependent Cox models.

Introduction

The sox R package provides a specialized toolkit for fitting and cross-validating time-dependent Cox proportional hazards models using structured regularization. By implementing nested and overlapping group lasso penalties, it allows researchers to incorporate prior knowledge about feature relationships—such as hierarchies or interactions—directly into high-dimensional survival analysis.

The Challenge of High-Dimensional Survival Analysis

In many biomedical and social science applications, modeling 'time-to-event' data involves predictors that change over time. While the Cox proportional hazards model is the standard approach, it struggles with high-dimensional data where the number of predictors is large. Standard regularization techniques like the Lasso handle high dimensionality by inducing sparsity, but they treat every variable as independent.

In reality, variables often possess inherent structures:

  1. Nested Structure: Interaction terms (e.g., Treatment x Age) should ideally only enter a model if their corresponding main effects are present.
  2. Overlapping Structure: A single biological feature, such as a gene, may participate in multiple pathways, requiring it to be penalized as part of several groups simultaneously.

The sox package addresses the computational complexity of applying these structured penalties within the mathematical framework of time-dependent Cox models.

Methodology: Structured Regularization

The package implements efficient procedures for structured learning via two primary penalty types: the Overlapping Group Lasso and the Nested Group Lasso. The latter enforces a hierarchy where 'child' groups are selected only if their 'parent' groups are active.

Automating Complex Structures

A significant hurdle in structured learning is the manual construction of group matrices and indices. The sox package mitigates this through automated helper functions like nested_structure() and overlap_structure().

The nested_structure function automates the generation of objects required for hierarchical penalties.

Technical Implementation

To handle the optimization of these complex penalties, sox utilizes Proximal Gradient Descent with a backtracking line search. Defining the relationships between overlapping groups and variables requires specific input structures, such as integer matrices and vectors to describe group subsets and variable indices.

Required input parameters for defining group and variable relationships in sox.

Stability in Cross-Validation

For model selection, sox_cv employs a deviance-based metric (2(P - Q) / R). By dividing the deviance by the number of events (R) rather than total observations, the package ensures more stable error estimates, particularly when the number of events in a test set is small.

Datasets and Evaluation

The package includes a built-in demo dataset, sim, which consists of simulated longitudinal data for 50 subjects. It features categorical and continuous variables along with interaction terms. Event times were generated using the PermAlgo package to ensure compliance with time-dependent covariate modeling requirements.

Evaluation is facilitated through:

  • Solution Paths: Visualizing coefficient profiles as the penalty parameter (lambda) varies.
  • Optimal Tuning: Identifying lambda.min (minimum cross-validated error) and lambda.1se (the most parsimonious model within one standard error).

Conclusion and Future Directions

By leveraging Rcpp for high-performance C++ computations, sox makes structured penalty approaches feasible for larger survival datasets. While currently focused on the Cox model and its proportional hazards assumption, future work may extend these structured learning techniques to other survival frameworks, such as Accelerated Failure Time (AFT) models.

Summary updated

Abstract

Efficient procedures for fitting and cross-validating the structurally-regularized timedependent Cox models.
BibTeX
@inproceedings{lian2023sox,
  author = {Yi Lian and Guanbo Wang and Archer Y. Yang and Mireille E. Schnitzer and Robert W. Platt and Rui Wang and Marc Dorais and Sylvie Perreault},
  title = {{sox: Structured Learning in Time-Dependent Cox Models}},
  booktitle = {Openalex Unknown},
  year = {2023},
  doi = {10.32614/cran.package.sox},
}
Figure