A Soft-Thresholding Operator for Sparse Time-Varying Effects in Survival Models

TL;DR

This paper introduces a novel soft-thresholding operator for Cox models with sparse, time-varying effects, enabling simultaneous estimation of non-zero effects and zero-effect regions with interpretable confidence intervals.

Contribution

It proposes a new soft-thresholding approach for sparse, piecewise smooth time-varying coefficients in survival models, improving estimation and inference of zero and non-zero effect regions.

Findings

Effective in identifying zero-effect time intervals.

Produces interpretable confidence intervals for effects.

Demonstrates good finite sample performance in simulations.

Abstract

We consider a class of Cox models with time-dependent effects that may be zero over certain unknown time regions or, in short, sparse time-varying effects. The model is particularly useful for biomedical studies as it conveniently depicts the gradual evolution of effects of risk factors on survival. Statistically, estimating and drawing inference on infinite dimensional functional parameters with sparsity (e.g., time-varying effects with zero-effect time intervals) present enormous challenges. To address them, we propose a new soft-thresholding operator for modeling sparse, piecewise smooth and continuous time-varying coefficients in a Cox time-varying effects model. Unlike the common regularized methods, our approach enables one to estimate non-zero time-varying effects and detect zero regions simultaneously, and construct a new type of sparse confidence intervals that accommodate zero regions. This leads to a more interpretable model with a straightforward inference procedure. We develop an efficient algorithm for inference in the target functional space, show that the proposed method enjoys desired theoretical properties, and present its finite sample performance by way of simulations. We apply the proposed method to analyze the data of the Boston Lung Cancer Survivor Cohort, an epidemiological cohort study investigating the impacts of risk factors on lung cancer survival, and obtain clinically useful results.