Prediction accuracy and variable selection for penalized cause-specific hazards models.

Area of Science:

• Biostatistics
• High-dimensional data analysis
• Survival analysis

Background:

• Competing risks data present unique challenges in high-dimensional settings, particularly in molecular data analysis.
• Understanding underlying biological mechanisms requires analyzing transition hazards.
• Existing methods like subdistribution hazards (SDH) models offer linking but sacrifice interpretability.

Purpose of the Study:

• To evaluate the effectiveness of penalized cause-specific hazards (CSHs) models for high-dimensional competing risks data.
• To investigate the utility of linking separate CSH models (penCR) for improved prediction and variable selection.
• To compare penalized CSH approaches (iCS and penCR) against the subdistribution hazards (SDH) model.

Main Methods:

• Implementation of independent penalized cause-specific hazards (iCS) models for each event type.
• Development of a linked penalized cause-specific hazards model (penCR) by optimizing combined penalty tuning parameters.
• Comparative analysis using simulation studies against the subdistribution hazards (SDH) model.

Main Results:

• Independent penalized CSH (iCS) models demonstrate competitive performance against penCR and SDH approaches in many aspects.
• Linking CSH models (penCR) shows advantages in specific situations, such as when opposing effects on CSHs are present.
• Penalized CSH models are a viable and often justified solution for high-dimensional competing risks modeling.

Conclusions:

• Penalized cause-specific hazards models provide a robust framework for high-dimensional competing risks data.
• Linking CSH models can offer benefits in prediction accuracy and variable selection in particular scenarios.
• Separately penalized CSH models are frequently adequate and maintain better interpretability of covariate effects.