Effects of Proportional Hazard Assumption on Variable Selection Methods for Censored Data

Area of Science:

• Biostatistics
• Survival Analysis
• Statistical Modeling

Background:

• The Cox proportional hazard (PH) model is a standard tool for analyzing survival data.
• Covariate selection is critical but relies on model assumptions like PH, Accelerated Failure Time (AFT), or proportional odds (PO).
• These models may not accommodate crossing survival functions or delayed effects, potentially impacting variable selection accuracy.

Purpose of the Study:

• To investigate the impact of the proportional hazard (PH) assumption on covariate selection when data deviates from PH.
• To explore alternative modeling approaches for improved variable selection under non-PH conditions.
• To evaluate the influence of stratifying by an off-treatment indicator on covariate selection.

Main Methods:

• Comparison of covariate selection under penalized PH (elastic-net) and linear spline hazard regression models.
• Application of models to the ACTG-175 dataset and simulated data.
• Analysis of survival times generated from Weibull and log-normal distributions.

Main Results:

• The study demonstrates that violations of the proportional hazard assumption can adversely affect covariate selection.
• Alternative models, specifically penalized PH and linear spline hazard regression, offer different approaches to variable selection.
• Stratification on the off-treatment indicator can influence the covariate selection outcomes.

Conclusions:

• The proportional hazard assumption's limitations necessitate exploring alternative models for robust covariate selection in survival analysis.
• Penalized PH and linear spline hazard models provide viable strategies when PH assumptions are questionable.
• Careful consideration of model assumptions and data characteristics is crucial for accurate risk factor identification in survival data.