Exploratory subgroup identification in the heterogeneous Cox model: A relatively simple procedure.

Area of Science:

• Biostatistics
• Survival Analysis
• Clinical Trial Methodology

Background:

• Identifying patient subgroups with differential treatment effects is crucial for personalized medicine.
• Existing methods may struggle to detect subgroups where treatment is potentially detrimental.
• The need for flexible and robust subgroup identification methods in survival data analysis.

Purpose of the Study:

• To propose a novel procedure, 'forest search,' for identifying subgroups with significant treatment effects, including those experiencing harm.
• To develop a method that is simple, flexible, and applicable to survival analysis.
• To compare the performance of the proposed method against existing approaches like virtual twins and generalized random forests.

Main Methods:

• Screening all possible subgroups using hazard ratio thresholds indicative of harm within the Cox model framework.
• Applying a splitting consistency criterion to identify subgroups 'maximally consistent with harm.'
• Utilizing numerical integration for approximating type-1 error and power, and employing bootstrap bias-corrected Cox model estimation with Jackknife variance approximation.

Main Results:

• The proposed forest search approach demonstrates favorable performance compared to virtual twins and generalized random forests.
• The method effectively controls type-1 error for falsely identifying heterogeneity.
• Forest search exhibits higher power and classification accuracy for substantial heterogeneous treatment effects.

Conclusions:

• Forest search is a simple, flexible, and effective procedure for identifying subgroups with significant treatment effects, particularly those at risk of harm.
• The method offers improved control of false discoveries and enhanced power for detecting true treatment effect heterogeneity.
• The approach is validated through simulations and real-world applications in oncology and HIV clinical trial data.