Penalized variable selection for accelerated failure time models with random effects.

Area of Science:

• Biostatistics
• Survival Analysis
• Statistical Modeling

Background:

• Accelerated failure time (AFT) models with random effects are crucial for analyzing correlated survival data, offering an alternative to frailty models.
• Existing literature lacks robust variable selection methods for these complex AFT models.
• Censored survival data and dependence structures require specialized statistical approaches.

Purpose of the Study:

• To propose a unified variable selection procedure for fixed effects in AFT random-effect models.
• To introduce penalized h-likelihood (HL) methods for variable selection in this context.
• To extend the methodology to multilevel and nested survival data structures.

Main Methods:

• Development of a variable selection procedure using penalized h-likelihood (HL) estimation.
• Evaluation of four penalty functions: least absolute shrinkage and selection operator (LASSO), adaptive LASSO, smoothly clipped absolute deviation (SCAD), and HL.
• Implementation via modifications to existing h-likelihood estimation procedures.
• Extension to multilevel and nested AFT models.

Main Results:

• The proposed method is easily implementable and extendable to complex data structures.
• Simulation studies indicate strong performance for adaptive LASSO, SCAD, and HL penalties.
• The HL penalty demonstrated a higher probability of selecting the true model compared to other methods.
• The method's utility was validated on real-world data from multicenter clinical trials.

Conclusions:

• The penalized h-likelihood approach offers an effective and unified method for variable selection in AFT random-effect models.
• The HL penalty is particularly recommended for its superior model selection accuracy.
• This methodology enhances the analysis of correlated and multilevel survival data, with practical applications in clinical trials.