Simultaneous variable selection for joint models of longitudinal and survival outcomes

Area of Science:

• Biostatistics
• Clinical Research Methodology
• Statistical Modeling

Background:

• Joint models for longitudinal and survival data are increasingly vital in clinical studies.
• Accurate specification of fixed and random effects is crucial for reliable data analysis.
• Simultaneous variable selection guards against joint model misspecification, yet remains understudied.

Purpose of the Study:

• To develop and present a novel penalized likelihood method for simultaneous variable selection in joint models.
• To provide practitioners with a computational tool for selecting fixed and random effects in complex longitudinal and survival data.
• To address the lack of existing methods for variable selection in joint longitudinal-survival models.

Main Methods:

• A penalized likelihood approach employing adaptive least absolute shrinkage and selection operator (ALASSO) penalties for simultaneous fixed and random effects selection.
• Reparameterization of random effects variance components using Cholesky decomposition to introduce group shrinkage penalties.
• A two-stage selection procedure to mitigate estimation bias introduced by penalization, with bias amelioration in the second stage.
• Approximation of penalized likelihood using Gaussian quadrature and optimization via an Expectation-Maximization (EM) algorithm.

Main Results:

• Simulation studies demonstrated excellent variable selection performance in the initial stage.
• The proposed method achieved small estimation biases in the second stage of the selection procedure.
• The methodology was successfully illustrated using a real-world dataset analyzing a longitudinal clinical marker and patient survival in heart failure.

Conclusions:

• The developed penalized likelihood method with ALASSO offers an effective solution for simultaneous variable selection in joint longitudinal and survival models.
• The two-stage procedure successfully balances selection accuracy and estimation bias.
• This approach provides a valuable computational tool for analyzing complex clinical data with both longitudinal and survival components.