Structured fusion lasso penalized multi-state models.

Area of Science:

• Biostatistics
• Survival Analysis
• Statistical Modeling

Background:

• Multi-state models analyze complex event data with multiple transitions.
• High-dimensional models with many covariates pose estimation challenges.
• Existing methods like LASSO regularize coefficients but don't fully address multi-transition relationships.

Purpose of the Study:

• To develop a novel estimation approach for sparse multi-state modeling.
• To simultaneously penalize L1-norm of coefficients and their differences for dimensionality reduction.
• To improve the interpretability and structure identification of multi-state models.

Main Methods:

• A new estimation approach combining multi-state model estimation with structured penalization.
• Simultaneous penalization of the L1-norm of covariate coefficients.
• Penalization of absolute differences between coefficients for the same covariate across different transitions.
• Implementation in the R package penMSM.

Main Results:

• The proposed method effectively reduces parameter space dimensionality in multi-state models.
• It facilitates the identification of sparse relationships between covariates and transitions.
• The approach was successfully illustrated using peritoneal dialysis study data.

Conclusions:

• The new sparse multi-state modeling approach offers a powerful tool for analyzing complex event data.
• It enhances model interpretability by identifying key covariate effects and simplifying model structure.
• The R package penMSM provides a practical implementation for researchers.