A comparative study of forest methods for time-to-event data: variable selection and predictive performance.

Area of Science:

• Machine Learning
• Biostatistics
• Data Science

Background:

• Survival forests are popular machine learning alternatives to Cox models.
• Random survival forests (RSF) suffer from selection bias.
• Conditional inference forests (CIF) reduce bias but are computationally intensive.
• Random forests with maximally selected rank statistics (MSR-RF) aim to improve upon RSF and CIF.

Purpose of the Study:

• Compare prediction and variable selection performance of RSF, CIF, and MSR-RF.
• Evaluate methods using simulation studies and real data.
• Identify optimal survival forest methods based on data characteristics.

Main Methods:

• Utilized simulation studies and real-world data applications.
• Assessed prediction accuracy using Integrated Brier Score (IBS) and c-index.
• Evaluated variable selection by calculating the frequency of correct variable ranking.

Main Results:

• All three methods showed comparable prediction performance.
• MSR-RF and RSF outperformed CIF with continuous or binary variables.
• RSF had lower selection frequency with categorical variables; MSR-RF and CIF performed better.
• MSR-RF demonstrated superior performance with continuous variables and robustness with increased dimensions.

Conclusions:

• MSR-RF shows practical value and warrants popularization.
• Each survival forest method has distinct advantages depending on data structure and research goals.
• Selecting the appropriate method is crucial for effective analysis of survival data.