Unbiased split variable selection for random survival forests using maximally selected rank statistics.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Cox regression is popular for survival data but can be misspecified, violating proportionality assumptions.
• Standard random survival forests (RSF) use log-rank statistics, introducing bias favoring variables with many split points.
• Conditional inference forests (CIF) avoid this bias but use linear statistics, missing non-linear covariate effects.

Purpose of the Study:

• To propose a novel random forest approach for survival prediction using maximally selected rank statistics.
• To address limitations of existing methods, including bias in variable selection and inability to detect non-linear effects.
• To evaluate the performance and computational efficiency of the new method compared to RSF and CIF.

Main Methods:

• Developed a random forest algorithm employing maximally selected rank statistics for split point selection.
• Utilized p-value approximations instead of Monte-Carlo methods for efficiency.
• Compared the new method against standard RSF and CIF using simulations and real-world datasets.

Main Results:

• The proposed method achieves unbiased split variable selection, though with a runtime trade-off.
• It outperforms RSF when combining dichotomous and multi-category variables.
• It surpasses CIF in scenarios with non-linear covariate effects.
• With simple p-value approximations, the new method is computationally faster than both alternatives.

Conclusions:

• The new random forest approach offers an effective alternative for survival prediction, balancing unbiased variable selection and computational efficiency.
• It demonstrates superior predictive performance, particularly when dealing with complex covariate structures and non-linear relationships.
• The method provides a valuable tool for survival data analysis, overcoming key limitations of traditional and existing machine learning techniques.