Random Survival Forest
Also known as: RSF, Rastgele Sağkalım Ormanı (RSF), survival random forest
Random Survival Forest (RSF), introduced by Ishwaran, Kogalur, Blackstone, and Lauer in 2008, is an ensemble machine learning method that adapts the Random Forest algorithm to time-to-event (survival) data. Trees are grown using log-rank splitting to handle censored observations naturally, and the ensemble aggregates cumulative hazard functions across hundreds of trees to produce predictions and variable importance rankings.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
RSF is appropriate when you have time-to-event data with right-censoring, a relatively large sample (at least 100 events is recommended), and you suspect that predictors interact or relate to the hazard in a non-linear way that a Cox model would not capture without manual specification. It does not require the proportional hazards assumption, making it useful when that assumption is violated. Because RSF is a black-box ensemble, it is best suited for prediction and variable screening rather than formal causal inference. In small samples, the tree splits become unstable and Cox regression is safer.
Strengths & limitations
- Handles right-censored data natively through log-rank splitting — no ad hoc adjustments needed.
- Captures non-linear relationships and predictor interactions automatically.
- Provides a data-driven variable importance ranking (VIMP) across all predictors.
- Does not assume proportional hazards, so it remains valid when Cox PH is violated.
- Robust to outliers and irrelevant predictors due to random feature selection at each split.
- Interpretability is low compared with Cox regression — it is essentially a black-box model.
- Requires at least 100 events for stable splits; performance degrades markedly in small samples.
- Computationally intensive: fitting hundreds of survival trees is substantially slower than fitting a single Cox model.
- Does not produce simple hazard ratio summaries; communicating results to clinical audiences requires extra effort.
Frequently asked
How does RSF differ from Cox regression?
Cox regression fits a single semi-parametric model that assumes proportional hazards and estimates a hazard ratio for each predictor. RSF builds an ensemble of hundreds of survival trees, makes no proportional-hazards assumption, and captures non-linear effects and interactions automatically. Cox gives interpretable coefficients; RSF gives better prediction in complex settings at the cost of interpretability.
What is VIMP and how should it be interpreted?
VIMP (Variable IMPortance) measures how much the out-of-bag prediction error (1 − C-index) increases when a predictor's values are randomly permuted. A large positive VIMP means the predictor contributes substantially to predictive accuracy; a VIMP near zero or negative means it adds little. VIMP is a measure of predictive relevance, not a causal effect estimate.
Does RSF require proportional hazards?
No. RSF uses the log-rank splitting criterion, which is sensitive to any difference between survival distributions in the two child nodes, not only proportional ones. This makes RSF valid even when the hazard ratio between groups changes over time — a setting where Cox regression is problematic.
How many trees and what minimum node size should I use?
The original authors recommend 500–1000 trees for stable estimates; increasing beyond 1000 usually yields diminishing returns. Minimum terminal node size is the key tuning parameter: the default of 3 is often too small for survival data, and a value of 6–15 is commonly recommended to reduce overfitting, especially when the event rate is low.
Sources
- Ishwaran, H., Kogalur, U.B., Blackstone, E.H. & Lauer, M.S. (2008). Random Survival Forests. Annals of Applied Statistics, 2(3), 841–860. DOI: 10.1214/08-AOAS169 ↗
How to cite this page
ScholarGate. (2026, June 1). Random Survival Forest. ScholarGate. https://scholargate.app/en/survival/random-survival-forest
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Kaplan-MeierSurvival↔ compare
- Nelson-Aalen EstimatorSurvival↔ compare