Überlebensbäume

A survival tree is a non-parametric method that models the relationship between covariates and time until an event occurs using recursive partitioning. Unlike parametric approaches, survival trees do not require assumptions about survival time distribution or functional relationships between variables and survival outcomes. They visualize complex covariate interactions through an interpretable tree structure.

Recursive partitioning splits the dataset into subsets at each node based on a covariate that best differentiates survival outcomes. The splitting criterion, typically the log-rank test, compares survival distributions between groups to identify the optimal split. This process continues recursively until stopping criteria are met, creating branches that represent variable value splits and nodes representing data subsets.

Survival tree construction requires four main parameters: covariates (predictor variables that can be continuous, ordinal, or categorical), a splitting criterion to choose the best split at each node, minimum node size to prevent overfitting, and pruning thresholds to determine when to stop pruning. These parameters control tree growth and ensure the model generalizes well to new data.

Pruning removes nodes that do not significantly improve model accuracy, preventing overfitting and ensuring the tree generalizes to new data. Without proper pruning, survival trees can memorize training data patterns and perform poorly on unseen observations. Pruning thresholds determine when to stop this process, balancing model complexity with predictive accuracy.

Survival trees accommodate censored data, where events have not occurred for some individuals by study end or exact event times are unknown. Terminal nodes in the tree are evaluated using Kaplan-Meier estimates of the survival function, which properly account for censoring. This provides survival probability estimates for subjects in each node despite incomplete event information.

Advantages include flexibility with various data types, robustness to outliers and missing values, and easy interpretability through visual tree structure. Disadvantages include susceptibility to overfitting without proper pruning and instability where small data changes cause significant tree structure changes. These limitations make survival trees less stable than ensemble methods like survival forests.

Terminal nodes represent final data subsets and indicate the number of subjects in each node. They provide predictions through Kaplan-Meier survival function estimates, which calculate survival probabilities for subjects falling into that node. This allows clinicians and researchers to estimate survival outcomes for new patients based on their covariate profile and corresponding terminal node assignment.