2015/07/11 by Matthias Schmid, Schmid, Matthias, Marvin N. Wright +3 · 3 citations
Biochemistry, Genetics and Molecular Biology · Mathematics · #Advanced Causal Inference Techniques #FOS: Computer and information sciences #Genetic Associations and Epidemiology #Machine Learning (stat.ML) #Statistical Methods and Inference
paper · pdf · doi:10.48550/arxiv.1507.03092
openalex publication_date 2015/07/11 · openalex created_date 2022/10/05 · openalex updated_date 2026/07/28
Random survival forests (RSF) are a powerful method for risk prediction of\nright-censored outcomes in biomedical research. RSF use the log-rank split\ncriterion to form an ensemble of survival trees. The most common approach to\nevaluate the prediction accuracy of a RSF model is Harrell's concordance index\nfor survival data ('C index'). Conceptually, this strategy implies that the\nsplit criterion in RSF is different from the evaluation criterion of interest.\nThis discrepancy can be overcome by using Harrell's C for both node splitting\nand evaluation. We compare the difference between the two split criteria\nanalytically and in simulation studies with respect to the preference of more\nunbalanced splits, termed end-cut preference (ECP). Specifically, we show that\nthe log-rank statistic has a stronger ECP compared to the C index. In\nsimulation studies and with the help of two medical data sets we demonstrate\nthat the accuracy of RSF predictions, as measured by Harrell's C, can be\nimproved if the log-rank statistic is replaced by the C index for node\nsplitting. This is especially true in situations where the censoring rate or\nthe fraction of informative continuous predictor variables is high. Conversely,\nlog-rank splitting is preferable in noisy scenarios. Both C-based and log-rank\nsplitting are implemented in the R~package ranger. We recommend Harrell's C as\nsplit criterion for use in smaller scale clinical studies and the log-rank\nsplit criterion for use in large-scale 'omics' studies.\n