Comparison of Random Survival Forests Split rules in selecting the determinants of under-five mortality in Kenya using 2014 DHS data.

Abstract

Background Survival trees and Random Survival Forests are extensions of classi cation and regression trees and random forests in analyzing time to event data. These methods are alternatives to Cox Proportional hazards models when Proportional Hazard assumption is violated. Survival tree methods are exible and can handle high dimensional covariate data as they are fully non-parametric. Random survival forests use the Brieman’s approach, rst, by employing a random selection of a bootstrap sample used for growing a tree then growing tree learners by splitting the nodes on the randomly selected predictors. The performance of the survival trees highly depended on the splitting method that is applied while growing the tree. 0.2 Methods In our analysis, we compare the performance of random survival forests in variable selecting based on the following split rules; Log-rank splitting, Log-rank score splitting, and and Conditioned Inference Forests. Our outcome variable is the under- ve child mortality in Kenya using 2014 DHS data. Covariates that were included in the models were chosen based on the existing literature. 0.3 Results Findings from this study show that Log-rank split rule outperforms Log-rank score split rule. Both split rules analyze time to event data based on the bootstrap cross-validated estimates for integrated Brier scores. 0.4 Conclusion As much as it is evident that Log-rank is the best, there is need to investigate other split rules and the nature of data that that best suit each split rule to be able to identify the best slitting method.