The state learner -- a super learner for right-censored data

In survival analysis, prediction models are needed as stand-alone tools and in applications of causal inference to estimate nuisance parameters. The super learner is a machine learning algorithm which combines a library of prediction models into a meta learner based on cross-validated loss. In right-censored data, the choice of the loss function and the estimation of the expected loss need careful consideration. We introduce the state learner, a new super learner for survival analysis, which simultaneously evaluates libraries of prediction models for the event of interest and the censoring distribution. The state learner can be applied to all types of survival models, works in the presence of competing risks, and does not require a single pre-specified estimator of the conditional censoring distribution. We establish an oracle inequality for the state learner and investigate its performance through numerical experiments. We illustrate the application of the state learner with prostate cancer data, as a stand-alone prediction tool, and, for causal inference, as a way to estimate the nuisance parameter models of a smooth statistical functional.