Boosted nonparametric hazards with time-dependent covariates (Q2054480)

The present article introduces a fully nonparametric boosted hazard estimator with time-dependent covariates. The estimate is based on an iterative gradient descent algorithm for minimizing a loss functional over a target function space. This is achieved by first providing an integral representation for the likelihood risk which in turn allows for obtaining an explicit representation for the gradient. By defining a suitable space of log-hazard functions on the time-covariate domain allows for deriving the gradient needed to implement gradient boosting. The authors establish the consistency of the procedure, by showing that the hazard estimator is consistent if the space is correctly specified. They further clarify the mechanisms used by gradient boosting to avoid overfitting. Finally, the proposed estimate can also potentially be used to evaluate the goodness-of-fit of simpler parametric hazard models.