Abstract: Boosting methods have been broadly discussed for various settings, and most methods handle data with complete observations. Although some methods are available for survival data with censored responses, they tend to assume a specific model for the survival process, and most provide numerical implementation procedures without rigorous theoretical justifications. In this paper, we develop an unbiased boosting estimation method for censored survival data, without assuming an explicit model, and explore three strategies for adjusting the loss functions, while accommodating censoring effects. We implement the proposed method using a functional gradient descent algorithm, and rigorously establish the theoretical results, including the consistency and optimization convergence. Our numerical studies show that the proposed method exhibits satisfactory performance in finite-sample settings.

Key words and phrases: Adjusted loss functions, boosting, consistency, empirical processes, machine learning, right-censoring, survival data.