A1814
Title: Robust estimation and inference for accelerated failure time models via double machine learning
Authors: Kun Ren - (Hong Kong) [presenting]
Abstract: Valid inference in accelerated failure time models with right-censored time-to-event data is challenging, particularly when infinite-dimensional nuisance functions are approximated via complex machine learning methods. A multi-stage estimation method leveraging double machine learning techniques is proposed to enable robust inference. The proposed approach mitigates biases derived from nonparametric nuisance estimation tasks and achieves asymptotically normal estimators under mild conditions. Furthermore, the approach generalizes prediction-powered inference to right-censored settings, providing a flexible guideline for maintaining statistical interpretability with black-box predictive models.
