In their paper, Yang et al. tackle the important challenge of identifying biomarkers that are predictive for a time-to-event response while taking into account relevant risk factors. The proposed solution is a doubly robust conditional independence test based on the model-X framework, that is, the test relies on sampling from the distribution of the biomarker of interest conditional on the relevant risk factors. However, the paper falls short of helping biometricians to make an informed decision on when the proposed test can be used and what alternative doubly robust tests exist in the literature with directly usable and open source software implementations. This comment intends to close this gap by (i) discussing the assumptions on the censoring mechanism that are sufficient for the test to be valid; and (ii) providing a small scale empirical comparison between the test by Yang et al. and another established doubly robust conditional independence tests for time-to-event responses based on the Generalised Covariance Measure. The results show that the test by Yang et al. performs on par with the existing test in terms of type I error control and power, while being computationally more expensive due to the refitting steps. Code to reproduce all results is available in the Supplementary Materials and on GitHub.
Lucas Kook (Tue,) studied this question.