I. Van Keilegom
Let P represent the source population with complete data which contains covariate Z and response T, and Q the target population, where only the covariate Z is available. Our objective is to leverage the information from P to conduct statistical inference in Q. To simultaneously allow distribution difference in P and Q, and potential response censoring in P, we consider survival analysis under label shift, where we allow the marginal distribution of T to vary between the two populations, while assuming the conditional distribution of Z given T to remain the same. Although we adopt a parametric model of T given Z in Q, we find that the problem is far from parametric due to the label shift setting and we propose a nonparametric maximum likelihood based method to estimate the parameters in the model. This method enables statistical inference in Q and is applicable to various classic survival models. The asymptotic properties of the estimator are derived and the effectiveness of the method is demonstrated through simulations and a real data application.
Keywords: Cox proportional hazards model, domain adaptation, parametric model, likelihood
Scheduled
Statistics Plenary Lecture I. Van Keilegom
June 10, 2025 10:00 AM
Auditorio 1. Ricard Vinyes