I. Willems, J. Beyhum, I. Van Keilegom
Modeling the time until a certain event of interest takes place is crucial in various applied fields, yet it often involves censoring mechanisms which challenge traditional approaches. In the current literature, many models require stringent assumptions in order to guarantee valid inference under censoring, especially when this censoring is dependent on the event of interest. In this research, we oppose such approaches by proposing a general model to study the effect of covariates on the distribution of the latent event time which makes minimal assumptions about the censoring mechanism. This leads us to a partially identified approach, in the sense that the parameters of interest are allowed to be not uniquely determined by the distribution of the observed data and the maintained assumptions. We achieve partial identification by exploiting Peterson's bounds on the conditional distribution of the event time and casting the problem into one defined by unconditional moment restrictions, which allows us to rely on a vast basis of established theory. As a special case, our approach can be used to study the popular Cox proportional hazards model while leaving the censoring distribution as well as its dependence with the time of interest completely unspecified. A simulation study illustrates good finite sample performance of the proposed approach and a data application demonstrates its practicability on real data.
Keywords: survival analysis, informative censoring, partial identification, moment inequality models, test inversion, Cox proportional hazards model.
Scheduled
FENStatS-SEIO: Statistics and Data Science
June 11, 2025 10:30 AM
Auditorio 1. Ricard Vinyes