Mediation Analysis of Failure Time Data Under Interval Censoring

Motivated by disentangling the direct and indirect genetic effects on the Alzheimer's disease progression, we propose a mediation approach for analyzing interval-censored failure time data. Under the counterfactual framework with sequential ignorability, generalized linear and spline-based proportional hazards (PH) models are employed to depict path-specific effects (PSEs) of an exposure on the interested failure time mediated through multiple mediators. In particular, the PSEs are derived on the basis of both survival probability and restricted mean survival time, bypassing the restrictive rare outcome assumption inherent in the PH model. A sieve maximum likelihood method is proposed to estimate all unknown parameters in the PH model. Simulation studies demonstrate the satisfactory empirical performances of the proposed mediation method. A real-world data analysis quantifies the mediation effects of Apolipoprotein E- $ϵ$ 4 allele on the risk of developing the Alzheimer's disease.