Abstract

Time-varying confounding complicates the causal survival analysis for longitudinal data. Traditional survival models that adjust for time-dependent covariates fail to estimate the intervention causal effect unbiasedly. The Structural Nested Accelerated Failure Time Model (SNAFTM) can address this challenge effectively. This model estimates the intervention causal effect as the acceleration factor of the survival time while controlling for the time-varying confounders. However, the SNAFTM model usually relies on the G-estimation, which lacks power and suffers from computational burden, especially when the model input data is high-dimensional with a temporally connected nature. This manuscript presents two Neural Networks based algorithms (GE-SCORE and GE-MIMIC) that estimate the SNAFTM. These two algorithms can handle high-dimensional input data while providing less biased and individualized intervention causal effect estimation, as demonstrated by simulations. The proposed algorithms were also applied to a real observational dataset (CARDIA), and we successfully identified and quantified subjects' smoking causal effects on the time to first cardiovascular events.