Neural Network Assisted Estimation for the Structural Nested Accelerated Failure Time Models.

Area of Science:

• Causal inference
• Survival analysis
• Machine learning

Background:

• Time-varying confounding in longitudinal data poses challenges for traditional causal survival analysis.
• Standard models adjusting for time-dependent covariates may not yield unbiased intervention causal effect estimates.
• The Structural Nested Accelerated Failure Time Model (SNAFTM) offers a robust framework but often involves computationally intensive G-estimation.

Purpose of the Study:

• To develop novel, computationally efficient algorithms for estimating the SNAFTM.
• To address limitations of G-estimation in high-dimensional, temporally dependent data.
• To provide less biased and individualized intervention causal effect estimations.

Main Methods:

• Introduction of two Neural Network-based algorithms: GE-SCORE and GE-MIMIC.
• These algorithms are designed to estimate the SNAFTM, handling high-dimensional input data.
• Validation through simulations and application to a real-world observational dataset (CARDIA).

Main Results:

• Simulations demonstrate that GE-SCORE and GE-MIMIC provide less biased and individualized intervention causal effect estimations.
• The algorithms effectively handle high-dimensional and temporally connected data.
• Application to the CARDIA dataset successfully identified and quantified smoking's causal effect on cardiovascular events.

Conclusions:

• GE-SCORE and GE-MIMIC offer powerful, data-driven alternatives for causal survival analysis with time-varying confounding.
• These neural network approaches overcome the computational burden and power limitations of traditional G-estimation.
• The methods have practical utility in identifying and quantifying causal effects in complex observational studies.