Mediation Analysis for Censored Survival Data Under an Accelerated Failure Time Model

Area of Science:

• Biostatistics
• Epidemiology
• Causal Inference

Background:

• Causal mediation analysis estimates direct and indirect effects of exposures.
• Accelerated failure time (AFT) models are common for survival outcomes in health research.
• Standard assumptions suggest difference and product methods for indirect effects are equivalent in AFT models.

Purpose of the Study:

• To formally investigate the validity of difference and product methods for estimating indirect effects in AFT models with censored data.
• To identify conditions under which these methods may yield different estimates.
• To provide guidance on appropriate methods for causal mediation analysis in survival data.

Main Methods:

• Formal theoretical analysis of causal mediation methods in the context of AFT models.
• Investigation of the impact of censoring and truncation on estimation methods.
• Simulation studies to evaluate method performance under various scenarios.
• Application to two real-world health datasets.

Main Results:

• The product method for indirect effects remains valid under independent censoring in AFT models.
• The difference method can be biased with independent censoring if the error distribution is not collapsible.
• Bias in the difference method occurs unless specific error distribution assumptions are met (e.g., normal-normal).
• Simulation studies and data applications confirm these theoretical findings.

Conclusions:

• The choice of method for estimating indirect effects in AFT models is critical when dealing with censored survival data.
• The product method is generally more robust to censoring than the difference method.
• Researchers should carefully consider error distribution assumptions and potential for bias when applying causal mediation analysis to survival data.