Identifiability and estimation of causal mediation effects with missing data

Area of Science:

• Statistics
• Biostatistics
• Causal Inference

Background:

• Mediation analysis is crucial for understanding intervention mechanisms in various scientific fields.
• Missing data, particularly missing not at random (MNAR) data, significantly hinders causal mediation analysis.
• Existing methods for handling missing data in causal mediation are insufficient.

Purpose of the Study:

• To investigate the identifiability of causal mediation effects under different missing outcome scenarios and missingness mechanisms.
• To develop and validate novel statistical approaches for estimating causal mediation effects with missing data.
• To provide a robust method for handling MNAR outcomes in causal mediation analysis.

Main Methods:

• Demonstrated the identifiability of causal mediation effects with various missing outcome types and missingness mechanisms.
• Developed an estimating equation-based approach for estimating causal mediation effects, particularly for MNAR data.
• Established asymptotic results for the proposed estimators and conducted simulations to evaluate performance.

Main Results:

• Causal mediation effects were shown to be identifiable under specified missing data conditions.
• The proposed estimating equation-based approach effectively handles different mediator and outcome types with MNAR data.
• Simulation studies indicated favorable performance of the new estimators in finite samples.

Conclusions:

• The developed methods enhance the applicability of causal mediation analysis in the presence of missing outcomes, including MNAR data.
• The approach is versatile, accommodating various mediators and outcomes.
• The study provides a valuable tool for researchers dealing with missing data in mediation studies, as illustrated by its application to Alzheimer's disease research.