Clarifying causal mediation analysis: Effect identification via three assumptions and five potential outcomes

Area of Science:

• Causal inference
• Statistical modeling
• Epidemiology

Background:

• Causal mediation analysis involves complex effect definitions with varying identification assumptions.
• Applied researchers often face challenges in understanding and selecting appropriate assumptions for mediation analysis.

Purpose of the Study:

• To systematically explain the assumptions required for identifying different causal mediation effect definitions.
• To provide a clear framework for applied researchers to select assumptions based on their target causal effects.
• To clarify the identification requirements for commonly used causal contrasts in mediation analysis.

Main Methods:

• Definition of five potential outcome types relevant to mediation effect definitions.
• Stepwise identification of mean/distribution assumptions, from weakest to strongest.
• Illustration using a running example to demonstrate assumption assembly for various causal contrasts.

Main Results:

• Systematic explanation clarifies why specific assumptions are needed for particular causal mediation estimands.
• Identification of commonly encountered causal contrasts requiring weaker assumptions than previously stated.
• Highlights differences in positivity assumption requirements across different estimands, with practical implications.

Conclusions:

• Enhanced clarity on identifying assumptions facilitates appropriate causal mediation analysis.
• Researchers can better interpret results by carefully considering the plausibility of identified assumptions.
• This work aids in conducting more rigorous and transparent causal mediation studies.