Understanding Interventional Effects: A More Natural Approach to Mediation Analysis?

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Traditional causal mediation analysis relies on "natural effects," which require empirically unverifiable assumptions.
• Natural effects face significant estimation challenges, especially with exposure-induced mediator-outcome confounding.
• Untestable assumptions limit the practical application and reliability of natural effects in mediation studies.

Purpose of the Study:

• To introduce and clarify the interpretation of "interventional effects" as an alternative to natural effects in causal mediation.
• To describe hypothetical randomized controlled trials for estimating interventional effects.
• To highlight the shift from individual-level to population-level interventions in mediation analysis.

Main Methods:

• Conceptualizing interventional effects based on population-level interventions.
• Describing hypothetical randomized controlled trials (RCTs) to estimate interventional effects.
• Contrasting the definitions and assumptions of natural versus interventional effects.

Main Results:

• Interventional effects overcome the limitations of natural effects, particularly regarding unverifiable assumptions and confounding.
• Two versions of interventional effects have been proposed to address exposure-induced confounding.
• The interpretation of interventional effects is grounded in population-level interventions, not individual-level ones.

Conclusions:

• Interventional effects provide a more empirically tractable and conceptually sound approach to mediation analysis.
• The shift to population-level interventions offers more tangible, average effects compared to individual-level natural effects.
• This framework has significant conceptual and practical implications for conducting mediation analysis.