Assessing moderated mediation in linear models requires fewer confounding assumptions than assessing mediation

Area of Science:

• Causal inference
• Statistical modeling
• Social sciences

Background:

• Mediation analysis typically requires strong no unmeasured confounding assumptions for direct and indirect effects.
• Even in randomized studies, unobserved variables can bias mediator role inference.
• Causal assumptions for moderated mediation analysis are under-explored in social sciences.

Purpose of the Study:

• To investigate the causal assumptions for moderated mediation analysis.
• To focus on the index for moderated mediation, quantifying how mediated effects change with a moderator.
• To demonstrate unbiased estimation of this index under less stringent confounding conditions.

Main Methods:

• Development of a statistical framework for moderated mediation analysis.
• Focus on linear models to estimate the index for moderated mediation.
• Application of the method to both randomized and observational study data.

Main Results:

• The index for moderated mediation can be estimated without bias in linear models despite unmeasured common causes of moderator, mediator, and outcome, under certain conditions.
• This allows for supporting mediation evidence under less stringent confounding assumptions.
• The findings are illustrated with real-world data from social psychology and relationship research.

Conclusions:

• The proposed method for moderated mediation analysis offers a more robust approach by relaxing traditional confounding assumptions.
• This advancement supports the use of moderated mediation tests for stronger evidence of mediation.
• The findings have implications for causal inference in behavioral and social sciences, enhancing the reliability of mediation studies.