Counterfactual graphical models for longitudinal mediation analysis with unobserved confounding

Area of Science:

• Causal inference and mediation analysis across psychology, medicine, social sciences, and public health.
• Application of counterfactual frameworks to understand mediated causal effects.

Background:

• Mediation analysis is crucial in numerous disciplines, with standard methods relying on regression models (difference and product methods).
• Existing parametric approaches are often unappreciated as special cases of a broader counterfactual framework for causality.

Purpose of the Study:

• To elucidate the advantages of the counterfactual framework for mediation analysis.
• To address limitations and assumptions of traditional difference and product methods.
• To extend mediation analysis to complex longitudinal settings.

Main Methods:

• Discussion of parametric mediation approaches as a subset of the Neyman-Rubin counterfactual framework.
• Identification and explication of underlying assumptions in mediation analysis.
• Development of a novel result for applying mediation analysis in longitudinal settings with unobserved confounders.

Main Results:

• The counterfactual framework makes implicit assumptions explicit, enhancing transparency.
• This framework circumvents issues like biased effect estimates present in product and difference methods.
• A new method enables mediation analysis in longitudinal data even with unobserved confounding.

Conclusions:

• The counterfactual framework provides a more robust and explicit approach to mediation analysis.
• This generalized framework offers significant advantages over traditional methods, particularly in complex scenarios.
• The novel result extends the applicability of mediation analysis to challenging longitudinal research designs.