Nonparametric causal mediation analysis for stochastic interventional (in)direct effects.

Area of Science:

• Causal inference
• Statistical methodology
• Epidemiology

Background:

• Traditional causal mediation analysis is limited to binary exposures and static interventions.
• Existing methods for effect decomposition require the absence of intermediate confounders affected by exposure.

Purpose of the Study:

• To develop a novel causal mediation analysis framework for population intervention effects using stochastic interventions.
• To define and evaluate (in)direct effects that are robust to exposure type and intermediate confounders.

Main Methods:

• Theoretical development of (in)direct effect decomposition for population intervention effects.
• Nonparametric efficiency theory for constructing multiply robust estimators.
• Development of inferential techniques for confidence intervals and hypothesis tests.

Main Results:

• Introduced novel (in)direct effects identifiable regardless of exposure type (categorical or continuous).
• Demonstrated that these effects remain well-defined even with intermediate confounders affected by exposure.
• Provided a theoretical basis for flexible, multiply robust estimation and valid inference.

Conclusions:

• The proposed causal mediation analysis framework overcomes key limitations of existing methods.
• The methodology allows for valid causal effect estimation and inference in more complex settings.
• The `medshift` R package facilitates the application of these advanced causal inference techniques.