Mediation analysis with multiple versions of the mediator

Area of Science:

• Causal inference
• Biostatistics
• Epidemiology

Background:

• Counterfactual definitions of direct and indirect effects are common in causal inference.
• These definitions often assume well-defined hypothetical interventions on the mediator.
• Multiple versions of a mediator can complicate the interpretation of these effects.

Purpose of the Study:

• To examine mediation analysis when multiple versions of a mediator are present.
• To investigate the decomposition of total effects into direct and indirect components under such conditions.
• To clarify the interpretation of estimated direct and indirect effects when only one measurement of the mediator is available.

Main Methods:

• Utilizing counterfactual frameworks for causal inference.
• Analyzing mediation with multiple, unobserved versions of a mediator.
• Comparing standard mediation estimators to the true causal effects.

Main Results:

• The natural indirect effect, estimated with available data, retains a valid interpretation as a specific type of mediated effect.
• The natural direct effect estimator incorrectly incorporates an effect mediated through unmeasured mediator versions.
• This misinterpretation occurs when the specific version of the mediator is unknown.

Conclusions:

• Standard mediation analysis requires careful consideration when multiple mediator versions exist.
• The interpretation of natural direct effects is compromised without information on specific mediator versions.
• Future research should account for multiple mediator versions in causal effect decomposition.