Robust inference on population indirect causal effects: the generalized front door criterion.

Area of Science:

• Causal Inference
• Epidemiology
• Biostatistics

Background:

• Standard causal inference methods for direct and indirect effects rely on strong no-unmeasured-confounding assumptions.
• These assumptions are frequently violated in observational studies, limiting reliable effect estimation.
• Unmeasured confounding is a significant challenge in determining the true impact of exposures and interventions.

Purpose of the Study:

• To introduce and define a new form of indirect effect: the population intervention indirect effect.
• To establish a method for non-parametric identification of this effect, even with unmeasured common causes of exposure and outcome.
• To generalize existing causal identification criteria, such as the front-door criterion.

Main Methods:

• Development of a new definition for population intervention indirect effect.
• Establishment of identification criteria that relax stringent no-unmeasured-confounding assumptions.
• Creation of parametric and semiparametric inference methods, including a doubly robust estimator.
• Validation through simulation studies.

Main Results:

• The population intervention indirect effect can be non-parametrically identified in the presence of unmeasured common causes.
• The identification criterion generalizes the front-door criterion by not requiring the absence of a direct effect.
• Parametric and semiparametric methods, including a novel doubly robust estimator, demonstrate strong performance in simulations.
• The proposed methods were successfully applied to assess the effectiveness of health program interventions.

Conclusions:

• The population intervention indirect effect provides a valuable tool for causal inference in complex observational settings.
• The developed methods offer robust and efficient estimation strategies for mediation analysis.
• This approach advances the ability to measure intervention effectiveness when unmeasured confounding is present.