G-computation demonstration in causal mediation analysis

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Causal mediation analysis has advanced, but a unified approach for effect estimation across different decomposition scenarios is lacking for epidemiologic research.
• Existing methods for estimating controlled direct effects and natural direct and indirect effects are varied, necessitating a more integrated framework.
• G-computation has shown promise for unifying effect estimation in total effect and joint controlled direct effect settings.

Purpose of the Study:

• To demonstrate the utility of parametric g-computation for estimating various components of total effect in causal mediation analysis.
• To apply parametric g-computation to estimate natural direct and indirect effects, standard and stochastic controlled direct effects, and reference and mediated interaction effects.
• To provide an intuitive framework for effect estimation under 3- and 4-way effect decomposition using nested potential outcomes.

Main Methods:

• Utilized parametric g-computation with Monte Carlo simulations in standard statistical software.
• Estimated nested potential outcomes for each subject under different mediator scenarios (counterfactual exposure, pre-specified distribution, fixed controlled value).
• Employed a final regression of potential outcome on exposure intervention for point estimates and bootstrapping for confidence intervals.

Main Results:

• Parametric g-computation successfully estimated natural direct and indirect effects, standard and stochastic controlled direct effects, and reference and mediated interaction effects.
• The framework provided an intuitive method for estimating effects under 3- and 4-way effect decomposition by contrasting potential outcomes.
• Demonstrated the flexibility of the g-computation approach in handling various mediation effect components.

Conclusions:

• Parametric g-computation offers a unified and intuitive framework for estimating diverse causal mediation effects in epidemiologic research.
• This approach facilitates the estimation of complex mediation components, including interaction effects, under different decomposition scenarios.
• The g-computation framework is extensible to more complex multivariable and longitudinal mediation analyses.