Simulation-based sensitivity analysis for causal mediation studies.

Area of Science:

• Statistics
• Causal Inference
• Epidemiology

Background:

• Causal mediation analysis requires assumptions about the absence of unmeasured confounders.
• Violations of these assumptions, particularly unmeasured pretreatment confounders, can bias results.
• Sensitivity analysis is crucial for evaluating the robustness of causal mediation findings.

Purpose of the Study:

• To propose a simulation-based method for assessing sensitivity to unmeasured pretreatment confounding in causal mediation analysis.
• To provide a tool for quantifying the potential impact of unmeasured confounders on treatment-mediator and mediator-outcome relationships.
• To enhance the reliability of causal inference in mediation studies.

Main Methods:

• Developed a simulation-based approach to model unmeasured pretreatment confounders.
• Quantified confounder strength via conditional associations with treatment, mediator, and outcome.
• Assessed impact on point estimation and efficiency by comparing analyses with and without simulated confounders.
• Provided a visualization tool and an R package (mediationsens) for implementation.

Main Results:

• The method allows for quantifying the strength of potential unmeasured pretreatment confounders.
• It demonstrates how unmeasured confounders can influence point estimation and estimation efficiency.
• The approach is applicable across various causal mediation methods, study designs, and data types.

Conclusions:

• The proposed sensitivity analysis method enhances the rigor of causal mediation inference.
• It provides researchers with a practical tool to assess the impact of potential unmeasured confounding.
• This facilitates more reliable conclusions in both experimental and observational research.