Sensitivity analysis for mistakenly adjusting for mediators in estimating total effect in observational studies

Area of Science:

• Epidemiology
• Causal Inference
• Biostatistics

Background:

• Estimating total effects of exposures on outcomes is crucial in observational studies.
• Misadjusting for mediators can introduce bias, especially with unknown causal diagrams.
• Understanding bias performance is essential for accurate effect estimation.

Purpose of the Study:

• To dissect bias performances when mistakenly adjusting for mediators in logistic regression.
• To evaluate how varying exposure-mediator and mediator-outcome effects influence bias.
• To compare bias under different mediator configurations and confounding scenarios.

Main Methods:

• Simulation studies based on six causal diagrams representing different mediator roles.
• Sensitivity analysis to assess bias from varying exposure-mediator and mediator-outcome effects.
• Logistic regression used to estimate effects, with bias defined as the difference from the true total effect.

Main Results:

• In scenarios without unobserved confounders, bias was more sensitive to exposure-mediator effects.
• In scenarios with unobserved confounders, bias was more sensitive to mediator-outcome effects.
• Specific mediator structures (series, parallel) influenced the magnitude and direction of bias.

Conclusions:

• Bias sensitivity to mediator effects differs based on the presence of unobserved confounders.
• Accurate estimation of total effects necessitates careful consideration of mediator roles and potential confounding.
• Findings highlight the importance of sensitivity analysis in observational research involving mediators.