Sensitivity analyses for parametric causal mediation effect estimation.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Causal mediation analysis estimates direct and indirect effects of exposures on outcomes via mediators.
• Sequential ignorability is a key assumption for causal interpretation but is not empirically testable.
• Existing sensitivity analyses are limited to linear models or complex nonparametric/semiparametric approaches.

Purpose of the Study:

• To develop novel sensitivity analysis methods for causal mediation analysis applicable to generalized linear models.
• To address limitations of current sensitivity analyses by accommodating unobserved confounders between mediator and outcome.

Main Methods:

• Proposed two alternative sensitivity analysis approaches for causal mediation analysis.
• Approach 1: Gaussian copula model with latent mediator and outcome variables.
• Approach 2: Hybrid causal-observational model extending association models with a novel sensitivity parameter.

Main Results:

• The proposed methods allow for sensitivity analyses under generalized linear models, expanding upon previous work.
• The models accommodate unobserved mediator-outcome confounders not affected by exposure, under a randomized exposure assumption.
• Applied the methods to analyze the effect of maternal education on adolescent dental caries.

Conclusions:

• The developed methods provide flexible and robust tools for sensitivity analysis in causal mediation.
• These approaches enhance the reliability of causal effect estimation when unobserved confounding is a concern.
• The study demonstrates practical application in epidemiological research, specifically concerning educational attainment and health outcomes.