Confounder detection in linear mediation models: Performance of kernel-based tests of independence

Area of Science:

• Statistics
• Econometrics
• Psychometrics

Background:

• Mediation analysis requires strong unconfoundedness assumptions for valid direct and indirect effect identification.
• Testing for unconfoundedness is crucial, but often relies on instrumental variables (IVs), which can be difficult to find.

Purpose of the Study:

• To present a novel method for detecting confounders in mediation analysis without relying on instrumental variables (IVs).
• To demonstrate the utility of kernel-based independence tests for identifying confounding under nonnormality.

Main Methods:

• Utilized kernel-based tests of independence to detect confounding in the mediator-outcome relation.
• Conducted a simulation study to evaluate the performance of these tests in terms of Type I error and statistical power.
• Applied the method to real-world data from the Job Search Intervention Study (JOBS II).

Main Results:

• Kernel-based independence tests successfully detect confounding when variables are nonnormal.
• The simulation study indicated good Type I error protection and statistical power, irrespective of confounder distribution or measurement level.
• The real-world data example demonstrated the approach's effectiveness in minimizing biased indirect effect estimates.

Conclusions:

• Nonnormality allows for confounder detection in mediation analysis without instrumental variables (IVs).
• Kernel-based independence tests offer a practical diagnostic tool for ensuring unconfoundedness in mediation models.
• The proposed method enhances the reliability of indirect effect estimates in various research fields.