Implications of M bias in epidemiologic studies: a simulation study.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Collider-stratification bias occurs when conditioning on a collider, opening an exposure-outcome path.
• M bias is a specific form transmitted through ancestors of exposure and outcome.
• Prior theory suggested M bias is smaller than confounding bias, but empirical data was lacking.

Purpose of the Study:

• To empirically assess the magnitude of M bias in realistic cohort study scenarios.
• To investigate conditions under which M bias becomes substantial.

Main Methods:

• Simulated data for large cohort studies with binary exposure, outcome, collider, and 2 predictors.
• Created 178 scenarios varying variable frequencies and associations.
• Calculated effect estimates, percentage bias, and mean squared error.

Main Results:

• M bias ranged from -2% to -5% in realistic scenarios.
• Substantial negative bias (>15%) occurred when collider-unmeasured factor relative risks were ≥8.
• Bias exceeded 20% if an unmeasured confounder/collider was not adjusted for M bias.

Conclusions:

• M bias generally has a small impact in typical cohort studies.
• Large associations (relative risk > 8) between colliders and unmeasured confounders amplify M bias.
• Controlling for confounding is prioritized over avoiding M bias when a collider is also a confounder.