Selection Bias When Estimating Average Treatment Effects Using One-sample Instrumental Variable Analysis.

Area of Science:

• Epidemiology
• Biostatistics
• Genetics

Background:

• Epidemiologic and genetic studies often lack true random samples, introducing selection bias.
• The impact of selection bias on instrumental variable (IV) analysis is not fully understood by practitioners.

Purpose of the Study:

• To clarify how selection mechanisms bias instrumental variable (IV) analyses.
• To demonstrate the utility of directed acyclic graphs (DAGs) in assessing selection bias in IV analyses.

Main Methods:

• Utilized directed acyclic graphs (DAGs) to model selection mechanisms.
• Conducted simulations to evaluate bias under various conditions (e.g., instrument strength, exposure-instrument association).
• Applied methods to UK Biobank data to estimate the causal effect of education on smoking.

Main Results:

• Simulations confirmed that selection bias can lead to biased IV estimates and confidence interval undercoverage.
• Bias magnitude varied with instrument strength, linearity of exposure-instrument association, and causality of the effect.
• UK Biobank analysis showed substantial differences in causal effect estimates when adjusting for selection versus ignoring it.

Conclusions:

• Selection bias can significantly distort findings from instrumental variable (IV) analyses.
• Further research is needed on sensitivity analyses for unmeasured selection bias in IV studies.