A Potential Outcomes Approach to Selection Bias.

Area of Science:

• Epidemiology
• Causal Inference
• Biostatistics

Background:

• Selection bias is a major challenge in analytic epidemiology, potentially distorting causal effect estimates.
• Existing definitions and methods for addressing selection bias are often disparate and may not cover all mechanisms.

Purpose of the Study:

• To propose a novel, unified definition of selection bias in analytic epidemiology.
• To provide a nonparametric framework for analyzing selection bias using potential outcomes and single-world intervention graphs.
• To demonstrate the adaptability of this approach to both analytic and descriptive epidemiology, and various study designs.

Main Methods:

• Utilizing the potential outcomes framework to define selection bias.
• Employing single-world intervention graphs for bias analysis under and away from the null hypothesis.
• Illustrating the approach with examples from epidemiological studies, including matched and case-cohort designs.

Main Results:

• The proposed definition encompasses both structural and traditional concepts of selection bias.
• The nonparametric approach allows for simultaneous analysis of confounding and selection bias.
• The framework explicitly connects participant selection to causal effect estimation.

Conclusions:

• The novel definition offers a unified perspective on the diverse mechanisms generating selection bias.
• This approach simplifies the analysis of selection bias in complex epidemiological study designs.
• The potential outcomes framework provides a robust foundation for causal inference in the presence of selection bias.