Causal simulation experiments: Lessons from bias amplification.

Area of Science:

• Causal inference
• Statistical modeling
• Epidemiology

Background:

• Theoretical work identifies variables that amplify unmeasured confounding bias (bias amplification).
• Existing simulations suggest bias amplification may be less critical in practice than theory implies.
• A tension exists between theoretical predictions and simulation-based evidence regarding bias amplification.

Purpose of the Study:

• To resolve the tension between theoretical and simulation-based findings on bias amplification.
• To develop a general characterization of bias amplification using semi-parametric regression.
• To propose a new framework for causal simulation experiments and sensitivity analysis.

Main Methods:

• Utilized tools from the semi-parametric regression literature.
• Characterized bias amplification based on the geometry of Ordinary Least Squares (OLS) estimators.
• Extended analysis to a broader range of Directed Acyclic Graphs (DAGs), functional forms, and distributional assumptions.
• Developed a novel framework for causal simulation experiments.

Main Results:

• Provided a general characterization of bias amplification, reconciling theoretical and simulation findings.
• Identified limitations in current simulation approaches for bias amplification.
• Proposed and evaluated a new framework for causal simulation experiments.
• Demonstrated the application of the simulation approach to a real clinical dataset with binary treatment.

Conclusions:

• The semi-parametric regression approach offers a more comprehensive understanding of bias amplification.
• The proposed simulation framework facilitates principled sensitivity analysis for unmeasured confounding.
• This work bridges theoretical insights with practical simulation methods for causal inference in clinical research.