Sharpening randomization-based causal inference for 2² factorial designs with binary outcomes.

Area of Science:

• Biostatistics
• Clinical Trials
• Causal Inference

Background:

• Randomized controlled 2² factorial designs are common in medical research for binary outcomes.
• Existing causal inference frameworks, like Dasgupta et al.'s, face challenges with unidentifiable sampling variance.
• This leads to over-estimation in traditional Neymanian variance estimators.

Purpose of the Study:

• To address the variance estimation issue in 2² factorial designs.
• To derive the sharp lower bound of sampling variance for factorial effect estimators.
• To develop a new variance estimator that improves finite-population Neymanian causal inference.

Main Methods:

• Utilized potential outcomes framework for causal inference.
• Derived the sharp lower bound for sampling variance in 2² factorial designs with binary outcomes.
• Developed and validated a new variance estimator.

Main Results:

• The proposed variance estimator sharpens finite-population Neymanian causal inference.
• Simulation studies demonstrated the advantages of the new estimator.
• Application to real-life clinical trial datasets provided new insights.

Conclusions:

• The new variance estimator offers a significant improvement over existing methods for 2² factorial designs.
• This methodology enhances the precision of causal effect estimation in medical research.
• The findings have practical implications for analyzing randomized clinical trials.