A causal data fusion method for the general exposure and outcome

Area of Science:

• Biostatistics
• Causal Inference
• Data Science

Background:

• Big data necessitates combining multiple datasets for robust causal effect estimation in medical and biological research.
• Individual datasets often lack sufficient confounders for accurate causal inference.
• Existing methods may not adequately address the fusion of more than two datasets or handle diverse exposure/outcome types.

Purpose of the Study:

• To extend causal data fusion methods for combining more than two datasets without external validation.
• To accommodate general (continuous or discrete) exposure and outcome variables in causal data fusion.
• To establish theoretical conditions for the identifiability of exposure effects from multiple data sources.

Main Methods:

• Developed a causal data fusion framework for integrating multiple individual data sources.
• Derived theoretical conditions for the identifiability of causal effects with continuous or discrete exposures and outcomes.
• Validated the method through simulations comparing it with regression, meta-analysis, and statistical matching.

Main Results:

• The proposed causal data fusion method yields unbiased causal effect estimates.
• The method demonstrates higher precision compared to traditional regression, meta-analysis, and statistical matching.
• The approach was successfully applied to investigate the causal effect of BMI on glucose levels in diabetic individuals by merging two datasets.

Conclusions:

• Causal data fusion is essential for leveraging multiple datasets in medical and biological research.
• The developed method provides a powerful tool for causal inference when merging diverse data sources.
• This work offers significant insights into the empirical analysis of combining multiple individual data sources.