Combining Multiple Observational Data Sources to Estimate Causal Effects.

Area of Science:

• Statistics
• Causal Inference
• Big Data Analytics

Background:

• Increasing availability of multiple data sources in the era of big data.
• Challenges in estimating causal effects with unmeasured confounders in large datasets.
• Need for methods combining big main data with smaller validation data containing supplementary confounder information.

Purpose of the Study:

• To develop a principled framework for consistent and efficient estimation of causal effects.
• To leverage information from both big main data and smaller validation data with supplementary confounder information.
• To improve upon error-prone estimators typically derived from big data alone.

Main Methods:

• Utilizing smaller validation data with supplementary information on unmeasured confounders.
• Combining information from big main data and validation data under the unconfoundedness assumption.
• Applying the framework to asymptotically normal estimators like regression imputation, weighting, and matching.

Main Results:

• Demonstrated that combining data sources can lead to consistent estimators for causal effects.
• Showcased improved estimation efficiencies by leveraging big main data information.
• Developed bootstrap procedures for straightforward implementation with existing software.

Conclusions:

• The proposed method allows for efficient and consistent estimation of causal effects even with unmeasured confounders.
• The framework does not require correct model specification for unmeasured confounders.
• The approach is practical and easily implementable for various statistical estimation techniques.