Causal meta-analysis by integrating multiple observational studies with multivariate outcomes.

Area of Science:

• Biostatistics
• Epidemiology
• Observational Study Design

Background:

• Integrating multiple retrospective cohorts for causal inference is challenging due to unrepresentative samples and covariate imbalance.
• Existing methods struggle with meta-analysis of multiple groups from diverse retrospective cohorts.

Purpose of the Study:

• To propose a general covariate-balancing framework for meta-analysis of multiple retrospective cohorts.
• To introduce the FLEXible, Optimized, and Realistic (FLEXOR) weighting method to maximize effective sample sizes.
• To develop weighted estimators for unconfounded population-level inferences across various outcome types.

Main Methods:

• Developed a covariate-balancing framework using pseudo-populations to extend existing weighting methods.
• Proposed the FLEXOR weighting method to optimize effective sample sizes in integrative analyses.
• Derived new weighted estimators for quantitative, categorical, and multivariate outcomes.

Main Results:

• Demonstrated the versatility and reliability of the proposed weighting strategy through simulations.
• Validated the method with meta-analyses of The Cancer Genome Atlas (TCGA) datasets.
• Showcased the effectiveness of the FLEXOR pseudo-population approach for integrative analyses.

Conclusions:

• The proposed framework and FLEXOR method provide a robust approach for unconfounded causal and descriptive comparisons using multiple retrospective cohorts.
• This strategy enhances the reliability of meta-analyses, particularly when dealing with complex observational data.
• The methods are applicable to a wide range of population-level features and estimands.