Methods for meta-analysis of individual participant data from Mendelian randomisation studies with binary outcomes

Area of Science:

• Epidemiology
• Genetics
• Biostatistics

Background:

• Mendelian randomization uses genetic variants as instrumental variables for causal inference in observational studies.
• Large sample sizes are crucial due to genetic variants explaining small proportions of risk factor variation.
• Meta-analysis of individual patient data from heterogeneous sources presents significant synthesis challenges.

Purpose of the Study:

• To develop and illustrate a Bayesian hierarchical framework for combining individual patient data from multiple Mendelian randomization studies.
• To efficiently synthesize heterogeneous data for robust causal effect estimation.
• To apply the framework to estimate the causal association between C-reactive protein and coronary heart disease.

Main Methods:

• A Bayesian hierarchical model was proposed to estimate causal effects from individual studies and combine them.
• The framework accommodates variations in genetic variants, study designs (prospective/retrospective, population-based/case-control), C-reactive protein measurement (timing, presence), and data sharing formats.
• The method was applied to data from the C-reactive protein Coronary Heart Disease Genetics Collaboration (CCGC).

Main Results:

• The Bayesian framework successfully combined heterogeneous data from multiple studies.
• Compared to traditional two-stage analysis, the Bayesian method incorporated 23% more participants and 51% more events.
• This integration resulted in a 23-26% increase in the efficiency of the causal association estimate.

Conclusions:

• The proposed Bayesian framework offers an efficient and robust method for synthesizing diverse individual patient data in Mendelian randomization studies.
• This approach enhances statistical power and precision for estimating causal relationships, as demonstrated for C-reactive protein and coronary heart disease.
• The methodology facilitates the integration of multi-source, heterogeneous data, advancing causal inference in epidemiology.