Multiple imputation for IPD meta-analysis: allowing for heterogeneity and studies with missing covariates

Area of Science:

• Statistics
• Biostatistics
• Epidemiology

Background:

• Multiple imputation is used for handling missing data in meta-analysis.
• Within-study imputation is often preferred but cannot address variables missing in entire studies.
• Sharing information across studies is beneficial for small contributing studies.

Purpose of the Study:

• To develop and evaluate a joint modeling approach for multiple imputation in individual patient data meta-analysis.
• To address limitations of within-study imputation, particularly for wholly missing variables.
• To incorporate between-study heterogeneity while allowing information sharing on covariance matrices.

Main Methods:

• A joint modeling approach with an across-study probability distribution for study-specific covariance matrices.
• Implementation of multiple imputation for individual patient data meta-analysis.
• Evaluation through simulation studies and illustration with a hypertension trials meta-analysis.

Main Results:

• The joint modeling approach performs comparably to within-study imputation where the latter is valid.
• The method shows good performance in scenarios with varying study sizes, significant heterogeneity, and wholly missing variables.
• Demonstrated effectiveness in a real-world meta-analysis of hypertension trials.

Conclusions:

• The proposed joint modeling approach offers a flexible and robust method for multiple imputation in individual patient data meta-analysis.
• It effectively handles missing data, including variables absent in entire studies, and accommodates between-study heterogeneity.
• This approach enhances the reliability and applicability of meta-analyses with complex missing data patterns.