Imputation of systematically missing predictors in an individual participant data meta-analysis: a generalized...

Area of Science:

• Biostatistics
• Epidemiology
• Medical Informatics

Background:

• Individual participant data meta-analyses (IPD-MA) are crucial for developing and validating risk prediction models.
• Systematically missing data in IPD-MA hinders the evaluation of between-study heterogeneity and accurate model development.
• Existing methods struggle to handle missing predictors or outcomes across studies.

Purpose of the Study:

• To introduce a novel multilevel multiple imputation (MLMI) approach for handling systematically missing data in IPD-MA.
• To extend existing imputation methods by incorporating between-study heterogeneity.
• To improve the development and validation of diagnostic and prognostic prediction models using IPD-MA.

Main Methods:

• Developed a novel imputation approach using generalized linear mixed models to account for between-study heterogeneity.
• Extended Resche-Rigon's method to relax assumptions on variance components and impute linear/nonlinear predictors.
• Compared MLMI with complete case analysis, traditional multiple imputation, and stratified multiple imputation in a deep venous thrombosis diagnostic model case study.

Main Results:

• MLMI fully accounts for between-study heterogeneity, unlike other methods.
• The approach successfully imputes systematically missing predictors in IPD-MA.
• Case study demonstrated MLMI's potential to improve estimation of heterogeneity parameters.

Conclusions:

• MLMI offers a robust solution for systematically missing data in IPD-MA.
• This method enhances the reliability of risk prediction models developed from IPD-MA.
• MLMI facilitates more comprehensive utilization of data in large-scale meta-analyses.