A Monte Carlo Comparison of Handling Missing Covariates in Structural Equation Modeling–Based Meta-Analysis

• In Event: Advanced Methods in Meta-Analysis: Handling Missing Covariates and Dependence

Abstract

Meta-analysis often suffers from missing data on the study-level covariates, each representing the characteristics of the research or study sample. Researchers may choose a method for handling missingness; nevertheless, the choice is usually innocuous and driven by convenience (e.g., software default) increasing the risk for biased inferences. This may be because there are few guidelines to help researchers understand, use, and evaluate available missing data techniques for meta-analysis.
This study conducted a Monte Carlo simulation to compare available missing data methods for SEM-based meta-analysis. Complete-case analysis performed the best; and the performance of FIML and Bayesian modeling was comparable to complete-case analysis in some conditions. MI demonstrated poor performance. This study offers methodological guidelines and suggestions for applied researchers.

Authors

• Ye Sil Kim, University of South Carolina

• Jaehoon Lee, Texas Tech University