Comparing methods of handling missing covariates in meta-analysis: Complete cases analysis, multiple imputation, full information maximum likelihood, and Bayesian analysis

Missing data on study-level covariates are inevitable in meta-analysis due to the different study characteristics in primary studies. Many meta-analysts have chosen the missing handling methods including complete cases analysis without considering underlying assumptions required for unbiased parameter estimate.
In this dissertation project, the performance of missing data handling methodologies (Complete cases analysis, Multiple imputation, Full Information Maximum Likelihood, and Bayesian analysis) was evaluated through simulation study under various conditions. As a result, complete cases analysis showed the best performance among the four methods with some limitation of sample size. FIML and Bayesian analysis provided results supporting the evidence that these methods can be used as an alternative to complete cases analysis under specific conditions. Multiple Imputation generally showed poor performance with large amounts of bias in the mean of covariate estimate and in the mean of covariate standard error estimate. Based on the results of findings in this dissertation, practical guidelines for applied researchers are discussed.