Evaluating the Efficacy of Correlation-Based Meta-Analytic Structural Equation Modeling Under Different Patterns of Missing Correlations
Abstract
Missing correlations often happen in the primary correlation matrices and create major problems for the performance of meta-analytic structural equation modeling (MASEM). In current literature, methodological investigation regarding the performance of MASEM methods paired with a large proportion of missing data is limited. This simulation study was designed to investigate the impacts of missing conditions on the correlation-based MASEM performance, utilizing weighted-covariance generalized least squares (W-COV GLS) with pairwise deletion (PD), W-COV GLS with multiple imputation (MI), and two-stage structural equation modeling (TSSEM) to pool the correlation matrices. Specifically, impacts of the study number (k), the within study sample size (n), the proportion of missing correlations (pm), and the proportion of studies reported full matrices (pf) on the performance of MASEM results were explored, with two different factorial models. Conditions where the correlation-based MASEM could produce accurate parameter estimations were identified. The results showed that larger sample size (i.e., k and n) and greater proportion of full matrices (pf) improve model fits, reduce bias in parameter estimates and their standard errors, and decrease the type I error rate. Whereas, the missing proportion (pm) has an inverse function on model fits, parameter and standard error estimates, and type I error rates. W-COV GLS with MI is superior than the other two methods with most missing conditions, but W-COV GLS with PD tends to perform better than W-COV GLS with MI when the missing proportion is extremely small paired with a large proportion of full matrices under the model with unequal factor loadings. Moreover, the outcomes demonstrated the necessity of including at least one study with full correlation matrices in the study pool for TSSEM and W-COV GLS with PD. Suggestions for future research were also described.
Degree
Ph.D.
Advisors
Maeda, Purdue University.
Subject Area
Educational psychology
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.