A Comparison of Approaches for Handling Within-Study Dependency in Correlations for Meta-Analytic Structural Equation Modeling

• In Event: AERA Roundtable Session Saturday 1:45 pm - Gold 4
  In Roundtable Session: RT3: Imputation Strategies and Synthesis/Meta-Analytic Designs with Mediation Effects or Complex Structures

Abstract

Meta-analytic structural equation modeling (MASEM) estimates an SEM using a correlation matrix with elements pooled across primary studies’ correlations. Correlations within a study’s correlation matrix are naturally correlated due to correlations among the matrix’s variables. There can be additional dependence in a study’s correlations when a study reports multiple estimates of one correlation. Most MASEM methods recognize natural not additional dependence in correlations. We evaluated and compared five methods for handling within-study dependence in correlations for MASEM: simple aggregation with two-stage (TSSEM), one-stage MASEM, multilevel multivariate pooling with constrained versus unconstrained random-effects variances, and use of robust variance estimation. Simple aggregation, and the unconstrained model with or without RVE worked well. Results are discussed in more detail in the paper.

Authors

• Sunyoung Park, California Lutheran University

• Tasha Beretvas, University of Texas at Austin