Within Study Dependence in Meta-Analysis

Multivariate meta-analysis methods typically assume the dependence of effect sizes. One type of experimental-design study that generates dependent effect sizes is the multiple-endpoint study. While the generalized least squares (GLS) approach requires the sample covariance between outcomes within studies to deal with the dependence of the effect sizes, the univariate three-level approach does not require the sample covariance to analyze such multivariate effect-size data. Considering that it is rare that primary studies report the sample covariance, if the two approaches produce the same estimates and corresponding standard errors, the univariate three-level model approach could be an alternative to the GLS approach. The main purpose of this dissertation was to compare these two approaches under the random-effects model for synthesizing standardized mean differences in multiple-endpoints experimental designs using a simulation study. Two data sets were generated under the random-effects model: one set with two outcomes and the other set with five outcomes. The simulation study in this dissertation found that the univariate three-level model yielded the appropriate parameter estimates and their standard errors corresponding to those in the multivariate meta-analysis using the GLS approach.