Model-Based (Linked) Meta-Analysis

Description

Pool the correlation matrices among a model's variables across studies and fit regression or path models to the pooled matrix to estimate direct, partial and indirect effects.

Instead of pooling one bivariate effect, model-based meta-analysis starts from a theoretical model drawn as a path diagram and collects all correlations among its variables from each study. Unreported cells are marked missing, and the covariances among each study's correlations are estimated. Homogeneity is tested, between-study variance estimated, and a fixed or random-effects mean correlation matrix is computed by generalised least squares. Standardised regression or path models are then fitted to the pooled matrix, giving partial and indirect effects, model comparisons and moderator analyses by subgroup. A useful by-product is a map of links that no study has examined. Limitations are missing correlations and fractionated evidence: different cells of the matrix may rest on different samples, so random-effects models are preferred.

Attributes

Loss type
Unmigrated
Stage
Strategy
Effort
Weeks

Steps

  1. Specify the model as a path diagram, and use its constructs and instruments as search terms and inclusion rules.
  2. Code each study's correlation vector, marking unreported cells as missing.
  3. Estimate the covariances among each study's correlations.
  4. Test homogeneity, estimate between-study variance and compute a mean correlation matrix by GLS.
  5. Fit the regression or path model to the pooled matrix, reporting slopes, standard errors and tests.
  6. Run moderator analyses by subgroup and map which links lack evidence.

Sources

  • The Handbook of Research Synthesis and Meta-Analysis (Cooper et al., 2009)

Also known as

  • Meta-analytic structural equation modelling
  • MASEM
  • Synthesis of correlation matrices
  • Model-driven meta-analysis

Categories

  • Analytics & Behaviour
  • Statistics & Modelling