Bayesian Structural Equation Modelling

Description

Fit structural equation and factor models by sampling parameters and each respondent's latent scores from their joint posterior, coping with small samples and categorical items.

Covariance-structure SEM (ML, GLS, ADF) relies on large samples and on the covariance matrix carrying all the information, which fails for small samples and for ordinal, nonlinear, multilevel, mixture or incomplete data. Bayesian SEM works with raw observations. Latent scores, and where needed underlying continuous responses and missing values, are treated as unknowns added to the data. Given them, the model becomes a set of regressions, so a Gibbs sampler alternates between drawing latent scores and drawing parameters, with Metropolis-Hastings steps where conditionals are non-standard. Conjugate priors (normal for loadings, inverted gamma for unique variances, inverted Wishart for factor covariances) also prevent Heywood cases. Lee's simulation found accurate estimates with samples of two to three times the number of parameters. Convergence is checked across chains, priors are perturbed to test sensitivity, and models are compared by Bayes factor or DIC.

Attributes

Loss type
Siloed
Effort
Weeks

Steps

  1. Specify the measurement and structural equations with identification constraints.
  2. Choose conjugate priors and set hyperparameters from expert knowledge, earlier studies, or an auxiliary analysis of part of the sample.
  3. Augment the data with latent scores and any underlying ordinal or missing values.
  4. Run the Gibbs sampler from several dispersed starting values and confirm convergence (EPSR below 1.2 for every parameter) before discarding burn-in.
  5. Estimate parameters and latent scores by posterior means, using posterior standard deviations as standard errors.
  6. Perturb the hyperparameters to check that conclusions do not depend on the prior.
  7. Compare models by Bayes factor or DIC and check fit with posterior predictive p-values and residual plots.

Sources

  • Structural Equation Modeling: A Bayesian Approach (Lee, 2007)
  • Applied Bayesian Modelling (Congdon, 2003)
  • Bayesian Statistical Modelling (Congdon, 2006)

Also known as

  • Bayesian SEM
  • BSEM
  • Bayesian confirmatory factor analysis
  • MCMC estimation of SEM

Categories

  • Analytics & Behaviour
  • Statistics & Modelling