SEM Specification Search and Model Modification

Description

A disciplined, theory-constrained search for a better-specified structural equation model, freeing or fixing one parameter at a time using diagnostics, then cross-validating the result.

When an SEM fits poorly the researcher may alter it, but every change risks capitalising on chance. The diagnostic tools are: significance of free parameters (to drop paths, cautiously, since significance depends on n); standardised residuals above about 1.96 or 2.58 flagging poorly reproduced covariances; modification indices (the minimum chi-square drop if a fixed parameter is freed) and expected parameter change; Lagrange multiplier tests as a multivariate forward search and Wald tests as a backward search. First localise misfit: fit a CFA with all latent variables freely correlated, and if it fits nearly as badly the problem is measurement. Measurement remedies include dropping weak indicators only if construct meaning survives, method factors or correlated residuals, or a composite single indicator. Simulation studies show automated searches rarely recover the true model, and fail entirely when the start is more than two or three parameters away, so the final model must be cross-validated.

Attributes

Loss type
Siloed
Effort
Weeks

Steps

  1. Let theory and prior research define the initial model, and list in advance which modifications would be substantively defensible.
  2. Fit the model; then fit a CFA with all latent variables freely correlated and compare chi-squares to see whether misfit lies in measurement or structure.
  3. Repair the measurement model first: inspect loadings, residuals and modification indices; consider method factors or correlated residuals only with a substantive reason.
  4. In the structural model, use Wald tests or non-significant paths to consider dropping parameters, and modification indices, expected parameter change and Lagrange multiplier tests to consider freeing them.
  5. Change one parameter at a time, rechecking several fit indices, parameter signs and the standardised residual matrix after each change.
  6. Report the initial and final models and every change with its justification and chi-square difference.
  7. Cross-validate the final model on a hold-out or new sample (or report ECVI) before drawing conclusions.

Sources

  • A Beginner's Guide to Structural Equation Modeling (Schumacker & Lomax, 2004)
  • Latent Variable Models (Loehlin & Beaujean, 2017)
  • Structural Equation Modeling in User Experience Research: Two Case Studies (Balaji et al., 2015)

Also known as

  • Model respecification
  • Modification index search
  • Lagrange multiplier and Wald test search
  • Forward and backward specification search
  • Exploratory modification of a misfitting latent variable model

Categories

  • Analytics & Behaviour
  • Statistics & Modelling