Latent Variable Interaction Modelling in SEM

Description

Tests whether the effect of one latent construct depends on another, or is curved, while keeping measurement error in the model rather than multiplying error-laden scale scores.

Multiplying two scale means to test moderation carries both scales' measurement error into the product and weakens the test. SEM approaches keep the constructs latent. In the Kenny-Judd product-indicator approach, products of the indicators of two latent variables become indicators of an interaction factor, with loadings and error variances fixed as functions of the original measurement parameters; Ping's two-step variant takes these values from a prior measurement model. Squares of indicators carry quadratic effects in the same way. Alternatives are a multiple-group model comparing slopes across moderator levels (simple, but it loses power and can create small groups), the product of latent variable scores with bootstrapped standard errors (recommended by Schumacker and Lomax), Bollen's two-stage least squares with instrumental variables, and full maximum likelihood or Bayesian MCMC estimators for nonlinear SEM (latent moderated structural equations), which avoid product indicators but need larger samples.

Attributes

Loss type
Siloed
Effort
Weeks

Steps

  1. Make each latent variable unidimensional, centre all indicators at zero, and fit a measurement model recording loadings and error variances.
  2. Choose an approach: product indicators (Kenny-Judd or Ping two-step), latent variable scores, multiple-group comparison of slopes, two-stage least squares, or a full ML or Bayesian nonlinear SEM estimator.
  3. For product indicators, form products of selected indicator pairs (not every pair is needed) and compute their loadings and error variances from the measurement estimates.
  4. Fix these values in a structural model containing the interaction (or quadratic) latent term and estimate it; re-run to check measurement parameters have not shifted.
  5. Compare the main-effects and interaction models with a chi-square difference test and the change in R-squared; use robust or bootstrap standard errors because product terms are non-normal.
  6. Probe a significant interaction by plotting the simple slope of the predictor at low, mean and high values of the moderator.

Sources

  • A Beginner's Guide to Structural Equation Modeling (Schumacker & Lomax, 2004)
  • Latent Variable Models (Loehlin & Beaujean, 2017)
  • The SAGE Handbook of Quantitative Methods in Psychology (Millsap & Maydeu-Olivares, 2009)
  • Structural Equation Modeling: A Bayesian Approach (Lee, 2007)
  • Applied Bayesian Modelling (Congdon, 2003)
  • Bayesian Statistical Modelling (Congdon, 2006)

Also known as

  • Kenny-Judd product indicator model
  • Latent moderation in SEM
  • Ping two-step latent interaction
  • Latent variable score interaction
  • Interaction effects in structural equation models
  • Nonlinear and interaction effects among latent variables (Kenny-Judd)
  • Nonlinear structural equation models
  • Latent moderated structural equations

Categories

  • Analytics & Behaviour
  • Statistics & Modelling