Structural Equation Mixture Modelling (Factor Mixture, Growth Mixture and Regime-Switching Models)

Description

Assumes users are a mixture of unobserved subgroups, each with its own factor, growth or structural model, and estimates class sizes, class-specific models and, over time, switching between classes.

SEM assumes respondents come from one population. If unobserved subgroups exist whose constructs or trajectories behave differently, estimates become uninterpretable. When membership is known, multiple-group, multilevel or MIMIC models suffice; when it is not, SEM mixture modelling treats each latent class as multivariate normal with its own SEM-structured means and covariances, optionally predicting class membership from covariates. Factor mixture models let classes differ in factor means or loadings; growth mixture models let them follow different trajectories; regime-switching models add a Markov transition model so people can move between classes over repeated measures. Mixtures can be used directly (substantive subgroups) or indirectly (to approximate non-normality or non-linearity). Choosing the number of classes is hard because the ordinary likelihood ratio test is invalid at the boundary; information criteria, bootstrap tests and the Lo-Mendell-Rubin test are used, and switching models grow quickly in complexity (three regimes over four occasions imply 81 components).

Attributes

Loss type
Siloed
Effort
Weeks

Steps

  1. Specify the within-class model (factor, growth or structural) and which parameters may differ between classes.
  2. Fit models with one, two, three and more classes by maximum likelihood, using many random start values to avoid local maxima.
  3. Compare solutions with BIC and AIC, bootstrap likelihood ratio or Lo-Mendell-Rubin tests, and class sizes and interpretability.
  4. Optionally regress class membership on covariates such as acquisition channel or plan.
  5. For repeated measures, add a Markov transition model and estimate the probability of moving between classes.
  6. Interpret class proportions and class-specific parameters substantively, and check that classes are not artefacts of non-normal data.

Sources

  • The SAGE Handbook of Quantitative Methods in Psychology (Millsap & Maydeu-Olivares, 2009)
  • Structural Equation Modeling: A Bayesian Approach (Lee, 2007)
  • Finite Mixture Models (McLachlan & Peel, 2000)

Also known as

  • SEMM
  • Factor mixture model
  • Growth mixture model
  • Finite mixture SEM
  • Latent Markov regime switching
  • Structural equation mixture modelling

Categories

  • Analytics & Behaviour
  • Statistics & Modelling