Vermunt-Magidson (2005) Structural Equation Models: Mixture Models

Summary

This encyclopedia entry introduces mixture SEM (equivalently latent class SEM) — a hybrid that fuses structural equation modeling with finite-mixture / latent-class modeling, proposed independently by Arminger-Stein (1997), Dolan-Van der Maas (1997), and Jedidi-Jagpal-DeSarbo (1997). The core idea is to fit a multivariate-normal (MVN) mixture in which each latent class's mean vector and covariance matrix are not left unrestricted but are instead constrained by a postulated SEM structure — a one-factor, latent-growth, or autoregressive model. The article develops the MVN-mixture baseline, the SEM restrictions that make it parsimonious, the role of covariates, ML/MAP estimation via EM, and the available software, closing with a longitudinal reading-skill example.

Key Claims

• Two complementary readings of mixture SEM. (1) A refinement of MVN mixtures: model-based clustering where each class covariance $\Sigma_k$ is smoothed by an SEM structure, yielding more parsimonious and stable clusters than unrestricted covariances. (2) An extension of standard SEM akin to multiple-group analysis, but with unobserved group membership — so it detects unobserved heterogeneity (e.g., equal factor loadings but class-specific error variances, or class-specific patterns of change over time).
• MVN-mixture baseline. $f(y_i\mid\pi,\theta)=\sum_{k=1}^{K}\pi_k\, f_k(y_i\mid\mu_k,\Sigma_k)$ with class-specific multivariate-normal densities; the unrestricted model has $K\!\cdot\!P\!\cdot\!(P-1)/2$ covariance parameters, which grows quickly in the number of indicators $P$ and classes $K$.
• Covariance-reduction hierarchy. Parameter growth is tamed by (i) local independence (diagonal $\Sigma_k$), (ii) homogeneity across classes ($\Sigma_k=\Sigma$), or (iii) a postulated SEM structure (factor, latent-growth, autoregressive) — the distinctive mixture-SEM route.
• Covariates / concomitant variables. Class-membership probabilities $\pi_k(z_i)$ are modeled by a multinomial logit in covariates $z_i$; a more general form lets covariates have direct effects on the indicators (Arminger-Stein), relaxing the assumption that covariate influence passes entirely through the latent classes.
• Estimation by ML and MAP via EM. Maximum a posteriori estimation with mild priors (e.g., Dirichlet on class sizes) is preferred over plain ML because it prevents boundary/terminal solutions — zero variances or zero probabilities — by keeping estimates in the interior of the parameter space.
• Software and example. Mplus, Mx, and Latent GOLD estimate these models; the empirical illustration fits a latent-growth / autoregressive mixture to four-wave PIAT reading-recognition data, recovering classes with different growth and autocorrelation structures.

Concepts Introduced or Extended

• Mixture Model
• Latent Class Model
• Consumer Heterogeneity

Entities Mentioned

• Jeroen K. Vermunt
• Jay Magidson

Quotes

"Mixture SEM restricts the form of… latent class clustering, by subjecting the class-specific mean vectors and covariance matrices to a postulated SEM structure such as a one-factor, a latent-growth, or an autoregressive model."

"An important difference between this and standard multiple group analysis is that in mixture SEM group membership is not observed."

My Take

This is the natural companion to Magidson-Vermunt (2004): where that chapter organizes cluster/factor/regression as special cases of one LC engine, this entry does the same one level up — an unrestricted MVN mixture is the permissive extreme, local independence and homogeneity are cheap restrictions, and an SEM structure is the principled middle path that buys stability without collapsing all covariance information. The MAP-over-ML argument is the practically important bit: boundary solutions are the recurring failure mode of MVN-mixture clustering, and a weak prior is a cheaper fix than restarting the EM. The main caveat is the usual one for this literature — everything hinges on correctly specifying both the number of classes and the within-class SEM, and the piece leans on proprietary software output rather than identification or standard-error detail.