Growth Mixture Model (GMM)
Growth Mixture Model · Also known as: Büyüme Karışım Modeli (Growth Mixture Model — GMM), GMM, latent class growth analysis extension, mixture latent growth curve model
The Growth Mixture Model, introduced by Muthén and Shedden in 1999, is a longitudinal latent variable method that identifies distinct subpopulations — latent trajectory classes — each following its own growth curve over time. It extends the standard Latent Growth Curve (LGC) model by allowing the sample to be composed of an unknown mixture of classes with different intercepts, slopes, and variance structures.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
GMM is appropriate when you have longitudinal or panel data with at least three — and preferably four or more — repeated measurements per individual, and you suspect that the population is not homogeneous with respect to how the outcome changes over time. The method requires a reasonably large sample (at least 200 overall, and at least 25–50 observations per expected class) because small classes produce unreliable parameter estimates. The outcome variable should be continuous and approximately normally distributed within each latent class. If you have strong theoretical reasons to believe that within-class variation is negligible, LCGA is a simpler alternative; if you know the class structure in advance, a standard multi-group LGC model may be more appropriate.
Strengths & limitations
- Captures unobserved population heterogeneity in developmental or longitudinal processes that a single growth curve would mask.
- Allows within-class individual variation through random effects, making the model more realistic than LCGA.
- Each latent class yields its own interpretable growth trajectory, which can then be linked to predictors or distal outcomes.
- Robust to some degree of non-normality in the overall sample because the mixture structure absorbs part of the distributional complexity.
- Requires large samples; with fewer than 200 observations, class solutions are often unstable and parameter estimates are unreliable.
- The number of latent classes is not uniquely determined by the data; different information criteria can favour different solutions, and the choice involves subjective judgement.
- Local maxima in the likelihood surface are common; multiple sets of random starting values should always be used to check that the best-fitting solution is not a local optimum.
- Latent classes are statistical constructs — whether they correspond to real, meaningful subgroups requires substantive validation beyond model fit.
Frequently asked
What is the difference between GMM and LCGA?
Both methods identify latent trajectory classes in longitudinal data, but they differ in how much within-class variation they permit. Latent Class Growth Analysis (LCGA) constrains within-class variance to zero, meaning all individuals in a class are assumed to follow exactly the same trajectory. GMM relaxes this constraint by including random effects, so individuals within a class can vary around their class mean trajectory. GMM is generally more realistic; LCGA is simpler and may be used when samples are too small for GMM or as a preliminary step.
How do I decide how many latent classes to retain?
Fit models with increasing numbers of classes (typically 1 through 5 or 6) and compare them on multiple criteria: the Bayesian Information Criterion (BIC, lower is better), entropy (closer to 1 indicates cleaner class separation), and the Lo–Mendell–Rubin adjusted likelihood-ratio test. A good solution balances statistical fit with substantive interpretability — each class should be large enough to be stable and meaningful. Classes containing fewer than about 5% of the sample deserve scrutiny.
How many time points and participants do I need?
At a minimum, three repeated measurements are required to identify a linear growth trajectory, but four or more time points are strongly preferred because they allow the shape of each class trajectory to be estimated more reliably and permit testing of non-linear growth. For sample size, a commonly cited guideline is at least 200 participants overall, with at least 25–50 observations expected in each class. Smaller samples tend to produce unstable solutions and convergence problems.
Can I add covariates or predict class membership?
Yes. GMM can be extended in two directions. Covariates (predictors) measured at baseline can be used to predict class membership via a multinomial logistic regression component embedded in the model. Distal outcomes measured after the observation window can be regressed on class membership to ask whether trajectory class predicts a later variable. Both extensions should be handled carefully to avoid biases from ignoring classification uncertainty; the three-step BCH procedure is a recommended approach.
Sources
- Muthén, B. O. & Shedden, K. (1999). Finite Mixture Modeling with Mixture Outcomes Using the EM Algorithm. Biometrics, 55(2), 463–469. DOI: 10.1111/j.0006-341x.1999.00463.x ↗
How to cite this page
ScholarGate. (2026, June 1). Growth Mixture Model. ScholarGate. https://scholargate.app/en/statistics/growth-mixture-model
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- EFAStatistics↔ compare
- Hierarchical Linear ModelingStatistics↔ compare
- LCAStatistics↔ compare
- Multiple ImputationStatistics↔ compare
- SEMStatistics↔ compare