Latent Class Analysis (LCA)
Latent Class Analysis · Also known as: LCA, latent class model, latent categorical analysis, finite mixture of multinomials
Latent class analysis identifies unobserved subgroups — latent classes — within a population by finding patterns of responses across a set of categorical observed indicators. It is the categorical-variable counterpart of cluster analysis, but grounded in an explicit probabilistic model, and is widely used in social, health, and behavioral sciences to discover typologies in survey or diagnostic data.
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.
+12 more
When to use it
Use LCA when you have a set of categorical observed indicators (binary yes/no, or polytomous nominal/ordinal items) and you hypothesize that the population contains qualitatively distinct subgroups whose differences explain the item associations. It is ideal for developing typologies in health behavior, psychopathology, educational assessment, or marketing segmentation when variables are categorical. Do not use LCA when your indicators are continuous — use latent profile analysis (LPA) instead. Avoid LCA when the sample is small (fewer than ~200 cases) because the EM algorithm becomes unreliable with many classes, and when the number of response patterns is sparse relative to the number of indicators. LCA also requires the local independence assumption; if items share specific variance beyond what the classes capture, the model will overextract classes.
Strengths & limitations
- Probabilistic model-based classification with explicit fit criteria, unlike heuristic distance-based clustering.
- Naturally handles binary and polytomous categorical indicators without requiring interval-level measurement.
- Yields soft assignments (posterior probabilities) that quantify classification uncertainty for each individual.
- Information criteria (BIC, AIC) and significance tests provide principled guidance on the number of classes.
- Classes can be interpreted substantively as latent typologies with distinct item-response profiles.
- Extends naturally to latent class regression, latent transition analysis, and multilevel LCA.
- The local independence assumption is strong and often violated; items within a class may still covary due to unmeasured item-specific factors.
- The EM algorithm can converge to local rather than global maxima, requiring many random starting values.
- Identifying the correct number of classes is challenging; different fit indices can disagree and there is no single gold standard.
- Large numbers of indicators with many response categories create sparse response patterns, making estimation unstable.
- Modal class assignment discards uncertainty and can lead to overstated classification accuracy in downstream analyses.
Frequently asked
What is the difference between LCA and cluster analysis?
Both aim to identify subgroups, but cluster analysis is an algorithmic procedure without an underlying probability model, making it difficult to assess fit or quantify uncertainty. LCA is a model-based approach that specifies an explicit probabilistic generative model, provides formal fit statistics and model-comparison tests, and yields posterior membership probabilities rather than hard assignments. LCA is preferred when interpretability, fit assessment, and uncertainty quantification matter.
What is the difference between LCA and latent profile analysis?
LCA is designed for categorical indicators (binary or polytomous). Latent profile analysis (LPA) uses the same latent mixture logic but models continuous indicators with within-class normal distributions. Applying LCA to continuous data or LPA to binary data is a model misspecification; always match the model to the measurement level of your indicators.
How do I choose the number of classes?
Fit models with K = 1, 2, 3, … classes and compare them using BIC (tends to favor parsimony), AIC (more liberal), the Bootstrap Likelihood Ratio Test (BLRT), and entropy. No single criterion is definitive; the preferred solution balances statistical fit with substantive interpretability of the class profiles. Classes that are very small or substantively redundant are signs of overextraction.
Can I use LCA with ordinal indicators?
Yes, though by default LCA treats each response category of a polytomous item separately (nominal). For ordinal indicators you can impose monotonicity constraints on item-response probabilities across classes, which reduces parameters and can improve stability. Alternatively, polychoric correlations combined with factor mixture models offer another approach for ordinal data.
How large a sample do I need for LCA?
There is no single threshold, but small samples (n < 200) make the EM algorithm prone to convergence problems and produce unstable class solutions. The required sample size grows with the number of classes and indicators. Simulation studies suggest n >= 300 as a practical minimum for models with three or more classes, and larger samples are needed when class prevalences are unequal or indicators are weakly discriminating.
Sources
- Goodman, L. A. (1974). Exploratory latent structure analysis using both identifiable and unidentifiable models. Biometrika, 61(2), 215–231. DOI: 10.1093/biomet/61.2.215 ↗
- Lazarsfeld, P. F. & Henry, N. W. (1968). Latent Structure Analysis. Houghton Mifflin. link ↗
How to cite this page
ScholarGate. (2026, June 3). Latent Class Analysis. ScholarGate. https://scholargate.app/en/statistics/latent-class-analysis
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.
- Cluster AnalysisStatistics↔ compare
- Confirmatory factor analysisPsychometrics↔ compare
- Discriminant AnalysisStatistics↔ compare
- EFAStatistics↔ compare
- Latent Profile AnalysisPsychometrics↔ compare
- Mixture ModelingStatistics↔ compare