In clinical psychology and applied research, it is common to encounter patient populations that, while sharing the same diagnosis, present very different symptom patterns, severity levels, and recovery trajectories. A group of patients with major depression, for example, may include individuals with predominantly somatic symptoms, others with marked anhedonia, others with prominent suicidal ideation, and others with a mixed anxious-depressive profile. Treating all these patients as a homogeneous group obscures clinically relevant differences that may have direct implications for prognosis and treatment selection in clinical research. Latent Class Analysis (LCA) is a statistical technique designed precisely to identify these subgroups or profiles within an apparently uniform population.
Unlike more classic clustering techniques such as cluster analysis, LCA is based on a formal probabilistic model. This means that it does not assign each individual to a group deterministically but instead estimates the probability of membership in each latent class, which allows quantification of the uncertainty in classification. Furthermore, as a model-based approach, it offers objective statistical criteria for determining how many classes exist in the data, something that in cluster analysis depends largely on subjective decisions by the researcher.
Foundations of latent class analysis
LCA starts from the premise that observed responses to a set of indicators (symptoms, behaviors, item responses) are generated by an unobserved, that is, latent, categorical variable. Each category of that latent variable represents a different class or profile. Within each class, the indicators are assumed to be independent of one another (local independence assumption), which implies that the observed correlations among the indicators are completely explained by membership in the different classes. If, after classifying participants, the correlations among indicators disappear within each class, the model fits the data well.
The parameters estimated by LCA are of two types. On one hand, the class membership probabilities, which tell us what proportion of the population belongs to each profile. On the other, the item-response probabilities conditional on class, which describe how the indicators behave within each class. For example, in an LCA with depressive symptom indicators, one class might be characterized by a high probability of insomnia and fatigue but a low probability of suicidal ideation, while another class might show high probabilities across all symptoms. It is these conditional probabilities that allow each profile to be interpreted and labeled.
When the indicators are continuous rather than categorical, the equivalent technique is called Latent Profile Analysis (LPA). The logic is identical, but instead of estimating response probabilities, means and variances of each indicator are estimated within each class. In clinical practice, it is common to work with continuous scores on questionnaires, which is why LPA tends to be the more widely used variant in applied psychological research. Ensuring the quality of those scores requires a rigorous questionnaire validation process.
How to determine the number of profiles
One of the most important decisions in an LCA is determining how many latent classes exist in the data. The standard strategy consists of estimating a series of models with an increasing number of classes (from one class, which is equivalent to assuming total homogeneity, up to as many as seems reasonable) and comparing their fit using several criteria. The most commonly used information criteria are the BIC (Bayesian Information Criterion) and the AIC (Akaike Information Criterion), where lower values indicate better fit while accounting for model parsimony. The BIC is usually preferred because it penalizes complexity more heavily and tends to select more parsimonious models.
In addition to information criteria, there are direct model comparison tests such as the Lo-Mendell-Rubin test (LMR) and the Bootstrap Likelihood Ratio Test (BLRT). These tests compare a model with k classes against one with k-1 classes, and a significant result indicates that the model with more classes fits significantly better. The BLRT is generally considered more reliable than the LMR, although it is computationally more expensive.
However, the decision about the number of classes should not be based solely on statistical criteria. The substantive interpretability of the classes is equally important. A model with five classes may fit slightly better than one with four, but if the fifth class contains only 3% of the sample and its profile does not make clinical sense, the more parsimonious model is probably preferable. The convergence between statistical criteria and clinical judgment is what produces the most useful solutions in practice.
Clinical and research applications
The applications of LCA in clinical psychology are numerous and clinically relevant. In the study of post-traumatic stress disorder, for example, several studies have identified patient profiles that differ not only in the overall severity of symptoms but in the specific configuration of symptom clusters (re-experiencing, avoidance, hyperarousal, cognitive and mood alterations). These profiles have been shown to differentially predict treatment response, which opens the door to more personalized interventions.
In the field of addictions, LCA has made it possible to identify subtypes of users who differ in their patterns of use, motivations for consumption, and associated risk factors. In health psychology, profiles of chronic patients have been identified based on their coping strategies, treatment adherence, and quality of life. And in educational psychology, latent profile analysis is used to identify groups of students with different patterns of motivation, academic engagement, and performance.
A particularly interesting application is the use of LCA in combination with auxiliary variables (covariates and distal outcomes). Once the profiles are identified, it is possible to examine what variables predict membership in each class (for example, whether age, gender, or family history predict belonging to the high-severity profile) and what consequences are associated with each profile (for example, whether a specific profile predicts poorer treatment response or higher risk of relapse). This approach turns LCA into something more than a descriptive technique, transforming it into a tool with predictive and potentially prescriptive value. A complementary approach for exploring these symptom relationships is network analysis in psychopathology.
Practical considerations
The most widely used software for LCA in psychological research is Mplus, which offers the greatest flexibility in model specification and the most advanced estimation methods. However, free alternatives such as the tidyLPA or poLCA packages in R allow latent profile analysis and latent class analysis to be performed with perfectly valid results for most applications. JASP has recently incorporated an LCA module that facilitates its use for researchers without programming experience.
The sample size needed for an LCA depends on several factors: the number of indicators, the expected number of classes, the relative size of the smallest classes, and the separation between classes. As a general guideline, samples below 200 participants tend to be problematic, and many methodologists recommend a minimum of 300-500 to obtain stable solutions. When a latent class is very small (less than 5% of the sample), even larger samples are needed to estimate it accurately.
Before you submit: the number of classes you kept is the first thing a methodological reviewer will question, and pointing at the lowest BIC rarely survives on its own. If your manuscript is already written, run it through the AI Paper Reviewer: a free Reviewer 2 style pre-review that tells you which of your class decisions they will push back on, while you can still add the justification.
A frequent error in the application of LCA is reifying the latent classes, that is, treating them as if they were natural and discrete categories that exist independently of the statistical model. Latent classes are model constructions that summarize patterns in the data, and their utility is fundamentally heuristic. Two researchers can analyze the same data and arrive at different solutions if they make different analytical decisions. This uncertainty does not invalidate the technique, but it does require interpreting it with the necessary caution and seeking replication in independent samples before drawing definitive conclusions about the profile structure of a clinical population.