psychology abstracts naming latent profile analysis, 2021–22 vs. 2025–26
29 → 70
in clinical psychology
9 → 27

Most analyses describe relationships among variables: does motivation predict achievement? Latent profile analysis (LPA) and latent class analysis (LCA) ask a different, person-centered question: are there subgroups of people who share a pattern, such as students who are highly motivated on every dimension, and others who feel competent but disconnected? In our analysis of recent psychology abstracts, the number naming latent profile analysis rose from 29 in 2021–22 to 70 in 2025–26, and in clinical psychology from 9 to 27. The numbers are still small, and the rise isn't large enough to rule out chance, but these models are appearing in more proposals.

This guide covers the difference between LPA and LCA, the steps from choosing indicators to relating profiles to outcomes, how to decide the number of profiles, example syntax, how to report the analysis in APA 7, and the questions committees ask.

LPA, LCA, and cluster analysis

LPA and LCA are both finite mixture models: they assume the sample is a mix of unobserved subgroups, each with its own pattern of means (LPA, for continuous indicators such as scale scores) or response probabilities (LCA, for categorical indicators such as yes/no items). Unlike k-means cluster analysis, they are statistical models, so each person gets a probability of belonging to each profile, and you can compare solutions with fit statistics and test the number of profiles.

A mixture model finds patterns that fit the data, but it can't prove that distinct types exist. A skewed or nonnormal continuous distribution can produce “profiles” even when everyone lies on a single continuum (Bauer & Curran, 2003). Treat profiles as a useful summary of heterogeneity, not proof of natural categories.

Step by step

  1. Choose indicators from theory. Use a small set of variables that define the profiles you are interested in, usually four to ten. Keep outcomes and predictors out of the indicator set; you will relate them to the profiles later.
  2. Prepare the indicators. Put them on a common scale (z-scores make profiles easier to read), check distributions, and use full-information estimation for missing data.
  3. Specify the model. The common starting point for LPA holds indicator variances equal across profiles and covariances at zero within profiles (conditional independence). Relaxing these can change the number of profiles, so say which specification you used and, if you can, compare alternatives (Masyn, 2013).
  4. Fit one profile, then two, three, and so on, stopping when adding profiles no longer improves the model or produces tiny, uninterpretable profiles. Use many random starts so the software finds the best solution rather than a local one, and check that the best log-likelihood is replicated.
  5. Choose the number of profiles using several criteria, described in the next section.
  6. Check classification quality with entropy and the average posterior probability of each profile.
  7. Describe and name the profiles. Plot the means, describe each profile by its pattern, and choose names that describe rather than overinterpret.
  8. Relate profiles to predictors and outcomes with a method that accounts for classification uncertainty.

Choosing the number of profiles

No single statistic decides. Report a table with every solution you fit and weigh these together:

  • BIC and sample-size adjusted BIC. Lower is better. In Nylund, Asparouhov, and Muthén's (2007) simulations, BIC was the best-performing information criterion. If BIC keeps falling, look for the point where the improvement flattens (an elbow).
  • The bootstrap likelihood ratio test (BLRT), which compares k profiles with k − 1. It performed best of the tests in the same simulations. The Lo–Mendell–Rubin test is a common alternative.
  • Profile sizes. Very small profiles, often below about 5% of the sample, may be unstable or reflect a few unusual cases.
  • Interpretability. Each profile should be distinct and meaningful. A new profile that is only a slightly higher or lower copy of an existing one adds little.

Don't use entropy to choose the number of profiles. Entropy describes how cleanly people are classified, with values of about .80 or higher usually read as good, not whether the model is right. Report it with the average posterior probabilities for each profile, which should be high, commonly above about .80 (Nagin, 2005).

Sample size. There is no simple rule. Simulations show that the number of profiles is harder to recover when the sample is small, profiles are small, or profiles are similar, and several hundred cases are often needed (Nylund et al., 2007; Tein, Coxe, & Cham, 2013).

Relating profiles to other variables

The tempting shortcut is to assign each person to their most likely profile and run an ANOVA or chi-square test. That treats the assignments as certain when they aren't, and it biases the comparisons. Use a three-step method that carries the classification error forward (Vermunt, 2010; Asparouhov & Muthén, 2014), such as the BCH method for outcomes or the R3STEP approach for predictors of profile membership in Mplus.

Running it

LPA in R with tidyLPA, which fits models through the mclust package:

library(tidyLPA); library(dplyr)
fits <- dat |>
  select(autonomy, competence, relatedness, amotivation) |>
  scale() |> as.data.frame() |>
  estimate_profiles(1:6, variances = "equal", covariances = "zero")
compare_solutions(fits, statistics = c("AIC", "BIC", "SABIC", "Entropy", "BLRT_p"))
get_estimates(fits[[4]])   # profile means for the 4-profile model

The same model in Mplus, where TECH11 gives the Lo–Mendell–Rubin test and TECH14 the BLRT:

VARIABLE: NAMES = id aut comp rel amot;
          USEVARIABLES = aut comp rel amot;
          IDVARIABLE = id;
          CLASSES = c(4);
ANALYSIS: TYPE = MIXTURE;
          STARTS = 500 100;
OUTPUT:   TECH11 TECH14;

For LCA with categorical indicators, the poLCA package in R, Mplus with CATEGORICAL indicators, and Latent GOLD are common choices.

Reporting it in APA 7

Report the indicators and why you chose them, the model specification, the software and number of random starts, a fit table for every solution you fit (log-likelihood, number of parameters, AIC, BIC, adjusted BIC, entropy, smallest profile size, and test p values), why you chose the final solution, a profile plot, profile sizes and average posterior probabilities, and the method for relating profiles to other variables. These made-up numbers are internally consistent, so you can use them as a template:

We fit latent profile models with one to six profiles to four standardized motivation indicators (N = 612) in Mplus 8, holding variances equal across profiles and covariances at zero, with 500 random starts and 100 final-stage optimizations; the best log-likelihood was replicated for every model. BIC was lowest for four profiles (6,791.6, compared with 6,842.1 for three and 6,803.9 for five), and the BLRT favored four profiles over three (p < .001) but not five over four (p = .09). The five-profile solution added a profile that differed from the moderately motivated profile only in level. We therefore retained four profiles.

Entropy was .84, and average posterior probabilities ranged from .86 to .93. The profiles were highly motivated (31%), moderately motivated (42%), low competence (18%), and amotivated (9%, n = 55; see Figure 1). Using the BCH method to account for classification error, GPA differed across profiles, χ²(3) = 41.2, p < .001, and was highest in the highly motivated profile (M = 3.42) and lowest in the amotivated profile (M = 2.71).

Our results chapter guide covers the general APA 7 formatting rules.

Questions committees ask

“Why not cluster analysis?” LPA and LCA are statistical models: they give each person a probability of membership, provide fit statistics and tests for the number of profiles, and handle missing data and uncertainty more formally.

“How did you decide on the number of profiles?” Show the fit table, name the criteria in the order you weighed them, and explain why the solutions with one more and one fewer profile were worse.

“Your entropy is high. Doesn't that confirm the solution?” No. Entropy measures classification precision, not whether the number of profiles is right.

“One profile is very small. Is it real?” Report its size and posterior probability, check whether it appears across random starts and specifications, and be cautious interpreting it.

“Can you just compare profiles with ANOVA?” Not on modal assignments without accounting for classification error. Use a three-step or BCH approach.

“Are these real types of people?” The profiles summarize patterns in this sample. Say that they may reflect a continuum or nonnormality rather than distinct types, and that replication in another sample would strengthen them.

Common mistakes

  • Choosing the number of profiles from one statistic, especially entropy.
  • Too few random starts, so the software reports a local solution.
  • Outcomes or predictors included as indicators.
  • Modal assignment followed by ANOVA without accounting for classification error.
  • Names that overclaim, such as “at-risk,” when the profile is only lower on some indicators.
  • Reporting only the chosen solution instead of the full fit table.
  • Treating profiles as proven types rather than a summary of heterogeneity.

Checklist

  • Person-centered research question and indicator choice justified
  • Indicators standardized or scaled, with missing-data handling
  • Model specification (variances and covariances) stated
  • Software, random starts, and replicated best log-likelihood
  • Fit table for every solution: LL, parameters, AIC, BIC, aBIC, entropy, smallest profile, test p values
  • Rationale for the chosen number of profiles
  • Profile plot, sizes, and average posterior probabilities
  • Three-step or BCH method for predictors and outcomes
  • Limits of interpreting profiles as types

How we did this

The guidance draws on Nylund, Asparouhov, and Muthén (2007), Structural Equation Modeling; Masyn (2013), in The Oxford Handbook of Quantitative Methods; Tein, Coxe, and Cham (2013), Structural Equation Modeling; Bauer and Curran (2003), Psychological Methods; Nagin (2005), Group-Based Modeling of Development; Vermunt (2010), Political Analysis; and Asparouhov and Muthén (2014), Structural Equation Modeling. The abstract counts come from our psychology methods overview, built from OpenAlex dissertation and thesis abstracts. The analysis code and result tables are on GitHub.