Statistics Glossary → Latent Profile Analysis (LPA)
Latent Profile Analysis (LPA)analysis
A method that uses a mixture model to find unobservable subgroups (latent profiles) sharing similar patterns of values on continuous measurements, and estimates the probability that each individual belongs to each profile.
Concept
Latent profile analysis (LPA) is the counterpart of latent class analysis for continuous indicators. In a group of students measured on anxiety, motivation and study time, for instance, it asks whether there are a few subgroups that differ in the 'profiles' these scores form together. Unlike clustering methods, it rests on a statistical model: the number of profiles is chosen by fit criteria and tests, and membership probabilities are given for each individual.
The model assumes that the observed data are a mixture of several normal (Gaussian) distributions: f(x) = Σ_k π_k · N(x | μ_k, Σ_k). Here π_k is the share of the k-th profile in the population, μ_k the profile's mean vector and Σ_k its variance-covariance matrix. The parameters are estimated with the EM algorithm, after which posterior profile probabilities are computed for each individual. Whether the profiles' variances are equal or free, and whether within-profile covariance between indicators is allowed, is set by a few standard model forms (EEI, VVI, EEE, VVV); freer structures are more flexible but need more parameters.
The number of profiles is chosen by comparing models with different numbers of profiles using information criteria such as BIC, the LMR and BLRT tests, and entropy; very small or uninterpretable profiles deserve separate scrutiny. The model assumes the within-profile distribution is normal; skewed distributions can lead it to suggest extra profiles where there is no real group structure. Because EM can get stuck in local solutions, many starts are needed. The profiles are a probabilistic summary of the data and should not be read as groups known to exist.
When to use it
- When individuals are to be divided into subgroups of unknown number by the patterns of values they show jointly on several continuous measurements, with membership probabilities also reported (student motivation profiles, health risk profiles, psychological scale scores).
- If the indicators are categorical (yes/no, or Likert options treated as separate categories), latent class analysis is used.
- If a quick, distance-based grouping without a probability model or fit comparison is enough, K-means or hierarchical clustering may be preferred.
An everyday example
A university counselling centre is examining 400 students' weekly study hours, average sleep time and exam anxiety scores. By looking at the patterns these three measurements form together, latent profile analysis might find three profiles: 'balanced' (moderate study, enough sleep, low anxiety; 50%), 'overloaded' (heavy study, little sleep, high anxiety; 20%) and 'disengaged' (little study, plenty of sleep, low anxiety; 30%). The probability of belonging to each profile is computed for every student, and the centre can offer a stress management workshop to those who appear to belong to the 'overloaded' profile with high probability.
Parameters in MerQur
Variance-covariance model (1 EEI / 2 VVI / 3 EEE / 6 VVV)
Concept: The assumption that sets the shape of the profiles. The first letter of the model name shows whether variances are equal (E) or free (V) across profiles; the last letter whether the within-profile covariances between indicators are zero (I), equal across profiles (E) or free (V). EEI is the most restricted structure with the fewest parameters, VVV the most flexible with the most.
Everyday example: In the EEI model, sleep times in each profile spread equally widely and, within a profile, study time and anxiety are assumed to be unrelated. In the VVV model the 'overloaded' profile's sleep may be far more variable, and a relationship can be allowed in which students in this profile who study more are also more anxious.
What it does: Equal or free variances across profiles; covariance zero or free (model names EEI, VVI, EEE, VVV).
Effect on the result: The number of parameters, the log-likelihood, the profile shapes and the selected number of profiles change.
On the validation data: (iris, K = 1..4 log-likelihood): EEI -741.0175, -488.9148, -361.4255, -310.1170; VVI -741.0175, -386.1853, -306.8605, -264.8476; EEE -379.9146 ... -223.0486; VVV -379.9146 ... -164.2839. For VVV K = 3 and 4, higher-likelihood solutions with near-singular components are spurious maxima; MerQur excludes them or reports them with a warning.
→ Covariance (correlation) structure (UN, VC, AR(1), exchangeable)
Maximum number of profiles and Number of profiles to report
Concept: A separate model is estimated for every number of profiles from 1 up to the maximum, and these are compared. The number of profiles to report sets which model's detailed results are given; if left empty, the best model by the selection criterion is reported.
Everyday example: The centre tries models with 1 to 5 profiles. Even if BIC suggests four, if the fourth profile has only 12 students the centre may prefer to report the three-profile model in detail.
What it does: Models with 1..K profiles are estimated; if the number of profiles to report is not given (0), the best K by the selection criterion is reported in detail.
Effect on the result: The range of models compared and the model reported in detail change.
On the validation data: iris maximum K 1, 4, 6 (K = 5 and 6: -277.0455 and -253.6222); when the number of profiles to report is set to 3, 3 profiles are reported (automatic BIC selects 4).
Number of random starts and Random seed
Concept: In mixture models the likelihood can have several peaks, and the EM algorithm may stay on a local peak depending on where it starts. Many random starts increase the chance of finding the highest peak; the seed makes the result reproducible.
Everyday example: Started once, the algorithm might fail to see the 'overloaded' students as a separate profile and remain in a weaker solution that mixes them with the 'balanced' profile. Choosing the best of 100 different starts reduces this risk.
What it does: With a multimodal likelihood EM gets stuck at a local maximum; the probability of finding the global maximum increases as the number of starts increases.
Effect on the result: Many starts raise the chance of finding the global maximum; with few starts a local solution may be reported.
On the validation data: In iris EEI, 1, 20, 100 starts and seed 7 give the same result. pisa100 VVV K = 3: with 100 starts and 500 starts -251.0452; the likelihood is highly multimodal, and there is a higher (-249.1638) but sparse peak (1 time in 1000 starts).
Standardize indicators (z)
Concept: Transforms the indicators to mean 0 and standard deviation 1. This changes only the scale: the selected model and the structure of the profiles stay the same, but profile means and variances are read in standard-deviation units rather than original units.
Everyday example: With study time in hours and anxiety on a 20-80 score, after standardisation a result such as 'the overloaded profile's anxiety mean is +1.2' says that this profile's anxiety is 1.2 standard deviations above the overall mean, which makes indicators in different units easier to compare.
What it does: Converts the indicators to z-scores (n-1) before estimation.
Effect on the result: The log-likelihood shifts by the scale Jacobian; the selected model and the profile structure do not change, means and variances are given on the z scale.
On the validation data: z (n-1) shifts the log-likelihood by the scale Jacobian (iris EEI K = 1: -849.3564, raw -741.0175); the selected model and the profiles are the same.
Selection criterion (BIC / SABIC / AIC / CAIC)
Concept: Information criteria combine fit (the log-likelihood) with a penalty for the number of parameters, and the model with the smallest value is chosen. AIC applies a light penalty and leans towards more profiles; BIC and CAIC apply heavier penalties; SABIC is a sample-size-adjusted, less conservative form of BIC.
Everyday example: A four-profile model may fit the data slightly better, but every new profile means new means and variances. BIC may decide the extra complexity is not worth it and choose three profiles, while AIC may choose four.
What it does: Information criterion for selecting the number of profiles; the K with the smallest value is selected.
Effect on the result: Can change only the selected K; the estimates for each K do not change.
On the validation data: (iris EEI, K = 1..4): BIC 1522.120, 1042.968, 813.042, 735.479; AIC 1498.035, 1003.830, 758.851, 666.234; all four criteria select 4.
Maximum EM iterations
Concept: The maximum number of times the EM algorithm is repeated within each start. Each iteration slightly improves the profile shares, means and variances; if the algorithm is stopped before converging, an unfinished solution is reported.
Everyday example: In the first rounds the profile means move quickly, then the changes slow down and stop. If the round limit is too low, the calculation may be cut off before the 'overloaded' profile's average sleep time has settled.
What it does: Maximum number of EM iterations in each start.
Effect on the result: With few iterations the model may not converge; after convergence it has no effect.
On the validation data: In iris EEI it converges at 50 iterations too (no effect).
BLRT bootstrap samples
Concept: The parametric bootstrap likelihood ratio test (BLRT) tests whether a K-profile model is significantly better than a (K − 1)-profile model. Many datasets are simulated from the (K − 1)-profile model, and the test looks at how unusual the improvement in the real data is compared with these datasets.
Everyday example: Assuming 'there were really only two profiles', the model generates hundreds of artificial groups of students and looks at how much a third profile improves them. If the improvement in the real data is seen in almost none of these artificial groups, the third profile is justified.
What it does: Compares the K-profile model with the K-1-profile model by a parametric bootstrap likelihood ratio test (0 = off).
Effect on the result: The fit does not change; the table gains a BLRT p column. p depends on the bootstrap samples and the seed.
On the validation data: iris EEI 50 samples: observed LRTS 504.205, 254.979, 102.617; p 0.0196.
Extra outputs: save
Concept: Adds to the dataset each individual's most probable profile (the modal profile) and their probabilities of belonging to each profile. The probabilities show how firmly the individual is assigned and let later analyses take this uncertainty into account.
Everyday example: One student may be assigned to the 'overloaded' profile with probability 0.97, another with 0.55. The centre can send workshop invitations first to students like the first one, whose profile is clear.
What it does: Adds the modal profile and the posterior profile probabilities to the dataset.
Effect on the result: The estimates do not change; new columns are added to the dataset.
On the validation data: iris K = 3: gives the 150 x 3 matrix of posterior probabilities.
LMR test (always reported)
Concept: The Lo-Mendell-Rubin (LMR) test compares a K-profile model with a (K − 1)-profile model through a corrected approximation to the likelihood ratio. Such special tests are used because the usual chi-square test is not valid for comparing numbers of profiles; it is much faster to compute than the BLRT.
Everyday example: If the LMR test gives p < 0.001 for three profiles and p = 0.30 for four, the third profile clearly improves the model while the fourth makes no meaningful contribution.
What it does: Compares the K-profile model with the K-1-profile model by an adjusted likelihood ratio; the MerQur correction is LR/(1 + 1/(dp·ln n)).
Effect on the result: The fit does not change; an LMR p column that helps to select the number of profiles is given.
On the validation data: (iris EEI, K = 2..4): 484.852, 245.192, 98.678.
Published tidyLPA vignette (pisa100, model 6, K = 3)
Concept: This is not a setting but the result of a comparison with a published example analysis: it is given to show whether the result obtained with the same data and the same model structure matches the result published for another piece of software.
Everyday example: It is like a student checking a solved problem against the answer key at the back of the book: if the results match, confidence in the method grows and it can be applied with confidence to one's own data.
What it does: The result of the VVV model (model 6) with K = 3 on the pisa100 data.
Effect on the result: Not an option but an example data result.
Example on real data
Data: Fisher 1936 iris, R datasets (n = 150)
Variables: indicators Sepal.Length, Sepal.Width, Petal.Length, Petal.Width (sepal and petal measurements)
- Question: Into how many latent profiles (subgroups) do the four continuous measurements of 150 iris flowers divide?
- Model selection: Model 1 (EEI: equal variances across profiles, covariance 0) was estimated with 1-4 profiles by EM (100 random starts). BIC 1522.12, 1042.97, 813.04 and 735.48; 4 profiles were selected by the smallest BIC (LMR p < .001).
- Result: Profile proportions .33, .28, .22 and .16 (with modal assignment 50, 42, 35 and 23 flowers).
- Classification quality: Entropy = 0.945; mean posterior probabilities 0.94-1.00, the profiles are well separated.
- Descriptive: The first profile is the setosa group with small petals (Petal.Length mean 1.46); the other three profiles are ordered by petal length (3.90, 4.78, 5.80).
- Limitation: The selected number of profiles equals the largest value scanned (4); more profiles should be tried. Because the EEI model assumes zero correlation between variables, it may suggest too many profiles for correlated measurements (iris has 3 species).
- APA: Latent profile analysis (model 1, EEI) was estimated with 1-4 classes (EM, 100 random starts, N = 150); the 4-class solution was selected (BIC = 735.48), with class proportions .33, .28, .22, .16 (entropy = .95, LMR p < .001).
Related concepts
Bootstrap likelihood ratio test (BLRT)Lo-Mendell-Rubin (LMR) testChoosing the number of clustersCovariance (correlation) structure (UN, VC, AR(1), exchangeable)EM (expectation-maximization) algorithmRandom starts and local optimaz-score (standardization)Information criteria (AIC, AICc, BIC)
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.