Purpose

Public preview: This page shows the analysis output, figures, model-selection guidance, and reporting example. It does not expose the underlying R code. The paid version includes the complete editable R Markdown source, reusable LPA module, all analysis code, synthetic data, documentation, and exported results.

Latent profile analysis (LPA) searches for subgroups whose members have similar patterns across continuous indicators. The method is useful when the research question concerns combinations of characteristics rather than the effect of one variable while holding the others constant.

This workflow emphasizes model enumeration and uncertainty. It does not treat a lower BIC, higher entropy, or significant likelihood-ratio comparison as a self-interpreting answer. The retained solution must also converge repeatedly, avoid inadmissible parameters, contain usable class sizes, remain interpretable, and fit the study’s theory and purpose.

The included data are entirely synthetic. Profiles are model-based summaries of this sample. They are not diagnoses, natural kinds, or fixed personality types. Naming and using profiles require theory, replication, uncertainty reporting, and careful attention to consequences.

Scope and learning outcomes

After completing the example, the user should be able to:

  • select and scale indicators based on the construct and research question;
  • fit one- through six-profile diagonal Gaussian models with many random starts;
  • identify local maxima, convergence failures, tiny classes, and variance problems;
  • compare AIC, BIC, adjusted BIC, entropy, bootstrap likelihood-ratio evidence, and profile sizes;
  • inspect posterior probabilities rather than relying only on modal assignment;
  • plot and label profiles using standardized and original-score summaries;
  • examine external variables without pretending that assigned classes were observed without error; and
  • write a cautious results section that documents ambiguity and model dependence.

The primary fitted model allows profile-specific means and variances while setting within-profile covariances to zero. Alternative covariance structures, categorical indicators, missing-data handling, complex sampling, multilevel mixtures, and longitudinal latent-transition models require extensions.

Analysis map

Stage Main question Evidence
Indicator preparation Do indicators represent the intended person-centered question? Theory, distributions, missingness, scale and redundancy
Enumeration How many profiles are plausible? Information criteria, BLRT, convergence, replicated likelihood, class size
Classification How clearly are participants distinguished? Entropy, posterior probabilities, uncertainty
Interpretation Are profiles distinct and useful? Standardized and raw means, theory, stability
External validation Do profiles relate to variables not used to create them? Probability-weighted descriptions and cautious modal-class models

User settings

Simulate and import the example

The four generating profiles are deliberately overlapping, and the smallest population profile represents about 7% of participants. generating_profile is retained only to evaluate the synthetic exercise; it is never supplied to the fitted models.

Indicator audit and scaling

Distributions and missingness

indicator n missing mean sd min max
academic_stress 900 0 50.79 10.20 23.14 79.31
advisor_support 900 0 3.27 0.73 1.00 5.00
research_confidence 900 0 49.62 11.12 16.77 77.79
belonging 900 0 3.06 0.65 1.00 4.72
burnout 900 0 46.04 10.14 10.08 73.95

Indicators should be selected because their joint configuration answers the research question. Highly redundant indicators can overweight one construct, while variables with weak theoretical relevance can create profiles that are difficult to interpret or reproduce.

Correlation and redundancy review

Standardization

indicator original_mean original_sd
academic_stress 50.789 10.200
advisor_support 3.267 0.732
research_confidence 49.618 11.116
belonging 3.065 0.646
burnout 46.037 10.136

Standardization places the indicators on comparable units and makes each indicator contribute on an approximately equal variance scale. It does not make indicators equally reliable or equally important. Report the scaling rule and retain the original-score summaries for interpretation.

Model enumeration

classes log_likelihood parameters AIC BIC adjusted_BIC entropy smallest_class_percent average_max_posterior replicated_best next_best_gap converged admissible
1 -6382.722 10 12785.44 12833.47 12801.71 1.000 100.000 1.000 100 0.000 TRUE TRUE
2 -5769.724 21 11581.45 11682.30 11615.61 0.826 36.600 0.951 1 0.001 TRUE TRUE
3 -5595.006 32 11254.01 11407.69 11306.06 0.838 9.026 0.928 1 0.000 TRUE TRUE
4 -5513.811 43 11113.62 11320.12 11183.56 0.800 7.697 0.887 6 0.000 TRUE TRUE
5 -5498.880 54 11105.76 11365.09 11193.59 0.824 1.910 0.886 1 0.123 TRUE TRUE
6 -5486.524 65 11103.05 11415.20 11208.77 0.797 2.444 0.854 1 0.239 TRUE TRUE

Each solution uses 100 random initializations. The retained likelihood should be replicated across independent starts. A solution that appears only once may be a local maximum or an unstable optimum even if its BIC is attractive.

Information criteria and entropy

BIC and adjusted BIC penalize model complexity differently. Entropy summarizes separation across the full posterior distribution; it should describe classification precision after a solution is substantively credible, not choose the number of profiles by itself.

Random-start stability

Parametric bootstrap likelihood-ratio comparisons

classes observed_lrt bootstrap_p bootstrap_reps
2 1225.996 0.048 20
3 349.436 0.048 20
4 162.391 0.048 20
5 29.862 0.143 20
6 24.713 0.524 20

The bootstrap comparison asks whether the improvement from k − 1 to k profiles is larger than expected when data are generated from the smaller fitted model. The p-value is simulation-based and should be interpreted with the number of replications, start quality, sample size, and other enumeration evidence. Twenty replications keep this teaching example runnable; publication work generally warrants substantially more replications and starts.

Select and interpret the retained solution

profile assigned_n assigned_percent model_estimated_percent
1 381 42.33 41.60
2 195 21.67 21.19
3 255 28.33 29.51
4 69 7.67 7.70

The automated rule identifies the admissible solution with the lowest BIC after excluding profiles smaller than 3%. That rule is a transparent starting point, not a substitute for reviewing neighboring solutions. In this synthetic example it retains 4 profiles.

Standardized profile plot

Original-score summaries

profile academic_stress advisor_support research_confidence belonging burnout
Profile 1 42.99 3.68 57.10 3.44 38.05
Profile 2 54.24 2.55 41.16 2.55 48.99
Profile 3 55.08 3.52 50.15 3.16 50.93
Profile 4 67.00 2.02 30.45 2.09 62.33

Profile names should be brief, descriptive, and nonstigmatizing. Report the indicator pattern before introducing a label. Avoid labels that imply clinical diagnosis, causality, permanence, or moral judgment.

Classification uncertainty

modal_profile n mean_max_posterior percent_below_70 mean_probability_gap
1 381 0.923 6.824 0.851
2 195 0.855 20.513 0.719
3 255 0.845 20.392 0.705
4 69 0.931 8.696 0.864

Modal assignment discards the remaining probabilities. Report entropy for the solution and profile-specific posterior summaries, and retain the full probability matrix for downstream methods that account for classification error.

External validation without treating class as observed

modal_profile n milestone_mean milestone_sd support_seeking_percent weighted_milestone_mean weighted_support_seeking_percent effective_n
1 381 74.85 8.05 33.07 74.72 33.10 374.44
2 195 60.74 8.38 58.97 61.09 58.84 190.66
3 255 69.49 8.28 45.10 69.29 44.88 265.61
4 69 54.28 7.79 66.67 54.85 67.35 69.29
term estimate std.error statistic p.value conf.low conf.high
(Intercept) 76.387 0.572 133.491 0.000 75.264 77.510
modal_profile2 -14.054 0.716 -19.638 0.000 -15.458 -12.649
modal_profile3 -5.390 0.660 -8.164 0.000 -6.685 -4.094
modal_profile4 -20.439 1.063 -19.225 0.000 -22.525 -18.352
program_stageCoursework -2.928 0.785 -3.731 0.000 -4.468 -1.388
program_stageData collection -1.895 0.745 -2.544 0.011 -3.357 -0.433
program_stageProposal -2.025 0.718 -2.822 0.005 -3.434 -0.617

The modal-class regression is intentionally labeled naive: it treats the assigned profile as if it were observed without error and can attenuate or distort external associations. Posterior-weighted means retain more uncertainty but are still descriptive and are not a full three-step correction. For confirmatory distal-outcome inference, use an appropriate BCH, Vermunt, or one-step method, predefine the external model, and avoid allowing the distal outcome to redefine the profiles unless a one-step model is the scientific target.

Sensitivity and stability checks

scaling log_likelihood smallest_profile_percent modal_assignment_agreement
Mean/SD standardization -5513.811 7.697 1
Median/MAD standardization -5257.457 7.726 1

The robust-scaling check asks whether unusual observations and scaling choices materially alter the profile solution. Direct likelihoods are not comparable across differently transformed data; focus on profile shapes, proportions, and assignment agreement. A stable result under one alternative does not establish population invariance or replication.

Dynamic results narrative

Latent profile models containing one through 6 profiles were estimated for 900 synthetic participants using five standardized continuous indicators. Each solution used 100 random starts. The retained 4-profile solution converged, reproduced the best log likelihood in 6 starts, and had BIC = 11320.1, compared with 11407.7 for the adjacent smaller solution. Its adjusted BIC was 11183.6, entropy was 0.800, and the smallest model-estimated profile contained 7.7% of participants. The bootstrap likelihood-ratio comparison with the adjacent smaller solution produced p = .048 using 20 replications. These indices were considered jointly with convergence, profile size, classification uncertainty, interpretability, and neighboring solutions. Posterior probabilities showed an average maximum assignment probability of 0.887. Profile-specific percentages below a .70 maximum posterior ranged from 6.8% to 20.5%. The profiles were interpreted from their indicator patterns and original-score means before descriptive labels were considered. External milestone and support-seeking summaries were examined using both modal assignment and posterior weighting; the modal regression was treated as exploratory because it ignores classification error.

Model-enumeration and reporting checklist

  • Explain why each indicator belongs in the person-centered question and why excluded variables were not used to form profiles.
  • Report original indicator distributions, missingness, transformations, and scaling.
  • State the distributional and covariance assumptions of the fitted profile model.
  • Report the candidate range, random starts, convergence criteria, best-likelihood replication, and inadmissible solutions.
  • Present AIC, BIC, adjusted BIC, entropy, bootstrap likelihood-ratio evidence, and profile proportions together.
  • Describe why the retained solution was preferred to both adjacent solutions.
  • Report model-estimated and modal-assignment profile sizes.
  • Show standardized shapes and original-score means before naming profiles.
  • Report average posterior probabilities or related classification-uncertainty summaries by profile.
  • State whether external variables were included during enumeration, examined with modal assignment, weighted by posterior probability, or modeled with a corrected three-step approach.
  • Avoid causal language for cross-sectional profile differences.
  • Treat small profiles, low entropy, local maxima, and sample-specific labels as limitations rather than details to hide.
  • Seek replication in an independent sample before presenting profiles as stable population subgroups.

When this model is not enough

  • Categorical indicators: use latent class models with appropriate item-response distributions.
  • Repeated profiles over time: use latent transition or growth-mixture models.
  • Nested participants: use multilevel mixture models or account for clustering in design-based inference.
  • Strong residual associations: compare covariance structures or model theoretically justified local dependence.
  • Missing indicators: use full-information or multiple-imputation methods compatible with mixture models and assess sensitivity.
  • Distal outcomes central to inference: use BCH, corrected three-step, or a prespecified one-step model.
  • Very small classes: increase sample size, simplify the model, or reconsider whether the class is stable and scientifically usable.
  • Clinical classification: evaluate calibration, decision consequences, fairness, replication, and individual-level uncertainty; exploratory LPA alone is insufficient.

Exports

file
bootstrap_likelihood_ratio.csv
external_validation_summary.csv
indicator_audit.csv
indicator_scaling.csv
model_enumeration.csv
naive_modal_class_regression.csv
posterior_probabilities.csv
profile_means_original_scale.csv
profile_means_standardized.csv
profile_sizes.csv
random_start_loglikelihoods.csv
scaling_sensitivity.csv
synthetic_data_with_profile_probabilities.csv
synthetic_lpa_data.csv

References

  • Berlin, K. S., Williams, N. A., & Parra, G. R. (2014). An introduction to latent variable mixture modeling (part 1): Overview and cross-sectional latent class and latent profile analyses. Journal of Pediatric Psychology, 39(2), 174–187.
  • Masyn, K. E. (2013). Latent class analysis and finite mixture modeling. In T. D. Little (Ed.), The Oxford handbook of quantitative methods (Vol. 2, pp. 551–611). Oxford University Press.
  • Nylund-Gibson, K., & Choi, A. Y. (2018). Ten frequently asked questions about latent class analysis. Translational Issues in Psychological Science, 4(4), 440–461.
  • Vermunt, J. K. (2010). Latent class modeling with covariates: Two improved three-step approaches. Political Analysis, 18(4), 450–469.
Dissertation Stats Helper · Reproducible statistical guidance for doctoral research · Enumerate carefully, retain uncertainty, and describe profiles without overclaiming.