SMART-LCA Checklist

This Vignette describes the SMART-LCA Checklist: Standards for More Accuracy in Reporting of different Types of Latent Class Analysis, introduced in Van Lissa, C. J., Garnier-Villarreal, M., & Anadria, D. (2023). Recommended Practices in Latent Class Analysis using the Open-Source R-Package tidySEM. Structural Equation Modeling. https://doi.org/10.1080/10705511.2023.2250920. This version of the checklist corresponds to the tidySEM R-package’s version , 0.2.6.

The paper discusses the best practices on which the checklist is based and describes specific points to check when writing, reviewing, or reading a paper in greater detail. However, this vignette may be updated to keep pace with tidySEM package development, whereas the print publication will remain static.

Note that, although the steps below are numbered for reference purposes, we acknowledge the process of conducting and reporting research is not always linear.

1. Pre-Analysis
   • Define a research question and study design suitable for LCA.
   • Determine whether the application of LCA is confirmatory or exploratory.
   • Choose valid and reliable indicators with an appropriate data type for LCA.
   • Justify sample size.
   • Optionally: pre-register the analysis plan.
   • Optionally: use Preregistration-As-Code.
2. Examining Observed Data
   • Assess indicator distributions and scales.
     • descriptives()
   • Does the observed distribution match the intended measurement level? E.g., continuous variables have many unique values, categorical variables have a well-balaced response distribution.
   • Are all indicators on similar scales?
   • Are there distributional idiosyncrasies that need to be accounted for in model specification or by transforming the data (e.g., extreme skew)?
3. Data Preprocessing
   • Make sure continuous variables have type numeric or integer and an interval measurement level.
   • Convert ordered and binary variables to mxFactor.
     • mxFactor()
   • Convert nominal variables to binary dummy variables.
     • mx_dummies()
   • Rescale indicators that are on very different scales to prevent convergence issues.
     • poms(), scale()
   • Optionally, transform data to account for distributional idiosyncracies that could cause the model to fit poorly.
     • sqrt(), log()
4. Missing data
   • Examine and report proportion of missingness per variable.
     • descriptives()
   • Report MAR test using mice::mcar().
   • Select and report appropriate method to handle missingness. In tidySEM, this is full information maximum likelihood (FIML), which is valid regardless of the outcome of the MAR test.
5. Model specification
   • For exploratory LCA, list all sensible alternative specifications
   • In confirmatory LCA, justify alternative model specifications on theoretical grounds
   • Clearly report the model specification, so it is clear what all parameters are.
   • Optionally, verify that the model can capture distributional idiosyncracies (see Data Preprocessing).
6. Number of classes
   • Justify the maximum number of classes \(k\).
   • In exploratory LCA, estimate 1:\(k\) classes.
   • In confirmatory LCA, estimate the theoretical number of classes and choose other models to benchmark it against. This should include a 1-class model, optionally competing theoretical models, and optionally models with a different numbers of classes.
7. Justify the criteria used for class enumeration, which can include:
   • Information criteria.
   • LR tests.
     • The BLRT() is best supported by the literature; lr_lmr() has been criticized but is also available.
   • Theoretically expected class solution.
8. Justify the criteria used to eliminate models from consideration, including:
   • Model convergence.
   • Classification diagnostics.
     • table_fit()
   • Local identifiability (acceptable number of observations per parameter in each class).
   • Theoretical interpretability of the results.
     • Columns np_global and np_local in output of table_fit()
9. Transparency and Reproducibility
   • Use free open-source software to enable reproducibility.
     • See the worcs package and its Vignettes, also Van Lissa et al. (2021)
   • Comprehensively cite literature, software, and data sources.
   • Share analysis code.
   • If possible, share data. Otherwise, share a synthetic dataset.
     • Synthetic data can be created for an entire dataset in a model-agnostic way via worcs::synthetic()
     • Synthetic data can be generated from the LCA model by running mxGenerateData() on the model object.
   • Share all digital research output in line with the FAIR principles.
   • Optionally, use the Workflow for Open Reproducible Code in Science and the worcs R-package automate creating a reproducible research archive.
10. Estimation and Convergence
    • Report the method of estimation. In tidySEM, this is simulated annealing with informative start values, determined via K-means clustering for complete data and hierarchical clustering when there are missing values.
    • Assess convergence of all models before interpreting them.
    • Optionally, if there is non-convergence, use functions like mxTryHard() to aid convergence.
      • E.g., if you have concerns about model 4 in output object res, you could run res[[4]] <- mxTryHard(res[[4]], extraTries = 200) to run 200 permutations.
11. Reporting Results
    • Report the fit of all models under consideration in a table.
      • table_fit()
    • Report classification diagnostics for all models under consideration.
      • class_probs()
    • Report the minimal class proportions for all models under consideration.
      • Column n_min in table_fit() results.
    • Report class proportions for the final model
      • class_prob(res, "sum.posterior")
    • Report point estimates, standard errors and confidence intervals for model parameters.
      • table_results()
    • Optionally, convert thresholds to conditional item probabilities.
      • table_prob()
    • Assign informative class names while clarifying that these names are just shorthand.
12. Inference
    • Standard p-values in table_results() test the hypothesis that parameters are equal to zero. Consider whether these tests are meaningful and relevant.
    • Control for, or reflect on, the study-wide Type I error level.
    • Optionally, use the standard errors to test other null hypotheses.
    • Optionally, use wald_test() to test informative hypotheses.
    • Optionally, to test a confirmatory LCA model, compare parameter estimates to hypothesized values.
    • Optionally, use lr_test() to compare parameters across classes.
13. Visualization
    • Visualize raw data to assess class separability and model fit.
      • plot_profiles(), plot_density(), plot_prob(), plot_growth()
    • Show point estimates of model parameters.
    • If possible, show confidence bounds.
    • Optionally, show multivariate distributions for multivariate models.
14. Follow-up analyses.
    • Account for classification inaccuracy in follow-up analyses.
      • BCH()