Description

Latent class analysis (LCA) and latent transition analysis (LTA) can be extended in a variety of ways. Advanced features, such as covariates, outcomes, moderators, and inverse propensity score weights, can be added to “baseline” static and dynamic latent class models. All of the models discussed here add advanced features to baseline models. Over time, these pages will includes information about recent research on latent class moderation, adding inverse propensity score weights to models, multilevel latent class models, and associative latent transition analysis, among other topics. Note that these models all fall under the broad classification of loglinear modeling with latent variables.

Extensions to Baseline LCA and LTA

LCA with a distal outcome

Latent class membership can be used to predict an outcome. There are a number of competing approaches proposed in the literature, but the “BCH approach” is currently recommended in most applications. This approach can be extended to include control variables in the regression model predicting the outcome.

Latent class moderation

Moderation analysis is typically conducted by incorporating a single variable (e.g., gender, baseline severity) as a moderator into a multiple regression model. By using LCA, researchers can identify subgroups of people exposed to a common set of factors and who, therefore, may respond differently to intervention. That is, latent class membership can be used as the moderating variable in multiple regression.

LCA with causal inference

LCA with covariates models the association between classes and predictors, but causation cannot be inferred unless people are randomly assigned to levels of the predictors of latent class membership. Modern causal inference methods, such as inverse propensity weighting, can be used to adjust for potential confounding in observational data.

Multilevel LCA with intensive longitudinal data

Multilevel latent class analysis (MLCA) was developed more than a decade ago to address nested data, such as individuals nested within a higher-order structure (e.g., students nested in classrooms). By applying MLCA to the analysis of intensive longitudinal data (ILD), we can better understand the heterogeneity of behavior patterns in daily life and identify within-person vs. between-person risk factors.

Advanced Models

LCA: Latent class moderation

This code demonstrates how to use a latent class moderator to examine heterogeneity in intervention effects among adolescents receiving treatment for cannabis use. First, the code identifies latent classes of contextual and individual risk at baseline using LCA. Then, it uses an adjusted 3-step approach with BCH weights to regress the outcomes on level of care, latent class membership, the interaction between them, and covariates.

View Model

MLCA: Marginal approach with covariates at day-level and person-level

This model estimates six day-level latent classes of substance use consequences using a marginal modeling approach to account for the nested data structure. One one day-level covariate (daily positive affect) and person-level covariate (student is under legal drinking age) are included to predict consequence latent class membership.

View Model

MLCA: Non-parametric approach with covariates at the day-level and person-level

This code fits a 2-level latent-class model with covariates using a “non-parametric approach.” This model simultaneously estimates day-level classes of substance-related consequences and person-level classes that group individuals based on proportion of each type of day. This model includes covariates at each level to estimate associations with both the day-level and person-level classes.

View Model

MLCA: Non-parametric approach without covariates

This code fits a 2-level latent-class model without covariates using a “non-parametric approach.” This model simultaneously estimates day-level classes of substance-related consequences and person-level classes that group individuals based on proportion of each type of day.

View Model

Let’s stay in touch.

We are in this together. Receive an email whenever a new model or resource is added to the Knowledge Base.