Three-Step Estimation for Latent Class Analysis • tseLCA

Overview

tseLCA (Three-Step Estimation for Latent Class Analysis) introduces bias-adjusted three-step estimation for structural latent class models in R. The package provides a comprehensive framework for estimating latent class models with covariates and distal outcomes while preserving the measurement structure established during class formation.

Building upon the efficient measurement-model estimation procedures implemented in multilevLCA, tseLCA extends existing functionality through modern three-step estimators, classification-error corrections, and variance estimation procedures that appropriately account for uncertainty from the latent class measurement stage.

Key Features

Bias-Adjusted Three-Step Estimation

tseLCA is the first R package to provide a unified implementation of modern bias-adjusted three-step estimators for latent class analysis. In contrast to traditional one-step approaches (implemented by the popular package, poLCA), where the inclusion of covariates may alter the underlying latent class definitions, three-step methods preserve the measurement model estimated in the first stage and subsequently adjust for classification error when estimating structural relationships.

The package implements both BCH- and ML-based three-step estimators with sandwich variance estimators that propagate uncertainty from the measurement model through the classification-error correction process.

Flexible Measurement and Structural Samples

Unlike conventional latent class software that uses a one-step estimation approach, tseLCA allows measurement and structural models to be estimated using different datasets. This flexibility enables researchers to calibrate a measurement model on a primary or reference sample and subsequently apply the resulting class definitions to an external dataset.

Support for Multiple Distal Outcome Types

tseLCA provides native support for a broad range of distal outcome distributions, including:

Continuous outcomes (Gaussian)
Count outcomes (Poisson)
Binary outcomes (Bernoulli).

Automated Model Optimization

Latent class estimation is often susceptible to local maxima and convergence to suboptimal solutions. To improve estimation reliability, tseLCA incorporates automated diagnostic procedures that monitor model quality during measurement-model estimation.

Missing Data Handling

Following a similar approach as multilevLCA, tseLCA employs Full-Information Maximum Likelihood (FIML) estimation to accommodate partially observed response patterns without discarding incomplete observations.

Installation

You can install the development version of tseLCA from GitHub like so:

# Install developmental tseLCA from the GitHub repository
if (!require("pak")) {
  install.packages("pak")
}

pak::pak("SamLeeBYU/tseLCA")

Once tseLCA is on CRAN, then you can simply install it from a CRAN server.

# Install tseLCA from CRAN
install.packages("tseLCA")

Then read the introductory vignette on this package’s webpage here: https://SamLeeBYU.github.io/tseLCA/articles/tseLCA-workflow.html

Example

This is a basic example which shows you how to simulate data and run a three-step LCA with a covariate in a single function call:

library(tseLCA)

# 1. Generate synthetic data 
# (3 classes, 6 dichotomous items, and a multinomial logit covariate 'Zp')
d <- generate_data(
  n = 500, 
  separation = "high", 
  scenario = "covariate", 
  seed = 1
)

# 2. Estimate the three-step model
# This automatically fits the measurement model and estimates covariate effects
fit <- three_step(
  data = d,
  Y.names = paste0("Y", 1:6),
  n_classes = 3,
  Zp.names = "Zp",
  # Proportional assignment is recommended for better uncertainty propagation
  use.modal.assignment = FALSE
)

# 3. View the measurement and structural model estimates
summary(fit)