Missing Data in Latent Transition Analysis: A Monte Carlo Simulation Study

Introduction
Latent Transition Analysis (LTA) is widely used to model longitudinal changes in unobserved categorical states. However, applied researchers frequently encounter missing data, and there is limited guidance on how different missing data mechanisms, and analytic techniques used to address them, affect the recovery of latent classes and transition parameters. This study evaluates the performance of common methods used to handle missing data within an LTA framework under varying conditions.

Methods
We conduct a Monte Carlo simulation examining a two-class LTA model with binary indicators. Simulation conditions vary sample size, number of time points, number of indicators, missing data mechanism (MCAR vs. MAR), and missing data rate. Class membership follows a time-homogeneous transition matrix, and indicators follow class-specific emission probabilities with measurement invariance across time. Missingness is introduced randomly under MCAR and based on prior observations under MAR. Models are estimated using Full Information Maximum Likelihood (FIML), Multiple Imputation (MI), and weighting-based approaches.

Results
Results indicate that model performance depends strongly on the missingness mechanism and missing data rate. Parameter recovery was more accurate under MCAR than MAR missingness. FIML demonstrated stable performance across most scenarios. MI and weighting methods produced comparable results but were more sensitive to simulation parameters. Larger sample sizes and more indicators substantially improved latent class recovery and reduced bias in emission and transition parameter estimates across methods.

Conclusions
These findings highlight that no single missing data analytic technique is universally optimal. Method choice should be based on the characteristics of data and missingness. The results demonstrate how study design features—such as sample size and number of indicators—interact with the missingness mechanism and rate to influence latent class recovery and transition parameter estimation. Together, these findings highlight the importance of selecting the most appropriate analytic approach when applying LTA to longitudinal data with missingness.