Addressing Missing Data in Latent Class Analysis When Using a Three-Step Estimation Approach
Sarah Depaoli, Fan Jia, Marieke Visser
Abstract
This study specifically focuses on addressing the challenges related to employing missing data techniques when estimating a conditional Latent Class Analysis (LCA) model. In the context of a conditional LCA, where covariates are incorporated, the process of estimation becomes more intricate, introducing an additional layer of complexity linked to missing data. The simulation design is structured to examine the performance of different methods of estimation in the presence of covariates and missing data. Our primary focus revolves around the ML three-step approach, and we delve into alternative missing data techniques, including full information maximum likelihood (FIML), Bayesian estimation, and multiple imputation (MI), as viable alternatives to the default listwise deletion approach. By evaluating their performance under various covariate and missing data conditions, we aim to provide valuable recommendations for applied researchers who are navigating the implementation LCA with covariates when missing data are present.