Power Analysis for Latent Class Identification Using the Bootstrap Likelihood Ratio Test

• Activity Number: 80 - Contributed Poster Presentations: Mental Health Statistics Section
• Type: Contributed
• Date/Time: Monday, August 3, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Mental Health Statistics Section
• Author(s): Alai Tan*
• Companies: Ohio State University College of Nursing
• Keywords: adverse childhood experience; bootstrap likelihood ratio test; effect size; latent class analysis; sample size

Abstract

Latent class analysis (LCA) is a statistical method to identify underlying subgroups in a population. The bootstrap likelihood ratio test (BLRT) outperforms other model performance methods to identify number of latent classes in LCA. However, insufficient sample size may lead to under-extraction of the latent classes. Statisticians often use simulation-based approach to calculate statistical power for the BLRT in LCA. Such simulation is computation intensive and thus is not feasible for studies with a range of possible indicators and unknown measurement strength. Alternatively, Dziak et al. has proposed an effect size based approach to calculate power for the BLRT in LCA. The present study illustrates an application of this method to the power calculation for a secondary data analysis aiming to identify latent classes of adverse childhood experiences with varying number of indicators (8-16), measurement strength (medium or strong), and distribution of latent class membership. Curves of maximum identifiable number of latent classes with sufficient power are generated to guide researchers to interpret LCA results and detect possible under-extraction with given sample size.