A comparison of three approaches for clustering polytomous data in the presence of masking variables

To uncover the heterogeneity in a population, it is common yet important to partition individuals into distinct subgroups based on their responses to items in measurement tools. Various approaches have been introduced to tackle this clustering problem in psychology and education. To provide more guidance to practitioners, in this study, we compareD the performance of three widely-applied approaches, including the latent class analysis (LCA), k -means and k -medians, in clustering polytomous items in the presence of masking variables. In the simulation conditions considered, we found that LCA coupled with Bayesian Information Criterion (BIC) outperformed other approaches and methods for determining the optimal number of subgroups. We also applied the three approaches to an empirical data set and obtained different conclusions regarding the number of subgroups. Additionally, we discussed the limitations of this study and future research directions.