LAUSR

Latent class analysis (LCA) is a popular statistical method used to group individuals into homogeneous subpopulations based on the responses to a set of variables. In the conventional LCA, the number of classes must be specified in advance. Various model selection criteria (e.g., information criteria) have been used to determine the optimal number of classes via the class enumeration process. However, previous studies have shown that different information criteria can favor different models because the penalty terms for the individual information criteria differ from one another, and there is no consensus on the optimal criterion. By incorporating the Dirichlet process mixture model (DPMM), the Bayesian nonparametric LCA is an alternative to the conventional approach. DPMM can circumvent the problems underlying class enumerating by assuming infinitely many classes and inferring the number of classes from the data. The extant literature in psychology and social sciences on DPMM-LCA is limited. The exception is Si and Reiter (2013), where a DPMM of multinomial distributions was proposed to perform multiple imputation for categorical data. Nonetheless, Si and Reiter (2013) specified a model for density estimation rather than clustering, which lacks a necessary post-processing procedure. To bridge the gap in research, we proposed an extended DPMM-LCA approach allowing for clustering individuals by items measured on different metrics, including categorical, continuous, and mixedtype. In combination with a post-processing procedure, the proposed DPMM-LCA approach allows one to simultaneously do the following: 1) determine the number of classes, 2) handle the label switching issue (see Figure 1 for an example), 3) classify individuals, and 4) obtain class-specific parameters for characterizing classes (see Figure 2 for an example). We