Title: Structural equation modeling considering cluster structure
Authors: Kensuke Tanioka - Wakayama Medical University (Japan)
Hiroshi Yadohisa - Doshisha University (Japan)
Ippei Takasawa - Doshisha University (Japan) [presenting]
Abstract: Structural equation modeling (SEM) is a method that clarifies the relationships between observed variables and latent factors using confirmatory factor analysis. Generalized structured component analysis (GSCA) is one type of SEM and is a component-based approach like PCA and is estimated using alternating least squares (ALS). GSCA can express many flexible models because it consists of three structural equations: from variables to variables, from variables to components, and vice versa. To consider the heterogeneity of data, fuzzy clusterwise GSCA (FCGSCA) is proposed as extended GSCA. FCGSCA is a method such as simultaneously estimating cluster labels for subjects and these path components by each cluster. It is difficult to assume a different path diagram for each cluster because forms of path diagrams are decided before estimation. Accordingly, FCGSCA uses same path diagram and each cluster feature is interpreted as differences of coefficients. However, there are situations in which path coefficients for each cluster are not very different although the data is heterogeneous and detection of clusters with different path coefficients is needed. Therefore, we propose a method that estimates each path diagram to be more different and makes it easy to interpret each cluster feature by constraints of the coefficient matrix.