Title: Mixture of factor analyzers for NMAR missing data
Authors: Yuki Morioka - Doshisha University Graduate School of Culture and Information Science (Japan) [presenting]
Kensuke Tanioka - Wakayama Medical University (Japan)
Hiroshi Yadohisa - Doshisha University (Japan)
Abstract: Probabilistic principal component analysis (PPCA) is a statistical multivariate method, widely used for research in several fields, that operates with the assumption of continuous latent variables. In PPCA, we can tackle missing values by using an EM algorithm. To deal with data with heterogeneous structures, PPCA was extended to a mixture model. To improve clustering accuracy, the PPCA mixture model is also extended to a mixture of factor analyzers, which can deal with observed errors of all variables are different structures. However, this method does not assume a missing data mechanism of NMAR. If the missing data mechanism is NMAR, the result of existing method is influenced by the missing values. Therefore, we proposed a new mixture of factor analyzers. We can obtain the parameters of our proposed method using an EM algorithm. We demonstrate the validity of the proposed method through a numerical example.