Factor analysis reduces high-dimensional data into a few latent factors, but applying it to categorical responses is challenging due to their discrete and nonlinear structure. While prior latent-factor research has focused primarily on binary response...
Factor analysis reduces high-dimensional data into a few latent factors, but applying it to categorical responses is challenging due to their discrete and nonlinear structure. While prior latent-factor research has focused primarily on binary responses, we extend this modeling framework to multinomial and ordinal data, providing a generalized approach for multi-category outcomes and incorporating an information-criterion method for automatic factor selection. By balancing model fit and complexity, the IC-based procedure helps prevent both under- and over-extraction of latent structure, improving the stability of factor determination and enhancing recovery of underlying loading patterns. Through simulation studies and real-data analyses, we confirm that the proposed approach offers consistent and reliable performance in both latent structure estimation and factor-number determination.