The author proposes a three-dimensional latent variable (trait) model for analyzing attitudinal scaled data. It is successfully applied to two examples: one with 12 binary items and the other with 8 items of five categories each. The models are exploratory instead of confirmatory, and subscales from which data were selected are clearly identified. For binary items, it gives similar results with factor analysis. For polytomous items, it can estimate category scores simultaneously with the internal structure. From that, another dimension of the degree to take moderate views is extracted. This is because conventional analyses usually fix category scores as numbers, while they are free to vary in latent variable models. Computational problems are discussed, and it is expected that more than three dimensions are possible given today's computing power and tailor-made methods such as adaptive quadrature points for numerical integrations.