This study was conducted to investigate the dimensionality of language tests by means of latent variable models for categorical data. It differs from previous studies by conducting the analysis at the level of the original items using the structural modeling approach of B. Muthen (1984) for dichotomous and ordered polytomous variables. In this approach, a multivariate regression model describes the relationship between a set of outcome variables, whether continuous, dichotomous, or ordered categorical, and a set of latent predictor variables. Data were test scores of the national sample of seventh and eighth graders who participated in the joint administration of Form 1 of Thinking about Language, a constructed response supplement to the Iowa Tests of Basic Skills (ITBS, and Form M of the ITBS and test scores of the national sample of seventh and eighth graders who participated in the ITBS fall 1992 national standardization of two ITBS forms. Fitting latent variable models to categorical data provides a direct means of assessing the extent to which conditional dependencies might exist among items with particular characteristics. The slightly better fit of the five latent variable model with one higher-order latent variable with paths to each first-order latent variable to all language tests in this study indicates the existence of some such dependencies unless latent variable models with content-specific dimensions are considered. This study advances the understanding of the dimensionality structures of different types of language tests and provides insights into using latent variable models for categorical data in the assessment of dimensionality. (Contains 4 tables and 15 references.) (SLD)