This study investigated the dimensionality of mathematics subscale scores from the National Assessment of Educational Progress for the assessment's Technical Review Panel, specifically for the data from the 1990 and 1992 main assessment in relation to students' instructional and noninstructional background variables. Discriminant analysis was applied to the math subscale scores using the background variables (questions) as grouping variables. For data from 1990 and 1992 it was hypothesized that: (1) five math subscales measure a general underlying mathematics ability; (2) five distinct subscales measure five different areas in math; and (3) in addition to the five subscales, there is a general math ability underlying part of the subscale scores. Confirmatory factor analysis tested the hypothesis of multidimensionality versus unidimensionality. Results, which are consistent across the data, indicated that when analyses were performed on the total group of students, almost perfect correlations between the subscale scores were obtained. When analyses were conducted on groups based on the background variables, lower correlations that were indicative of multidimensional subscale scores were obtained. Results indicate that models with subscale scores as latent variables exhibit a better fit to the data than models that had only one general mathematics score. An appendix presents detailed tables of findings. (Contains 70 appendix tables and 103 references.) (SLD)