The purpose of this paper is to further the cause of clarifying construct interpretations of tests, by proposing that non-metric multidimensional scaling may be more useful than factor analysis or other latent structure models for investigating the internal structure of tests. It also suggests that typical problems associated with scaling dichotomous variables can be avoided by using tetrachoric correlations as input to the multidimensional scaling. Finally, it demonstrates the utility of the suggested procedures by applying them to actual data from the Reading subtest of the Metropolitan Achievement Tests. (BW)