The present study tested the assumption that self-ratings, such as those used for measuring state anxiety, do not measure a one-dimensional transcendent entity but involve decisions based on a multi-dimensional judgment. Two groups of subjects were presented with a balanced nine-item state anxiety questionnaire. Each group received a different set of instructions (a standard set and an altered instruction set suggesting unidimensionality of the questions in the questionnaire). It was hypothesized that this change in instructions would impact the structure of (...) the data. The impact of the instructions was detectable at the level of the observed correlation between the negative and positive composites, Cohen's (1988) q=-0.27$. A multigroup confirmatory factor analysis showed that positive and negative wording factors correlated more strongly when unidimensionality was suggested than when standard instructions were used, q=-0.54. Theoretical and practical implications are discussed. (shrink)