Do interval-class vector-based measures of similarity among pitch- class sets, such as Robert Morris's similarity index, predict how listeners relate sets? In a series of experiments listeners were asked to judge differences between pitch- class sets. The results did not argue convincingly for the similarity index as the primary determinant in the perception of similarity. They suggested, however, that an underlying pattern might show the relationships among set classes along a number of dimensions. In an experiment involving comparison of the twelve trichords, multidimensional scaling techniques were used to model the data. The dimensions of these models can be interpreted in terms of intervals, although none of the interpretations involve total interval content or correspond to the similarity index. The results suggest two conclusions. First, the perception of similarity among pitch-class sets is tied to context and manner of presentation. Second, theoretically plausible measures of similarity may not be immediately perceivable and may require considerable training to be heard.