This study investigates four widely used methods of collecting direct similarity judgments for nonmetric multidimensional scaling: Rating Scale, Triads, Rank Order of Pairs, and Conditional Rank Order. The results indicate that: (1) when class structure is very striking in the attributes to be scaled, respondents under the Rating Scale method place significantly more weight on class structure than on continuously varying attributes in determining similarity; (2) nonmetric multidimensional scaling solutions account for a greater proportion of the variance in Rank Order of Pairs data than in data collected by the other methods studied, although the differences are so small that they would have little operational impact; and (3) the four methods do not differ significantly in test-retest reliability.