A study concerned with identifying sources of interrater variation in ratings posed the following questions: Are ratings decomposable into a single nonerror component with interrater variations representing individual error components, or is a better fit to the data provided by multiple nonerror components representing generalized rating styles? And if multiple rating styles are found, what are their characteristics? Rated events were 10-minute segments from videotapes of high school classes in four different subjects. The 50-minute composite videotape was viewed by 83 subjects (teachers, teacher trainees, school administrators, and graduate students) using a 21-item questionnaire synthesized from a variety of sources to sample three aspects of teaching behavior: intended objectives, teaching style, and interpersonal climate. The data from ratings of the four classrooms with the 21 scales formed an 83 x 21 x 4 data array. Two analyses were performed on the extended matrix: principal component analysis of covariances and correlations between rows. Additional analytical procedures were employed to characterize generalized rating styles. Conclusions are methodological rather than substantive: The analytical procedures offer the possibility of providing more information about the quality of ratings than is provided by more traditional reliability estimation procedures, and provide a basis for selecting raters having rating styles of particular interest. (Observation schedule and data tables included.) (JS)