When researchers implement student ratings of college instruction, one problem that must be faced is the assessment of the reliability of the measure used in the rating system. Measures of instructor performance (as rated by students) contain a number of potential sources of error associated with courses, with semesters, with sections, with students, with items of instrument, and with their combinations. The present study examined the section effect, that is, the consistency of ratings between two sections of a course taught by an instructor within a given semester. More importantly, since the component behaviors could be rated more objectively than the global domain of behaviors, the generalizability of instructor ratings obtained used specific attribute items. Once the magnitude of each source of variation have been identified, their effects can be compared global(very general) instructor items with attribute items (dealing with different aspects of the instructor's performance). Specifically, this study examined the magnitude of variation due to course (within teacher), section (within course), items and the interaction among these factors for three global items and nine attribute items. Second emphasis of the study was to present various combinations of numbers of items, students, and sections such that a satisfactory level of generalizability was attained and compared across different type of items. The evaluation form used in the study was composed of 25 items, 16 of the 25 items on the form were relavant to this study ; four very general (global) items and twelve more specific items examing a reasonably broad sample of instructor attributes, that is, clarity of presentation of course material, ability to answer questions, vocal delivery and speech clarity, use of the blackboard, attitude toward students, enthusiasm in teaching the course, etc. The results of the study indicated that teacher performance in a given course was very consistent from section to section. The study also presented the four out of five instructors were remarkably consisted for both global and attribute items. For both global and attribute items, the teacher and student variance components were large and approximately equal. The study examined several combinations of the factors in order to determine the number of items, students and sections that are needed to obtain an acceptable level of generalizability for a measure. For both global and attribute items, the generalizability coefficients increased considerably as the number of students increased, but the number of items had a strong effect only for attribute items. Also, because the class and item-class interaction variance components were zero or small, increasing the number of class(section) had approximately the same effect as adding a corresponding number of students to a single class. The findings suggested that if an instructor's performance in a given course was being assessed, generalization was enhanced by combining ratings from more than one section, but mainly because this effectively increased the number of students sampled. In addition, there was little gain in generalizability from using more than one global item, but substantial gain from using additional attribute items. This indicated the generalizability of ratings was substantially influenced by item specificity.