Most previous studies of the dependability of student ratings of instruction have used the class as the unit of analysis, thus confounding the effects of the course, the teacher, and the interaction between the two. This paper presents two studies in which generalizability theory is applied to data collected within designs allowing for independent estimates of the variance components for the teacher effect and for the course effect. Results show that generalizing over items and students but not over teachers or courses yields highly dependable results. Generalizing over courses as well as items and students, with the teacher as the unit of analysis, yields moderately dependable results. Generalizing over teachers, items and students with the course as the unit of analysis yields very low dependability. Implications of the results for tenure and curricular decisions are discussed. (Author)