Generalizability theory is described and applied to data collection methods for student ratings of instruction. Three generalizability coefficients (generalizing over both students and items, generalizing over students only, and generalizing over items only) are defined, and the appropriateness of each for various purposes is discussed. Generalizability theory is applied to data derived from the University of Washington Instructional Assessment System. Results show that adequate dependability for evaluating teacher or course effectiveness is obtained when fifteen or more students rate a class on four or five items. A single item is reliable for diagnosing instructional problems with 15 or more student raters. (Author/MV)