When classes are the units of analyses, estimates of the reliability of class means are needed. Using classical test theory it is difficult to treat this problem adequately. Generalizability theory, however, provides a natural framework for dealing with the problem. Each of four possible formulas for the generalizability of class means is derived from two points of view. Each of the four generalizability coefficients is shown to be the expected value of a variation on the split-half method for estimating reliability. Finally, the four coefficients are related to estimates of the reliability of class means that have been used previously. (Author)