Several methods of analyzing random samples of teacher behavior are discussed. Their assumptions and sensitivity, and the inferences they permit, are compared. Data of this kind might be obtained whenever program implementation is assessed by monitoring specific teacher behaviors. The quasi F ratio can be used in analyzing such data, but error rates are difficult to estimate and the generalization of the quasi F ratio to multivariate observations is difficult. Two alternatives are presented: (1) Observations within teachers may be averaged, leaving teacher as the only random effect. (2) Teacher may be treated as a fixed effect, but this limits the kinds of inferences which may be drawn. Results of analyses by both methods are presented. (Author/MV)