Although the evaluation of innovations in teacher training is often dependent on observational data, the problem of the reliability of the observations collected by a team of observers has often been treated in a superficial manner, partially because of the difficulties of maintaining an observer team intact over an extended period of time and of being permitted to observe each teacher a number of times. An analysis of variance model has been developed which permits the calculation of an overall reliability coefficient and the partitioning of the sources of variation for the typical observer team situation in which the team visits a number of different teachers only once and where the team does not necessarily contain the same members for all visits. The paradigm is developed for the situations in which there are "n" observations per item per observer as well as when there is only one observation per item per observer. The model has been tested using data based on the School University Teacher Education Center (SUTEC) observation schedule which is designed to investigate seven aspects of classroom behavior. Since the proposed model permitted the partitioning of the variance associated with the component parts of the schedule, it may provide useful as a test of the homogeneity of the items in an observation schedule as well as for reliability calculations. (Author/JS)