The purposes of our study were: (a) to train a set of observationally determined social behaviors via peer initiation; (b) to determine if effects generalized across classroom settings and to directly intervene if generalization did not occur; and (c) to analyze components of the peer-initiation intervention. After baseline, nonhandicapped preschool children (confederates) were taught to direct social initiations to the three handicapped preschool-aged students. Teachers prompted the confederates to engage the students in social interaction when necessary and rewarded the confederates with tokens. Confederates' initiations to the students resulted in increased frequencies of positive social interaction. There was no generalization to other classroom settings, and the intervention was subsequently implemented in a second and third classroom. Next, the confederates' token reinforcement system was withdrawn, with no apparent deleterious effects on the confederates' or students' social interactions. When teachers substantially reduced their prompts to the confederates, students' social interactions decreased. Finally, reinstatement of teacher prompts resulted in increases in the confederates' social initiations and, consequently, the positive social interactions of the students.