Social and emotional learning programs are designed to improve the quality of social interactions in schools and classrooms in order to positively affect students' social, emotional, and academic development. The statistical power of group randomized trials to detect effects of social and emotional learning programs and other preventive interventions on setting-level outcomes is influenced by the reliability of the outcome measure. In this paper, we apply generalizability theory to an observational measure of the quality of classroom interactions that is an outcome in a study of the efficacy of a social and emotional learning program called The Recognizing, Understanding, Labeling, Expressing, and Regulating emotions Approach. We estimate multiple sources of error variance in the setting-level outcome and identify observation procedures to use in the efficacy study that most efficiently reduce these sources of error. We then discuss the implications of using different observation procedures on both the statistical power and the monetary costs of conducting the efficacy study.