The possibility that "coder drift" may be a source of bias in classroom observation systems was investigated in two parallel experiments using the Flanders System of Interaction Analysis. Coder drift may be described as the differences in coding which result when observers create their own idiosyncratic systems of classifying teacher behaviors. In this experiment, observers all underwent the same training and then worked in pairs coding teacher behaviors in videotaped junior high school social studies lessons. When within-pair reliabilities were higher than between-pair reliabilities, coder drift was judged to have occurred. Results of analyses on the relationship between coding-related talk between coders within pairs and similarities in their coding, analyses on the differences in coding which resulted when coders worked alone, and discriminant analyses on differences among pairs of coders, are reported. The discussion focuses on the implications of these results for classroom research on teacher effects. (Author/CTM)