One of the challenges many learning scientists face is the laborious task of coding large amounts of video data and consistently identifying social actions, which is time consuming and difficult to accomplish in a systematic and consistent manner. It is easier to catalog observable behaviours (e.g., body motions or gaze) without explicitly attempting to identify their social relevance for the participants. We explore the potential for using a multimodal learning analytic approach to identify whether clusters of observable behaviours can be used to identify and characterize behavioural frames in rich video data of student interviews. We argue that by conducting a systematic analysis of behavioural frames using computerized algorithms we can model student frames as a latent class variable. We explore whether those behavioural frames overlap in productive ways with epistemological frames, thus supporting our efforts to interpret rich video data. We believe that a positive feedback loop between methodological approaches and theory will emerge as we further our understanding of framing by developing analytical models leveraged by multimodal learning analytics.