Engagement (i.e., presence, immersion) is important in the development of educational media; engaged users are putatively more likely to grasp pedagogical content than disengaged users. Previous psychophysiological approaches to measuring engagement, however, have been largely counterproductive in that they rely on comparisons of physiologic signals outside of the multimedia context (e.g., simple baselines), or experimentally constrain the context in which one hopes users to be engaged. We investigated whether context-dependent approaches to modeling physiological data taken from users during unconstrained media exposure was informative beyond context-independent physiological data in accounting for post-hoc self-reports of engagement. In our experiment, 39 participants were exposed to various genres of interactive (video-games) and passive media (videos), for comparison, while monitored with eye-tracking, electromyography, and autonomic nervous system sensors (i.e., electrocardiogram, photoplethysmograph, electrodermal activity), followed by a self-report battery of published measures of engagement (Brockmeyer, et al., 2009; Procci & Bowers, 2011). For context modeling, events within each media exposure were coded for content and time-stamped from screen-captures. Results indicate that context-dependent (r = .56, p < .01) and -independent physiological (r = .31, p < .05) responses were significantly correlated with engagement measures, but context-dependent responses independently accounted for substantially more variance in self-report measures (∆R2 = 23%, p < .001) as compared to context-independent responses, which were not independently significant predictors. Our findings address shortcomings in current methods for measuring engagement across different media types and genres, and show promise in providing granularity for metrics that may inform the design of engaging media.