Managing individuals' attention and interruptibility is still a challenging task in the field of human-computer interaction. Individuals' intrinsic interruptibility preferences are often established for and across different social roles and life domains, which have not yet been captured by modeling short-term opportunities alone. This paper investigates the applicability of social role theory and boundary management as theoretical underpinnings for analyzing social roles and their associated interruptibility preferences. We conducted an in-the-wild study with 16 participants for five weeks to collect individuals' social roles, interruptibility preferences, application usage and spatio-temporal information. A paired t-test shows that interruptibility models are significantly improved by incorporating individuals' self-reported social roles, achieving a F1 score of 0.73 for classifying 4 different interruptibility preferences. We design and evaluate social role classification models based on spatio-temporal and application based features. We then combined social role and interruptibility classifiers in a novel two-stage interruptibility model that first infers individuals' social roles to finally predict individuals' interruptibility preferences. The two-stage interruptibility model achieves a F1 score of 0.70. Finally, we examine the influence of multi-device data on social role and interruptibility classification performances. Our findings break new grounds and provide new insights for the design of future interruption management systems.CCS Concepts: • Human-centered computing → Ubiquitous and mobile computing; HCI theory, concepts and models.