Triaxial accelerometers have been used in the study of human behavior for over twenty years, providing the basis for diverse methods of classifying and quantifying physical activity. Historically, implementation of these methods for extended follow-up periods and in large cohorts has been limited by the cost of wearable research-grade sensors, and by the behavioral modifications associated with their use. This limitation has been overcome by recent innovations in smartphone sensor data collection, which permit inexpensive observation of physical activity patterns via accelerometry at minimal inconvenience to participants. These innovations offer a potential basis for novel methods of monitoring patient behavior and delivering personalized interventions in a wide range of clinical and research contexts. However, these sensor data are subject to frequent interruptions arising from diverse habits among smartphone users. We present methods for the analysis of physical activity in the presence of missingness typical of smartphone sensor observations, with applications to data from ongoing studies at Harvard Medical School teaching hospitals.