Passive and ubiquitous sensing has proven to be a valuable way to detect and diagnose mental health conditions, assess health and well-being, and has shown promise in the future of work. In the past decade, the quality and sophistication of sensing devices: wearables, smartphones, beacons, etc. have dramatically improved. The widespread adoption of these devices by the public has given researchers the opportunity to conduct less intrusive research longitudinally on a larger scale outside of the laboratory. However, researchers face a number of challenges when designing effective studies. In this thesis we focus on three: data quality, appropriate sensor selection and deployment, and bridging the gap across disciplines when bringing research developed in the lab into the real world. First, researchers can only draw conclusions from their studies if the data quality is acceptable. But gathering data in the real world can be unpredictable. Conditions are less than ideal, and human error is a significant factor. The consequences of poor data quality can have devastating effects on modern health and well-being research. In this thesis, we present a thorough analysis of the problem of data quality in ubiquitous sensing studies and suggest methods to increase participant compliance. Second, as the number of sensing modalities increases and researchers find new ways of sensing a number of traits and behaviors, the challenge of carefully selecting the sensors to use and how to use them in a study has grown accordingly. To address this problem, we present a framework for conducting a cost-benefit analysis when selecting sensing modalities for a longitudinal study. This novel framework takes into account the costs to participants, researchers, and IT infrastructure of using a set of sensors in a study, and considers the benefits in the form of prediction performance or feature importance in including these sensors, in order to obtain an optimized subset of modalities to include in a study. Third, we found that often there are unchecked assumptions about how research that was conducted in the lab can apply broadly in the context of passive sensing in real-life settings. In particular, we find examples of this problem in the context of stress prediction from heart rate variability in real-life settings. In this work, we show why heart rate variability cannot be considered a proxy for stress, as it is often assumed in the field based on assessments under controlled conditions, exemplifying the gap between the engineering side of the ubiquitous computing field and the psychology side of the field. Additionally, we show to what extent heart rate variability from wearables can be a predictor of stress in real life settings. Conversely, we show that a cross-disciplinary approach can be used in the field in the problem of measuring sleep with wearables. We show an improvement on the accuracy and reliability of the measurement of sleep duration by fusing phone activity and wearable measurements through leveraging the advantages of each sensor with regards to the aspects that each can measure from sleep.